Saltar al contenido

oh-my-hermes Historial de versiones

hope (rlaope) · Operating layer for Hermes Agent that runs coworkers in parallel lanes, plus the omher desktop app.

Descarga desde la webHerramientas de IAGratis y de código abierto 5,8 mil
Las descargas vienen de github.com · OpenNavo no aloja instaladores
  1. 4.0.9

    ActualizadoParche

    GitHub Release

    • Codex units can write inside the fanout fence on Linux and macOS. codex exec is read-only by default, and workspace-write cannot start Codex's own sandbox inside an enforced bwrap or sandbox-exec fence, so every write a codex unit tried was refused. When the fence is enforced with one of those backends, the codex argv now gets --sandbox danger-full-access before the prompt and the fence stays the write boundary. An explicit sandbox choice in the argv is kept; unfenced runs, other owners and native Windows (#2471) are unchanged.

    • omher: terminal splits without a title bar per pane. A split pane no longer carries a bar with its name, path and x, and the focused pane no longer gets an accent frame (the others an inset shade). Panes sit flush like Ghostty or iTerm splits: a 1px line is the only edge, the cursor and the dim show which pane has the keyboard, and a pane's zoom and x appear in its top right corner while the pointer is over it. The bottom panel's panes lose the same frame and shade.

    • The OMH TUI runs on native Windows, not only under WSL. Bare omh on Windows used to open Hermes' own terminal every time, because the OMH TUI needed termios. Where there is no termios its terminal layer now enters the Windows console instead: VT input and output modes with UTF-8 code pages (all put back on exit, a failed start included), console input records read on a thread and handed on as the same UTF-8 escape bytes a POSIX terminal sends, the window size polled in place of SIGWINCH, and a closed console ending the session as a hung-up pty does (closing the window or Ctrl+Break arrives as SIGBREAK, which now unwinds: console restored, gateway reaped). The gateway's wake pipe now holds at most one byte, so it never blocks without the non-blocking mode Windows pipes lack before Python 3.12, and the console peeks at it instead of select. omh doctor reports the real OMH TUI state on Windows. It needs Windows Terminal or a VT-capable console (an old conhost falls back with one line saying why); a Windows user still has to try it.

    • omher: ideas from kasaterm. Behaviour and numbers borrowed from 2rami/kasaterm (MIT), code written fresh.

      • Undo a closed terminal. Closing a running TERMINALS row can be undone for 10 seconds: the row leaves at once, a toast offers Undo (⌘Z / Ctrl+Shift+Z), and the same terminal comes back with its program and scrollback, in its split if it was in one. Settings › Terminal › Undo window after closing: Off / 5 s / 10 s (default 10 s).
      • Quieter notifications. Only approvals, questions and ends that went wrong ring, and the body is the bot's own first sentence (greetings skipped, code and markdown removed, paths shortened to their name, URLs to their host). A turn error only marks its chat; Settings › General › Also notify on errors (off by default) brings the note back. A clean exit 0 only marks its row, and so does the bell of a plain shell (tab completion, a stray \a); a terminal's bell notifies only while a coding agent runs in it. A wordless ring waits 8 s so the texted question that follows replaces it.
      • Disk space warning. When the disk holding your home folder drops under 20 GB (Settings › General › Warn when disk space is below: 5/10/20/50 GB or Off), a sidebar chip shows the free space and one notification goes out. omher checks once a minute with one system call and never deletes anything. The terminal agent mark also reads a node/python program's command line about 4× less often (60 → 13 reads a minute).
      • Close keeps omher running. Closing the window hides omher to the menu bar / notification area (on by default; Settings › General › When the window closes). Terminals and agents keep working; reopen from the tray icon or the Dock, quit from the icon's Quit omher or ⌘Q.
      • Keep awake while agents work. omher keeps the computer from idle-sleeping while a TERMINALS agent prints or a bot answers, until a minute after (on by default; Settings › General). Closing the lid still sleeps the computer.
      • Chat and split view. Chats get a Jump to bottom pill when you scroll up (with "· N new"; End or ⌘↓ / Ctrl+End also jump), and your last message stays pinned as one line at the top while you read a long answer. A zoomed terminal split now follows focus, and splitting, closing or moving a pane un-zooms, so typing never goes to a hidden pane.
      • Shell picker. The TERMINALS + menu's Shell row has a Shell ▸ list of the shells installed on this computer (Windows PowerShell, PowerShell 7, Command Prompt, Git Bash and WSL on Windows; the login shell and /etc/shells on macOS and Linux). Settings › Terminal › Default shell sets the one a plain Shell row opens, a row keeps its shell after a restart, and a missing shell falls back to the default with a note.
      • Theme import (macOS only). Settings › Terminal › Import from… brings the colours and font of an iTerm2 or Apple Terminal profile into omher as your own terminal theme. It reads the other app's prefs once when you press it, keeps a copy, and never changes the other app.
    • show-reel's cut-sheet lane checks a remake's genre and pace by numbers. Remaking a supplied reference now starts by naming its genre from a 5 fps sheet and 30 fps strips (the material, the motif that survives every cut, art piece or promo), and measures pace on the reference and each render: cuts per second, mean frame difference, motion inside shots and near-still share, with the targets written into the contract. The page also lists the devices that carry pace (2-4 frame bursts, scale motion with ease-out, shard or smear frames from the outgoing image, exit glitches, speed ramps, re-framed sub-shots), asks that a recurring motif act at every appearance, and says to add real feature chapters, not stretched shots, when a film gets longer.

    • A bot answers a mid-turn message before it acts on it. A message sent while the bot works (a steer) used to change only its next command, with nothing said, so the person could not tell it was read. OMH's reply line, which every session carries once in its system prompt, now asks for one line before the next tool call: what was taken from the message and what changes. The primer grows by 117 characters (compact rail 1550 -> 1667).

    • The run summary shows how much of the prompt came from cache. omh_run_summary adds a line after the token count, such as Prompt cache: 62% of input read from cache (1,234,567 tokens) (en/ko/ja/zh), and breakdown.cache_read_share. The share is cache reads over uncached input plus cache reads, which is how Hermes stores them. No line appears when nothing was read from cache.

    • A one-line list of three code changes is answered in the session, like a listed backlog. "migrate the store to sqlite, update the three callers, and add tests" or "프론트 버튼 추가랑 백엔드 엔드포인트랑 테스트 다 해줘" used to land on a lane picked from one word (toolbelt-readiness, browser-operator, backend). They now get the code_backlog answer: make the changes here, one after another, not as a dag-ulw graph. One bug's symptoms, one feature's fields, a list of questions, an errand list with a single code word, and work the request itself splits keep their routes.

    • A dag-ulw node that writes files is told to commit only its own paths. Its task text now says to stage the paths in its SCOPE by name and never to run git add -A, git add ., git commit -a or git stash, so parallel nodes in one checkout stop committing each other's work. Verify, review and read-only nodes get no such line. This is guidance; the write-scope guard on file tools stays the enforcement.

    • The pi-family coding lane runs on the omp CLI too. With only omp on PATH, a pi-family coding unit now runs omp --print --no-session [--model <id>] [--thinking <level>] -- <prompt>; when pi or its distribution host is also installed, those still win. omp has no --permission-preset, so no approval flag is passed (omp's default applies) and writes stay fenced by the worktree and fan-out confinement as for every host. Model discovery also reads ~/.omp/agent/sessions. omp keeps its sign-in in a SQLite store, so its login marker reads absent; that does not block dispatch.

    • Hermes Bug Hunter: code review keeps only findings it can show, and every review says what it did not check. code-review's finding validation gains a drop gate: a candidate stays out of the findings when it cannot be traced step by step from a concrete input to a wrong result, when the behaviour is intended or documented, when its trigger needs a privilege the user cannot have, when the harm is only theoretical, or when it is a style nit dressed as a bug. code-review, production-audit and verification-gate now end every result with a Not checked list (each entry with its reason, or Not checked: none only after full coverage), so a short findings list no longer reads as "the rest is clean".

    • Qwen4-Max is onboarded. Alibaba released it on 2026-11-16. Before this, every Qwen4 id (qwen4-max, qwen/qwen4-max, qwen4-max-preview) classified unknown and got generic discipline; they now classify qwen. qwen4-max has an exact contract read from docs.qwencloud.com: 1M context, hybrid thinking on by default, and a low / high / xhigh ladder with high as the default. The vendor maps minimal, medium and max onto those three without an error, so OMH raises nothing and the contract records the table. Cost estimates use the Singapore list price, $1.50 / $4.50 per million tokens with cached input at $0.075. The Qwen calibration is now version-aware: the family block still describes the non-thinking Qwen3-Coder, and Qwen4-Max gets its own block that does not forbid thinking. That block is editorial and unmeasured. No shipped chain names Qwen4-Max, because Qwen3.8-Max held no slot either; to try it, use omh model-chains set <category> "qwen4-max:high, ...". The Token Plan preview is recognized only, with no contract and no price. Hermes through origin/main fe8b920cae sends no effort on its DashScope profiles, so the model runs at high there whatever effort is routed.

    • A cheap dag-ulw node that hits a hard judgement asks for a stronger model instead of guessing. When a node's child cannot settle a design choice or a root cause, it stops before editing and writes ESCALATE: <question> into its handoff; the node is settled escalated and offered again one category up (quick, unspecified-low, unspecified-high, ultrabrain), at most twice, briefed with the question the first child stopped on. Nodes already on ultrabrain, review or deep never get the option, nor does a node pinned to an identity, which keeps that identity's own model chain.

    • Hermes Bug Hunter: a trap benchmark counts the bugs a reviewer invents. benchmarks/review-false-alarms holds 11 traps (changes that look wrong and are correct: SQL built from a constant allowlist, eval over a literal table, an index guarded two lines up, a lock the caller holds) and 6 controls with one real, reachable bug each. A reviewer ends with VERDICT: clean or VERDICT: bug <path>:<line>; a control counts as hit only within ±3 lines, and a missing or garbled verdict is its own column, never clean. The table prints the false-alarm rate and the hit rate side by side for omh-on and omh-off, and OMH wins only with fewer false alarms and no fewer hits. bench.py --dry-run grades a recorded fixture with no model; live runs use hermes -z on two profiles.

    • The Start Guide lists long-term memory among the main features. It moves from the closing list to a fourth card in "What OMH does" (owner request, 2026-11-18): nothing is remembered silently, and the next session gets only the memories that fit its task, within a token budget. The closing list is now one row of three.

    • A fail-closed leak guard keeps private identifiers out of the public repo (Hermes Bug Hunter). CI now scans every tracked file on each pull request and main push against a committed denylist of sha256 hashes, so the list never carries the names it protects. A hit prints path:line and the hash; a missing or empty denylist fails the job instead of passing. tools/qa/identifier_guard.py --staged runs the same check as a pre-commit hook, and --hash "name" prints the line to add (see CONTRIBUTING.md).

    • Four OMH tool descriptions are shorter, and the model still picks the same calls. omh_todo, omh_loop, omh_document_plan and omh_delegate_route say the same rules in fewer words, so each costs fewer tokens the first time Hermes loads it (24,943 -> 23,718 characters together; the registered OMH schemas 63,324 -> 62,099). An A/B of the first tool call on 11 scenarios (Claude Haiku and Sonnet, 3 runs per arm) passed 33/33 on both the old and the new text. A first draft that also cut "a route set in an earlier turn runs on the previous model" and "OMH mirrors each todo_list write" lost passes and was restored.

    • Gemini prices re-read ahead of Google's two increases. Google's pricing page, re-read on 2026-11-18, still lists Gemini 4 Argon and Gemini 4 Carbon at $2 / $10 per million tokens through 2026-11-30 and $4 / $20 from 2026-12-01, and Gemini 3.8 Flash and 3.6 Flash at $0.75 / $3.75 through 2026-12-31 and $1.50 / $7.50 from 2027-01-01, cached input moving with each. Hydrogen stays at $1.50 / $7.50. Cost estimates keep today's prices until each date; the price rows now point at the current page, and Argon's row is pinned by the same list-price test as Carbon's.

    • A dag-ulw re-run can skip the nodes the last run already confirmed, and only when asked. omh_dag action=define takes an opt-in carry_forward with the base revision. A node an earlier run of the same goal at that base accepted, with the same id, prompt, write scope and dependencies and its handoff still on disk, is recorded carried and not handed out; its dependents read the earlier handoff. Verify nodes, nodes that failed or were blocked, changed nodes and anything after a node that runs again always run. define and status list each carried node with the run it came from, the status line says carried N of M nodes from <run>, and the goal verdict still waits on the verify node, which is told which nodes were carried.

    • omher.app has a Start Guide at /guide/, linked beside Docs in every header (owner request, 2026-11-18). One page for someone new, with plain section titles that say what each part covers: what OMH does in three cards (a model per kind of work, dag-ulw, expert skills), install in three commands (curl … | sh, omh setup, omh model), categories and fallback shown on three real chains, plan versus dag-ulw as a to-do list versus a flowchart (an animated graph: two steps at once, fix, test, a separate goal check, and the loop back when it fails), how OMH stays current (a new model is checked before it joins a chain, the plugin loads into stable and main Hermes every night, reconstructed skills record the upstream commit they were reviewed against), the omh TUI, and omher desktop as an ADE for bots with a terminal panel. Everything else (ulw-maestro, memory, evidence labels, LSP) sits in one short list at the end. It reads in English, Korean, Japanese and Chinese through the site's language switch. The model names on the page come from the shipped recommendation catalog through tools/render_site_guide.py, which Pages re-runs on every deploy, so a model onboarding never leaves the guide naming a retired model; tests/test_site_guide.py pins the sections, the installer commands, four-language coverage of every key, the header links and the catalog match.

    • GPT-6.2 Astra and GPT-6.2 Sol take every slot GPT-6 Astra and GPT-6.1 Sol held (OpenAI released both on 2026-11-17; owner decision, 2026-11-18). In the Hermes chains, the Codex lane table and the shared final order (Claude Opus 5.5, then GPT-6.2 Sol), each new id sits where its predecessor sat, at the same effort, and no chain changes order. ultrabrain is now Claude Fable 5.2, GPT-6.2 Astra, then Claude Opus 5.5; forensics and the Codex ultrabrain, architect, deep-work and forensics rows lead with GPT-6.2 Astra, and every Sol slot names GPT-6.2 Sol. A child with no route record is labelled by its new pair: GPT-6.2 Astra at xhigh reads forensics, GPT-6.2 Sol at xhigh review. omh coding model-contract --model gpt-6.2-sol (or gpt-6.2-astra) prints each record from OpenAI's pages: the API rejects none, so a request for off, none or minimal is raised to low on record, and tool calling needs the Responses API. Astra lists at $8 / $40 per million tokens against GPT-6 Astra's $10 / $50, Sol at GPT-6.1 Sol's $2 / $10, both with cached input at 5%, and cost estimates use those prices. Both get GPT-6 Astra's calibration text unchanged; OpenAI's "persist until the task is complete" prompts are not used. GPT-6 Astra and GPT-6.1 Sol leave the shipped chains but stay recognized, priced and routable, so a chain you set yourself can still name them. Hermes builds up to origin/main of 2026-11-17 do not know the 6.2 ids yet and clamp max to xhigh on the Codex route.

    • OMH starts faster in a fresh Hermes process and its tools cost fewer tokens to open. The chat router no longer imports the whole loop engine for one loop-command check, and the dag-ulw cue patterns compile on first use instead of all at import, so the plugin's start before a one-shot chat -q, a cron run or a delegated child's first reply drops by about 70 ms (register, import warm-up and first routed turn together: median 327 -> 254 ms over 10 interleaved cold starts against the previous main; the first turn pays about 20 ms of it back compiling the patterns it reads). The OMH tool schemas Hermes loads on a tool's first use shrink from 72,940 to 63,324 characters (-13%): the observation field on 20 tools stops describing metadata a model never fills, and the profile-override fields say their refusal in fewer words. No tool, parameter or rule a model acts on changed.

    • An OMH TUI reply no longer hangs on a rail. The ⚚ Hermes · <model> header starts at the left edge and the reply's text sits two cells in, with no ┃ beside it, so a reply reads apart from your prompts and the tool calls, which keep their rails. The two cells the bar took go to the reply's text, and a copy still leaves the margin out.

    • An OMH TUI approval opens on allow once, not deny. Enter alone now allows that one command; ↑ then Enter, d, n, Esc or Ctrl+C still deny and stop the running turn. Keys typed for the composer as the approval appears still land nowhere, since every question ignores keys for a moment after it opens, and a request that does not offer allow once still opens on deny.

    • A confirmed paired run now compares Hermes alone with Hermes plus OMH for real. omh coding paired-run dispatch --confirm-dispatch prepares one disposable Hermes home per arm: the variant gets the packaged OMH plugin and exactly its frozen exposed_skills, the baseline neither, and both run the same file,skills flags. A gateway the harness starts and stops serves each home, and an untimed warm-up keeps first-start work out of every cell's seconds. The dispatcher signs the arm setup it reads from disk into each receipt, and fan-in refuses a cell whose setup contradicts its arm. New --provider-base-url, --provider-api-mode and --provider-key-env declare a custom provider whose key arrives only by environment. Every other Hermes child keeps safe mode.

    • A paired-run receipt now says what each cell's child did, not only how long it took. Before the arm home is removed, the dispatcher reads the child's own rows from that home's state.db and signs a metadata-only cell_trace into the receipt: API calls, input/output/cache/reasoning tokens, the ordered tool names, how many skill_view calls opened which skills, and the system prompt's and first request's size in characters. No prompt, reply or tool argument is kept, a caller cannot supply or keep one, and a store that is missing, locked or holds no row of the child leaves the trace null rather than zero. The execution report lists each cell's verified trace.

    • omh_gather_evidence can check the task's own test, and a run of no tests is no longer a pass. A unittest probe may now name test modules, classes or methods inside the working directory (python3 -m unittest tests.test_x, tests/test_x.py), limited to what default discovery would run there, so a repository whose other tests still fail no longer hides whether the change at hand passes. Ran 0 tests, and exit 5 from unittest or pytest, now read passed: false with reason: no tests ran. The tool's line in Hermes' deferred-tool catalog names tests and shows a whole call, so a model looking for a test runner can call it from the listing. Every existing refusal stays.

    • A paired run can give both arms Hermes' shell tool, and only both at once. omh coding paired-run dispatch --arm-toolsets file,skills,terminal runs both the baseline and the variant with file,skills,terminal instead of the default file,skills, so an arm whose skills say "run the tests" is no longer measured against an arm that cannot run them. Only file, skills and terminal are accepted, file and skills are required, and --yolo is never passed, so Hermes' dangerous-command approval stays on. The toolsets each cell ran with are signed into its receipt's arm_setup; fan-in refuses a cell whose toolsets differ from the run's declared ones and a matrix whose arms differ from each other. Every other Hermes child keeps safe mode and --toolsets file.

    • A paired run can let both arms delegate, so OMH's job decomposition can be timed against Hermes alone. --arm-toolsets file,skills,terminal,delegation gives both arms Hermes' own one-level delegate_task. Both arm homes then carry the same child settings (up to 6 children instead of Hermes' one-shot default of 2, no deeper children, no per-child worktree) and run every child on the arm's own model; the variant's OMH routing chains are pinned to that model too. Each cell's signed cell_trace now says how many delegate children ran, their API calls, how many ran at once, and which models the cell's sessions used, and fan-in refuses a delegating cell whose trace is missing or names another model. A delegating cell's usage reports no cost rather than one that may count the children twice; tokens and API calls still sum. Without delegation nothing changes.

    • A listed backlog of separate code changes runs as a dag-ulw graph. A request that hands over four or more list items, each its own change, and asks in its own words for all of them to be made in code ("these six issues are open: fix all of them", "아래 버그들 전부 고쳐줘") is dispatched to dag-ulw by the router, and Hermes' per-turn hint starts it with each item as its own build node and a verify node after them. Before, such a backlog went to whatever lane one item's word scored (a "stock" lot took a six-bug job to live-info-operator) and the skill's own size cue never ran. A list of three or fewer, one bug, a list to review, prioritize, explain, plan or file as issues, one held or put off, one handed to a named coding agent, a fenced list, failing test files by path, and chores or reading stay where they were. The dag-ulw skill now also says what to do where delegate_task returns its children inside the call (a finite one-shot session): settle and dispatch the next layer in the same turn. "current" said of the workspace (the current directory, branch or file) is no longer read as a live-information cue.

    • Another OMH TUI no longer exits when Hermes' gateway restarts, and typing shows while it reattaches. Where no service runs the gateway, hermes gateway restart (also /gateway restart from another TUI) stops it as hermes gateway stop does, so its record says stopped with no restart asked, and starts the new one a few seconds later. Every other TUI attached to it read that record as a stop and left at once with the resume hint. A stopped record is now watched for up to 30 s, with no try: a new live record reattaches the session as a restart run elsewhere, a failed start ends with Hermes' reason, and a record still stopped then leaves as before, so no try ever starts a gateway the person stopped. While any respawn or reattach waits, text and its editing keys go into the composer at once and show on the next frame; before, every key was held unseen until the gateway was back, which read as a frozen TUI. Enter, a paste, a key that opens a list and every key after one of those still wait for the gateway, in order.

    • A request to fix a code defect now gets OMH's debugging method in the turn, not only a skill name. Hermes' skills index tells a model not to open process skills such as debugging, so omh-app-debugging and omh-verification-gate stayed closed on exactly the turns they exist for. On a turn that asks to fix code (a change verb such as fix, debug, "make ... pass", 고쳐, with a code object such as a test, a traceback, a source path or a function), the plugin now adds a short block once per session: reproduce first, trace the bad value to where it is produced, try other valid inputs, run the whole suite. Questions, reviews, features and fixes of anything that is not code get nothing new.

    • OMH TUI: Ctrl+Enter with an image attached sends it now. Before, a prompt with an image never went into a running turn: Ctrl+Enter left it as a ⧗ queued row over the composer until the turn finished. Now Ctrl+Enter stops the turn and sends the prompt with its image the moment the turn ends. Meanwhile it shows above the working row as your prompt, not a queued row. Esc again gives it back to the composer. Enter with an image still queues for the turn's end, and Ctrl+Enter with text alone still steers without stopping anything.

    • OMH TUI: the working spinner stays in the theme colour on terminals that draw ✳ as an emoji. Zellij Desktop on macOS painted its two ✳ frames as a green emoji, so the spinner flashed green twice a cycle. Those frames now ask for text presentation (U+FE0E) and stay one cell in the accent colour. Contributed by @DanRioDev (#2923).

    • An OMH update asks one restart question, not two. The TUI still asks once whether to restart on the new version. The reopened TUI no longer opens a second question about Hermes' gateway: when the gateway still runs the old OMH plugin code, it says so in one line and names /gateway restart, which loads the new code.

    • What you type while the OMH TUI loses its gateway is never lost. When the gateway does not come back and the TUI leaves (its restarts spent, a stop it watched, a reattach given up), what you typed and never sent, the draft and anything typed while it waited, is printed under its last line instead of going with the screen.

    • omh model-chains and /omh-model now look like every other OMH picker. The chain picker paints through the OMH TUI's theme roles instead of the old skin palette: no selection band, the cursor row is ❯ with its name bold in the menu colour, the rules and the effort bar's filled cells take the launch look's colours, words are neutral greys, and hue is kept for what a cell says (customized, edited, enabled, disabled, an effort off its model's list, warnings). It follows /theme, the saved launch look, a light background (COLORFGBG on the command line, the measured background in the TUI) and the terminal's colour depth. Text, columns and keys are unchanged; only the cursor mark moved from ▍ to ❯.

    • A code backlog that names a monthly report is no longer read as a schedule. A period word such as monthly, weekly or nightly counts as a recurring schedule only when it says how often the work runs ("fix these monthly", "rebase them nightly, then report"). Naming a thing ("the monthly report", "a weekly digest module") no longer stops a listed backlog of code changes from running as a dag-ulw graph.

    • DeepSeek coding units write their answer in the shape the task asked for. The deepseek calibration block gains one sentence: the answer file is one JSON object keyed by the task's own field names, not a bare list and not a wrapper key of the model's own. On og deepseek/deepseek-v4.1-flash-eco the search tasks of benchmarks/live-model-tools/v1 had been writing [...] or {"results": [...]} where the task asked for locations; with the sentence the six fixable templates passed 90 / 90 against 76 / 90 for a control run in the same window (3 seeds x 5 reps, gate fixed before the run). The held-out product-ab corpus showed no difference either way (19 / 40 against 20 / 40, one run per task). The block also reaches DeepSeek V3.2, V4 Flash and V4 Pro, which were not measured; V4.1 Flash keeps its own override, which gains the same sentence in the entry below.

    • DeepSeek V4.1 Flash units also write their answer in the shape the task asked for. The V4.1 Flash override replaces the deepseek family block, so it did not get the answer-shape sentence the family block gained; it now carries the same sentence verbatim. On og deepseek/deepseek-flash the six fixable live-model-tools templates passed 85 / 90 with it against 76 / 90 for the shipped override in the same window (3 seeds x 5 reps), at 0.97x the tokens.

    • A subagent that is not part of the running dag-ulw run no longer shows up inside it. A subagent sent without the run's [dag:<run>/<node>] marker (a hand-made repair, a child of an earlier cancelled run) used to sit in layer 1 of the DAG pane, where Desktop labelled it "Start node". It now shows below the run as Outside DAG: Desktop cards, the text tree and the TUI block alike. Usage and cost still count it, and it no longer holds a finished run's one-line summary open. While a run is active, such a dispatch is stopped once per turn with a pointer to omh_dag (ready, settle, repair, extend); sending it again goes through. Contributed by @yqm-307 (#2914).

    • omh setup ends by offering a guided model setup in the OMH TUI. On a terminal, the last question is "Set up your models now?" (default Yes). Yes opens the OMH TUI in a new session with a fixed onboarding prompt as its first turn, and Hermes walks you through providers, keys, a main model and at most one fallback by conversation; if Hermes has no main model yet, your own hermes model runs first. No prints how to come back later. The prompt is sent only under the new omh --onboard model-setup, never on a resume, and omh setup --model-setup --fallback-model PROVIDER/MODEL previews and writes one Hermes fallback_model under the same digest-bound approval.

    • A dag-ulw node can pick a built-in role or a custom identity by name. omh_role action=list now shows built-in roles and custom identity profiles together, and a node sets identity: builtin:<name> or custom:<name> independently of its role and category. The chosen definition supplies the child's prompt guidance and its model chain; retries count per node, so two nodes no longer use up each other's place in the chain. A missing, disabled or changed identity stops the dispatch instead of quietly swapping in another one. Contributed by @yqm-307 (#2939).

    • omh_delegate_route set refuses a model id pinned to a provider that does not serve it. A set with model=ANT-claude-opus-5-5 provider=openai-codex used to reach config.yaml, and every delegated child after it died on its first call with HTTP 400 from the vendor. When routing/model-providers.json records that exact model id only under other providers, the call now returns an error naming the alias, every provider that serves the id and the two valid ways to route it, and writes nothing. Aliases may still be pinned to another provider, and ids no route records pass unchanged. Contributed by @rafaelmoreiraai (#2915).

    • omh setup and omh uninstall can edit a config that Hermes' installer just wrote. Hermes' config template has comment lines with quotes in them (# order: ["anthropic", "google"]), and OMH's config editor read those as quoted keys it must not touch. On a fresh Hermes, setup therefore never recorded plugins.entries.omh.settings.omh_home, and omh doctor warned about it with a fix that hit the same refusal. Uninstall could not take back plugins.enabled or memory.provider. Comments are now ignored by that check; a quoted key in the YAML itself is still left alone.

    • dag-ulw waits for a helper Hermes will resume instead of running its node twice (Hermes 0.23.2 delegation.auto_resume). Hermes now continues a helper a usage limit or a restart stopped, under the same subagent id and session. The delegation ledger took that second subagent_start for a duplicate and dropped it, so the helper kept its old error stop and a node whose resumed run finished could not be settled accepted (child_reported_failure); after a restart omh_dag also marked the node lost and handed it out again beside the resumed helper. A repeated start with the same session now counts as a resume (the old stop is dropped, the process it resumed in is recorded, and only a stop since the resume answers for it), and omh_dag reads Hermes' own resumable_subagents record read-only: a node whose helper Hermes promised to resume stays running with waiting_resume (at, attempt), is not offered again or named stuck, refuses a settle with child_resume_pending until it reports, and every reply says not to dispatch it again. A Hermes without that table, or a helper it gave up on, keeps the restart and failure rules as they were.

    • The OMH TUI can proofread a draft on a local LanguageTool server. Press Alt+P in the main or clarify text field and issues are underlined; Alt+P on an underlined word offers replacements, Enter applies one (normal Ctrl+_ undo), Skip or Esc closes, and nothing is ever sent or rewritten on its own. Automatic checks after 600 ms of quiet are off by default and turn on with OMH_PROOFREAD=1. Only plain HTTP to a loopback address is accepted (OMH_PROOFREAD_URL, default http://127.0.0.1:8081/v2/check); masked fields, commands, code, images and paste chips are never sent. OMH does not install or start LanguageTool (brew install languagetool). Contributed by @DanRioDev (#2948).

    • A listed backlog of code changes is made in the session, not split into child agents. A request that hands over four or more separate code changes and asks for all of them is still recognized, so it no longer lands on an unrelated workflow because of one item's word. It now gets an answer in the same session (code_backlog): the changes are made one after another, with no dag-ulw graph and no child agents. In paired runs on a five-part, six-thousand-line job, splitting such a backlog finished later than Hermes alone at the same quality. Asking for dag-ulw by name still runs the list as a graph.

    • OMH skills load whole on every host they ship to, and omh skill-eval checks any skill folder the same way. Every situation clause now ends inside Hermes' 60-character index window. The portable descriptions (.agents/skills, .claude/skills) are the situation clause plus the start of what the skill does, with the trigger list kept in the body, so the pack fits Claude Code's and Codex's 8,000-character skill list. A portable body over OpenClaw's read cap moves its Catalog Metadata into references/details.md, and the eight toolkit skills name toolkit/README.md in SKILL.md itself. Measured with NVIDIA SkillEvaluator 1.0.1 on a5649293 → this branch: Hermes clipped situation clauses 20 → 0; .agents pack descriptions 42,581 → 7,442 characters; .claude copy 16,474 → 7,442; OpenClaw over-read-cap bodies 7 → 0; toolkit README named 0/8 → 8/8. The new omh skill-eval PATH (offline, stdlib) reports per host whether Hermes, Claude Code, Codex, OpenClaw, OpenCode and Copilot CLI load a skill, load it degraded, or skip it, with each finding's source, and marks a row unverified when no recorded host fact covers the input (unparsed YAML, say); --catalog checks a whole folder, category folders included, and --strict exits 1 on any degraded or unverified row. The omh-skill-eval skill runs it before a skill ships, and a test now runs the same checker over the three shipped packs (#2995).

    • show-reel has a cut-sheet lane for films written as a shot contract. A new opt-in reference, cut-sheet.md, covers cut-driven sequences of 30 s or less handed to a builder or a text-to-video generator, and remakes of a supplied reference video. It opens with one contract line (ratio, duration, exact cut count, fps, palette), locks the subject in observable markers with a never-redesign clause repeated in every prompt chunk, and writes one timecoded row per cut, each compositionally different from its neighbours with a named exit transition, inside a closed editing vocabulary. Sound is written with the rows (a named peak, a cold ending, SFX onsets aligned to the cut), a remake measures the reference's cut times and audio onsets before any row is written, and a delegated build is reviewed in numbered, measurable FIX rounds. The brief, rights and the short-promo one-scene rule still apply.

    • The OMH TUI's proofreader keeps a local dictionary. On a spelling issue, Alt+P now offers Add to dictionary beside the replacements: the word's spelling marks hide in every OMH TUI on that OMH home, and Alt+P on it later offers Remove from dictionary. The picker's Skip row is gone (Cancel and Esc already did the same). Grammar corrections are never hidden by a saved or unsaved spelling, the word list is read and saved off the screen's thread, and a damaged dictionary file is never overwritten. Words are kept in <omh_home>/proofread/dictionary.txt; a word with an apostrophe or hyphen gets no dictionary action. Contributed by @DanRioDev (#2996).

    • The OMH TUI's waiting row says what it waits on. Between turns the row above the composer named nothing (✻ Waiting… (33m 47s · 1 agent · terminal)). It now names the longest wait from structure only: a background process by the words in command position of its command (Waiting for CI with the PR number, Deploying, Testing, Building, Installing, Rendering, Pushing, Downloading, Sleeping with the time left), a process woken on a match by its first pattern (Watching for "READY"), a dag-ulw run by its name with its nodes done and failed, and a single child as Reviewing, Researching or Writing with its model. Other waits are counted after it (+1 test run · 1 agent); a narrow terminal falls back to the old row. As the wait ends, one muted row says so for 3 seconds (✓ Review finished, ✗ … failed, · Tests ended for a process, which has no exit code). During a turn, ? waiting on you now reads ? waiting for your approval or ? waiting for your answer.

    • A tool call no longer re-parses an unchanged profile config.yaml to bind its OMH home. Every pre_tool_call/post_tool_call callback resolves the profile's OMH home, and each resolution ran Hermes' validating parse of the whole config.yaml. The plugin now keeps that one validated parse per file, keyed on a SHA-256 of the bytes it read on that call, and reuses it only while the file still holds exactly those bytes and the same host validator is installed. A parse is kept only when Hermes' own YAML reader turns exactly those bytes into the same mapping, so a file swapped and restored while the validator ran (A, then B, then A again) is never remembered as B's parse. An absent or unreadable file, a config backend other than the local file, or a write during the parse goes through Hermes' own validator. Path resolution, the managed and effective config comparison, launch fallback and secret ownership still run on every call. Measured on a Mac16,7 (Apple M4 Pro, macOS 26.6.2, Hermes' CPython 3.11.16) with Hermes' real file-backend validator and a 46,614-byte synthetic config.yaml, 12 alternating fresh processes per arm, 20 warm calls each, medians of per-process medians: a warm resolve_homes() went from 5.76 ms to 0.22 ms, and the two bindings pre_tool_call resolves before any gate from 5.75 ms to 0.24 ms; the cold first call, which now also parses the bytes it keyed on with Hermes' reader, went from 19.8 ms to 25.4 ms. The contributor's version, which cached the whole binding, measured 0.03 ms warm in the same harness, but it could hand back a stale home after a same-size rewrite or a changed path alias. Thanks @zangetsu787 (#2973).

    • A delegation route write no longer drops a setting saved while it ran. OMH built the new config.yaml from the copy it validated, so a hermes config set landing in between (observed: context_file_max_chars reset to unset, repeatedly) was overwritten. The route writer now re-reads the file just before replacing it and applies only the three route keys (delegation.model, provider, reasoning_effort) to that read, retrying up to five times while the file keeps changing, and leaves the file untouched when nothing would change. Every check (a fallback's or restore's expected route, the endpoint guard, the unsupported-mapping refusal) runs on that same read; a change to the route keys themselves is refused rather than overwritten, and the baseline put back at turn end is the route the write really replaced. This is best effort: Hermes' CLI takes no OMH lock, so an edit landing in the last moment between OMH's final check and the rename is still lost. Fixes #2998. Reported and first fixed by @OptionalCoin in #2999.

    • Hermes Bug Hunter: the false-alarm benchmark gets traps a strong reviewer actually falls for. In the first live run (3 runs per case per arm) both arms raised 3 false alarms in 33 trap runs, all on the one trap whose proof sits outside the hunk (a lock the caller holds); the other 10 traps were judged clean every time, so the set could not show a difference. benchmarks/review-false-alarms adds 12 traps of that shape across Python, JavaScript and Go, each with the files that prove it safe in context/ (a sort done two files upstream, an Atoi error a regex rules out, a map written only before the server starts, a depth the settings loader bounds), and 4 controls whose bug also shows only with a file outside the hunk, so an over-cautious reviewer still loses hits. The corpus goes from 11 traps and 6 controls to 23 and 10.

  2. 4.0.8

    Parche

    GitHub Release

    • The OMH TUI shows a compression as one gauge row, and a prompt typed meanwhile goes on by itself. /compact (or /compress) no longer prints the long "Compressing: Hermes summarizes the older turns…" notice above a second spinner row. One row now reads ✻ Compacting conversation… ━━━━━━──── (30s), with a blank row between it and the composer. Hermes sends no progress for a compression, so the gauge follows the time waited, never fills on its own and shows no percent; Hermes' ✓ Compressed: … report closes it. A prompt sent during the compression waits in the ⧗ queue (Ctrl+Enter has no turn to steer, so it queues too) and is sent by itself, oldest first, once the compression is answered, refused or given up on.

    • omher.app shows the bot faces as a moving scene under the desktop app. Between the omher block and the OMH hero there is now a band with nine soft 3D characters in the style of omher's built-in bot faces, bouncing and blinking on a 4-second loop, with one line of copy in English, Korean, Japanese and Chinese and one Download omher link. The link downloads the build for the visitor's system, the same file as the first download button above it, and opens the releases page when the system is unknown. The loop pauses while it is off screen or the tab is hidden, and shows a still picture under reduced motion. On phones the copy sits above the picture. Four of the nine faces ship in the app today (blob, ghost, mochi, star); the other five are in the picture only.

    • The OMH TUI sends its desktop notification only while the terminal window is not in front. A turn that ran 10 s or more, a question and a finished background process each rang the bell and, in iTerm2, WezTerm, Ghostty, foot and rxvt, raised a notification (OSC 9 or 777) even with the window focused, where the person already sees it. Once the terminal has reported focus and has it, only the bell rings; with focus gone, or from a terminal that never reports it, both go as before. OMH_TUI_BELL=0 still keeps everything quiet.

    • A resumed session no longer shows Hermes' own turns as your prompts. When omh --resume or /resume redrew a session, every user-role row in Hermes' history was drawn as a ❯ prompt, including the ones Hermes wrote itself: an async delegation's results ([ASYNC DELEGATION BATCH COMPLETE …]), a background process finishing, an internal notice, a resumed turn, a model switch. They now draw as one dim ↳ row with the title Hermes stored for them (↳ Subagent Tasks Completed: Review the TUI branches), as Hermes' own TUI draws them. Your own prompts, skill lines and steers are unchanged.

    • A successful shell run in the OMH TUI shows the end of what it printed. A terminal or execute_code call that exited 0 (a long build, a test run piped through tee) showed its first three lines and cut the summary or the errors at the bottom. It now shows its last three, with … +N lines above them, as a failed call already did; every other tool (read_file, search_files, web_extract) still shows its first lines, and Ctrl+O details hold the last 40 lines of a shell run instead of the first 40.

    • Hermes' own terminal opens as Hermes' own when omh falls back from inside an OMH TUI session. When the OMH TUI could not start, omh handed Hermes the shell's whole environment, so an omh run from a shell that had inherited the OMH TUI's launch state (its HERMES_TUI_DIR among it) could point Hermes' launcher at the OMH frontend again. The fallback now drops the names the OMH TUI launch sets for its own frontend, on both the hermes --tui and plain hermes paths; OMH_HOME and the chosen HERMES_HOME still reach Hermes.

    • Bare /model in the OMH TUI lists the models you last switched to first. Each switch Hermes confirmed puts the model and provider Hermes confirmed (an alias you typed is kept as the model it resolved to, and a /model without --provider as the provider Hermes switched on) at the front of ~/.omh/tui/recent-models.json (at most five, no repeats); a switch Hermes deferred or asked to confirm records nothing. With an empty filter the picker opens with those still on offer under Recent, then the whole list under All models; typing filters as before, with no group.

    • The slash menu and the pickers match letters typed in order. /rld now offers /reload-skills, and snt finds a claude-sonnet-… row in /model. Names and rows that start with or contain what you typed keep their place and order; the in-order matches follow, best fit first (word starts and runs of neighbouring letters rank higher). /sessions and /rewind filter the same way, and an empty filter is unchanged. After opencode's fuzzy dialog-select.

    • The OMH TUI renames and deletes sessions from /sessions. In the session picker, Ctrl+R turns the list's last row into the highlighted session's title (Enter renames, Esc cancels) and Ctrl+D asks on the row, then deletes on a second press; the session you are in is never deleted from there. Hermes' stdio entry gets session.title / session.delete; attached to a shared gateway, the gateway child sends them as the owner's revision-fenced session.mutate rename / delete, as Hermes' Ink TUI does. A refusal is said on the row.

    • A dag-ulw node can be read-only. omh_dag define, repair and extend take read_only: true on a node; its child's write_file and patch inside the run's workspace are blocked, except its own handoff report, with a message telling it to put its findings there. Before, a review, verify or explore node with no write scope could change any file. A read-only node that also declares a write_scope is refused (read_only_with_write_scope); a node without the field keeps today's rule, and shell writes stay unjudged.

    • An active workflow's reminder rides a session once, not on every turn. [OMH Active Workflow] was appended to every turn's message while a workflow was active, and Hermes replays each turn's injection, so N turns carried N copies. It is now sent only when the history Hermes replays does not already end on that exact line: after a compaction dropped it, or when the workflow changed. When the workflow ends, one line says the earlier reminder no longer applies. Image turns count too: there the reminder is read after the ask Hermes stamps on the turn, so a quoted reminder never does. The omh_active_workflow metadata still comes with every call.

    • A toolcall rule can ask you instead of blocking. With "effect": "ask", a matching call goes to Hermes' approval prompt with the rule's message, keyed omh-toolcall-rule:<name> so [a]lways covers that one rule. Where nobody can answer (a delegated child, hermes chat -q, a webhook or API platform) the rule blocks as before. An ask never overrides OMH's own refusals (evidence-store writes, edits to OMH-installed skills, a dag-ulw child writing outside its scope, and the rest). omh ops toolcall-rules-validate refuses any other effect. See docs/TOOLCALL-RULES.md.

    • The OMH TUI shows a background wait on its own runtime again. With omh theme runtime own (the default), the HUD read the bots' Hermes home rather than the TUI profile the session runs on, so it never found the session's background terminals: no ✳ Waiting… (… · terminal) row stood over the composer between turns and no background row sat on the info line, though the process ran and Hermes would wake the session for it. The HUD now reads the home HERMES_HOME names; the bot board alone keeps the bots' home.

    • Claude Code lists a description for every OMH skill it may pick on its own. Claude Code fits skill descriptions into a listing budget of 1% of the context window and drops the rest. With the whole pack in .claude/skills, 62 of 123 OMH skills (omh-plan and all eight ulw-* among them) were listed by name only. The Claude copy is now its own generated pack, .claude/agent-skills/, which .claude/install.sh, install.ps1 and omh install --target agents write to .claude/skills. It keeps all 123 skills and cuts each description to its lead, at most 160 characters (a cut one ends with …); the trigger phrases stay in the portable pack. On Claude Code 2.1.322 with claude-opus-5-5 and an empty config directory (2026-11-13), /context shows 123 of 123 OMH skills with a description in an 8.5k listing, where it showed 61 of 123 in 9.9k. In four task prompts, what the model loaded on its own did not differ from the old pack in the runs made (three prompts loaded the fitting skill in every run on both; the fourth in 0 of 1 runs before and 2 of 8 after); see docs/AGENT-SKILLS.md#claude-code-skill-listing. A listing already full from other skills stays full. .agents/skills and the Hermes skills keep their bytes. An existing .claude/skills mirror from omh install --target agents reports stale until that install is rerun (a user-scope one is also refreshed by omh update). tools/claude_context_skills.py counts the skill rows of a saved /context output.

    • The OMH TUI no longer pins ↳ you: … above the composer. While a long turn ran, the prompt you sent was echoed as a ↳ you: row over the working row once it had scrolled off; the owner found it noise. The prompt stays in the transcript and scrolls off with it (scroll up to read it), and the two rows the echo and the blank row over it took go to the turn's own output.

    • Enter steers a running turn in the OMH TUI, as Claude Code does. Text you type and send with Enter while Hermes is working now goes into that turn at once (session.redirect, read at the next tool call) and shows as your prompt above the working row, instead of waiting as a ⧗ queued row with a ctrl+enter to send now hint. A slash command, a ! line, a prompt with images, a prompt sent in plan mode, one sent behind lines already queued, or one sent before Hermes has started the turn still queues for the turn's end as before; Ctrl+Enter, Ctrl+J and /steer keep working.

    • A question's Type something. row is where you type. Picking it (its number, or Up/Down then Enter) or typing a character that picks nothing opens the answer field on that row of the question, as Claude Code's question dialog does, instead of moving the typing to the composer under the question; the caret stands on the row and a long answer scrolls there. Up and Down still go back to the choices with what you typed kept.

    • A drag in the OMH TUI copies clean by default, long commands as one line. The main screen takes the mouse again (the owner's choice of 2026-11-11, back from 2026-11-06): a drag copies the transcript without the ┃ rail, and rows the TUI wrapped to fit the screen are joined back into the line they were written as, a space where the wrap dropped one, nothing inside a long word or path, code exactly as it is. A long command or path pastes as the one line it is; a line break the text had stays. The wheel up scrolls the terminal's own scrollback, as in any shell: its first notch hands the mouse to the terminal, and the next key or paste takes it back. A drag that starts on a row the OMH TUI did not print selects nothing, and the mouse never opens the transcript view; only Ctrl+O does. OMH_TUI_MOUSE=0 gives the wheel and native selection back to the terminal.

    • The Desktop pane stays on the focused chat and folds large DAGs by layer. The yolo chip follows the focused chat instead of whichever session's hook ticked last (a subagent or a cron run); where OMH holds no observation of that chat while some session of the gateway was observed ON within the ledger's freshness window, the chip still shows. In its DAG section, a pane that names its chat no longer draws a fanout whose record names another chat as its origin; one with no recorded origin, which includes any fanout still running, is drawn as before. The pane always draws the same five sections in one order, USAGE, CATCHES, PLAN, DAG and AGENTS, each with an empty line when it has nothing to say; the subagent list and the routes are one AGENTS section, and a plan that Hermes' todo_list holds gets a short PLAN block saying where its steps are. DAG cards fold by layer, with Auto, Focus, Unfold all and Fold all on a bar over the cards and a layer's chip to open or fold it by hand; the fold mode and the View menu's choices are kept in the app's local storage. A graph too large for the pane comes as a bounded layer view that counts what each layer left out by state: running and returned lanes are kept first, then failed, blocked and stuck ones, so neither is cut while a finished lane is shown, and a lane whose parent was cut hangs from its nearest shown ancestor by a dotted edge. A run too deep for even its layer rows reads as its counts by state instead of "No DAG". Layers held open are asked for again within a larger bound, and a layer bigger than that still comes back cut. USAGE reads the gateway's session.stats counters where they count more than the chat's recorded row, and done DAG work is drawn green. The TUI's DAG block and the chat's end post get the same graph as before. Thanks @yqm-307. (#2745)

    • instinct-ledger can start from git history, and code-review gains two rules. A fix commit that replaces lines a different pull request landed shortly before is a re-fix, and nothing turned such a pair into a lesson. instinct-ledger now ships references/refixed-commits.md: how to find the pairs (blame the removed lines at the fix's parent and split by first-parent merge, listing the non-merge fix commits reachable from the default branch), that the count is an upper bound to read pair by pair, that two pairs from one day count once, what a ledger line carries, and where a promoted rule goes. Two requests that reached the generic router and code-review (어제 머지된 커밋 중에 다시 고친 것들 찾아서 교훈으로 정리해줘, review yesterday's merged commits for fixes of fixes and write the lessons) now dispatch instinct-ledger, when the message holds both a re-fix phrase (fix of fix, fixes of fixes, re-fixed commits, refixed commits, 다시 고친 커밋, 다시 고친 것들) and a lesson cue (lesson, lessons, 교훈); nightly lessons from merged work keeps its base route. Messages that use fix, lesson, commit, merge or review for other work, or hold a re-fix phrase without asking for lessons, keep the route they had; a message that negates the request while naming both still reaches the ledger. The code-review quality bar adds two rules: a fix reaches every place of the same shape, and a read, parse or hash that failed means unknown, never empty, none or equal. The idea of mining re-fixed commits comes from yetone/magpie (MIT), concept only, reimplemented here.

    • Claude Fable 5.2 now leads the visual-engineering chain. Frontend and visual lanes routed through Hermes try Fable 5.2 first, then Claude Opus 5.5, Gemini 4 Argon and Kimi K3.1 Pro, all at high (before, Opus 5.5 came first). A child on Fable 5.2 at high with no route record is labelled visual-engineering, as before; a child on Opus 5.5 at high is still labelled deep. The Maestro lane's chain is unchanged. The setup examples in the installation guide now name Fable 5.2 instead of Fable 5.1, so no shipped recommendation names Fable 5.1 any more. A chain you set yourself in ~/.omh/routing/model-chains.json is not touched.

    • dag-ulw plans fewer, larger nodes. It splits where work can run side by side and keeps a small follow-on step in the node it follows, so a run starts fewer children; a job with nothing to run side by side still runs as a graph.

    • dag-ulw stops telling a parallel plan to split further. Its advice to split a run into parallel lanes now looks only at the build nodes: it comes when at least 80% of two or more build nodes sit on one chain, in a run of five or more nodes, and it now also says a small step that only follows another can be folded into the node it follows. A run whose only chain is its baseline, review and verify steps no longer gets it, nor does a chain of three builds beside one other build (3 of 4).

    • dag-ulw checks each part of the goal, not only the goal as a whole. The model splits the goal into the outcomes it can check separately (you type nothing new), and the verify node marks each one met, not met or not checked, with what it ran. A check that leaves one not met or not checked is not accepted, and the run counts as verified only when every outcome was met; the status line, the end of the run and the TUI say how many were met (2/3 criteria met). A run defined without them works as before.

    • dag-ulw's handoff checks on a Docker or SSH terminal ignore shell aliases and functions. The commands that check a node's handoff file and read its goal-criteria lines now run past any alias or function in your shell that shares a tool's name (an alias grep='grep --color=always', a function named echo or [), so they read the file as it is. When the criteria search finds no line, that is a read with nothing in it; any other failed step leaves the handoff not checked.

    • omh_jev_ask can reach a self-hosted Jev-compatible server over local HTTP. With {"local_http_endpoint": true} in <omh_home>/jev/settings.json, TYPESAFE_BASE_URL may be a plain-HTTP root on localhost or a loopback, RFC 1918, ULA, or link-local IP literal, and "typesafe_model" names the model that endpoint serves (an Ollama, vLLM or llama.cpp server exposing /v1/systemone). The local ranges are named exactly, so documentation, reserved, Teredo, 6to4 and NAT64 addresses and an IP with a trailing dot are refused; a plain-HTTP ask never goes through HTTP_PROXY, so the key is not sent in clear to a proxy; and typesafe_model never applies to TypeSafe's own host however it is spelled. Thanks @brentkearney. (#2716)

    • omher 0.2.10. Bots on another computer join over SSH, the sidebar starts with one bot and adds the others from +, and you can add your own ACP coding agent to SESSIONS, by hand or from a bot's draft.

      • Connect a Hermes on another computer over SSH. The sidebar + has "Connect over SSH": pick a host from ~/.ssh/config (or type user@host), press Check, and a card shows the exact ssh command, the Hermes found there and its bots; nothing is saved until Connect. Those bots join the list marked "SSH · " (the host in the tooltip), with the link's state when it reconnects or fails. Asking a bot to connect one makes it answer with a "Connect over SSH" chip that opens the same card. omher uses your own ssh with key login only, writes only ~/.omher/peer/ on the other computer, and Settings › Remote machines removes one. (#2881)
      • A bot's SSH or agent draft shows as its chip only, and the SSH marker is readable. The raw JSON block (with its user@host) no longer sits in the reply above the chip; the SSH · <name> marker is 11px, keeps the bot's name whole and ends a long computer name in an ellipsis. (#2889)
      • omher starts with one bot; the others join from + in one click. A new install starts with the default bot, an existing one with the bot last open (plus pinned ones). + is "Add a bot", listing the profiles not in the sidebar yet; right-click "Remove from list" keeps the profile. Settings › All bots still shows every profile. A bot with no AI account is no longer offered a Claude sign-in (it needs Claude Max extra-usage credits); ChatGPT / Codex, Nous Portal and API key… stay. (#2880)
      • Add your own ACP agent, after a confirm card. Settings › Agents lists SESSIONS' agents and yours; Add agent takes a name, a program and its ACP arguments, shows the exact command, and saves only on Add (~/.omher/custom-agents.json). A bot can draft one with an omher-agent block that opens the same card with "Proposed by "; a shell or interpreter is refused for bot drafts. The SESSIONS + menu lists Claude Code, Codex and OmO first, then the other built-ins, then generic ACP agents. Kilo Code is no longer offered; a saved Kilo Code session comes back ended with one line saying so. (#2876)
      • The model picker completes a switch Hermes asks to confirm (large context, cost or data policy): the pick is the confirmation, and the picker shows the chat's live model. Reported and first fixed by @Pluronymous-ODK in #2883 and #2886. (#2887)
      • Plugin cards have a Details button that opens the plugin sheet without installing and shows the whole description. Reported and first fixed by @soares-daniel in #2833. (#2887)
      • omher on Windows finds Hermes' Python through the package-manager launcher (hermes.exe or hermes.cmd, read without running it), so a PM or source install no longer reads as "Hermes Agent isn't on this computer". Reported by @MarinWasAsleep in #2885. (#2887)
      • ⌘T opens the New session picker on macOS, the same menu as ⌘N; Settings › Keyboard lists both. On Windows and Linux only Ctrl+N opens it (Ctrl+T stays the program's). Both keys work with a non-Latin layout active. (#2890)
    • omher 0.2.9. A new bot without an AI account signs in from its own chat, the + in the message box has one way to add files and the bot's plugins beside it, and an empty part of the Sessions list starts a new session.

      • A new bot that cannot answer yet says so and signs in from its chat. When Hermes' readiness check says a bot has no AI account, its chat shows one line with ChatGPT / Codex, Claude, Nous Portal and API key…. A sign-in opens the system terminal running exactly hermes auth add openai-codex, anthropic or nous; the line asks again every 3 seconds for up to 10 minutes and when omher comes back to the front, and goes away once the bot can answer. API key… opens Settings › Providers. The first install of Hermes from the app lands in the default bot's chat, and with no bots the empty window points at the + menu. (#2874)
      • One way to add files, and the bot's plugins under the same +. The menu with "Photos & videos", "Files" and "Paste from clipboard" is gone. In a bot's chat the + has two rows: "Add photos and files" (one system picker for any kind of file, also ⌘O / Ctrl+O) and "Plugins ›", whose submenu lists the bot's enabled Hermes plugins and Manage plugins; picking one puts "Use : " at the start of the message without sending it. In a session the + opens the file picker straight away. Paste and drop attach files as before. (#2870)
      • Right-click an empty part of the Sessions list to start a new session. It opens the agent picker without starting a process; the chat and terminal marks beside session rows are smaller. (#2869)
    • omher 0.2.8. Six more coding agents in SESSIONS, update checks that recover after a failed request or sleep, and a session header that says when background work is waiting for you, with clearer terminal find.

      • Six more agents in SESSIONS. Kilo Code, Gemini CLI, Qwen Code, Kimi Code, Copilot CLI and Cline open as session chats through the ACP adapter: answers stream, tool steps show, and approvals offer Deny and Allow once. Installed agents come first in the + menu; More agents unfolds the rest with whole install commands that copy on click. Continue a conversation stays reachable when the menu first opens; the folded menu keeps its text clear and shows a scrollbar when it needs one. An installed agent that is signed out ends with its own sign-in command, Open in Terminal and Restart, not raw protocol output; omher never signs in, sends authenticate, or takes a key. (#2858)
      • Kilo Code and OpenCode check the rules they will start with. omher builds ask rules from each CLI's own agent listing and starts only when its conservative comparison accepts the listed permissions and agent modes. Unsupported changes are refused; file/env substitution tokens are refused before a second listing. Windows matching is handled with its own case rules. Rules are a session-start snapshot: a deny added later applies from the next session. Private prompt/model metadata not exposed by the listing is not preserved by that comparison. Actual Windows CLI behavior and the other five agents after sign-in were not verified for this release. (#2858)
      • An omher left open finds updates without a restart. A failed check retries after 2 minutes, backs off to at most once an hour, and stays silent on failure. Showing or focusing the window checks again when the last good check is at least 30 minutes old, including after sleep. There is still only one request at a time, to the same updater feed. This retry behavior starts once 0.2.8 is installed; a 0.2.7 app keeps its old launch/6-hour schedule until it updates. (#2857)
      • The session header says "waiting" about background work that waits for you. If every known background subagent waits on an approval or question, the header uses waiting rather than running in all four languages. A live command or running subagent still reads running. In a narrow chat the label shortens and wraps with Stop under the name, while the path gives way first so the session and model stay readable. (#2862)
      • Terminal find stays readable and follows the theme. The active match's text and fill have stronger contrast on all 38 themes, including thin Korean strokes on a 1x screen. Changing the terminal theme recolours open matches at once; leaving a terminal clears highlights when its find bar is no longer shown. An open terminal now follows an OS light/dark change when Appearance is System. Solarized Dark and Light keep bold green, yellow, blue and cyan in their colours instead of turning them grey. (#2860)
      • omher.app's preview. The live preview is pinned to omher 0.2.7; Show in Finder and Open in Terminal explain their preview-only limits rather than touching a real profile. (#2853)
    • omher 0.2.7. Find and continue sessions, choose how their terminals look, answer and stop background work, read a plan before allowing it, and open a Codex subagent's file change to its diff.

      • A subagent's approval that comes after its turn ended reaches you in omher. When Claude Code ran a subagent in the background, the turn ended first and the subagent asked afterwards (or asked during the turn and the turn ended with the ask open). omher showed that ask as "Not answered" with no buttons, left it out of Needs you and kept the header on Ready, while the subagent waited for an answer nobody could give. Now it is a live card with Deny, Allow once and Allow for session, the session reads "Needs you" in its header and its sidebar row, and the answer reaches Claude Code with no turn running; the subagent then finishes. The ask reads "Not answered" once its subagent ends, Claude Code withdraws it (Stop), or the session's program ends. A Codex subagent's approval follows the same rule. (#2845)
      • You can stop what a Claude Code session still runs in the background after its turn. While no turn runs, the session header's "1 still running in the background" now ends with a quiet "Stop". It sends Claude Code's own interrupt and, for a background command, its own stop_task: the subagent's row reads "stopped", the count goes, an open late ask reads "Not answered" and the header is Ready. Esc still stops a running turn only, and the composer keeps Send. Codex, OmO, OpenCode and DSH get no such Stop, because their own Stop does nothing to that work. Also: the subagent card and its ask end at one right edge; a card whose rows all wait on you says "Waiting", not "Running"; the needs strip no longer ends on a lone "·"; and a background subagent's row shows its model. (#2849)
      • An omher session that closed normally no longer gives a warning as why it ended. A Codex session you closed showed "Ended · user (~/.codex/config.toml): mcp_servers.playwright.type is ignored.", a line Codex prints at start about your own config. A session whose program ends with code 0 now reads only "Ended", for every chat agent. A program that really fails shows the last line it printed about the failure, not a warning from its start. (#2850)
      • Claude Code's plan-mode card shows the plan as text. When Claude Code finished a plan in plan mode, omher asked "Allow ExitPlanMode?" over a cut-off JSON string. The card now asks "Start working on this plan?" and shows the whole plan the way an answer reads, scrolling inside the card, with Deny and Allow once (Claude Code offers nothing to keep for the session there). Other tools that are neither a command nor a file edit (a skill, a web fetch, MCP tools) show one readable line instead of their input as JSON. (#2850)
      • An edit Codex hands to its subagent opens to its diff. The subagent's row showed the edit only as a line like README.md +2, calc.py +4. Opening that step now shows the removed and added lines per file, as an edit Codex makes itself does. (#2850)
      • omher's session terminals take Orca's themes and options, find, links and bounded output. Settings › Terminal is a new page: 38 themes (omher's own two and 36 carried over from Orca, their MIT notices in apps/omher/THIRD_PARTY_NOTICES.md, which now ships inside the app), one picked for the dark app theme and one for the light, with a search field and a live preview drawn by the real terminal; font, size, line height, cursor shape and blink, scrollback, minimum contrast, padding, Option as Meta on macOS, and the renderer. A change reaches every open terminal at once, one that is hidden when it is shown again, and the settings are one record on this computer. In a terminal, Cmd+F (Ctrl+F on Windows and Linux at a prompt, Ctrl+Shift+F anywhere) finds in the scrollback, with a count and the active match as the strongest thing on the screen in colours taken from the terminal's own theme. A web address in the output opens in the browser on Cmd+click or Ctrl+click, and a hint beside the pointer names the key; a hyperlink the program marks itself (OSC 8) follows the same rule and no longer goes through a confirm dialog. Copy on select and Right-click pastes are two switches, both off by default. The shell now reads a program only 2 MiB ahead of what the terminal has drawn, so a runaway program no longer fills memory and Ctrl+C is answered at once. Measured against a model of the page written in Rust, not the app's window: with a page that needs 20 ms a message, ten seconds of yes held 15 MB instead of 1,211 MB, and Ctrl+C stopped the output in under half a second instead of 69 s; with a page that draws 100 MB/s, after 14 ms instead of 865 ms. A session that is not the one on screen keeps drawing, so a background agent is not held at that mark. What did not get faster, as measured: throughput against the 100 MB/s page is unchanged (75.6 and 75.7 MB/s), a keystroke's echo is unchanged (31.7 and 31.1 ms run alternately), and a page that needs 20 ms for every message draws 5.3% fewer bytes a second because its messages are smaller. The GPU renderer is 1.2 to 3.4 times slower than the standard one on a full-screen redraw in WebKit, over two measurement runs (2.2, 3.4 and 1.3 times in one; 1.7, 3.0 and 1.2 in the other), so Auto uses the standard renderer on every platform (nothing was measured on Windows or Linux) and On still tries the GPU. No number here was taken in the Tauri window; the tables are in apps/omher/bench/render/README.md and apps/omher/bench/pty/README.md. (#2841)
      • omher sessions are easy to find, reopen and switch. A session row's right-click menu now holds Rename, Show in Finder / Explorer, Open in Terminal (your own system terminal in that folder), Copy conversation id, New session here and Close session. The sidebar search also filters sessions by name and folder. The SESSIONS + menu ends with "Continue a conversation" for Claude Code and Codex: their recent conversations in your recent folders and home, read-only from their own files, and picking one continues it in omher with its last messages shown. ⌘N (Ctrl+N) opens the + menu and ⌘⇧] / ⌘⇧[ move to the next and previous session. The Dock shows how many approvals and questions wait (Linux launchers that show counts too; Windows gets a dot on the taskbar button). BOUNDARY rule 30. (#2839)
      • omher.app's preview. The hero preview runs omher 0.2.6; restart, terminal and rewind controls explain their preview-only limits instead of acting on a real profile. (#2846)
    • omher 0.2.6. A session chat shows what its agent did and the subagents it starts, a session can run in the agent's own terminal, OpenCode joins SESSIONS, a bot restarts with one press, and omher works without oh-my-hermes.

      • What the agent did. In a session chat a file-edit step opens to the real change (removed and added lines per file), the agent's todo list is one checklist that updates in place, its thinking stays folded under the activity line until you click it, and answers and code blocks have a copy control, in the bot chat too. Edit on your last message takes it back through the agent's own rewind (Claude Code, Codex) and puts the text back in the box; files are not undone, and a quiet line says so when that turn changed files. (#2816)
      • Subagents in sessions. Claude Code, Codex and OmO subagents show as one card per fan-out with a row per subagent that opens to its own steps and answer, and the session header opens Plans on that session's run. A Codex session stays on your turn while its subagents work, and Stop stops your turn, not a subagent's. When OmO runs subagents in the foreground, each row ends with that subagent's own answer, model and time, and OmO's todo list shows as the turn's checklist. omher only shows subagents; it never starts, steers or stops one. (#2800, #2788, #2838)
      • Chat or terminal. Beside the agent in a New session, a quiet Chat | Terminal choice runs claude, codex, omo, pi, omp, opencode or omh as the real CLI in a terminal in the folder. Every session row marks whether it is a chat or a terminal and shows "waiting", "error" or "done" with a system notification while you look elsewhere. (#2813)
      • OpenCode. OpenCode runs as a session chat: answers stream, tool steps and loaded skills show, its models are listed, and a conversation continues after a restart. It asks before edits and shell commands, with ask rules omher builds from OpenCode's own, and offers Allow once only, never "Always allow". (#2801)
      • Restart a bot. A bot row's right-click (or ...) menu has Restart, or Start for a bot that is not connected. It runs Hermes' own gateway restart; a busy gateway shows what is running and asks before "Restart anyway", and every bot on the same gateway is named first. After a plugin install omher asks the gateway to reload its plugins and offers Restart only when Hermes still needs one. omher never restarts a bot by itself (BOUNDARY rule 29). (#2826)
      • Hermes without oh-my-hermes. With Hermes alone, the SESSIONS + menu's row reads Hermes and opens a Hermes chat in the picked folder, and a quiet "!" by the + explains oh-my-hermes and installs it. After that install your bot's chat offers the model setup once: it puts the command in the message box and sends nothing until you press Enter. The skill picker now lists oh-my-hermes' skills, and the first-run cards over an empty chat are gone. (#2825, #2822)
      • Session controls. The composer row shows the session's permission mode as a quiet label (Claude Code: Ask, Plan, Accept edits; Codex: Read-only, Ask, Auto in folder), the model pill shows the agent's reasoning levels, / lists the agent's commands and skills, and the line under the box shows the git branch and the context in use. A message sent during a turn reads Waiting, and Not sent when Stop drops it. (#2795)
      • Skills in a bot chat are $name. $ anywhere in the message lists the bot's skills and puts in $name; / stays a Hermes command and goes out exactly as typed, so /new, paths and ./name are plain text. (#2821)
      • The agent you picked, and never stuck on Working. omher starts the agent's own install even when another program with the same name comes first on PATH, and finds Grok and pi where their installers put them. DSH now means DeepSeek Harness and opens as a chat. An agent that does not answer in chat mode within 20 seconds ends with a line that names it and its install command, and Restart; a Codex chat whose conversation was deleted starts a new one in the same folder. (#2797, #2787, #2802)
      • What really happened in a turn. An OmO retry is one turn that reads "Retrying 1 of 5", a step still running when you press Stop reads Stopped, and a Codex session whose own config asks for no approvals says "No approvals · full disk access". A running subagent card counts "1 of 2 done", a subagent waiting for you reads "waiting" in the sidebar, and a bot's chat header keeps the model name whole from 900px up. (#2824, #2812)
      • Accounts that reset soonest go first. A new switch under Auto-switch in Settings › Accounts (off by default) hands the next session to the account whose limit resets soonest when the one in use runs out, and the chat line says so. Codex also sends one notification when a window under 50% used resets within the hour, and each CLI's version shows beside its name. (#2834)
      • Files you attach to Claude Code. Claude Code reads a file attached to its session without a "Run this command?" card; a read anywhere else outside the work folder still asks. (#2818)
      • First launch on macOS. The README and the release notes give the step that works on macOS 15 and later (System Settings › Privacy & Security › Open Anyway). (#2820)
      • omher.app's preview. The live preview on omher.app runs 0.2.5, opens Automations and Plugins on staged data, and keeps the whole model name in its chat header. (#2803, #2804)
    • omher 0.2.5. omher reads fully in Chinese, Korean and Japanese, the Linux AppImage reaches its bots, and SESSIONS chats take files.

      • Chinese, Korean and Japanese. Every word omher itself shows is in all four languages, Simplified Chinese written for a developer in mainland China; brand, model and command names stay in Latin and what the bots say is never translated. Enter that commits an IME composition (pinyin, kana, hangul) no longer submits a field, a first run takes the language from the system, and CJK text falls back to PingFang SC, Microsoft YaHei and Noto Sans CJK. In Chinese a bot's reply reads 回复 everywhere and every Settings description ends with 。. (#2777, #2786)
      • Linux. The 0.2.4 AppImage opened its window but no bot connected: omher passed the AppImage's own PYTHONHOME and PYTHONPATH to the host Hermes Python. Every host program omher starts now gets the host environment, and PR CI starts the built AppImage and checks that the app, its window and its bridge are up. (#2778)
      • Files in SESSIONS chats. Paste, + and a drop attach files in a Claude Code, Codex, OmO, pi or oh-my-pi chat, with the bot chat's chips and limits; files are saved under ~/.omher/session-files/<session>/ and removed when the session closes, and a picture reaches the agent as an image. An oh-my-hermes session takes no files (BOUNDARY rule 26). (#2781)
      • The bot's face stays put. After you send, the bot's ringed face no longer blinks away when the first status line replaces the typing dots; one face stays in place from send to the end of the reply, and a SESSIONS chat's agent mark does the same. (#2780)
      • Finished turns go idle. A session whose agent has answered no longer stays on Working with Stop (OmO, pi and oh-my-pi sent statuses after the turn ended), an agent that dies mid-turn ends its turn, and a running OmO chat continues after omher restarts. DSH and Grok sessions that exit get the Ended line and Restart a chat has. (#2782)
      • Subagents are visible. The bot chat's header has a panel button right of the model control: it opens the right panel, on Plans when the bot has live work, and carries a live dot while subagents work. A turn that starts a delegate_task opens Plans on that run once by itself (windows 1100px and wider), Plans re-reads itself while a run is live and says what it is for when empty, and a dag run nothing changed for a day reads "no update for 9d" instead of pulsing as live. (#2784)
      • Subagents in the chat. A turn that starts a delegate_task shows one card per fan-out ("2 subagents running · 1 of 2 · 41s", then "Ran 2 subagents · 3m") with one row per subagent: its model's mark, its goal, its state, its current step or result, its time. The rows keep updating after the bot's own turn ends, and the card's run name opens that run in Plans. The word is "subagent", not "helper", in all four languages; omher still never assigns, steers or stops one. (#2796)
    • omher 0.2.4. omher runs on Linux, starts OmO and DSH sessions, and shows every Claude login you have.

      • Linux. An x86_64 AppImage (and a .deb) ships with each release; the AppImage updates itself through the Update button like the other builds. The omher@beta installer puts it in ~/.local/bin/omher with a menu entry, and omher.app has a Linux download button. (#2765)
      • SESSIONS + menu, OmO and DSH. + opens a menu of agents with their marks: installed ones first, the one used last on top, a missing one dimmed with its install line. Picking one opens its New session chat at once. OmO (npm i -g omo-ai) runs as a chat and Restart continues it; DSH (deepseek-bot's dsh) opens in its terminal and Resume continues it. (#2766)
      • Every Claude login. Settings › Accounts and the Claude usage card list each Claude login a local CLIProxyAPI holds, read-only and by name alone; omher never opens those files. (#2764)
      • Providers. Settings › Providers shows one compact row per provider that opens to its models grouped by vendor, a long list scrolls inside a capped box, and Add key opens your terminal with hermes auth add. (#2763, #2767)
      • Automations and Plugins. The two sidebar pages that said Soon now work, by hand or by asking a bot. Automations lists every bot's cron jobs with a status mark, an on/off switch and a detail panel; pause, resume, delete and New go through the bot's own gateway, and schedules read in words. Plugins shows a bot's plugins and the curated catalog; toggle, update, remove and a catalog-pinned install go through the bot. docs/OMHER-BOUNDARY.md gains rules 24 and 25. (#2768)
      • Language. Settings › General has a Language row (English, 한국어, 日本語, 中文) that switches at once. (#2768)
    • omher 0.2.3. New sessions open a chat at once, sessions have names, and an update asks first and brings you back where you were.

      • New session. SESSIONS + opens a chat right away, like the Codex app: the agent is picked in the composer and the folder in a chip under it, and nothing runs until the first message. (#2759)
      • Session names. A SESSIONS row shows the session's name (yours, the agent's own title, or a short summary of your first message) above its folder and agent; double-click to rename. (#2756)
      • Update and resume. Update asks before it restarts. After any restart omher reopens the chat, panel and unsent text you had, and a running Claude Code or Codex session continues its conversation. (#2757)
      • Pull requests on plans. Each plan part shows the pull request it opened with GitHub's state mark and number; a click opens it. (#2755)
      • @bot names. @name of a bot in a chat opens that bot's chat; it never sends anything. (#2753)
      • Link chips. A bare link reads as the service's mark and a short label (GitHub, Slack with its channel, Google, X, YouTube, Notion, Linear, Figma, npm); the full URL is in the tooltip and omher fetches nothing from it. (#2758)
    • omher 0.2.2. Codex from the ChatGPT app counts as installed, and an empty SESSIONS section is one quiet row.

      • Codex. On macOS, the Codex CLI inside the ChatGPT or Codex desktop app (under /Applications or ~/Applications) is found after PATH and every install folder, so a Mac whose only Codex is the app's copy no longer shows it as not installed. An npm or Homebrew install still wins.
      • New session. An empty SESSIONS section is one muted New session row with a +, like the Automations and Settings rows; the hint and the installed agents are in its tooltip. The first-run card's two buttons share one outlined look.
    • omher 0.2.1. The plan graph looks like part of a desktop app, and the first launch says what bots and sessions are.

      • Plans. Each part shows its model's logo and name (Opus 5.5 · high) in the app's own font, with rounded boxes, small status icons and curved lines; several lines into one part no longer pile up on one arrowhead. The plan's title leads its header, with a quieter status line under it.
      • First run. Two small cards over an empty chat explain Bots (coworkers you hand work to) and Sessions (a coding agent working with you in one folder), each with one button. One × hides them for good.
      • Sessions. The Terminal button is a plain icon with a tooltip, so it no longer looks like the model button. A session whose agent exits cleanly reads Ended, not Failed. Two folders of the same name show their parent folder. "Not installed" says how to install that CLI.
      • Needs you. The collapsed strip names who is asking and their first words.
    • omher 0.2.0. Your bots, coding sessions and plans in one window.

      • Plans. A dag-ulw run and a coding agent's sub-runs open as a plan. Its graph shows each part with the model, effort and category it ran on, the parts it waits on, and the latest lines of the one you pick. There is a full view (⌘⇧G).
      • SESSIONS. Claude Code, Codex, pi and oh-my-pi sessions read their real command, approval and question events. Approvals go Deny, Allow once, Allow for session. Stop works, and an ended session can Restart. A new session asks for its folder and offers the five recent ones.
      • Accounts and usage. Settings › Accounts lists the coding-CLI logins on this computer, adds one through the CLI's own sign-in and switches between them. The sidebar shows each CLI's usage.
      • Files. You can attach files to a message, and files a bot sends show as images, video, audio or file cards.
      • More. A ./ skill picker in the composer, official agent marks, and four short sounds (Settings › Sounds).
      • Updating. From this version, later releases arrive through the sidebar's Update button.
    • omher updates itself. When a new omher release is out, an "Update to 0.1.x" button appears at the bottom left of the sidebar; one click downloads it, checks its signature and restarts omher on the new version. The app checks at launch and again while it runs. People on omher 0.1.1 or earlier have no button yet: install the next version once from the Releases page (or the omher@beta one-liner), and every release after that arrives through the button.

    • The OMH TUI queues commands sent during a turn, and Enter never interrupts it. What Enter does during a turn no longer depends on Hermes' display.busy_input_mode, so a config set to interrupt no longer stops the turn. A slash command typed during a turn waits in the prompt queue and runs in order when the turn ends (a prompt after a /compress goes once Hermes has answered it); only /stop, /steer, /queue, /btw, /bg, /gateway, /quit and the commands about this terminal run at once. Interrupting is Esc or /stop, never Enter, and the send-now key never sends a queued command into the turn. One Esc now stops a running turn even with a draft typed, as in Claude Code, and the draft stays in the composer (vim mode keeps Esc Esc; Ctrl+C still clears the draft first).

    • omher 0.1.1. The app shows SESSIONS: Claude Code, Codex, pi, omp and Grok open as chats in a folder you pick (or as a terminal), and their questions and approvals join the bots' asks in a "needs you" strip that ⌘J walks. A bot row rings while it answers and shows how long it has been at it. A message from Slack, Discord or cron gets "Reply there", which asks the bot to answer in that thread with its own tools; omher never posts there itself. The side panel has tabs, the model picker has search, and the chat keeps one face per run of bot bubbles. (#2695, #2702, #2704, #2706, #2708, #2709, #2710)

    • Claude Fable 5.2 takes every Claude Fable 5.1 slot, and heads ultrabrain (owner decision, 2026-11-10). The ultrabrain chain is now Claude Fable 5.2, GPT-6 Astra, then Claude Opus 5.5, all at xhigh; it was GPT-6 Astra alone, so a user without OpenAI access fell straight back to the parent model. A child on Fable 5.2 at xhigh is labelled ultrabrain, and one on GPT-6 Astra at xhigh is labelled forensics when no route record names its lane, unless the machine's recorded providers put GPT-6 Astra first in ultrabrain. Fable 5.2 also replaces Fable 5.1 in architect, visual-engineering, artistry, capable, the Hermes main suggestion, and the Claude Code chain (claude-opus-5-5 -> claude-fable-5-2 -> opus), each at the effort 5.1 had. Claude Code needs version 2.1.290 or newer to run Fable 5.2 (seen on one machine, not stated by Anthropic). omh coding model-contract --model claude-fable-5-2 prints its record from Anthropic's pages: thinking is always on, so a request for off or minimal is raised to low. It lists at $8 / $40 per million tokens, against $10 / $50 for Fable 5.1, and cost estimates use that price. Fable 5.1 left the shipped chains but stays recognized, priced and routable, so a chain you set yourself can still name it. Claude Mythos 5.2, the same model for organizations Anthropic has verified, is recognized and priced but in no shipped chain. The calibration text a Claude model receives is unchanged.

    • One command installs Hermes, OMH and the omher desktop app (beta). curl -fsSL https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.sh | sh -s -- omher@beta (and the PowerShell script-block form with the same argument) installs Hermes with its official installer when it is missing, installs or upgrades OMH, runs omh setup when OMH or Hermes is new on the machine, then installs omher from the newest omher-v* release (dmg on macOS, silent per-user installer on Windows, the AppImage on Linux). The plain one-liner is unchanged. An omher installed first offers the same install, minus the app, on its first launch when Hermes is missing, and streams the installer's output on that screen. The README shows the one-liner and a screenshot. A failed download or install is reported as a failure, on the app's setup screen too (it used to read as success); the app image is checked against the release's SHA256SUMS.txt and swapped in only once copied whole; a profile HERMES_HOME finds Hermes under its root; the release lookup reads one small API page instead of 700 KB and stops with a clear message when GitHub's rate limit is hit.

    • Skill replies say verdicts and labels in plain words. Live replies copied skill vocabulary to people: release-cut's "version surface", build-failure-triage's handoff and its FIX_READY / NEEDS_MORE_LOGS verdicts, code-review's no_ship, not_assessed, spec_axis and verification_gaps, and "run record" inside commit text. The build-failure-triage, code-review and commit-pr-authoring checklists now give each verdict a plain reading (the review verdict makes no merge claim) and keep the tokens in the record or structured return, which a dispatched reviewer still returns in full; release-cut says "every file that carries the version". No schema, tool contract or record field changes, and routing is unchanged.

    • Planning work runs on ultrabrain. The brain role chain for Codex, Claude Code and a category-maestro table now heads on ultrabrain and keeps deep and unspecified-low behind it, so omh coding model-route --role brain names GPT-6 Astra at xhigh on Codex and the Opus 5.5 chain at xhigh on Claude Code. A Hermes-owner brain, research depth deep or scale large unit reads ultrabrain and, when no generative model there is confirmed active, steps to deep before the shared last resort, never into a cheaper tier: a set without an ultrabrain model keeps deep's effort. No route moves down, every other unit resolves as before, and a requested category is never widened. On the Hermes lane, ulw-work routes planning, decomposition and hard-debugging lanes to ultrabrain and ideation to deep, and ulw-plan and omh-plan set omh_delegate_route to ultrabrain before dispatching a planner or critic child and clear it afterwards. Planning itself still runs in the parent session.

    • A maintainer script lists fix commits that replaced lines another pull request landed shortly before, and the first lessons file is written from its output. uv run python tools/refix_report.py [revision] [--days 7] [--window-hours 48] [--json] blames the lines each fix/hotfix commit replaces and prints the pairs from a different pull request; its count is an upper bound, since a replaced young line does not say why. Run on 65796e68f for 7 days it found 179 fix commits, 80 of them in that class, 217 pairs, 0 files that could not be blamed (the script reads paths with git diff -z, so names with a space, a quote or non-ASCII letters are kept, and a failed blame is counted and printed); 8 were kept after reading them and give three lessons in docs/LESSONS.md. One met the three-separate-days bar and is now in REVIEW.md. Nothing runs the script on a schedule. docs/REVIEW-SWEEP.md gains a step for pull requests from forks with the changes that stop the review until a maintainer has read them.

    • dag-ulw names the node whose judgement decides the run, and gives it ultrabrain. The skill told the model to climb quick -> unspecified-low -> unspecified-high and capped ultrabrain at one node, but never said when to use it; on the owner's machine no node in recent dag-ulw graphs used it. The skill now says that when one build node's judgement decides the run (a design decision, a root cause, a contested trade-off), it takes ultrabrain where the user's routes serve it, at most one; every other node keeps the ladder, and review, baseline and verify nodes keep their own rules. omh_dag define adds the advice no_judgement_node to a graph of 4+ nodes with no ultrabrain, architect or forensics node, with an exit (go on when no node decides the run). Repair and extend rounds do not raise it, and the reply leaves it out when the user's model_map options have no ultrabrain option. Like all lint it never refuses a run.

    • The one-line installer can remove OMH. curl -fsSL https://omher.app/install | sh -s -- --uninstall (and install.ps1 --uninstall on Windows) finds the installed omh without relying on PATH, runs omh uninstall and shows its output and exit code; --purge becomes --all and --dry-run previews. An uninstall run never downloads or installs anything, and with no command found it says where it looked and exits nonzero. The idea of removal through the install one-liner comes from yetone/magpie (MIT, commit 5e933e1; concept only, reimplemented here). install.ps1 and its test were written without a PowerShell host; CI is their first run.

    • omher.app shows the real GitHub star count. The header used to show a 4.9k baked into the page on 2026-10-31 whenever the visitor's own, unauthenticated GitHub API call was rate-limited. The Pages deploy now stamps the current count into the page (tools/stamp_site_counts.py, with the workflow's token; a failed call keeps the committed value and never fails the deploy), and a schedule rebuilds the site every 6 hours so it stays current without a merge. In the browser, one code path in app.js updates the count, and a cached or fetched count can only raise the baked one, never lower it.

    • The Desktop pane keeps a finished DAG and lets you switch between a conversation's DAGs. With two or more dag runs in one chat, a menu in the DAG header picks which one to draw (the newest by default); a picked run is drawn alone, without another run's children. A finished run keeps its cards, node summaries and figures under its run line instead of collapsing to one line (under the pane's Auto fold, a run of nine lanes or more shows them as folded layer chips); the chat end post and the TUI keep their linger. Edges are routed by the rows the cards actually wrap to, implied dependency edges are hidden unless Implied edges is on in the DAG's View menu, and repair/extend edges are drawn faint. USAGE is a token ring with the total and the cache hit as a percent in its middle, and can add the drawn graph's subagent lanes; a sum with any unpriced part, the chat included, shows as a lower bound (≥$). Thanks @yqm-307. (#2690)

    • A short install command: curl -fsSL https://omher.app/install | sh. omher.app now serves install and install.ps1, small wrappers that download install.sh (or install.ps1) from main and run it with only the arguments you pass, so the short form installs exactly what the long raw.githubusercontent.com/.../install.sh | sh one-liner does; on Windows, irm https://omher.app/install.ps1 | iex. The README, its translations, the website and docs/INSTALLATION.md show the short form; the install guide keeps the long one as "same as".

    • A pytest run no longer makes omh update skip the plugin refresh. The plugin bundle ships a module named test_expectation_check.py, which pytest's assertion rewriting takes for a test file, so a pytest process that imports the installed plugin leaves a <module>.<tag>-pytest-<version>.pyc beside it. The refresh read that file as something the person had added and left the plugin directory unchanged. It is now recognised as bytecode of a shipped module, under the same content check as any other cache. Reported by @junjuny0227 (#2791).

    • OMH TUI settings changes say what they did, and do what they say. /yolo and /reasoning <level> went to Hermes' slash.exec, which runs them in a worker apart from the live agent: /reasoning high said it was set and changed nothing, and /yolo said approvals were off while Hermes kept asking, then left a flag that turned them off at the next resume. Both now go through the TUI's own per-session calls (Ctrl+T's and Shift+Tab's), and say the outcome under the command; /reasoning with anything but a level still goes to Hermes as typed. Typed during a turn, /yolo, /effort and /reasoning <level> run at once, as Shift+Tab and Ctrl+T do; bare /reasoning and its other forms wait for the turn's end. Shift+Tab presses made while Hermes has not answered the last one are kept and shown at once, dimmed with …, so two quick presses end in plan mode instead of in accept all; Ctrl+T's level shows the same way until Hermes confirms it. /model <name> and the picker show switching to <model>… under the command until Hermes answers, say when Hermes also made the model the default for new sessions, and a switch held for the next turn that Hermes drops as that turn starts is now a lasting transcript row naming it. Bare /model opens its picker at once: what is typed while the list loads filters it, Enter waits for the list, and Esc cancels it; picking the current row switches nothing, and the echo names the model without the flags sent. A typed /effort <level> says the level set. The info line keeps the model's name beside the mode label at 120 columns (the context goes first), and the launch look's file is named as ~/….

    • omh update no longer waits forever on a stalled package manager. The step that runs npm, bun or Homebrew had no time limit and printed nothing while it ran, so a registry that never answered left [1/2] Updating omh command package... on screen for good. The step now prints a still running line with the seconds passed about every 30 seconds, and after 900 seconds it stops the package manager and names the command to run yourself before rerunning omh update. OMH_UPDATE_PACKAGE_TIMEOUT sets another limit in whole seconds, up to one day; --json output is unchanged.

    • omher finds Hermes installed by its package manager. Hermes installs made with the unified package manager have no venv, so omher said Hermes was not on the computer. omher now starts its bridge on the Python the hermes command itself runs on (read from hermes-agent/.hermes/bin/hermes, including paths with spaces) and loads Hermes' installed dependencies before it starts, so the bridge no longer stops on No module named 'websockets'. An install upgraded from a venv uses the same Python as hermes, not the old venv. Thanks to @sykuang (#2740).

    • Vim editing in the OMH TUI composes like Vim, and Ctrl+N / Ctrl+P move every list. With display.vim_mode on, . and [n]. repeat the last completed change; W B E ge gE join the word motions; d c y take j k gg G as logical hard lines with multiplied counts; v, V and gv select, change and reselect; r waits on a steady underline (REPLACE). Ctrl+N and Ctrl+P move down and up in approvals, questions, pickers, pagers, /dag, /bots, and the CLI theme, model-chain and setup pickers, and Ctrl+N on an empty composer enters the agent list, as Down does; Ctrl+J and Ctrl+K keep editing a typed draft. Contributed by @DanRioDev (#2770).

    • Custom identities show up in the model picker, and Space turns one off. omh model and /omh-model list each identity from identities.json under the built-in categories; left/right and -/+ edit its model chain and Space disables or restores it, without touching its prompt. A disabled identity stays in the picker but is refused by routing, prompt injection and fanout. A new session's system prompt also names the enabled identities, so the main agent can pick one. Saving only the on/off state leaves the stored chain as written, and a document without the new enabled field keeps the fingerprint it had, so a conversation already routed under an identity keeps its fallback chain across the update. Contributed by @yqm-307 (#2799).

    • frontend judges a phone-used surface on density as well as hygiene, and ships a script that measures it. A new references/mobile-density-gate.md keeps two verdicts apart: hygiene (no sideways scroll, no clipped label, no small target, no input that zooms the page) and density (the y of the first data row on a named phone viewport, against a default limit of 300 CSS px that a project may change). It asks for touch and mobile user-agent emulation and names a width-only capture as width-only, and it reports each change as a before and after pair in px. The bundled toolkit/scripts/phone_check.py loads a page in Playwright at 390x844 with touch and a mobile user agent and prints omh_phone_check/v1 JSON: horizontal overflow, the first data and action y, touch targets under 44 px, inputs under 16 px and console errors, measuring overflow from the page's start side, so a right-to-left or vertical page is measured on the side it overflows (Chromium only). OMH never runs it; when no browser is available the measurement is not_observed. visual-qa records in the capture manifest whether emulation was on.

    • frontend names the layout before it builds a new page, and reviews what the page says about itself. A new references/editorial-layout.md asks for one value and one reason per axis (skeleton, first screen, section boundary, image placement, signature moment, density, screen corners and navigation), a change of at least three axes from the last page built for the same user, read from a one-line-per-page record, one frame with varied interiors across a site, and a static build reviewed before any motion is added. taste-foundations.md gains a screen-copy section: hero sentences, section descriptions, "not X but Y" headlines and numbers on the page are reviewed against the captured screen, and a number without a source goes or gets one.

    • A design contract names the rules the project's build holds. design-system-contract.md gains a section on turning a rule that can be judged from source into the project's own test or lint step, with one escape marker for exceptions, closed sets whose steps each name a role, a dated history per rule, decided and open lists kept apart, and a styleguide page built from the real components. A check counts only when it ran. reference-token-extraction.md separates a reference's skeleton from its brand and asks for a searched, not read, removal of the reference's traces before work leaves the team. design-orchestration shows direction options as rendered images and runs to the end without asking only when nobody can answer or the user handed the choice over.

    • Korean redesign and phone-layout requests reach frontend. "대시보드 다시 디자인해줘", "폰에서 가로 스크롤 생기는데 고쳐줘" and their English forms now dispatch frontend. The guard reads the whole request against closed sentence shapes (a redesign whose only object is a named UI surface, or a sideways scroll or breakage tied to a phone viewport with a repair ask), and checks the message as sent as well as after masking. A qualifier, a second clause or a question mark stands it down, so a car dashboard, a database redesign or a cracked phone screen keeps the verdict it had. Thirty routing-precision cases pin both sides.

    • show-reel's references state more of the craft rules they relied on. The brief records motion_limits per genre before scenes; the storyboard writes a motion theme as one sentence with a verb and an entity legend every renderer draws from; numbers count up only when every intermediate value was true, and a value one scene ended on carries into the next; the first frame shows a value at its final state; a vertical frame is composed for its height; the no-context judge is named in the record and runs two rounds by default before an open rebuild goes to the person; and every name the kit's checkers print maps to an edit. The skill body is unchanged.

    • The Windows installer skips an MSYS2 or MinGW Python instead of failing halfway. When python on PATH was MSYS2's mingw64 build, the one-liner and omher's Install button built %LOCALAPPDATA%\omh\venv with it and stopped at "Install OMH package" with ...\venv\Scripts\python.exe is not recognized: that build puts a venv's python.exe in bin\. The installer now asks each candidate where its venvs put python.exe, skips one that answers bin with a line saying why, and moves on to py, the next name, Hermes' own interpreter or uv. A Python running from a venv, such as Hermes' own, is still accepted. Reported and first fixed by @kinanshihab in #2847.

    • The toolkit scripts refuse a path under an unreadable folder on Python 3.14 too. From 3.14, Path.is_symlink() answers "not a link" for an entry inside a folder that cannot be searched, where 3.11 to 3.13 raised. The frontend, biomedical-evidence, geospatial, sprite-forge, molecular-docking and show-reel helpers walked past such a path unchecked on 3.14 and stopped with a bare traceback on older Pythons. Each now checks the entry with lstat first and refuses with path could not be checked. Missing paths and the symlink rules are unchanged. Reported and first fixed for the frontend helper by @Kenmege in #2855.

    • omh update no longer refuses a config whose webhook route, or any other section, carries its own skills:, memory:, plugins: or display: key. The section shape guard counted every line that named the section as a mapping key, and its pattern let \s* eat indentation, so a route's nested skills: list read as a second top-level skills section and the refresh stopped with duplicate skills sections are ambiguous; that profile's OMH skills then went stale. Only a column-0 line whose own root key is the section now counts, so plain, quoted, explicit and flow forms of a real top-level section are unchanged, two column-0 sections are still refused as duplicates, a flow or explicit-only section is still noncanonical, and only a document whose whole root is indented stays user-owned. A profile with no skills: yet beside a route's own skills: list now gets its first block. The uninstall report reads its section.key presence the same way, so a name inside another root key's flow mapping is no longer reported as OMH's own. Reported by @PabloJustDevelops in #2829.

    • omh coding model-routing status and the setup profile pack now read the model aliases Hermes itself writes. hermes config set model_aliases.<name>.model <model> stores each alias as a mapping with model, provider, base_url and key_env children, under the root model_aliases key or model.aliases. The shared alias parser read only the flat alias: target form, so such a config showed Hermes aliases (corrupt), an empty alias map and a "Repair the Hermes config" action in omh doctor. A nested entry's target is now its model child. The other children are read past, and their values are never kept or returned. The flat form is unchanged. A nested entry with no model child now names its error as model alias entry shape is unsupported instead of a malformed entry. A duplicate model child, a deeper or misaligned child, a child key named aliases, and a credential-shaped model value are still refused. Reported in #2863 by @MarinWasAsleep; first fixed by @SHURIKA6 in #2867.

    • The OMH TUI's /sessions picker no longer loses older conversations behind newer ones, and orders sessions by when they were last active. Attached to Hermes' shared gateway, the gateway child asked both the live and the stored listing for only the picker's limit, kept the live rows first, and cut the joined list there, so a session held open and its own stored history could take two rows while a conversation you resumed yesterday fell off the page. It now asks each listing for up to 200 rows, merges a live session and the stored row of the same conversation into one row (by id, resolved id or lineage root), orders the result by the conversation's latest activity (last_active_at, which current Hermes answers when asked for its picker fields), and applies the limit last; a title lookup is forwarded unchanged. The day headings and ages follow the same last activity. A row Hermes marks resumable: false / owner_runtime: "other" (a session another runtime holds open) opens read-only: Enter shows its newest messages through session.transcript in a pager instead of resuming it, and Ctrl+R and Ctrl+D are refused on its row. Only a Hermes with the companion listing change (DanRioDev/hermes-agent@01ba3239b, not yet in upstream Hermes) lists such rows; against stock Hermes every row resumes as before, and a Hermes without session.transcript answers that read with a notice, never a lost session. Contributed by @DanRioDev (#2865).

    • /resume in the OMH TUI looks like Claude Code's resume picker. It takes most of the screen, headed Resume session (2 of 12) with a rounded search box under it (the caret typed in it), and each session is a card: its name on one row and 2 days ago · 14 messages · tui muted under it, a blank row between two cards, under the same Current / Today / Yesterday / Previous 7 days / Earlier headings. ↑ and ↓ in the pointer's column say more cards sit above or below. A long name or its facts wrap onto a second row (cut there with …), and the keys are listed under the cards, wrapped between two keys, as Claude Code lists its own. Ctrl+R turns the search box into the title field and Ctrl+D asks on the card's facts row, as before. Closing it gives every row back to the transcript, with no band of blank rows left above the composer. On a terminal too short for two cards and the keys, or under 30 columns, it stays the one-row list.

    • The OMH TUI's ✳ Waiting… (… · terminal) row shows while a notify=true background process runs. Hermes writes its process record at the spawn, before it arms the completion notice, and does not rewrite it until the process exits, so the record said nobody would be told for the whole run. The TUI read it that way: no waiting row between turns, and the info line said 🔕 Hermes won't be pinged. The TUI now takes who will be told from the terminal result Hermes sent back as it started the process (notify_on_complete, or watch patterns), so the waiting row stands and the line says 🔔 will ping you when done. A result that says no one will be told (a subagent, a session that cannot take a completion) changes nothing.

    • An OMH TUI question never opens with nothing above it. Some models write the explanation that should come before a clarify question only into their reasoning and send no reply text, so with Hermes' display.show_reasoning off the picker appeared alone and the person could not tell what to pick. The question now reads that turn's row from the session's state.db (read-only, matched by the clarify's own call id) and shows the reasoning dim above the asked sentence, at most three rows, Ctrl+O for the rest. A turn that wrote a reply, an older turn's row, or a store that cannot be read draws the question exactly as before.

    • A paired run now measures how long each cell took and says whether the variant was faster at equal quality. The Hermes-child dispatcher times every spawned child from spawn to exit on its own clock and signs those seconds into the cell's receipt; a hand-edited figure no longer verifies. New paired_run_decision/v2 decisions carry the signed seconds on every result row and a speed_at_equal_quality block recomputed from the rows: the quality gate regresses if any task went from a baseline pass to a variant fail, and then no speed is claimed at all; timing covers only tasks both arms passed with seconds on both sides, and every other task is counted under a named exclusion (one arm failed, both failed, missing seconds, infra error, not observed). A missing figure stays unmeasured, never zero, and the share of timed tasks the variant won carries its denominator. v1 decisions and store history still parse (timing reads as unmeasured), and a v1 decision can be superseded by v2. No speed figure has been measured with it yet.

    • Running OMH from your home directory no longer stalls every later tool call. The consistency check's file index never starts a walk at the home directory or above it (/, /Users): from a background gateway such a walk reaches macOS privacy-protected folders, where one open() can wait forever on a prompt no one sees. That root is treated as too large to index; project roots below home index as before. Reported and first fixed by @skihyeon in #2854.

    • OMH tool results keep non-ASCII text as written instead of \uXXXX escapes, which cuts their token cost for Korean, Chinese, Japanese and other non-Latin text on every later model call. Reported and first fixed by @yqm-307 in #2884.

    • omh setup --model-setup accepts a provider configured under Hermes providers: when its key variable is set or a credential pool exists. When neither can be seen, the refusal names the variable to set or hermes auth add <id> instead of "lacks safe observed provider auth". A key kept only in Hermes' .env file is not visible to OMH; export it in the shell that runs omh. Reported by @MarinWasAsleep in #2882.

    • Vim NORMAL in the OMH TUI takes bracket text objects. With display.vim_mode on, d, c and y take i( / a(, i[ / a[ and i{ / a{ (the closing bracket works too, and counts reach outer pairs), so di(, ci[ and ya{ act on the balanced region; an unmatched bracket changes nothing. Contributed by @DanRioDev (#2714).

    • Gemini 4 Carbon and Gemini 4 Hydrogen are priced and mapped to their providers. Runs on gemini-4-carbon and gemini-4-hydrogen now get a cost estimate from Google's own list prices, cached input included, and provider filtering knows Google, Gemini and OpenRouter serve them. Neither is in a shipped chain. (#2762, #2848)

    • bench.py --extra-model no longer needs family=. Left out, the family is read from the model id. Contributed by @Vedant-200 (#2798, carried in #2806).

    • The README and omher.app open with the omher desktop app. Its screenshot comes first, with per-OS download links and the one-line install, and the site's hero window runs a live, clickable omher on staged data that follows each omher release. (#2712, #2721, #2743, #2750, #2875 and follow-ups)

  3. 4.0.7

    Parche

    GitHub Release

    • Bare omh opens on a Hermes runtime of the TUI's own by default, with your existing login and no extra command, so the bots' gateway restarts and drains never pause it. The TUI's runtime is a Hermes profile, <root>/profiles/omh-tui, whose copied config says gateway.standalone: true: the host gateway never serves it, Hermes starts an idle-exiting gateway for it when the TUI attaches, and a profile without an auth.json of its own reads the root's store and rotates OAuth tokens back into it (Hermes' profile rule), so nothing is copied and nobody logs in twice. Each launch and each omh update refresh it from the root Hermes home: the config, .env without messaging credentials, SOUL.md, one link per skill at its own <category>/<skill> path over the root's and every served profile's (the root winning a clash), and plugins/omh. A root config.yaml that connects a messaging platform keeps the launch on the bots' gateway, with one line saying why. omh theme runtime shared puts the TUI back on the bots' gateway. An explicit --hermes-home or -p stays on the home it names, and --resume opens where the session is, with one line when that differs from the choice. On the TUI's own runtime, /gateway restart restarts that runtime alone and loads new OMH code without a bot reconnecting. omh doctor's omh_tui_runtime row says which runtime bare omh opens on, what the profile was built from and whether its gateway runs, and fails on a messaging credential or a connected adapter there; the menu bar's settings carry TUI runtime: own (gateway pid N) or shared gateway. omh uninstall removes the profile, and the bot board, skill registration, the curator ledger and the skill union leave omh-tui out. The builder stages a new profile under a dot name and renames it whole, and writes Hermes' .no-bundled-skills marker. On the bots' gateway, an omh launched while Hermes' gateway restarts now waits through the drain and the restart on one shell line naming the work the gateway lets finish and the cap it allows, and opens once it is back, where it used to fall back to Hermes' own TUI and show only error: runtime_draining; a gateway that is there but not answering past the wait is reported with what to check and exits 1. Measured on a live host (tools/qa/tui_own_runtime/RESULTS.md): a prompt answered on the root's OAuth login with the root store unchanged and the host gateway untouched. Known cost, Hermes' own: on a host whose bots have scheduled cron jobs the TUI gateway stays up between sessions. (#2667, #2670, #2676, #2691)

    • omher, a desktop app for your Hermes bots, runs on macOS and Windows. It ships as open-source installers on GitHub Releases under omher-v* tags: an ad-hoc signed .dmg for Apple Silicon and Intel Macs and an unsigned per-user setup and .msi for Windows, with SHA256SUMS.txt; the app's README gives the first-launch steps. Each Hermes profile is a bot in the sidebar with one main chat. A turn shows as one activity line whose steps name what actually ran (the command, the path, the query or a helper's goal) and collapse to "Worked for 2m 41s · 6 steps"; approval and clarify cards, a model and reasoning picker and "Connect a Hermes profile" (nothing copied) sit beside it. The sessions a bot runs in Slack, Discord or cron show read-only, each as a bot bubble with its last answer and a link to the thread (Discord #cron) that opens it in its own app. Settings → All bots shows a card per bot with its status, model and sessions per place; a bot's page lists its skills and routines, and there you can rename the bot or change its picture, the only things omher writes to a profile (Hermes' display_name and avatar asset, so Hermes' own desktop shows the same). Each thread link carries its app's mark (Slack, Discord, Telegram, or a clock for cron). While omher is in the background it notifies you of a bot's answer, an approval or question waiting, or a failed turn, never for the chat you are looking at, and Settings turns that off. Keyboard shortcuts open a bot, search, Details and Settings, and Esc stops the running answer. An approval for a risky command (rm -rf, sudo, a force push, DROP TABLE and the like) gets a red rail and one plain line saying what it does, with Deny still first. Appearance offers six colour sets, Soft, Solid or Open bubbles and three text sizes, kept on that computer. omher reaches each profile's own gateway over loopback with Hermes' own Python, sends prompts only to a bot's main chat and makes no model or API call; docs/OMHER-BOUNDARY.md states the boundary. (#2677, #2685, #2688, #2689, #2693, #2694)

    • Codex and Claude Code units dispatched by OMH now come back accepted, and one dag-ulw run can send different coding nodes to both owners. Untracked executor plugin state under .omx/ and .omc/ no longer fails committed-result intake; tracked changes and unrelated untracked files still do. Claude Code's temp directory stays inside the fence, and its result-file directory is explicitly allowed. A readiness decision stale only because it came from another checkout is re-observed at dispatch. Naming Codex or Claude Code in a dag-ulw request in Hermes chat chooses each owner for that request's run; coding nodes can name their owner, and each owner's batch exports separately while cross-owner dependencies wait.

    • In Slack and Discord group chats, telling the bot to be quiet now mutes it in that thread, and a message meant for someone else no longer opens a turn. OMH registers Hermes' pre_gateway_dispatch hook. A whole-message quiet word aimed at the bot (shut up, 조용히 해, 黙って, 闭嘴, cállate and the rest of eleven languages' closed lists) mutes the thread with a 🫡 reaction and no text; one that ends in a question in any order of marks (shut up?!, shut up!?, 닥쳐?!, ‽, ¿) does not, and shut up! still does. A mention of the bot, a reply to it, or a call by its name unmutes it. A message that opens with a real mention of someone else, or calls a configured other name at its start or end without naming the bot or saying "you", is skipped. In a room where Hermes lets a silence marker stand (the serving profile sets platforms.<slack|discord>.channel_overrides.<channel id>.allow_silence: true, Hermes 0.22.6 and later, with streaming off), the model decides instead of a word list: the turn carries a short instruction to answer [SHUTUP] when asked to be quiet in any language or wording, [SILENT] when the message is clearly meant for someone else. OMH turns [SHUTUP] into Hermes' [SILENT], mutes that thread and adds the 🫡 reaction; Hermes drops a [SILENT] reply, and a message addressed to the bot is never offered it. A reply made only of [SILENT] and [SHUTUP] lines counts as that marker, and any [SHUTUP] among them mutes. Tokens in a list, a quote, a table or a fence around several lines stay an answer, a token leaves the end of an answer only after sentence-ending punctuation (The marker - [SILENT] stays as written), and only one word after a reply's opening fence is read as its label (text, c++, 완료), so prose in a fence above a [SHUTUP] line is posted as written. In those rooms the quiet-word list and calls by other names are off; a leading mention of someone else, muted threads and the unmute on a mention, reply or name stay. Every other room, an older Hermes, streaming on, or a profile on Codex's app-server runtime keeps the word-list gate and gets no instruction, so Hermes' "returned only a silence marker" notice never appears. Only senders Hermes authorizes are gated, and every skip is logged under the owning profile's OMH home; messenger_gate: off or mute_only turns it down per profile. DMs, slash commands and other platforms are untouched. (#2602, #2609, #2612, #2621, #2646)

    • Host-readiness skills now say what their commands do and stop promising what they cannot. Turning off a capability family reports what the disable actually does (memory stops; the other families are only withheld, which omh capability-policy status shows), names the command, and drops a flag that never existed. omh uninstall, from a chat phrase or a wrapper card, is previewed with --dry-run and confirmed first. Wrappers pass the chosen owner with --from-plan. A web-search setup is not called verified while a keyed backend has no key. Doctor, toolbelt, executor, provider-profile and prompt-import skills name the command or probe that produces their evidence, a started run record says how to state its outcome, and generated references no longer end in a literal \n or print an ellipsis followed by a period.

    • Several features or stages asked to be built are offered as a dag-ulw graph, in Korean as in English. A request that names three or more features as one build verb's object, with a feature noun after the list or a verb that builds software on its own ("메일 가져오기, 파일 업로드, 검색 기능 개발해줘", "장바구니랑 주문이랑 배송조회 기능 구현해", "build the upload, preview and download screens"), a count with its list after the verb ("기능 네 개 개발해줘: 검색, 필터, 정렬, 페이징", "build these four features: search, filter, sort and export"), a range from one product feature or stage of the work to another ("회원가입부터 결제까지 개발해줘", "implement the flow from cart to order confirmation") or several counted features ("여러 기능 나눠서 다 해줘") got no route hint, or a hint for a rule one item's word matched ("검색" for web-research), so the agent ran the parts one subagent at a time. The route hint now offers dag-ulw first (ask once), ahead of web-research where it sits and without pushing out a hint already there, and the router asks with dag-ulw as its candidate where it would ask anyway. Nothing starts a run on it. A list that fills one thing ("우유, 계란, 빵 장보기 목록 만들어줘") or names the fields of one screen ("이름, 이메일, 전화번호 입력 화면 만들어줘"), three plain nouns and a make verb ("make pancakes, waffles and toast"), non-software objects ("build confidence, momentum and credibility"), build targets ("build dev, staging and prod"), work quoted after a colon or a full stop, a range between any two words ("봄부터 여름까지 만들어줘", "build it from soup to nuts"), of numbers, pages or dates, or from the start to the end of the work at hand, a count that is not the verb's own object of features ("세 페이지 보고서 만들어줘"), verbs that change what exists, stated obligations or intent, a comparison, a question, one fix, a message over 1000 characters and every hard exclude of the many-targets reading stay off it. Hint text and routes are otherwise unchanged. (Reported and first fixed by @dhddldid in #2610 and #2611; #2614, #2620.)

    • Ctrl+Enter in the OMH TUI sends now into the running turn, not after it. Ctrl+Enter, Ctrl+J on an empty composer and the text typed after a denied approval went as Hermes' session.steer, which only holds the text until a tool batch ends or the next model call. When the model was already writing its final answer neither came, and Hermes sent the text as the next prompt once the turn had ended. They now go as session.redirect, as Hermes' own TUI sends busy input: a model request under way is cut where it was, its reply sealed, and retried with the text as a user message, so the retried reply streams under the steer row; a tool batch running gets the text with its results, and a foreground terminal command moves to the background so the batch can end. The row reads as a sent prompt at once, and stopping the turn afterwards no longer puts it back in the queue. A redirect Hermes queues as its own next prompt (the turn still starting, or compressing) comes back to the queue on a stop, which clears Hermes' queue. An agent without redirect, or a redirect refused between a tool batch and the next model call, gets session.steer. A redirect that cuts reasoning while it streams seals that reasoning, so the retry's reasoning reads as its own block. When a turn ends on its own with a redirect or a steer Hermes passes on as its next prompt, a queued line waits for the turn that runs it; if none starts within a few seconds (the turn's last model call read it), or the turn fails, the queue goes on, and a line sent during that wait stays behind the older ones. /steer <text> stays Hermes' steer, read after the next tool call without interrupting. (#2603, #2609, #2621)

    • omh update keeps a stable install on stable releases, and --dry-run says when a real run would update the command package. The installer-managed self-update still defaulted to --channel preview, so a plain omh update after the documented stable install downloaded the main archive, while omh update --dry-run printed "Command package: unchanged (preview only)". With no --channel, --version, or --package-url, the self-update now reads the installed package's PEP 610 record, as omh setup has since #2294: a stable install moves to the newest published release and is left alone when it is already on it, when it is newer than npm's latest (the GitHub release is created before npm publish, so latest can trail it for a while), or when the newest release cannot be read; a main install, or one that records nothing, keeps tracking main. A dry run now plans like a real run and its card reads "Command package: a real run would update it to v4.0.7". (Reported and first fixed by @aikaizen in #2606, carried in #2607.)

    • omh coding model-routing status names each discovered model once. Session history yields one observation per session, so the human line repeated one model id hundreds of times. It now reads claude-opus-5-5 (x412), claude-sonnet-5-5 (x61); --json keeps every observation. (Reported and first fixed by @aikaizen in #2606, carried in #2607.)

    • A route can name commandcode once its key is in Hermes' .env. omh_delegate_route refused every route naming the provider. With COMMANDCODE_API_KEY in .env such a route now applies, and a Hermes login or credential pool row for the provider, or a providers.commandcode block, counts as well; without any of those, or with the key only exported in a shell, the route is still refused. The key or login alone does not reorder any model chain or change which category routes work. The isolated Hermes child lane passes the key to a commandcode child. Separately, a line with an empty value (X=, X="") no longer counts as a key for any provider, because Hermes has no credential to use from it. (Reported and first fixed by @levinkin in #2636.)

    • A reply that opens with a code block keeps its ⚚ Hermes header in the OMH TUI's scrollback. When a reply's first line was a code fence and it streamed in small pieces, the fence line alone used the reply's first take and drew no row, so the header never reached scrollback; /history had it. The header now waits for the reply's first drawn row. (#2598)

    • The OMH TUI waits out a Hermes gateway restart run elsewhere instead of exiting. hermes gateway restart from another terminal, omh update, launchd or another screen's /gateway restart closed this screen's connection, and the crash respawn then spent its three tries within a second: a draining gateway answers "not ready" (the attach child's exit 69) until its running work is done, so the TUI exited with a resume hint. Now a drop where the gateway's record (gateway_state.json) says a restart was asked or a new process is starting, or where the attach child exits 69, keeps the screen open with one status row (Hermes' gateway is restarting… 0:42 · ctrl+c leaves) and reattaches the same session, keeping the draft and holding prompts as before. While the old process drains for the restart, a try goes every 5 s (Hermes' after-turn wait lets a client answer a parked question) and the wait lasts as long as the drain, at most Hermes' restart_after_turn_timeout plus 150 s; otherwise tries pause 1, 2, 4, then 5 s for up to 150 s. Once back it says Hermes' gateway is back and, when the restart loaded a newer OMH, which version runs now. A gateway stopped elsewhere (hermes gateway stop: no restart asked) or a new one that failed to start ends the screen at once with the resume hint, so no try starts a stopped service again; a stopped gateway is said plainly, as a state, with hermes gateway start to start it again. The wait reads HERMES_RESTART_AFTER_TURN_TIMEOUT before the config key, as Hermes does. A try that fails any other way, or a resume the gateway refuses, still spends one of the crash respawn's three; a gateway that never comes back ends after the budget with the same resume hint. /gateway restart's own reattach shares the pauses, the drain wait and the status row. (#2596, #2609)

    • A tick-box question in the OMH TUI no longer sends on the first press. When Hermes asked a multi-select question ([ ] boxes), Enter on a choice sent the question at once: with nothing ticked it answered with the highlighted choice alone and moved on, so a person who pressed Enter to tick lost the rest of their picks. Now Enter, Space and a digit on a choice only tick or untick it, as in Claude Code's question. The list ends with a Submit row (Next while more questions follow), and Enter there sends the ticked choices with any text typed in the Type something. field, which stays on that row when the arrows leave the field; with neither, the key row says Enter to skip and the question reads as skipped. Enter in the open field still sends the typed answer with the ticked choices. A question with several parts shows each one before the last is sent. A single-choice question is unchanged: a digit or Enter still picks and sends.

    • The OMH TUI draws Markdown tables as boxes, and a table wider than the terminal wraps inside its columns. Every table is now a box like Claude Code's: ┌─┬─┐ borders in the rule colour, the header bold in the look's brand colour, a ├─┼─┤ line under the header and under every body row, and a cell of padding inside each side. This changes the look of tables that fit too: they no longer use the ─ band under the header. A table wider than the terminal used to show its raw source (| a | b | rows and the |---|---| divider). Now its short columns keep their whole width, and the long columns share the rest of the row. Each cell wraps inside its column, and a word too long for the column breaks across rows (a path or URL after a /). Cells show their inline markup (code, bold, links), and a code span keeps its style when it wraps. The borders are not copied by a drag; NO_COLOR keeps them. Only when a terminal is too narrow for columns of 8 cells is each body row drawn as header: value lines instead. A table taller than the live region is now also released by the rows it draws, not only by its source lines. In scrollback, its last released row has no bottom border until the table ends; the live region shows the box closed in the meantime. A row that arrives after the release keeps the columns already drawn at least as wide as they are, and takes the rest of the row's width for the cells that want it. Its border line meets both rows' columns. /history lays the whole table out again. A table that follows a released one, with no blank line between or with a carriage return (a CRLF reply), a no-break or ideographic space or a terminal escape in its divider, gets its own box in scrollback, as in /history: while a tall table streams, its last row waits until the next line shows it is no divider. Alignment colons (:---:) are still not drawn. (#2590, #2592, #2598)

    • A memory captured again while an earlier copy waits for review leaves one copy, not two. The earlier held candidate with the same normalized summary, the key the duplicate check already reads, in the same scope and observed of the same agent, is retired with retirement_reason: recaptured and recaptured_as naming the new one, so the same summary observed of codex and of claude-code stays two memories. Approval and rejection refuse the retired copy, the review queue drops it, a demotion of the same Hermes entry reads it as held, and the capture result lists it. The retirement runs only after an auto-safe approval succeeds. Approved and rejected candidates are left as they are, and no person is asked. (#2652, #2659)

    • A Homebrew Python patch upgrade no longer breaks a self-updated omh. omh update created each generation's venv from the running venv's interpreter, so python -m venv linked it to the resolved base, the version-pinned keg (/opt/homebrew/Cellar/python@3.15/3.15.0/...); after brew upgrade and cleanup the launcher failed with bad interpreter. The update now reads pyvenv.cfg in the running venv and in the bootstrap venv, follows each command back to the interpreter the installer ran, then tries each home, and uses the first that runs the same major.minor and ABI flags; otherwise it keeps the old choice. The bootstrap venv matters after three or more older updates, when cleanup has deleted the middle of the chain. No vendor paths: python.org, pyenv, uv and Windows (Scripts\python.exe) venvs are read the same way. omh doctor reports a generation whose base is gone or version-pinned as generation_base_python and prints the repair command; Installation gives the same repair for an omh that no longer starts. (Reported by @junjuny0227 in #2683; #2687.)

    • show-reel asks through the spend gate before it spends the user's ElevenLabs or codex credits. Narration, voice, transcription, sound, music and image runs on the user's own key or login, and each codex exec text run for script review and concept candidates, now go through omh-spend-gate first, at every step that runs them: a card names the exact lines, voice and model or the job and its length, with the expected credits from the provider's own price for that model (or says the cost is unknown), and the person's typed yes to it is that run's exact-target approval. omh-spend-gate now says a call billed to subscription credits or plan usage is a spend, with its own row in the rails table, a receipt of the output and the credits used, and a ledger row kept in credits. When no price gives a call's credits or cost, the person sets a spend-only cap, which the rails page names, in the call's own units (images, characters, seconds or calls), and the call runs only when its planned count fits; its ledger row keeps that count and leaves the period total unknown, so the next spend asks what was already spent. The scripts are unchanged: they run only when invoked and exit 2 without a key. (#2653, #2662, #2678)

    • show-reel renders faster. The toolkit's renderer no longer makes headless Chrome wait on a vsync tick for each captured frame, so a light 6-second 1080p scene renders in under a fifth of the time it took, and heavy scenes, whose paint cost is real, gain less. Renders with the same settings stay byte-identical run to run, and the quality settings are unchanged. (#2622)

    • The round-end checks also run after edits made through the shell. Hermes calls OMH's round-end hook only after a patch or write_file, so a turn that edited files only with sed, a heredoc script or cat > got no export, test or caller check at all; about half of Sonnet's coding turns edit that way. Now, after a terminal or execute_code call that changed a Python file in a turn with no patch / write_file, OMH builds the same three notes from the files on disk and adds any that name a broken line to that call's result. Each check speaks at most once per turn, so an early caller note does not silence a later test note. When one message holds several commands, only the last one is looked at, so a note never names a line a later command of the same message already fixed. A command that changed no Python file, and every turn that used patch or write_file, is left as before.

    • The OMH TUI briefs each subagent as it finishes, in the transcript. A finished subagent only changed its mark in the agent list under the composer, so during a long dag-ulw run nobody saw what each child concluded without opening its log. Each subagent.complete now writes one block above the composer, as Claude Code does for a teammate: ● <name> finished 1m 35s · ↓1.8k, then the first lines of the child's reply in up to two dim rows, cut with …, with the reply's Markdown drawn as the reply is (no **, backticks or #). A dag node's child reads ● build @notify finished: the role the node was defined with (verify, review, baseline), which the delegation graph now carries, then the node's id as a handle, bold in a hue of its own, with the node's label on one dim row under it. The hue is picked by the node's place in its run's definition order, from seven hues kept clear of the running, done and failed colours, so the first seven nodes of a run never share one and neighbours never do; an agent keeps its colour on every row, every redraw, in /history and when the run is read again, and each hue reads at 4.5:1 on every launch look and scene ground, light and dark. A child whose node the HUD has not bound yet takes the id its goal's [dag:<run>/<node>] marker names, with no role until the graph holds the node; a graph from an OMH that exported no role shows none, never one guessed from the label. Any other agent keeps its bold short title, without a role or @, since Hermes' subagent frames carry no name to make a handle of. Only a completed agent's mark is green: a failure reads ● <name> failed in red with its error, an interrupted child ▪ <name> interrupted, a cancelled one ○ <name> cancelled. A briefing that arrives while a reply streams or a call (delegate_task) runs waits under it and is written at the next block boundary, so it never splits a paragraph or a table (/history and Ctrl+O show it meanwhile); past three waiting under one running call, or past what the rows above the composer hold beside it, the rest fold into one ● 9 more finished · 1 failed row until the boundary writes them all. On a screen too short to hold the running call, the fold row and the spacer between them, the fold row first drops its spacer and is then not drawn at all, so the call stays on screen. Parallel children are briefed in the order they finished, and main never is. Briefings are not part of the session Hermes stores, so a resumed or rewound session is drawn without them. The rows erase on a scene look's ground and keep every mark and word under NO_COLOR, where the handle is plain bold. The agent list and the dag tree under the composer keep their state colours, and the messenger is unchanged. (#2586, #2591)

    • dag-ulw no longer runs silently with Hermes' delegation.worktree_isolation on. With that setting each delegated child works in a git worktree of its own, made from the parent's last commit, so a dependent node and the verify node started without the earlier nodes' edits, and verify could pass a tree without the work. omh_dag action=define is now refused as worktree_isolation_on with a line the model relays, naming hermes config set delegation.worktree_isolation false, or do the work without dag-ulw. A run defined before the setting was turned on still settles its nodes but hands out no delegate_task call while it is on, and status says why. OMH asks Hermes' own getters first, so its managed config and environment count; outside Hermes it reads the profile's config.yaml the way Hermes reads the key (a quoted 'false' is on). It never writes the setting. On a Docker or SSH terminal, where Hermes skips isolation, nothing changes; a config that cannot be read counts as off, as before.

    • The code-quality skills now say what to do where they used to dead-end or contradict. Part of one round over all skills, split per lane for independent review and rollback. codebase-uml and codegraph-refresh stop on a repository the Python-only generator cannot read, and a stale codegraph is defined by generated_at against the newest commit. native-debugging picks the instrument by fault class (sanitizer build, Valgrind, rr, core file) and scripts lldb or gdb in batch because Hermes has no debug-adapter client. build-failure-triage no longer calls a flake introduced on one green merge-base run, takes the origin of a CI-only failure from CI history, and notes that a pull-request run builds the merge ref. verification-gate defines its thorough lane and where the required checks come from, and flaky-fix runs are sized from the observed rate. git-workflow adds the bisect exit codes, a dirty-tree check before reset --hard, and cherry-pick, detached HEAD, lost-commit and committed-secret procedures. refactor-plan, ai-slop-cleaner, frontend-refactor, tech-debt-audit, rust, commit-pr-authoring, github-event-ops, github-issue-intake, repro-audit, agent-debug, app-debugging, code-review and codebase-onboarding get their own smaller fixes. No backward-compatibility layer: the producers and their generated files ship together.

    • Making a model smaller now has its own skill, omh-model-compression. "Quantize this model", "distill this model into a smaller one", GPTQ, AWQ, GGUF, FP8, bitsandbytes, pruning and their Korean phrases (모델 양자화, 지식 증류, ...) reach a plan that sizes the target from the hardware, memory budget and runtime, tries the cheapest method first, keeps calibration data apart from the held-out eval, and keeps a smaller model only when it holds quality against the uncompressed baseline on the person's own task, with perplexity or KL divergence as a secondary signal. Its reference quotes each command with the page it was read on. A phrase that other fields share ("compress the model", "structured pruning", "8-bit quantization") routes here only beside a word of model weights, so a Blender asset, a business model, fruit trees and audio stay where they were. inference-serving, model-finetuning, model-optimization and rules-distill now name the new skill in their Do Not Use lines.

    • The research and knowledge skills say what an expert needs and stop pointing at dead ends. omh-long-document-reading runs the Hermes pdf skill's own environment instead of telling the model to pip install, keeps its range ledger in omh_document_plan so it survives compaction, gives delegated children the path and command they need, and names the encrypted-file, printed-page and lookup cases. analyze.py in omh-geospatial-analysis refuses Web Mercator as the measurement CRS for real layers, since it inflates area by about 1.6x at 38 degrees; the synthetic fixture still runs. omh-data-analysis computes with code and states n, effect size and an interval; omh-live-info runs the lookup and checks provider time, units and staleness; omh-systematic-review says its delegated reviewers share a model and asks for deduplication first; omh-molecular-docking says what an exhaustiveness-4 score cannot support and what to do with a SMILES; omh-wiki hands to llm-wiki only when installed (it is optional on current Hermes).

    • A research run answers only from sources on the exact entity the question names. Before answering from a source, ulw-research now confirms it studies the gene, enzyme, strain or product the question names: a different name is a different entity unless a source says they are the same, so a source on a related one is a lead, not an answer. (#2600)

    • Literature screening for a systematic review has its own skill, omh-systematic-review. It runs the screening step of a systematic, scoping or rapid review with two independent reviewers: freeze the criteria, prepare seeded blind batches, give each to its own delegate_task reviewer, merge, adjudicate. omh research screening freeze, prepare and merge do the parts a model cannot vouch for itself: a protocol frozen once and refused when edited, batches that carry nothing from the other reviewer, and a merge that never invents an exclusion. The merge lists agreements, conflicts and pending records with the reason per reviewer, and exits 1 until adjudication and missing answers are done. Agreement comes with Cohen's kappa, PABAK and what the figure left out (unsure and pending records), and same-model reviewers are flagged as broken independence. Everyday "review", "screen" and "criteria" requests (pull requests, job applicants, support tickets) stay with their own skills. (#2589)

    • omh research eval litqa2 scores the research skills on a frozen LitQA2 sample. prepare checks the dataset's digest and writes a seeded sample of task files, with the answer key kept apart; grade reads each answer's last ANSWER: letter with no model judge and reports accuracy, precision, unsure and unanswered rates and citation-receipt coverage per arm, with and without ulw-research. The dataset is never shipped; the docs say how to fetch a pinned revision. (#2578)

    • A setup-profile default coding owner is no longer taken for one named in the request. omh chat interact and omh chat session start used to resolve the profile's default_executor before the wrapper saw it, so the wrapper labelled it explicit: it skipped the genre veto that holds a non-coding request back from an external coding CLI, and it was recorded as an explicit choice the person never made. The wrapper now reads the profile itself and labels it setup_profile, the route decision is the recorded-owner one, and nothing is recorded. An owner named in the message now beats the profile default on every entry point: with the profile on claude-code, "fix the login bug with codex" runs Codex through chat interact, chat session start and coding delegate, as it already did through the plugin omh_interact tool (those three ran Claude Code before). With no profile set, coding delegate now takes that named owner too, where it fell back to generic. --executor still wins over both. A CLI counts as named only as the one to do the work ("use codex", "with claude code", "have codex fix", "codex로", "codex한테", "클로드 코드로", "open this in codex"): "fix the codex adapter tests" or "update the claude code docs" names what the work is about and chooses no owner, on every entry point and in the route decision, where a bare product name used to pick the owner. A dag-ulw request reads the same cue: "/dag-ulw fix the codex adapter tests" no longer sends a node that names owner: codex to Codex. The CLI addressed as the one to act counts too ("codex가 고쳐줘", "codex, fix ...", "get codex to ...", "codex please ...", "can codex fix ...", "codexに…させて"). When a request mentions a coding CLI but not as the one to do the work, and the profile default is a different CLI, nothing runs: the owner is asked for, with the mentioned CLI named first ("fix the login bug in codex", "codex에서 로그인 버그 고쳐줘"). A CLI the request excludes ("don't use codex", "codex 말고"), or names only inside a link or a filename, is no such mention, and naming two owners as the doer asks on every entry point. The plugin tool and the turn's route hint now say owner_set_in_setup_profile when the profile holds the owner, where they said the owner had to be asked for.

    • The operations and board skills now name the commands and Hermes surfaces they depend on. cancel lists what exists before clearing and hands a running scheduled job to hermes cron pause and runs, and to hermes cron cancel where Hermes has it; running-work-board names omh coding status-board; run-efficiency names omh runtime efficiency-report and its input; agent-board ships a reference with the prepare fields, the supported operations and the refusal reasons instead of pointing at repository docs; gateway-intent-card states the platform:chat_id:thread_id target, the MEDIA: attachment line and when [SILENT] is dropped in a live chat; harness-session-inventory says where a previous coding session lives and what a snapshot file holds; achievements says a snapshot exists only after the dashboard tab ran. context-budget-review, achievements, ops-review, feedback-triage and gateway-intent-card write their own completion checklist in place of a category default that did not fit them, the must-keep file in context-budget-review now carries item_classes so the class comparison can run, feedback-triage ranks by distinct customers and masks personal data in quotes, and spend-gate stops treating an unknown amount already spent as zero and names the pay, pay-cli and hermes-mpp plugins instead of presenting the removed mpp-agent skill as current. gateway-intent-card and automation-blueprint, harness-session-inventory and workspace-audit, agent-board and running-work-board each name the other for the requests that belong to it.

    • Memory, learning, agent and Jev skills now give steps that run. A fact with a relative date ("next Friday") is captured again with an absolute date in the same turn instead of waiting for a review nobody does, and the memory tool names the right next step for each reason it holds a capture (protected content says there is nothing to approve). Recovering from one wrong memory source archives live records with omh memory attention --tier archive or the omh_memory retire action (the old omh memory retire step refuses a live record), and says that hard delete accepts only an expired volatile record. decision-recall names its command and required flags and says an empty result is normal on a default install; "what was decided" now goes to omh_memory recall, not to the rejected-only store. jev-action-check treats a missing tool or a non-answer as a hold, jev-review-gate sends the matching test diff with each file and says what to do past eight files or an over-size diff, and jev-done-check and jev-route cover the over-limit and no-route-question cases. The instrumentation ladder grades one way, workflow-learning names omh learning route-signal, instinct-ledger and rules-distill count observations the same way, and agent-instructions gains a size rule for Hermes' context-file cap. The full-profile skill-body ceiling is re-derived by its written policy.

    • Several too-wide routing words stopped taking turns that were not theirs. Bare priorities, 리스크, 블로커, 우선순위, risk analysis, release gate, accessibility and 접근성 on voice input, and the video words on the connector card no longer pull a turn. Missing or unset words alone no longer open toolbelt readiness unless a connector or tool is named, "주간" no longer reads as a stock price, and "time instead" no longer reads as a time lookup. native-debugging's routed text stops promising a debug adapter Hermes does not have. A few phrases the skill audits named now reach their skill: account takeover and suspected unauthorized access, telemetry audit, write a handover guide and 인수인계 문서 써줘, and in Korean 시장 규모 and 경쟁사 분석 right before a request to size or analyse.

    • A capability family turned off with omh capability-policy disable is no longer routed to. Its skills used to lose their tools but kept being dispatched, offered as candidates, hinted and listed on the skill-candidate line. Now none of those name them. A request the router would have dispatched to one, or that names one, is answered with the family being off and omh capability-policy enable <family> instead of being routed elsewhere. When the router was only asking about one, it asks an open question instead, noting in the reply's own language that the family is off and how to turn it on, rather than offering a different skill in its place; the per-turn hint names no skill for it either. omh_interact, omh chat interact, omh chat route-hint, omh_recommend (plugin and MCP), route-answer checks and the per-turn hint read the policy from the OMH home on every call, so a profile-scoped home answers with its own, and a policy file that cannot be read is reported. A home with nothing turned off routes exactly as before. The operations family's card now also lists operating-rhythm, ops-review and reliability-review, which belonged to it but were missing from the card.

    • A drag copy in the OMH TUI inside a herdr pane now reaches the clipboard of the machine you are looking at, including over SSH. herdr strips TMUX from its panes and keeps the SSH variables its server started with, so on a Mac the TUI used to put the copy on that Mac's own clipboard with pbcopy and still say "copied", even for a person attached over SSH from another machine. In a herdr pane the copy now goes out as OSC 52, which herdr hands to its client's clipboard. A copy past the 192 KiB herdr takes says it was not copied, and so does /copy. A tmux, screen or zellij started inside the pane, or another terminal opened from it, keeps its own rule. Reported by @foxion37 in #2672.

    • A turn whose message matches no stored memory gets no memory block. The live pack used to list such records by name in one English <record_index> line, which Hermes wraps as context that "should inform all responses". With a persona that sets no reply language, that line alone moved a Korean-writing user's replies to English. Those records are now withheld and named nowhere; with nothing else to serve, the provider returns nothing. Records from this conversation and pinned records still show in full, and so does a record whose subject or tag the message names in another form ("tickets" for ticket-tool-db-choice). omh_remember recall finds a withheld record by subject or words, and a recall that matches nothing, or an empty one, now lists what is stored (subject, type and id, without text), so "what did we decide last time?" reaches the decision in two calls.

    • The per-turn primer names only skills a model can load. The generic tool map, workflow context cards, common cues and each lane's Key: list named catalog names (research, img-summary, deep-interview) and shorthand that was no skill (automation), so one primer showed both research and ulw-research, and only the rendered rail page rewrote one of those lines. They now write ulw-research, omh-image-cards, ulw-interview, omh-routing and the other labels skill_view loads. (#2647, #2663)

    • The per-turn hook and omh commands start faster. The hook that runs before every model call no longer reopens every run's unchanged progress files, so its cost stops growing with the runs kept under the OMH home. omh <subcommand> loads only that subcommand's code, so the commands agents call during workflows (omh hud, omh runtime record, omh memory status and the rest) no longer spend most of their start-up loading the whole command tree; help, --version and abbreviations still load it. (#2644)

    • omh coding delegate --from-plan no longer picks Codex when no owner was named. Leaving out --executor used to select Codex with --from-plan and generic without it. Both now resolve the same default: the setup profile's default executor, else generic. Codex is selected only when it is named or is the setup profile's default executor. The hermes plan wrapper template and plan-accept --write-context-pack follow the same default instead of writing codex, and the omh-routing text that said the flag selects Codex now says what it resolves.

    • Business and operations skills now check what a practitioner checks. The morning brief no longer claims a Gmail connection is draft-only from its scopes (every Gmail scope that creates drafts also sends); it enforces read-and-draft through the mail server's tool list and takes that list as evidence. Pipeline review asks for the quota and keeps stage movement unavailable from a single snapshot. Lifecycle experiments check sample size before approval. Legal review inventories the clauses of a contract and no longer writes redline language without a playbook. Control testing starts with design, tests the reports it relies on, and says what one deviation means. Data sharing scans national-id shapes, counts names only against a known list, ties the scan to the file digest and no longer reads as anonymous. The award score uses CSSDA's published thresholds and adds Awwwards. Buzz setup names the real access keys and the attachment command. Document-skill pointers work on Hermes before and after it folded them into one Office skill. Completion checklists, examples and boundaries of the sales, product, support, people, meeting, cto and content skills are specific to each job.

    • Design, media and operator skills now gate and describe what they actually do. The sprite-forge helper measures alignment on the source sheet, so a misaligned sheet fails the gate instead of reading back the row it had just set, using the largest deviation from the median so one lifted frame cannot hide. --free-frames frees only the named airborne frames and the rest stay gated. It also reports key-colour spill. Visual QA and the accessibility audit return only their own verdicts and gain rule IDs for focus, reflow, target size, status messages and non-text contrast. Show reel no longer points at flags and downloads its toolkit does not have. The browser, terminal, files and connector operators confirm the exact payload and the paths a bulk change touches, and image cards read generated text back.

    • The build, infrastructure and security skills give safer, more exact steps. A leaked secret that is compromised (pushed publicly, or used) is revoked before its replacement is deployed, matching the rotation reference, and a planned rotation keeps its overlap window. Model optimization means onboarding a model into OMH's router only; making weights smaller or faster is pointed to inference serving and fine-tuning. The backend error table gains rate-limit and payload rows and a correct timeout status. The database skill warns that EXPLAIN ANALYZE runs the statement and covers index rebuilds. The LLM app skill drops a retired model id, plans for model retirement, accounts for run-to-run variance and permission filtering in retrieval. The serving runbook gains a startup probe, endpoint authentication and a measured quantization delta. Data pipelines cover late data, exactly-once and topic retention; infrastructure changes treat a saved plan as secret; mobile release adds the Play closed-test and toolchain gates and a rejection path. Deploy, idea-to-deploy, reliability and production-audit lose their generic fallback text, and the incident, security-event, threat-model and safety-review references close their contradictions and gaps.

    • A skill now hands work only to a sibling it can load, ends with its hand-back line, and agrees with its siblings about who does what. A hyphenated sibling a rendered skill names in backticks, anywhere in the body or a reference, a hand-off written in plain text and each body's Current lane sample of names are written as the label skill_view loads; a one-word one (research, plan) only when its line hands work to it (use, route to, hand back to, from, workflow), so a config key, a command or a DOM event that shares a skill's name stays as written. Trigger lists, metadata lines, fenced recipes, lines crediting an upstream project and the router's registries keep the catalog names, and the shared rail's generic-tool checkpoint speaks loadable labels too. The workflow-artifact and omh-jev-* bodies, which ended on a stray fragment, carry the same lane section as every other skill, and a full contract names references/procedure.md only for a skill that ships it. omh-meta-router points at the catalog index through skill_view, because the file belongs to omh-routing. An accepted plan leaves omh-cto-loop for ulw-work instead of going back to planning. omh-decide keeps Rejected records in the decision log and no longer says omh-decision-recall reads them. dag-ulw keeps an accepted plan in omh_todo. closing-a-story.md maps ship, no_ship and QA pass or fail onto the PASS, HOLD or BLOCK a record accepts, and says what each does to completeness, checked against the store. omh-rules-distill names the memory skills, omh-instinct-ledger the artifact omh-memory-new declares, and omh-workflow-learning every route-signal destination. Script paths in omh-sprite-forge and omh-show-reel references resolve from the skill's folder, and portable readers are told which steps need Hermes. The payload-density floor is re-derived, because those sample names had been counted as payload. (#2623, #2643)

    • A coding-owner choice now counts only for the request it was made in, in dag-ulw as in ulw-maestro. dag-ulw sends a coding node to Codex or Claude Code only when the person named that owner in the dag-ulw request that defines the run; the run keeps that choice for its later turns, and nothing else carries over. A choice from an earlier request, from omh chat or omh coding, or from any stored record no longer starts an external coding CLI, so the one-day window is gone. A coding node that names no owner runs on delegate_task. ulw-maestro's owner rule says the same: an earlier, recorded or learned owner is not a choice, and it asks again.

    • omh hermes plan --record can record the reviewed plan, and omh hermes plan-accept refuses a plan that is only the generated scaffold. The recorded file held the same generated goals, options and risks whatever the plan was; the reviewed text stayed in chat, and accepting the file accepted the scaffold, so a coding handoff could carry the scaffold as the accepted plan. --plan-file <path> records that text in place of the generated sections (plan_body: reviewed). Recording without it still works, and stdout and the recorded file both say it recorded the scaffold only: quality_gate.readiness: needs_reviewed_body, coding_handoff_ready: false, next_action: record_reviewed_plan, and decision_gate.acceptance_requires, repeated on coding_delegate, whose available still means available after acceptance. Accepting such a file exits 2, leaves it a draft, and names the fix: record the reviewed plan with --plan-file and accept that file. --allow-scaffold accepts the scaffold on purpose and returns reviewed_body: false, and re-accepting an already accepted file still succeeds. The hermes plan accept <path> alias refuses the same way. --input-bundle records a consensus bundle as hermes_plan_input/v1, planning input that plan-accept refuses. ulw-plan, omh-plan, omh-adversarial-consensus, the omh-routing wrapper reference and docs/INSTALLATION.md use and describe these. (#2638, #2650, #2660)

    • dag-ulw review nodes route to the review category, and baselines stay low. The skill's difficulty-ladder rule now says a role=review node takes the review category where the user's routes serve it, a baseline node stays quick or unspecified-low, and a verify node, which judges the run goal and baseline exemptions, climbs the ladder like any judgement. Before, the planner put code reviews on unspecified-high in some runs and on review or unspecified-low in others. A review node now costs more: the shipped review chain runs the same models as unspecified-high at xhigh instead of medium. omh_dag define, repair and extend also return the advisory lint review_not_review_category:<id> for a declared review node on another category. Like every lint finding, it never refuses a run. It is not shown when Hermes' model_map backend offers no option for review but does for unspecified-high. Verify nodes are not linted, and neither are nodes whose label only reads like a review.

    • With OMH_TUI_MOUSE=1, a drag in the OMH TUI copies the transcript without the rail. The main screen then takes the mouse: a drag over the transcript or the live rows selects the text (shown reversed) and copies it without the thin ┃ rail or a reply's margin, which stay drawn as before; the row above the composer says copied N chars, and a double click copies a word, a triple click a row. The wheel up opens the transcript view (Ctrl+O's), one notch up from where the screen was, and scrolling back down to its end closes it, so the composer is back without Esc or q (a view opened by Ctrl+O or /history still stays open at its end). It works on every terminal that reports the mouse, inside tmux with mouse on or off; a row the TUI cannot map for certain (the shell's output above the session) is never guessed at, a drag whose live rows change under it is dropped and copies nothing, Ctrl+C right after a drag only drops the selection, and the terminal's own selection is still there with its bypass modifier (Shift, Option in iTerm2, Fn in Terminal.app). It is off by default: without it the terminal keeps its own wheel scrolling and selection on the main screen, and the wheel never opens the transcript view. (#2575, #2597, #2601)

    • The OMH TUI puts two blank rows between two tool calls of a turn. A run of calls now reads apart, as Claude Code's does; a folded row of loaded skills counts as a call. Every other pair of blocks keeps one blank row, and what a call printed under ⎿ stays right under it. When the running calls do not fit, the oldest gaps in the live region give up their second row first, only as many as needed, so no call row is lost and no blank rows gather over the working row; scrollback has both.

    • The OMH TUI shows a preview of an image you attach. An image sent with Ctrl+V, a drop or /image is drawn under the prompt that sends it, through Hermes' own image.render: real pixels on kitty and Ghostty, truecolor half-blocks on other terminals (Terminal.app, iTerm2, WezTerm, tmux, screen, VS Code), at most 64×16 cells. The [Image #N] label becomes a caption with the file's name, size and bytes ([Image #1] shot.png · 1280×720 · 979 KB). NO_COLOR and Hermes' display.inline_images: off turn the preview off, and a Hermes without image.render keeps the label row. Images in tool results and in a resumed session's history are not drawn. (#2645)

    • The router no longer picks a coding owner from earlier choices. Three accepted answers to the coding-owner question used to make that owner a visible default for later requests. Now every request asks for an explicit choice unless it names an owner or the setup profile holds one, so the "Override learned coding agent" and "Reset learned coding agent" card actions are gone. The stored choice history stays readable in omh coding model-routing status and clearable with its reset, but it selects no owner. ulw-maestro now names the setup-profile default as a choice that stands, and dag-ulw says a newly named owner needs a new run id. omh_dag define now refuses to define a cancelled run id again instead of handing back the cancelled run, and omh coding model-routing status reports stored choices as recorded history that selects no owner.

    • The round-end check continues a turn less often. A changed function's name in another file no longer continues the turn when every mention there is proven to need no change: another object's method of the same name, or an import whose calls bind the same way before and after the edit.

    • OMH recognizes GPT-6 Cyber and the other models earlier rounds missed, and puts none of them in a shipped chain. GPT-6 Cyber (2026-11-05) gets an exact contract from OpenAI's pages: a 1.05M context, 128K output, and OpenAI's 20/100 list price. Its effort ladder and calibration come from GPT-6 Astra, because the system card calls it a variant of Astra. OpenAI serves it only to approved Daybreak Red organizations, so like Mythos 5.1 it is recognized and priced for someone who asks for it by name. GPT-5.6 Cyber (12.5/75, Daybreak Red) and GPT-6 Rosalind (15/75, approved Rosalind Research organizations) are access-gated too. They get a price and a label, and they use the gpt family calibration. They get no provider check, because no vendor page says whether Codex serves them and OMH does not refuse an approved account on a guess. Cost estimates now cover GLM-5.3-FlashX-Legal, Mistral Large 4 (mistral-large-4, mistral-large-4-0) and Solar Mini 4, at each vendor's list price. deepseek-v4.1-flash-1017, the version deepseek-flash has served since 2026-10-17, uses the V4.1 Flash contract, price and calibration, and its runs keep the pointer's category.

    • The model-chain picker steps each model through its own efforts. -/+ in omh model-chains and /omh-model now follow the ladder the model's contract documents, so Kimi K3.1 Pro walks low, high, max, Claude Fable 5.1 low to max and Grok 4.7 low to xhigh, and none appears only where thinking can be turned off. A route asking Fable 5.1 or Grok 4.7 for no thinking is raised to low on record, since their APIs refuse or ignore it. A model OMH has no contract for (GLM 5.3, GLM 5.3 Flash and Grok Build 0.1 among them) keeps the shared steps low to xhigh, marked ? because its ladder is not known, so nobody on a custom model has to write a file to change an effort. An optional ~/.omh/routing/model-efforts.json replaces a model's list, and an empty list turns its keys off. Pressing Enter on a value the head can't take (+ to xhigh, then a switch to Gemini 4 Argon) keeps the picker open on that row, with every other edit intact, until you fix it or cancel. Stored values nobody touched are left as they are and marked. (#2640, #2658)

    • Every Hermes skill body carries the reply-voice and turn-ending rules as one line. The paragraph each body repeated becomes one line that keeps the rule's core: reply in the user's own words and the host's own voice, the SOUL.md persona owns reply language, tone, speech level and sentence endings in progress updates too, the user's language where the persona sets none, and OMH shapes structure and content only. The same line keeps the turn-ending rule, keyed on the request: when the requested output is delivered, open points go inside the result and the reply ends with no question or offer; only when the person asked to choose or decide and that choice is still theirs does it end with one specific question about it, not a list, a generic offer, or a footer. The record-term clauses live once in the shared rail's Reply Language And Host Voice and Turn Ending sections, which already held them. Portable agent-skills bodies keep the full paragraph, since their hosts have no shared rail. (#2657, #2680)

    • Vim NORMAL gets Space leader shortcuts in the OMH TUI. With display.vim_mode on, Space then s, b, p, r or /, m, t, c or Enter opens sessions or bots, folds the plan, searches history, cycles the input mode, steps reasoning effort, interrupts or clears, and submits or steers. The info line says NORMAL · SPC while the leader waits. Esc or an unmapped key cancels it; Ctrl+C, Ctrl+O, Ctrl+L and Ctrl+Z cancel it and do what they always do. Without vim mode nothing changes. Contributed by @DanRioDev in #2545.

    • Vim NORMAL in the OMH TUI finds characters on the line and takes iw and aw. With display.vim_mode on, f F t T and a character move to its next or previous copy on the line, or just before it; ; repeats the last find and , repeats it the other way, and a count works with each. d, c and y take a find (dt), cf,) and the word objects iw and aw (diw, caw, y2aw), on one line and with words between blanks, as w reads them. A find that misses changes nothing, and a c whose find misses stays in NORMAL. While a find waits for its character, Esc, Ctrl+C or any other key cancels it, and on a terminal that reports Shift with the key (kitty, modifyOtherKeys) fX and f( find as they do elsewhere. Contributed by @DanRioDev in #2576.

    • A dag-ulw verify node names which producer's scope holds each failure, and a repair node is told the failed attempt's edits may still be on disk. A producer accepted earlier cannot be repaired, so a failing check is repaired through the verify node, and nothing said which producer the failing path belonged to. The verify node's task now carries a compact map of its upstream producers' declared write_scope (their repairs and its own included, cut at 600 characters on whole paths and producers) and asks it to name in its handoff, for each failing check, the producer whose scope holds the failing path, or unattributed: attribution by declared scope, not proven cause. A repair node that writes files, in its first cycle and in the re-plan cycle, now reads that the failed attempt's edits may still be on disk and, in a git checkout, to inspect git status --short and git diff HEAD on the failed nodes' scopes (its own scope when the failed node was a verify node) before editing. There is still no rollback: the worktree is shared. Every other node's task reads byte for byte as before.

    • The coding-core skills no longer contradict each other or dead-end. A fix pass over the skills most coding runs load: ulw-interview's question budget, dimensions and option count now agree with its taxonomy; ulw-work and ulw-plan no longer send an unplanned request back and forth, and ulw-work names dag-ulw for an ordered plan; omh-plan and ulw-plan say which one an ordinary or a risky change belongs to; ulw-loop handles a north-star goal and a direct invocation that matches a Do Not Use line; ulw-perf runs its edits in the Hermes harness unless an external owner was chosen, and asks for repeated runs, spread and the metric direction before a regression budget; ulw-qa caps fix rounds and requires a cleanup receipt; ulw-research takes defaults when it cannot ask, names the briefing renderer and the budget a declared depth gives; ulw-context names the terminology capture command; omh-ask runs an advisor read-only and out of the live repository. Smaller edits in omh-decide, omh-adversarial-consensus, dag-ulw, ulw-maestro and the TDD reference.

    • A dag-ulw run that waits shows whole and moves, and says what it waits on. A run of up to 12 lanes (the transcript's own small-run measure) shows every layer and every node that runs; a layer with work moving shows at most three of its quiet nodes as rows, and more fold into its ▸ N more row from the start, however tall the terminal is (opening that row still lists them). A done layer of one node is one dim row (✓ layer 2 · <title> model · accepted · ↓tokens), consecutive ones stack, and one blank row stands between layer groups. An open layer row no longer repeats the counts its node rows say (▾ layer 1 · 2 nodes, ▾ layer 4). A bigger run folds consecutive done layers into one ✓ layers 1–3 · 11 done row, which opens to its layers, and opens only the layers with a live node and the next layer still to come, which fitting never shuts. The tree is at most 16 rows and still at most a third of the screen; a 6-node run is 14 rows at 150x42. Rows that do not fit become one +N more · /dag row. A parked node was drawn ? and counted unknown, because the graph reader had no entry for the parked outcome. It is now a waiting node with the status parked and the time since it parked, and the header and its layer row count it as parked. One row under it names what its park recorded it waits on, and nothing else: the handle, named by its call (terminal(uv run …)) when it is one of the session's background processes, then until HH:MM from its wake_at, then its reason, and the process's own running time when the handle is one. The parked and running times count on every second. The spinner turns on a running or parked node's row. A waiting row, and a shut layer of waiting nodes, carries a slow, faint dot cycle (· ·· ···, one step a second). Done rows, the done layers' row and a run that is not alive stay still, and a done node's title is muted. This motion runs between turns too while the run waits. While the terminal window is unfocused, the spinner and the dots hold, as the wallpaper does (#2542), while the times keep counting, redrawn once a second. An unspecified-* or unknown category no longer draws a cell. Between turns, while something the session waits on still runs, the working row's place above the composer says so, as it says a turn runs: ✻ Waiting… (33m 47s · 1 agent · terminal). That is a delegated child or dag-ulw node that wakes the parent as it ends (the running lanes of the delegation graph the HUD reads), or a background process Hermes will wake the session for (notify_on_complete, or watch_patterns on a match). Each count shows only when it is not zero, and the time is the longest wait. It has the working row's own spinner and clock. A turn's working row takes over when a turn starts, and the row goes when nothing is left to wait on. It is read from structure, never from what the model said, and the info line is unchanged. (#2561, #2595)

    • dag-ulw refuses two nodes that write one file under different spellings. omh_dag define, repair and extend refused a shared write scope with no direct edge only when both nodes wrote the same entry character for character, so src/ beside src/x.py, ./src/x.py beside src/x.py and src/*.py beside src/x.py ran in parallel and could overwrite each other. Entries are now compared as paths: \ reads as /, letter case, ./, repeated or trailing / and x/.. change nothing, and an entry covers every path under it by whole segments (src/a and src/ab stay apart). A * or ? glob beside a file name with a suffix must match it, with **/ matching no directory too (*.md meets docs/a.md, not src/x.py; src/**/*.py meets src/x.py). Beside a directory, a name with no suffix or another glob, it covers the directory before it. A [ stays a literal, so app/[slug]/page.tsx names one file. The refusal still reads write_scope_overlap, now names every overlapping pair at once (up to 20, then a count) and which node wrote which entry, and a repair, like an extend, is checked only for pairs with one of its own new nodes, so two earlier nodes it cannot change never refuse it. The coding fanout contract's boundary check still compares as written.

    • A model can no longer write OMH's evidence stores, and a dag-ulw child is held to its node's write scope. A write_file or patch aimed at a store OMH keeps under <omh_home>/runtime/ (decision gates and events, delegation, dag, research receipts, runs, todos) is refused, naming the store, since a line written there would read as one OMH observed; the dag-ulw handoff report a model is told to write still passes. A dag-ulw node's child, and any child it spawns, is refused a write outside the node's declared write_scope, with the node, run and scope named and the file to report in its handoff so the parent can repair or extend. Parents, ordinary children, nodes with no scope, settled nodes, cancelled runs and writes outside the workspace are unchanged, and a path either guard cannot resolve passes with one warning. (#2585)

    • The area under the composer is exactly as tall as what it draws. On the screen's bottom rows, when the agent view gives rows up (a dag-ulw layer settles or folds, or the run ends and its tree goes), the composer moves down by them and the output above is painted again that many rows lower, in the same synchronized update. No row stays blank under the tree or above the composer. This replaces the tree's high-water hold during a turn and the blank rows a finished run left under the info line. The lines painted again on the screen's first rows are also still in the terminal's scrollback, one per row given up, so they show twice there once they scroll off again.

    • A finished subagent shows a green ✓ in the OMH TUI's agent list instead of the hollow ○. Four finished agents in a row looked the same as four that had not started. A completed agent now leads with ✓ in the done green; the collapsed N done row is green only when every agent it counts completed, else a muted ▪. Interrupted, failed, waiting and running agents keep their marks.

    • A skill that fails partway now says what it already proved. It lowers the affected result to tentative, unresolved, or not_observed, and hands back what the person needs to continue instead of presenting partial work as finished. Procedure steps name what survives each failure, Recovery Notes cover each skill's own blockers, and catalog checks keep both in place. The idea comes from PortSwigger Burp AT research skills.

    • A bot profile that fails its refresh no longer rolls omh update back, and a pending restore no longer blocks every later update. In one measured case four bench profiles whose plugins/omh links pointed at deleted worktrees made the re-entered update exit 1. The staged update rolled back, the restore failed on the same profiles, and from then on every omh update only retried that restore and staged nothing, while the report read "the known-good generation remains active". Now an update whose only failures are bot profiles exits 3: the staged update keeps the activated generation and names the profiles to repair. A run that finishes a pending restore goes on to stage the update it was asked for. A failed restore leads its report with "could not restore the previous generation: recovery is required" and says no update installs until it succeeds. A rollback names the generation it restored, with its commit. omh doctor reports an open marker as staged_update_recovery. A plugins/omh link whose target is gone is now replaced with the released bundle without --force, and the update prints a Repaired … line naming where it pointed; a link to an existing tree OMH did not write still needs --force.

    • omh update names the commit it moved between, and omh --version names the commit it runs. A generation built from one of this repository's GitHub archives (the installer's main.zip, and tag or commit archives) now knows its commit: src/omh/_archive_commit.txt is marked export-subst, so git archive writes the commit into it, and the ref comes from pip's PEP 610 record. omh --version prints omh 4.0.7 (main@1a2b3c4d). An installer-managed update ends with oh-my-hermes 4.0.7 main@1a2b3c4d → main@5e6f7a8b, or already at main@5e6f7a8b when main has not moved, even when the release version did not change; release wheels still name versions only. omh update --json carries the same facts as identity_change. EN, KO, JA and ZH rows. No new network call: the commit is read from the archive pip already downloads.

    • After omh update, the open OMH TUI and the running gateway each offer a restart, and nothing restarts on its own. An OMH TUI started by omh from the installer-managed environment reads the current generation pointer at most every 5 s, with no network call. Once an update has finished moving it, it shows restart to update on the info line and asks once, between turns: OMH moved to 4.0.7 main@1a2b3c4d. Restart the TUI now to use it? This session reopens.. Not now comes first. Restart the TUI waits for the running turn, for a turn about to start, for what the session waits on or still has out, and for what you typed (a /bg or /btw answer still out does not hold it: the question says that answer will not show), then exits for omh to open the new generation's omh --resume <id> on the same session, which sends nothing: a notice names any work left and says to send continue. A pointer that moves back closes the open question and says so. If the shared gateway still runs the old OMH plugin code, that TUI then offers the existing /gateway restart once, with Cancel first. At the end of an activated update on a terminal, omh update names what each stale running gateway's restart interrupts, then asks Restart the Hermes gateway now so the new OMH plugin code loads?, defaulting to no; yes runs Hermes' own hermes -p <profile> gateway restart --yes --detach. --restart-gateway and --no-restart-gateway decide without asking. With --json, --no-interactive, --yes, off a terminal, and in a re-entered update, nothing is asked. A restart that was asked for and failed exits 4. Both start with the update after the one that installs this: the question runs in the omh update you started and in the TUI's own generation, and both are the previous version during that first update. (#2616, #2619)

    • --channel main names the channel that tracks main. It is another spelling of --channel preview, which a plain installer-managed omh update uses on an install from main. The re-entry into a restored, older generation spells it preview. npm, Bun and Homebrew installs now refuse --channel preview instead of updating to a release and recording it as a main build.

    • A turn Hermes interrupted itself goes on by itself in the OMH TUI. After the machine sleeps past the session's turn lease, Hermes stops the running turn, and the TUI said "Hermes interrupted this turn; send continue to go on." It now says it sends that continue in 3 s, with Esc to cancel, and then sends it as a typed continue goes. It never does so for the person's own Esc, /stop or interrupting prompt, for a turn that ended with an error, or for a turn this TUI did not send (another client's on a shared gateway, whose stop looks the same from here). Anything the person sends in the wait (a prompt, /stop, /help, /new), a turn that starts, or another session opened cancels it; one Esc does too, in vim INSERT as well. A draft, a queued line or an open question holds it, and a notice says to send continue. At most 2 go between two prompts of the person's own; past that the old notice asks.

    • /gateway restart restarts Hermes' gateway from inside the OMH TUI and picks the session back up. On the bots' gateway: after omh update the running gateway kept the old OMH code until someone ran hermes gateway restart in another terminal and came back. Now the TUI first names what a restart affects (the profiles the gateway serves, its connected messenger bots, its running work, this session's own), runs Hermes' own restart only once you pick Restart, and reattaches to the same session with your draft kept and nothing sent again. Each choice says what it stops. A gateway a service runs lets running work finish first, up to Hermes' restart_after_turn_timeout, and the TUI waits for it; since Hermes' own systemd or launchd service drains a gateway without --external-supervisor in its argv too, which the record cannot show, that slot says the work finishes first if a service runs the gateway and is otherwise interrupted. In a named profile the shared gateway serves you pick that profile alone, whose running turns stop now and which does not load new OMH code, or the whole gateway, which does. Once the gateway is back it says so, and that it now runs the newer OMH when its files show it was behind. /image and /paste wait in the composer during the restart, as the other commands that call the gateway do. A refused or failed restart says why and the command to run instead. Bare /gateway is still Hermes' platform status. (#2588, #2591)

    • A prompt queued behind a ! line in the OMH TUI waits for a turn that started while the command ran. The next queued line after a ! command used to go once Hermes answered the command, even into a turn that had started meanwhile; it now goes when that turn ends, or after the turn Hermes runs next when that turn left it one. Esc on a prompt still staging its images holds the lines queued behind it, with the interrupt's notice, instead of leaving them unsent. A rewind still goes ahead of lines already queued: queued behind them, its cut would drop the turns they start.

    • A resumed session with work left goes on by itself in the OMH TUI. omh --resume <id>, /resume and the session picker opened the session and waited for the person to write, so an interrupted turn or an open dag-ulw run sat still. Now the TUI says "Resuming: {why}; sending continue in 3 s · Esc cancels" and sends continue as a typed one goes, when Hermes interrupted the session's last turn itself (a sleep past its turn lease), or when the session holds a dag-ulw run that a turn would move with no node running: nodes ready, back to settle, parked past their wake time, or out with no child after a restart. The continue's turn then brings the run's own turn-start line. Hermes keeps no cause for an interrupt, so each OMH TUI writes how a session's last turn ended to <omh_home>/tui/turn-ends.json (the newest 128 sessions), the person's Esc or /stop as they ask for it. It never goes on for the person's stop, a session with a turn running (its own or another client's), a turn Hermes resubmits itself after a crash, a run that needs repair, a draft, a queued line or an open question.

    • A push no longer ends the reply with "the session has stopped, send a message". After a git push, gh pr create, gh pr merge or gh workflow run with no watcher armed, OMH appended that sentence to the person's reply in the reply's language. It landed on very nearly every successful push, a README edit pushed straight to main included, where nothing was ever going to be waited on, and it was appended even right after the model said it would stop writing it. The model still gets the in-loop note on the push result (arm a background watcher, or say the session has stopped); the reply the person reads is left as the model wrote it. The end-of-turn sentence remains only on a turn a failed watcher's notice opened.

    • A Kilo Code or Vercel AI Gateway key no longer turns chain shaping off. KILOCODE_API_KEY or AI_GATEWAY_API_KEY in Hermes' .env, or a Hermes login or credential pool row for kilocode or ai-gateway, counted as a relay serving every model, so chains fell back to the shipped order and category routes the session's own provider served were refused (an Anthropic session routed 7 of 18 categories instead of 15, a Kimi session none instead of 11). They now work the way a commandcode key does: a route may name the relay, and a routing/model-providers.json route to a relay this machine holds counts its alias as served, so a chain head sent to the relay still leads and routes there. Nothing else about routing moves. A key with a recorded gateway in routing/providers.json no longer hides the relay, so a route naming it applies too. The same routes document now works for a held commandcode relay, whose routed aliases were demoted and reported as routed to an unknown provider. A providers.kilocode or providers.ai-gateway block, or model.provider: kilocode / ai-gateway, in config.yaml still counts as a gateway the person declared. A key or login no longer lists either relay in the omh setup provider step; the "a provider not listed here" row still adds one. A routing/providers.json that already recorded the relay ("kilocode": "gateway", from an earlier omh setup pre-tick) keeps turning shaping off until that row is cleared in omh setup or with omh model-chains provider clear.

    • A reply written in Traditional Chinese now ends with a Traditional Chinese stop sentence. On a turn a failed watcher's notice opened and nothing re-armed, OMH appends one sentence in the reply's own language. Every Chinese reply used to get the Simplified copy, so a reader in Taiwan or Hong Kong saw a script they had not written in. A Chinese reply now gets 「沒有任何工作會再次喚醒這個對話。如需繼續,請傳送一則訊息。」 when its Traditional-only characters outnumber its Simplified-only ones, read from Python's own Big5 and GB2312 codecs; Simplified replies are unchanged. A kanji-only Japanese reply, already read as Chinese, now usually gets the Traditional copy. (Reported and first fixed by @sykuang in #2654.)

    • The OMH pane now draws an active delegation as cards and shows what the focused chat has spent. A delegation or fanout run of up to 16 lanes renders as layered node cards with edges, the critical path, bottleneck chips and per-model stats, and a click on a card shows that node's details; a larger run keeps the drawn graph. A USAGE section at the top of the pane shows the chat's own tokens, cache hit, cost, context share, rolling tokens per second and first-token wait, read from its recorded session row and its auxiliary calls across compaction segments, and — for any figure nothing reported. Contributed by @yqm-307 (#2665).

    • Vim NORMAL in the OMH TUI takes quoted text objects. With display.vim_mode on, d, c and y take i or a with a double quote, single quote or backtick (di", ca', yi and a backtick), on one line. Inner takes what is between the quotes, and with a count of 2 or more the quotes too; around takes the quotes and the blanks after them, or before them when none follow. A quote after a backslash does not close the pair, and a missing pair changes nothing. The second Ctrl+C of an exit now leaves over any half-typed command, and Esc over a pending find holds a waiting continue as it does over any other command. Contributed by @DanRioDev (#2668).

    • A finished dag-ulw run says what it used. Once a run is done, stopped or cancelled, omh_dag adds usage (its children's uncached input, cache read, cache write and output tokens, and the cost Hermes recorded) and a usage_line such as DAG feat · 2 children · 13.2k tokens · cache 80.0% · $0.0500 (2/2 priced). Every child the run bound counts, an earlier retry's too, each with its compaction segments and auxiliary calls. The cost is only what Hermes priced: a child it did not price makes the figure a floor (≥), an estimate keeps its ~, and no token is converted into a price or compared with a model the run did not use. omh_run_summary prints the same figures as a DAG usage line before the goal check, and the Desktop pane keeps the run's Tokens / Cache hit / Cost / Node time figures under the summary line once the cards collapse, where before they vanished at the moment the run ended. Asked for in #2684.

    • A dag node can be kept on a local model, and parallel lanes prefer a fast speed tier. Taken from Hermes' own model routes (the advanced-delegation plugin documents a local route and a per-route service_tier). Under a model_map delegation backend, a node marked local: true runs only on an option whose base_url is on this machine or its private network (or provider lmstudio/llamacpp); a provider error moves it to the next local option, and with none left ready holds it in held_for_local_model and says why. It never goes to a cloud model, and under any other backend it is held. In swarm, quick and simple-work, an option with a fast service_tier is tried before the same model's normal one, and the normal one leads elsewhere; a tier the model does not offer is refused by Hermes and falls over to the next option. OMH only reads the options and writes none. omh_delegate_route no longer tells the parent Hermes has no child fallback: a child takes delegation.fallback_providers when the user sets them (Hermes 2026.9.7 and later).

    • omher.app has a skill guide: every OMH skill outside the ULW engine, in plain words. docs/skills/ lists the 145 skills by work lane, each as a card that opens to what the skill is for, when it fits, an example request, what it is not for, and what has to be true before it calls the work done, with search, lane filters and deep links (#omh-finance-analysis). The page is rendered from skills/*/SKILL.md by tools/render_site_skills.py, and the Pages workflow re-renders it on every deploy, so it follows skill edits without a manual regen. Checklist lines shared by three or more skills (the common rail) are left out so each card shows only its own checks.

  4. 4.0.6

    Expandir

    GitHub Release

    • In a console session a choice goes through Hermes' clarify picker, and the OMH TUI draws that picker the way Claude Code draws its question dialog. On the first turn of an attended cli, tui or desktop session, the model is asked to use clarify where its reply would stop for the person to pick between named options or approve an action it will not take without their answer, instead of listing the options in prose and asking for a word back. A closing offer of optional further work stays prose. After a clarify timed out, the person's next message brings the same ask once more. Messaging platforms, single-query runs, kanban task processes and delegated children get nothing and render the same bytes as before. In the OMH TUI a question with choices now stands between two rules: the context dim, the question itself bold, a label — description choice as its label with the description under it, Type something. for an answer of your own, and the keys on a row of their own with how long the question has waited. The tool row says how many choices it offers instead of repeating the question. A pick still answers with Hermes' own choice string.

    • The Hermes Desktop pane reads the Windows Hermes home in its fallback (#2569 by @k1lgor). When the dashboard backend could not import Hermes' get_hermes_home() and no HERMES_HOME was set, it read ~/.hermes, while Hermes itself defaults to %LOCALAPPDATA%\hermes on Windows, so the pane's HUD could read an empty home. That fallback now uses Hermes' platform default, with HERMES_DATA_DIR_SUFFIX honoured; a parity test keeps it equal to the bundle's own platform_default_hermes_home. The host's answer and an explicit HERMES_HOME still win, and POSIX stays ~/.hermes.

    • A drag selects text in the OMH TUI again after another program left the mouse on. Hermes' own TUI turns on mouse modes 1000, 1002, 1003 and 1006, and a crash or kill -9 leaves them on in the tab. The OMH TUI never asked for them, so it dropped every report, and the terminal kept the drag for itself: no selection, nothing to copy, for the whole session. Entering now switches every mouse mode off (9, 1000, 1001, 1002, 1003, 1005, 1006, 1015, 1016, the list Hermes resets), and every exit, a stop with Ctrl+Z included, switches them all off again, where it used to switch off only 1000, 1002 and 1006 and let a leftover 1003 outlive the session.

    • Research citations are checked against what was actually fetched. Every successful web_extract, browser_navigate and web_search call now leaves a metadata-only fetch receipt under the session that made it, subagent sessions included: a hash of the normalized URL, its scheme, host and path, and the content's hash and length. Page text, titles, query strings and arguments are never stored. A new agent-facing command, omh research receipts verify --citations <ledger> --draft <draft>, checks every [n] citation in a draft against the receipts of the session, its earlier compaction segments and its subagent sessions. It exits 0 only when every cited source was fetched; a never-fetched citation exits 1 and is named. ulw-research, web-research and research-brief tell the agent to run it after sources.py verify. A receipt proves the page was opened, never that it supports the claim. Receipts are written only inside Hermes, so a Codex or Claude Code session outside Hermes reads missing.

    • dag-ulw is offered where a plan's follow-on is chosen. On Hermes' own /plan turn, whose reply offers to carry the plan out, the model is asked to also name dag-ulw as one option when the plan's tasks form a graph (some can run apart, with disjoint write scopes) ending in a verification task, with the fit reason in one line. On the turn right after a plan was made, a short go ("구현하기", "looks good, implement the plan", "플랜 승인. 진행해") gets a conditional note: if the message accepts the plan without choosing how to run it, the model may name the same option; if it declined, held, edited, asked or changed topic, the note is ignored. Neither line starts anything. "Just made" is read from structure, once per plan: the turn before was /plan's, invoked or loaded a planning skill, or wrote a plan file under .hermes/plans/ or .omh/plans/, or this is the first turn to see a plan, ralplan or refactor-plan state active. Reading or mentioning a plan file does not reopen it, so carrying a plan out never brings the note back. Before, none of the six acceptance messages measured saw dag-ulw: four got no OMH context at all after /plan. A turn with no plan just made renders the same bytes as before, and routing is unchanged.

    • After a plan is accepted, the planning skills offer dag-ulw beside the usual follow-ons, and taking the offer works. An accepted plan or ralplan whose tasks form a graph (ids, dependencies, write scopes and a closing verification task) now names dag-ulw with a one-line fit reason, next to carrying the plan out directly or an ultrawork cycle. refactor-plan names it on a whole-plan go whose phases split into groups with disjoint files: the phase order becomes the graph's order and each phase's check its node's check, while the per-phase commits and rollback points stay with the user. It starts only on the user's go-ahead. Before, the workflow state refused to start dag-ulw while plan, ralplan or refactor-plan was active, so the offer could not be taken even when it was made. dag-ulw now builds its graph from the plan the user accepted instead of planning a new one. ralplan's recovery note, the /plan reference, the router's lane map and engine list, the intent-to-plan card, and the safe-feature and picker hints name it too. Routing is unchanged.

    • The OMH TUI's working glow is removed. The light that went round the screen's edge while a turn ran read as too big and uncomfortable, and its trail framed the screen so the terminal's padding outside it looked like a black margin. It is gone, with everything that existed only for it: the /glow command, omh theme glow, OMH_TUI_GLOW, the faster 50 ms paint while it moved, and the column on each side and the blank last row the screen kept free for it, so rows start at the screen's first column again. A glow choice an older release saved in <omh_home>/tui/theme.json is ignored and dropped the next time a choice is written. The living wallpaper, the working row and its spinner are unchanged.

    • A look with a scene paints its ground on every terminal. The eight looks with a scene set their ground only with OSC 11. Otty gives that colour only to cells written after it, so every cell an erase cleared kept the terminal's own black. That covered the chat area, every row's tail, the composer and the blank rows: about 2,500–3,200 of the 6,300 cells on a 150x42 screen. Every erase and blank fill on those looks now carries the ground as its background. A terminal fills those cells with the ground through background colour erase. A move in the /theme picker paints the new ground over the whole visible screen in one synchronized update: the region, the rows above it that are still on screen, and the blank rows under it. Nothing is written into the scrollback. Pagers (/logs, /dag, Ctrl+S) erase on the ground too. The default look, 256 and 16 colours, NO_COLOR and /wallpaper off write the same bytes as before. At exit and at Ctrl+Z, the rows still on screen are painted again on the terminal's own background. One cost, accepted by the owner: rows already in the scrollback keep the look's colour until they are printed again. The session rows left above the prompt after a Ctrl+Z resume are on the terminal's own background again, so Otty shows its own colour in their tails until they are reprinted, as before this change. The background restore at exit is unchanged.

    • On a dark look with a scene, the OMH TUI's diff and outcome bands and the pager's footer bar follow the look. These bands were the skin's fixed dark green and maroon on every look, and the red one sank into aurora's teal at 1.01:1. Each band now keeps its green or red hue and moves onto the look's ground. It sits 1.27:1 or more against every ground, and its text stays at 4.5:1 or better. The footer bar of the pager, /dag and the Ctrl+S bot board takes the sent prompt band's colour instead of the skin's teal-black, with its words re-lifted to 4.5:1. The /theme picker's tile for the default look now shows the terminal's own background, which is what picking it gives. The default look and the light side are unchanged.

    • The Hermes Desktop pane takes the OMH TUI's mood, and draws the DAG as the TUI does. The pane stays a Desktop pane: the app's sans, its components and its theme tokens. It takes on the TUI's default launch look as Desktop colour mixes: a rose brand, warm section labels, and muted violet rules and frames. Sections have more room between them, and agent rows get a hover. While a delegation graph runs, the pane shows the box drawing the TUI prints into its transcript, in a card. The backend makes it with the same renderer, at the width the pane measures in cells. A run of up to 12 lanes is drawn as boxes. A bigger run, or a pane too narrow for boxes, gets the tree. Each box takes its state's colour: the accent for running, yellow for stuck, red for failed, grey for done or waiting. Every word the pane colours now reads at 4.5:1 or better on Desktop's light and dark themes; the lowest before was about 2.2:1.

    • The OMH TUI draws Markdown tables as aligned columns while a reply streams. A | header row used to reach scrollback as plain text one delta before its |---| divider arrived, so a table could never be drawn properly. The trailing run of | lines is now held out of scrollback until the reply moves past it, and shown in the live region as the same table the finished reply draws: bold header, a ─ band, wide characters aligned by cell width, \| kept as a literal pipe. A table wider than the row, and a | line with no divider under it, stay plain text, quitting mid-reply still prints the rows held back, and a table taller than half the terminal is released to scrollback as it grows, on the columns it has. Carried forward from #2550 by @kfatyuip.

    • The Windows loader probe keeps LOCALAPPDATA. The isolated probe that omh doctor runs dropped it, so the Hermes launcher could not find the home directory and doctor reported hermes_runtime_command_failed although hermes --print-runtime-command worked in a normal shell. The variable is forwarded only when the parent has it. Carries #2548 by @DefinitelyNotSpammy.

    • A changed Python function's broken callers are named at the turn's first stop. When a turn changes a Python function's parameters, OMH reads the files that import it, following package re-exports and aliases. It names, by file and line, each call that matched the old parameters and no longer does, with the new signature. It needs no language server and is on by default. It says nothing about a call it cannot prove broken, such as one passing *args / **kwargs or one written to raise. It runs only after an edit made with a file tool: Hermes does not run the turn-stop check after a shell-only edit. In the live bench this took gpt-6-luna on a signature change from 0/6 to 6/6, and Haiku 4.5 from 5/6 to 6/6.

    • A dated snapshot in memory says how to retire it, in the session that saves it and in later ones. When the model saves a fact like "As of 2026-10-30 (v0.1.7), nanran has ..." or "nanran v0.1.13 added ...", the result tells it to name this memory in replaces when it saves the new state. The old memory then leaves recall instead of being served beside the new one. The owner's store held 8 such snapshots side by side, none ever replaced. Because the state usually changes weeks later in another session, recall also marks snapshot facts and decisions: snapshot="as-of" on the record in the memory pack, "snapshot": "as-of" on a recall tool row, and one shared line per pack. The line says the record still holds until that same thing changes, and only then is it named in replaces; something newly added stays its own record. Episodes are not marked, since an event never changes. This reaches snapshots already stored, without a migration. The memory itself is unchanged: it stays permanent and nothing is aged or hidden. Versions stated as rules ("requires Hermes v0.22.5 or newer", "pin Node v20.11"), a "Decided :" line with no version change in it, lessons and procedures get neither the note nor the mark. The mark never costs a record: when it does not fit the pack's bound, the pack renders unmarked, exactly as before. A marked pack is 287 characters longer for the line, plus 17 per marked record. The cue is English only.

    • Vim mode changes the cursor shape. With display.vim_mode on, the terminal's cursor is a steady beam in INSERT and a steady block in NORMAL (DECSCUSR). This is an OMH addition: Hermes keeps a steady block in both modes. It is written only when the mode changes, and the terminal's own shape comes back when a question or picker takes the keys, when vim mode goes off, and on exit or Ctrl+Z. On exit it is the terminal's default shape, not whatever the shell had set before. Carries the cursor part of #2510 by @DanRioDev.

    • Memory recall costs a turn far less. The memory pack is rendered in on_turn_start, before the turn's model call, so its cost is turn latency (a code comment and docs/MEMORY.md said it was rendered off the hot path; both now say what happens). Every render ran the credential and safety scan on every stored record again, twice on a console turn, and that scan was most of the render. Each record's verdict is now kept, keyed by the digest of the record's exact content, so an unchanged record is not scanned again and a changed one is scanned on its next render and refused there if it became unsafe. The credential check on the short metadata strings a render redacts is kept per string the same way. A turn also reads each review file once instead of three times and lists the records directory once less. On a read-only copy of the owner's store (39 records), on_turn_start plus prefetch went from 77-124 ms to 13-17 ms and a repeat render from 42-52 ms to 10-14 ms; on a synthetic 500-record store, from 3.0-3.3 s to 0.13-0.52 s per turn. The first render in a process still scans every record (measured on a machine under heavy load, two interleaved rounds).

    • The OMH TUI's living wallpaper holds still while its window is in the background. The TUI now asks the terminal for focus reports, as Hermes' own TUI does. While the window does not have focus, the living wallpaper's scene stops where it was, and nothing wakes the TUI to paint it. The spinner, streamed text and everything else in a running turn keep updating. When focus comes back, the next frame is the one the animation would have drawn at that moment, and with focus the look is byte for byte what it was. On an idle screen with a scene, an unfocused TUI went from 4 wake-ups and about 6 KiB written a second to 1 wake-up and nothing written. A terminal that does not send focus reports is treated as focused throughout, so nothing changes there. The focus reports type nothing in any mode and are switched off on every exit and while the TUI is suspended.

    • A full redraw keeps 4,000 rows of the session, up from 3,000. A resize, a tab switch, Ctrl+L, /redraw and a resumed session erase the terminal's scrollback and print the session's last rows again, because a tab switch can reach the TUI as one same-size signal after the terminal has already reflowed its rows. After that, a 1,100-turn session's scrollback went from 27,504 rows to 3,006; it now keeps 4,006. The count was chosen by measurement on a session of distinct replies. A redraw at the width the session was printed at takes 13 to 18 ms instead of 11 to 14 ms. The first redraw at a new width takes about 235 to 290 ms instead of 170 to 240 ms, and holding what that redraw reuses takes 1.2 to 1.5 MiB more. Raising the count further made both slower in proportion: 6,000 rows took 23 to 29 ms and 375 to 450 ms. The rows above the cut are still in /history. A terminal's own scrollback limit can keep fewer rows than this.

    • The recall switches and records say what they do. Under recall_mode: on_demand or off, a save that resembles a held memory no longer tells the model that either one "recalls": no turn is served memories unasked there, and the hint now says so. Tool recall's recall_mode, demote, recall origins and the omh_memory consolidation status read the profile the memory provider was started with, rather than the process's HERMES_HOME, so the two cannot disagree about whether recall is off, and a brief raised in one profile is not reported from another. In the TUI, an omh_remember recall that found memories lists them and gets the same sweep as an omh_memory one. A test now pins that a consolidation brief raised in the same second by another trigger is served. A pinned memory still suppresses the list of the rest's names on a message that matches nothing: listing them would make the live recall select differently from coding handoffs, which the recall suite and incident replay hold equal.

    • Skill hints read the request, not the gateway's busy-path header. A Discord or Slack message sent while the agent is busy reaches the turn with Hermes' origin block in front of it: one line of JSON ids between two fixed lines. OMH read that block as the request. The router then dispatched gateway-intent-card at high confidence whatever the person wrote, and the per-turn "Skills that may fit this request" line named gateway-intent-card, backend and llm-app-dev on every such turn. Because those words are ASCII, the block also kept a Korean request off the Korean line. OMH now cuts the block, in its exact shape and only at the start of the turn, the same way it cuts the other gateway material. The fix was replayed on a frozen, hand-labelled set of the owner's turns since 10-15. On the held-out 40%, turns given a wrong line went from 49 to 26, and the share of offered skills that were right went from 3.3% to 5.3%. Recall did not change.

  5. 4.0.5

    Expandir

    GitHub Release

    • Matching a coding handoff to a specialist domain is about 2.7 times faster on English text. Preparing a handoff folds the task message and the plan text before it matches specialist domains, and ran the full Unicode normalisation on every pass even for plain ASCII, which that work cannot change. ASCII text now takes a direct lowercase fold, and text that is still ASCII after the first normalisation step returns early. Other text keeps the original steps in the original order, so routing decisions, exclusions and errors are identical. On the domain-routing benchmark 772.81 ms fell to 290.94 ms (median of three runs). Contributed by @dexhunter (#2536).

    • Memory recall shows what the turn asked for, once per conversation. A message with nothing to search for (an empty turn, "야") used to get every saved memory in full; it now gets pinned ones in full and the rest as the one-line list of names a non-matching message already gets. A background-process or delegation notice that nobody typed gets nothing at all. A memory the conversation already holds is not sent again until it changes or a compaction drops it, and a list of names it was already shown is not repeated. The recall line counts only what is new, so a turn that adds nothing posts no line. omh_remember and omh_memory recall return nothing_found for a query that matches nothing instead of unrelated memories, and look up a memory by the id or subject a list names. Coding handoffs and omh memory recall are unchanged. Replaying stored turns on read-only copies of the owner's stores, the largest per-session memory text went from 111,665 to 6,649 characters on the console profile, repeats from 91% to 0%, and messenger recall lines from 447 to 61.

    • Turns that only read, and a session's first turns, are faster. The round-end consistency check indexed every Python file in the workspace at a turn's first tool call, which cost about 110-175 ms in a checkout of a few thousand files even when the turn only read files or searched. The index is now taken at the turn's first call that can change a file (an edit, a command, a delegated child, any tool that is not a read), still before that call runs, so a read-only turn pays nothing and the check decides as it did for the turn's own edits; a file changed by something else (an editor, a background process) before the turn's first non-read call is now part of the baseline. The capability families and the router, which a session's first user turns imported on the spot (about 60 ms on the first turn and 70-100 ms on the next new kind of request), are now imported on a background thread when the plugin loads, so after a normal pause before the first message the first turn takes about 25-40 ms and the second about 2 ms. Plugin load time is unchanged. A first message sent the moment the plugin loads waits for the remaining imports.

    • One broken part of the OMH TUI's screen no longer ends the session. If the info line, the live transcript, a picker, the plan or HUD panel, the agent view or the sticky prompt fails to draw, that part becomes one row in the warning colour that names it, and the rest of the screen and your keys keep working. Its traceback is kept once, and /logs shows it after the gateway's lines. While the TUI runs, a background thread's traceback, a Python warning or a stray print now goes to the same in-memory log instead of landing on the screen and shifting every later frame. A crash that does end the TUI still prints its traceback once your terminal is restored.

    • OMH notices when Hermes cannot load its skills. Skills could be installed while Hermes could not load them by name: a stale generation directory registered beside the pointer made every name ambiguous, and then a cached path to a deleted generation left none. The launch card still said "skills 89 installed", and the skill suggestion kept telling the model to load skills that came back "not found". The suggestion and the route hint now name only skills Hermes resolves to one copy, read from the skill directories Hermes itself cached, so they go quiet instead of steering into a failed load. When a profile that registers OMH's skills sees fewer than half of them, the launch card says how many Hermes sees and suggests a gateway restart, or /new for an older session. omh doctor has a new hermes_skill_visibility row, per profile, that reports an ambiguous registration and a missing directory separately. A healthy install shows the same suggestion and the same card as before.

    • The OMH recall line stays off Discord and Slack, and recall can be turned off. 🧠 OMH — recalled N memories now shows on the CLI, TUI and Desktop only; on a messenger the memory is still used, with no line posted. The new profile setting recall_line (console by default, all, off) chooses where it shows, and recall_mode (auto by default, on_demand, off) chooses whether memory is recalled every turn, only when the model asks, or not at all. Set them with omh memory recall-line and omh memory recall-mode; omh memory status shows both. A memory_mode of off, or a disabled retain_knowledge family, now actually stops recall: before, the provider and the recall tool kept serving records under it. It also stops the consolidation brief and the save reminder, which asked for saves that off refuses.

    • The OMH TUI's first frame no longer waits on the skill catalog. When the shared gateway names every skill in its first answer, the launch card used to import the whole skill catalog to label them before the first frame was drawn; it now takes the labels from a small module of their own. The card is drawn the same, byte for byte. In a clean process, starting up against a shared gateway went from about 55 ms to 7 ms and imports 24 modules instead of 114. On a stdio gateway, the first regrouping of the skills when they arrive went from about 50 ms to under 1 ms.

    • The memory consolidation brief no longer rides almost every console turn. A session that ended with one to four unconsolidated turns used to raise a brief at its end and again as a recovery at the next start, so 213 of 216 console turns on one machine carried the capture request (153 briefs came from session-start recovery), and every capture grew the store that recall draws from. The session-ending reasons now need the same five turns as the turn interval, and a given brief is shown once per session instead of on every later turn. Headroom, duplicate and expiring-record reasons keep their own thresholds, dreaming_mode is unchanged, and the model still captures through its own tool with no approve step.

    • The omh_recommend tool result is about 12% smaller. The handoff policy used to ride on every recommended skill, and one policy can run past 1 KB. It now stays on the top recommendation and on the skills the route selects, which are the skills the agent acts on. When the top row is a router (meta-router, oh-my-hermes), the first concrete skill below it keeps its policy too. Every other row leaves it out, and the result says that omh_capabilities inspect with section skills returns it for any skill. The ranking, the top skill, the route and every other field are unchanged. Over the routing corpus the mean result drops from 6,794 to 5,985 bytes.

    • A multi-line paste no longer sends half a prompt where the terminal strips paste markers, and arrows hold together over SSH and WSL. When tmux, SSH or the terminal itself drops bracketed-paste markers, the OMH TUI now reads a burst of pasted lines as one paste: the lines land in the composer (as a chip when the paste is long) instead of the first line break submitting the first line. Over SSH the burst may arrive in chunks up to 100 ms apart. An Enter pressed on its own still sends at once, and so does an Enter followed by the next key typed. Two or more prompts typed ahead into one read while the TUI was stalled now arrive as a draft in the composer instead of being sent, since those bytes are the same as a paste. Over SSH and in WSL a lone Esc now waits 100 ms instead of 40 ms for the rest of its sequence, so an arrow key split by a slow link is the arrow and not Esc and two letters, which could clear the draft or open rewind. Locally nothing changes.

    • Typing in a long message no longer slows down. The input box measured the whole text again on every key, so a key cost about 5 ms at 5,000 characters and about 17 ms at 20,000, longer than a frame. It now measures only the words and lines an edit touched, which brings the 20,000-character key to about 5 ms and the 5,000-character one to under 2 ms. This holds for text with spaces between words (English, Korean and so on); a long line of unspaced or Cyrillic text is still measured whole. What is drawn, the caret, wrapping, wide characters, paste chips and image tokens are unchanged.

    • Running work survives a context compaction. When Hermes compacts a session, OMH now hands its summarizer the work this conversation still has open, stated as facts: each unfinished dag-ulw run (its goal, run state, nodes done, and the open nodes with their states) and the session's open omh_todo plan items. The summary that replaces the old turns then still knows about the run and the plan, so the model picks them up instead of re-deriving them from scratch. It is bounded (at most 3 runs, 6 nodes a run, 10 plan items, 1,500 characters in all), adds nothing to a session with no open work, and leaves a plan written by Hermes' own todo_list to Hermes, which already carries that list across compaction.

    • A redraw at an unchanged width no longer stalls the TUI. Switching tabs, Ctrl+L, /redraw and every other full redraw at the width the session was printed at reprinted the session by wrapping every finished reply's markdown and painting every row again: 119 ms for a 1,100-turn session, 109 ms for a 100-turn one. A finished reply in the redraw's tail now keeps its rows for the width and theme it was drawn at, and a row printed above the live region keeps its bytes, so the second and later redraws take 6 to 9 ms and the first after a reply streamed about half what it did. What is written is unchanged, byte for byte. A redraw still erases the terminal's scrollback and prints the last 3,000 rows, because a tab switch can reach the TUI as one same-size signal after the terminal has already reflowed its rows. The row over the rest, and the /history footer, now count "entries" (⋯ 2940 earlier entries): the number counts prompts, replies, tool groups, notices and done lines one each, which is not what "messages" said.

    • A TUI performance bench lives in the repo. tools/benchmarks/tui_perf.py measures frame cost by transcript size, full redraw, streaming per chunk, keystroke cost by composer length, idle wake-ups and bytes, plugin hook latency and launch to first frame, offline with temp homes, and --compare <tree> judges two checkouts back to back. It changes nothing you see in the TUI.

    • Ctrl+N/P and Ctrl+J/K move through lists and menus. In the /model, /sessions, /rewind and /theme pickers, the agent list (roster and DAG tree), history search (Ctrl+R) and the slash and @ path completion menus, Ctrl+N and Ctrl+J move down one row and Ctrl+P and Ctrl+K move up, as Down and Up do; in history search Down and Up now step through matches too. While a completion menu is open those chords move the menu instead: Ctrl+P does not fold the plan, and Ctrl+J and Ctrl+K do not add a newline or delete to the end of the line (Alt+Enter or Shift+Enter still add a line, Esc closes the menu). With no list or menu open, Ctrl+J keeps its newline and steer behaviour, Ctrl+K deletes to the end of the line and Ctrl+P folds the plan. A Shift-reported Ctrl+N/P/J/K does not move a list; terminals without key reporting cannot tell Shift apart. Contributed by @DanRioDev (#2513; #2524, carried by #2525).

    • A running turn with the glow and a living wallpaper uses less CPU. The same frames are drawn, byte for byte, but the colours the glow and the wallpaper scenes mix are no longer parsed again for every cell, and the loop no longer wakes a second time per frame to wait out the frame interval when nothing arrives from the gateway. On a 60x200 screen with the aurora look, a busy frame went from about 9.9 ms to 5.6 ms and the loop from 38 to 19 wake-ups a second.

    • "Improve the architecture" has an owner. tech-debt-audit gains an architecture-deepening reading: it looks for shallow modules (an interface nearly as large as what it hides), proves each one with the deletion test rather than its size, classifies its dependencies to say how the deepened module will be tested, and ranks the candidates as ledger rows with one top pick, proposing no interface. refactor-plan then shapes the chosen module's interface: at least three designs under contrasting constraints, compared on depth, locality and seam placement, one recommendation, and the approval gate unchanged. Its tests phase now deletes tests that reach past the new interface once an interface-level test covers them, and ai-slop-cleaner keeps a single-caller module that hides real complexity instead of inlining it. Requests that name the code ("improve the codebase architecture", "improve the architecture of this repo", "find deepening opportunities in this codebase", "shallow modules in this codebase", "코드베이스 아키텍처 개선해줘") now name the audit in the clarifying question, where 4.0.4 named codebase onboarding or unrelated lanes; they dispatch it only beside its own words, such as "tech debt". An architecture that is not code (a model, cloud infra, an org chart, a building) routes exactly as before.

    • LSP navigation works on Windows. lsp_definition, lsp_references and lsp_rename answered no_result on Windows while the language server was active: OMH named the file file://C:%5CUsers%5C... while Hermes had opened it as file:///C:/Users/..., so the server answered for a document it never opened. OMH now builds the same URI Hermes does, and reads back the server's forms of a drive (file:///c%3A/... included) as the local path, so the workspace-scope checks see real paths (#2470).

    • The OMH TUI and omh doctor say what Hermes' gateway needs, and why. Three cases used to leave a slow TUI, an empty /resume, refused messages or missing skills with no reason given. Each line is said as Hermes', only while its gateway_state.json is fresh:

      • A session store failure is said when the record says unavailable and Hermes' error log holds a store failure from after the last start. A deleted profile still referenced (a cron job, a kanban board or a served profile) is named, with no restart advice, since a restart does not clear it. A replaced store gets hermes gateway restart, with what that restart does to every bot it serves.
      • A gateway started before the last OMH update still runs the old OMH plugin code, and the line says so.
      • A draining gateway starts no new turns, and Hermes holds or refuses new messages as runtime_draining. A restart is named with its turns, its seconds left and agent.restart_after_turn_timeout; a drain request with the fact that the gateway keeps running; anything else as a stop.

      The TUI shows each once at launch; doctor shows them as the hermes_gateway warning row. omh update --json now carries the update restart hints as gateway_restart_hints (null on a dry run, which computes none). Local files only, read in bounded reads; OMH stops and restarts nothing.

    • Code skills tell the executor how to write a test, not only in TDD mode. The questions before a new test and the shapes of tests that guard nothing loaded only when the user asked for TDD, so a backend or frontend handoff that said "add tests" let an executor mock its own repository class and assert that save was called. They now live in one shared page, omh-verification-gate's references/test-authoring.md, which adds tests coupled to internals (own-module fakes, call-count assertions, side-channel reads), fakes only at system boundaries chosen by dependency category, and a regression_seam: none_correct outcome when no seam can run the pattern that produced a bug, where the verifying reproduction loop stands in place of a regression test. omh-backend, omh-frontend and omh-app-debugging load it when a handoff asks for tests, and ulw-work's TDD page points at it instead of keeping a second copy. omh-backend also gains references/contract-tests.md: one negative test per auth row and per error-table row, the same database engine in tests, writes read back through the interface, and migrations run on populated data.

    • backend and relational-db now cover concurrent writes, cursors, and inbound deliveries. A backend contract names the guard on every write two requests can reach at once: a unique constraint (mapped to a 409 row), an idempotency key stored with its request fingerprint and response, a lost-update guard, and side effects sent only after commit, through an outbox or after the commit returns. Cursor pagination ends its ordering key in a unique tie-breaker, binds the cursor to its filter, and caps the page size; each endpoint states its expected rate, latency target, and payload bound or unknown; dependency timeouts shrink down the chain; and a cache names its key, TTL, and invalidating write. A new references/inbound-and-async.md covers webhooks (signature over the raw body, replay window, event-id record), queue consumers (at-least-once, bounded attempts, dead-letter, poison messages), and connection-pool sizing; an unsigned inbound webhook is a blocker. relational-db gains a concurrency section (default isolation per engine, FOR UPDATE, SKIP LOCKED, upsert, lock order, whole-transaction retry by error code), the keyset-pagination index row, and the rule that CREATE INDEX CONCURRENTLY gets its own non-transactional migration. The backend references no longer end with an attribution block about Rust unsafe and native debugging that did not belong to them.

    • omh-llm-app-dev and omh-rust cover the code around the model and the command line. The LLM provider boundary now has a deterministic fake adapter beside the real one, so parsing, repair, retry classification, tool handling and budget termination get offline unit tests while evals stay a separate suite against the pinned model. A response cut off at the output cap is classified as truncation and never re-asked under the same cap; tool-call arguments are validated before the tool runs, and invalid calls and tool failures return to the model as tool results; a streamed call is cancelled upstream when its client goes away and is validated only once complete. The Rust gate list now lints every target with all features, tests both feature extremes, checks the declared MSRV, adds --locked in CI-shaped runs and no longer claims cargo doc runs doc tests; new sections cover binary and CLI contracts (stdout for data, named exit codes, no panic on user input, binary integration tests) and adding a dependency (license, MSRV, default features, cargo tree -d).

    • Review, commit, rule and onboarding skills check what they actually rely on. omh-code-review now searches the reviewed repository for its standards files (REVIEW.md, CONTRIBUTING.md, AGENTS.md/CLAUDE.md, *STANDARDS*.md, linter and formatter configuration) before the smell baseline applies, hands that list to a dispatched reviewer, and skips smells the configured tooling already enforces on the diff. omh-commit-pr-authoring has every PR body say whether a revert undoes the change or what it leaves behind, and which surfaces a bad merge would reach. Without a PR template that goes in the fallback body; with one, it goes under the template's risk or rollout section, or, when the template has neither, in one sentence at the end of its summary or description section. A structural change's PR body shows the smallest file tree, call tree or diff sketch of its new shape. omh-rules-distill classes each recurring miss as mechanical (it becomes a check, proven by failing on an injected violation) or judgement (prose). omh-codebase-onboarding checks its own repo map, reading path, glossary, risk map and first-task runway instead of a generic planning checklist.

    • A green regression test no longer proves a fix on its own, and a reproduction has to be the reported bug. omh-verification-gate now asks, for a "fixed" claim whose evidence is a new or changed test, for an observed red of that test, for the bug's reason, on the tree before the fix (a worktree run, a CI job, or the debugging session's own pre-fix run, recorded with its source); an intermittent fault or a test-only fix is witnessed by a failure rate beside 0/N at the same N. A fix test only ever seen green holds the claim with regression_witness_missing. A fix recorded with regression_seam: none_correct has no fix test, so the witness rule does not apply. omh-app-debugging's reproduction record gains symptom_match (the observed failure beside the reported one; a nearby failure is another bug), loop_cost (time per run and what was pinned, which is only what neither the report nor a hypothesis already written implicates), and minimised (the reproduction cut down until every element is load-bearing). Its method reference gains a ladder for faults no test reaches, ending in numbered steps for a person with named fields to report, and a session tag on every temporary probe with an observed search for it before handover. Commands read credentials from environment variables and quoted output is redacted. The confirmed hypothesis and its eliminated rivals go into the fix's commit or PR body. omh-native-debugging's fault statement takes the symptom match, the minimise step and the probe tag too.

    • plan cuts code tasks as vertical slices, and decision-prototype says what a good prototype looks like. A plan's code tasks each deliver one behavior through every layer it needs, with a done check that drives it through the surface a user or caller reaches, so an integration failure shows in the task that caused it instead of first in the closing gate. Slices that write the same file are ordered rather than re-cut by layer, each later slice listing every earlier one that writes it as a direct dependency; a file every slice only appends to (a changelog, a generated index) may instead be written by one task after them. That task keeps a generated, drift-gated file's gate out of each slice's done check. A layer-only task is a prefactor placed first, or a wide mechanical change handed to refactor-plan (the new form beside the old, callers moved in groups while both stay green, the old form deleted last). decision-prototype gains references/artifact-shapes.md: a logic or state question gets a pure module a later plan can lift, with the full state shown after each action; an interaction or layout question gets structurally different variants mounted in the real host page inside a temporary worktree, with writes stubbed; and prototype code the receipt lets later work reference is kept on a named throwaway ref the receipt cites.

    • The round-end check names a test whose expectation a keep-behaviour turn rewrote, and diffs a Python file a command names. A new always-on test check names an existing test that lost an expected value when the request asked for the behaviour, the output or the result to stay the same as a whole, and did not name that test, ask to update a test, or state a new result. A test a command appends to (printf >> tests/...) is now diffed instead of firing as an unseen edit. The copy check is unchanged: a look-alike that says it differs on purpose still fires it, because four review rounds showed a note can hide a real copy. The bench gains the C-TEST-UPDATE control.

    • The frontend skill now covers data, forms, and input, not only the look. A page that reads or writes remote data loads the state rules that until now only a refactor loaded (server state in a query cache, each mutation naming what it invalidates, an optimistic write with its rollback), and the route state matrix gains submitting, submit-failed, optimistic-pending, stale, and conflict rows. A new forms-and-input reference, loaded for a form, a text composer, or a dialog, covers labels, when errors appear, mapping server field errors onto fields, one submit in flight, dialog focus, and an Enter-to-send that waits for IME composition, so a Korean, Japanese, or Chinese user no longer sends half-composed text. The brief names a browser-support target before newer platform features are used, the taste reference gains interface-copy rules, and the component inventory records each primitive's interface. frontend-refactor now requires a before/after capture when a change alters DOM nesting or class composition, and inlines only abstractions whose deletion makes their complexity vanish.

    • A running gateway keeps working through omh update. Three causes are fixed. GC could delete the generation an older gateway still served: the first activation it recorded had no time, and GC read that as earlier than every gateway start, so it held the wrong generation. When the history cannot say which generation a gateway started on, GC now keeps every generation created before that start and prints a note: saying so. An open OMH TUI holds its own generation the same way, through an OS lock under generation-holds/. Every update also rewrote the plugin bundle and restamped installed_at even when no bundle file changed, so the gateway restart hint fired after every update; a moved package directory now updates only the manifest's package_site, and the hint fires only when the bundle's files change. Last, a gateway's omh finder never looked at the manifest again. It now keeps its directory while that exists. Once an update removes it, the finder moves to the recorded directory when that holds the version already loaded (package_site status moved), so later omh.* imports keep working; code that reads files through an old module's __file__ still points at the removed directory until a restart. The recorded directory holding another version is reported as stale_until_restart, and nothing is loaded from it.

    • The OMH TUI rides out a slow or busy Hermes gateway. Its connection to the shared gateway sent websockets' default keepalive ping, so a gateway whose event loop was busy for about 40 s (a long /compress, a heavy turn) had a healthy connection closed under it. The TUI then said the gateway exited, marked the running turn lost, and gave the resume 30 s before leaving. Now:

      • The connection sends no keepalive ping, as Hermes' own loopback connections do not. During a turn, once nothing has come from the gateway for 45 s and no tool is running (Hermes sends nothing while a terminal tool runs), the working row says "no word from Hermes' gateway for N s"; the TUI closes nothing.
      • A connection that does drop says "the connection to Hermes' gateway dropped; the turn may still be running". The resume reads whether the turn still runs (or Hermes has started it): one that does goes on with its clock and any stop already asked, and one that ended meanwhile says so. A gateway that started again (a new authority epoch) says the turn was lost.
      • The launch's and a respawn's session calls wait 120 s, as Hermes' own TUI and the TUI's in-session calls do, not 30 s.
      • omh --resume <id> that Hermes refuses (not_found, permission_denied, ...) says why and exits 1; a session Hermes keeps for another app is to be opened where it was started. It used to open Hermes' own TUI, which starts a new session in its place without a word.
      • session.replay_gap, sent when the gateway stops sending this screen a session's events because it fell behind, resumes the session, as Hermes' own TUI does. Before, the turn spun with no end. The turn's own events that arrive before the answer win over it. A resume refused or unanswered closes the turn and holds the queue; one that names a newer id for the session moves the screen to it and closes the turn shown.
      • skills.changed asks for the command catalog again, once per burst, so a skill installed from another terminal reaches completion and /help.
      • A turn Hermes interrupted itself says "Hermes interrupted this turn; send continue to go on" instead of "Interrupted.". A stop the person asked for (Esc, /stop, a prompt in busy mode interrupt) still says "Interrupted.".
    • /resume always answers, and omh --resume opens the picker. A bare /resume (or /sessions, Ctrl+X) drew nothing until Hermes answered the list, which during a turn can take a while. Now Loading sessions… stands where the picker opens, keys typed meanwhile stay in the composer (the picker opens with an empty filter and the draft is there when it closes), and no stored session, a refusal in Hermes' words, or no answer within the call's wait each say so. A list that arrives after you typed into the composer shows its rows but leaves your typing in the composer, pauses included, and the next prompt too after Enter sends one; ↑, ↓ or Esc hands it the keys, and its filter row says so. omh --resume with no id (short form -r) starts the TUI on that picker; omh --resume <id> and omh --continue are unchanged.

    • A prompt sent while a session switch waits goes to the session it opens. While /new, /resume <id> or a pick waits on Hermes, switching to <title>… (or starting a new session…) shows under the queue, and a prompt sent then is held and sent in the new session once it is live; before, it went into the session being left, which was then closed. That holds for plan-mode prompts, slash and skill commands and ! lines too. Prompts still queued behind the old session's turn come back to the composer with a notice instead of vanishing, and so do prompts held for a switch that failed. Of two switches out at once only the newer lands: a session the older one opened late is closed again, with a notice on the new screen, but never the session the newest switch goes to. Asking again for a switch already out sends nothing, and a switch answered after its wait ran out has the session it opened closed again. A second /resume while the list is out sends no second call, and an answer that comes while the picker is open no longer resets its highlight or filter.

    • Reports on OMH's routing reach workflow learning, and tasks that only name OMH do not. "OMH가 왜 내 메시지를 code-review로 보냈어?" dispatched the code-review it names, because the fold reads the past tense 보냈 as the request cue 보내; a Korean execution cue now matches whole syllables. Three English report shapes reach workflow learning: an adverb between OMH and its verb ("OMH apparently chose ..."), a why-question about what OMH did ("Show me why OMH picked them"), and OMH ignoring the request. The other way round, "OMH picked the config. Find out which keys are unused" no longer gets the ~1.1k-char workflow-learning hint: "find out" with a question of its own is the user's task when what OMH picked is a concrete artifact (by its head noun, never a catalog skill name), while a bare "find out why", a question about the choice itself, or a choice made for the user's own message still reads as a report. Nor does a code edit led by "OMH: migrate" or "using OMH, rename" on a vocabulary or terminology module; a report that quotes such an ask ("Migrate the cache was all I asked, but OMH chose ...") stays a report. 9 of the 12 router misses #2455 listed are fixed on the route hint. The other 3 are Korean reports that name OMH with a particle ("OMH가 라우팅을 잘못했어, 원인을 같이 보자"); reading that form as OMH was tried for seven review rounds and dropped, because each closed list it needed let coding requests through. The particle mention itself gives no routing context, as before; the 보내 fix still changes Korean reports that use 보냈. "OMH ignored ..." is a report only when OMH ignored the user's own ask or a routing target, not a file or a setting. Measured offline on the earlier held-out sets and eleven fresh sets frozen for this change (1,083 scored messages, hint and chat route), no message the previous router answered right is answered wrong.

    • Hermes picks up the new OMH skills after an update without a restart. A running Hermes resolves skills.external_dirs once and keeps the path until config.yaml changes, so after omh update moved the current pointer, chats kept the previous generation's OMH skills, or lost them all once that generation was collected, until hermes gateway restart. The update now moves the modification time of every Hermes config.yaml (the primary home and each bot profile) that names current/skills, without rewriting it, and says Hermes skills refreshed in <N> homes; a rollback and a --recover-known-good do the same after their last pointer switch. New conversations and every skill load use the new pack; a conversation that is already open keeps the skill list it started with until /new. A home it could not touch is a warning and a failed row under hermes_skill_refresh, and does not fail the update. The gateway hint now says what a restart is still for: skills reload on their own, the new OMH plugin code loads only after hermes gateway restart (and, for a home whose touch failed, the skills too). The bot-profile line no longer asks for a Hermes Desktop restart to reload skills.

    • A same-size rewrite in one clock tick no longer hides from the round-end check. The workspace index kept each Python file's modification time and size. A sed, execute_code or delegated child that rewrote a file at the same size within one tick of the filesystem's clock (about 16 ms on Windows) left both unchanged, so the edit went unseen and the check was skipped. A file modified within two seconds of the index is now also indexed by a digest of its bytes and compared by content. This was also why two test_consistency_trigger tests failed now and then on Windows CI.

    • An answer Hermes sends late no longer acts on what the person moved on to. A Hermes slash command's answer (a skill, /retry's prompt, /undo's draft, an alias) that arrives after a session switch is dropped with a notice instead of running in the new session, and its command.dispatch fallback is not sent there. Output it prints is still shown, a /branch it made is not switched to, a side task stays the left session's, and images typed with it come back to the composer. A prefill or a rewind pick that lands on a draft typed meanwhile puts that draft in the session's history first and says Up brings it back. The /model and /rewind lists do not open during a turn that started while they were on their way (the busy notice is said instead), nor over another open picker. Asking for the same list twice in a session sends one call, a list still out for a session left never holds up the new one, /model --refresh while a bare /model's list is out says one is on its way, and a second answer never resets the open picker. While a list is on its way the slot says waiting for Hermes…. A late /model <name> answer for a session already left says nothing in the new one. Keys always go to the picker the screen draws. An approval or question arms only after a quiet gap with no key, so a person still typing when it opens answers nothing. What they type, and their edits to it (Backspace, Delete, Left, Right, Home, End, Ctrl+A/E/U/K/W), go on into the field they were typing in, the tell field too, and a path-shaped paste is still a drop. A key that would send, interrupt, complete or walk history does nothing then, and nothing is typed while a mouse report may still arrive. A Shift+Tab approvals switch Hermes never answered no longer leaves Shift+Tab dead: accept is shown as possibly on, and the next Shift+Tab sends an explicit off.

    • omh says what it waits for while the Hermes gateway restarts. The launch waited for the gateway and for the session before drawing anything, so a busy or restarting gateway left the shell silent for up to a minute, and a gateway mid-restart sent the launch straight to Hermes' own terminal. Now, after a second and a half of quiet, one line on the shell says connecting to the Hermes gateway… (Ns), the Hermes gateway is restarting; waiting… (Ns) or waiting for the Hermes gateway to open a session… (Ns). A gateway that is starting or draining, whose control socket is missing, refuses or closes without answering, or whose WebSocket connection is refused or times out is tried again for up to about a minute; so, at most twice, is one that closes the connection as going away (a restart) while the session opens. A gateway that crashes while it opens the session, refuses this client, or denies its socket falls back at once, as before. Once the minute is spent the launch says how long it waited, that the gateway did not answer, and to check hermes gateway status, then falls back. The line is erased however the launch ends, Ctrl+C included. A normal launch prints nothing new. The attach child exits 69 for a gateway that is not answering yet; every other failure still exits 1.

    • Show-reel judges the hero frame before the full render. A product film that passes every lint and contact-sheet check can still read as cheap on sight, so show-reel now asks for a launch-frame verdict on the first frame: a 3D product is a lit model rather than stacked CSS layers, light surfaces are judged by measured luminance, and the title uses the brand's own live typeface and colour. Brand looks are gathered from the brand's site and repo with provenance, a film pairing two products names both from frame 0, and a model's critique of a downscaled sheet is checked on a full-resolution crop before anything changes. design-quality-gate's rubric gains a matching material-credibility axis.

    • The info line's badge carries the version. It reads [⚚ OMH Native v4.0.5], the version faint inside the brackets. On a narrow line the version is the first part of the left side to go, before the branch name is cut and before the badge shortens.

    • The OMH TUI says when a newer OMH is out. At most once a day, on a thread the screen never waits for, the TUI asks https://registry.npmjs.org/oh-my-hermes/latest for the latest published version and keeps only that and the time of the ask in <omh_home>/tui/update-check.json. While it is newer than the one running, the info line says so beside the badge: [⚚ OMH Native v4.0.5] v4.0.6 available · omh update. On a narrower line it is the first thing to shorten, to ↑ v4.0.6, which stays beside the version until the version itself goes. Offline, a timeout or a bad reply says nothing. OMH_TUI_UPDATE_CHECK=0 holds it off in one terminal, omh theme update-check off in every later session, and so does an omh update-check you set to off (in omh setup or with omh update-check set --mode off).

    • What's new, once, after an update. The first fresh OMH TUI session on a new version shows What's new in v<version> and a few one-line items under the launch card, and once they are drawn records the version in <omh_home>/tui/whats-new.json so they are not shown again. An install that ran the OMH TUI before this release is told too; a new install shows nothing. /whats-new shows them whenever you ask.

    • The composer's rules are violet in the default look. The lines above and below the composer, and the launch card's frames, were spectrum's grey edge (#4A4F66); they are now a muted violet (#6E5A9E), lifted for contrast on a light terminal. The other looks keep their own edges.

    • A dag-ulw run with no node running is named at the next turn. A parent re-enters a run only when a child completes, so a run with no child running sat open: a parked node past its wake_at, ready nodes never handed out, a node out with no child (a restart ended it, or none took it), a node a restart ended past the re-offer bound, a node that came back and was never settled, or a failed node with a repair cycle left; so did a node past the stuck mark, which was named only on an omh_dag call. The next turn's context now carries one [OMH dag] line naming the run and the call to make (omh_dag action=ready, settle, repair or status), once per situation; when every delegate_task slot of the process is taken, it says to call ready at a later turn. Not named: a node out within dag_stale_after_s (its child may still be starting), a parked node whose wake_at is ahead, and a node exported to a coding owner or left to the parent, which its fanout result or the parent's own work settles. It blocks nothing, OMH still starts no timer, and delegated children, finished runs and conversations without a run get nothing.

    • A plain knowledge question no longer pulls in an OMH skill. "Explain TCP vs UDP briefly", "What's the difference between git merge and rebase?" or "클로저가 뭐야?" used to get a route hint or a "Skills that may fit" line from one shared word (explain -> paper-learning, merge -> ulw-work, explain + vs -> finance-analysis), and the model then spent an extra call loading a skill it did not need. A one-sentence concept question with no work object (no my/our, no file, PR, CI or agent word, no URL or code) now gets neither a token-only hint nor a candidate line; phrase matches such as "explain this paper" keep theirs. On a 55-question English and Korean probe, suggestions went from 13 to 1, and none of the 1,098 dispatch cases in the routing corpus lost a hint (#2472).

    • A fanout check that could never pass is refused before dispatch, and finished work that nobody verified is unverified, not failed (#2460, #2461, reported by @vince2930). The dispatcher runs every verification_commands entry without a shell, so test -f X && grep -q Y X handed && to test as an argument: it exited 2 on correct work and the unit landed verification_failed, yet fanout validate had accepted it. Freeze, validate and prepare now refuse an unquoted &&, ||, |, ;, &, >, <, $( or backtick in a verification command, reproduction_command or task_linked_test_runner, and name both fixes: one command per step, or sh -c "...". Quoted text such as grep -E 'a|b' is unaffected. A unit that exited 0 with a valid result and no observed check now ends unit_state: unverified (reason unchanged, verification_not_observed) instead of failed, so failure counts and retry automation stop treating finished work as broken; failed is kept for a check that ran and failed. omh coding run gains a repeatable --verification-command, which implies --run-verification, so the shorthand can reach verified.

    • Hermes can save a stated preference or decision to OMH memory in the same turn, and a rule the user restates updates its record instead of adding a second one. The memory provider now exposes a small omh_remember tool with save and recall actions outside Hermes' deferred plugin-tool catalog, so a chat with many loaded tools no longer has to search for a tool before it can save. save runs the existing capture path (admission, the session lens, surface refusals, bound channel scope, duplicate skip) and recall the existing recall path. The tool description and its one-line prompt ask for the save before other tools, even when the message also asks for other work, limit it to the user's own words, and exclude pasted text and facts that only matter for the current task. save takes replaces, the id of a record already shown to the conversation: the merged summary goes through the correction path, the old record moves to history, and only the merged record recalls; an id outside the session's lens still saves the fact as a new record. It also takes tags, a few topic words in the user's own language, so a later message in that language can match a record whose summary is written in another. On Discord, a thread Hermes auto-created from a channel message uses its parent channel's memory scope, so a fact saved in one conversation thread can be recalled in the next thread started in the same channel; other threads keep their own. omh_memory is unchanged.

    • omher.app pages carry X's full share-card tag set. Every indexable site page now carries twitter:site, twitter:creator, twitter:image:alt and og:image:type beside the existing share card, and test_site_seo pins them so a home-page redesign cannot drop the share block unnoticed.

    • A code-change plan from plan carries its execution order. Each task now has an id, the tasks it depends on, the files it may write and its done check: the dependency and write-scope fields a dag-ulw node takes, and the check its VERIFY and STOP WHEN come from. For a code task the done check is a command with the result that counts as passing (the test's run command, a gate, a grep for the signature), not a description such as "an exact signature and its file", which a converter to dag-ulw had to turn into a command by guesswork; a task that is not code may keep an observable check. Of two tasks that write the same file, one lists the other directly as a dependency, as dag-ulw requires. Converting a plan to a graph still adds each node's label, category and full prompt. The last task is a closing gate, named as the verification task: real verification commands run by someone other than the builder, then a review, and an unclean review adds a fix task (at most two rounds) instead of ending the plan. After acceptance, plan and ulw-plan now offer dag-ulw for a plan whose tasks form a graph.

    • A dag-ulw run that finished without a passing check no longer reads as a verified goal. Each node now declares a role (build by default, baseline, verify, review) instead of OMH guessing it from the label, which misread checks such as "independent verification vs baseline" and producers such as "visual-qa motion reference". A finished run carries a goal verdict apart from its state: verified only when every producer sits under a verify node that was accepted with its handoff checked, otherwise unverified (with the reason) or verify_failed. omh_dag, the run summary, the messenger end block and the TUI end view all show it, and an unverified done run is told to add a verify node through extend. A node with no role and a write_scope whose label reads like a check past its first word ("Re-verify the fix", "Final verification pass", "Re-review the diff", "Double-check the migration") still runs as a build node and still counts as a producer, but it stays on delegate_task instead of going to the external coding owner, where the builder would approve itself; the lint names the role it reads like, and role=build sends it to the owner. Build work that reads the same way ("Add input validation to the API") stays native too until the parent names its role. Such a node that reads like a verify node and fails runs again after its repair, the way a verify node does. An advisory lint, no_review_node, names a run whose build nodes write more than docs with no review node; it never refuses a run or moves its state or goal verdict. Runs stored before this update keep reading their nodes as before.

    • omh_jev_ask can reach a TypeSafe endpoint you configure. It honors TYPESAFE_BASE_URL, the API-root variable TypeSafe's own SDKs read: an HTTPS root gets /v1/systemone appended (/systemone when it already ends in /v1). A value with plain HTTP, no host, credentials, a query or a fragment is refused before any connection opens, redirects stay refused, and the OpenRouter route and its opt-in are unchanged. omh doctor says a custom root is configured without printing it, and a custom endpoint that reports no cost is shown as unpriced instead of at TypeSafe's list price. Contributed by @yqm-307 (#2473; requested in #2468).

    • Short promo clips get their own design rules. A launch clip of 15 seconds or less is one segment, so show-reel no longer applies its multi-segment habits (scene variety, transitions, a giant number scene, an end card) to it. The new Short promo clips section in craft-design.md asks for one hero scene held throughout, no meta text (labels, IDs, URLs and prices go in the post body), one number smaller than the logo, and the official logo at 25% or more of the frame height, checked on a still sheet before the render. design-quality-gate scores the same items as a Short-clip discipline axis. Generalized from six review rounds on one launch clip. Contributed by @bellman-sionic (#2488, carried by #2490).

    • Codex handoffs on native Windows can write (#2471). Four changes:

      • The child environment now passes SYSTEMDRIVE with the other Windows base variables. Without it, Codex's MXC sandbox refused every shell call (MXC requires nonempty Unicode host SystemDrive), and the only workaround was --owner-env codex:SYSTEMDRIVE on every run.
      • A new coding/fanout-settings.json stores environment grants and a Codex Windows sandbox backend. Dispatches an agent starts, which never see a command-line flag, now get them too. "codex": {"windows_sandbox": "mxc"} passes --config windows.sandbox=mxc to codex exec.
      • A unit the sandbox stopped is reported as executor_sandbox_read_only or executor_sandbox_start_failed, with the fix in the journal summary, instead of malformed_result with exit 0.
      • On Windows, omh coding executor-readiness --executor codex reads the effective backend and says the session will be read-only when none is set, before a dispatch spends a run finding out. Its PATH report no longer lists npm's extensionless shell shim, which logged [WinError 193] before falling back to codex.CMD.
    • The Desktop pane says when a delegated lane failed, keeps a narrow pane readable, and darkens the text you read. A failed lane or route used to be counted as blocked; the header and the status bar now say 1 running · 1 failed with failed in red, apart from blocked (stuck or interrupted lanes, blocked rows) in yellow, and a red dot when nothing runs. At the default 320 px pane a delegation row puts its model on a second line under the title, so neither is cut to a few letters. The scope line, plan notes, the subagent summary and backend notices use a text tone between Desktop's secondary and tertiary roles, readable in the light theme where tertiary alone was faint. In Korean the count reads 1개 실패.

    • show-reel has a launch-film lane for 60-120 s feature releases. A new opt-in reference, launch-film.md, gives the act sheet (mark, title becomes the product, one honest loop, escalating range, turn, reach, climax and close) with time shares, and the devices that carry it: the feature name stacking and bending back into the captured window, typing anchored to the measured line so it never re-centres, two-tone escalation copy, word cards around compressed agent runs, close-ups only on decisions, a window on a blurred ground instead of a device mock, and one declared breath before the heaviest montage. Short promos keep the one-scene rule; every capture and citation rule still applies. Method notes come from a frame-by-frame study of a public 99 s launch film; nothing was copied.

    • Cancelling a dag run names the children still running. omh_dag action=cancel now lists each node whose delegate child is still out, with its subagent id and how long it has been open, under still_running, and the reply names each one with its delegate_task action=stop route. A repeated cancel names the same children instead of saying none are left. Contributed by @ryanrodrigues25200525-svg (#2512, carried by #2515).

    • The LSP opt-in has a visible enable path. When the impact check finds new errors and the omh-lsp toolset is off, its note says lsp_diagnostics re-checks directly once OMH_LSP_TOOLS=1 or the lsp_tools setting is on and Hermes restarted; with the toolset on it no longer repeats that. omh doctor advisories gain lsp_tools_disabled (advice when unset or off, unobserved when the config shape cannot be read). Contributed by @ryanrodrigues25200525-svg (#2511, carried by #2515).

    • Korean requests that mention OMH by name stop drawing unrelated skill suggestions. The per-turn "skills that may fit" line ranked Korean on syllable pairs, and "헤르메스"/"오마이헤르메스" sit in three skills' triggers, so "오마이헤르메스 홍보 영상 만들어줘" was offered parallel-tools and executor-runtime-readiness. The host name, the particles "-에서"/"-으로" and the request verb "확인" now count as filler. Replayed over 722 Korean owner messages, lines naming those three skills fall from 24 to 11; each skill still matches its own nouns ("병렬 도구", "코덱스", "기억 점검").

    • Codex fanout units get OMH's own skills, and can edit, commit and reach the network inside the fence. OMH looked for Codex skills only in ~/.codex/skills, while Codex also loads ~/.agents/skills, where .codex/install.sh puts OMH's pack. On a Mac with six personal skills there, every Codex unit was told to start with $generate2dsprite and $game-video-feedback-loop (11 of 11 units in a real run, about 32k characters of unrelated skill text each). Discovery now reads the repo .agents/skills, $CODEX_HOME and ~/.agents/skills too, so the same unit gets $omh-plan, $omh-build-failure-triage, $omh-code-review; a name found in two roots is never suggested, since Codex refuses a plain $name that matches two skills. Inside the macOS fence, zsh heredocs and uv now work: TMPPREFIX and UV_CACHE_DIR point into the unit's scratch dir (before, every git commit -F - <<EOF failed with "can't create temp file for here document" and uv run could not open its cache). The fence's write boundary is unchanged. On macOS the fence also allows the two trustd mach lookups (com.apple.trustd.agent, com.apple.trustd), so Codex's TLS certificate check can ask the trust daemon; before, a unit connected, could not verify the certificate, and ended with workspace routing discovery failed or retried until its timeout. They are lookup rules only: no file or preference write is allowed (#2456).

    • A stale Codex skill pack is now reported and refreshed. omh update refreshes an installed ~/.agents/skills pack in its recorded layout (it never creates one), omh doctor warns with omh update when that pack is stale or edited, and omh install --target agents --scope user --status reports the pack's real state instead of missing when only the unused ~/.claude/skills mirror is absent.

    • Memory that does not match a message is listed, not injected; a stuck run changes approach before it stops. When a chat message shares no word with any stored record, the memory pack used to serve every active record in full, and Hermes frames that pack as reference data for the reply, so rules about unrelated work steered it. Those records now arrive as one <record_index> line naming each record's subject (at most eight, then +K more), and omh_remember recall reads one in full when it bears on the request. Records saved in the same conversation still arrive in full, a matching message gets the same pack as before, and CLI recall, omh_memory recall and coding handoffs are unchanged. In dag-ulw, a node's last repair cycle is now a re-plan: its tasks carry the earlier failure reasons and ask for two or three approaches that differ in kind before any edit, unless every failure was an infrastructure cut. The loop stop ladder's first no-progress stop now asks for one re-plan from the goal ledger (replan_stuck_loop_once) and resets the window; only a stop after that names the blocker step. The dag-ulw run loop reference gains a "When the path is unknown" section (cheap scout nodes on different hypotheses first, then extend with the build layer), and the interview skill keeps at least one answer candidate from outside recalled memory.

  6. 4.0.4

    Expandir

    GitHub Release

    • Name your own identity: a prompt and a model chain. Write ~/.omh/routing/identities.json (omh_identity_profiles/v1) with one or more named identities, then select one by putting [omh-identity:name] in a request for its prompt, or --identity name on omh coding model-route, run and delegate (an identity field on fanout prepare units, an identity argument on omh_delegate_route) for its model chain. An explicit --model still wins; the identity chain beats category, role, depth and scale. Nothing is inferred and nothing carries to the next session, and with no file routes are unchanged. Schema and limits are in the Installation guide. Contributed by @yqm-307 (#2459, carried by #2462; requested in #2458).

    • On Windows, OMH uses the Hermes home Hermes uses (#2453 by @k1lgor, carried by #2463). With no HERMES_HOME set, omh, omh doctor, omh setup and omh release hermes-smoke took Hermes' home to be ~/.hermes, while Hermes itself defaults to %LOCALAPPDATA%\hermes on Windows. omh exported that empty home to the terminal it launched, so Hermes opened it and refused every session with storage_unavailable. They now resolve Hermes' own platform default, %LOCALAPPDATA%\hermes with HERMES_DATA_DIR_SUFFIX honoured. POSIX and an explicit HERMES_HOME are unchanged.

    • omh_context returns a compact view. With Hermes' tool_search off, a model could call the plugin tool first and carry its result through every later model call of the session, and two of its sections were what omh_capabilities action=summary already returns. The result now drops capability_families and workflow_context_cards, trims each generic-tool checkpoint route to its family, workflow, next action, fallback action and not-evidence list, and adds view: "compact" with an expand line naming omh_capabilities for what was dropped. Replayed on 209 recorded calls, the mean result went from 40,162 to 18,909 characters. omh context brief, the wrapper contract and the hooks still build the full brief.

    • A coding request that mentions OMH no longer gets the workflow-learning hint. "Use OMH to fix the failing test in billing.py" or "OMH, 로그인 API에 대한 통합 테스트 작성해줘" carried the intent=meta_discussion; selected=workflow-learning route hint on every model call, and a CI-workflow edit through OMH dispatched workflow-learning; they now route as the coding work they are, and a throwaway test message ("I'm testing OMH right now, ignore this message") gets no hint either. A report on OMH's own routing ("OMH picked deep-interview when I only wanted a quick fix; why?") still reaches workflow-learning, on the hint and the chat route. Everyday meta words now match as whole words, so "pytest", "terminal", "changelog", 로그인 and 블로그 no longer count as test, term, log or 로그, and the product name alone is no longer routing context. On the bench prompt the injected OMH text fell from 2,612 to 1,761 characters a message.

    • Opt-in language-server tools: lsp_definition, lsp_references and lsp_rename. With OMH_LSP_TOOLS=1 or the profile setting plugins.entries.omh.settings.lsp_tools, the omh-lsp toolset drives the language servers Hermes already runs for its post-write diagnostics. A symbol matches as a whole identifier only; lsp_rename writes only if the file still holds the bytes it read (stale_file otherwise) and answers already_renamed on a line that already holds the new name. Python with pyright only, and off by default. On an 8-model A/B of 9 rename tasks per model, the rename pass rate went from 47/72 with plain patching to 61/72, the gain on aliases, re-exports and shadowed names (measurement).

    • Opt-in impact_check names the callers an edit broke. With OMH_IMPACT_CHECK or plugins.entries.omh.settings.impact_check on, a turn's first edit to an existing Python file records the errors in the workspace files that import it, and at the turn's first stop OMH asks the language server Hermes runs about them again; if any gained an error, the turn continues with at most 15 lines of path:line: message. It says nothing otherwise and is off by default. lsp_diagnostics(path|glob) joins the opt-in omh-lsp toolset, so the model can re-check those files.

    • A rename's stale export is named before the turn ends. When a turn leaves a Python module exporting a name it no longer defines, OMH names the line at the turn's first stop as path:line, so the model can fix it. That covers an __all__ entry, and a from module import name whose workspace module no longer has the name. It needs no language server and is silent otherwise. Measured on four models, the rename template that models kept missing went from 8/12 to 12/12, with +3 of that from the note itself. The pre-registered gate failed on its token clause (1.097x against 1.05x), and the owner overrode it because the gap was run-to-run variation (measurement).

    • Vim mode in the composer. With Hermes' display.vim_mode on (hermes config set display.vim_mode true, then start omh again), the OMH TUI's composer edits modally, read from the active profile's config. It starts in INSERT, where every key works as before except Esc, which goes to NORMAL and closes the / or @ menu, and does nothing else: it no longer arms the clear or the rewind, cancels nothing and interrupts no turn. In NORMAL a letter is a command, never text: h l 0 ^ $ w b e j k gg G, i a I A o O, x X dd D C cc S s r, d c y with a motion, p P, u and Ctrl+R, with counts, and Esc there does what it always did, so Esc Esc still clears or rewinds; during a turn, Esc on an empty composer arms the interrupt and a second Esc sends it. Enter sends from either mode, and in NORMAL always sends, a \ before the caret included. Paste chips and image tokens are one character to every command, p puts them back with their text or image, and the info line says INSERT or NORMAL. Questions Hermes asks, masked input, the tell row, pickers and the agent list keep their own keys. The supported set and how it differs from Hermes' own Vi mode are in the Installation guide. With the setting off or absent nothing changes. Requested by @DanRioDev (#2434).

    • agent-debug sees a turn that ended without a reply. A new finding, silent_turn_end, cites a turn whose last tool result was followed by an empty assistant row and then nothing: no reply, no further call, until the next user message. It cites the tool row and the empty row, the call, and the result's exit code (130 for an interrupted command). When one is found, the incident leads with turn_ended_without_reply ahead of correlations such as a compaction or an unwatched background process, because the silent end is the symptom itself, and proposes reading the host's agent.log Turn ended: reason= line for that timestamp before changing anything. Both state.db and --session-record are read; without tool_calls the kind is listed unavailable.

    • A Kanagawa Wave theme and launch look. omh theme use kanagawa selects an indigo palette with warm ivory text and muted blue and violet accents, adapted from rebelot/kanagawa.nvim (MIT, Copyright (c) 2021 Tommaso Laurenzi). Like the other four it sets no background, so the terminal's own shows through. /theme kanagawa-wave selects the matching launch look: a rule above and below on a dark sumi-ink ground (Kanagawa Lotus on a light terminal), with an ocean swell for its living wallpaper. Contributed by @DanRioDev (#2436, carried by #2441).

    • A failed CI or merge watcher no longer leaves the session silently stopped. When a watcher you armed exits non-zero, its notice opens a turn; that turn now starts with a short directive: if the work it watched is still the session's to finish, do it and arm the next wait, otherwise say plainly that the session has stopped and what you have to decide. If the turn ends with nothing armed anyway (no notifying background process, background delegation or scheduled follow-up job), the reply ends with one sentence in its own language saying nothing will wake the session again. Both halves read only records: the host's row kind and notice header, and the turn's own tool results. A killed watcher, a zero exit and a person's own message get nothing.

    • A steer reads as the prompt it is. A steer you send during a turn shows as a prompt row (❯ <text>), as Claude Code shows one, in place of the dim ↪ steer queued by Hermes · <text>. It stays dim right above the working row until the turn takes it in, at the first tool that finishes or when the turn ends, then reads as any prompt you sent where the model got it. Hermes reads a steer only between steps, so one sent during a long tool call waits there until that call ends. A turn you stop before taking it in says not read — the turn stopped first.

    • A coding turn ends with one consistency check, when the edit has a copy or a reference elsewhere. When a turn in the coding posture edited source code, OMH can continue it once at its first stop. It asks the model to look for other copies or call sites of the logic it changed that have to change with it, and to fix them, without re-running tests for the check. That is one extra round, so after a Python edit it is skipped when none of the four signals below is found in the workspace's other Python and config files. It fires when:

      • a changed definition is also defined in another file;
      • a removed or renamed name is still used in another Python or config file;
      • a function whose signature changed is named in another file, or a class whose constructor, annotated fields or abstract methods changed;
      • a replaced line appears in another file.

      It also fires, as before, when that search cannot be made:

      • a non-Python source file changed;
      • a Python file changed with no pre-edit text kept for it: through sed, execute_code or a delegated child, or past OMH's snapshot limits;
      • a changed file does not parse or is too large;
      • a change touches no definition, import or docstring;
      • the workspace walk passes 5,000 entries or its time limit (a workspace that ran over is not walked again for five minutes);
      • a Python file or a directory in the workspace is a symlink that leaves it;
      • an unedited Python file is too large or unreadable;
      • any error.

      There is no other fallback. A change the four signals cannot see is skipped: a function body that changes behaviour without changing its signature, and a reference that exists only in prose such as docs. The search skips hidden, version-control, virtualenv, cache and node_modules directories, and build output: a build directory holding a lib, lib.*, bdist.*, temp.* or __editable__* directory or no .py file in its first two levels, or a dist directory with no top-level .py file. Any other build or dist, such as a scripts directory or a namespace subpackage, is searched.

      On the bench corpus the always-on check fixed a second copy of the changed logic far more often: 76 vs 38 of 90 runs, pooled over five measurements. It cost about 1.37x the median tokens. The trigger kept those fixes and stopped the check on edits with no copy: 0 of 54 sessions against 54 of 54, with the same passes. Strong models' median tokens fell to 0.68x and 0.75x of the always-on check in two runs. One clause of the pre-registered gate failed in each run; the owner overrode it. The decision is recorded in the bench README.

      If nothing more changed, the model replies with the single line OMH: no further changes., and OMH puts the answer it had already given back in that line's place, so --oneshot, API and delegating callers get that answer unchanged. Hermes's own CLI shows the restored answer a second time; the OMH TUI keeps the one line. Delegated children never get the check. agent.max_verify_nudges: 0 turns it off, together with OMH's other end-of-turn nudges.

    • omh update no longer deletes the skill pack a running Hermes gateway still uses. A gateway follows current/skills to the generation behind it once and keeps that directory until it restarts. Update kept only the active and the previous generation, so two updates in a row deleted the one the gateway was serving, and Hermes showed zero OMH skills (skills 89 installed on the launch card, OMH's 0) until a restart. Update now records when each generation was activated and also keeps the one that was current at each gateway's last recorded start, in the default home and every profile, until a later update sees that gateway started again. The hermes gateway restart hint still names the restart that picks up the new pack. With no recorded start, retention is unchanged. The old command runs the cleanup, so this holds from the update after the one that installs it.

    • The info line shows the context as a gauge with its token counts. The bare ctx 37% under the composer is now ctx ━━━━──────── 37% 74k/200k: twelve cells, the filled ones and the percent in the context's colour (warning from 80 % full, error from 95 %), the rest dim, then the tokens the context holds of the most it can, in k and M. On a narrow line the token counts go first, then the gauge, leaving ctx 37%, before the context goes. Under NO_COLOR the filled and the empty cells still tell how full it is.

    • Korean typed from Termius keeps its spaces where they were typed. Termius sends every syllable a Korean (or other) input method commits as its own bracketed paste, and on a slow link the space typed after a syllable arrives in the same read. The OMH TUI asked Hermes whether each one-line paste was a dropped file and typed it only once Hermes answered, while the space went in at once, so 안녕 하세요 came out as 안 녕 하세요 and the last syllable of each word slid past its space. Only a paste shaped like a path (/, ~, ./, ../, a drive, file://, the same first test Hermes' drop detection makes) still waits for that answer; any other text paste is typed at once, in order with the keys around it.

    • omh coding run --owner codex finishes when an agent shell starts it (#2402, reported by @vince2930). Every unit inherited the dispatcher's stdin. An agent shell such as Hermes' terminal tool keeps an open pipe on that stdin and never closes it, and codex exec "<prompt>" reads a piped stdin to EOF even with the prompt in argv, so the unit waited until its deadline and failed as timeout (exit 124) whatever --timeout said. A unit, a declared verification command and the dispatcher's git calls now get the null device as stdin; the workspace preflight's git blob, the one call that passes input, still goes through its own pipe.

    • A browser-game request reaches frontend, which looks at the scene after every change and passes the game only when it boots, answers input and keeps running. A request to build or polish a browser game ("three.js로 간단한 3D 게임 만들어줘", "make the game feel juicier") went to an interview, a clarifying question or a fallback; it now dispatches frontend. A native engine (Unity, Unreal, Godot, Roblox), a how-to question and topic uses such as game theory, a trailer or tonight's score stand that down, and a direct code fix ("fix the collision bug in my game") keeps the delivery lane. While building, frontend captures the affected surface (for a 3D or canvas scene the default camera and each declared preset) and reads the capture before the next change; a change not looked at is not done, and the final 1440/768/375 loop still runs. visual-qa gains references/playable-check.md: a game or interactive scene passes when the booted first frame is not blank or one flat colour, the frame after the declared input differs where the input should move something, idle frames seconds apart differ (background-only animation does not count), and the console shows zero errors across boot and input. A missing capture is BLOCK, reading code never passes, and a run with no WebGL context is recorded as environment: no_webgl instead of being judged a blank frame.

    • A Kimi main session now gets a short execution calibration of its own. The new omh.main-calibration system-prompt section renders once at session start (no per-turn cost) for a main session whose model resolves to the Kimi family: decide minor choices and note them, ask only before irreversible or externally visible actions, call the obvious tool instead of describing it, read the request's verb (explain or check answers, build or fix does the work), and bound repeated searches. Delegated children (platform == "subagent") get nothing from it and keep omh.child-calibration; every other family renders nothing until an A/B shows it needs a block.

    • A steer sent while a reply streams no longer erases that reply. When Hermes read the steer before a second model call, the OMH TUI replaced the whole streamed text with the last call's at turn end. Once you steer, what streamed now stays as it was, and only what the final adds past it (a footer) is added. A long one-paragraph reply now wraps only its last rows each frame, so a frame no longer costs more as the paragraph grows; agent names drop a goal's trailing colon (Interrupted child alpha.); and /title X shows the title bold instead of **X**.

    • Every OMH TUI launch look but the default has a living wallpaper: its own background colour and a pixel-art scene that keeps moving. While the TUI runs, a look with a scene sets the terminal's background to the look's colour with OSC 11 and puts the terminal's own back on every way out: on exit, after a crash or a signal, while the job is stopped with Ctrl+Z, and as soon as the wallpaper is turned off. The default look, spectrum, draws no scene and keeps the terminal's own background. Every scene is drawn in half blocks, two square pixels to a cell (▀ in the top pixel's colour over the bottom one's, ▄, and a space on the colour where both match), in the language of the oh-my-hermes pixel-aura short: a mosaic of solid square tiles, dense in the middle and scattering out to single tiles, with four-point + sparkles that grow and fade. Each look keeps its motif in that language: moonlit a clean crescent moon in a pixel halo with soft tile clouds on deep navy; sakura blossom clusters with falling petals on dusky plum; aurora ribbons of scattered tiles on deep teal; ember a flickering flame in a warm glow with rising embers on dark brown-red; violet pastel jellyfish with pulsing bells; gold a sun in a breathing glow; graphite a cloud raining into a rippling puddle; and kanagawa-wave an ocean swell. The motifs sit low on the right, where the launch card and a working session's rows leave the screen blank, and a big screen (140×40 or more) draws the sprites' art pixels two pixels square. A light terminal gets each look's pale ground. The theme is measured against the ground, and every word and mark also keeps its floor against the terminal's own background, where the session's rows stay after exit; a look changed mid-session prints the session again on its ground. At most a sixth of the screen is lit and no pixel passes 3:1 against the ground; colours keep their hue under that cap (mixed from a dark shade rather than into the ground) and a motif's shades are spread within it instead of collapsing onto it. The scene is drawn on blank cells only, never over text, a band or into the scrollback: blank rows whole and the tail of a row with text two cells past it, the transcript's rows too, so the motifs stay in view over a full session (a selection across rows then copies those glyphs; /wallpaper off keeps it clean). It repaints only the cells that changed every 0.25 s, pauses while the composer holds text, and stands still with reduced motion. It shows at truecolor on a terminal that answered the background query, not under NO_COLOR, and not where /theme light|dark names the other side from the terminal's own. It is on by default; /wallpaper on|off (alone, it switches) and omh theme wallpaper on|off save the choice as wallpaper in the theme record, and OMH_TUI_WALLPAPER=0 holds it off in one terminal. The /theme picker shows each look's colours on its ground.

    • Copying from the pager after an emoji copies the right text. The pager's drag, double click and highlight counted columns by code point while the row is painted by glyph, so after a ZWJ family, a skin-tone or VS16 emoji the cells under the pointer were the wrong ones (a double click on the next word copied nothing) and the highlight could change style inside one emoji, breaking it. They now count by glyph, as do the recall glint and the running box's aura in /dag. A copy hands back the text's own code points, so a decomposed macOS path from /logs still matches its file, and a subagent's log row and a bot's message are cut between glyphs, never inside an emoji, a flag or a letter with its accent. /logs shows a tab as spaces and no longer lets a colour code in stderr shift the columns a click or a drag lands on.

    • The second Ctrl+C of the OMH TUI exits, and the launch, a resize and the import cost less. After a Ctrl+C interrupts a turn, the queued prompt it puts back in the composer no longer eats the second press: that press exits as the info line says, with Shift or Alt held too, and only while the line says so; after an interrupt it says so over an open steer, rewind or tell as well. What the composer holds is not kept on exit; prompts already sent or queued are in Hermes' history, where Up finds them, and a submit still waiting on its pasted image keeps the press from exiting. The launch logo's contrast lift parses each colour once and is cached (first paint about 12.6 -> 10.3 ms, same bytes); a full redraw prints the rows it measured instead of rendering the kept blocks twice (a long session's settled resize about 270 -> 205 ms); and the bot board's module loads when the board opens, not with the TUI.

    • The OMH TUI has four more launch looks, and a look now tints the chrome. /theme adds sakura (pastel pink, sky blue and cream), moonlit (moonlight yellow and silver over navy), aurora (emerald and teal into violet) and ember (amber through orange to crimson) to spectrum, violet, gold and graphite, each with a set for dark and one for light backgrounds. Every look, the four earlier ones too, now also paints the composer's rules and the frames drawn like them in its card-edge colour, and a tool call's rail and the working row's spinner in its label colour, each lifted to 3:1 against the background; a failed or denied call's rail stays red. Body text, the grey details, the reply's rail and the state colours keep the theme's, and spectrum stays the default with its logo unchanged. The picker, /theme <look>, Tab completion and the saved launch key take the new names; an unknown saved name still warns and falls back to spectrum. (A ninth, kanagawa-wave, is described above.) On a terminal too short to list every look, the /theme picker scrolls so the highlighted look always shows, moving only when the highlight would leave it, and a click lands on the look drawn in that row.

    • A light goes round the OMH TUI's screen while a turn runs. It runs along the terminal's whole edge, clockwise, once every four seconds, with a round head and a trail in the launch look's logo colours that fades out over a seventh of the way round, half a row thick along the top and bottom rows; nothing else of the edge is drawn. It goes round a session of any length, over the blank rows under a short one, and is gone when the turn ends. The top row's own text stays painted under it, and only the cells the light reaches cover it. It moves by the clock and repaints the edge every 0.05 s, so it glides round corners instead of stepping. It needs the truecolor depth with motion on. /glow on|off (alone, it switches) is saved beside the theme under the OMH home, so it holds in every later session; omh theme glow on|off saves the same choice from the shell, omh theme glow shows it, and OMH_TUI_GLOW=0 holds it off in one terminal. /glow on says when something still holds it off: OMH_TUI_GLOW=0, reduced motion, a colour depth below truecolor, or a screen smaller than 20×8. To keep the edge free, the TUI draws its rows one column in from each side, a prompt's band included, and keeps the screen's last row blank, whether a turn runs or not, so nothing moves as the light comes and goes; /glow off gives that margin and row back. The light is taken off before anything scrolls, so none of it reaches the scrollback.

    • Gemini 4 Argon replaces Gemini 3.1 Pro in the shipped chains. Google made Argon generally available on 2026-10-13. Every slot 3.1 Pro held now names gemini-4-argon at the same effort: writing and google-search at medium; visual-engineering, artistry, x-search, and synthesize at high; and the X-platform affinity. The id has an exact contract read from ai.google.dev: 1M unified context, 65,536 output per synchronous request (1M with long decode), and a low/medium/high thinking ladder with medium as the default and no off switch. off and minimal are raised to low on record; xhigh and max pass through, because nothing documented sits above high. Dated snapshots project onto it. Cost estimates use the 2/10 list price with 0.10 cached input, which rises to 4/20 on 2026-12-01. Provider checks know that Google's API and OpenRouter serve it. On Hermes, 3.1 Pro's medium slots were sent as low, because Hermes allows only low and high for a Gemini 3 Pro id. Argon's medium is sent as medium. Gemini 3.1 Pro is retired under the superseded-generation rule. It stays recognized, priced, and provider-mapped, so omh model-chains set can still name it. The gemini family calibration applies unchanged. Both site pages, the generated install table, and the model-setup skill say so. This is an editorial change (owner decision) and has not been measured: no served Gemini route was available to probe.

    • Six more model categories you can route a subagent to. google-search (Gemini 4 Argon first, for research grounded in Google Search), x-search (Grok 4.7 first, for live search on X), review (GPT-6.1 Sol at xhigh, so a different model tries to break the diff), swarm (DeepSeek V4.1 Flash at low, for many parallel workers), forensics (GPT-6 Astra at xhigh, for root-causing a live symptom or a trace) and synthesize (Claude Opus 5.5 at xhigh, for merging many results into one answer) join the eighteen shipped chains, in the Hermes lane, the Codex and Claude Code tables, the delegate-route tool and the public chain table. Each chain ends on a model from another provider. They are opt-in: no role or task size picks them, so every existing route is unchanged. Two heads repeat an earlier category's (forensics shares ultrabrain's, synthesize shares architect's), so an unrouted child on those models is still labelled with the earlier name; a lane routed through omh_delegate_route keeps its own. Six HUD labels move with the new chains: DeepSeek V4.1 Flash at low now reads swarm, Gemini 4 Argon at medium google-search, Kimi K3.1 Pro at high x-search and at xhigh review, and GPT-6.1 Sol at xhigh and Grok 4.7 at high, which had no label, read review and x-search.

    • The pager copies on drag, without the transcript's left bar. In a pager (Ctrl+O, /history, /logs, /dag, an agent's log, the bot board) the TUI now takes the mouse: the wheel scrolls three lines, a drag selects and copies on release without the rails or a reply's margin, a double click copies a word and a triple click a row, and the footer says ✓ copied N chars to clipboard. The copy is OSC 52 for kitty, Ghostty, WezTerm and Alacritty, pbcopy on a local Mac otherwise, and else OSC 52 the footer says was only sent. The main screen keeps the terminal's own selection, since taking the mouse there would take the wheel and the scrollback, and there each block keeps its thin ┃ rail, which that selection copies: a painted full cell that copied as a space read too heavy. Under NO_COLOR your prompt's rail is the light │, so it is told apart from a reply's.

    • Fourteen more models cost out at their vendors' list prices. OMH already recognized Grok 4.7, 4.6, 4.5 and 4.3, DeepSeek V4 Pro, Gemini 3.8 Flash, 3.6 Flash and 3.5 Flash-Lite, GLM-5.3-FlashX, Qwen3.8-Max and Qwen3.8-Flash, Solar Pro 4, MiniMax-M3 and MiniMax-M3.1-Flash, but it had no price for them, so a run on one that the host recorded no cost for showed no estimate. Each now carries the first-party list price read on 2026-10-28, with its cached-input rate where the vendor publishes one. The rows also note what the table cannot represent: Grok's 200K whole-request tier, DeepSeek's off-peak half rate, the Gemini Flash introductory price that doubles on 2027-01-01, and MiniMax-M3's 512K tier. The xAI, DeepSeek, Gemini and Z.ai ids also get provider checks; Qwen, Solar and MiniMax have no provider family to name. No shipped chain named any of them until Grok 4.7 joined x-search (above). GLM-5.2 was costed at the GLM-4.x rate (0.6/2.2) and now uses its own 1.4/4.4. Its Ultrafast serving moves from a halved figure to the same base rate, as no separate tier rate is published.

    • The route hint stops saying the same thing twice. Route hint text is about 18% of the OMH text a model reads per session, and part of it repeated other text in the same block: a hint's adjacent_workflows= line that equals the block's top-level one, and a fallback_action_label= or next_action_label= that only respells its action's underscores. Each one now renders only when it differs from the text it would repeat. Rendered before and after on the same 68 routing-intervention hints, the block went from 106,957 to 98,121 characters (8.3%), and no hint grew. Selection, the payload, first_response_shape= and not_evidence_yet= do not change, and the per-turn ratchets drop with the render (pre_llm_call 5786 -> 5632, fallback 7338 -> 7184).

    • A long turn in the OMH TUI no longer ends with blank rows over the composer. Once a turn's output scrolled the prompt off the screen, the ↳ you: reminder and the blank row over it took two rows on the bottom rows. That scrolled two transcript rows into scrollback, where nothing can bring them back. When the turn ended, the region kept its height, so the done line (✻ 7m 11s · 12 tools) stood three blank rows above the composer, and with the plan shown any turn left one more over the plan, whose own blank row the running turn had. The frame that ends a turn now gives up the rows nothing in it takes: the reminder, its blank row, and the blank row over the plan. The done line sits one blank row above the composer or the plan, as between any two blocks. The composer rises by those rows, at most three, left blank under the footer, and the next turn's output takes it back to the bottom rows.

    • The answer-first plan line stops repeating its whole rule. While a plan is open, every turn a person opens carried the 445-character TODO_ANSWER_FIRST_RULE, and Hermes keeps each turn's injection in history until compaction (11,125 characters over 25 turns in one owner's three days). The rule now spends the reconciliation rule's turn budget: whole on the three turns after any plan write, then TODO_ANSWER_FIRST_RULE_SHORT (253 characters), which keeps the precedence, both incident guards and the record-first resume. A live A/B over five models, a checkable claim in a person's turn after 20 quiet plan turns, counted unchecked agreement or deferral in 1 of 119 turns with the short line, 2 of 120 with the full rule and 4 of 48 with the pre-#1705 drive.

    • A routed session carries the route hint's boundary sentence once. Every [OMH Route Hint] closed on the same 197-character Boundary: line, and Hermes replays each hint, so a session routed on several turns held one copy per turn (3,090 characters over 15 hints in one owner's three days). A later hint now drops the line when the host-injected part of history already carries it; a first hint, an unreadable history, or a person quoting the line keeps it, and the route fingerprint is still taken from the whole hint.

    Image: Sprite Forge patch notes

    • Character sprite sheets become aligned game frames (sprite-forge). An image model draws a character sheet whose frames differ in scale and foot position, so the character swells and bobs in game. The new sprite-forge skill keeps the sheet a grid on flat #FF00FF (never a 1xN strip, wide effects in their own sheet) and ships an optional standalone Pillow helper, toolkit/scripts/forge.py, that keys the background, splits cells, applies one common scale and puts every body's feet on one shared baseline. qc.json records body_scale_cv and anchor_y_std against 0.08 and 0.05, and a REGENERATE verdict exits 3 so the sheet is redrawn rather than patched. Without Pillow, --help still works and any processing request exits nonzero with no output. OMH generates no images; the host's own image tool draws the sheet. (#2335)

    • visual-qa checks motion in animated and simulated output. A new references/motion-continuity.md checks spawn position against the declared origin, per-frame displacement against the declared speed, force direction against observed motion and monotonic HP bars, and it replays the deploy, travel and engage flow independently instead of reading it off cost or clear numbers. (#2335)

    • show-reel captures reels from a running game. references/game-capture.md keeps events, sound cues and titles in one timeline.json, pins the game build, holds a reel whose gameplay event has no sound cue, and records a real voice's source URL, hash and excerpt; impersonating a real voice is refused. (#2335)

    • deliverable-package tracks archived and published. A file is archived only with its hash, version and catalog row read back, published only when the destination's own URL resolves to the same hash, and the final file is the one the person approved, never the newest mtime. (#2335)

    • An approval in the OMH TUI shows every risky command, not just the one Hermes named. Hermes names a command line by the first dangerous pattern it matches, and the slot printed the line as one wrapped block. One real request chained 14 commands and read as "recursive delete", while the same block force-removed a worktree, force-deleted two branches and deleted a remote branch. The slot now splits the line at its top-level &&, ||, ; and line breaks, keeping a here-document with its command and reading no comment. Each command is listed on a row of its own, and each one that runs rm recursively or on a path from /, deletes or force-pushes a remote branch, force-deletes a local branch, force-removes a worktree, hard-resets or force-cleans is marked ▲, with the home directory written ~ where the shell would read it so. So are the cases Hermes' approval table flags of a download run as a script (curl … | sh, bash <(curl …), eval "$(curl …)", sh -c "$(wget -qO- …)", and inside a -c script, a substitution or after then), decoded text piped into a shell (base64 -d | sh, xxd -r, echo | tr, openssl … -d), find -delete or -exec rm, and chmod with a mode that lets others write (777, 666, o+w, a+w); echo "$(curl …)", x=$(curl …), diff <(curl a) <(curl b) and find … -print are not. A quoted here-document body (<<'EOF') inside $(…), such as a commit or PR body, is read as text; an unquoted one's $(…) and backticks are read for what they run, a << inside $((…)) is a shift, and a here-string (<<<) or a nested subshell (((cd a; rm -rf b); ls)) hides nothing after it.

      When rows run short and Hermes' own description names a risk the slot marked, the spacing and the rule go first, then the rows fold step by step into one + N other commands (echo, git, …) · ctrl+o shows all row: first the commands that change nothing (echo, ls, grep, read-only git and gh and the like, writing no file, with no assignment in front and no risk inside quotes; a cd with nothing listed after it folds with them), then directory changes (cd, pushd, popd) and variable assignments (T=/, printf -v, ${T:=/}), the last one before each risky command going last, then every row that is not risky. A ▲ row is never cut while another command's row shows. Hermes knows far more dangerous patterns than the slot, so when it flagged one the slot does not know, or named none, nothing folds and a long list is cut from its end with the usual Ctrl+O marker, risky rows too. Risks Hermes did not name get a ✦ OMH also sees: row under its header. Quoted text and the arguments of echo or grep never match. Each choice now has an icon and a tone: ✗ deny, ✓ allow, ★ always allow, ✎ tell Hermes. The highlighted one is bold on a band of its tone. Ctrl+O still pages the command whole, as Hermes sent it. Deny stays first and selected, and the keys do not change.

    • The OMH TUI says when OMH caught a model mistake. OMH's catches reached only the model, so an install looked like it did nothing. Now the moment one happens the transcript gets one row, ✦ OMH stepped through the launch look's logo colours and the words in its label colour, so it follows /theme: ✦ OMH 🔁 stopped a loop of N identical calls (the repeat guard's first refusal of a streak), ✦ OMH 📖 pointed the re-read at line 1667 (the first note on a truncated read's refused re-read), ✦ OMH 📋 kept going, plan has 2 open items (a turn continued over an open plan), and ✦ OMH ⏰ flagged a wait with nothing armed to wake it (the turn's unarmed remote-wait directive). Nothing prints otherwise, and a resumed session does not replay earlier catches. The hooks record each catch beside their unchanged decision in <omh_home>/runtime/catches.json (omh_catch_log/v1): a kind from a closed set and small integers, never model text, arguments or paths; the HUD payload carries the reading session's newest few as catches.

    • The OMH TUI shows a session's background processes on the info line, and offers Enter when one ends unreported. A terminal command Hermes ran in the background with notify_on_complete off ended without waking the session, and nothing on screen said work was still running. While this session owns one, the info line under the composer says, after the badge and the branch, ✳ background terminal(<command>) 1h 07m 13s and, once the turn is over, who will be told: 🔔 will ping you when done, 🔔 will ping you on a match (watch patterns), or 🔕 Hermes won't be pinged · OMH is watching it. When a process in that last state ends after the turn, the row turns green, ✓ … finished 0:12 ago (+N more when others ended too), the bell rings once, and Enter on an empty composer sends one prompt asking the session to check the results; nothing is sent on its own. As the line narrows the title goes first, then the row shortens and goes, and only then do the model, the context and the effort shorten. The TUI reads Hermes' processes.json read-only and matches a process on the session's own key, its stored id or its local route agent:<namespace>:local:dm:<stored id>, never a prefix or suffix of either; with compression.in_place: false the row and an unreported end survive the rotation to a new stored id. A messaging session resumed in the TUI (a bot profile's Telegram chat) shows the processes its own turns and its subagents started, matched on the route Hermes records on the session's row and on the spawning session's id; after a /new on the same chat, neither conversation shows the other's. It asks a host process by pid and recorded start time whether it still runs. A process it cannot ask that way (a sandbox, or off POSIX) says 🔕 won't ping you · ask Hermes for the result instead, shows only while its row is listed, and is never called finished. The row's spinner steps once a second, like its clock, so an idle session with a process redraws once a second, and not at all under the pager.

    • A generation's own skills pack is never registered, and a home left on one is moved to the pointer. 4.0.3 stopped omh setup from registering the running generation's pack (generations/<id>/skills) beside the shared current/skills pointer -- once the next update moved the pointer, every OMH skill resolved to two files and Hermes listed none of them -- and retires such entries from every home. omh chat --auto-apply-target-change, the last writer of that path, now registers the pointer too. A primary home naming only an older generation's pack is migrated to the pointer the way a bot profile already was, instead of being read as the opt-out and losing every OMH skill once the pack is collected; only the installer-managed command, which retires the pack in the same write, counts it as a registration. A relative entry, which Hermes anchors at its own home, never matches.

    • A skills entry spelled through ${VAR} is matched the way Hermes reads it. Hermes expands ${VAR} and ~ in skills.external_dirs before it resolves an entry, but setup, update and omh doctor expanded only ~, so ${XDG_DATA_HOME}/omh/generations/<id>/skills beside the pointer was neither retired nor reported and the home lost every OMH skill once the pointer moved. Such entries are now expanded first; one whose variable is unset is left untouched, as Hermes leaves it. A Windows entry rooted without a drive (\omh\...), which Hermes loads from its home's drive, is matched on every Python version instead of through 3.12 only. A relative entry, which Hermes reads under its own home, is no longer resolved against the directory omh runs from, so update cannot retire it or read the home as registered through it, and doctor cannot count it as a second managed directory. Uninstall's check for a store another Hermes home still names, omh hermes readiness and the installed-skill weight advisory expand entries the same way and read a relative one under the home that names it, so uninstall keeps a store a home names as ../omh/skills.

    • Setup, update and doctor read every skills entry from the home that names it. The registration migration, the generation-pack retirement, the two-directories finding and the registered check now read a relative entry under that Hermes home and a Windows entry rooted without a drive on the home's drive, as Hermes does, instead of skipping the first and resolving the second on the drive omh runs from. So ../.omh/skills or ../omh/generations/<id>/skills beside the pointer is retired and reported, and omh update run from D: still retires \omh\... when the home is on C:. Registration no longer adds the pointer beside an entry that already names it another way (C:\omh\current\skills on Windows, ${VAR}, ~, or relative to the home), and omh uninstall --registration-only removes every spelling setup retires (${VAR}, ~, relative to the home, or through a symlink). Nothing is retired or removed that Hermes does not read through that home.

    • /compress works in the OMH TUI, and /compact runs it. Hermes' shared gateway (Hermes v0.22) refuses /compress as a typed command, so the TUI said Hermes could not run it in a TUI session. The gateway child now runs it, and /branch, /retry, /undo, /stop, /bg, /btw, /rollback, /reload, /reload-mcp, /tools, bare /plugins and /config, through the gateway call Hermes' own TUI uses. /topup goes to Hermes only on a shared gateway, and the commands only Hermes' classic CLI runs leave completion. While a compression is out and no turn runs, the working row shows a spinner and Compressing… with the time it has taken so far, until the report or the unanswered notice lands; it offers no Esc, since Esc between turns does not stop a compression. The TUI waits 660 s for it, Hermes' own budget, and past that says the compression may still land rather than inviting a second one. Hermes answers a connection's calls one at a time, so every call sent meanwhile (a command, a prompt, an interrupt) waits for the compression's time left plus its own budget, at most 780 s, and a typed command says Hermes runs it once the compression ends; a call that still goes unanswered says how long it really waited. A command Hermes refuses while it compresses says so. /stop reports an interrupt only when it stopped a turn, and stops the session's processes even when its read fails. A /bg or /btw answer for a session left meanwhile names that session, and leaving a session with one still running says it will not show.

    • The OMH TUI no longer shows a blank screen when you switch to its terminal and back. A resize signal (a tab shown again, a split laid out again) redraws the session from memory, and the erase that starts that redraw was written on its own before the session was printed again, so the screen went blank for a moment. The erase now goes out in the same synchronized update as the frame drawn after it, so a terminal that holds the update (DEC 2026) swaps the old screen for the new one in one step.

    • The TUI's menus lose their bands, and a sent prompt gains one. The completion menu, the session, model, rewind and launch-look pickers, the agent list, the delegation view and the bot board's sessions draw no fill: the highlighted row is a ❯ and a bold name in the skill hue, other commands keep that hue and their descriptions are muted. A prompt you sent runs on a band to the right edge, tinted by the launch look. The working row, the plan and the composer get a blank row between them where the terminal has one to spare. During a long turn the blank row over the working row stays, and the oldest row goes instead, while six rows still show above it; on a shorter terminal it gives way first, so it never costs one of the few rows.

    • The live region reads more calmly while a turn runs. A mid-turn ↳ you: row and a steer run on your sent prompt's band, with a blank row over the sticky prompt. Hermes' background review is one dim row naming each skill once. The plan's done rows are quiet and fold into ✓ N done before a row still to do is cut, and Ctrl+P folds the plan to its header and current item (the header names the key, and offers ctrl+p expand only where expanding would show more). The working row keeps Hermes' spinner verb and shows a status such as 🧠 OMH — recalled 4 memories after its facts, dim. Each reply opens with ⚚ Hermes · <model>.

    • A redraw due while a pager is open no longer leaves the session twice in scrollback. A catalog, rewind or resume answer that redraws the session while a pager is open used to erase the pager's screen, and leaving the pager printed the session again under the main screen's rows. The redraw now waits until the pager is left.

    • A promo or demo video asked for in a shared channel reaches show-reel only when the channel's backfill is greetings and thanks. On Discord, Hermes puts the messages others posted since the bot's last turn before [New message], and a closed request ("Make a demo video for the app") was admitted whatever that backfill said, even one describing the fault the video would record. It now routes as it would without the backfill only when the backfill is the Discord adapter's own shape -- the [Recent channel messages] header, the [unverified] notice exactly when a sender is tagged, and [<name>] <text> lines with a one- or two-word name -- and every message is acknowledgement-only chatter: greetings, thanks, acks, times of day and "all", "everyone" or "folks" ("morning all", "thank you", "ok, nice!"), with no noun, verb, number, question mark, link, code or emoji and no pronoun but the "you" of "thank you". Any other backfill makes it ask, as a reply pointer or an attachment note already did. A closed request beside a symbol or modifier character (Unicode So or Sk, such as ˜) or a character that folding removes entirely (U+2017, U+203E, a bare combining mark) is no longer admitted either.

    • An image you attach in the OMH TUI is an [Image #N] token in the text, as in Claude Code. Ctrl+V, a dropped file and /image put the token at the caret, in place of the 📎 shot.png Backspace removes row above the composer. Images are numbered upward through the TUI run, past any number a resumed session already holds. The TUI keeps the image's bytes and Hermes holds it only while its prompt is sent: the prompt stages exactly its own images once Hermes has admitted the turn before it. So deleting the token (Backspace, Cmd+Backspace, Ctrl+W, Ctrl+U, Ctrl+K, Ctrl+C, Esc Esc) removes the image; before, Cmd+Backspace cleared the text and left the image attached. An undo puts both back, and so does Up/Down through history over a draft. The sent prompt keeps the labels and lists under its echo each image whose staging Hermes answered (⎿ [Image #1]). A queued prompt, and each command in flight, goes with its own images. A ! line or a command that sends no prompt gives its tokens back to the composer, with a notice. Typing the label yourself attaches nothing.

    • A queued prompt can be sent into the running turn now, and one line sits under the composer. As in Claude Code, a muted row under the queued prompts says ctrl+enter to send now · ↑ to edit while a turn runs; on an empty composer the key sends the oldest queued prompt into the turn as a steer and takes it off the queue. A queued ! line, skill prompt or plan prompt keeps its place and shows no row. Where the terminal has not shown that it tells Ctrl+Enter from Enter, the row names Ctrl+J, which every terminal sends apart from Enter; with text typed, Ctrl+J still adds a line and Ctrl+Enter steers that text, as before. A sent-now steer Hermes refuses goes back to the head of the queue as it was, or to the composer with a notice when the next queued prompt has already started. One Hermes took counts as read at the first tool that finishes after it, since Hermes joins it to that batch even when the batch is interrupted; one it had not read when the turn is stopped is queued again, with a notice. A prompt held for the default-model check is never replaced: the queue waits behind it. The info line no longer says /steer <text> steers; /steer <text> still works and stays in /help and the ? list. The info line starts with [⚚ OMH Native] and the session directory's git branch with its changed and untracked counts (branch:<name> | !N ?M), then the mode and the keys that matter, with the title, model, context and effort on the right. The badge's OMH is a pale silver blue (#C3D2E3, moved toward black on a light background as far as body-text contrast takes) beside the launch look's Native. Between turns it no longer says ? for shortcuts: the launch card names ? every shortcut, and ? still opens the list. As the line narrows the right side's fields go first, then the branch's name is cut, then the badge shortens to [⚚ OMH] and ⚚; the mode's label and a question's or picker's keys keep their words. The start-up query now asks every terminal for its kitty keyboard flags (CSI ? u): one that answers gets them pushed, Otty with its kitty-keyboard setting on among them, and names Ctrl+Enter from the first frame; a terminal the environment does not name (over SSH) must also name itself through XTVERSION as one known to read them, never Ghostty. The push is undone on every exit but kill -9.

    • The exit card is the session's title and how to resume it. Leaving the OMH TUI printed five lines: the title, a line of id, turns, messages and duration, a line of tokens and model, omh --resume <id> and omh -c "<title>". It now prints the title (untitled session when there is none) and, after a blank line, omh --resume <id> alone, ready to copy, with the whole id and -p <profile> for a profile launch (omh -p work --resume <id>), then a blank line so the shell prompt does not sit right under it. The counts and the -c "<title>" line are gone, since the id always opens this session and a title can be shadowed by a newer <title> #N.

    • The /sessions picker reads as a table, and /resume is /sessions. The session you are in heads the list under Current (named this session until it has a title), and the rest sit under a muted heading for the day each started (Today, Yesterday, Previous 7 days, Earlier), in Hermes' order inside a day. Age and message count are right-aligned in columns of their own, and a rule sets the list off from its header and its filter row, in the /rewind and /model pickers too. The arrows and Page keys move past the headings, a heading never ends the visible list with its sessions below, and a filter keeps only the headings with a match. The / menu offers /resume under its own name once what you type starts it and is not in sessions. /res, which highlighted Hermes' /compress, now starts with /resume; every other short prefix keeps the top row it had. On Hermes' shared gateway, which resumes ids only, /sessions <title> and /resume <title> were refused not_found. A refused title is now looked up through the gateway's exact-title session.list, over the whole store, then among the listed sessions ignoring case. The one session with it opens; several (titles that differ only in case) or none open the picker filtered by the title.

    • omh uninstall removes an OMH home under /var or /tmp on macOS again (#2359, reported and fixed by @Kenmege). The uninstall guard that keeps a home configured through a symlink walked every ancestor of the configured path, and on macOS /var, /tmp and /etc are themselves root-owned links into /private. Every home under them, including the per-user temporary roots in /var/folders, was therefore kept with "passes through a symlink or directory junction". A root-owned link directly under the filesystem root now counts as an OS alias; a link anywhere below it, or one a user owns, still refuses, and Windows never takes the exemption. The same rule applies to the menu-bar refresh that omh update skips for a linked home.

    • Delegated agents no longer stop on a repository convention the brief overrides. The goal echo-back told an agent to stop whenever its brief conflicted with anything, so a brief saying "do not create a branch" met AGENTS.md's "branch first" and the agent declined without making a change. The brief now outranks repository workflow rules, and only a brief that contradicts itself is stopped on.

    • An image pasted or dropped right before Enter goes with that prompt. The OMH TUI used to send the text and leave the image in the next draft when Enter beat Hermes' answer to the paste. Now the prompt waits for the answer and goes with the image, as in Claude Code, and the composer shows a muted waiting for the image… row meanwhile, with the seconds after three. A one-line text paste sent at once joins its prompt the same way. Esc, or Ctrl+C, gives the held prompt back. A paste Hermes never answers holds the prompt only until it expires.

    • omh doctor and the probe read the skill pack you actually run, on a migrated install too (#2358, reported by @yqm-307; #2378, reported by @mazzy89). A managed install serves its skills from a generation behind the current pointer, and a host migrated from the 3.x layout keeps ~/.omh/skills as a frozen copy. Three read-only diagnostics looked at that copy instead. The workflow_engine_reach advice measured it, so a migrated host was told dag-ulw was not installed (it postdates the copy) and a host without the copy was told no install existed; neither report could change, since omh update never touches the copy. It now measures the pack doctor resolved, and an unmeasurable install names the file or directory that was missing. The plugin's standalone probe matched only the copy's path in config.yaml, which registers the pointer, so it reported the managed skills unregistered; it now accepts the pointer, a generation's pack, the recorded pack and the copy, matched by real path, and checks for omh-routing in the pack in use. The chat status card carried the same mistake. Separately, omh doctor run from the 3.x bootstrap venv stopped at once with support refuses symlink component: .../generations/bootstrap-legacy/skills, because it read the support files through that generation's link and the support reader refuses every link. Doctor now reads through the link the way omh update does, so the full report prints. A link the installer would also refuse is a blocking skills_dir_link check that names it, and any other refused path fails local_modifications and guidance_projection with the reason, while every other check still runs. The advice still never moves doctor's status or exit code.

    • show-reel now checks a film's motion the way an editor would, and has a kinetic-type lane. motion_check.py reads every frame at the film's own frame rate and reports the share of still frames, each still stretch, cuts per 10 seconds and back-and-forth flashes. The default general profile flags a still stretch over one second; the kinetic profile, for fast launch films, asks for 40% still frames or fewer and no still stretch over 0.6 seconds. --endcard leaves the end card out and --reference prints a reference film's numbers beside yours. music_tools.py analyze --bpm reads fast tracks at their real tempo instead of halving them, and reports where the music drops, which timeline.py exposes as music.drop1 cues. Hard cuts listed in project.json cuts keep render.mjs's motion blur inside each shot, so the first frame after a cut is never a blend of two shots. qa_check.py gains max_h264_level and macOS hardware_decode policy keys, finish.py keeps 1080p share copies at H.264 level 4.2, and mix.py lists sound effects that sit under the music. The opt-in references/kinetic-type.md covers beat-unit authoring, the drop and the kinetic gate, and stays out of explainer and evidence films. The method notes were informed by Scenario's Kinetic Film exploration; OMH wrote its own wording and code.

    • show-reel makes any video from a text idea, not only a product film. Explainers, motion graphics and kinetic type, logo and title intros, 3D scenes, data-viz animation, music visuals and short or vertical cuts now reach it without naming it, in English, Korean and Japanese ("make a 10 second logo intro animation for our podcast", "DNS 설명하는 30초 영상 만들어줘", "DNSを説明する動画を作って"). An idea lane runs on defaults, each recorded as an assumption, keeps every render step and skips the source and rights registers and the three-concept comparison; references/genres.md holds the genre recipes. It renders with what the machine already has, in this order: HyperFrames when its CLI is installed, the show-reel toolkit, a frame renderer already installed, else HOLD with the exact install command and one consent ask. It never installs anything itself, and delivery is an ffprobe-measured MP4 or an honest HOLD, never a storyboard or an HTML file alone. Admission takes closed shapes only: a bare "animation", a logo or animation on an app surface (a splash, a loading screen, an icon), loop phrasing, a topic of two or more words before 설명, and some one-word fault topics (서버다운, サーバーダウン, from a fixed list) do not get in. A recording that already exists (summarize, transcribe, caption, trim) goes to media-input-operator.

    • A delegated child no longer dies on a model its inherited provider does not have. A route written with no provider runs on the session's own, and omh_delegate_route checks that provider can serve the model first. Two gaps let a known-bad route through: a session on custom:<name> was looked up under that spelling while OMH records the bare <name>, and a relay block counted as serving every model even when its providers.<name> block in config.yaml lists the models it has. Both now resolve, so a session on a Claude-only relay sending a gpt-6.1-sol category head is caught. When exactly one provider you configured is known to serve the model, the route pins it instead of refusing (provider_pinned in the result names the inherited and the pinned provider), on set and fallback; otherwise it is refused as before, naming only providers that could serve it.

    • A package installed into Hermes' own Python is now named when the plugin still cannot see it. On a Hermes install its package manager prepares, copying oh-my-hermes into the managed interpreter's site-packages made omh import when that interpreter ran alone, even with Hermes' -I flag, while omh_probe, omh_loop, omh_decision_gate and omh_interact kept reporting the package absent. The cause is Hermes, not the install: when Hermes starts, its package manager takes the interpreter's own site-packages off sys.path and imports from the dependency environment it prepared. An absent package_backend block now carries off_path_install, the directory where that skipped install sits, and its text says why the plugin cannot import it and that there is still no supported way into that environment, instead of telling the person to install the package they already installed. omh_interact's chat fallback now carries the same block beside package_backend_unavailable, so it says which failure it met. docs/INSTALLATION.md explains the case and how to check what the plugin sees (#2415, reported by @trend0522).

    • Native Loop, the decision gate, chat recording and the other package-backed tools now work on a Hermes install its package manager prepares. Hermes loads the OMH plugin with its own interpreter, and on those installs nothing could put the oh-my-hermes package where that interpreter imports from, so omh_loop, omh_decision_gate, omh_team, omh_agent_board, omh_source_trust and omh_interact's session recording were off. omh setup and omh update now record where their own omh package is installed (the installer venv, Homebrew, uv tool, pip, or the npm/Bun cache) in the plugin bundle's manifest, and the bundle imports omh from there when Hermes' interpreter has none of its own. Only omh itself is imported from that directory, only from a wheel install at the bundle's exact version, without writing bytecode into it, and an omh the interpreter already has always wins. Run omh update, then restart Hermes. When the recorded package is missing or another version, the refusal says which and names that same repair, and omh doctor reports the bundle as stale (#2166, #2415).

    • Memory recall works on long pastes, saved memories lead back to the original words, and hermes backup includes OMH memory. A message that pastes a log or document next to a question used to rank memory against the whole paste; the chat-turn recall query now keeps the person's own words (other speakers' Discord channel backfill, host notices and loaded skill bodies are left out, a replied-to message and the same speaker's recent lines are kept) and cuts a long turn to its question, else its last sentence, else its tail, at most 250 characters. CLI recall and coding-handoff packs keep their raw task text, and the prefetch receipt records only the method and lengths. A capture made through omh_memory now records its Hermes session, and omh_memory recall rows carry an origin the model can open with session_search to quote the person's message; it is read-only from Hermes' state.db, and OMH still stores no transcript. The memory provider's backup_paths() hands <omh_home>/memory to hermes backup, and a restored principal.key is tightened back to 0600 on its next read.

    • The OMH pane in Hermes Desktop reads idle when nothing is running and keeps the plan and status bar in agreement. Old runs no longer keep it looking busy. OMH catches and the delegation graph appear alongside the plan, board and active fanout graph, while individual subagent rows stay in Desktop's Agents panel. Calmer colours keep running work distinct from waiting or finished work, and the pane follows Desktop's language when Korean is selected.

    • A dag-ulw node left open for hours says so. omh_dag status, ready and settle give every open node its age (open_for_s, from when it first went out, so re-offering it does not reset it). One open past an hour is listed in stuck, counted on the run's status line (1 open past 1h (Docs body 2h 10m)), and named in say with what to check: no child took it, its child still runs (delegate_task action=list, then settle failed and repair when none is live), or it came back and was never settled. A run whose wave never started or whose child went silent used to look the same at two minutes and at eight hours. The marks are profile settings under plugins.entries.omh.settings: dag_stuck_after_s (default 3600, 0 turns the warning off), dag_stale_after_s (default 600) and dag_max_lost_reoffers (default 3); a value that is not a number is named in the result's limits and the default kept. OMH still starts no timer, so the warning shows only when the parent calls the tool. Proposed by @yqm-307 in #2427.

    • Category routing works under Hermes 0.22.5's delegation backends. Hermes 0.22.5 can pick a child's model through a delegation.backend of model_map options or profile_routes profiles, and OMH wrote its category route into delegation.* regardless, so it fought the backend or reported a route Hermes never ran. OMH now reads the backend first. Under model_map it writes no config: omh_delegate_route set answers option_selected with one of your own option names for the category (no_matching_option when none fits), and omh_dag puts that option on each task it hands out and steps to the next option on a fallback retry. Under profile_routes it answers not_applied and picks no profile for you. The delegation ledger records the route Hermes actually ran beside the one OMH handed out, and the graph and HUD show category@option when the two agree and flag a mismatch when they do not. Under default, and on an older Hermes, nothing changes.

    • A dag-ulw run under hermes chat -q stays inside the one-shot child budget left (#2276, reported by @yqecea). Hermes counts a one-shot run's delegated children per process and refuses a whole delegate_task call that would pass delegation.oneshot_max_children, so a wave larger than what was left was refused whole and its nodes fell back to inline work. omh_dag now counts the top-level children already started in the run, a delegate_task outside the graph included. define refuses a graph larger than the budget left; ready and settle hand out no more tasks than remain and list the rest in held_for_slots; once the budget is spent, the held nodes come back as for_parent with their handoff paths, for the parent to do and settle itself. repair and extend refuse oneshot_budget_exceeded before touching the record. In a one-shot session slots.oneshot shows {limit, spent, remaining}; interactive and gateway sessions are unchanged.

    • The website moved to https://omher.app/. The old rlaope.github.io/oh-my-hermes/ address redirects there, and the package homepages (PyPI, npm, Homebrew), the README in every language and the Installation guide link to it.

    • A redesigned home page that search engines and link previews can read. The home page is rebuilt around Kami: a hero with the Kami logo, the pixel wordmark and the install tabs, a strip of the providers omh model setup covers, a scroll-driven /dag-ulw run, every skill name falling as rain, a Kami you can poke and pull, and the community links. The language switch, every install command, the agent prompt, the shipped model chains and the ulw-* list carry over. Every indexable page now has a canonical URL, Open Graph and Twitter card tags with one 1200×630 share card, and a title naming its topic; the site gains robots.txt, sitemap.xml, a favicon set, a web manifest and a styled 404 page, and the home's JSON-LD names the project, omher included, beside an IndexNow key for the search engines that read it.

  7. 4.0.3

    Expandir

    GitHub Release

    • plan and ralplan name the goal beyond the request, the raw material, and an outside judge. Before splitting tasks a plan states what should happen once the requested result exists and whether it aims there, and adds the access an executor needs to work at that level unattended as its own tasks; it starts from raw logs, traces or data rather than a digest, names the methods a leading practitioner would use, and gives every verification step a judge other than the planner or executor, recording a missing one as an evidence gap; it restates the deliverable as verbs and brings other fields in as practitioner actions, never as voices to imitate. The lines sit after each bar's first four items, so capability rows do not move, and they reach the portable projection.

    • 3D and canvas scenes have a craft and QA bar. A user ran the same voxel pagoda garden prompt through two agents: both consoles were clean, and the OMH build was judged less colorful. Its first frame was flat noon under its own "golden hour" subtitle, and its birds read as floating black cubes, and nothing in the catalog named a 3D scene, so neither the build nor its own visual review checked for either. frontend gains references/rendered-scene-craft.md (the first frame keeps the scene's own promise; a warm-or-cool key light, a sky fill, explicit tone mapping, shadows and fog before more geometry; small animated objects read as what they are; the capture matrix), one body line pointing at it, and the phrases three.js scene, threejs scene, webgl scene, voxel scene and voxel art scene, with scene, voxel, art and three held back to whole phrases. visual-qa captures a scene's booted first frame, its camera presets and each advertised time of day, and records a claim the frame does not show, an unreadable small object, or flat light as a difference.

    • A finished /goal leaves its whole time in the OMH TUI. When a goal shown above the composer turns done, the transcript keeps ✓ goal done · 46m 2s · 7/20 turns · <title>, timed from the goal's own created_at to updated_at, so a goal that ran across many turns is not read off the last turn's ✻ line. It is said once, sends nothing, and leaves the time out when either stamp is missing.

    • The gateway-attach child no longer aborts when Hermes' gateway hangs up first (#2327, reported and first fixed by @skihyeon in #2328). Its omh-attach-stdin thread read the TUI's input through sys.stdin.buffer and so held that buffer's lock while it waited. When the owner's side of the WebSocket ended before the TUI closed its input, interpreter finalization closed stdin on the main thread, waited a second for that lock and aborted the process (Fatal Python error: _enter_buffered_busy, exit -6), on every Python from 3.11 to 3.14. The TUI then reported the Hermes gateway exited (code -6) instead of the child's own code 1. The thread now reads the stdin descriptor directly, which holds no lock, and frames lines as before. The child exits with the code main() computes, and normal shutdown still runs, including the atexit handlers that Hermes' in-process stdio entry registers.

    • GPT-6.1 Terra is onboarded. OpenAI released it on 2026-10-26. The id now has an exact contract read from developers.openai.com: 1.05M context, 128K output, and a none to max effort ladder with medium as the default. That is GPT-6 Luna's shape, so none is sent as is and only minimal is raised to low on record. Before this, OMH passed every rung through unadjudicated. gpt-6.1-terra-pro and dated snapshots project onto it. Cost estimates use the 1/5 list price with 0.05 cached input. Provider checks know that Codex, the OpenAI API, and OpenRouter serve it. The gpt family calibration applies unchanged, and no shipped chain names it. To try it now, use omh model-chains set <category> "gpt-6.1-terra:medium, ...". On the live-model-tools corpus at high, Terra passed the same 18 / 30 instances as GPT-6.1 Sol with about 5% fewer tokens, at half the list price. Hermes through origin/main a382af23f4 clamps its max to xhigh on the Codex route.

    • The OMH TUI shows up in herdr's agent list. herdr lists each pane's agent as idle, working or blocked, but reads Hermes' state from Hermes' own title and screen, which the OMH TUI does not draw, so a pane running omh was missing from the list. Inside a herdr pane (HERDR_ENV=1 and HERDR_PANE_ID), the TUI now reports the state its title shows through herdr pane report-agent (source omh:tui, agent omh): a question or approval waiting on you is blocked, a running turn working, otherwise idle. It reports only on a change, off the render loop, sends only the newest state when herdr is slow, and releases the pane on exit. The Hermes gateway the TUI starts does not get the pane's HERDR_PANE_ID, so herdr's own Hermes plugin cannot claim the pane and lock those reports out; herdr's CLI and skill still work from Hermes' tools, without a current-pane default. Outside herdr nothing changes and nothing is spawned; a missing or unresponsive herdr is ignored.

    • Setup no longer writes a second plugins.enabled when a column-0 comment sits inside plugins: (#2304). Hermes' own config separates the children of plugins: with # ==== banners written at column 0. OMH's config readers and writers ended a top-level node at the first unindented line, comment included, so an enabled: list below such a banner was never seen: the guard passed, setup inserted a second enabled: under plugins:, and Hermes then refused the whole file as a duplicate key, so no plugin selection was read at all. A comment now never ends a node, in every line reader and writer of the config adapter (plugins, skills, memory and display), so setup adds omh to the one existing list and uninstall takes it back out. Setup also reads its composed text back before returning it and refuses, with the file untouched, when it would not hold exactly the earlier plugins.enabled names plus omh.

    • The installers no longer run Python modules from the current directory or PYTHONPATH. install.sh and install.ps1 start venv creation, pip, the command-discovery probes, and the OMH CLI with -I, or with -P -E in direct-Python mode so pip --user still sees the user site; the printed fallback commands carry the same flags. A venv.py, pip.py, or omh.py in the directory the installer runs from is no longer imported.

    • omh uninstall --remove-files / --all refuses to delete an OMH home it cannot prove OMH created. OMH_HOME is removed only when it holds a readable, unlinked oh-my-hermes schema-1 manifest; otherwise it is kept and listed in kept_paths with the reason, and --dry-run reports the same. HOME, HERMES_HOME, their ancestors, and the filesystem root are never removed, and a home configured through a symlink or junction (by --omh-home, the persisted omh_home binding, or OMH_HOME) is kept even with --force. Linked plugin directories are not followed, team manifests authorize single files rather than directories, the command venv is removed only when it carries a venv marker and an OMH entry, and the menu-bar app directory only with a matching OMH LaunchAgent.

    • omh_delegate_route refuses providers you have not configured. set returns an error and writes nothing when the explicit provider, or the one a routing/model-providers.json route resolves to, is not linked through auth.json, config.yaml, .env, or routing/providers.json, is not a configured custom provider, and is not the session provider. fallback skips such candidates, lists them in skipped_unconfigured, and falls back to inherit when none remain.

    • Temp and lock files no longer use predictable names that a planted symlink can redirect. The jev ask ledger writes through mkstemp; the awareness delivery lock opens with O_NOFOLLOW and refuses anything that is not a regular file; the self-update pointer and launcher swaps use random names created exclusively; and browser skill promotion temp names gain a random token.

    • A dag-ulw run that ended leaves the area under the composer on its own. Once a run was over its tree folded into one line, DAG · <name> · done 9/9, which stayed under the info line for as long as the HUD kept the settled graph (15 minutes). The whole finished tree also came back while the next turn ran. The run's end print in the transcript already holds the final graph, and /dag draws the whole of it, so the view under the composer now empties when the turn that ended the run ends. It goes in the frame that prints the end, from the end print's own read, even when the panel's slower read still shows a lane running. A run that later turns move on (lanes still running, a repair or an extend round) keeps its tree between turns and into the next one, as before. On the bottom rows the rows the tree gave up stay blank under the info line until output fills them, so the composer does not move up and no blank band opens between the run's last lines and the composer. The run's finished children leave the agent list with it, so they are not listed in its place once the graph leaves the HUD. Subagent rows outside a dag run are unchanged: a finished turn still shows each agent's verdict until the next prompt.

    • The TUI's working row counts tokens while the turn streams. The ↓N tokens on the working row moved only when Hermes sent a usage report, which it does after an API call ends, so a one-call reply showed nothing until it was over. It also read the wrong number. A fresh session's first turn took its first report as the starting point and showed no tokens at all, and the done line said ✻ 4s with no ↓. Hermes' gateway now starts its counters over at every turn, so every later turn subtracted the last turn's total: nothing showed until the new count passed it, and then a small wrong number (↓48 tokens after 2m 37s). The row now counts up from the reply, the reasoning (when display.show_reasoning sends it) and each tool call's arguments as they arrive, at about four characters a token and one per CJK character, and moves to Hermes' count at each report without ever going back during the turn. A counter that starts over is detected and counted from zero, and the done line carries Hermes' count.

    • Leaving the OMH TUI says which session it was, only the omh way, and Ctrl+C leaves during a turn. Closing a session omh opened now prints an OMH card in place of Hermes' exit summary: the title, the session id, turns, messages, how long omh ran, tokens split and summed over a /branch chain as Hermes does it, the model, and omh --resume <id> plus omh -c "<title>" when Hermes' title lookup would open that same session (it prefers a newer <title> #N) and the title does not start with -. After a /new with nothing said, the card names the newest earlier session of the launch that has messages. Hermes' summary, which named hermes --tui --resume, no longer prints: the OMH TUI exits with a code of OMH's own, which Hermes' launcher prints no summary for, and omh returns 0 to the shell as before. A resumed session gets the card too: the old line came from the state database's end stamps, and Hermes leaves a resumed session's old ended_at in place until a prompt reopens it, so reopening a session and leaving printed no omh line. The OMH TUI now records the open session and its Hermes home in a private file omh names for the launch, and the end stamps are only the fallback for Hermes' own terminal. A session nothing was said in gets one line saying there is nothing to resume. The Ctrl+C that interrupts a turn also arms the exit, so a second press within 2 s leaves while the turn is still interrupting… (it used to be swallowed, with no hint), and the info line says ctrl+c again to exit then too. The info line also leads its right side with the session's title, cut to 24 cells, and drops it first on a narrow terminal.

    • Esc, Ctrl+C and an approval deny stop the turn again on a Hermes whose TUI attaches to the shared gateway. On Hermes bf0203a794 the OMH TUI sent session.interrupt the way the stdio entry takes it, with only the session id. The canonical gateway stops a turn only for the execution generation it runs, so it refused every stop with stale_generation. The TUI said "Hermes did not interrupt the turn: stale_generation", and the turn ran on in the shared gateway, after the TUI exited too. The attach child now sends the stop with the generation it tracks from the session's snapshot, events and prompt receipts, as Hermes' Ink TUI does. Refused as stale, or with no generation known, it reads the session once and sends the stop once more at the generation read, only while that read says a turn runs. The stdio wire is unchanged.

    • The OMH TUI shows this machine's Hermes bots on a read-only board. Ctrl+S, or /bots, opens it full screen. The first page lists the Hermes home the TUI runs under and each named profile beside it, each with its gateway as gateway_state.json records it (state, pid, running agents, every platform's state), its sessions counted by source, how many were active in the last 24 hours, and when the last one was. Enter opens a profile's sessions, newest activity first, and a session's last 40 messages; ← or Esc goes back, and r reads again. It reads again every 5 s off the screen's thread. A gateway reads as running only while the process its record names still runs and is the one that wrote it, and a platform's state left by a stopped or restarted gateway shows as last recorded. It counts active sessions, not open ones: a Slack gateway never ends a chat, so a bot idle for weeks still had dozens open. Each state.db is opened read-only, and immutable when nothing has it open, so not even SQLite's -wal and -shm appear beside it; the only signal is a signal-0 liveness probe, and nothing is sent to the gateway.

    • The OMH TUI switches models the way Hermes' own TUI does. Attached to a shared Hermes gateway, every /model the TUI sent (a picker pick, a typed /model <name>, /model --refresh) was refused as Unsupported_command., because that gateway runs no /model through slash.exec. A pick and a typed /model <name> [--provider <slug>] now go as the config.set model Hermes' own TUI sends, which the gateway child turns into the owner's session.mutate model. A pick is for this session only; a typed one keeps Hermes' own model.persist_switch_by_default rule on the stdio entry, as before; /model --refresh opens the picker over a list Hermes asks the providers for again. When Hermes' configured default model changes while a session runs (the person, a bot or the agent edited the config), the session switches to it at the next turn boundary, before the next prompt goes, and says so in one line, and the info line shows the new model; a model picked with /model stays until the default itself changes. Hermes' refusal codes (unsupported_command, model_resolution_failed, session_busy and the rest) now read as sentences that say what to do instead. A model Hermes asks to confirm (cost, data policy, a large context) is asked about in the prompt slot, as Hermes' own TUI asks it: keeping the session's model is the first choice and the one selected, and only "switch anyway" sends the switch again with Hermes' confirmation.

    • Byte-identical managed plugin files no longer block setup or update (#2300). A managed plugin file whose current bytes match the incoming bundle no longer needs --force merely because its hash differs from an older installed manifest. The normal install rewrites that manifest. A file whose bytes diverge from the incoming bundle is still protected, including by the staging revalidation, and a non-forced setup or update refuses to replace it. Managed skill SKILL.md files follow the same rule: one that already holds the bytes the install would write is not reported as a local modification; support files keep the manifest-only comparison.

    • omh doctor adds independent personal-file and state-write privacy controls (#2300). --no-personal-files never opens or reads the bodies of SOUL.md, memories/MEMORY.md, or memories/USER.md; the soul_missing_or_starter and hermes_memory_staleness advisories are reported as skipped, with the reason, rather than as passing checks. --no-state-write does not create or update runtime/state.json, leaves JSON state_log empty, and skips the runtime_artifacts and workflow_state .doctor-write-test probes. Those two checks instead report a non-writing permission result with write access explicitly not probed. No writes under the OMH home were observed in this mode, but the Hermes loader-registration smoke can still copy the plugin into a temporary directory when a Hermes install is found. Both flags work with human and --json output; --no-personal-files does not disable Hermes config.yaml reads.

    • omh setup --dry-run no longer applies model aliases, and the model step reports a broken Hermes instead of a traceback (#2299). With --model-setup --apply-model-config --model-config-digest D, a dry run ran hermes config set model.aliases.* and wrote the alias while the payload said dry_run: true. It now runs no mutating Hermes command, still checks the digest, and reports apply.status: "dry_run" with the exact commands it would run; the human summary prints each as a "would run" line, and the interactive question asks whether to list them rather than to apply. When Hermes' config.yaml does not exist (a dry run on a fresh home, or --skip-apply), the step reports status: "hermes_config_missing"; when it exists but cannot be read, status: "hermes_config_unreadable" with a next action to fix its permissions. Either way setup exits 3. When a hermes command exits nonzero or times out, it reports status: "hermes_command_failed" and exits 1. All of these used to stop on a HermesCommandError traceback. The setup JSON ok and the recorded last_setup.ok now follow the exit code, so a model step that could not run no longer reads as success.

    • omh setup --interactive without a terminal refuses instead of crashing or taking every default (#2295). With stdin closed (<&-) it stopped on an AttributeError traceback; with stdin from /dev/null or a pipe the wizard read EOF at every prompt and recorded the defaults as the user's choices. It now writes nothing, prints one line naming the same command with --interactive swapped for --no-interactive --json, adds that as next_command in a JSON record when stdout is not a terminal, and exits 1. Piped answers into --interactive therefore stop working; pass the choices as flags instead. A closed stdin no longer crashes bare omh, omh theme, omh model-chains or omh model-chains interview either.

    • The OMH TUI no longer opens a blank band above the running calls. On the bottom rows, a frame that got shorter (an approval or other question answered mid-turn, a picker or menu closing, live thinking folding while details are shown) turned the freed rows into blank rows at the top of the live region. The running calls, the working row and its tip jumped down, leaving up to a third of the screen blank between the last settled call and the running ones until more output filled it. The freed rows now go under the live block and above the working row. The running calls stay right under the transcript they continue, the working row and its tip stay on the composer, and the composer still holds its row. Output arriving later fills the blank rows downward, as printed output would.

    • omh setup records the release that is installed. After the documented stable curl install, a plain omh setup recorded release_channel=preview, the main archive URL, source_ref=main, and no version, although the release wheel was installed, and a SHA-pinned install lost its commit the same way. With no --channel, --version, or --package-url, setup now reads the package's PEP 610 direct_url.json: a release wheel or tag archive records stable and its version, the main archive or a commit (including pip/uv tool install git+...) records preview with that ref. npm, Bun, and Homebrew installs record the package version and no channel. A uv tool install pinned to a commit, a release tag, or the release wheel URL is now told to reinstall at the new pin with uv tool install --force ...; uv tool upgrade answered "Nothing to upgrade" for those. The stable installer's failed-lookup hint no longer suggests OMH_VERSION=1.0.6, and install.ps1 now prints the interpreter it picked and probes it with -I, as install.sh does.

    • Russian ships as a trigger language pack, and a later turn in a user-installed pack's language re-arms the route hint (#2284). src/routing/trigger_packs/ru.json, contributed in #2284, carries 278 phrases across 49 skills, such as сделай ревью кода, найди баг, сделай лендинг and составь план реализации. A Russian word scores on its own only when it is a software word (баг, лендинг, рефакторинг, деплоя); every other word scores only inside its phrase, so a phrase made of everyday words (составь план, сборка упала) asks one question instead of dispatching; a skill credits at most one such Russian phrase per message, so two of them in one sentence, overlapping or apart, ask as one (the Korean pack's phrases still add up, as they were written to); and phrases an everyday sentence contains (починить, по расписанию, в чём проблема, убедись) are not in the pack. A test re-derives the pack's hold-back from its phrases and a reviewed list of software words. Cyrillic and Greek now get the word edges Latin has, so a phrase no longer matches inside a longer word: ревью inside превью, выкати inside выкатил. ultrawork carries no Russian phrase: its body would exceed the 27,300-byte shipped-body ceiling, as with vi. On later turns the awareness gate read only shipped cues and ASCII tokens, so a message written wholly in a user pack's non-Latin script never recomputed its route hint, while the same message on a first turn was hinted. The gate now arms when that user pack earns the message a hint, and not on a pack phrase alone. Shipped packs, ru included, stay outside this gate on purpose: a later turn in Russian does not re-arm the hint through it. The phrases have not been reviewed by a native speaker. ru is an input language only: it is not a --language / OMH_LANG output choice.

    • A Jev mention that is negated no longer counts as consent to ask Jev. The consent gate used to grant consent whenever a message named Jev, so "ask jev? no, review locally" let omh_jev_ask send the request. Consent now requires a Jev mention with no negation or refusal cue anywhere in the person's message (English cues and contractions, a few other Latin-script words, and Korean forms such as 말고, 안 써 and 제외), and a message in a script the gate cannot read for negation is denied. This fails closed: an affirmative ask that carries its own negation, such as "claude 말고 jev로 물어봐", is denied and has to be rephrased.

    • The memory store and the principal key refuse symlinked or out-of-store targets. Every directory from <omh_home>/memory down to a write target is checked: a symlink or Windows junction, a non-directory, or a path that resolves outside the store root is refused, and a target file that is a link or not a regular file is refused before writing. The principal key is created with O_EXCL|O_NOFOLLOW, read with O_NOFOLLOW, and must be a 32-byte regular file owned by the current user. Block, dreaming, reminder, session, receipt, journal and counter-lock writers go through the same guard. The checks run before each write rather than through a directory fd, so a swap between check and write is still possible, and the Windows junction and no-O_NOFOLLOW branches have not been run on Windows.

    • omh uninstall --remove-files / --all removes only the entries OMH created and keeps the rest of the OMH home. Once the home is proven OMH-owned, uninstall removes the top-level entries OMH writes (the OmhPaths locations plus a fixed list of other OMH writers, their lock and temp sidecars, and OMH's * .gitignore) and lists every other entry in kept_paths with the reason "not created by OMH". The home itself is removed only when nothing else is left, and --dry-run reports the same split. An OMH-created directory such as runtime/ is still removed with everything inside it.

    • The Jev ask ledger is written through a pinned directory fd on POSIX. The ledger directory is opened once with O_DIRECTORY|O_NOFOLLOW, checked to be a directory owned by the current user, and every later step (lock, read, temp create, replace, cleanup) runs relative to that fd, so swapping the parent directory for a symlink mid-write no longer redirects the ledger, and a symlinked ledger file is neither read nor replaced. Windows keeps the path-based write, where a parent swapped between the check and the write is still followed.

    • Replies read like a person who did the work wrote them. OMH now ships a human-voice pass adapted from blader/humanizer (MIT, v3.1.0): the always-on reply rule adds one sentence (result first, no chatbot openers or closers, staged contrasts, or decoration), the common rail points every skill at the new on-demand omh-routing/references/human-voice.md before prose a person reads, and that reference covers the strongest English tells plus Korean habits (번역투 nominalisation, register drift in a banmal persona). The persona still owns the voice; the pass edits content and structure only. omh quality-evidence reply-lint reports the high-precision shapes as a new ai_writing_tell finding, skipping code and quoted lines. The awareness primer budgets move 1440 -> 1560 and 3520 -> 3630 for the one sentence.

  8. 4.0.2

    Expandir

    GitHub Release

    v4.0.2 · Kami

    Oh My Hermes 4.0.2 — Kami

    Kami makes dag-ulw the default for plain multi-part jobs, while it reads the remarks after a request more strictly. It also rebuilds the OMH TUI launch and transcript, keeps each conversation's delegation route its own, onboards Kimi K3.1 Pro into every Kimi K3 slot, and adds new science, job-search and show-reel skills.

    dag-ulw is the default for a job over many targets

    A plain request for a change over many targets now runs as a dag-ulw graph, with no dag ulw needed (update every service to node 22, 열린 PR들 전부 리베이스해줘). The router starts one only on positive evidence: a change verb from a closed allow list, on an object that names independent units, or two named parts asked for in parallel. Work over many targets without that evidence (review all open PRs) is at most offered, as one question. A request that reads, reports, stops, holds, is conditional, is in the past, or is scoped to one PR does not start a run. Measured on the frozen acceptance sets; prompts outside them are not measured. (#2232) Hermes' route hint never starts a dag-ulw run on a split the router would only ask about, and a split that names its work (결제, 주문, 배송 모듈을 병렬로 나눠서 리팩터링해줘) dispatches on every surface. (#2243)

    A remark after the request is read by its kind, more widely. A retraction (…, calling the whole thing off), a stated precondition (…, need to confirm it's safe on my side) or a contrast after a wish (…, fingers crossed but let's not) holds the run. An aside, a request on someone's behalf, a toast or a back-reference keeps it. (#2265)

    An explicit /dag-ulw <request> runs as a graph. After any explicit skill invocation the skill runs, unless the words after it stop it, ask how its run stands, or ask what it is. (#2221, #2239)

    The OMH TUI: launch looks, a folding dag tree, and attach

    • Launch. A fresh launch draws the logo, the banner and a launch card in one of four looks: spectrum (the default), violet, gold or graphite. The card shows the model and effort, the sessions, the first keys, and the skills Hermes has. Bare /theme opens a picker that previews each look in place. (#2256) The card draws its skills from the last launch until Hermes names them, so they show in the first frame instead of loading…. (#2275) Scrolling up stops at the top of the session, not in the shell history from before omh. (#2219)
    • Transcript. Each block hangs on a ┃ rail. A call's state shows in its glyph (running blue ○, done green ●, failed red ✗), and its name is bold in its tool family's hue. (#2223, #2225) Skin inks reach 4.5:1 contrast on mid-grey backgrounds. (#2257) NO_COLOR turns off colour in the whole TUI. (#2230, #2233, #2234)
    • dag runs. The agent view under the composer is a folding tree. Each node shows its state, model, status and tokens, then two rows with what its child is doing. A repair hangs under the node it repairs. (#2224, #2228) A run over 12 lanes prints a condensed graph to the transcript, and /dag still shows the whole graph. (#2250) Typing dag-ulw in the composer draws a moving purple aura. (#2252)
    • Commands. Rewind (Esc Esc or /rewind) followed by a skill command replaces the picked prompt. (#2222) A skill command typed once shows as one input row. (#2217) The TUI's own commands work in any case. (#2226)
    • Attach. Bare omh opens the OMH TUI on current Hermes whether or not a gateway is already running: it attaches to the profile's shared gateway where Hermes has the attach API, and older Hermes keeps the stdio entry. A reason the OMH TUI could not start now stays readable after the Hermes TUI it falls back to closes. (#2249, #2259)

    A delegation route belongs to its conversation

    A route set with omh_delegate_route now belongs to the conversation that set it. Before Hermes reads delegation.* for a delegate_task, OMH puts that conversation's route in place, so another session's set no longer decides where this child runs. A child that spawns workers passes on its own route. A conversation that set no route gets the profile's baseline. When OMH cannot put the right route in place, the call is blocked with a typed status and no child starts. (#2254) A clear or an exhausted chain stays gone, and a newer route stays newer after another session's set, including when a write failed partway. (#2258, #2264, #2268, #2270, #2273)

    Kimi K3.1 Pro in every Kimi K3 slot

    Kimi K3.1 Pro and K3.1 Pro Highspeed are onboarded, with exact contracts and list prices. A medium or xhigh request maps to an effort the API accepts. (#2266) Every shipped chain slot that named Kimi K3 now names Kimi K3.1 Pro at the same effort. K3 is retired from the chains, but it stays recognized and priced, so an override that names it still resolves. This is an editorial change and has not been measured. (#2267)

    New skills

    • show-reel: a multilingual show-reel and launch-film workflow with an optional motion and audio toolkit. The engine matches upstream revfactory/showreel (MIT). Contributed by @frirenai. (#2212, carries #2195; #2220)
    • geospatial-analysis, biomedical-evidence and molecular-docking: executable, opt-in domain skills. Their helpers ship as inert files. Installing a skill neither runs a helper nor installs scientific dependencies. Contributed by @frirenai. (#2235, #2241, #2245, #2246)
    • ai-job-search: an evidence-bound job search. An application counts as submitted only on confirmed evidence. Contributed by @frirenai. (#2253, #2261)

    Also in this release

    • research, source-finder and web-research share a bounded multi-source search procedure and an optional omh_search helper. The PR calls itself a draft candidate; search quality is not yet accepted. (#2255)
    • ulw-research records inspectable evidence, and python -m omh.workflows.research_dossier validates a dossier. (#2227)
    • Code review keeps the spec axis, coverage limits and verification gaps apart. (#2231) Briefings end on their result, not on unrelated non-actions. (#2238) A DAG run's goal reaches every worker, and adaptations keep their quality goals. (#2240)
    • Hermes-native delegate_task children get their own model calibration. (#2215)
    • omh setup's provider step ends on one Enter. (#2247) Setup reads a config key that holds a dot, such as z.ai-coding-plan, as one key. (#2248)
    • Windows staged self-updates pass an absolute junction target, which PowerShell 7 requires. (#2218)
    • Every fromisoformat that reads outside text refuses hour 24 on Python 3.14 too. (#2263)
    • The four READMEs and the GitHub Pages site are rebuilt as a lean landing page, with a spectrum banner, the app icon and a dag-ulw clip. (#2269, #2271, #2272, #2274)
    • omh_agent_board states the fields a prepare needs, and a prepare that omits one is refused as invalid_prepare_input naming the missing fields. (#2278)
    • The Vietnamese trigger pack is reworded after a native review, and the routing effect of each change is measured. Contributed by @PhucNguyen-Dev. (#2262)
    • Install, update and uninstall keep shared state until a profile's teardown succeeds. Detaching one profile no longer deletes stores, records or the command another Hermes home still uses, and a failed child detach keeps them and reports a partial, nonzero result. An unreadable profile config is reported by name while healthy profiles go on. Update prints one final JSON result after every profile, with a nonzero exit when a required refresh fails. Contributed by @frirenai. (#2280)
    • omh setup is shorter on first use and names omh as the next step. Fresh noninteractive output falls from 36 to 24 lines, the optional coding, provider and model choices are explained more briefly in all four languages, and with no terminal or no Hermes it prints an 11-line first-use guide. Both launch paths now get the resolved --hermes-home instead of inheriting another home. Contributed by @frirenai. (#2282)
    • The TUI and omh hud --graph no longer draw an old fanout as live work. A settled fanout leaves after 15 minutes, and one with a prepared or running unit stays for 24 hours. The record on disk is left alone. (#2283)

    The full entry list is the 4.0.2 section of CHANGELOG.md. #2262, #2280, #2282 and #2283 have no entry in the tagged section; they are listed above.

    Upgrade

    omh update
    

    Restart or reload Hermes after updating. Launch with omh.

    Release record

    • Version and channel: 4.0.2, stable (.release-channel = stable; GitHub release not a prerelease, npm latest, Homebrew tap).
    • Hermes skill tap / install wording: unchanged from 4.0.1: hermes skills tap add rlaope/oh-my-hermes; default install curl -fsSL https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.sh | sh.
    • Update path tested: not run for this cut.
    • Workflow docs generation: all 11 docs ... --check gates passed locally on both release prep heads, b670d6d98 and 553fa6961 (PR #2281). PR CI was 11/11 green, and main CI run 40126304216 on the tagged commit cfb962514 passed 12/12 jobs.
    • Harness catalog validation: no separate harness validate step ran for this cut; the catalog is covered by the full suite in CI.
    • Source-checkout console script smoke: Installed OMH command smoke passed in main CI on cfb962514 (Python 3.11 and 3.13).
    • Runtime validation: Runtime artifact smoke passed in main CI on cfb962514 (Python 3.11 and 3.13).
    • Workflow learning review queue: no workflow-learning contract changed in this release.
    • Capability probe: Capability probe smoke passed in main CI on cfb962514.
    • Install script smoke: plan-only (Dry-run install smoke and Stable-channel dry-run smoke) in main CI on cfb962514. The Windows installer end-to-end step ran live on test-windows (0) and passed, under PowerShell 7 and Windows PowerShell 5.1.
    • Hermes CLI install smoke: plan-only (Hermes release smoke plan) in main CI on cfb962514.
    • Plugin bundle: plugin.yaml version 4.0.2, requires_hermes unchanged (>=0.21.1); provides_tools adds omh_search (#2255).
    • GitHub Pages: site rebuilt in #2271 and #2274, hero badge to v4.0.2. The Pages run 40126304242 on cfb962514 passed, and the live site at https://rlaope.github.io/oh-my-hermes/ reads v4.0.2.
    • Tag and publish: annotated tag v4.0.2 ("v4.0.2 Kami") on cfb962514, the head of main, published by Distribution Release run 40154073716. The npm environment was approved by sionic-khope, the account that did not push the tag.
    • First publish attempt: the tag first pointed at b02148c91, the merge commit of the release prep PR #2281. Run 40106573742 failed at "Validate tag and distribution contracts", because #2282, #2283 and #2280 had merged after it and the tag was no longer the head of main. With the owner's approval the tag moved to cfb962514, so those three PRs ship in 4.0.2.
    • npm / Homebrew: run 40154073716 succeeded at 2026-10-25 02:28 UTC. It published oh-my-hermes@4.0.2 to npm with tag latest and a signed provenance statement, and the rlaope/homebrew-tap formula moved to 4.0.2 (c44d40230).
    • Known manual Hermes checks not automated: attaching the OMH TUI to a running Hermes gateway, the four launch looks and the folding dag tree in a real terminal, and a live dag-ulw run were not re-run for this cut.
    • Machine sync after the cut: none has synced yet.
    • Claims resting on wrapper evidence: the launch card, the dag tree under the composer and the transcript's graph prints are OMH-rendered from Hermes' recorded state; they are not Hermes-native rendering.

    Merged pull requests since v4.0.1

    #2212, #2215, #2217, #2218, #2219, #2220, #2221, #2222, #2223, #2224, #2225, #2226, #2227, #2228, #2230, #2231, #2232, #2233, #2234, #2235, #2236, #2238, #2239, #2240, #2241, #2243, #2245, #2246, #2247, #2248, #2249, #2250, #2252, #2253, #2254, #2255, #2256, #2257, #2258, #2259, #2261, #2262, #2263, #2264, #2265, #2266, #2267, #2268, #2269, #2270, #2271, #2272, #2273, #2274, #2275, #2278, #2280, #2282, #2283, and the release prep PR #2281. (#2212 carries #2195; #2233, #2234, #2236, #2246 and #2259 are test-only.)

    The complete version-bound changelog follows below.

    Full changelog

    • omh_agent_board states the fields a prepare needs, and names the one that is missing (#2277). The schema required only action and request_id, so a prepare built from it left out coordination, operation, board, profile and arguments and came back as invalid_request_or_board_store, the same reason an unreadable board store gives. The schema now carries a top-level oneOf with the two closed shapes, and the tool description lists the prepare fields, because Hermes removes a top-level combinator from every tool schema before any provider sees it. A prepare that still omits one is refused as invalid_prepare_input with fields naming the missing names and no values; a missing board keeps board_required, a host without the board engine still answers omh_agent_board_core_unavailable first, and status is unchanged. docs/AGENT-BOARD.md lists the refusals that come before an engine record.

    • Kimi K3.1 Pro replaces Kimi K3 in the shipped chains. Every slot K3 held now names kimi-k3.1-pro at the same effort: deep, architect, unspecified-high, quick, writing, visual-engineering, artistry, capable, the main suggestion, and the X-platform affinity. Both site pages, the generated install table, and the model-setup skill say so; the four lean READMEs link to the canonical table. K3 is retired under the superseded-generation rule. It stays recognized and priced, so omh model-chains set can still name it. The Highspeed tier joins no chain. The quick chain therefore no longer offers a Kimi Ultrafast swap in omh model-chains interview, because no K3.1 Ultrafast tier is priced. This is an editorial change (owner decision) and has not been measured yet.

    • A remark after a request is read by its kind, more widely (#2243 follow-up). A retraction is read by its verb stem (…, side note, I'm cancelling that, …, fyi backing out of this, …, calling the whole thing off), a stated precondition holds (…, need to confirm it's safe on my side), and a contrast after a wish turns to what follows (…, fingers crossed but let's not): each holds where it started a graph run before. Who asked needs someone other than the speaker as its subject, or a passive (…, approved by the CTO). More remarks keep the request: an aside marker of any kind (…, also, …, …, off topic, …), a request on someone's behalf or for a purpose (…, the PM's request, …, at the request of QA, …, on behalf of legal, …, it's for the audit), a toast (…, here's to a clean run, …, may the build be green), a back-reference (…, consistent with the RFC, …, like we did for billing), …, no deadline and …, bear with me. A toast, a wish, a back-reference or an attribution keeps only when what it carries is clean: a negation of doing the work holds wherever it is said (…, here's to not doing it, …, here's to letting sleeping services lie, …, the team lead asked us to leave it, …, cheers don't bother), while a negated outcome keeps (…, hope nothing breaks, …, btw it won't touch prod, …, fwiw no exceptions, …, hope we don't break imports: a don't that governs an outcome verb asks for the work all the same). A flat refusal holds (…, btw, no). What a third party said, wanted or told us is read by the kinds a bare remark is: a clean approval keeps, with intensifiers, punctuation and a when or what for around it (…, legal said a big yes, …, the lead said absolutely, go for it, …, the client wants it by friday), and so do thanks, a greeting, a sign-off and a wish (…, the lead said thanks a ton, …, the PM says hi); after an aside marker a plain statement keeps too (…, fwiw QA says the suite is faster now). A refusal, a hedge or anything unknown holds (…, QA says hard no, …, fyi the lead said meh), and so does sarcasm: a mocking idiom or an approval dressed in filler or said twice (…, the lead said yeah right, …, ops said sure, in your dreams, …, the lead said oh sure, …, QA said yeah yeah). A negator before the reporting verb holds whatever was said (…, nobody said yes, …, ops never said go, …, nobody signed off), and so does one before an attribution verb (…, no one on the team gave the go-ahead). A subordinating condition opener holds where it opens a clause (…, assuming QA says it's fine, …; as soon as the lead says go); inside a clause it constrains the work and keeps, and an adjective never counts (rename every module as long as the import paths stay the same, … using the provided helper). A reported wish keeps only as a wish alone or that the work goes well (…, the lead said hope it goes well; …, the PM said good luck lol holds), a doubled assent is no thanks (…, QA said great, great holds), and under an aside a reported negated state, reluctance or blocking status holds (…, btw legal says the contract isn't signed, …, btw ops says staging isn't up, …, btw ops says prod is frozen), while a negated problem keeps (…, fwiw QA says it isn't a problem). A status word that is also good news holds only of a person away or a system down (…, fyi the lead said she's out today, …, btw ops says the cluster is down); of a work item it keeps (…, btw ops says the fix is out, …, fyi ops said the error rate is down). Thanks for a negation mocks wherever it is said (…, thanks for nothing holds). A relative what is no reported speech (…, hope it's what the PM wants keeps as a wish). Korean reported speech is unchanged and reads as on main.

    • The launch card shows skills from its first frame. The skills section said loading… for about 1.4 s on a warm launch and 1.8 s on a cold one. Hermes names its skills only in the session.info it sends after building the session's agent: plugin discovery, AIAgent setup and its skill scan. The OMH TUI's own share is under 0.1 s. The card now keeps the skills it drew in <omh_home>/tui/launch-skills.json, written atomically and best-effort, with the Hermes home and profile they were drawn for. The next launch on that home and profile draws them at once, with a dim last-known mark on the title row. When the session's own skills arrive they replace the snapshot whole, so a skill removed since then leaves the screen. The card keeps its height through the swap. The snapshot holds the names as drawn, so the first frame imports no skill catalog. A first launch, a launch on another Hermes home or profile, or a snapshot that is unreadable or from another schema still says loading…. Hermes is unchanged.

    • Kimi K3.1 Pro and K3.1 Pro Highspeed are onboarded. Moonshot released them on 2026-10-23. Both ids now have exact contracts read from platform.kimi.ai: 1M context, thinking always on, and a low / high / max effort ladder with max as the default. A request for any other rung is raised on record (medium to high, xhigh to max, and off/minimal to low), the same way Hermes' Kimi provider sends it. Before this, OMH passed medium through and called it unadjudicated. Cost estimates use the list prices: 2/10 per MTok, and 4/20 for Highspeed, which keeps the 0.20 cached-input rate. Provider checks know which catalogs serve each id. The kimi calibration applies unchanged. K3.1 Pro takes every slot Kimi K3 held (the shipped-chains entry above), and Highspeed joins no chain. The K3-vs-K3.1 benchmark pair is the follow-up and needs a served route.

    • A newer delegation route stays newer after another session's set (#2251 follow-up). When an omh_delegate_route set wrote its route but failed to write the conversation's entry, the conversation's older entry was ignored only while the profile-wide restore record still named that writer. Once another session's set replaced the record, the older route came back at the conversation's next delegate_task. Each set now first records the route it wrote in ~/.omh/routing/conversation-writes.json, one stamp per writer (the session, or task:<id> for a host that names no session). Another writer's set does not replace it. The stamp stands in for the older entry, so the conversation's next dispatch runs on the route it set last, and so does a delegate child inheriting from it, even when the older entry was from an earlier turn or from the session a compaction continued. A stamp never brings back a route that a clear, a chain exhaustion or a turn end took out, even when that removal's own write failed: each marks the conversation's stamps released first. A reapply, a repeated set of the same route and a child inheriting its parent's route are unchanged. Stamps are kept for six hours, at most 64. A missing or corrupt file reads as "no newer write", which is the behaviour before.

    • A cleared or exhausted delegation route stays gone (#2251 follow-up; #2258, #2264, #2268). clear and chain exhaustion now release the conversation's own route inside the same lock as the restore, after it. Before, the release took a second lock. When that lock timed out, the cleared route, or the model that had just failed, came back at the next dispatch of the same task. A turn-end restore whose write is refused now keeps the route recorded, because the keys still hold it. A session-start restore that cannot release a gone writer's route reports a degradation instead of leaving it to the six-hour bound. A delegate_task that runs on another session's route because its own state.db row cannot be read now says so, in the hook result and in Hermes' log. A corrupt _delegate_from cycle can no longer make a session inherit its own stale route. omh_delegate_route action=status reports the conversation's own route beside the profile-wide keys. Before it releases, a clear or a chain exhaustion stamps the entries it takes out in ~/.omh/routing/conversation-take-outs.json, inside the restore's lock and whatever the restore decided, so a failed release cannot bring the cleared route back at the same task's next dispatch. When the 2 s dispatch-lock wait times out and the unlocked read finds nothing to change, the delegate_task goes ahead instead of being refused.

    • A delegation route belongs to the conversation that set it (#2251). The three delegation.* keys are shared by every session on a profile. When session A ran omh_delegate_route set and session B set its own route before A's delegate_task, A's child ran on B's model and nothing reported it. B's turn-end restore could also take the route away from A. Each route is now also recorded per conversation in ~/.omh/routing/conversation-routes.json. Just before Hermes reads the keys for a delegate_task, pre_tool_call puts the dispatching conversation's own route back. A delegate child's workers get the route of the conversation the child was delegated from, as they did before. A conversation that set no route gets the profile's default, not another session's live route. If OMH cannot do that, the call is refused with a typed status naming both sessions, and no child starts. A dispatch on a session's own route still writes nothing. A small window remains, because Hermes reads the keys after the hook returns. A wait for a person to approve the call is part of that window.

    • TUI text stays readable on a mid-grey background with every skin. A colour that needed lifting for contrast stopped at the skin's own light ink. Crimson's is #FFEDE6, which reaches only 4.36:1 on #707070, so its labels, headings, warnings and errors read below the 4.5:1 floor; between #6E6E6E and #747474 every skin did. The lift now goes on to white or black where that ink falls short, so every text ink meets 4.5:1 and every glyph meets 3:1 on any grey.

    • The launch has a look of its own, and a card. The owner asked for new logo colours and for more than the logo, like the skills box in Hermes' own session panel. A fresh session's launch now draws the logo, the banner's title and a launch card in one of four launch looks: spectrum (the default: the logo swept through the five tool-kind hues, a box per section), violet, gold and graphite. The card names the model and effort, the new session, the last session with its /resume id, the first keys, and the skills Hermes has. OMH's skills are grouped by capability family, with the ulw lanes on their own row, and Hermes' own skills by category. It fills in as session.info and session.list answer, sits under the banner until the first prompt, then stays in scrollback above it. /theme alone opens a picker that previews each look as you move (on the logo and the card at launch, in a compact preview later); Enter keeps it under the OMH home, and /theme <look> keeps one at once. The logo is coloured cell by cell, with a darker drop shadow, at 3:1 or more on dark and light backgrounds. The skin file Hermes' banner reads is unchanged.

    • A dag-ulw in the composer glows. Typing dag-ulw, dag ulw or /dag-ulw in the TUI's composer runs the word's letters through violet, magenta, pink and lavender with a brighter band sweeping through them, and draws the composer's rules heavy in the same gradient with a band travelling along them, the way Claude Code marks ultrathink. It adds no rows and leaves the caret and the cells alone, ticks only while the word is there on a truecolor terminal, stands still at 256 or 16 colours, and is bold without colour with NO_COLOR. The word takes the glow in place of its /name colour.

    • The dag-ulw TUI reads at a glance on a big run. The owner's GIF review and an audit at 100×30, 150×42 and 200×50 found the transcript flooded by 147–219 rows per print under the 48-node cap, and a tree whose short titles left the figures ragged at the right edge. A print now leaves the Dependencies list and the Node details to /dag. A run of up to 12 lanes still prints every box and arrow; a bigger one prints its header, a row per layer, the boxes the moment changed and … N more nodes · /dag shows the whole graph, and its end is the whole run's tally and its failed or stuck nodes. A layer the node cap emptied keeps its row. Tree titles take the task line of a node's prompt after a short label, cut at a word past 52 cells, and the figures stand in fixed columns. Boxes grow to 48 cells and cut a long title in its middle. Counts are the run's lanes and the edges between them, with thousands separators and B/T token tiers, a capped layer reads 28 of 640 nodes, and the agents chip counts the running lanes. The launch fills the screen with the composer on the bottom rows, the composer keeps its rules and a ↑↓ browsing · esc to type hint while the tree has the keys, and a run that settles or shrinks mid-turn keeps its rows until the turn ends instead of leaving a blank band above the composer.

    • omh setup's provider step ends on one Enter. After the provider list, a second prompt, Add another provider by name (Enter to skip), always followed and waited for its own Enter. The step looked finished while it was still blocked. The list now has an Other provider -- type its name row just above skip. The name prompt appears only when that row is ticked; it keeps the duplicate and invalid-id checks and the kind menu, and an empty line still ends it. The row is never pre-ticked, it is refused beside skip like any other row, and it is never written to ~/.omh/routing/providers.json. In the typed menu, typing only that row's number adds it to the ticked rows; it does not replace them, so found providers are not recorded as excluded. In the arrow-key menus (the provider list and the single-choice menus, such as the provider-kind menu), digits now build a row number: 1 then 7 is row 17. Before, the list toggled row 1 and then row 7, which silently excluded a found provider. The kind menu returned row 1 and left the 7 to be read by the next "Provider name" prompt. While a digit could still be extended, a note under the menu shows it (Row 1_ -- type a second digit, or Space/Enter ...). Menus of up to nine rows behave as before.

    • /omh-doctor status runs doctor (#2221 L4). After a slash skill, a status question shows OMH's status card. The exception is a skill whose own work is that status, and doctor is now one of them. The card counts setup gaps but not install health: on an install whose managed skill file had drifted, it read "OMH setup gaps: 0" while doctor reported local_modifications and guidance_projection. A stop after /omh-doctor is still a stop, a question about the skill is still answered, and the bare /omh-doctor runs as before. Every skill in that status-is-the-work set now has a routing case and a test that turns red if it is dropped from the set.

    • Preserve adaptation outcomes and native DAG goal context. Planning, implementation, DAG and verification guidance share one on-demand procedure for reuse decisions, user-agreed losses, early/final output comparisons and separate technical/output-quality verdicts. Native DAG tasks now carry the saved run goal through reload, retry, repair and extension; verifiers are told to check the whole goal, not only local VERIFY. Contract tests establish instruction delivery, not live-agent compliance or artistic quality.

    • The route hint no longer starts dag-ulw on a split the router only asks about (#2232 follow-ups). A split into parallel parts ("fan out across the services …", "split … into parallel lanes", "병렬로 나눠서 …") started a dag-ulw run from the hint on 36 survey rows while the router only clarified. Now named work asked to split runs as a graph on both: a work verb after the split (fan out across the services and migrate each one, …, then integrate, 모듈을 병렬로 나눠서 리팩터링해줘) or a work noun as what splits (the migration is big, split it into parallel lanes), in an imperative, a polite request, or after tell me when done:. A split of text to translate, a proposal or a plan said (how about we …, my plan is to …, maybe …), a question how to split, a report, breadth alone, or one named object (이 함수 병렬로 나눠서 구현해줘, README 병렬로 나눠서 …, fan out the image resize into parallel workers …, the checkout page: split it into parallel parts …) is asked about on the router and offered in the hint, never started; a count or 들 after a demonstrative is several objects (이 세 서비스 …). More Korean request forms count (닫아, 고쳐, 올려라, 주라, 고고싱, 놔라, 해주십쇼, 요망, …해 주시면 감사하겠습니다), breadth words 쭉 and 하나도 빠짐없이, … fix them all, and rewrite and reimplement are changes. What follows a request in a later clause keeps it only on positive evidence, in Korean and English: a keeper by shape (an urgency or a deadline, a start cue, thanks or politeness, with an intensifier or a request ending: …해줘, 최대한 빨리, …해줘, 금요일까지, …해줘, 지금 바로, …해줘, 부탁할게, …, by friday please, …, no rush, a per-unit spec such as …, one commit per package or …, a separate PR per repo), or a follow-on after a sequencing connective that asks for more work (…해줘, 그리고 README도 갱신해줘, …해줘, 그 다음 테스트 돌려줘, 모든 패키지 업데이트해줘, 패키지마다 changelog도 써줘, …, then run the tests), with no hold or joke in any clause after the request. A final …, ok? on any request and a report on the result, by its opening (tell, ping, notify, message, show or send me, let me know, report back, then which, any, the ones, each, when, once or a result noun: after a quantifier the clause turns on its subject (the targets with any finite verb report, … which ones can't be upgraded; a person with a modal or an intention asks, … which ones you'd skip; a bare infinitive asks), when/once needs a completion that ends the tail, and deliberation is no report: … and report any that fail, … and let me know which ones break, …, and notify me when it's done; not …, and send me any concerns, …, then tell me which ones you'd recommend), are read off before the request is; any other let me … and a tail that asks for anything before the work (…, then tell me why we shouldn't, …, then tell me the risks first) hold it. Every hard exclude reads the whole text, and a deferral is one (… later, … in a bit, … tomorrow, … tonight, … next week, PR들 전부 나중에 …; … by tonight and …, 내일까지 are deadlines and run, and a time inside a past clause, … we deployed on monday, is none either, while a time ahead of now is, … we deployed next week). A colon list after a breadth head is one request when every item opens on a verb of work (for each open PR: rebase, squash, merge); an item that opens on a status or carries a hedge ends it (…: done, done, conflict, …, too risky though). After a Korean request anything else keeps nothing, and the message is read whole, as on main (…해줘, 아 잠깐만, …해줘, 해도 되나, …해줘, 확인해줘). After an English request a remark keeps it only when it is of a keep kind, in any words and with any intensifier: gratitude (…, thanks a bunch, …, thx so much!!, …, much obliged), a sign-off (…, have a good one), a wish (…, fingers crossed, …, wish me luck), a back-reference (…, as discussed, …, same as last time), who asked (…, my manager asked for it, …, approved by the CTO), an aside marker of any kind and a harmless statement (…, fwiw the tests are flaky, …, also, the new keyboard is lovely, …, off topic, the sofa arrived), or no pressure (…, no worries, …, take your time, …, whenever works). An unknown remark holds, so the request is offered, not started, and a refusal class holds even inside a keep kind: a new request (…, run the tests), a hold, a joke or a retraction (…, never mind, …, forget I said that), a status, something asked before the work, a warning (…, fwiw last time this broke CI), a hedge or thinking aloud (…, not sure though, …, spitballing), deliberation or the decision handed back (…, what do you think, …, your call), a person's modal (…, we'll see), a question, a negation (…, not yet), a deferral or a condition. A past inside an embedded question (어떻게 나눴는지 설명해줘), a report after a -고 work verb (… 고치고 뭘 고쳤는지 보고해줘) and a change aimed so a thing does not happen a few words off (… 안 되게끔 수정해줘) count too. 일정 간격으로 is a schedule.

    • NO_COLOR turns off colour in the whole OMH TUI (#2223 review item 17). With NO_COLOR set to any non-empty value (no-color.org), only the motion and the delegation graph used to honour it. The transcript, composer, agent view, pager and pickers still painted truecolour: 326 colour codes in the redesign's scene A and 486 in scene B. Now the app's theme drops every colour, and so do the skin's logo, a box graph drawn in the transcript, and the rows on an approval's, a notice's or a command's band. Both scenes write none. Bold, dim, italic, underline and reverse stay, and the bands themselves are not drawn. Two rows that only a background marked are reversed: a menu's or list's highlighted row, which is also bold, and the pager's and /dag's footer bar. Every state keeps the glyph it already carries. With NO_COLOR unset or empty nothing changes: both scenes are byte-identical to before.

    • A plain request for a change over many targets can run as a dag-ulw graph. "rlaope로 올린 hermes upstream pr들 전부 확인해서 닫을건 닫고 리뷰반영할건 리뷰반영하고 rebase할건 리베이스하고" went to web-research on the router and ulw-work in the route hint; it now dispatches dag-ulw. The router dispatches dag-ulw only on positive evidence: a change verb from a closed allow list (fix, update, rebase, close, migrate, 고쳐, 올려, 리베이스, 반영, …) in request form, on a breadth object that names independent units (열린 PR들 전부 리베이스해줘, 열린 PR들 전부 리베이스 해줘, 모든 패키지 버전 올려놔, fix all the failing tests, update every service to node 22), or two named parts asked for in parallel. rebase, merge, bump and fix count inside a Korean request (rebase까지 ㄱㄱ), and a push counts only after a change (리뷰 코멘트 반영하고 force push ㄱㄱ); 모든 브랜치 push 해줘 is one command. More request forms count: 해줘요, 해주세용, 해 놔줘, 시켜줘, 올리자, 리베이스 고, 부탁, upgrade해줘, 몽땅, 최신화, and for each of my PRs: …. A count is many only for PRs, repos or branches with a breadth word after it (PR 3개 전부 리베이스해줘, PR 5개 다 머지해줘); PR 3개 리베이스해줘, 파일 2개 수정해줘 and 테스트 3개 다 고쳐줘 stay small single tasks, and PR 3, 3번 PR and #3 stay one PR. A past in a modifier names the units (내가 올렸던 PR들 전부 리베이스해줘, 충돌났었던 브랜치들), and a report after an English change is its tail (fix all the failing tests and tell me which ones were flaky). A stop or a hold, a condition, a past or a completion as the predicate, a recurring schedule in any of its shapes (매일 아침, any noun with 마다 but a unit of work such as 반나절마다 or 배포마다, 매 분기, 2시간에 한 번씩, 하루 두 번, 주 2회, 3일 간격, 하루 걸러, 격주로, 자정에, every hour, each morning, twice a day, on every push, cron으로, 스케줄 걸어서; 서비스마다 and PR마다 still mean each of them), one pull request (its number, PR number 2240, 2240번 풀리퀘, pull 951, my latest pull request, PR 하나, its link, or one singled out by a modifier: 어제 내가 연 PR 테스트 전부), one branch (feature/x, 현재 브랜치, this branch), one fault with one fix in any words (안 돌아가, 안 먹어, 돌지 않아, 페이지가 하얗게 나와 고쳐줘; 안 되게 고쳐줘 is a change), work until a condition, and a request that opens with a sigil or names a skill, ulw or a retired engine, also with a Korean particle (loop로, github-event-ops에서, 루프로) or as the instrument anywhere (… git-workflow로 리베이스해줘, … with git-workflow), never qualify, and neither does an explicit invocation of a skill or the loop (use git-workflow: …). The route hint starts dag-ulw only on that same reading; work over many targets without it (review all open PRs) is at most offered, as one question with dag-ulw as the candidate. Reading, summarising, measuring, sharing or checking verbs are not on the list, so those requests route as before. Coordinated workers on one shared task list stay on ulw-work. Measured on the frozen sets under .omc/research/dag-default-routing/acceptance/; prompts outside them are not measured. A Korean loanword or counter after a stop verb is its object (/ulw-work 멈춰 리뷰어, 멈춰 서버 두 대 run), and 멈춰 그거 아님 is a stop.

    • The dag tree under the composer holds still when Down gives it the keys, and keeps a run's lanes until a payload shows the run over (#2224's delta review). With the keys the tree cuts the same rows as without them, so Down only highlights a row; the done layers that gave way at the top stay out, and the last row says ↑↓ to browse when it counts one of them as well as rows below. A new turn's kept lanes now go only on a payload whose graph shows the run over (its one summary line); a payload with no run in it (no session, a graph that lingered out of the HUD, an error) drops nothing. A window shrunk under three rows gives the keys back to the composer, and INSTALLATION.md says → opens a row and Space flips it.

    • Code review separates response shape, evidence, and review depth. Native and portable skills now ship on-demand finding-validation, architecture-delta (including shared-code extraction proof), and risk-selection methods. Reviewer returns require structured spec_axis, explicit could_not_assess, and a separate unranked verification_gaps array; old returns omitting these fields, using a spec string, or ranking gap rows no longer satisfy the contract. Ordinary findings keep P0–P3 and low/medium/high confidence. Located evidence rejects malformed line syntax and blank text; real paths, ordered/in-bounds ranges, actual reads, and claim truth still require reviewer checks. These are instruction and schema contracts, not a new review runner or measured gains in model detection, false positives, semantic accuracy, or review cost.

    • A dag run's view under the composer is a folding tree that says what each child is doing (the owner's ask of 2026-10-22). Each node is one header row: a ▾/▸ fold triangle, its state glyph (the running spinner in the running blue, a red ✗ once failed), its label in bold, and category · model:effort · status · ↓tokens at the right edge. Under an open node, two › rows show its child's latest call (terminal(gh pr checkout 138300)) and its latest other line: a thinking line, Hermes' progress line (the tool names of its last five calls) or its reply, whichever came last, never repeating the call above it. A repair hangs under the node it repairs as ↻ repair N, N its repair cycle (the delegation graph now carries a repair node's cycle), so a second repair reads repair 2 after the first one's nodes leave the drawing. Running and failed nodes open, and so does a node with one of them under it; done and waiting ones stay shut, and a blank row sets each open layer apart. The agents no node names close the tree as other agents. A run taller than the view's third of the screen shuts its quiet layers, folds a busy layer's quiet nodes into one ▸ N more row, then shuts open nodes, before a last row counts what is cut. Down gives the tree the keys: ↑↓ move, ←/→ and Space fold, and Enter, x and s open the highlighted node's child's log, interrupt it and steer it as the agent list does. A new turn keeps the open run's lanes, so their rows stay without waiting for the next frame, and drops them on the first payload that shows the run over. When rows run short, the shut all-done layers are cut before anything else, so running and failed nodes keep the rows. A fold lasts as long as its run: a new run, or another session, starts with none. A click on a tree row highlights it, and on a node's row opens its agent's log. In a dag run Down opens the agent list only when fewer than three rows are left for the tree, and the info line's count leaves out a superseded repair attempt's lane.

    • The OMH TUI's transcript tells blocks, states and tools apart at a glance (the rail and kind colours the owner picked on 2026-10-22). On v4.0.1 a tool's name, its arguments, its duration and its ⎿ results were all one teal, calls were stacked with no gap, and nine of the fifteen rows in the owner's screenshot were skill loads and tool-detail reads. Now each block hangs on a continuous ┃ rail with no fill: grey for your prompt, the skin's accent for a reply, and a call's state colour for the call. A call's mark and rail follow the dag view's rule, each state in its own glyph: a running blue ○ (the same blue as the working glyph, the agent spinners, ● main and a running lane in the dag tree), a green ● once done, a red ✗ when it failed or was denied (the one mark: the answer row under a denied call keeps its words and drops its own); a failed call no longer sits on a red band. Its name is bold in its tool family's hue (skills orchid, shell amber, files peach, web lime, delegation and the dag magenta), each at 4.5:1 on dark and light terminals; arguments are grey and results a fainter grey. A blank row separates two calls. A run of skill_view and tool_describe calls is one row, ● Loaded 4 skills + tool details · …, unfolded again by Ctrl+O, and an interpreter fed its script through a heredoc reads terminal(python3 script · 23 lines) while any other command shows its own first line. The composer sits between two full-width rules behind ❯ with no side edges, as in Claude Code, and your echoed prompt and a command's echo hang on the same ┃ ❯, so a bare ❯ is only ever the composer. A /name that names one of the TUI's commands or Hermes' skills is bold in the skill hue in the composer, in the echoed prompt and on the ↳ you: row; a path, an unknown /foo and a slash inside a word stay plain. Agent rows read a short title (the goal up to its first field label such as DELIVERABLE: or the end of its first sentence; a dag lane keeps its label) and lead their figures with the model, which stays longest as the row narrows. The agent log's tool rows take the same hues.

    • A resumed /skill prompt, the heredoc command count and short agent titles read right in the OMH TUI (#2223 review follow-ups). A session resumed at launch printed its history before Hermes named the session's commands and skills, so a resumed /dag-ulw go stayed plain for the whole session while the TUI's own /model was bold. Once that catalog answers, a resumed prompt whose /name it resolves otherwise is drawn again, with the full redraw a resize does; when no resumed prompt the redraw would print changes, nothing is redrawn. The heredoc summary counts one command a line around the script, and another heredoc's body, a # comment and a \ continuation count none: python3 - <<PY … PY followed by cat > notes.md <<EOF with three lines read + 5 commands and now reads + 1 command. An agent's short title no longer ends at an abbreviation (e.g., i.e., etc., vs., cf.): Fix e.g. the bug. Then … read Fix e.g and now reads Fix e.g. the bug.

    • Native Hermes children receive their own model calibration (#2214). A separate child-only system prompt section selects one exact-model override, family block, or generic fallback from the child's own session model, never parent configuration. It applies at every native-child effort, including low, unknown, or absent effort, because section metadata does not reliably expose reasoning effort. Fanout retains its high/xhigh/max threshold, recorded family fallback, and family-only measurement arm. The standalone plugin and coding control plane share the same stdlib-only calibration definitions and resolver; protocol text, goals, and existing calibration wording are unchanged. The section reaches newly built prompts, not already frozen resumed prompts.

    • Scrolling up in the OMH TUI stops at the top of the session. The launch scrolled your shell's rows and the launcher's warnings into the terminal's scrollback and kept them there, so the wheel walked from the TUI's logo straight into the shell history from before omh started; in Terminal.app all 80 rows of a test shell history were still above the session. The launch now erases the scrollback (ED 3) right after that push, as every full redraw (a resize, Ctrl+L, /redraw, /theme, a session switch, a rewind) already does and as Hermes' own TUI does in its default alternate-screen mode, and the same run measured 0 of them. The TUI still runs on the main screen, so scrolling, selecting and copying stay your terminal's own, and when you quit the session stays above your shell's prompt, as any output does. A terminal that ignores ED 3 keeps the earlier rows one scroll up, as before. A default iTerm2 asks once whether a control sequence may clear the scrollback and keeps it until you choose Always Allow (a resize or Ctrl+L could already ask the same), and inside tmux ED 3 clears that pane's history, as tmux clear-history does.

    • A skill command typed once in the OMH TUI shows as one input row. /dag-ulw <request> printed > /dag-ulw <request> twice: the command's echo, then the prompt the skill sends, which shows Hermes' display for the turn. Hermes got one command.dispatch and one prompt.submit either way; only the screen doubled. Once command.dispatch answers with a skill's prompt, the TUI takes the echo back and the prompt row is the input. That row shows Hermes' display, the same row a resumed or rewound session draws for the turn, or the line as typed when Hermes names none. A dispatch notice, such as stacked skills loading, goes under that row, also when the prompt waits behind a running turn. A refused prompt reads as one for a typed prompt. A built-in command that sends a prompt with no display (/plan, /retry, /learn, /init) keeps its echo above what it sends, as Hermes' own TUI shows it, and commands that print output keep their echo. The same line sent twice still shows twice and is sent twice.

    • A skill command sent over a picked rewind replaces the picked prompt. With a rewind picked (Esc Esc or /rewind), a line starting with / ran as a command. A skill's prompt then went as a plain prompt.submit without truncate_before_row_id, so Hermes kept the prompt you meant to replace and everything after it, and the rewind stayed picked for the next prompt. A prompt a command sends (a skill, /plan) now goes in the picked prompt's place, as a typed prompt does: the session is cut there, the screen is redrawn from the cut, and the rewind clears; /steer <text> sent between turns does the same. A command that only prints (/help, /usage) runs and leaves the rewind picked. /retry and /undo take the last turn back in Hermes themselves, so they cancel the rewind and say so: /retry resends plain, and the prompt /undo puts back in the composer goes plain when you send it. Before, a rewind left picked after /undo sent that prompt as a cut at a row Hermes had removed; Hermes refuses that cut, and the prompt was gone from the composer. A rewind picked while either was out, or the rewind list left open, goes too. For any other command, a rewind cancelled before Hermes builds the command's prompt, or picked (again) after the command went, cuts nothing and stays picked for the next prompt.

    • A Hermes slash-skill turn routes on what the person typed, not on the skill body Hermes pasted in. Hermes sends /dag-ulw <text> to every hook as the skill's whole SKILL.md inside a frame, and OMH routed that body as the person's words: the owner's /dag-ulw turn got ulw-work and ulw-plan hints, and never the "call omh_dag define" one. person_request_text now reads the frame back the way Hermes' own describe_skill_invocation does (one skill, a bundle or stacked /a /b, a bare invocation, a runtime note, a shared session's sender prefix), so a skill of OMH's own routes as /<skill> <instruction>. Another skill's turn is a host-command turn and gets no OMH route hint: the person chose that skill. Only a turn that opens with the frame's exact first line is read this way, so typed messages route as before. A dag-ulw turn's first delegate_task with no [dag: marker, before the conversation has a run, is now sent back to omh_dag action=define once; omh_dag status with no run says to start with define; and the dag-ulw skill says never to delegate_task before define returns its calls.

    • A question, a status word or a stop after /dag-ulw starts no run, in Korean, Japanese and Chinese as in English. Every Hermes /dag-ulw turn now routes as the sigil form, and that form read only English question words: /dag-ulw 이게 뭐야?, /dag-ulw 어떻게 써?, /dag-ulw CI 통과했어?, /dag-ulw /ulw-plan what is it?, /dag-ulw 停止 and /dag-ulw cancel order dispatched a graph run. A question after the sigil (and after any skills stacked on it) is now answered; status, show me the run, 뭐 남았어?, still running?, 状态 and /dag-ulw /ulw-plan 상태 보여줘 read as status; 停止 and 止めて cancel; an English stop word with other words asks which, as it does after the bare name; and a capability question that names dag-ulw gets the answer, not dag-ulw's run card. A request that holds a question word, opens on a request verb, or is asked politely still runs (/dag-ulw 왜 깨졌는지 찾아서 고쳐줘, /dag-ulw migrate the billing service?, /dag-ulw 나눠서 구현해줄 수 있어?).

    • After an explicit skill invocation the skill runs, unless the words stop it, ask how its run stands, or ask what it is. /ulw-work stop, /omh-plan stop, /ulw-work 그만, /ulw-plan what does this do? and /ulw-work status dispatched the skill they named, typed or as Hermes' slash-skill turn: the router's explicit-invocation rule read a sigil as a command before its stop and question guards. One rule (invocation_cues.invoked_instruction) now reads the words after a sigil for the router, the chat wrapper and Hermes' route hint alike. A stop with no task object (stop, 그만, 이거 취소, 停止, 止めて, 取消) and a question about the skill (what is it?, 이게 뭐야?, 어떻게 써?, 何これ, 这是什么) are answered, and a request to see how the run stands (status, show me the run, 상태 보여줘, 뭐 남았어?, 끝났어?, is it running?, 状態, 状态) shows OMH's status card. Everything else runs the skill: a stop verb with a task object (/ulw-work kill the zombie processes, /omh-git-workflow abort the rebase, /ulw-loop 로그 폭주 멈춰줘), a question with one (/omh-code-review is it safe to merge?, /ulw-plan how should we migrate billing?, /omh-app-debugging why did the run stop?), a report (/ultrawork broke my build), and a status request to a skill whose own work that status is (/omh-running-work-board status, /omh-deploy-and-monitor status, /omh-verification-gate is it done?, /omh-ops-observability-card status). A question about an omh- skill is answered about that skill, not with OMH's intro card.

    • A stop of the invoked skill's own run is a stop. /ulw-loop stop the loop, /omh-code-review 리뷰 취소, /ulw-interview 인터뷰 그만, /omh-plan cancel the plan, /omh-app-debugging stop debugging, /ulw-work stop ultrawork and /ulw-loop stop ulw-loop started the skill they stopped. The skill's names (its catalog name and label) and the words for its own run, picked by the words of those names (loop, review, plan, interview, debugging, work, ulw, research, audit, qa, analysis, triage, onboarding, learning), in Korean, Japanese and Chinese too (/ulw-loop ループを止めて, /omh-code-review 取消审查, /ulw-plan 计划取消, /omh-code-review 코드 리뷰 취소), now count as the stop's object, so these are answered; a stop of anything else is still work (/omh-data-pipelines stop the pipelines, /ulw-loop stop the loop when tests pass, /ulw-loop 무한 루프 멈춰, /omh-plan cancel the plan for the free tier). A Korean stop said with why (그만해 비싸, 멈춰 너무 오래 걸려), a stop after no (no stop), and the Chinese 别做了, 不要做了, 算了 and 停一下 are stops too, and so is a Korean stop followed by fillers, a negation or what the person is in the middle of (멈춰 일단, 취소해 이거 아니야, 그만해 지금 회의 중); a Korean stop verb before a noun is that noun's stop, work (/ulw-work 취소해 주문, /ulw-work 멈춰 서버). After /omh-cancel a stop is that skill's own work and dispatches it, as on origin/main.

    • After a skill that explains or judges what the person points at, a question about "this" or "it" is its work. /omh-code-review how is it?, /omh-code-review 这是什么, /omh-codebase-onboarding what does this do? and /omh-paper-learning what is it? run the skill; only a question that names the skill (what is omh-code-review?, what does this skill do?, 이 스킬 뭐야?) or asks how to use it (어떻게 써?, how do I use it?) is answered about it, after any skill. Only the skill's hyphenated or coined name names it (what is omh-rust?, what is ulw-loop?): "a skill", a bare 스킬, a name after an article and a name that is a single common word name a thing, so /omh-docs what is a skill?, /omh-frontend what does the frontend do? and /omh-rust what is rust? run their skill; /ulw-work what is ulw? names ultrawork by its short name. /omh-context-budget-review status runs the context budget review, whose work that status is; only a skill whose job is reporting a state runs on a status request, and after any other the status card shows (/omh-cancel 상태 보여줘, /omh-doctor status, which ran their skill on origin/main). A sigil's name is read in any case, as Hermes reads it (/ULW-LOOP stop the loop). An invoked skill's omh- label no longer makes its words an OMH intro or quickstart question: /omh-paper-learning what is it? got OMH's intro card on origin/main. After /dag-ulw, a stop verb with an object (/dag-ulw kill the zombie processes) asks once whether to stop the run or start one on that work, dag-ulw's own reading of a stop word with other words; origin/main started a run.

    • In the frozen acceptance set's 3,097 labeled explicit invocations, typed and as Hermes' turn, no stop, status or question-about-the-skill row starts a run (origin/main started one on 2,183 of 2,241), and none of the 856 labeled work rows is answered instead of run; four questions about the run after /dag-ulw are answered and three stop verbs with an object after it ask once, by dag-ulw's own reading. Rows outside that set are not measured.

    • The OMH TUI's own commands work typed in any case. /Rewind, /HELP or /Steer <text> used to miss the TUI's command table, which matched the name exactly, and went to Hermes as typed. Hermes matches its commands whatever their case, but it cannot open the TUI's rewind list or show its help, so the command did not do what it does in lower case. The TUI's commands and their aliases now resolve whatever their case, as Hermes' do; the echo keeps what you typed. The completion menu, the /name highlight and /help's Hermes list match names the same way, so /Dag-ulw is highlighted as the skill Hermes runs for it. A name the TUI does not have, and a skill, still go to Hermes as typed.

    • The README and the GitHub Pages site are a lean landing page (#2269, #2271, #2272, #2274). All four READMEs open with the spectrum OH-MY-HERMES wordmark, then the app-icon logo, the tagline "All-in-one package for Hermes Agent. Install once, then just type dag-ulw.", the two install lines (curl …install.sh | sh and omh setup), five one-line bullets, and the dag-ulw clip. The longer install options and details sit behind a link and two collapsed toggles. The site's landing page is rebuilt around the same copy, with the real TUI recording, every install path, the community and sponsor sections, a spectrum aura on the hero and the DAG-ULW band, and a header button with the live GitHub star and npm install counts.

    • Every timestamp read from outside OMH refuses hour 24 on Python 3.14 too (#2263). CPython 3.14 reads T24:00:00 as midnight of the next day, where 3.11 to 3.13 raise. 85 of the 108 fromisoformat calls in src/ read text from outside OMH's own clock; they now go through strict_fromisoformat (or a standalone script's own copy, tested equal) and refuse it on every supported Python. Before, on 3.14, a model-supplied todo done_at was stored verbatim, memory records stayed eligible until the next midnight, and connector receipts and routing observations were admitted. Valid timestamps, including a real T00:00:00, are unaffected. A source gate pins the remaining raw calls, each with its reason.

    • Bare omh opens the OMH TUI with a Hermes gateway already running (#2249). On current upstream Hermes, the stdio gateway entry refuses a home a running gateway owns, so bare omh fell back to Hermes' own TUI every time. The TUI's gateway child now attaches to the profile's shared gateway where Hermes has the attach API, and starts Hermes' own gateway with idle exit when none is running, as Hermes' Ink TUI does. Older Hermes, with only the stdio entry, works as before. When the OMH TUI cannot start, its reason now prints again after the Hermes TUI it falls back to closes.

    • omh setup finishes with a config key that holds a dot (#2248). A provider key such as z.ai-coding-plan: under providers: stopped setup at [2/5] Connecting OMH to Hermes with too many values to unpack. Such a key is now read as one key, not a level of nesting, in the dry run and the real run.

    • omh-ai-job-search is a new installable skill (#2253, #2261). An AI-assisted job search for any profession, adapted from the upstream ai-job-search methodology to Hermes' own research and document tools. It confirms the person's profile and preferences, shortlists real vacancies with their sources and fit reasons, drafts truthful, versioned application material for review, and prepares interviews. A stdlib-only job_state.py tracks applications locally; an application counts as submitted only on confirmed evidence, and a malformed timestamp (hour 24 included, on every Python) is refused before any write. An employer's hiring request is not routed to it. Contributed by @frirenai.

    • research, source-finder and web-research share a bounded multi-source search (#2255). A "find this" request is decomposed into queries, searched and read with the host's own tools, fused into requirement-level coverage, and followed up within a bound. An optional omh_search plugin tool helps with that. There is no new command, mode or required provider key. An unavailable, failed or unsearched source is not reported as zero hits. This is a draft candidate: search quality and installed-model behaviour have not passed acceptance. Contributed by @frirenai.

    • Three executable domain-science skills: omh-geospatial-analysis, omh-biomedical-evidence and omh-molecular-docking (#2235, #2241, #2245). CRS-aware polygon intersection and area in projected metres (Shapely/PyProj); captured BioMCP public-source responses reconciled offline into an evidence matrix that keeps disagreements and gaps; and standalone AutoDock Vina docking with reproducible settings and per-model integrity checks. They are opt-in. Their helpers ship as inert files: installing a skill neither runs a helper nor installs a scientific dependency. On macOS the helpers' path guards, and show-reel's, accept the system /etc, /tmp and /var links into /private, and every other symlink is still refused. Docking refuses one file passed as both receptor and ligand, however it is spelled. Contributed by @frirenai.

    • Skill replies end on their result, not on a list of things not done (#2238). The shared closing guidance asked for a list of what was not checked or executed and for a question at the end, so a briefing could close on "I did not install or change settings" when neither was asked for. The shared closing rule, in every skill body, now orders a reply as result, relevant evidence and limits, the required observed run summary, then one specific question about the content only when a real decision or a useful discussion remains. Completed work needs no question, and unrelated non-action footers and generic offers to continue are left out. Requested work that was not done, a failed retrieval, an unavailable check and a permission decision stay explicit, and a discussion question grants no permission. Contributed by @frirenai.

    • Windows staged self-updates work under PowerShell 7 (#2218). The generation-pointer junction now gets an absolute target, which PowerShell 7 requires; a relative one staged the update and then failed at activation. POSIX symlinks keep their relative targets. Contributed by @frirenai.

    • ulw-research records inspectable evidence (#2227). Claims and sources carry typed provenance, ownership and counter-search checks, so correlated republications cannot count as independent corroboration. An exhausted budget is kept apart from completed research. python -m omh.workflows.research_dossier <json-file> validates a dossier's record structure, with distinct exit codes for valid and invalid input. Passing it is not factual truth or evidence that a source was read. Contributed by @frirenai.

    • omh-show-reel is a new installable skill (#2212, #2220). A multilingual show-reel and launch-film workflow ("Create a launch film.", "제품 쇼케이스 영상 만들어줘", /show-reel …), with an optional motion and audio toolkit that ships as inert files. Its engine, project layout and QA tools match upstream revfactory/showreel (MIT) at a pinned revision, and the render receipt records its quality settings. Narration reaches a network service only when the person runs the opt-in script with ELEVENLABS_API_KEY set; without it, the reel falls back to music and captions. A film named inside another request (an issue title, a Slack post, "make the launch film shorter") keeps that request's owner. Contributed by @frirenai (carries #2195).

  9. 4.0.1

    Expandir

    GitHub Release

    v4.0.1 · graph master

    Oh My Hermes 4.0.1 — graph master

    graph master is the graph-engineering release. A big job can now run on its own as a graph of delegated nodes, the work Hermes hands out is drawn as a graph you can open and steer, and the routes that start paid runs got stricter about when they start one.

    dag-ulw: a big job runs as a graph

    Say dag ulw <goal> (also ulw dag, /dag-ulw, dag ulw해줘), or ask for a big job whose words say how it splits or how wide it reaches. Hermes plans it as a graph of small nodes, runs the ready ones in parallel on its own delegate_task children, settles each from its handoff file, verifies the goal with a separate node, repairs failed nodes (two cycles each), and adds the work the goal check finds missing to the same run (two rounds) before it reports. It asks only for scope, spend or an irreversible step. (#2185)

    dag-ulw also works on a Docker or SSH terminal whose user has a writable home: each node writes its handoff under the terminal user's own home, read once through Hermes' live environment, so a remote run no longer fails every node with handoff_missing. (#2205) Each delegated goal ends with its run marker, so a preview shows the task. (#2203)

    The delegation graph you can see

    Delegated work (delegate_task children, omh_team parts, fanout units) is drawn as a graph in the OMH TUI and in omh hud --graph, and chat replies get the run's start, its count changes and its end, never the boxes. (#2179) The boxes stand in frontiers with state colours, Dependencies lists every recorded edge exactly, and /dag opens an interactive full-screen view. (#2191, #2199) The graph survives Hermes' rotating compaction (#2201), and each lane carries its subagent's output tokens, as does the agent view that now sits under the composer. (#2188, #2192, #2194, #2200)

    Routing safety: a stop or a question never starts a paid run

    A complaint, a question or a stop that opens with a skill's name (ultrawork broke my build, dag ulw는 뭐야?, ultrawork stop, loop 멈춰) is answered directly; neither the router nor Hermes' route hint runs the skill. (#2207) dag-ulw routing reads remarks, bug reports and questions as such, and a polite request as a run. (#2197) A delegation route is refused when Hermes would send the child somewhere other than where the route says (#2180, #2202), and fallback advances only the calling conversation's own route. (#2177)

    Restarts and compaction hand work back

    A Hermes restart no longer leaves an omh_team part waiting four hours: the first team_start after a restart hands back each part whose process is gone, same attempt, no try spent. (#2198) A Hermes compaction no longer reads as the end of the work for the delegation route or omh_team. (#2176)

    Read-only Hermes stores through a quoted URI

    A Hermes home whose path holds #, ?, or a % followed by two hex digits is read, and no stray file is created beside it: every read-only open of state.db and kanban.db goes through one helper that quotes the path. (#2209)

    Also in this release

    • Every Hermes delegate_task child gets OMH's delegation protocol, with one verification rule and a doctor row for the platform they are marked with. (#2178, #2208)
    • A child's model verdict rests on its own recorded calls, not Hermes' auxiliary ones or today's config. (#2173, #2196)
    • Vietnamese ships as a trigger language pack, contributed by @PhucNguyen-Dev. (#2187, carries #2182)
    • Bare omh opens the OMH TUI on Bun and npm installs. (#2170)
    • A plugin-only install says why native Loop has no backend, and omh_probe names its own import failure. (#2169, #2171)
    • A skill body edit no longer rewrites the six host installer directories. (#2189)
    • omh doctor's real-loader probe can observe a load on Windows and on the install omh setup makes. (#2210, carries #2204)
    • A new installable market-research skill: scoped TAM/SAM/SOM sizing with sensitivity, attributed customer evidence, and comparable competitor and pricing profiles, contributed by @frirenai. It offers itself only for a request that opens with one of its method phrases. (#2211, carries #2190)

    The full entry list is the 4.0.1 section of CHANGELOG.md. #2211 merged after the release prep PR was cut and has no entry there; it is listed above.

    Upgrade

    omh update
    

    Restart or reload Hermes after updating. Launch with omh.

    Release record

    • Version and channel: 4.0.1, stable (.release-channel = stable; GitHub release not a prerelease, npm latest, Homebrew tap).
    • Hermes skill tap / install wording: unchanged from 4.0.0: hermes skills tap add rlaope/oh-my-hermes; default install curl -fsSL https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.sh | sh.
    • Update path tested: not run for this cut.
    • Workflow docs generation: all 11 docs ... --check gates passed locally on the release prep head 158ecbb8f (PR #2213), and in PR CI (11/11) and main CI on the tagged commit 7ceaf0607 (13/13 jobs).
    • Harness catalog validation: no separate harness validate step ran for this cut; the catalog is covered by the full suite in CI.
    • Source-checkout console script smoke: Installed OMH command smoke passed in main CI on 7ceaf0607 (Python 3.11 and 3.13).
    • Runtime validation: Runtime artifact smoke passed in main CI on 7ceaf0607 (Python 3.11 and 3.13).
    • Workflow learning review queue: no workflow-learning contract changed in this release.
    • Capability probe: Capability probe smoke passed in main CI on 7ceaf0607.
    • Install script smoke: plan-only (install --dry-run, stable channel included) in main CI on 7ceaf0607; the Windows installer end-to-end step ran live on test-windows (0) and passed.
    • Hermes CLI install smoke: plan-only (release hermes-smoke) in main CI on 7ceaf0607.
    • Plugin bundle: plugin.yaml version 4.0.1, requires_hermes unchanged (>=0.21.1). The setup-path command smoke (--install-path setup) passed in main CI on 7ceaf0607.
    • GitHub Pages: site copy changed (hero badge to v4.0.1, plus site copy from PRs since 4.0.0). The Pages run on 7ceaf0607 passed, and the live site at https://rlaope.github.io/oh-my-hermes/ reads v4.0.1.
    • npm / Homebrew: npm latest = 4.0.1, observed on the registry at 2026-10-22 04:15:54 UTC (published with provenance). The rlaope/homebrew-tap formula omh.rb moved to 4.0.1 in 6d504603c, pointing at the release wheel.
    • Known manual Hermes checks not automated: a live dag-ulw run on a Docker or SSH terminal, and the TUI graph in a real terminal, were not re-run for this cut.
    • Machine sync after the cut: none has synced yet.
    • Claims resting on wrapper evidence: the chat posts of a delegated run's start, counts and end, and the TUI graph, are OMH-rendered from Hermes' recorded state; they are not Hermes-native rendering.

    Merged pull requests since v4.0.0

    #2169, #2170, #2171, #2173, #2176, #2177, #2178, #2179, #2180, #2181, #2183, #2184, #2185, #2186, #2187, #2188, #2189, #2191, #2192, #2194, #2196, #2197, #2198, #2199, #2200, #2201, #2202, #2203, #2205, #2206, #2207, #2208, #2209, #2210, #2211, and the release prep PR #2213. (#2206 is the dev-only ruff 0.17.0 → 0.17.1 bump; #2210 carries #2204; #2211 carries #2190.)

    The complete version-bound changelog follows below.

    Full changelog

    • omh doctor's real-loader probe can observe a load on Windows, and on the install omh setup makes. The probe wrote a synthetic config.yaml naming no plugins.entries.omh.settings.omh_home, and the plugin refuses to load in a native host whose profile names no store (#2037) -- so every install omh setup registers reported loader_probe_failed and the row could only ever read real Hermes loader not observed. It now carries the binding the real home declares. The child process also gets the OS baseline (SystemRoot, SYSTEMDRIVE, windir, TEMP, TMP): without SystemRoot, import asyncio dies with WinError 10106 before the plugin is reached, which read as a plugin load failure on Windows.

    • A Hermes home whose path holds #, ?, or a % followed by two hex digits is read, and no stray file is created beside it. Ten read-only opens of Hermes' state.db and kanban.db built file:{path}?mode=ro from the bare path, which SQLite parses as a URI. A # or ? cut the path: the open lost mode=ro and created an empty writable file at the cut, and a %41 opened another directory's file, or none. The resume line after a TUI exit, the cost receipt and omh_run_summary, the delegation ledger's root walk, the native delegate and kanban HUD readers, and omh_dag's free slots then found nothing. They now open through one bundle helper, sqlite_readonly.connect_readonly, which quotes the path's file-system bytes. A policy test fails any sqlite3.connect(..., uri=...) in src/ whose URI is built from an unquoted path.

    • A complaint, a question or a stop that opens with a skill's name no longer runs the skill, in the shapes measured. ultrawork broke my build, dag ulw failed again, ralplan 때문에 빌드 깨졌어, maestro failed again, ultraqa 또 실패했어, code-review 결과가 이상해, ralph broke my build, ultrawork 이거 어떻게 써, ultrawork still running?, ultrawork stop, ultrawork stop, too expensive and loop 멈춰 are answered directly, and Hermes' route hint neither offers the skill nor calls it explicitly invoked. The reading is name_report_cues in the plugin bundle, which the router, the retired-engine aliases and the route hint share:

      • the name, or the skill's own output, is the subject of a finite verb (broke, failed again, is too slow, output was wrong);
      • a Korean predicate after it is a complaint, negated or past (너무 느려, 안 돼, 또 실패했어);
      • the name is the cause (때문에);
      • a stop word is all there is, said once or twice or with its reason after it;
      • a question asks about the skill. A report counts only after a name that can mean nothing but the skill (a coined name such as ultrawork, or a hyphenated id such as code-review): a name that is also a word may open a sentence about the person's own thing (frontend is broken on safari, code review takes forever), which keeps its route. After dag-ulw, ultrawork and maestro a question that opens with a question word is answered whatever it asks (dag ulw 결제 api 뭐야), and after ralplan unless it deliberates the work (ralplan how should we split the billing refactor still plans). After a skill whose work is answering what is running, wrong or left, that question is its work (app-debugging what's wrong). A request still runs the skill: ultrawork fix the broken build, dag ulw the build that broke, ultrawork failed again, rerun it, … one more time, … 다시, the name said again (ultrawork failed again, ultrawork it again), a request after a use clause or credit (ultrawork 하고 나서 배포해, ultrawork 덕분에 빨리 끝났어, 결제 모듈도 이렇게 해줘), and every sigil form (/dag-ulw broke the build, fix it). A stop of the dag-ulw run said as an act after words that may name the work (dag ulw 결제 취소 처리해줘, 결제 작업 중단 걸어줘, 취소 버튼 눌러줘) and an English stop word beside such words (dag ulw cancel order) now ask which (owner decision O7).
    • dag-ulw works on a Docker or SSH terminal whose user has a writable home. There a child's write_file lands inside the backend, so OMH's host-side check refused every node with handoff_missing, and on SSH a child often could not even create the host's OMH path on the remote. omh_dag now records the terminal backend Hermes resolves for the profile when a run is defined. Before the run's first node goes out, it reads the terminal user's $HOME once through Hermes' live environment and pins <home>/.omh/dag on the run, where every node writes its handoff and its dependents read it; with no terminal environment up yet, nothing goes out and the result says to run one terminal command first, in the same turn (OMH never starts one). A home that is a link is judged, and pinned, as the directory it points at. A home another user can write, one in a shared temp directory, one with whitespace in it, or one OMH cannot read keeps this host's path, as before. settle reads the handoff inside the backend under the host's rule: a regular file and no link, so a missing file, a directory or a link is still refused. When the terminal cannot be reached (its environment is gone, Hermes' registry cannot be read, or no answer comes), the node is accepted with handoff_checked: false and the reason, and every view says handoff not checked: omh_dag status and the run's status line, the boxes and the tree, omh hud --graph, the TUI's transcript block and its one-line summary, and the chat count line and end post. A read the parent's interrupt stopped settles nothing. The read runs only once every other check of the settle passed, through Hermes' base execute, so it never enters Docker's recovery of a gone container. A local terminal, or a backend OMH does not read (modal, daytona, singularity), keeps the host check as before.

    • dag-ulw routing reads remarks, bug reports, questions and polite requests as what they are. The router and Hermes' route hint read the same dag_ulw_cues, so each change holds on both paths. can you dag ulw the migration? and could you please dag ulw … now start a run. A polite question about it (can you explain dag ulw?), or one that weighs it against another way (can you dag ulw this or should I just fix it myself?, would you dag ulw it, if you were me?), is answered. A cache that breaks when two named modules request at once (api 모듈과 ui 모듈이 동시에 요청하면 캐시가 깨져, 고쳐줘) stays one ultrawork fix, while failing tests fixed 병렬로 (각각) still reach dag-ulw. A question or explain request about splitting work (explain how to split …, show me how to split …, is it possible to split …, how do I split …?, 병렬로 나눠서 하는 방법 설명해줘 / 알려줘 / 방법이 뭐야) no longer starts a run or hints "Start at once". A proposal such as how about we split …, can we split …? or … 구현해줄 수 있어? still asks for the work. A remark whose subject holds the split (the fan out across the services is too high) and a split whose object is one paragraph or sentence (split the paragraph into parallel parts for the translation) no longer start a run or hint "Start at once" either. A Korean stop word after the name no longer starts a run unless the words around the last stop word make it work: a work noun or verb after it (… 취소 기능 구현해줘), a request verb of its own ending the message (… 취소 내역 확인해줘, … 종료 훅 등록해줘), a condition before or after it (… 실패하면 중단해줘, … 취소 시 환불되게 해줘), or a work word as the object of 꺼줘 (다크모드 꺼줘, not … 작업 꺼줘). With none, a stop noun after words that may name the work (dag ulw 주문 취소, dag ulw 주문 취소요, dag ulw 취소 요청, a verb-form cancel after more than three words, or a stop of the work on a cancel feature such as … 취소 기능 만드는 거 그만해줘) neither cancels the run in progress nor starts one, and neither does a cancel followed by a restart (… 지금 거 취소하고 결제 모듈부터 다시 해줘): the router clarifies and the hint asks one question naming both readings (owner decision O7). A question about stopping (dag ulw 취소 어떻게 해, dag ulw 멈출 수 있어, dag ulw 중단하면 어떻게 돼), or whether it stopped (dag ulw 취소됐어?, dag ulw 멈췄어, dag ulw 확인해줘, 취소됐는지), is answered, unless a re-run request comes first (dag ulw 다시 돌려줘, 아까 취소한 거 runs). So is a sigil before a question (/dag-ulw what is it?, /dag-ulw how do I cancel the run?), unless it goes on to ask for work (/dag-ulw explain the plan, then run the migration runs). These cancel. On main the polite endings, 그만 돌려, 멈춰라, 취소요, 정지, 스톱 and 꺼줘 started a run, 취소해줄 수 있어? was answered, and can you stop dag ulw clarified:

      • a verb ending, plain or polite, after up to three words (주문 취소해줘, 취소해줘요, 취소하세요, 취소 부탁해, 취소해줄 수 있어?, 그만 돌려, 멈춰라);
      • the bare stop word, with a bare 요, 좀 or ㄱ or after stop adverbs (취소, 취소요, 정지, 스톱, 꺼줘, 지금 바로 취소);
      • a word for the run, with or without a modifier (작업 취소, 진행 중인 작업 취소, 그것도 취소, 다른 거 취소, 저거 말고 이거 취소);
      • the act of stopping after those (작업 취소 처리해줘, 런 중단 요청 보내줘, 취소 눌러줘), and a stop that did not take (취소했는데 계속 돌아가, 취소 눌렀는데 계속 돌아가);
      • an English stop after can you (can you stop dag ulw).

      dag ulw 주문 취소 기능 구현해줘 still runs.

    • The delegate route tool answers right at its edges. An alias (*name), <<, or a tag Hermes cannot build (!!merge) given as a value in config.yaml's model or delegation section, on a nested line or inside a flow collection too, or in a root base_url, api_base or provider, now refuses every route, with or without a provider (config_unloadable). Hermes cannot load the file when an alias has no anchor, a value is <<, or a tag is not one it builds, so a route written there never took effect while OMH answered routed; a defined alias refuses too, because the check does not follow anchors. In the managed config the same values make their section unreadable. A route that names no provider over a direct delegation.base_url is checked against the provider Hermes names that endpoint (anthropic, openai-codex) instead of the main session's, since Hermes builds the child there; a custom endpoint is not checked, and one the check cannot read keeps the session's check. fallback with no category now continues the category this conversation's own route was set under, so a model in two chains no longer fails as ambiguous. It refuses when the model is not in that chain, and a route set by model alone still needs the category. docs/INSTALLATION.md now says that false, no, off and zero for model.provider read as unset only up to Hermes v0.21.3 and as set from v0.21.4, and why OMH still refuses on them.

    • The OMH TUI's composer stays still when a turn ends with agents listed under it. On the bottom rows the agent view took the working row's place once the turn ended, so at 12 rows with the plan and seven agents the composer's box moved up a row and came back down when the next turn started. Once a turn ends there, the view no longer takes the working row's place: it gets the rows the slot, the menu and the composer leave it, less the working rows, and never fewer than the turn's last frame gave it. That place stays blank above the composer until output fills it. A draft, a menu or a question that squeezed the view as the turn ended gives the rows back when it closes. A view that wants more rows than it did drops the hold, so it does not snap back when it wants fewer again, and so does a frame drawn above the bottom rows. The cost, measured over 1,287 cases: at 8, 9, 12 and 13 rows with the plan, 54 idle views keep 1 or 2 fewer rows than before, the same 54 where the composer used to jump, and each keeps at least one row (a header, a running agent or the +N count). A test now pins that a finished DAG run counts at its one summary line in the agents block's budget. The docs say what the DAG tree takes: up to a third of the rows (8 of 24, 13 of 40), and no more rows than the tree has, so a five-lane run's seven-row tree takes 7 at 24 and at 40.

    • A delegated goal ends with its run marker, so a preview shows the task. Hermes previews each delegate_task goal by its first 40 characters on a Discord tool-progress post, the CLI and the TUI, so a dag-ulw node's leading [dag:<run>/<node>] marker and an omh_team helper's leading [omh-team:<team>/<unit>/attempt-<n>] key took the whole preview (🔀 Delegating 1 tasks: [dag:billing-refactor/invoic...). omh_dag and omh_team now write the marker at the end of the goal, on the label's line (Render invoice PDFs from the new template [dag:billing-refactor/invoice-pdf]), and the binding reads the last whole marker in the goal, on any line, and binds only when it is its own lane's: a goal handed out before the update still opens with it and still binds, and text a model adds after the marker no longer leaves the child unbound and its node offered again. A node label or a part title that holds either [dag: or [omh-team: is refused, so a label cannot name another node or team part. The OMH TUI drops a whole marker wherever it is from a delegate_task row and an agent row, also a marker head Hermes' cut left at the end of a goal (… to API [dag:ex...), even where Hermes' cut of the whole row follows it; a head too short to name its kind ([, [d, [o) only after a space. Hermes' own lines (the Discord post, its completion line) keep what its cut leaves: after a short label, the marker or its head (2 tasks: Export API [dag:export/api] ...).

    • One verification rule for Hermes delegate_task children, and a doctor row for the platform they are marked with. The protocol section's verification rule now binds a child whose task changes files or other state (a deploy, a migration, remote or config state), not only files. Its goal echo-back asks for the numbered criteria only when the goal has them, so a /review reviewer is not asked for criteria it was never given; fanout unit prompts keep the unconditional line. The coding-handling harness line and omh coding composition-guide now state the same conditions: the harness no longer makes the verification pass mandatory on every prompt. omh doctor gains plugin_delegation_protocol, which reads the installed Hermes tools/delegate_tool.py for platform="subagent", the platform the section renders for, and warns when it is gone. The unknown-role warning in pre_tool_call is removed: Hermes reads only action from that hook's result, so it never reached anyone, and the child's own pre_llm_call turn already warns about an unknown [omh-role:...] marker in its goal.

    • A conversation keeps its delegation graph across Hermes' rotating compaction. With compression.in_place: false, Hermes moves a conversation to a new session id mid-run. The HUD graph, the TUI's agent tree and omh hud --graph then read only the new session: its lanes vanished, and a child that stopped afterwards (Hermes names the new session as its parent) never read done. The graph now reads the ledger, the omh_team record and the omh_dag run under every session of the conversation, as the compaction links in state.db join them. A stop the ledger could not join is matched to its child, and omh_dag reads it the same way: a dag-ulw node whose child came back before a Hermes restart reads returned instead of being offered, and run, again. A chat post goes on from the last post any of those sessions recorded, so the run is not announced again. A child whose own session rotated is read through its continuation, so it shows running while its continuation runs, instead of unknown (compacted), which also posted it as Stopped mid-run. Its output tokens (the agent view and the graph) add up over every session it wrote in, instead of stopping at the rotation. Hermes' default in-place compaction keeps one session id, and its graphs are unchanged.

    • A Hermes restart no longer leaves a team part waiting four hours. A helper runs inside the Hermes process that started it, and a restart ends it without the subagent_stop callback, so omh_team held the part as out until its four-hour bound blocked it as delegation_not_returned. The dispatch now records that process (pid and start-time fingerprint, through Hermes' own gateway.status probes, imported when called, and where the pid was read: on Linux the boot id and pid namespace, on macOS the hardware UUID, on Windows the MachineGuid), and the first team_start after a restart hands back each part whose process is gone: same attempt, no try spent, said as "Hermes restarted while 1 part was out; it starts again." A part with no recorded process, one recorded on another machine or in another container, or one whose process cannot be told alive or gone, waits for the bound as before. On Linux the boot is part of where the pid was read, so a reboot, wsl --shutdown, docker restart or a recreated container also reads as unknown: the team part waits the four-hour bound, and a dag node stays running with no bound, as before. Where that id cannot be read, or on another platform, nothing is recorded and the hand-back is off: the team part waits the four-hour bound and a dag node stays running, as before. Hand-backs are counted apart from re-emits for a slow start: after three, a part blocks as helper_lost ("Hermes kept restarting while its helper was out"). A dag-ulw node whose child a restart ended no longer stays running with no bound: its ledger entry records the same process, and omh_dag ready offers the node again at once, at most three times per attempt, then says on every omh_dag reply to settle it. The delegation graph no longer draws an ended team helper as a lane of its own. Start lines name how many parts start: "Started 1 part", "1 starts now and the rest wait for the parts they depend on", and no line says "1 parts".

    • The delegation graph keeps a long duration whole, draws no ? title, and puts a repair above what waits on it. A duration of 100 hours or more (111h 06m) is measured as it reads, so a box widens by one cell at the 100-hour mark instead of cutting it to 111h 0…, and the tree's columns give it the cell. A node no record titles reads its own name: a team part its unit id, a child session <tail>; a drawing or chat post given no title reads untitled. Where a dag-ulw node failed and was repaired, the repair takes its own frontier between that node and what comes after it (Verify export no longer sits beside the repair it waits for); the parents line and Dependencies keep the recorded edges. An end post says (sent, no child seen), not (dispatched), for a stopped team part whose child never appeared. OMH_SUBAGENT_GRAPH is read in any case (OFF, Off), in the HUD and for chat posts; the ledger keeps recording while it is off.

    • A child's answering model is judged on its own calls, not Hermes' auxiliary ones. Hermes records a compression or vision call in session_model_usage under a non-empty task, on the aux route's model. The HUD's served-model check counted those rows. One compression call on another model turned a child's agreement into unknown/multiple_usage_models, and a child whose only rows were auxiliary read as a disagreement with the model it asked for. The answering model, its count, and the model the HUD row and the delegation graph show for a child whose session row names none now read only rows with an empty task. Tokens, calls and recorded cost still sum every row, because auxiliary calls are real spend, and a child whose only calls were auxiliary is still priced at the model they ran on. One approximate figure moves: a child whose session row names no model, with its own calls on one model and an auxiliary call on another, now gets the model, category and ~$ estimate that a child whose session row names that model already got. Before, it got none, or $0.0000 (unknown). A state.db without the task column reads as before.

    • A skill body edit no longer rewrites the six host installer directories. Every .<host>/install.sh, install.ps1 and manifest.json embedded the full per-file SHA256 list, the installers twice, so a one-line SKILL.md change moved all 18 files. The per-file digests now live once, in agent-skills/.omh-agent-skills-manifest.json: the exact receipt omh install --target agents writes, generated beside the pack and pinned by docs agent-skills --check as a canonical file. The installers are static per host and read that inventory. As before, they refuse a symlink, an extra file or a missing file in the source and verify every digest before the first write, then copy the inventory verbatim as the target receipt. They accept only the exact layout the producer emits, sorted and unique, with no row outside the files block. Without that, a row one reader skipped could make the copied receipt own a user's own skill, which omh install --target agents would later delete as retired. Each host manifest.json, and the host_adapter block of omh install --target agents --host <host> --json, drops files, source_digest and digest_format and names the inventory in a new inventory field; the schema stays omh_host_adapter/v1. omh quality-evidence agent-skills compares scan paths with that inventory, and the source_digest its receipts record keeps its definition and value, so stored receipts still compare.

    • The OMH TUI shows each subagent's output tokens. A row in the agent view under the composer ends 12s · 3 tools · ↓1.2k tokens, and the agent's log header carries the same count. While a subagent runs, the count is read from its own Hermes session in state.db (one read-only query per HUD poll, only while one runs), so it grows a poll behind each model call; once it finishes it is Hermes' own count. There is no token part until there is a count, and a narrow row drops it first. The delegation graph carries the same fact on each lane bound to a child (output_tokens in delegation_graph/v1, dag nodes included), and /dag's Node details, its boxes and the tree show it where the line has room; once a lane carries a count, the boxes make room for one through the first hour of a run (● Running 59m 59s · ↓999.9k). Chat posts do not change.

    • dag ulw: a big job runs on its own as a graph. Say dag ulw <goal> (also dag ulw해줘, ulw dag, /dag-ulw), or ask for a big job whose words say how it splits (parallel lanes, a fan-out, parts integrated afterwards) or how wide it reaches (every module, the whole codebase), and Hermes plans it as a graph of small nodes, runs the ready ones in parallel on its own delegate_task children, settles each from its handoff file, verifies the goal with a separate node, repairs failed nodes (two cycles each; a failed verification runs again after its repair, and only that re-run passes it), and adds the work the goal check finds missing to the same run (two rounds) before it reports. It starts at once and asks only for scope, spend or an irreversible step. A question about it ("dag ulw는 뭐야?", "how does dag-ulw handle repairs?") is answered, not run, and a stop, cancel or status word said to it ("dag ulw stop", "dag ulw 상태 보여줘") starts nothing: Hermes' route hint points that turn at the run's own omh_dag cancel or status.

      • The new dag-ulw skill carries the run loop and the planning doctrine; the new omh_dag tool keeps the run (omh_dag/v1), refuses cycles, unknown or self edges, shared write scopes without an edge and more than 48 nodes, and never hands out more delegate_task calls than Hermes' background pool has free (read from delegation.max_concurrent_children and this process's running units in state.db). The run record keeps each node's prompt (at most 3,000 characters a node, created mode 0600), so a compacted parent can still hand nodes out; it keeps nothing a child returned. A node after a repaired one reads the repair's handoff files. extend adds nodes to a run whose nodes all passed, under the same refusals over the whole run, and every node records where it came from (define, repair:<root>, extend:<round>).
      • Big work reaches it by intent, in omh chat and in Hermes' own per-turn route hint alike (one reading, plugin_bundle/omh/dag_ulw_cues.py): on a route the router already gave ultrawork, never when ultrawork is named (ulw, ulw-work, ultrawork), and never for ultrawork's own coordinated workers or delivery cycle. In Hermes, a request no hint rule matched reaches it only when it says outright how the work splits. A concurrency word that names the bug or the feature ("fix the bug where uploads run in parallel", "동시에 두 요청이 오면 캐시가 깨지는 버그 고쳐줘") keeps the request where it was.
      • Graph topology is decided once, in plugin_bundle/omh/dag_core.py. omh_team, the fanout contract, a dag run in the delegation graph and the graph drawing read it there, with their refusals, layers and merge order unchanged (5,000 random graphs, old and new code, no difference). The delegation graph's own layers and a team's stages keep their walks: on a repeated node id or a hand-edited cycle the core answers otherwise.
      • The TUI panel, /dag and omh hud --graph draw the run under its name; on Discord, Slack, Telegram, Mattermost and Matrix a reply carries the start block once, one status line when counts change (none added when the parent's reply already is one), and the end block once the run is over. Nodes added mid-run hang from where they came from in the drawing, and the next status line names them once (+1 added (repair: Docs body)). A delegated child's own turn never gets a post. In the TUI the plan holds the run's layers, the graph its nodes, and the agents block only the children the graph does not draw; the agent view and a delegate_task tool row name a node's child by its label. The transcript keeps the run's graph as boxes (the tree where boxes cannot be drawn) at define, at each repair and extend with the added nodes marked, and once at the end, when the turn that ended the run is over; a chat never gets the boxes. /dag closes a run's graph with its critical path and the nodes that block the most others.
      • Coding nodes go to Codex or Claude Code only when the person chose that owner within the last day through omh chat or omh coding (the router's explicit-choice record, read exactly as the router reads it). Then omh_dag lists the ready coding nodes under export instead of handing them to delegate_task, and omh_dag action=export writes the fanout units and returns the omh coding fanout commands; verification and review nodes stay native, and an exported node is never handed to delegate_task again. The day is dag-ulw's own rule: an external CLI spends the person's money and quota, so a stale choice starts nothing. Hermes chat on its own records no choice, so there every node runs on delegate_task.
    • The delegation graph is drawn as frontiers of boxes, and /dag is interactive. The boxes stand in frontiers (a node's longest path of recorded edges), side by side over a · Same frontier label, with one connector from each frontier to the next; each box holds the title, the category and model, the status and the parents (← Export API, or Start node). The drawing implies no edge: under the boxes, Dependencies lists every recorded edge exactly, and Node details gives a card to each running or stuck node and, in /dag, to the highlighted one. A frontier goes on one row while its boxes can narrow to fit, and wraps only past that. In the OMH TUI a box's border, title and status take its node's state: running blue, handed out indigo, failed red, done the terminal's foreground, waiting dim, selected violet; every state keeps its glyph and word. /dag opens the graph full screen and redraws it from the HUD's newest read: Tab and Shift+Tab move the highlight, Space folds a box, d shows or hides Node details, and nothing is sent anywhere. This replaces the #2179 boxes in /dag and omh hud --graph; the panel under the info line keeps the tree, and the chat posts are unchanged.

    • Vietnamese ships as a trigger language pack. src/routing/trigger_packs/vi.json, contributed by @PhucNguyen-Dev (#2182), covers 20 of the 21 skills the id pack covers, with phrases such as rà soát mã nguồn, kiểm tra khả năng tiếp cận, build thất bại and lập kế hoạch thực hiện. The rule that makes a shipped pack phrase count only as a whole phrase now keys on script rather than on ASCII. The router folds diacritics, so cần, bạn and thêm score as can, ban and them; keyed on ASCII, no Vietnamese phrase was phrase-only, its words scored alone, and English sentences such as "can I ban this user from the chat" named a pack lane. A pack phrase whose letters are all Latin, with or without diacritics, now scores only as a whole phrase and merges after the ja, ko and zh phrases; Indonesian routes as before. With its words out of scoring, the pack can use the loanword Vietnamese developers say: merge replaces hòa nhập, and the backend phrase becomes thiết kế contract cho endpoint. A bare phrase clarifies with its skill named and dispatches where an English word the skill carries comes with it. ultrawork carries no Vietnamese phrase: its always-loaded body has 3 bytes left under the ceiling. The phrases have not been reviewed by a native speaker. vi is an input language only: it is not a --language / OMH_LANG output choice.

    • A delegation route is refused when Hermes would not send the child where it says. omh_delegate_route wrote model, provider and effort, and answered routed even when Hermes would build the child elsewhere. That was measured with Hermes' own resolver functions, imported from git archive copies of Hermes main 8b9ac5773e, v0.22.2 and v0.21.1:

      • With delegation.base_url, delegation.api_key and delegation.api_mode set, a route to anthropic built a custom child on that endpoint, using delegation.api_key.
      • With a local model.base_url and no model.provider, Hermes' local-endpoint fallback ignored the requested provider. The child went to OpenRouter with the OpenRouter key, or the dispatch failed for want of one.

      set and fallback now refuse these cases with status: error, a reason, and conflicting_keys:

      • a direct delegation.base_url that overrides the provider, or that names it but has no delegation.api_key of its own, so the parent session's key would be sent;
      • a route with no provider that would remove an SDK provider beside that endpoint;
      • the local-endpoint fallback;
      • a managed config ($HERMES_MANAGED_DIR or /etc/hermes) that sets the endpoint or pins a route key;
      • a managed config that gives delegation one non-null value, which replaces the whole section;
      • a config line OMH cannot read safely, including a repeated key, a merge key, a value continued past its line, and a model: or delegation: given as an alias, anchor or tag. A one-line flow mapping there is read, the shape Hermes' own save_config keeps.

      openai is refused under the local-endpoint fallback, because Hermes v0.21.1 to v0.21.3 send it to OpenRouter. A spelling Hermes versions read differently, such as n or 0e0, is undecidable. A route that names no provider still writes over delegation.base_url, and status names that endpoint under child_endpoint.

      A refusal changes nothing in config.yaml and records nothing: no restore record and no provenance. The route lock's file is the only file left. OMH changes none of the keys it names. Restoring the person's own values is never refused. A leading byte-order mark before delegation: is now kept, and that section is reused instead of a second delegation: being appended. Hermes' loader rejected the duplicate.

    • A Hermes compaction no longer reads as the end of the work, for the delegation route or for omh_team. Hermes' rotating compaction (compression.in_place: false; the default in-place compaction keeps one session id and was never affected) closes the running session row with end_reason compression and goes on in a new row that names it as its parent. Three OMH readers took that for an ending or a different session. The session-start restore of a delegation route read the writer's closed row as "not live" and put the person's baseline back while the compacted conversation was still dispatching; it now follows the continuation, keeps the route while any continuation is running, restores once Hermes closed every one, and holds it as continuation_unverified when the continuation cannot be read, for at most six hours of the route's age (continuation_unverified_expired after that). A team helper that compacted came back under its new session id, which matched nothing, so its part stayed dispatched until the four-hour bound blocked it as "never came back"; the return now links to the attempt through the earlier segments the new id continues. A parent that compacted after team_start lost its team: team_reconcile and team_status from the new segment answered team_not_found, and a helper's start or stop naming the new segment was dropped; the team is now found under the segment that started it, and a team_start there binds to the plan accepted before the compaction. All three read only the links Hermes itself writes on a compaction (the closed row's end reason, the new row's parent_session_id, and its inherited _delegate_from), never the same parent or a nearby time, and delegate children under a segment do not count toward the lineage bound. A return whose links cannot be read links to no part: it is kept on the team as unlinked_returns and said by team_reconcile while a part is still out, and a part still out at the four-hour bound stops as return_not_linked instead of "never came back". A team asked for from a session whose earlier segments cannot be read is refused as team_lineage_unverified, with the reason, instead of team_not_found. The conversation identity the HUD, board lanes, cost receipts and fallback own records by now decides each link with the same predicates, so the five row shapes where it used to answer differently agree: a rotated delegate child keeps its earlier segment (and its route position), a stale sibling Hermes left beside a continuation under any segment is counted in the conversation, a reset fork is never a continuation whether or not it sits beside one, and a rotated branch keeps its inherited _branched_from lineage. Every segment of a conversation, the middle ones included, answers the same set. Past 64 rotations the identity, like the walkers, stops at 64 links and keeps what it verified up to there, so which end drops out depends on the segment asked from; it used to have no bound.

    • A delegation child's model verdict rests only on what the run recorded. A HUD row's model_attestation says whether the model that answered a child is the model it asked for. It had two ways to report a match it could not prove. A child whose session row names no model was judged against the current delegation.model, which OMH rewrites for each dispatch and puts back at turn end, so the same finished child read agreement or disagreement depending on when the HUD was read. And the request was read from sessions.model, which Hermes rewrites to the answering model on a session's first accounted call when that call names a billing provider, so a substitution on a child's first call read as agreement. The request now comes from the Model: line of the identity trailer Hermes stored with the child's first system prompt (requested_source: prompt_identity), written before its first call. Hermes rewrites that stored prompt from the running model, which after a fallback is the fallback, at an in-place compaction (its default) and at the start of a later turn, so the line is read only on a first segment with no archived message and at most one user message. Not covered: a row Hermes creates late (a retry after a transient create failure, or a re-create after the row disappeared mid-run) carries the in-memory prompt, and nothing in the row marks it. Without an attested request no agreement is reported. These cases read unknown, and the basis says which:

      • requested_model_not_attested: no stored first prompt names a model (none is stored, the row is a compression continuation, the schema lacks the columns, the read failed, or the trailer has no Model: line), and sessions.model agrees with the answering model.
      • requested_model_prompt_may_be_rebuilt: a first segment's stored prompt was withheld (an archived message, or more than one user message), and sessions.model agrees with the answering model.
      • no_requested_model: neither an attested stored prompt nor sessions.model names a model.

      In the first two cases, a sessions.model that disagrees with the sole model the usage rows name reads disagreement. The answering side is judged before any of the three: no usage row (no_usage_observation) or more than one usage model (multiple_usage_models) also reads unknown. delegation_route is no longer emitted. Measured read-only over every delegation child row in two local Hermes databases (a profile and a pre-update backup, 550 children; rows still only in the WAL were not read): 518 requests attested by the stored prompt, 32 withheld (28 compacted in place, 4 with more than one user message); 513 agreement, 30 unknown because the prompt may be rebuilt, 7 unknown with more than one usage model. Before the change the same rows read 543 agreement and 7 unknown.

    • Delegated work is drawn as a graph in the OMH TUI and omh hud --graph, and reported in chat replies at three moments. A conversation's delegate_task children, its omh_team parts, or a fanout's units are drawn as boxes with their title, category, model and effort, and state, joined by the edges OMH or Hermes recorded: a dispatch from parent to child, or a depends_on. Lanes one turn dispatched form a wave; waves are not drawn as dependencies, and a wave that settled more than 15 minutes ago leaves the drawing while newer work moves. Where boxes do not fit, or a shape cannot be drawn without implying an edge, the same graph is a compact tree. The OMH TUI draws it live above the plan, with two-line boxes where four-line ones do not fit its third of the screen, and /dag opens it whole in the pager. The [Agents] block stops repeating a native lane the graph draws, and its header then counts only the rows it still lists, so the lanes have one count line, also past the eight rows it lists. On Slack, Discord, Telegram, Mattermost and Matrix a reply never carries the boxes. The reply Hermes sends gets a code block naming the run (DAG · <name> · <N> tasks · <W> layers, then one line per layer) once when the run starts, a plain count line (DAG <name> · 1/3 done · 1 running) on a turn its counts changed, and a short block once when every lane has settled, or when nothing has moved for 15 minutes: each lane under Done, Failed or Stopped, the lane that failed first and the totals. Where a block and the reply do not fit one message, the count line goes in its place; where even that does not fit, a later reply carries it. OMH sends nothing itself. A category is shown only where the route OMH recorded matches the model the child ran, and no field reads current config: what has no record is ?. The structure comes from a new metadata-only record, delegation_ledger/v1, written from Hermes' subagent_start, post_tool_call and subagent_stop hooks; a conversation's record that no hook wrote for a day is removed. OMH_SUBAGENT_GRAPH=off turns it off everywhere. The TUI no longer draws the fanout-only DAG · N ready block; a fanout is drawn by the same view. A native child with no news for 30 minutes is unknown, an omh_team repair is repairing only while its child runs, and a stuck fanout unit (its dispatcher's unit_state) is drawn stuck. In the TUI the graph is in the theme's colours: edges step from violet to blue down the layers (dag_edge_start, dag_edge_mid and dag_edge_end, which a palette can name), a box's border shows its state, and every colour is measured against the terminal's background. It is painted at the depth the terminal takes (24-bit, 256 or 16 colours, none with NO_COLOR), and colour changes no character or width. A chat post is always plain text.

    • omh_delegate_route fallback advances only the calling conversation's own route (#2175). With two sessions on one profile, fallback took its chain position from whichever session had routed last: the newest provenance record, or the live delegation.* keys whenever OMH had written them. One session's route then decided another's fallback: an error naming a model the caller never routed, a fallback from the other session's model, an exhausted chain while the caller's own chain still had candidates, or the other session's live route advanced one step with the caller recorded as its writer. The position now comes from the newest route-leaving provenance record whose session_id belongs to the calling conversation, which is the host's session id plus its compression continuations, the set the HUD already owns route records with. The live keys count only when the restore record names a writer in that conversation and the file still holds that write. A conversation with no route of its own gets the existing no active route to fall back from error with position_source: "none", and nothing is written. A record that names no session is nobody's, so a call that reaches the tool without the host's session id gets that error too. Like set, a fallback still writes into the profile-wide keys, and it now does so from its own position. Where another session's route is live, the fallback that 4.0.0 refused with a category mismatch now replaces that route with the caller's next candidate, and when the caller's own chain is exhausted it clears that route, as the issue scopes. The leaked-route check before a baseline is captured still reads every session's records.

    • Every Hermes delegate_task child gets OMH's delegation protocol. The coding-handling harness requires the protocol in every delegated or reviewer prompt, runtime-native spawns included, but a Hermes-native child got it only when the parent model pasted it into the goal. The plugin now registers a second system prompt section, omh.delegation-protocol, which renders only for a session Hermes builds as a delegate_task child (platform subagent). It carries the goal echo-back, verification-stop and re-review-cap lines that omh coding composition-guide returns, which now read the same definitions. Only the goal echo-back applies to every child. The verification rule is framed for a task that changes files, and the review rule for a child that reviews work against stated success criteria. The section also reaches children no OMH-guided parent wrote, such as Hermes' own /review reviewer, whose brief from Hermes states no success criteria, so the review rule's condition does not match it. The section renders once in each child's system prompt. The coding-handling harness and omh coding composition-guide (its applies_to and its printed header) now tell the parent not to paste the protocol into a Hermes delegate_task goal, and still to embed it for Codex, Claude Code and other runtime-native spawns. A goal that pastes it anyway carries it twice. The parent's own prompt, the dispatched goals and the bot persona do not change. Each child's system prompt grows by 1,007 characters: the 920-character section plus Hermes' framing (#2174).

    • omh_probe says why its own package probe did not import, and a broken import hook cannot make the classifier raise. When omh.probe alone fails to import, the probe's bundle answer now carries a package_probe row with that failure, so the package_backend row, which reports the Loop backend, is no longer the only reason given. An import hook that raises for the name omh itself now reads broken instead of raising out of omh_loop, omh_decision_gate or omh_probe (#2166 follow-up).

    • A plugin-only install says why native Loop has no backend. Every documented install puts the oh-my-hermes package in OMH's own environment, so the interpreter Hermes runs usually cannot import it. omh_loop still refuses there with service_unavailable and reads or writes no Loop state, and now carries a package_backend block: absent when the interpreter has no omh package, incompatible when it finds an omh that lacks a module or name the bundle imports, and broken for any other failure while importing it, including one that is not an ImportError. A present omh never reads as absent, and the block names the plugin's own version beside the package's. The detail used to say "install oh-my-hermes or use the omh loop CLI". It now says the plugin is registered, that an environment Hermes' package manager prepares has no supported way to add the package yet, that an environment you manage yourself can take it, and that omh --omh-home <that home> loop status reaches the same Loop store. omh_decision_gate raised ModuleNotFoundError there for a payload of the required shape; it now answers unavailable with the same block and records nothing. omh_probe reports the same state as its own package_backend row, apart from registration, and answers from the bundle whenever the package's own probe cannot be imported, where it used to raise for a present but broken package. A sweep that called each of the 22 tools and 10 hooks on a plugin-only host, with one or more payloads each, found no other raise. The Installation page now lists native Loop, the decision gate, omh_team and omh_source_trust among the features that need the package, and no longer tells every reader to install it into Hermes' environment (#2166).

    • Bare omh opens the OMH TUI on Bun and npm installs (#2168). The npm and Bun launcher puts its extracted cache on Python's import path only inside its own process, and the TUI's interpreter, started -E -P by the frontend shim, could not import omh: bare omh said the OMH TUI could not import itself: No module named 'omh' and opened Hermes' own TUI. omh now hands the TUI the directory it imported its own omh package from, in OMH_TUI_SITE, and the shim's bootstrap puts that directory first on the import path when the interpreter does not already import omh from there. It then writes no bytecode: the launcher checks its cache against a digest on every run, and rebuilds a cache that has bytecode in it. -E -P stay, so Hermes' PYTHONPATH still never reaches the OMH interpreter, and omh sets the variable on every launch, so a value left in your environment does not choose the tree.

  10. 4.0.0

    Expandir

    GitHub Release

    v4.0.0 · Zeus

    Oh My Hermes 4.0.0 — Zeus ⚡

    Zeus brings three major changes to OMH: a dedicated terminal, long-term memory you can maintain through conversation, and specialist workflows connected to Hermes’ native capabilities.

    🖥️ A dedicated OMH terminal

    Work with Hermes through OMH’s own terminal interface. Follow tool calls, inspect diffs, check subagents, and steer a running task without leaving the conversation. Session pickers, history search, rewind, and clearer approval and question controls make longer sessions easier to navigate.

    Run omh for the OMH terminal or hermes for the stock Hermes interface. OMH’s plugin capabilities remain available in Hermes.

    🧠 Long-term memory, maintained through conversation

    Search, confirm, correct, and retire memories by talking to your agent. When Hermes’ built-in memory fills up, OMH can preserve approved copies in its long-term store before freeing native memory space.

    Retrieval prioritizes relevant context and current values. Memory support extends across TUI, Desktop, and supported messenger conversations, with user, channel, project, and profile boundaries governing what can be recalled.

    🔗 Specialist workflows, connected to native Hermes

    Bring domain-specific methods into research, planning, coding, design, and everyday work without maintaining a second set of controls.

    OMH workflows now use more of Hermes’ native capabilities: one plan checklist, source registration and citation checks, structured review results, Automation Blueprints, and recorded outputs from document and image tools.

    More powerful ULW workflows

    Set up subagent model routing with /omh-model in the OMH terminal. Choose the model and reasoning effort for each task category, then use ULW to research, plan, and run independent work in parallel.

    Upgrade

    omh update
    

    Restart or reload Hermes after updating. Launch with omh.

    The OMH terminal falls back to Hermes’ own interface on unsupported environments, including Windows. Updates release OMH-owned styling from the stock Hermes interface while preserving user-owned customizations.

    The complete version-bound changelog follows below.

    Full changelog

    • Indonesian ships as a trigger language pack. src/routing/trigger_packs/id.json, contributed by @fauzibrevy-hub (#2154), covers the 21 skills the ja and zh packs cover, with phrases such as tinjau kode, audit aksesibilitas, build gagal and lakukan riset mendalam. It is the first pack in a Latin-script language, and that shares the English token space: Indonesian borrows bug, crash, error and file. So a shipped pack phrase written entirely in ASCII now scores only as a whole phrase. Its words never join the skill's trigger tokens, and the lexical shortlist leaves it out of the skill's English document and anchors. Tests hold every skill's trigger tokens, lexical document and anchors equal to the view with its ASCII pack phrases taken out, and refuse an ASCII hold-back entry that none of the pack's own non-ASCII phrases carries. ASCII pack phrases merge after the ja, ko and zh ones, so SKILL.md descriptions and capability rows, which show only a skill's first triggers, keep the cues they showed before. Phrases that routed ordinary Indonesian chat (a team work plan, "ask me one at a time", a medicine reminder) did not ship, and neither did a few that read as word-for-word translations. ultrawork carries one Indonesian phrase, implementasi paralel: its always-loaded body had room for one under the byte ceiling. id is an input language only: it is not a --language / OMH_LANG output choice.

    • OMH TUI tool calls read as in Claude Code. Each call is one row, ⏺ read_file(README.md) 0.2s: a ⏺ in the accent while it runs, green once done, red when it failed or was denied, in place of the ✓ and ✗ marks, then the tool's name and what it ran on in parentheses. What it printed, and an edit's diff, hang under ⎿ , as a command's output does. A read or a load no longer echoes the first lines it read: it says Read 715 lines (read_file), loaded skill omh-docs (skill_view) or 12 matches (search_files). A resumed session's earlier calls read the same way, with a dim ⏺ (Hermes' record keeps no outcome) and an edit's diff under it, in place of the · read_file README.md rows with a blank row between each.

    • An answer you gave Hermes reads under the call it was about. The record of an approval or a question (✓ allowed once · <command>, ✓ blue · Which colour?) was written above the turn's live tool calls, so with one row per call it stood over the first read of the turn. It is now the first row under the call that asked, ⎿ ✓ allowed once · <command>, on the same green or red band; a denial reads ⏺ terminal(…), then ⎿ ✗ denied · …, then Hermes' denial line. With no call running (a subagent's question after its turn), it is still a row of its own. A clarify question Hermes timed out, or one you cancelled, says so once under its call (⎿ ✗ timed out · Which colour?), not again in a bare ⎿ timed out row under that.

    • An OMH TUI diff no longer doubles each changed line. Hermes pads every added and removed line of an edit's diff with spaces, to 143 columns in a live patch, so on any narrower terminal each green or red band wrapped into a second, empty band row. A diff line's trailing spaces are dropped now; the band still reaches the right edge.

    • The OMH TUI's working row reads as Claude Code's. It was a braille spinner and Hermes' kaomoji text (⠋ (⌐■_■) analyzing... (12s)). It is now ✶ Analyzing… (12s · ↓1.2k tokens · thinking with high effort · esc to interrupt): a glyph turning through · ✢ ✳ ✶ ✻ ✽ (still under reduced motion), Hermes' verb without its face or dots (Working… when it gives none; a stall or retry notice, No response from provider in 120s — reconnecting…, keeps its words), and the effort when the session reports one. A tip naming a key sits under it, one a turn, on a terminal 16 rows tall or more (⎿ Tip: ctrl+o shows the whole transcript); it is local text and nothing is sent. interrupting… lines up under the same glyph column.

    • Your prompt and a turn's end read as in Claude Code. A prompt in the OMH TUI transcript is a dim > and dim text, with no tinted band behind it, and a command's echo takes the same > (> /model). The ↳ you: … reminder over a running turn loses its band too: ↳ you: in the accent, your words dim. A turn's done line is one dim line under the reply, ✻ 5s · 3 tools · ↓1.3k, without the rule that ran to the right edge.

    • The OMH TUI's composer is a rounded box over one dim info line, as in Claude Code. The composer's two rules, the mode row and the dark status band are gone. The composer sits in a box (╭─╮, │ > … │, ╰─╯) that keeps its colour in every mode, and the line under it, two cells in under the box's text, says the mode or the keys that matter now on the left (? for shortcuts while the composer is empty, ⏵⏵ accept all on (⇧⇥ to cycle) in accept's colour, ⏸ plan mode on (⇧⇥ to cycle) in plan's, a question's or a picker's keys) and the model, ctx 5% and effort medium on the right, with the OMH HUD's repeat and in-flight tool counts first when there are any. The ⚚ OMH brand and the idle / running words are gone: the working row, the strip and the queue rows already say them, and during a turn the left names only the steer keys, since the working row already offers esc to interrupt. On a narrow terminal the right side gives way first, the model shortening to deepseek-v4.1-flash and then v4.1-flash before it goes. The line is on every terminal of 3 rows or more, so the mode is never hidden, and the agent list now opens from 6 rows (and 20 columns) instead of 10. A ! shell line keeps its ! shell label, and its > mark turns the accent colour.

    • The OMH TUI opens on a clean screen. As it starts, what your terminal showed above the cursor (your shell, the warnings hermes printed) scrolls up into the terminal's own scrollback, and the TUI starts on the top row. Nothing is erased: scroll up to read it, until a full redraw (a resize, Ctrl+L, /redraw) clears the scrollback. A launch that falls back to Hermes' own terminal leaves the screen as it was.

    • Hermes' note about a skill named like a built-in command is no longer an OMH TUI notice at every start. Hermes' own bundled plan skill shares its name with /plan, so every start said Hermes' command catalog: slash command /plan unavailable — name taken by built-in; use /skill plan. /help now lists those notes, in Hermes' words, under the skill count; a real discovery failure (skills, quick commands, plugins) is still a notice.

    • omh handoff-risk-scan --repo completes on ordinary repositories. The scan charged the raw bytes of every Git listing to its 1 MiB metadata budget, and the index listing with its stat lines alone prints about 190 bytes per short-path entry, so a clean repository of 4,500 short-path files, or 6,538 files with 59-byte paths, ended in scan_error / metadata_oversized, far under the documented 10000 entries (#2157, reported by @jhlee0409). The budget now charges what the scan keeps: the HEAD and branch output, each committed path with its mode and object id, each index path the commit lacks, and the untracked listing; a path in both the commit and the index is kept and counted once. The 10000-entry and 1 MiB limits are unchanged. The tree and index listings are read in 64 KiB pieces up to a fixed cap derived from them, and a longer listing still kills the Git process and ends the scan with metadata_oversized. Git output the scanner failed to read is a scan error, never an empty listing.

    • A party named before a Korean data-share place, or after an English one, keeps the review. Korean puts the data, and any party, before the place, and only the place was read: "보험사가 원하는 고객 엑셀 개인정보 마스킹해서 개발팀에 보내줘" read inside and named no review. Every word outside the places and before the requested verb must now be a closed word, as English already reads the span before a send: data containers and field names in either script, a verb the rule knows, an inside place word, a determiner or a softener, and a word that says whose records the data holds (고객, 회원, 환자, ...) only bare and right before a data word. A party acting ("보험사가 원하는"), owning ("보험사의", "고객의"), a purpose ("협력사용", "보험사를 위한") or any other noun ("삼성 고객 파일") keeps the review, and "-기 전에" is a step rather than a place. Korean also read only the first sentence; a later one now counts when it sends, names a recipient or a place outside, or names a party acting ("... 개발팀에 보내줘. 벤더한테도 보내줘.", "... 보험사가 볼 거야."), and remarks alone ("급해요.") say nothing. In English a place now ends with its sentence (".", "!", "?" or ";" before a space), so "... send it to the dev team. It's urgent." names no review, while every sentence is still read for sends and the words after a place must be remarks -- thanks, please, urgency -- so "... to the dev team. Acme will review it." and "... Also cc Acme." keep it. A line break alone ends no sentence, because the router reads it as a space; a period before a space ends one even after an abbreviation ("Acme Inc. today", "the U.S. team"), which only cuts a place short. Between a send and its "to" only closed words may stand, read as the span before the first send is: the data, also as protected, counted or the version the redaction made ("the anonymized copy", "both files", "the updated file", "the customer export"), the verb's particle, "straight" or "directly", and remarks; so "send it where Acme wants it, to the dev team" and "send the Acme copy to the dev team" keep the review, and so does any other word there ("send the sorted export to the dev team"), the safe direction. The Korean reader no longer opens a quote at an apostrophe after a Latin letter or digit, nor closes one before a Latin letter, so the router and the hint read "고객 정보 마스킹해서 Acme's team에 보내줘. It's urgent." alike; a quote a Korean particle follows ("'고객 정보'를") is still a quote. Measured against main on sets written before the change (104 Korean outside rows, NFD copies included; 80 Korean inside controls; 57 English sentence-end rows) and every earlier data-share set: no request to an outside party lost the review on admission, router offers or the hint, and the router and the hint never disagree with every cache cleared (three disagreements on main). Korean outside rows gained the review on admission and the hint from 16 to 98 of 104, and on router offers from 2 to 26; the six left are two that name a regulator, which reads as a legal question, and one that names no personal data, each with its NFD copy. Korean requests sent on in a later sentence gained it from 0 to 12 of 12. English inside requests followed by a remark stopped firing (20 to 5 of 20), and no English inside request on these sets or the review's starts firing. Of 196 Korean inside rows that read inside on main, 8 now fire, the safe direction: an order history, survey responses, "this month's" orders, and "고객의" (a customer's), each with its NFD copy. On the review's own sets no outside request lost the review either, and 19 more Korean inside rows now fire for the same kinds of word: time words ("이번 달", "지난달", "어제 받은", "오늘 안에"), descriptors ("주문 내역", "거래 내역", "설문 응답", "가입 정보"), an inside team acting ("우리 팀이 만든", "개발팀이 요청한"), and "고객의".

    • A plural possessive no longer hides the rest of a message from routing. The routing masker read the ' of "the customers' export" as an opening quote. Nothing closed it, so everything after it was masked and no matcher read the verb or the place: "Redact the PII in the customers' export and send it to Acme." got no data share review from the router or the hint, and in "explain the users' '$ulw-work' snippet" the possessive paired with the next quote and left $ulw-work to dispatch. A ' that ends a word now opens no quote: one right after a letter or digit that is not East-Asian wide (CJK, kana, Hangul or fullwidth), and right before whitespace, one of .,;:!?)], an East-Asian wide character or the end of the message. It still closes one that is open. A ' after a space or an East-Asian wide character (고객' list), or glued to more text (f'$x', API''), still opens one. The data share review also reads ’ between two letters ("it’s", as Mac and iOS type it) as an apostrophe, as it reads ', not as a quote.

    • Two OMH TUI subagents whose goals start alike read differently on the strip. The strip names an agent by the first words of its goal, and two goals that began the same were numbered with those words, so two shell tasks read 1·Run the shell 2·Run the shell and nothing said which was which. Agents that share a name now show the word where each goal parts from the one nearest it, out of any quotes or backticks around it (with the next word when it is one like in), so goals that part at `echo alpha` and `echo beta` read …alpha and …beta; the numbers stay only for goals whose words still read the same.

    • A fresh OMH TUI session no longer shows the last session's subagents. Hermes writes a session's state.db row on its first prompt, so a new session's key mapped to nothing until then, and the HUD reader fell back to every conversation's recent delegate children: a session nobody had typed in yet opened with [Agents] │ [global] 2 done and the previous session's rows. The TUI's own key now owns no children until its row exists, and the panel stays empty. Other readers that name an unmapped reference (the desktop pane, the plugin tools) keep the documented global fallback.

    • Denying an OMH TUI approval also stops the turn, as Esc does in Claude Code. A deny went to Hermes and the turn went on, so the model often tried the next thing straight away. Every deny the person gives (Esc, Ctrl+C, d, n, Enter on deny, and Up then Enter typed ahead) now sends the explicit deny first, so the waiting tool ends denied, and then interrupts the running turn (session.interrupt, the call Esc makes on an empty composer); with no turn running only the deny goes, and the mode row then says d or esc denies. An approval a subagent asked for stops the session's whole turn. The tell row is the one deny that does not stop: it redirects the running turn instead. The mode row says so in plain words, ↑↓ pick · enter confirms · d or esc denies and stops, down to esc denies and stops on a narrow row.

    • An OMH TUI approval can be denied with what to do instead. Telling Hermes what to do after a deny took a deny, then a separate Ctrl+Enter steer. The approval's last row, no, and tell Hermes what to do instead (last, as Claude Code puts its "No, and tell Claude" row), sends the explicit deny at once, so the turn is not held, and opens the composer labelled ↪ tell Hermes what to do instead · enter sends · esc cancels. What you send goes into the running turn as a steer through the same path as Ctrl+Enter (a refused steer waits in the queue, an unanswered one comes back after 120 s), or as an ordinary prompt once the turn has ended, which waits behind a prompt you queued and never steers it; the tell row of a later approval keeps that composer open with what you typed. Esc sends nothing more and brings your draft back. Deny stays first and highlighted, the highlight still stops at both ends, and no digit or letter but d and n answers.

    • An OMH TUI approval offers Hermes' "always allow" when Hermes does. The OMH TUI left Hermes' always choice to hermes --tui. When Hermes lists it (it leaves it out where a permanent allow is refused, such as a command its smart approval flagged), the approval now has a fourth row under allow this session, labelled with what it allows, always allow recursive delete (Hermes saves it): Hermes saves that kind of command in its config (command_allowlist), so later ones run without asking. Like every allow it takes the arrows and Enter; the transcript records ✓ always allow sent (Hermes saves it) · <command>.

    • Every OMH TUI question asks for your attention, not only an approval. An approval rang the bell; a clarify question, a sudo password or a secret signalled nothing, and Hermes drops a clarify question nobody answers after 300 s by default. Each request now rings the bell once as it opens and, on a terminal that shows them, sends the notification a long turn's end sends (OSC 9 in iTerm2, WezTerm and Ghostty, OSC 777 in foot and rxvt), saying what Hermes asks (Hermes asks: Which colour?, never a value). OMH_TUI_BELL=0 keeps all of it quiet.

    • The OMH TUI stops billing your answer time to Hermes. While an approval or a clarify question was open the working row kept spinning, ⠼ working… (16s · ↓262 tokens), beside the status line's waiting on you; a command you took 20 s to allow then read ✗ terminal rm -rf ./tmp-a 21.8s, as if it were slow. The working row now reads ? waiting on you (11s) with no spinner and one clock, that wait's (the OMH tool count beside the model shows no seconds of its own meanwhile, and leaves with the turn instead of lingering into idle); a tool's row leaves the wait out of its time; and the done line keeps the turn's wall clock with the wait beside it, ✻ 30s (27s waiting on you) · 1 tool.

    • The OMH TUI's transcript records how each question ended. A clarify question left only Hermes' tool row (✓ clarify Which colour should I use? 2.5s), a ✓ even when it was cancelled, and the answer appeared nowhere but in the model's reply; a sudo or secret prompt left nothing. Like an approval's ✓ allowed once / ✗ denied, the answer now goes above the call it was about: ✓ red · Which colour? (ticked labels joined, or what was typed, as Hermes reads the answer: without the (Recommended) tag it shows on a choice, and with empty items of a multi-select answer left out), ✗ cancelled · <question> or ✗ skipped (also a multi-select answer with no item left), and for sudo or a secret only ✓ answered or ✗ declined, never the value. A question Hermes withdrew reads ✗ timed out, ✗ interrupted or ✗ withdrawn (<reason>) with what it asked, in place of the amber Hermes withdrew its <method> request (<reason>). A clarify call that got no answer is a red ✗ clarify row with cancelled, timed out or not delivered under it, not a ✓.

    • OMH TUI clarify questions that take several choices are answered with several. Hermes' multi_select was never read: the question looked like any other and went back as one label, and an answer typed as 1,3 reached the model as one string. Such a question now says (choose one or more) and boxes each choice; Space or the choice's digit ticks it, Enter sends the ticked ones (the highlighted one alone when none is ticked), and an answer of your own joins them. The answer is the JSON list Hermes' Ink TUI sends (["ham","cheese"], in the order ticked), in the same answers frame as before; a single-select question is unchanged.

    • An OMH TUI question says when others wait behind it. A second request that arrived while one was open queued with nothing on screen to say so; the open question now ends its first words with · 1 more question waiting (or the count). A request the TUI cannot show (a terminal read, a vault prompt) that arrived behind an open question waited for it to close before its decline was sent; it is declined at once, with its notice, as it is when nothing is open.

    • OMH TUI clarify choices wrap instead of being cut to one row. At 80 columns ❯ 1. Blue/green with a full traffic switch after smoke tests pass on the gree… lost the (Recommended) tag Hermes puts at the end of the first choice, and the rest of a choice up to Hermes' 8000 characters. Each choice (an approval's too) now wraps under its number, the slot counts rows rather than choices, and a terminal too short for every choice keeps the highlighted one whole, or its first rows, with … ctrl+o shows all. Choices it cannot list are marked beside the list on the side they are cut (… +3 more choices, 1 ticked · ctrl+o shows all; an approval's … more choices, which numbers nothing), so a cut tell row or a ticked choice scrolled away is never silent, and the marker no longer ends the question's text.

    • An OMH TUI question you cannot see whole says so, and Ctrl+O shows it. A long approval command showed its first 6 of 20 lines at 24 rows, and Ctrl+O did nothing while the approval was open; below 10 rows a cut had no marker, at 6 rows no command row was drawn at all, a narrow terminal cut the approval's header, and under 6 rows or 20 columns a sudo, secret or typed clarify question showed no label. Every cut now ends with … ctrl+o shows all (… +14 more lines · ctrl+o shows all under a command's first lines), from 6 rows a command's first row is drawn, the header and a value prompt's label wrap, a tiny terminal keeps one row for a typed question's label, and Ctrl+O while a question is open pages it whole: it answers nothing, and q or Esc returns to the question with the same row highlighted.

    • Enter on an OMH TUI / or @ menu takes the highlighted row, as in Claude Code. A command typed in full still runs on the first Enter. Esc closes the / menu, and both menus stop at their ends.

    • The OMH TUI's Ctrl+R search and agent list stand where the composer was, as the pickers do: the composer is hidden, the search's caret is at the end of its query, the agent list takes every key (only its own act, so typing never turns into a steer; Esc, Left, q or Ctrl+C leaves it), and the mode row says their keys.

    • An interrupted OMH TUI turn says so until Hermes stops, and the queue waits. After Esc or Ctrl+C the working row says interrupting… until the turn ends, and the queue does not start: queued prompts come back into the composer unsent, a queued ! line waits until your next prompt.

    • The OMH TUI's /model picker opens on the session's model. It marks that model (first, when no provider lists it) and shows Model (12/162) once it scrolls.

    • OMH TUI list commands leave no echo in scrollback. Bare /sessions, /rewind, /keys, /history, /logs and /redraw leave no > /name row.

    • New OMH TUI keys. Ctrl+X opens the session picker, Ctrl+_ and Ctrl+Y undo and redo in the composer, and Ctrl+Z suspends to the shell on POSIX (Hermes' gateway stops with it; fg draws it again under what the shell printed). Where Shift+Enter would send, the mode row names the newline keys while a draft spans lines.

    • /omh-model no longer stops the OMH TUI's gateway. A reply keeps streaming underneath and a question closes the picker with nothing saved; the picker stops at its ends and says esc cancel.

    • A crashed Hermes gateway comes back. When Hermes' tui_gateway exits with a failure, the OMH TUI starts it again from the launcher's environment and resumes the session, saying so (The Hermes gateway exited (code 1): starting it again and resuming session <id> (1 of 3 in 60 s); what you type waits until it is back, and ctrl+c leaves.), at most three times a minute. Each attempt waits up to 15 s for the new gateway (HERMES_TUI_STARTUP_TIMEOUT_MS when set, as in Hermes' own TUI), Ctrl+C or Ctrl+D while it waits leaves with the new gateway stopped, and one that exits before it resumed the session uses an attempt, as in Hermes' own TUI. An open question closes unanswered, a running turn is reported lost and not sent again, a steer the old gateway had not answered, and what was typed in an approval's tell field, come back to the composer at once, the subagents on the strip stay there (those still running marked interrupted) rather than reappearing as look-alike rows in the HUD's [Agents] block, and what you queued comes back as after an interrupt. AGENTS.md and CONTEXT.md now say the frontend spawns one child at a time and respawns it so.

    • The OMH TUI's composer sits right under the output until the screen fills, then stays on the bottom rows. A fresh session showed the banner, then 20 blank rows (110x34) to 36 (200x50), then the composer, and 6 to 12 blank rows stayed above it after a short turn, during an approval and after a resize: the live region was padded to the screen's height at every start and redraw and never got shorter. It is now drawn right under the last line printed (one blank row under a turn, the spacer every block opens with) and follows the content until its last row is the screen's last row; from then on it keeps its height, so a list or question closing above the composer leaves blank rows above it, filled by the next output, and does not pull it up. A resize or a redraw starts again under the output. Short of the bottom rows the / and @ menus open under the composer where they fit, as in Claude Code, so typing / no longer moves the composer down by the menu's height and Esc back up; on the bottom rows they open above it, and output that brings the composer there while one is open takes it above with that output, so closing it moves nothing. The startup query that asks for the background colour also asks for the cursor's row (CSI 6n, inside the same 0.1 s wait), so a start from the last row of a terminal already holding output is on the bottom rows from its first frame; only a terminal that does not answer that cursor-position query follows the content until the TUI's own output has filled the screen once. Lines that leave the top of the streaming tail are written over the rows they were drawn on, so the rows under them are not painted again, and while the screen is not yet full a new tail line is inserted where it goes (IL), so the composer and the rows under it move down without being painted again (a reply streaming onto an empty 40-row screen writes 232 bytes a frame, where following the content had doubled it to 605).

    • Each tool call is its own row, and a cut output names the key that shows the rest. A turn's calls were folded under one ⏺ Tool calls (2) header; each call is now a row of its own (its mark, name and what it ran on) with what it printed under it, as in Claude Code, and the parallel ×K badge that header carried is gone with it. A tool's output, a diff and a command's output cut short end … +N lines · ctrl+o shows all (rows for a command), where they named /history or nothing: Ctrl+O opens the details view at its end.

    • A finished plan leaves the OMH TUI's panel. The plan stayed above the composer into later, unrelated turns. Once every item is done it now shows ✓ 3/3 until the turn ends and then goes, as Claude Code's list does, with its final list written to the transcript above the done line under the header the panel showed ([Plan] │ ✓ 3/3, as /todos now heads it too); one already finished when the session opens is not drawn, and a new plan is. The [Plan] │ 2/6 . dot was the header's walking ellipsis, caught at one dot; it is gone, and the active item's colour wave still shows work moving.

    • The OMH TUI states the approval mode once. After Shift+Tab the mode row said ⏵⏵ accept all on · approvals off for this session while the status line said yolo mode: off: that chip read Hermes' saved state through the HUD, which lags behind a switch. The status line no longer carries it; the composer's rules and mode row (the status line on a short terminal) say the mode, as Hermes confirmed it for the session, in the words the docs use. The mode row also says what each mode means and what Shift+Tab turns on next, Claude Code style: between turns ? shortcuts · ⇧⇥ accept all/plan · ctrl+t effort (medium) (it said ⇧⇥ mode · ctrl+t effort with a separate effort medium on the right), in accept all ⏵⏵ accept all on · Hermes asks no approvals in this session (⇧⇥ plan/normal), in plan ⏸ plan mode on · prompts go as /plan (⇧⇥ normal/accept all), with shorter forms of the same words on a narrow row. The effort is said once, as ctrl+t effort (medium), and the keys list uses the same words for Shift+Tab and Ctrl+T. It names the level Hermes' config sets from the start of each session, /new and switches included (read from the settings the TUI already reads), until Hermes names the session's own once it has built the session's agent, and where none is configured it says ctrl+t effort (default) (the provider's own default applies) instead of no level at all; Ctrl+T steps from the level shown.

    • A first Ctrl+C between turns says ctrl+c again to exit on the mode row for the 2 s a second press exits in, as Esc's first press does, instead of leaving a Ctrl+C again to exit notice in the scrollback every time.

    • The OMH TUI shows three things Hermes sends it and it dropped: what the background review saved to memory or skills (review.summary, a transcript line), a standing /goal (session.control.update, one row above the composer worded as Hermes' Ink goal bar: ⊙ goal · 3/20 turns · <title>, parked or paused with its reason), and a mixture-of-agents turn (moa.*: each reference model's answer as a collapsed ◇ Reference 1/3 — <model> row, and MoA: refs 2/3 then MoA: aggregating… on the working row). The gateway already wrote these to the TUI; showing them sends nothing.

    • The status line follows a light terminal. It kept the palette's dark bar on a light screen (sky's #0A2229 under white). On a background the terminal reports light (or COLORFGBG names, or /theme light) the bar is now a light tint of the palette's own bar colour, its words still 4.5:1 and its │ dividers 3:1 against it; a dark terminal, and one that reports nothing, keep the palette's bar.

    • The keys list names the /model picker beside the /sessions and /rewind lists, which take the same keys (?, /keys, and INSTALLATION's table).

    • A team inside the company is internal for data-share-review, in English as in Korean, and only when it is proven inside. Owner decision (2026-10-16): sending personal data to a team in the same company is internal, so the review does not fire; it fires only when the data leaves for an outside party (a vendor, provider, partner, contractor, third party, or an external service). English admission read any verb that sends a file as the file leaving, wherever it went: "redact pii in the export before we send it to the dev team" dispatched the review. The gate fails safe: a place is inside only when every word of it is an inside word, in closed tables both languages read -- an anchor (a teammate, a colleague or coworker, internal, in-house, QA, staging, a test or dev server; 사내, 내부, 동료, 스테이징), a team or department with an owner (our, my, another, or an internal function such as data, dev, marketing, finance, hr; 우리, 다른, 개발, 재무, 인사), a determiner, or a system beside an anchor (the internal wiki, the staging server; 사내 위키). Any other word keeps the review: a company ("the Acme team", "Acme's data team"), a client or customer ("the customer success team", "고객지원팀"), an organisation ("the Department of Health", "은행 담당 부서에", "다른 회사 동료에게"), a relative clause ("a former colleague who now works at Acme"), a place after "at", "of" or "for" ("the data team at Acme", "the dev team for review"), a team with no owner ("the team", "팀 전체에"), and Hangul in an English place. In English, only closed words -- where the data is, a determiner, the speaker, data containers and personal-data field names -- may stand from the lead verb to the first send, its object included ("the export Acme will get", "the CSV the bank wants" and "before Acme sees it" keep the review). A word that names whose records the data holds (customer, user, patient, client, member, employee, staff, student, donor, applicant, subscriber, account) may stand there only right before a data container or a field name, not possessive and not after "for", "to" or "from": "the customer export", "the patient CSV" and "the user logs" describe the data, while "the customer's export" and "the export for the customer" keep the review; every later send must name such a place after its "to" or "with", and so must every place it joins ("the data team and Acme", "cc the auditor"); "give", "hand" and "pass" without "to" send nothing to the person and otherwise take their first object as the place ("give Acme a copy" keeps the review). A Korean place is read back, however many words, to a step or a joined form of a verb the rule knows, so a noun ending in 서 or 고 is read on ("경찰서 보안팀에" keeps the review); every place in the sentence must be inside, and an English recipient word ("vendor team에") keeps the review. Both surfaces mask quoted, code, link and relayed text before they call the rule, and each passes the masker's own report that it removed something, so no place is then proven inside ('the "Acme" data team', "the dev team https://acme.com"). Measured against main on 160 requests written for this change, 140 written blind by a separate agent, and the reviewer's sets written before reading the code (mine2.tsv and its NFD copy, adv.tsv to adv4.tsv, r3.tsv, r3b.tsv): on these sets no request to an outside party lost the review on admission, the router or the hint, and the two surfaces never disagree with every cache cleared. Korean outside requests gained it (29 to 40 of 40, 32 to 34 of 35, 34 to 53 of 53, 10 to 31 of 31, 1 to 18 of 18), and 47 adversarial rows written for the compound-modifier rule ("the customer export the regulator wants", "the patient CSV for the insurer", "the customer's export") lost nothing. Inside requests that still fire, the safe direction: English 2 of 36, 12 of 35 and 14 of the reviewer's 31 (each for where the data goes: a manager, HR, the intranet, a bare "engineering", "in Seoul"); Korean 2 of 36, 14 of 35 and 5 of 20. The three requests the decision named without a file ("redact the pii and send it to the dev team") were never admitted, since English admission needs a file-shaped carrier; they are pinned as negative controls. Every table is pinned by a literal snapshot, and a list of outside words must keep the review. Files and folders stay inside only in Korean ("엑셀로 내보내줘").

    • Korean spend and data-share requests reach their gates. "내 카드로 결제해줘", "이거 결제해줘" and "송금해줘" fell back with score 0, "고객 정보 마스킹해서 벤더한테 보내줘" dispatched gateway-intent-card, and "개인정보 지우고 외부에 공유해줘" asked about automation-blueprint: both gates' admission rules read only an English lead. A Korean request puts its verb last, so a new bundle reader (korean_request) reads the request clause's last predicate, and only when a request ending closes it ("-해줘", "-해 주세요", "-해줄래?", "부탁해"), the way "please" marks an English one. It skips a greeting or an address before the request ("안녕하세요.", "헤르메스,"), removes a label's text ("버튼 라벨: ...") and quoted text, and drops what comes before a negation it knows (말고, 않고, 마시고, 마세요, 마요, 없이): "카드로 결제하지 말고 견적만 보내줘" asks for the quote. spend-gate admits it when that predicate spends (결제, 송금, 이체, 구매, 주문, 예매, "사줘", ..., but not "결제 처리해줘"); books, tops up, subscribes, renews or donates something one pays for with a price, money, the person's wallet or account, or a card as the means in its own clause ("카드로 호텔 예약해줘", not "가격 5000원으로 갱신해줘" or "제 카드 정보 갱신해 주세요"); or sends money as its object ("엄마한테 10만원 보내줘", not "결제 금액 10,000원 영수증 이메일로 보내줘"). data-share-review admits it when it protects personal data (익명화, 가명 처리, 비식별화, or masking, deleting or covering personal data) and the data leaves: a recipient named as where it goes ("벤더한테", "외부에"; not "업체 목록에서" or "업체별로"), or a verb that sends it to no team, file or folder of the person's own ("스테이징에", "개발팀에", "엑셀로"). Each gate scores only its own phrases, which an admitted request always says, with no boost: one phrase scores below the dispatch line, so the router asks, as it does for "please send money", and a second one (the person's card, a recipient) dispatches the gate unless another skill outscores it. The Korean phrases live in the bundle beside the English ones, since the rules read them, and "예약" and "보내" are declared as shared with automation-blueprint. Measured on seven sets of everyday-sense Korean negatives (90, 80 and 42 of my own sets, and a reviewer's 82, 34, 75 and 27), no sentence names either gate on either surface. Recall is partial: on a set written blind by a separate author and first measured at an earlier version of the rule, the router offers 23 of 40 spend requests and 19 of 40 data-share requests (the hint 24 and 21). Every set has since been read while the rule was revised, so none is held out for this version.

    • Korean diagram, LLM-wiki and token/cost requests stop going to the wrong skill. codebase-uml's diagram check compared the composed words "다이어그램" and "시각화" with the router's NFKD-folded tokens, so no Korean diagram request reached it: "코드베이스 다이어그램 그려줘" asked about codebase-onboarding while the in-session hint named codebase-uml. Both words are folded before they are compared and read as prefixes, because Korean attaches the particle ("다이어그램이"). On the router the English boundaries hold in Korean too: 클라우드, 인프라, 쿠버네티스 and 서브넷, bare or with a particle ("클라우드의"), mark an infrastructure diagram that is the bundled architecture-diagram skill's job unless a Korean or English code word is also there. Because the check now runs, "시각화해줘" joins the skill's Korean whole-phrase-only tokens: the trigger "코드베이스 시각화해줘" made it a scored token of its own, which would have named the skill for "매출 데이터 시각화해줘". "LLM 위키" and "LLM위키" now name the pattern Hermes' bundled llm-wiki skill builds, as "llm wiki" did: the wiki blueprint withdraws, and so does the greenfield interview, which Korean, exempt from the weak-evidence check, dispatched on that guard alone. English llm-wiki requests still clarify; their leading candidate moves from deep-interview to the next scored skill. ops-observability-card loses its bare "비용" and "토큰" triggers, which dispatched it for "토큰 저장 방식 바꿔줘" and "이 API 호출 비용 줄이는 코드 짜줘"; its Korean trigger count drops from 9 to 7. The card's guard keeps them as telemetry words, as it keeps "token" and "cost", but no longer reads "알려줘" as a status request, a word its English list never had. Korean token and cost usage is read from "토큰 사용량", "토큰 사용 추이", "토큰 소모(량)", "토큰 소비(량)" and "비용 모니터링", each as a whole word ("토큰 소비" is not "토큰 소비자"), and only when the sentence's final predicate asks to see the usage -- show, tell, check, compare, monitor, track, how much, how is it, a report, or "is a lot" / "is fast" -- or the sentence ends on the phrase itself ("이번 주 토큰 사용량"). The final predicate is the last word, or the word before it when the last word is the auxiliary written apart ("보여 주세요", "알려 줘", "확인해 주세요") or a spaced ending matched whole ("모니터링 해줘", "추적 해 줘", "얼마 나왔어"), so 해제해줘 and 하지마 are predicates of their own. Every other final predicate leaves the phrase uncounted: work (만들어줘, 추가해줘, 고쳐줘, 파싱해줘, 작성해줘, 짜줘, 줄여줘), so "토큰 사용량 API 만들어줘" and "토큰 사용량 버그 고쳐줘" stay on ultrawork, and any other question shape too. A modifier earlier in the sentence ("기능별 토큰 사용량 알려줘") does not stop a usage question. The guard's English-shaped token arm is unchanged: 토큰 or 비용 beside 보여줘, 상태 or 요약 still opens the card, as on main ("토큰 사용량 제한 로직 보여줘", "토큰 사용량 요약해줘"). So does a how-to or a check under a report predicate ("토큰 사용량 파싱하는 방법 알려줘", "토큰 사용량 버그 확인해줘"), as on main, where English asks or answers directly. Usage requests whose final predicate is outside the report set no longer reach the card; this is a class, not a list. Measured examples that dispatched it on main and now ask: "토큰 사용량 어떻게 돼?", "토큰 사용량 얼마나 돼?", "토큰 사용량 궁금해", "토큰 사용량 확인 가능해?", "토큰 소모 심하네", "에이전트 비용 추적해줘", "오늘 LLM 비용 얼마 나왔어?", "이번 달 API 비용 얼마야", "이번 달 비용 얼마 나왔어" and "토큰 얼마나 썼어?"; "토큰 사용량 리포트 만들어줘" dispatches the greenfield interview. English asks or answers directly on the same shapes ("how is my token usage?", "how much did the API cost this month"), and questions about what was spent belong to Hermes' /usage and /insights per the card's Do-Not-Use. Measured with build_chat_interaction_payload (source="discord"), awareness_route_hint and the skill candidate line at main and at this head, over both corpora (unchanged except three English llm-wiki clarifies whose leading candidate moved), the triage probes, 134 Korean probes of our own and the review's probe sets: Korean diagram requests that dispatch codebase-uml go from 1 to 5 of 25, plus 4 that now ask with it leading, and no Korean diagram negative moves toward it; of 21 LLM-wiki requests, which went to wiki (18) or dispatched the interview (3), 20 now ask and one dispatches materials-package on "PDF"; ops-observability-card dispatches none of our 24 everyday-sense negatives (20 before), 4 of the review's 52 code and everyday requests (44 before; the 4 dispatched it on main too), 22 of our 24 usage requests and 31 of the review's 36 (all before). A held-out set of 48, written after the first fixes were chosen, moved the same way: codebase-uml dispatches 0 -> 4 of 8, wiki reaches 7 -> 0 of 7 LLM-wiki requests, and ops-observability-card dispatches 9 -> 0 of 10 negatives and 7 of 8 usage requests (8 before). Adds 21 negative controls and 29 interventions.

    • The OMH TUI catches up with Claude Code and oh-my-pi. It reads on a light terminal: it asks the terminal for its background colour before the first frame (OSC 11, at most 0.1 s, with COLORFGBG as the fallback) and measures its theme against it, so prompts and replies are the terminal's own text colour, every coloured word reaches 4.5:1 and every rule or mark 3:1 on that background (the status line's against its own bar), words that sat on the 3:1 decoration colour moved to colours that read, and a light background gets light diff bands (on white the body text measured 1.08 to 1.13:1 before); /theme auto|light|dark picks the background for the open session. A slash command is echoed as > /name arg and what it prints stays in the transcript under it, a long output as its first 12 rows and a count of the rest for /history; nothing a command prints opens the pager any more, so typed-ahead keys are no longer swallowed and the output is still in scrollback afterwards. Hermes' own notices (notification.show: the agent still starting, credit warnings) show on one row above the composer, replaced by the next, withdrawn by key, or gone after their time, where they were dropped before; and a refused call reads as words (/undo on an empty session: No user messages to undo. instead of command.dispatch failed (4018): no user messages to undo). Agents show in one place: the HUD panel's [Agents] block steps aside while the strip and the list hold agents (the panel said "2 running" minutes after both had finished), agents whose goals start alike are numbered on the strip, and the status line says idle · 2 agents running while agents outlive the turn. Mouse reports are on only while the agent list is open, so during a turn the wheel scrolls and a drag selects; the strip is reached with Down. Up and the new Ctrl+R search reach earlier sessions' prompts: the TUI reads Hermes' own input history ($HERMES_HOME/.hermes_history, the file hermes --tui keeps) and appends each line you send to it in that file's format, opened append-only and without following a symlink, created mode 0600. Esc twice clears a draft into the session's history, or, on an empty composer between turns (and /rewind), lists your prompts: the one you take comes back to edit, and sending it truncates the session before it (prompt.submit with truncate_before_row_id and confirm_truncate) and redraws from the cut. A line that starts with ! runs in Hermes' shell (shell.exec, 30 s at most, dangerous commands refused) and shows > !cmd with its output and exit code; it never reaches the model, and during a turn it waits in the queue. A paste of more than 3 lines or 800 characters is one atomic chip, [Pasted text #1 +60 lines], sent whole and shown compact in the transcript. The working row counts the turn's tokens, each turn ends with ✻ 34s · 3 tools · ↓1.2k, a turn of 10 s or more ends with the terminal's notification (OSC 9, OSC 777 or the bell; OMH_TUI_BELL=0 silences it), and the terminal title names the session and its state. A tool row shows the first three lines of what the tool printed (a failed one its error's first line, or the last three lines when only its exit code says it failed), and Ctrl+O opens the whole transcript with every detail. In the agent list, x twice interrupts the highlighted agent and s steers it with the next line you send. The plan panel follows Hermes' todo.updated at once instead of waiting 2 to 8 s for the HUD's next poll. Every repaint is one synchronized update (DECSET 2026). While Hermes asks something, the mode row names that question's keys, and neither row offers Esc for an interrupt while anything else takes Esc (a question, a picker, the search, the agent list, a picked steer or rewind, a draft); the strip ends with ↓ agents; a narrow approval keeps a key hint and a wrapped command its indent; the banner is one row (the model is on the status line, the keys on the mode row) and a fresh session names the last one with its /resume command. Claude Code's names work: /cost and /permissions run Hermes' /usage and /approvals, /todos prints the whole plan, bare /model opens a picker of the models Hermes offers and switches this session to the one picked (it printed the current model before), and /export first says that Hermes exports a profile archive, not the transcript. Alt+Backspace (and Ctrl+Backspace) deletes the word before the caret, and Alt+B and Alt+F move by a word under the kitty keyboard protocol too, where they did nothing. A start on a terminal of 100 columns and 24 rows or more shows the OH-MY-HERMES block logo above the banner, from the shipped skin and in its gradient, lifted to read on a light background. Motion follows one rule set: a reply opens with ⏺ in the text colour and a tool call's name in the tool colour; a running tool's ⏺ pulses (each call on its own clock) and settles to ✓ or ✗ before it reaches scrollback; a highlight sweeps across the working row's verb; OMH_TUI_REDUCED_MOTION=1 or NO_COLOR holds everything still and repaints once a second. Your prompt sits on a tinted band to the right edge (an erase in that background, so a copy picks up no spaces), a prompt the queue sent says ⧗ from queue, and a steer shows as ↪ steer · <text> under what is running until Hermes answers; its row then lands at the next block boundary. A steer answered while a tool ran used to close the tool group and mark the running call ▪ cut off, and one answered mid-reply split the reply in two. A command typed during a turn (its echo and output) and the agent list's steer and interrupt rows wait under what runs the same way, and two commands out at once each print under their own echo. Each turn's done line sits on a thin rule to the edge, and the answer you gave an approval is written above the call it was about, ✓ allowed once · <command> on a light green band or ✗ denied · <command> on a light red one; failed and denied tool rows with their lines and error notices sit on the red band too, a denial shows the one red ✗ (no amber ⊘), and with no background reported nothing is banded. A turn's last lines and its done line sit right above the panel on a full screen, where blank rows used to stay under them, more the taller the last tool group was (23 for 12 calls at 40x100; origin/main had the same gap). A call Hermes only drafted (tool.generating) and never started, such as a streamed attempt it discards and retries, leaves no row and is not counted in the done line, where it settled as ▪ tool_call; a deferred tool's tool_call draft becomes the call it reached. The plan header's parallel shot ×N badge, and the desktop pane's chip, show only a count of 2 or more, so calls run one after another no longer read parallel shot ×1. The composer wraps between words instead of splitting a word at the edge, as the echo of the sent prompt does, with one difference: a space stays on the row it ends, so a word with no room left for the space typed after it moves to the next row (the echo keeps the word there and drops the space), and a tab with no room takes a row of its own. While you pick from a list, the composer is hidden, as in Claude Code: an approval, a clarify question with choices, and the /sessions, /rewind and bare /model pickers stand where the composer was, with no rules (a picker keeps the caret at the end of its filter row, where its typing goes), and the draft comes back as you left it when the list closes (a question waiting behind the first opens before the composer is drawn again). An approval lists deny, allow once and allow this session one per row, deny highlighted (❯); Up and Down move as Left and Right do, stop at either end so Up then Enter still denies, and no number picks one. A clarify question numbers its choices and ends with Other (type your answer), Hermes' wording; that row, or any character that picks nothing, opens the composer as the answer field, and Up or Down goes back to the list. A picker's filter is its last row. A clarify or sudo answer used to be typed into the composer in place of the draft, which came back as plain text with its caret at the end: a paste chip in it no longer expanded, so sending it sent the chip's label instead of the pasted text.

    • A portable host is no longer told to load a skill only Hermes ships. On main the Agent Skills bodies and references named 17 such skills 63 times: by label (omh-research-department), by backticked or hyphenated canonical name, or by a one-word name right after "route to" or "routed to". None of these remain. A Do Not Use When line that hands its request only to such skills now keeps the request and loses the pointer, so the skill still says what it is not for. The rule reads the shipping set, so the pointer comes back once one of its siblings ships portably. Every other sentence is rewritten for that projection only: a list line through the new PortableLine, which keeps its portable wording beside the Hermes text; a description or other single-text field through portable_overrides (the portable frontmatter now reads a description override); and the wiki blueprint reference through PORTABLE_REFERENCE_REWRITES. A gate reads which skills each projection ships and fails, by path and line, when a portable body or reference names one the portable projection does not ship. A second test fails when a skill that a rewrite leaves out starts shipping portably, so the rewrite is re-read. Hermes bodies, references and their digests do not change.

    • A host turn or a thank-you gets a direct-answer card. omh chat route-hint and the example shims showed the generic no-hint card, with "Open omh" and "Clarify" buttons, beside a route decision that said to answer directly: on a turn Hermes wrote from one of its session-command templates, on "thanks!", and on every other route the router answers in chat. Where the hint rail had named a workflow beside such a decision, the card offered to open it ("how long does it take to become a doctor" offered omh-doctor). The card could not tell these routes apart: the compact route_decision it reads cuts the reason to thirteen characters. The router now names each answer-in-chat fallback with a stable reason_code (direct_answer, host_command_turn, native_loop_cadence, typed_loop_goal, loop_construct, and file_lookup) on the route and on route_decision. The key is additive and absent on every other route, and reason keeps its prose. A direct-answer code gives a new direct_answer_route_hint card with no actions and no tool checkpoint, which says no OMH workflow is needed; its next_backend_commands no longer lead with a hinted workflow, and the human omh chat route-hint summary reads the card instead of the hint beside it. The router's next action, claim boundary, clarification and routing instruction, and the wrapper's file-lookup and direct-answer cards and its loop-timer check, read the code instead of matching the reason's wording. A turn Hermes wrote around an OMH status, quickstart or intro question ("omh status" under a /plan or heartbeat frame) became that OMH card in omh chat interact, beside the router's decision to answer it directly; it now gets the direct-answer card the router decided, as it already did for a catalog question. Where the wrapper does turn a route into an OMH card, it drops the replaced reason's code from the route and from route_decision. Measured over both corpora on Discord, plus 18 messages on four sources, at origin/main dc6933c3a and at the head: no route, interaction card or next action moved. The route-hint card moved on exactly the 271 direct-answer routes and nowhere else, and its backend commands on the 58 of them where the hint had named a workflow. Of the 271, the 211 corpus rows are all negative controls, 165 from the no-hint card and 46 from a workflow card. Over four Hermes frames around six OMH questions on four sources, the interaction card moved on the 72 host turns that had become OMH cards, and on none of the same questions typed by a person. host_owned_command_turn, a pass-through over the bundle's host_command_turn, is gone; the router and both wrapper payloads call the bundle's predicate.

    • spend-gate and data-share-review admit the same requests on the router and the in-session hint, whatever courtesy opener is used. Four readers of a request's opening -- spend-gate's clause rule, data-share-review's lead, the loop-on-a-timer verdict and the router's command word -- each kept their own list of openers, and the lists had drifted: "hey, redact pii in the export before we send it to the vendor" reached data-share-review and the same sentence opened with "hi,", "pls" or "would you" did not. One bundle list, request_lead.COURTESY_OPENERS, now serves all four; a "." right before a "/" stays with what follows, so ok./loop keeps OMH's sigil. "pls", "plz" and "kindly" mark a spend as "please" does, and so carry its known residual: "pls pay with my card in safari is broken" reaches the gate as "please pay with my card in safari is broken" already did. "hermes," before a data-share request is the address it is for spend-gate; Hermes named after it still withdraws the review ("before hermes uploads it"), a trailing ", hermes" included, and the docstring now says so instead of claiming only a Hermes setting does. The hint also stops naming spend-gate where the router withdraws it: the everyday-sense pair for its cap and wallet phrases moved from the router into spend_signals, whose clause rule applies it, so "can you check out the spending cap for the city council budget" gets no gate on either surface. A loop's prompt and the gate named as a command are exempt on both surfaces, as they are from the request marker, so "run spend gate on my spending limit" dispatches the gate instead of a clarify without it. The gates' spaced names ("spend gate", "data share review") are no longer direct invocations for the hint, which follows the router's verbs ("use spend gate") and the hyphenated names. data-share-review's hint reads the router's own trigger table as whole words, so "anonymize the dataset before we upload it to the provider", which the router dispatches, gets its hint and arms awareness on a later turn. Measured with build_chat_interaction_payload(source="discord") and awareness_route_hint at dc6933c3a (identical on every row to the branch's first base, cd8a879db) and at this head over both corpora, the follow-up probe sets, every string in the two skills' test files and a matrix of 33 openers over 17 requests (2,632 sentences): neither corpus moved on the router or the hint; router/hint agreement on whether the gate is offered went from 260/281 to 297/297 for spend-gate and from 61/78 to 152/154 for data-share-review (the two left are one transcript-excerpt sentence, in two of the sets, that the router dispatches to its owner, media-input-operator). Outside the two gates, an opener new to a reader now reads as the bare request does, or as the request after "please" for an opener that marks one. Behind 11 openers over every corpus case that opens with a Latin word (12,740 routes), 166 routes outside the gates moved and 161 now match that reference: 76 requests that open on a fix, build or change verb, behind "hey,", "hi,", "so", "just", "let's" or "i need you to", dispatch (46) or ask (30) with ultrawork leading, where they went to feedback-triage (30, among them three negative controls such as "fix the broken wheel on my shopping cart") or a clarify led by another skill; 12 cadence reports behind four openers each (48) clarify instead of dispatching automation-blueprint; loops on a timer get Hermes' /loop (20) or, after a cadence naming ulw-loop, OMH's loop (8); and two pay-on-my-behalf feature requests after "pls" and "kindly" clarify as after "please" (2). The other 5 are "configure web search" after "hey,", "hi,", "so", "just" or "let's", which now asks with websearch-setup leading, as "please configure web search" does, where the bare request dispatches websearch-setup (before: dispatched web-research). The money cues of a send or transfer are read as words, so "please transfer the balance with my card-on-file" stays admitted. Internal: one offers-itself table row per gate, spend-gate's clause tokenized like every other skill's message, and the misleading _WHOLE_WORD_ROUTE_HINT_RULES replaced by the hint's phrase readers.

    • Every skill says one thing about spends, retired siblings, review retries and the one plan checklist. A loop step that would move money goes through spend-gate first while the loop carries on, in that skill's own terms: an agent run on a schedule never spends, and a no-agent script that pays is a spend set up now. ultraperf stops sending a declared metric, baseline and benchmark to performance-goal, a skill #1691 folded into ultraperf itself; a declared goal starts from its baseline there, as its recovery note already said. decision-prototype's compact body now carries the line that hands a multi-question feasibility probe to Hermes' bundled spike, which only its on-demand full contract held. code-review's two references agree on a failed review return: when the host checked the return against the schema, the same reviewer is not asked again, because Hermes' completion message carries no retry count to wait for. The route hint's plan_todo line and the context tool hint name all three plans only omh_todo holds (a planning run, an accepted plan or a code-story template), so a code-story run is no longer pointed at todo_list, whose write would be held. The grounded-citations ledger and verify rule, the blocked-page recovery note, the curator-ledger reading rules, report-package's powerpoint rungs, the native-tool presence clause and the shared-rail pointer are each one string their skills compose, where they had been retyped.

    • Memory recall no longer fills a pack with records that share only the project's name. A query term that more than half of the records the pack could serve hold, when there are at least five of them, never makes a record match on its own. A subject's replaced values, stale records shown only for inspection, and extra copies of one summary are not counted. The term still orders records of equal coverage, a record that carries it as a tag still matches on it, a subject's current value that shares it still takes an older value's match on another word, and a query whose every term is common keeps them all. A record that met the query only on such a term is excluded as common_terms_only, and that overlap no longer triggers the active-tier fallback; a query that overlaps nothing, or overlaps only through stemming, still falls back. It is not idf, and a rare term gains nothing. Measured against origin/main at 03741a6a8 by replaying the 2026-10-14 memory A/B store (21 records, 11 tasks) through the prefetch selector: the two control tasks, which no record answers, served 12 records before and 4 after; each of the nine memory tasks kept its record in the pack, and one more of them now leads with it. On blind rounds 3 to 6 and their tie-flip twins (176 cases, the lead-record rule), 0 wins, 176 ties, 0 losses; one round-5 case and its twin now hold back a second topic record that shared only the common word. Those rounds were seen while choosing the rule.

    • What a Hermes background-review or /btw fork still does to OMH on the 0.21.1 plugin floor is written down, and the floor stays at 0.21.1 (a #2112 follow-up). On Hermes 0.21.1 such a fork's turn end still calls OMH's on_session_end under the person's session id, and that call carries no parent_session_id; Hermes 0.21.2 stops calling it for a fork. CONTEXT.md now defines the Detached fork and records what that turn end does, measured by calling OMH's hooks with the arguments the fork sends as read from Hermes source. A route written during a Hermes turn stays in place, because the restore is scoped to the task that wrote it and the fork's task id is its own. A route recorded without a task id but under the person's session id is put back at the fork's turn end. The metadata-only plugin_session_end checkpoint can carry the fork's time. None of it was observed on a live 0.21.1 host. The entry also names the tool hooks a fork reaches on every release, transform_tool_result included, and the unbuilt OMH-side fix (correlating the turn id pre_llm_call saw). Raising requires_hermes to >=0.21.2 was rejected because it would drop every 0.21.1 install. Docs only; no code changes.

    • On a gateway turn OMH routes on the person's own words. With Hermes' Discord tools loaded, every Discord message reaches OMH behind [Triggering message id: `…` — use as `message_id` for reply/react/pin via the discord tools.]; a reply adds a pointer that quotes the message replied to, a channel can add backfilled history before [New message], and a shared session puts [<name>] before the text. The in-session route hint, the candidate line, the chat router and the wrapper read all of it as the request. On origin/main 42ad99ec7 the trigger note alone changed the hint on 102 of 787 intervention cases and 41 of 748 negative controls, and the route on 159 and 404 of them: "please pay with my card for the summary report" went to gateway-intent-card instead of the pre-spend card, and "what is a web crawler" dispatched web-research. One helper, person_request_text, now removes that material before any route is decided; a reply's quoted text is context, not the request. What OMH hands on still carries the whole turn: the routing prompt a bridge forwards to Hermes, the coding handoff and the plan keep the quote, an attachment's path and the backfill. The sender prefix is removed only on a turn from a messaging gateway (the platform in the hook, the discord, slack or telegram source in the router and the wrapper), since Hermes writes it nowhere else, and even there a bracket that names a skill, omh or ulw, an invocation, a relay header, or a second bracket stays: [omh] status, [deep-interview] … and [ulw] fix the flaky tests route as before on every surface. A role marker is read from the person's own text. A reply pointer ends at the "] after its quote and an image description at the ~] after its closer-look note, so a ] and a blank line inside either no longer cut them; QQ's [Quoted message]: block, which has no close, leaves the turn whole. The context brief and the plugin tools' standalone fallbacks now skip a turn Hermes wrote from a session-command template, as the hook and the wrapper already did. Measured over both corpora: behind a trigger note, a reply pointer in either Hermes order, or a reply quoting a skill request, nothing differs from the bare request on the hint, the route or the wrapper's next action; behind a sender prefix on a messaging turn, only a request that itself opens with a bracket differs, and it reads as it did before. Routed bare, no route moved; the hint moved to none on the 8 Hermes host frames in the corpora, and two loop requests behind a sender prefix now start the cycle, as the same request without the prefix does.

    • A refused or Hermes-recovered fanout unit ends in a true state. A unit refused after a capacity-reuse reservation (no provable write fence, or the reused worktree's preflight), or one whose dispatch raised before its spawn, left its in-flight marker behind, so fanout status had to compare write times to avoid reading it as in flight; every path out of a unit dispatch that got its worktree now clears the marker, and a refusal for a worktree it did not get still leaves another dispatch's marker alone. The opt-in Hermes recovery lane (--on-failure=hermes) sent its file-tools-only child a prompt asking it to run a verification pass and commit, which it cannot do, and nothing did either afterwards, so a completed recovery left edits unchecked and uncommitted. The child now gets an edit-only prompt. What the failed attempt left in the worktree is committed on its own before the child starts; when the child completes, the dispatcher commits its edits on the unit branch and, under --run-verification, runs the unit's declared checks on that commit. The dispatcher's commits and checks run inside the unit's write fence (the child does not), and the lane refuses before the child starts when no fence can be proven or the unit's branch is not checked out. A new file -- a tracked file moved to a new path included -- is committed only inside the unit's file_scope; one outside it (a .env, a build artifact) is never committed and is named in the record, and the lane refuses before the child when the failed attempt left one. The attempt's unit_state is verified only when the child changed something, all of it was committed, and those checks ran on it and passed. fanout status --wait's documented codes and bounds, and the deadline ulw-maestro names, are now pinned by parity tests against the command's constants.

    • A profile is served only its own consolidation brief, and plugins.entries.omh.settings.dreaming_mode turns the brief off for that profile (#2087). Profiles that bind one OMH store, such as every profile of a gateway running gateway.multiplex_profiles: true against ~/.omh, shared one memory/dreaming.json and one memory/consolidation.json. A nearly full MEMORY.md in one profile raised a brief that every other profile was then served on every turn, and all profiles' turns moved one interval counter. The counters and the brief now live under memory/profiles/<sha256 of the Hermes home>/ (the slot the group-activity status already keys by). The pack, the chat notice, omh_memory, omh memory dream and omh doctor read the brief of the Hermes home they are bound to. Measured with two profiles on one store, profile A over its headroom floor, 20 interleaved turns each. Before, B's pack carried a brief on 20 of 20 turns, 14 of them with A's headroom_below_floor reason. The shared interval fired 6 times for B and once for A, and A raised a brief led by its headroom reason 6 times, because each brief B raised replaced the last_reasons that suppression compares against. After, B's pack carries A's reason on 0 turns, each profile's interval fires 4 times in its own 20 turns, and A's headroom-led brief is raised once. dreaming_mode: off in a profile's own config.yaml stops that profile's brief from being raised, served or reported, including one written before; bare off and false both mean off. Unset keeps reminder, the default. Any other value, or a config OMH cannot read, also keeps reminder but is reported as mode_setting_unrecognized, in omh memory dream's mode_setting and as omh doctor advice. On the first attended session start after the upgrade, a single-profile host moves its old home-wide files into its profile directory and keeps a pending brief. Two cases leave them unread: the profile already has its own files (omh memory dream --evaluate raised a brief first), or the profile is routed, which every profile of a multiplexed process is, even a single profile under gateway.multiplex_profiles: true. omh memory retire --apply retires an expiring-only brief in every profile's directory. Thanks to @kanzaka110 for the report and @sanjkrsna for pointing at the per-profile key.

    • omh install --force reads Hermes' curator ledger once, and names the ledger it could not read when it reports what a removal lost. Under --force the install read and parsed the whole ledger again for a replaced Hermes edit, again for a skill it cleared, and again for every retired directory it removed, beside a comment saying it read the ledger once. The read it already made now also keeps every write the ledger records when a local change may be overwritten, and each of those reports takes that list. On a synthetic 190.8 MB ledger (16,000 rows in the shape Hermes wrote before 0.21.4, 20 files each, 30% of them under the OMH pack), a --force install that replaced two skills and removed three retired directories went from 6 ledger reads in 8.1 s to 1 read in 2.9 s with the same result; a --force with nothing to lose went from 2.9 s to 2.7 s, and a plain install from 2.3 s to 2.0 s. With no row under the pack, that --force with nothing to lose is slower, 0.19 s to 0.54 s, because it now parses every row where it used to skip them all, while the one that replaced and removed went from 5.4 s to 0.54 s. When a ledger could not be read, each replaced edit or --force removal that lists what Hermes still holds a copy of carries ledger_copy_partial: true, and one summary line after those lists names what could not be read and says they count only what could be; a --force removal OMH could not judge for that reason now names it too, and a --dry-run result carries unreadable_ledgers. A ledger with more than one torn line is named once, with a count, instead of once per line. The ledger reader no longer drops every write when given a generator of directories, resolves those directories once per read instead of once per write, and builds a row's after list once instead of once per file the row listed before; omh doctor judges Hermes edits once instead of twice.

    • A checked team no longer checks a part while its helper's work sits unmerged in the worktree Hermes gave it, and the board fan-in asks Hermes only for transitions it lands. With Hermes' delegation.worktree_isolation on, an omh_team helper edits a worktree of its own, and its work reaches the session's workspace only when the session merges it. team_reconcile still ran each part's check in that workspace, so a check could pass on a tree without the part's work and accept it. It now reads each returned helper's branch (hermes-subagent/subagent-<id>, from the subagent id Hermes' subagent_start carries) and its worktree (.worktrees/subagent-<id>, found by path even when the helper left the branch for a detached HEAD or a branch of its own) with read-only git. While the branch's last commit or the worktree's HEAD is missing from the workspace's history, or the worktree holds uncommitted changes or untracked files, it checks nothing and is refused as helper_branch_unmerged, naming each branch in unmerged. OMH never merges. A branch git cannot read makes that part's check inconclusive, so the team still reaches a stop. Work held by neither the branch nor the worktree (a branch renamed or deleted after its worktree was removed) is not looked for, and a host that names them differently leaves the guard off. The helper's own instructions no longer say the workspace is shared: they say the check runs in the parent session's workspace, and that a helper in a worktree of its own commits on its branch. In ultraqa's board fan-in reference the main session no longer prepares request_review on a done fixer row or request_changes on a row no review run holds, both of which Hermes refuses: the fixer's own run ends with kanban_request_review, and a failed re-check this session ran is a comment on the row, with sending it back left to the operator's hermes kanban reopen-review.

    • omh_gather_evidence bounds the paths a probe's arguments name. The allowlist matched a command prefix and left the rest free, so python3 -m compileall /any/path compiled files anywhere and printed their paths and the source line of any syntax error (#2137, reported by @simongonzalezdc). A compileall argument, the value of an omh option that opens a path, and a git diff path argument must now be relative, free of .., and resolve inside workdir after symlinks; a path refusal names the rule, never the path or where a symlink leads. compileall also needs at least one such path, because with none it compiles every directory on sys.path; it refuses -i (paths read from a file or stdin), -b (bytecode written beside the sources, past the private pycache) and -d/-s/-p; and it runs with -e ., so a symlink below a compiled directory that leads out is skipped, and with PYTHONSAFEPATH, so a compileall.py in workdir no longer runs in place of the stdlib module (Python 3.11 and later; an older python3 ignores it). git diff refuses -O (an order file read from anywhere), and its path bound stops the switch to --no-index that two paths with one outside the repository caused, which printed both files; a :/ pathspec or <rev>:<path> still reaches the whole repository, as a bare git diff already does. omh doctor refuses --profile-isolation, and omh release evidence-bundle refuses --write and a --verify PATH. Evidence probes and omh_team checks now run with no stdin.

    • The plugin's route hint reaches the router in Hermes chat, so your own trigger packs hint there too. Hermes imports the plugin bundle as hermes_plugins.omh, one package deep, and there the bundle's from ...routing imports climb past the top-level package and raise ImportError even when Hermes can import omh. Four of them had no absolute fallback, so in every Hermes chat the bridge to the router for user trigger packs (#1539) returned nothing: a phrase from <omh-home>/routing/trigger-packs/ got no_hint from the plugin while omh chat route-hint hinted it. Coding-delivery hints also asked the user to choose the executor instead of carrying the router's owner decision, and the visual-QA and office-material cue tables were the bundle's copies. Each import now falls back to its absolute omh. form, and the bridge still resolves on each call. Over both routing corpora (1,535 messages, measured against 42ad99ec7), the plugin's hint with omh importable differed from the package layout's on 118 messages before, each in the coding-owner decision, and on none after. Where Hermes cannot import omh, the usual install, nothing changes; docs/INSTALLATION.md now lists the trigger-pack hint and the coding owner among what that turns off. A test loads the bundle under Hermes' own package name, and a gate requires every bundle import that climbs above the bundle root to be paired with its absolute form and, at module scope, to bind the router's object there. Reported with a reproduction by @fauzibrevy-hub (#2129).

    • omh coding status-board shows Hermes' delegate children and kanban workers beside the fanout units. The board listed only fanout units, so the command could not show the work omh_hud already showed. It now reads the same two readers the HUD uses (hermes_delegation over state.db, kanban_board_reader over kanban.db), opening both read-only, and lists each worker with its source and state in a second table under one headline. The rows keep the board's running-first order, and a kanban lane whose heartbeat stopped is counted with the stuck units. A Hermes database that is missing, unreadable, refused or read only in part prints a line saying so instead of an empty list. To tell those cases apart, both readers now report a read_status, which the HUD payload does not carry. A reader that reaches its selection bound now says so (selection_capped), and the board then shows its counts as lower bounds ("at least 96 running") instead of as totals. A kanban lane that never started reports no elapsed time instead of 0; the HUD still draws it as before. The fanout rows render as before. The run-efficiency route reason now says the report reads the supplied timings and context budget plus Hermes' recorded usage for a named session, with provider, billing and cron not observed; it still described the supplied-metadata report from before #2114.

    • omh chat route-hint names no workflow on a turn Hermes wrote for its own session command. On such a turn (a /plan frame, a /loop wakeup, a /goal continuation, a heartbeat, a goal-mode kanban nudge, a /blueprint seed) the route decision already said fallback, but the hint beside it was read from the frame's text and named a workflow: the /plan frame came back hinted command-operator, and each of the 16 seeds Hermes' own build_blueprint_seed writes for its built-in blueprints automation-blueprint. The hint now reads no request on such a turn, as the plugin's pre_llm_call hook already did. Measured against origin/main at 03741a6a8 over both routing corpora, every host frame behind each gateway prefix, and the frames' bodies without them (1,678 messages): 101 hints moved to none, every one on a host turn, and no route decision moved on either payload. The router's reason for these turns no longer lists commands, so it cannot fall behind the frames OMH recognizes (it had left out /blueprint and /initiate-setup). The spend-rails.md reference both targets ship no longer names OMH's browser adapter, which another host does not have; the Hermes body still does.

    • The OMH TUI keeps the keys typed right after Esc, and stays fast in long sessions. Esc followed by a key within about 50 ms lost both keys (Esc then /quit left quit in the composer, for Enter to send as a prompt); a character sent in the same burst as an Esc now types (tmux before 3.5 sends them that way), and Esc then a key in separate reads acts as both. Option+← and Option+→ in macOS Terminal.app move by a word. A bracketed paste whose end marker never comes ends after 0.5 s of silence instead of swallowing every later key, Ctrl+C included; for the second after, or while the rest of it keeps arriving, a line break adds a line instead of sending the prompt. /new no longer keeps a resumed session's banner for the next redraw, and an image Hermes attached (/image, /paste, a drop) after you left the session it was sent for is a notice instead of a chip in the new one. A call the gateway leaves unanswered for 120 s (Hermes' own TUI's limit) is dropped, with a notice for one a person waits on. A full redraw (a resize, Ctrl+L, /redraw, a switch or a resume) prints the last 3,000 rows under ⋯ N earlier messages · /history shows all, and /history renders a page at a time. The TUI holds a session's last 2,000 messages and the composer its last 1,000 prompts, and an edit's diff is held as its UTF-8 bytes (one 200-line truecolor edit: 16,508 bytes, 28,558 as parsed lines). Measured against origin/main at 14cc8731a with the review lanes' in-process harnesses (load average 10 to 30): at 1,000 turns a redraw took 14.8 s and 118 MB at its peak and now takes 0.27 s and 4.5 MB, and /history opened in 5.7 s holding 65 MB and now in 6 ms; a key in a 500-turn session takes 0.19 ms (was 1.35), a key in a 2,000-line paste 13 ms (169), Up there 0.35 ms (254), a spinner tick over a streaming 20,000-character paragraph 0.26 ms (40), and a tick with an agent's 400-row log open 0.06 ms (95). The loop wakes for gateway events and HUD changes instead of reading every 50 ms: event to frame 0.1 ms median (30), and an idle TUI wakes about once a second (19). Bare omh reaches hermes in 46 ms (445) without importing its command tree, and the first frame no longer waits for the busy-mode read. A HUD read over 5,000 run directories checks only the runs it shows and reopens only binding files that changed: 190 ms (708).

    • Skill text decides which skills are OMH's by the installed label, and the spend, browser, schedule, and parallel-tool skills say what Hermes does. skill, rules-distill, and skill-health no longer treat any omh-* or ulw-* name as OMH's: one shared sentence says an OMH-managed skill is one OMH installed under a label its catalog renders, and a companion the person or an agent created is theirs whatever its name. The skill_manage refusal suggests a companion name without OMH's prefix (plan-notes for omh-plan). parallel-tools reads each MCP server's supports_parallel_tool_calls and offers that flag, with an update only on request, instead of describing a version-currency check. A recurring job or a page step that would move money goes from automation-blueprint or browser-operator to spend-gate; loop sends an unattended schedule through automation-blueprint instead of straight to Hermes' cron tool; and browser-operator carries the unattended-run rule the other operator skills state, with Hermes' card prompt recorded beside the spend gate's typed yes. command-operator and spend-gate say what Hermes does with a pending_approval command at v2026.9.7 and on upstream main: it is stored and never run, and /approve resumes only a run still waiting on an approval. spend-gate's portable body drops the hermes chat -q parenthetical, the OMH browser adapter line, and the pending_approval rule, which a host-neutral held-command rule replaces; prose that named a sibling by its catalog name now writes the label skill_view loads, and the spend and data-share references name only states their skills keep.

    • A /blueprint seed is Hermes' turn, and a spend is spend-gate's wherever it is asked. When /blueprint matches, Hermes rewrites the turn to a seed that asks for the blueprint's settings and then creates the cron job itself; OMH now reads that seed's frame as a host turn, so it gets no route, no route hint and no candidate line, where before it dispatched a file, toolbelt or observability skill. A person typing the seed's first line alone still routes to automation-blueprint. A loop on a timer whose prompt spends the person's money ("loop every hour and buy the tickets for me with my card if they drop under $50", or the same as a typed /loop 1h ...) opens spend-gate instead of being answered with Hermes' /loop, since a scheduled run never spends; Hermes' own /loop wakeup turn is unchanged. Naming the gate as a command ("use spend gate") now asks for it. The router, the route hint and the awareness markers read one phrase table for the gate, so "go ahead and pay the 402 challenge for the dataset" no longer hints a research skill while the router opens the gate. A renewal, tip, booking, subscription, donation or top-up paid "from my wallet" is no longer dropped as everyday wording, and paying with the person's card, or reviewing the pre-spend card, no longer opens the image-card skill, while an image card of "my card game stats" or "my card collection" still does.

    • Memory recall no longer lets a colloquial query's particles, endings and fillers decide which record leads. A Korean query run loses the particles and request or verb endings it ends in while two syllables stay (자동으로 is scored as 자동, 확인해주세요 as 확인, 경로 stays whole), and a run that is one of them whole (해주세요, 에서 after a Latin word) adds nothing. Korean discourse fillers, intensifiers and demonstratives (그냥, 진짜, 그거 ...), English hedges (honestly, literally, kinda, really ...) and the Japanese and Chinese wrappers って, なってる, 什么时候 and 一般 join the closed query-side lists. All of them still order records of equal coverage; records are tokenized in full. The strip never makes a query fall back to every word (which would promote 어떻게 or "what" to content): the unstripped content is tried first, and a two-letter word written in capitals (PR, CI, DB) is scored as a name, so "what about CI" no longer falls back either. Measured against origin/main at 4d3c10141 with the blind-round harness and its lead-record rule (distractor ids sort first, so a tie goes against this change): round 4, 20 cases, 5 wins, 15 ties, 0 losses; its tie-flip twin 4 wins, 16 ties, 0 losses; round 3 and its twin 0 wins, 20 ties, 0 losses each; the 40-case pre-registered corpus served identical packs. Those rounds were seen while writing the change. On two rounds frozen by a writer who never saw the code (24 owner-style cases each), it scored 2 wins, 21 ties, 1 loss and 2 wins, 22 ties, 0 losses. The loss is a query whose topic record shared the particle form (모델로) that now only orders records.

    • Bare hermes is stock Hermes again. With bare omh opening the OMH TUI, OMH stops styling Hermes' own terminal: setup and update write no display.interface, display.skin or display.sections, install no status widget and no skin files, and ask no display question (--no-omh-tui is accepted and does nothing). The OMH plugin stays loaded in hermes, so its tools, hooks, memory and skills still work there; the HUD and /omh-model are the OMH TUI's. An install an earlier version styled is migrated by the next omh setup, omh update or omh apply, primary home and every bot profile: a display key OMH's record says it wrote goes (or goes back to the value consent replaced), an omh* skin goes with or without a record after a theme the person picked is saved under the OMH home, and a value the person changed or OMH cannot attribute stays. The widget and skin files go where a manifest (or, before the release, the shipped bytes) proves OMH wrote them; an edited one, and the skin the config still names, stay. A key in a YAML shape the editor refuses holds the release, and is named, until a later pass can take it; once a pass leaves none, display_released_at in the write record keeps later updates from taking back a skin, or a skin file, the person picks afterwards. plugins.enabled, memory.provider and the registration are never touched by it. A home left unregistered keeps its config and loses only the old files. omh theme use now saves to <omh_home>/tui/theme.json (a profile's own store when it names one) and the OMH TUI opens with it, omh --profile <name> included; omh theme repair is retired (it says so and exits 0). Where the OMH TUI cannot start, omh says why in one line and opens hermes --tui for a reason of the OMH TUI's own, or plain hermes under TERM=dumb, OMH_NO_TUI=1 or without Node. The hermes_tui_* doctor checks and the post-update OMH TUI: verdict are retired with the surface they judged, and omh setup --json moves to hermes_native_setup/v2. See "Returning hermes to stock Hermes" in docs/INSTALLATION.md.

    • omh quality-evidence agent-skills reports Agent Skills host evidence per host and scope. For every Agent Skills host adapter, in repo and user scope, it records five stages (installed, discovered, nominated, loaded, executed) and never infers one from another. installed is observed by OMH: it reads the host's scan path under --project-root or --user-home and compares every file with the omh_host_adapter/v1 manifest; the read writes nothing. The later four stages stay not_observed with each host's probe_unsupported reason, since OMH has verified no credential-free, non-interactive listing command, makes no model call, and starts no host session. The only host process it runs is the readiness module's approved --version probe, for a CLI in that probe's set; it runs whenever such a CLI is on PATH, where readiness ran pi and opencode only as omo hosts. Each omh_agent_skills_discovery/v1 receipt binds that host version, the OS, the scan path, the manifest's source_digest, and the pack catalog_revision; the parser rejects a receipt that fails the checks docs/AGENT-SKILLS.md lists. --receipts compares an earlier --json run with the current pack and host versions; an unreadable scan path exits 3, and a missing root is refused. Built on the stage model, host descriptors and fixture tests @blrain3 contributed in #2093.

    • Bare omh opens the OMH TUI, OMH's own terminal over Hermes' gateway. omh runs hermes --tui with OMH's frontend in place of Hermes' Ink app, so Hermes' launcher still resolves the profile, the resume target and session ownership, and every model call, tool run and approval decision stays Hermes'. Finished lines go to the terminal's own scrollback and only the region under them repaints; once the composer has reached the bottom rows it does not move when the plan and agents panel, or a question Hermes asked, grows or shrinks. Nothing is answered for the person: Esc, Ctrl+C, d or n declines, and allowing a command takes the arrows and then Enter. When the TUI cannot start (Windows, no termios, no Node, an OMH that cannot import itself, gateway drift) it creates a file omh named for that launch and exits 75, and omh opens Hermes' own terminal. The boundary is in CONTEXT.md ("OMH TUI", "Hermes gateway protocol"); the gateway slice it relies on is declared in src/tui/gateway_contract.py. The composer sits between two rules drawn in the input mode's glyph and colour, over a mode row that names the mode or the keys that matter now; replies open with ⏺, a diff numbers its lines in a gutter, and the status line reads brand, model, context, state. On the terminals Hermes' own TUI trusts with it, it asks iTerm2, kitty, WezTerm, tmux, VS Code and Windows Terminal for the kitty keyboard protocol and xterm modifyOtherKeys, and Ghostty for modifyOtherKeys only, so Ctrl+Enter and Shift+Enter arrive as keys of their own, while every key the TUI binds decodes as before and so does each Ctrl chord xterm sends as one control byte (Ctrl+2 a NUL, Ctrl+3 Esc, Ctrl+8 Backspace); OMH_TUI_KEYBOARD=legacy keeps the plain bytes. Shift+Tab cycles normal, accept all and plan: accept all turns off Hermes' approvals for that session only (Hermes' per-session switch, never approvals.mode), is shown on the composer once Hermes confirms it, or when Hermes confirms neither it nor the switch back off, and the TUI sends that session an explicit off before it leaves it, and plan sends each prompt as /plan <text>; where Hermes has approvals off on its own, Shift+Tab sends nothing and turns plan on and off beside accept all. Ctrl+T or /effort <level> sets the session's reasoning effort, Ctrl+Enter or /steer <text> adds typed text to the running turn (a steer Hermes does not take waits in the queue for the turn to end), queued prompts are listed above the composer and Up takes them back to edit, ? or /keys lists every key from the table the docs are held to, and an approval rings the terminal bell once (OMH_TUI_BELL=0 silences it). A turn's subagents show as a strip on the mode row; Down opens them as a list, and Enter or a click opens one agent's read-only log, with mouse reports on only while an agent is there to click.

    • In the OMH TUI, /reload-skills makes a newly installed skill runnable, and a turn of unknown gateway events says so. /reload-skills (or /reload_skills) has Hermes re-scan its skills in the gateway itself, then reads the command list again, so a skill installed after the TUI started (omh update, /skills install) runs as /<name> without a restart. The command list is read for each session the TUI opens, so a project's own skills follow its workspace; its warning (skill, quick-command or plugin discovery failed, or skills that run only as /skill <name>) shows once; /help lists Hermes' commands by category with a count of skill commands and is paged when long; and completion also matches part of a name, so /uml finds /omh-codebase-uml. A turn that brings no event the TUI shows, only types it does not know, gets one notice pointing at omh doctor, whose OMH TUI check now prints the installed Hermes commit next to the one the TUI was declared against. omh opens Hermes' own terminal only on exit 75 together with a file it named for that launch, so a 75 Hermes returns for a reason of its own ends the launch like any other exit code. The HUD's agent panel shows a Hermes-native subagent whose session has ended as finished at once, where it showed it running for up to 150 seconds. A plan item carrying an escape sequence is stored without it, where the sequence's parameters ([31m) used to stay behind as text.

    • The OMH status widget and the Hermes Desktop pane keep ordinary punctuation and drop whole escape sequences. Their character allowlists dropped apostrophes, commas, question marks and quotes, so a plan item person's rendered as persons, and they removed only an escape's ESC byte, leaving [31m behind as text. The widget, the Desktop pane and the OMH TUI's panels now share one allowlist and one escape rule and render the same text.

    • Temp directories, RFC 4122 UUID directories and Hermes session ids no longer read as credentials. The opaque-value check took the hyphens of a UUID, a run of Python tempfile names, or a numbered temp directory among words for a secret split into fragments, and a path segment that opened with _ or - (_temp, -Users-name-project) or held an uppercase UUID made a whole path suspect. A memory capture whose source_ref was such a path stored it as [redacted] and went to review instead of auto-approval, and work-resume items and completion summaries carrying one became [redacted]. Over 2,000 generated paths per shape, a UUID directory went from 100% flagged to 0%, three chained tempfile names from 51.8% to 0%, a Hermes request dump name from 7.8% to 0%, and a Hermes session directory from 1.7% to 0%. A value split into fragments beside one of these names is still caught at the rate the same value is caught with nothing beside it, and detection over 25,000 generated split secrets moved from 17,571 to 17,568.

    • A new spend-gate skill shows what a spend will cost before the agent moves any money, and calls it done only on a receipt. It covers every rail: a card checkout in a page, an HTTP 402 or pay-per-request call, a wallet transfer, a booking or deposit, a paid data endpoint, and a renewal that charges again. Before money moves the person sees a pre-spend card with the payee and its exact origin or address, the item, the total with fees and currency, the rail, the cap and what remains of it, the refund terms, whether the spend can be undone, and who approves it (the person whose money moves), and the person's typed yes in that session is required whether or not Hermes shows an approval prompt of its own. A host prompt that appears is recorded beside that yes; one that did not appear, or an auto-approval, is no confirmation, and a choice picked in clarify counts as the typed yes. A run nobody can answer -- a scheduled job, a webhook, an API-server or single-query run, a background worker, or a delegated subagent -- never spends: it prepares the card for the person's next interactive turn. A payment set up to recur without the agent (a subscription, a standing order, a capped auto top-up, a script with no agent in it) gets its card, typed yes and cap now. A quote is not a purchase; only an observed order id, transaction id, or receipt header completes a spend, an outcome nobody saw is unknown and is reconciled before any retry, a payment Hermes holds for approval (it never runs on its own) is never re-run or reworded and is recorded as not submitted, and a declined card is not retried. The cap ledger keeps metadata only. Hermes' saved-card fill, clarify, and the optional stripe-link-cli and mpp-agent skills are used only when present; without the fill tool the person enters the card and the agent never types one into a page. OMH moves no money and holds no credential. The skill is guidance only: no hook holds a payment tool call.

      Routing is deliberately conservative and closed-world, because the gate's phrases are everyday words that also name payment features, labels, bug reports, product ideas, articles and questions. The gate is offered only when the message is marked as a request outside the gate's own phrase: "please", "can you", "could you", "would you", "will you", "go ahead", "I need you to", "I want you to", an address to Hermes, or an amount or "for me" as the spend's object: outside the gate's phrases and before "at", "for", "from", "in", "on" or "to", and for "for me" after a word of its own since the last phrase ("transfer 40 dollars", "buy the cheapest one for me", "pay for me with my card"; not "pay with my card for $0 orders is broken", not "book the flight with my card for me never shows up", and not the "for me" of "buy it for me"). An imperative with none of these -- "pay with my card for the hotel", "buy it for me" -- routes as before the skill existed: that recall is given up so "Pay with my card in safari is broken" never reaches the gate. A marked message still has to lead with the spend: the sentence's first words after filler, up to the first comma, colon, question mark, "and", "if", "when", "which", "about" and the like, must open with a spend addressed to the agent -- "buy", "pay", "order", "book", "renew", "place the order", "complete the checkout", "use my card", and the like, not an idiom ("pay attention to the spend cap", "pay a visit") -- and one of the gate's phrases must be in that clause, each followed only by the end of the clause or the start of the next ("when it is back in stock"), "for", "on", "at", "from", "to" or "in", "with" the person's own card or an amount, an amount, "please", "now" or "right away"; "or" after the phrase makes a comparison, not a request. "send" and "transfer" also need money, an amount, or the person's own card or wallet; "did the payment go through" counts only when the agent paid. The skill scores only on that clause, there is no boost, and nothing stands down another skill's guard: a spend dispatches only when its clause scores high on the gate's own phrases, otherwise the gate leads a clarify, and every other sentence routes as before the skill existed. The route hint and the awareness markers read the same clause, as whole words. The bare words spend, pay, checkout, wallet and their siblings score only inside a complete phrase, and budget is not one of the skill's words at all. Budget variances, payment failure, checkout, 3DS and crash reports, a feature request, browser logins, spend-approval research, checkout test failures and issue triage, purchase acceptance criteria and a saved-card settings toggle keep their owners.

    • Session usage and the skill-draft check recognize an OMH skill by its catalog label, not by the [omh] description prefix (#2120). canonical_skill_for_label answers from every label the catalog renders or rendered, retired skills included, and omh quality-evidence session-usage reads it for a SKILL.md load, a reference-file load, a compaction placeholder, and a frontmatter name line. A prefixed description still counts under a name the catalog does not render, so no session that counted before stops counting. The 3.0.0 entry for session usage, which names the [omh] prefix as the OMH-skill signal, no longer describes it.

    • A new data-share-review skill prepares a dataset, log, ticket export, or transcript excerpt to leave the team, and calls it redacted only after the file that leaves was scanned. It asks where the file goes and what that destination needs, classifies every field as a direct identifier, quasi-identifier, secret, free text, or not personal, reads a row only when per-column shape counts leave a field unclear, drops what is not needed before masking anything, and keeps any pseudonym mapping at a local path outside the shipped file. The verdict is redacted only after an observed residual scan of the exact file that leaves counts every class, free text included; without one it is not_observed. Counts come from Hermes' search_files in its count mode or a detector script run through execute_code that prints counts only; that script's stderr and any traceback return too, so it catches its own errors without echoing a value. The skill says plainly that Hermes' privacy.redact_pii and security.redact_secrets cover gateway chat identifiers and tool output, never a file that leaves. What left is recorded as metadata only: destination, date, fields, row count, and digest. OMH uploads nothing, keeps no raw value, and leaves the lawful basis to legal-compliance-review. Chat routing to it is deliberately conservative: the words it lives on (redact, PII, pseudonymize, strip, export, vendor) also name code, Hermes settings, legal questions, issue triage, messages and posts, so it takes nothing another skill holds. It has no routing boost, no other skill's guard stands down for it, and it is scored only when a sentence leads with redact, pseudonymize, anonymize, or de-identify on a file leaving the team, such as "de-identify this export for the researcher". Logs for a GitHub issue, a transcript excerpt, or a dataset for a contractor stay with the skill that holds them, while Hermes' own skill index, or naming the skill, can still select it.

    • The coding lane works with Hermes' own delegation, board, and debugging skills instead of beside them. ulw-work loads a Hermes-only native-delegation reference before a delegate_task dispatch. Each lane it writes itself carries an output_schema holding only the fields integration reads; the host's schema_valid is read only from a result returned in the same turn, a later completion message is checked against the schema by hand, completed is never read as validated, and a lane's answer stays its own report until the verification fan-in runs. With the user's delegation.worktree_isolation on, each lane works on its own hermes-subagent/… branch, so integration is a merge the parent performs before any check, after committing whatever a lane left uncommitted in its worktree: the checked team lane now says a check run before that merge tests a tree without the lane's work, and the file-ownership manifest is compared per lane branch. A one-shot run, which is also how a kanban worker runs, that is refused for its delegation budget finishes the remaining lanes inline in dependency order instead of retrying. The kanban lane recipe adds completion_contract (the board refuses completion until the repository's required checks pass on the PR's current head, a record the closing brief still checks against the PR head), goal_mode with goal_max_turns, kanban_request_review into the review column, and kanban_schedule where the worker has it; omh_agent_board now admits the two goal fields on a create and reads back a card in the host's scheduled status instead of failing the receipt. agent-instructions documents how Hermes loads context files: one file type per session, .hermes.md first, every AGENTS.md from the repository root down, and a per-directory AGENTS.override.md the inventory has to list. plan treats a file saved by Hermes' built-in /plan as its draft after that turn, with pasted code reduced by the proportion check and acceptance still taken only from the user. On Hermes, decision-prototype keeps its decision record when the bundled spike builds the experiment, app-debugging writes its records at each phase of the bundled systematic-debugging and counts a python-debugpy or node-inspect-debugger read as one observation, and ask takes only the invocation from the bundled claude-code, codex, or opencode skill for a CLI advisor the user named, with no CLI as the default.

    • Generated skills follow more of Hermes' own skill-authoring conventions. Every Hermes skill body now opens its trigger section with ## When to Use, the heading Hermes' skill linter looks for, and carries license: MIT and author in its frontmatter (the Agent Skills projection carries license only, the field that spec defines). Pointers to the references the router skill owns (the shared rail, the Jev rail, structural code search, workflow artifacts) are written as the call that loads them, such as skill_view(name='omh-routing', file_path='references/skill-common-rail.md'), instead of a bare path the model had to translate. A test now holds every shipped SKILL.md free of Hermes' inline-shell shape, which Hermes would run as a command when skills.inline_shell is on.

    • Three request shapes that belong to a skill Hermes bundles, or to a work board, stop dispatching an OMH skill. An architecture diagram of cloud or infrastructure pieces described in words, with no codebase, repository, module, folder, src tree, CDK or Terraform named, no longer dispatches codebase-uml; Hermes bundles architecture-diagram for it. An English message naming an LLM wiki no longer reaches the wiki blueprint, nor the research department unless one is named outright; Hermes bundles llm-wiki. A review of what sits on a kanban, board, task, Trello or Jira card no longer opens an image card; a dashboard, scoreboard or storyboard card still does. Each of those turns now asks instead. Measured on the chat-interaction surface over both routing corpora, every sentence literal in the routing tests, the sentences this change was probed with, and its reviews' sentences: those are the only routes whose action or selected skill moves, each from a dispatch to a clarify. arXiv lookups, citation checks and paper research route as they did. The research-department body maps Scout to the research and web-research lanes its prepared plan emits, and handoff policies, like Do Not Use When lines, name siblings by the labels skill_view loads.

    • Skill bodies name their siblings by the name Hermes can load, and say which requests belong to a skill Hermes bundles. A ## Do Not Use When line that hands a request to a sibling used to name it by its catalog name (connector-operator, strategy-brief, research), which skill_view cannot load because the installed directory carries the display label (omh-apps, omh-decide, ulw-research). The renderer now writes the label in both the Hermes body and the Agent Skills projection; the catalog keeps canonical names, so routing and the deference graph are unchanged. A sibling that is no longer installed keeps its catalog name. The Hermes frontmatter gains metadata.hermes.related_skills, derived from those same lines, which skill_view returns beside the body; Hermes' per-turn skill index does not read it, and the index size is unchanged. A skill whose job overlaps a skill Hermes bundles now carries one line naming which request belongs to the bundled one: requesting-code-review, systematic-debugging, simplify-code, spike, dogfood, llm-wiki, grounded-citations, arxiv, architecture-diagram, github, claude-code / codex / opencode, sdlc-review, meeting-action-items, humanizer, pdf / docx / powerpoint / xlsx, hermes-agent-skill-authoring, and competitor-news-monitor. Those lines are Hermes-only, no other host bundles those skills: the deferral lines web-research, research-department, visual-qa, ultraqa and code-review already carried for grounded-citations, arxiv, competitor-news-monitor, dogfood and sdlc-review keep their wording and now leave the Agent Skills projection too, each counterpart named once per skill.

    • The memory and learning skills now read what Hermes' own learning loop wrote or is holding: pending memory writes, staged skill writes, the skill audit ledger, curator run reports, and the one memory-provider slot. memory-sync lists the memory writes Hermes holds for approval under pending/memory/ as proposals the person decides with /memory approve or /memory reject, reports a native write that comes back staged: true as waiting rather than saved, and reads each file's cap from the inventory, with the fallback figures taken from the plugin bundle's own Hermes memory reader rather than a second copy. Hermes' write_approval is described as Hermes' gate on Hermes' own writes: a safe omh_memory capture is still approved by OMH's auto-safe policy with no person in the loop, OMH never approves or rejects a pending write, and it never proposes turning the gate on. external-connector-readiness states that Hermes runs one external memory provider, so adopting another provider where memory.provider reads omh stops OMH's per-turn recall and dreaming reminders and leaves omh_memory refusing to recall or write records; the switch stays a proposed diff the user applies. It also records the plugin install scan and the OSV malware check for npx, uvx, and pipx MCP servers only from output that was seen, since a quiet install or launch is not a clean result. workflow-learning reads the skill audit ledger and reports agent or background-review skill writes that no store route accounts for as unreviewed_native_write (no OMH record of the write, not proof nobody reviewed it), and counts an approved update as landed from the ledger only when an agent or user entry creates, edits, patches, or writes the approved file after the approval; instinct-ledger keeps such a write out of an instinct's evidence and promotes through store routes; rules-distill lands an approved candidate in the user's own Hermes skills through skill_manage, never an OMH-managed skill, and reads curator run reports as one observation each. skill-health reads the ledger for managed skills changed outside OMH, lint_warnings from skill writes, and hermes skills check for hub updates, and never presents use or load counts as a reason to prune, archive, or retire a skill. skill-scout fills its candidates from hermes skills search and inspect with each source's trust level, and a scan verdict only from install or audit output. prompt-import-readiness starts from the hermes import-agent --dry-run preview, which skips Claude Code's slash-command prompts and does not read Codex prompt files, makes a skill the conversion target, and checks names against built-in commands, quick commands, plugin commands, skill bundles, and skills in the CLI's order. The seven feature-surface skills among these state their own Why and Do-Not-Use.

    • The run-efficiency report now carries what Hermes itself recorded for a named session, and the tracker skills point at Hermes' own usage, compaction, tool, agent, stop, and skill surfaces. omh runtime efficiency-report --hermes-session <id> adds a host_usage section from the part of the cost receipt omh_run_summary returns that Hermes itself accounts for in state.db (the session, its delegated children, and its kanban workers; executor-reported fanout units are left out), read-only: observed only when that usage recorded tokens, with each amount's basis kept beside it (a Hermes estimate is not a charge), while a session Hermes has no row for, one with no recorded tokens, or a Hermes home whose state.db is missing or unreadable stays not_observed with that reason instead of reading as zero, and provider billing stays not_observed either way. The cost receipt now names which of its reads hit a bound (truncated_sources). run-efficiency, ops-observability-card, and context-budget-review answer "what has this conversation used or cost" by what each Hermes surface shows: /usage this session's tokens and context, /insights an estimated cost across recent sessions, and omh_run_summary this conversation's recorded tokens and cost; context-budget-review stops sending the question to the ops card, which cannot answer it. context-budget-review turns its summarization checkpoint into a /compress <focus> the person runs, offers --preview only in the classic CLI or on a messaging platform and never when the surface is unknown, because a modern TUI that does not parse the flag compresses for real, and counts a preview only from Hermes' own "no changes made" reply. toolbelt-readiness counts a tool deferred behind tool_search as present, never missing. running-work-board adds this conversation's Hermes delegate children and kanban workers from omh_hud. cancel clears OMH state with OMH's own commands and hands host work to the person's own Hermes stop commands (/stop, /loop stop, /goal pause or /goal clear, /heartbeat clear, /pause), never ending a host process itself, and reports a host stop as requested until Hermes answers. skill lists, reads, and edits the person's skills with Hermes' skills_list, skill_view, and skill_manage, leaves search, install, and hiding to the person's hermes skills commands, and never edits an OMH-managed skill in place. capability-toggle sends hiding one skill on one platform to Hermes' own hermes skills config.

    • Hermes' background review and /btw side question no longer use up the once-per-session OMH lines meant for the person's next turn. On Hermes 0.21.1, the plugin's floor, both run as a fork of the live agent that keeps the session's id and names that same id as its parent, and the fork's turn still ran OMH's pre_llm_call. A review runs right after a turn, so it took the board card naming this chat's open lanes, counted toward the plan line's turns with the full reconciliation rule, and replaced the session's last route-hint and skill-candidate claims; a side question asked mid-turn cleared that turn's Jev marker, so an omh_jev_ask the person had asked for returned consent_not_observed. The fork was handed OMH context of its own as well: with Hermes' review prompts, the primer, a route hint, the candidate line, the plan line, and the board card. pre_llm_call now returns nothing for a turn whose parent_session_id equals its session_id, before it reads or records anything, which is what Hermes 0.21.2 and later do by not calling the hook for such a fork. A delegate child, which names its parent's id, and a turn that names no parent are unchanged. The held-write line's own fork check moves to that one place. Other hooks a fork reaches are unchanged.

    Bounded for publication: the first 79 of 113 entries. A GitHub release body accepts 125000 characters and this section is longer. The remaining 34 are in CHANGELOG.md at tag v4.0.0.

  11. 3.0.2

    Expandir

    GitHub Release

    oh-my-hermes 3.0.2 · stable

    Everything merged since 3.0.1 (2026-10-07), 33 pull requests. Per-entry detail is in the 3.0.2 section of CHANGELOG.md.

    Why this release

    Two themes. Skills now hold a stricter line on how work is planned, debugged, tested and approved (SkillEx), and parallel coding agents can no longer run, reach the network, or leave settings behind outside their fence (FenceGuard). Memory capture also stops admitting instructions disguised as facts.

    Highlights

    • SkillEx: skills that work more carefully. (#2052–#2056, #2060–#2065, #2072)
      • A workflow name inside a block quote, a relayed line or a link no longer starts that workflow, and a delegated child is not told to start an orchestrating engine its brief names.
      • plan lists up to 5 inputs or failure modes no planned test covers; ralplan plans for the target state of whoever consumes the change and counts acceptance only from the user's answer.
      • app-debugging stops after three failed fixes on the same reproduction and re-examines the architecture.
      • loop reviews its design before launch: a finish line a program can decide, and a fence of paths checked against the diff.
      • ulw-work asks what a new test guards before it counts, settles small reversible questions as recorded Rulings, and caps review rounds.
      • deep-interview: a yes to an idea approves only the idea.
      • agent-debug reads past sessions read-only with every read length-capped; backend names one authoritative contract per service boundary.
    • FenceGuard: fanout stays in its fence.
      • Fanout no longer runs a unit unfenced without being told to. (#2026)
      • Git the dispatcher runs inside a fence gets no credentials and no network; host git no longer runs a unit's planted core.fsmonitor, core.hooksPath or filter driver. (#2034, #2039)
      • Diagnostics analyzers run fenced to their snapshot; the workspace preflight writes nothing into the unit's git directory, and a preflight the fenced tree failed is no longer reported as a host failure. (#2038, #2044, #2046)
      • A pre-spawn refusal no longer exits 0; Codex starts inside the macOS fence. (#2033, #2025)
    • Memory.
      • A model capture phrased as an order to its reader (you must…, Korean request endings) is refused with a next_action to restate it as an observation — no review step. (#2064)
      • Spelled-out, vague and past-relative time ("two weeks ago", "last Monday", "il y a une semaine") routes a capture to review. (#2043, #2051, #2059)
      • Recall reads French: accented Latin letters no longer split a word. (#2028)
      • A default-store install names ~/.omh for the plugin, so multiplexed processes load it. (#2047)
    • Also. Concurrent TUIs share one fanout scan and an idle widget polls less (#2066); model setup binds aliases on Hermes custom-provider machines (#2023); omh quality-evidence file-activity fails closed on symlink escapes (#2032); omh_agent_board survives Hermes dropping the bundle's modules (#2024).

    Updating from 3.0.1

    Run omh update, then restart any running Hermes session (TUI, Desktop, gateway). omh doctor confirms the active generation.

    Install

    curl -fsSL https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.sh | sh      # macOS / Linux
    irm https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.ps1 | iex           # Windows
    omh setup
    

    Release evidence

    Check Status
    Local release checks on the bumped tree compileall, docs workflows/roles/navigation/claims --check, release checklist --json, and the auto_release, build_identity, capability_impact, cli, hook_manifest, release_smoke and plugin_bundle_catalog_scan tests (500) pass; full suite at main d1e26086c: 15,765 tests OK
    Live Hermes smoke not run for this cut (neither --live nor plan-only); install path covered by test_release_smoke and the shipped-skill scan gate
    CI release prep #2073 green 11/11 (one Windows shard rerun on the known O4 observer-drain flake); main at 117f67a23 green (run 38115832238)
    Distribution run 38115835509 success (tag push; npm environment approved by rlaope, the tag was pushed under sionic-khope)
    Wheel oh_my_hermes-3.0.2-py3-none-any.whl sha256 2df56ebd191b290ec44ecb408f98354ba99a24544e835457185bd1b3174ae647, attached with omh-release-evidence-3.0.2.json
    npm published with signed provenance (sigstore log index 3202426721) at 06:57Z; npm reported "being processed" and the registry served 3.0.2 as latest from 07:01Z, about four minutes after the publish step
    Homebrew tap rlaope/tap/omh at 3.0.2, wheel sha256 2df56ebd… matching the release asset
    GitHub Pages ran on the version-bump merge (38115832244) and succeeded; site/index.html badge reads v3.0.2 on main
    Plugin bundle changed: routing reference regions and delegated-child route hint, memory instruction-shaped admission, awareness, default-store binding
    Workflow-learning review queue not applicable, no learning contract changed
    Machine sync after the cut none yet

    Not observed in this release run

    • Hermes's real delegated-child process model: the route-hint gate is documented to fail open for a child running in another process.
    • Every runtime claim above the plugin install is wrapper-observed evidence, per docs/RELEASE.md. A routed model is not proof that a provider serves it.
  12. 3.0.1

    Expandir

    GitHub Release

    oh-my-hermes 3.0.1 · stable

    Everything merged since 3.0.0 (2026-09-29), 56 pull requests. Per-entry detail is in the 3.0.1 section of CHANGELOG.md.

    Why this release

    Long-term memory was being written but never read back: the turn-start hook order left every prefetch empty. 3.0.1 fixes that and removes the human approval step from the memory loop, then takes the opportunity to give several things that were copied by hand a single owner — and fixes the two defects that work uncovered.

    Highlights

    • Long-term memory that works without an operator.
      • The memory provider renders its pack in the order Hermes calls its hooks, so recall reaches the model again. (#2001)
      • The model records memories itself through omh_memory capture; safe content is approved automatically, review-first stays available as a policy. (#2001, #2002)
      • Admission (capture, duplicate check, auto-safe approval, durability receipt) lives in the plugin bundle and the tool calls it in-process — no more locating an installed omh executable. (#2012)
      • omh memory pack scopes records the way recall does: explicit thread and project-identity packs are populated, a thread record travels only with its own session, and a pack for the resolved identity is the default pack under its own name, conflicts included. (#2016, #2019)
      • The bundle's store lock retries Windows permission denials with the one backoff definition the stores already used. (#2017)
    • One owner instead of hand-kept copies (no behaviour change). The router's prepared query, the run-record file names and schema versions, the per-turn budget ledger, the per-skill record, the executor-name phrase groups, the portable overrides and the per-turn text budgets each have a single definition that everything else reads. (#2003–#2011)
    • Import order can no longer change routing. The plugin bundle's standalone fallbacks bound themselves whenever the router was still mid-import, so /loop … was classified differently depending on which module a process loaded first. The guards and classifiers resolve per call, and the cycle itself (skills/__init__ loading the renderer eagerly) is gone. (#2013, #2015)
    • Every shipped skill installs under the Hermes install scanner. hermes skills install refused omh-routing and six other skills on wording alone (setgid, a > .claude/settings shape, "do not tell the user", a cryptocurrency name, sudo). The sentences are reworded at the source, and CI now runs the Hermes scanner itself over every shipped skill. (#2014, #2018, #2000)
    • Fanout units stay inside their worktree: Git runs there, commits land on the unit's own branch, Git LFS repositories work, a busy unit exits 75, and one function assembles every unit prompt. Windows dispatch stays unconfined as a measured finding.
    • Models. GPT-6.1 Sol, Claude Sonnet 5.5 and Grok Build 0.1 are registered and measured; Bedrock Claude ids resolve to the claude family; named custom providers route through delegation; the codestral calibration has a live measurement.
    • Install and update. omh setup --omh-home X binds the plugin to X; omh update keeps full-only skills and takes back a bot profile's store setting once its .env names the key; install.ps1 works on Windows PowerShell 5.1; the plugin declares a Hermes floor, not a window.
    • Replies and diagnostics. The persona owns the voice (MODEL_OPTI.md records which models keep a non-English persona's language); the per-turn route hint names only installed skills; one-word English replies decline the route question; agent-debug produces a bounded, cited incident report.

    Updating from 3.0.0

    Run omh update, then restart any running Hermes session (TUI, Desktop, gateway). omh doctor confirms the active generation. Hermes skill tap: hermes skills tap add rlaope/oh-my-hermes then hermes skills install rlaope/oh-my-hermes/skills/omh-routing --yes — this now passes the install scan.

    Install

    curl -fsSL https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.sh | sh      # macOS / Linux
    irm https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.ps1 | iex           # Windows
    omh setup
    

    Release evidence

    Check Status
    Local release checks on the bumped tree the Cut Release run 37589768350 ran the full unittest suite and compileall on the bumped tree (success) before the main ruleset refused its push; the identical bump landed through #2020 with CI 11/11 green; release checklist --json, every docs … --check gate and the version-surface, release-smoke, build-identity and plugin-scan tests pass locally on the bumped tree
    Live Hermes smoke not run for this cut (neither --live nor plan-only); the install path is covered by test_release_smoke and the shipped-skill scan gate, and no install surface changed since 3.0.0's live smoke
    CI main at 9f24f873e green (run 37588105351); every PR in the range merged on a green CI
    Distribution run 37597439093 success (tag push; npm environment approved by rlaope — the tag was pushed under the sionic-khope credential, which the self-review guard then excluded)
    Wheel oh_my_hermes-3.0.1-py3-none-any.whl sha256 2750387eee094a2ef12dc9dd23bdbc08b0b24a3d08d08cf241cb9a715d02021e, attached with omh-release-evidence-3.0.1.json
    npm published with signed provenance (sigstore log index 3127998552) by run 37597439093; npm processed the publish asynchronously ("being processed") and the registry served 3.0.1 as latest from 09:25Z, about twelve minutes after the publish step
    Homebrew tap rlaope/tap/omh at 3.0.1, wheel sha256 2750387e… matching the release asset
    GitHub Pages ran on the version-bump merge (37597431234) and succeeded; site/index.html badge reads v3.0.1 on main
    Plugin bundle changed: memory provider hook order, memory admission + store I/O, awareness guards per call, turn budgets, run records, executor cues, store lock retry
    Workflow-learning review queue not applicable, no learning contract changed
    Machine sync after the cut none yet

    Not observed in this release run

    • The reworded skills were scanned with the same skills_guard.py Hermes ships; a hermes skills install from the published tap against 3.0.1 has not been run from a user machine yet.
    • Every runtime claim above the plugin install is wrapper-observed evidence, per docs/RELEASE.md. A routed model is not proof that a provider serves it.
  13. 3.0.0

    Expandir

    GitHub Release

    oh-my-hermes 3.0.0 · stable

    Everything merged since 2.0.5 (2026-09-22), 98 pull requests. Per-entry detail is in the 3.0.0 section of CHANGELOG.md.

    Why a major

    OMH has changed what it does with work, not only how it routes it. Until 2.0.5, OMH prepared plans, handoffs and routes and then stepped back. In 3.0.0 it keeps going until the work is checked:

    • a todo item closes on a recorded result;
    • a failed fanout unit is repaired in place;
    • an accepted plan can run as a team of Hermes delegates, each task closing on its own check.

    Hermes users will see the change in the reply as well.

    Highlights

    • Work that finishes.
      • A todo item marked done without a recorded result (tool call, file write, PR, CI run or team check) stays done_unverified and keeps the loop going. (#1928)
      • A fanout unit that fails its checks is repaired in its own worktree. This repeats until the checks pass or the repair budget is spent (default 2, cap 3); a spent budget stops with the reason named. (#1929)
      • A unit's "done" is held to the tests its own changes reach. (#1876, #1866)
    • Teammates on Hermes' runtime. omh_team turns an accepted plan into a checked delegate team:
      • each task has an owner, its dependencies, and one verification command;
      • that command is shown in full when the plan is approved;
      • a task closes only when its check was observed passing;
      • a forged record runs nothing;
      • check commands are held to a refusal list for shell escapes and invisible characters. (#1931)
    • Replies in the host's voice.
      • OMH's record vocabulary (schema ids, "contract", "receipt", "lane") stays out of the reply.
      • Orchestration decisions each carry one plain sentence in the reply language: the team split, which specialist took a task, the goal and done criterion, and a blocked reason. (#1930, #1837, #1851, #1834)
      • omh quality-evidence reply-lint checks those rules against real Hermes sessions. (#1848)
    • Hermes Desktop.
      • The bundle ships a Desktop half: an OMH pane and a status item carrying the TUI widget's information. (#1856, #1864)
      • Work started on one Hermes surface is picked up on another. (#1878)
    • Routing and skills.
      • Dispatch happens only on strong evidence; otherwise Hermes gets a lexical shortlist of candidate skills. (#1879, #1880, #1886, #1891)
      • Everyday-sense phrases no longer reach a skill. (#1910)
      • Every installable skill is gated on measured routing reach. (#1875, #1899)
      • New skills: app debugging, commit/PR authoring, git workflow, relational databases, security-event response, release cuts, agent instructions, IaC changes, data pipelines, model fine-tuning, mobile release, internal audit. (#1912, #1913, #1916, #1918, #1919, #1920, #1921, #1925, #1927, #1934, #1936, #1937)
    • Cost and evidence.
      • "How much did this task cost?" is answered from recorded usage, including the kanban workers a conversation started. (#1877, #1897)
      • Served Ultrafast tiers are priced. (#1854)
      • quality-evidence reports on:
        • per-surface utilisation (#1855);
        • file activity (#1922);
        • agent-debug findings: tool errors, identical retries, unnotified background processes and compactions (#1924).
    • Models.
      • Claude Opus 5.5, GPT-6 Luna and GPT-6 Sol are onboarded. (#1833, #1841, #1859)
      • OMH can ask Jev itself, only when the user names Jev in that turn. (#1844)
      • omh_jev_ask no longer refuses every call as consent_not_observed. (#1941)
    • Install and update.
      • Reinstalling keeps files you added to the plugin directory. (#1940, contributed by @junjuny0227)
      • omh update tells you what changed for you. (#1874)
      • A home on the pre-pointer skills path is migrated rather than registered twice. (#1860)

    Updating from 2.0.x

    Run omh update, then restart any running Hermes session (TUI, Desktop, gateway). The plugin and widgets load at session start. omh doctor confirms the active generation.

    Install

    curl -fsSL https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.sh | sh      # macOS / Linux
    irm https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.ps1 | iex           # Windows
    omh setup
    

    Hermes skill tap: hermes skills tap add rlaope/oh-my-hermes then hermes skills install rlaope/oh-my-hermes/skills/omh-routing --yes.

    Release evidence

    Check Status
    Local release checks on the bumped tree 33/33 pass: compileall, every docs … --check gate, ruff, harness validate, release checklist, skill-content-smoke, product-readiness, cases demo/readiness, demo routing-precision, probe, setup --dry-run, hermes-smoke, install-smoke, uv build, wheel install + omh --help/--version + wheel hermes-smoke --include-command-smoke, install.sh from the local wheel
    Live Hermes smoke release hermes-smoke --live --install-path setup with the 3.0.0 wheel against the real hermes binary in temporary homes: 5/5 steps (omh_setup, tap_list, skills_list, skill_check, setup_doctor)
    CI #1942 and #1943 11/11 green; main at dba9e54c green (one Windows quarantine timing test re-ran green)
    Distribution run 36515352264 success. The first run (36513602745) stopped at notes_source_mismatch before anything was built or published: the evidence check compared the size-bounded notes with the unbounded section. Fixed in #1943, and the tag was moved to that merge commit before any artifact existed
    Wheel oh_my_hermes-3.0.0-py3-none-any.whl sha256 6a05f40a…, attached with omh-release-evidence-3.0.0.json
    npm published with signed provenance (sigstore log index 2992410550); registry latest observed at 3.0.0
    Homebrew tap rlaope/tap/omh at 3.0.0, wheel sha256 6a05f40a… matching the release asset
    GitHub Pages ran on the version-bump merge and succeeded; live badge observed at v3.0.0
    Plugin bundle changed: omh_team, todo evidence, reply voice, Desktop half, Jev consent
    Workflow-learning review queue not applicable, no learning contract changed
    Machine sync after the cut none yet; no machine has run omh update against 3.0.0

    Not observed in this release run

    • No live Hermes turn ran omh_team with a real delegate team. The lane is proven by tool payloads, hook payloads and the pinned attack strings, not by a model driving it.
    • The reply-voice rules are verified by lint over recorded sessions and by tests. No post-release session has been linted yet.
    • Every runtime claim above the plugin install is wrapper-observed evidence, per docs/RELEASE.md. A routed model is not proof that a provider serves it.

    The full entry list for this release is longer than a GitHub release body accepts, so it lives in the 3.0.0 section of CHANGELOG.md.

  14. 2.0.5

    Expandir

    GitHub Release

    oh-my-hermes 2.0.5 · stable

    Everything merged since 2.0.3 (2026-09-12), 142 pull requests.

    2.0.4 was tagged and never published. Its cut ran green and pushed the tag, then the publication step was refused by GitHub — HTTP 422 ... body is too long (maximum is 125000 characters) against a 228,212-byte release body — and every step after it was skipped, so nothing reached npm, Homebrew or a release page. The tag stands as a record; 2.0.5 carries all of its work. That is why this release is large: 133 of these pull requests are the 2.0.4 range, described in the 2.0.4 section of CHANGELOG.md (103 entries), and the nine below it are new.

    Why this release

    Most of the 2.0.4 range was reported by someone who could not get past it. An install on a machine whose python3 is 3.9.6 failed with a resolver message about a version bound, and the omh that failed to install was then not on the path to say so. A menu redraw threw away the arrow keys pressed during it, and a free-text prompt answered the escape bytes instead. omh uninstall reversed one of the seven config.yaml keys setup writes. A skill named with an ordinary English word claimed every sentence that happened to spell it. A model that had made the same tool call four times running was warned, and made it again.

    The nine that follow are mostly the release pipeline itself, found by the 2.0.4 failure and by sweeping the surfaces it touched.

    Since the 2.0.4 tag

    • A planning run asks before it starts editing. A plan stamped awaiting_acceptance that reaches a file-mutating tool raises the host's own approval prompt rather than a refusal the model can ignore. Escalation is bound to the session that owns the plan — a stamped record with no owner escalates for nobody, which a reviewer caught after the full suite had passed over it. (#1794)
    • A release body is bounded by what it is published to. The bound that existed was 262,144 bytes against a surface that accepts 125,000 characters, so it could never have caught the 2.0.4 overflow. An oversized section is now trimmed to whole entries plus a trailer naming what was left out and where; a section inside the bound is returned byte for byte. Counted in UTF-16 code units, the wider of the two readings of "characters". (#1798)
    • A cut asks for the site rebuild its own push cannot start. A cut pushes with GITHUB_TOKEN, and GitHub creates no workflow run from an event that token caused; Pages listens for a push touching site/**, which the version bump always rewrites. So the one push that changed the advertised version was the one push Pages never saw, and the public badge kept the previous version until an unrelated merge happened to touch a watched path. The cut now dispatches pages.yml, with continue-on-error so a failed rebuild cannot repaint a finished release as a failed one. (#1800)
    • Non-generative models are recognized and refused coding work by name, omh doctor reports which Jev-class plugins a machine holds, an undecidable route records the typed question it could not answer, and the routing corpora can be asked as typed questions anything can answer and scored. (#1804, #1806, #1811, #1812)
    • Tool identity is complete or it is unknown. Argument fingerprinting digests the whole argument set or marks the identity unknown rather than fingerprinting a prefix. (#1805, from an outside contributor)
    • The HUD route label drops a comparison its reader could not check. A lane routed to the model the parent session runs read category:deep(model =parent); the token compared the row's model against the parent's, which no row on that screen shows. The tokens that remain each name an event in the route. (#1797)

    Highlights of the unpublished 2.0.4 range

    • Install works on the Python you have. The installer picks an interpreter the wheel can be installed into, names the ones it rejected and why, and provisions one when none qualifies; install.ps1 mirrors it, held by an environment-variable parity gate. (#1771, #1636)
    • The setup wizard stops fighting the keyboard. Menus keep keys pressed while redrawing and read a whole keypress instead of its first three bytes; control input typed into a free-text prompt is stripped rather than answered and echoed back; the wizard says where you are in its questions, and asks once whether OMH should watch for its own updates. (#1779, #1774, #1775, #1784, #1586)
    • Uninstall reverses what setup wrote — all seven config.yaml keys rather than one, under a single write discipline, with no emptied containers left behind. (#1764, #1765, #1770)
    • A repeated tool call is refused, then put to the person. The guard is cycle-aware and result-aware, an identical repeat is refused outright, an ignored refusal reaches a person, and the HUD shows repeat xN. (#1708, #1745, #1760)
    • Routing stops claiming sentences it does not own. A skill's own name no longer claims every sentence that spells it; three reported misroutes reach the skill that owns the question; ulw plan reaches the plan engine; saying a plan is fine no longer restarts it. (#1758, #1606, #1608, #1661, #1645, #1659, #1631, #1759, #1747)
    • The HUD tells one session's work from another's — each TUI's plan panel renders its own session, a delegate row gets its route and label back, and the HUD attests the model that answered against the model requested. (#1673, #1791, #1776, #1744, #1736, #1588)
    • A plan is a thing you can finish. A ten-phase code story ships as a stampable template; the deep guide, ELI5 pass, quiz, manual test guide and the close that ends a story each got an owner; a blocked item's recorded reason renders on the row. (#1780, #1781, #1785, #1787, #1660, #1686, #1616, #1647, #1549, #1705, #1751)
    • The plugin bundle survives its own failures — it loads without the omh package, and a session no profile owns is no longer vetoed out of every tool call. (#1669, #1677, #1762, #1763, #1761)
    • Long documents are planned before they are read — numbered read ranges, one frozen unit per range, a very large PDF read in page-anchored ranges, and a refused re-read that names the offset. (#1532, #1537, #1533, #1531, #1735, #1530)
    • Audit, fences and readiness. Declared hook semantics classified before you enable a plugin, plugin self-replacement paths named, fanout's write fence fixed on Linux, and three readiness checks — model data policy, plugin enforcement, profile isolation. (#1555, #1685, #1702, #1601, #1684, #1630, #1587)
    • Models and providers. A model-chain picker in the CLI and /omh-model in the Modern TUI, both editing the store a bot profile dispatches from; providers Hermes is linked to counted without an interview; DeepSeek V4.1 Flash's override measured and revised. (#1506, #1507, #1680, #1519, #1524, #1628, #1526)

    Updating from 2.0.3

    omh update, then restart any running Hermes session — the plugin and TUI widgets load at session start, so a running session keeps the old ones. omh doctor confirms the active generation.

    Install

    curl -fsSL https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.sh | sh      # macOS / Linux
    irm https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.ps1 | iex           # Windows
    omh setup
    

    Hermes skill tap: hermes skills tap add rlaope/oh-my-hermes then hermes skills install rlaope/oh-my-hermes/skills/omh-routing --yes.

    Release evidence

    Check Status
    Full unit suite + compileall on the bumped tree passed inside the Cut Release run (35707043022)
    CI on the tagged tree's parent (2c09f7b6) 13/13 jobs green (35706998373): sharded suite on 3.11 and 3.12, both Windows shards, three quarantine lanes, two distribution lanes, plan and aggregate
    Wheel built, verified immutable, attached oh_my_hermes-2.0.5-py3-none-any.whl + omh-release-evidence-2.0.5.json (35709525437)
    npm publish published with signed provenance (sigstore log index 2908869234); registry latest observed at 2.0.5
    Homebrew tap rlaope/tap/omh formula at 2.0.5, wheel sha256 98c33a89… pinned and matching the release asset
    Every version surface at the tag pyproject, version.py, plugin.yaml, the site badge and its four locale strings, and .release-channel all read 2.0.5, each checked against the committed tree at v2.0.5 rather than a working copy
    GitHub Pages ran on the cut commit and succeeded — the first cut for which it did, and the live badge is observed at v2.0.5
    Workflow docs, roles, capability families, ULW inventory and site regions byte-exact, checked in CI
    Workflow-learning review queue not applicable, no learning contract changed
    Plugin bundle changed: the plan-stage edit gate, the repeat guard, tool-argument identity, and the Jev model class
    Machine sync after the cut none yet; no machine has run omh update against the 2.0.5 tag

    Not observed in this release run

    • No cut has yet failed its Pages dispatch, so the continue-on-error path that keeps such a failure from repainting a finished release is proven by workflow text and its gate, not by a failure that happened.
    • The release-body bound did not fire here: this section is 30,266 bytes, well inside the limit. It is verified against the real 228,212-byte 2.0.4 body regenerated from that tag, not by this release exercising it.
    • No live Hermes turn exercised the plan-stage approval prompt; it is proven by hook payloads and the host's own approval path.
    • omh release hermes-smoke --live against a real Hermes profile was not run.
    • Every runtime claim above the plugin install is wrapper-observed evidence, per docs/RELEASE.md; a routed model is not proof that a provider serves it.

    • The cut now asks for the site rebuild its own push cannot start. A cut pushes with GITHUB_TOKEN, and GitHub does not create workflow runs from events that token caused. Pages listens for a push to main touching site/**, and the version bump always rewrites site/index.html and site/i18n.js — so the one push that changes the advertised version was the one push Pages never saw. Observed on the v2.0.4 cut: Distribution Release, dispatched explicitly, was the only run on that commit, with no Pages run and no CI push run beside it.

      The public badge therefore kept the previous version until some later merge happened to touch a path Pages watches, which left the page advertising a release as the last surface still naming the one before it. The cut now dispatches pages.yml next to the distribution dispatch it already makes for the same reason. No new permission — the job already holds actions: write. That step carries continue-on-error: true: by the time it runs the tag, the release and the distribution dispatch have all happened, and a site rebuild that fails for a reason of its own must not repaint a finished cut as a failed one. It still goes red on its own row. The distribution dispatch above is deliberately not treated this way — if that one fails nothing publishes, and red is true.

    • A release body is now bounded by what the surface it is published to accepts. The 2.0.4 cut tagged and pushed, ran its gates green, and then failed at publication: HTTP 422 Validation Failed / body is too long (maximum is 125000 characters) against a 228,212-byte section. Nothing was released, and the npm and Homebrew steps after it were skipped, so the version existed as a tag with no installable artifact. The bound that existed, MAX_NOTES_BYTES, is 262,144 — twice the limit of the surface the artifact is published to, which is not a bound at all, and it grew more likely to fire with every release that shipped more work than the last.

      omh release notes now trims an oversized section to whole entries plus one trailer stating how many it kept, how many remain, and where they are; a section already inside the bound is returned byte for byte, so an ordinary release publishes exactly its authored notes as before. The CHANGELOG stays the record and is never rewritten — the bounded body is a view of it. The length is counted in UTF-16 code units rather than code points, because GitHub says "characters" without saying which and the UTF-16 count is the larger of the two readings. A single entry that cannot fit raises rather than truncating inside it, since half a sentence published as a whole claim is worse than a refusal a person can act on.

    • A route OMH cannot decide now carries the question it could not answer, and an answer to it can be recorded and measured. When the deterministic router reaches an undecidable route -- a script its trigger tables do not cover, a near-tie between two candidates, or a low-confidence score -- it already hands the shortlist to model selection. It now also types that shortlist as a question: one relative Choice over the candidates plus none, and one absolute yes/no per candidate, built by the same builder the routing corpora are exported with, from the handoff's own candidates so the two cannot name different shortlists. omh chat route, omh chat interact, omh chat route-hint (as a new top-level key of chat_route_hint/v1) and the omh_interact plugin tool carry it. omh_recommend, omh_context and the awareness rail do not, and that is a statement about wiring rather than policy: those three build their own payload from the awareness hint and never see the core route.

      Nothing about the route changes. A decidable route carries no question at all, an unanswered question leaves the deterministic route in force, and a recorded answer does not re-route anything either. What an answer is FOR is measurement: omh_route_answer writes one route_question_answer/v1 per session and question under the OMH home, embedding the same routing_question_answers/v1 row that omh chat route-questions score --answers <dir> already reads, so an answerer is scored against OMH's own corpora instead of writing into a ledger nothing reads. An answer whose digest names a different question is refused rather than recorded, because a record joined to the wrong question measures the wrong thing.

      The record names the request it answers for, not only the question. It carries message_sha256 over the raw request and repeats it on the embedded row, so a row read out of a JSONL file still identifies its own request and a scorer can join on it. A caller sends the request and OMH derives the hash, or sends the hash when the text is not at hand; sending both with different values is refused, because they name different requests and choosing one would record the answer against a request nobody meant.

      confidence_source is derived from who answered, never taken as an argument and never called calibrated: self_reported when the host model answered about itself, answerer_declared when a plugin reported a number OMH did not observe. OMH sees only which answerer the caller named -- not the request, not the response, not whether anything was called at all -- and a field asserting a vendor's calibration would be a claim about a model written from a tool argument.

      The answerer ladder reports what could answer the question on this machine, one provable tier at a time: a Jev-class plugin whose manifest declares a jev_ tool is installed, one Hermes' own config also enables is enabled, and one whose tool has reached dispatch here at least once is observed. That last tier is a durable timestamp in the burst ledger rather than a scan of it, because the entry ring holds 200 calls and a scan would report a plugin as never used after 200 unrelated ones. Every tier under-claims by construction: a config shape the bundle's reader cannot follow reports installed rather than enabled, and an unreadable plugins directory drops the rung instead of inventing one.

      The ladder is a report and never a recommendation. No text OMH injects into a turn names a third-party tool, and no rail was added for this: OMH warning about a plugin's egress in omh doctor while nominating it in the prompt would be the same product saying two things. A test pins that the prompt context is byte-identical to the awareness text it was always built from.

      One rollout note: the HUD reads its required tool list from the bundle's own PROVIDED_TOOLS, so an install that has not run omh update will report the new tool as missing until it does.

    • The cases that gate OMH's router are now questions anything can answer, so a second opinion can be measured on the same ground. omh chat route-questions export projects both shipped routing corpora into typed questions: one relative Choice over the router's own candidate shortlist plus none, and one absolute yes/no per candidate. The two are deliberately separate questions. The Choice picks WHICH workflow and always returns one of its options, which is why none has to be one of them; each yes/no decides WHETHER that one workflow is what the request asks for, answerable without reference to the others. An answerer may pick a candidate and still say no to every fit, and the scorer reports that as it stands rather than repairing it.

      omh chat route-questions score reads answers and always includes the deterministic arm, so no arm is ever reported alone. It takes a JSONL file of routing_question_answers/v1 rows or a directory of route_question_answer/v1 records, which join a corpus item by question digest. A malformed row is counted and named by its line, never dropped: an arm that answered badly and an arm that did not answer are different results, and only one of them may be excluded from a denominator. Every rate is a reported_rate payload naming what it divided, so an arm with nothing to score reports percent: null rather than a confident zero.

      The deterministic arm's verdicts are the routing-precision corpus's own, read through two new accessors rather than re-derived. That is the whole load- bearing decision here. clarify with a named candidate is a PASS in the negative corpus -- the router asks one question instead of opening a workflow, picker, or handoff -- and the EXPECTED intervention in the positive one, so a single re-invented predicate cannot serve both. A measured draft that used one reported the shipped router as over-routing on more than a third of the negative controls, on a corpus this repository gates at zero, which would have put two OMH surfaces on record disagreeing about OMH's own router with the benchmark holding the wrong number. A test now pins the projection against build_routing_precision_demo case by case, and shows the pin failing when the verdict is replaced. That corpus's pass verdict on each case is also read rather than only carried: a score report fails by name when the corpus records a case the routing-precision gate fails, so a report cannot read clean while the gate reads red.

      A recorded answer joins back to its question by digest, and the digest names the request rather than the shortlist it produced. A digest over the question block alone is shared by every request the router shortlists the same way, which on these corpora is most of them, so a reader joining by digest would have scored an answer against whatever the first case behind it expected -- an arm that answered correctly recorded as having hijacked a negative control. The digest covers the request and its candidates, and where a digest still reaches more than one case, which is where the two corpora record the same request verbatim, the record is reported as ambiguous and scored against none of them.

      The message is the join key, and the digest reports the shortlist. A question digest covers the candidate list as well as the request, and the candidate list is cut per surface -- a route hint asks about two candidates where a full route asks about three -- so the same request asked on two surfaces produces two digests and a digest-only join would drop the answer. A recorded answer joins by message first; the digest then says whether the shortlist was the same one, reported per answer as digest_match and counted per arm as digest_mismatch. A mismatch is scored, because the answer is about that request; it is reported, because an arm asked about a shorter list was not asked quite the same question.

      A corpus question is now built from the same place a live route builds one. A live route attaches a question only where it could not decide, and it builds it from the undecidable route's candidate handoff; the projection read the route's public recommendations instead, which drop each candidate's description and carry the route's prose reason rather than the handoff's machine codes. The two questions therefore never shared a digest, and an answer recorded on a live route scored as having answered nothing. Both sides now call one builder with one set of inputs. The cases a live route never questions -- the ones it decided -- still carry a question for the offline arms, and an arm answering recorded live routes is denominated on the undecidable cases alone, with the rest named not_live_joinable in each rate's excluded list rather than counted as questions it failed to answer.

      An answers file is written by a model or by whoever ran an external arm, so it is read as untrusted input. Every file is bounded before it is read whole, every string that reaches a report line is stripped of control characters and capped, because a report is an artifact somebody attaches to a PR and an embedded newline in an arm name forges a line in it. A row claiming the reserved deterministic arm name is refused by name rather than merged into a tally the report then replaces.

      No routing case was added. The projection reads the two tuples and adds nothing to either, because their lengths are pinned across several test files and a benchmark is not a reason to move a gate. The expected answer for each case comes from the case's own record, including the one intervention case whose correct answer is to open nothing and the clarify cases that pin no candidate; both map to none, which the corpora say and the projection does not decide.

      benchmarks/routing-questions/v1 is the lane. It answers batches through omh coding hermes-child dispatch --confirm-dispatch, with the prompt on stdin and the answers collected from a file the prompt names, because that boundary reports usage metadata and never model text. It is offline by default and refuses a live run without --allow-paid-live, --max-paid-calls and --confirm, with a second refusal inside the library so importing it and calling a harmless-sounding function cannot spend anything. It never imports omh: the lane measures the product a person installed, through its executable, not the checkout it happens to sit in. The arm that runs the operator's own authenticated Hermes rather than the isolated child has its process directory and its TERMINAL_CWD both pinned to the batch workspace, so a model with the file toolset resolves relative paths there and not in the checkout the lane was launched from. Answers are written as each batch returns, so a run that stops partway keeps every answer it has already paid for, and a model-written row is narrowed to the documented answer keys before it is persisted. No live arm has been run, and running one is an operator's call.

    • omh doctor now says which Jev-class plugins a machine holds, and what each one's catalog entry declares. Jev (TypeSafe System One) is a non-generative decision model: it answers typed questions and cannot write code, so OMH never routes it and never calls it. Community Hermes plugins do call it, and several ask for hooks the OMH bridge also registers — most sharply pre_llm_call, where OMH injects its route hint while a Jev skill router nominates a skill, so one message can reach the model carrying two nominations. Nothing in the CLI said that was happening.

      The new plugin_jev_sidekick check is appended on every run, in both branches: a machine with no signal reports optional: no Jev-class plugin installed, and a machine with one reports the tier reached — installed, enabled, or credential name present — plus one note per plugin naming the tools and hooks its manifest declares, the hooks it shares with the OMH bridge, and, verbatim, whatever its catalog entry discloses about what leaves the machine. Every note says "declares" and carries the artifact and date OMH read it from, because a catalog description is its author's statement and a plugin.yaml is whoever installed it: neither is evidence that the plugin ran or that any request left the machine. omh doctor --json carries the full jev_sidekick_posture/v1 payload.

      Three local reads and nothing else: the plugin directories under $HERMES_HOME/plugins, Hermes' plugins.enabled list, and the variable NAMES in $HERMES_HOME/.env. Manifests go through the existing bounded subset reader, so a construct it does not model is reported as unread rather than guessed at — an unread provides_hooks leaves the hook overlap unestablished, which the check says instead of reporting no overlap, and an unread plugin directory stops the check from reporting absence at all. The credential scan reaches its names through a new allowed parameter on env_key_names rather than by enrolling them in HERMES_ENV_KEY_PROVIDERS, which feeds provider entitlements and would reorder mixture chains around a model that can never be a chain member. No value is read, anywhere.

      A read that did not happen is reported as itself everywhere, not only for hooks. Enablement is enabled, not enabled, or unknown with the reason attached: Hermes' plugins node written as a flow mapping is valid YAML Hermes loads and a form the shared block reader walks past, and it produces the same empty list a config that enables nothing produces, so the check says which of the two it read rather than asserting the commoner one. A manifest that declares a name OMH could not read is classified on its jev_ tools alone, because the directory name is then a guess and a guess that lands on a catalog entry would attach that maintainer's verbatim egress disclosure to this install.

      Everything the sweep reads is bounded and untrusted. A symlinked plugin.yaml is refused and named for the reason the symlinked-directory guard already states, the manifest cap is taken at the read rather than from a preceding stat that a zero-length file walks past, Hermes' config is bounded like the manifests beside it, and a directory name — which passes no reader and is chosen by whoever created the directory — is stripped of control characters and length-capped before it reaches a payload field or a report line an operator pastes elsewhere.

      ok stays True in every branch, including the warning one: a third-party plugin an operator installed deliberately is not an OMH install failure and must not flip the doctor exit code. The next action is built from the branches that fired and names the .env path OMH read rather than a default spelling of it, and it describes an overlap as the declaration it is: the bridge registers eight hooks, only one of which is the nomination surface, so what a shared hook supports is "declares the same hook OMH registers" and not a sentence about what the plugin does. The check groups into optional_surfaces by its plugin_ prefix. Check gains an optional detail payload, None for every check that carries no structured finding, so observed keeps its one meaning as a boolean.

    • OMH now recognizes a model class it cannot route coding work to, and refuses that work by name instead of preparing it. TypeSafe's Jev answers a typed Choice, Score, or Noul over options the caller supplies and its own jaggedness page states it "is not trained to generate text". Until now jev-1.13.0 fell to family unknown, inherited generic prompt discipline, and would have been routed an implementation unit like any other id, with the failure arriving from the provider rather than from OMH.

      The class is model_class(), read off the family, with generative as the default so every model the catalog has never met routes exactly as it did before. non_generative is refused at each surface that would hand a model work to write: omh coding model-route --model jev… answers status: model_refused and prepares no model, no effort, and no chain; omh model-chains set exits 2 and writes nothing; the operator category-maestro config rejects such an entry by name on read and on write; and a chain entry that reaches the resolver anyway, from a hand-edited chain document or an operator recommendation document, is skipped with a record naming the class while the next generative entry takes the head — on the catalog lane and on the Hermes editorial lane alike. The refusal of a requested id is decided before any catalog is consulted, so neither lane can route around it.

      This narrows one stated invariant and the docstring that carried it moved in the same commit: an explicitly requested model still wins over every catalog and is still never adjudicated on quality, but it is now refused when its class cannot do the work at all. The refusal payload carries a refusal record with the id and the kind, so a JSON consumer never parses a sentence to learn what happened, and the key is present only on a refused route, leaving every payload that routes today byte-identical.

      Recognition is the rest of the onboarding, done honestly rather than skipped. jev-1.13.0 is an exact contract read from docs.typesafe.ai on 2026-09-21, with jev-latest and jev-preview as declared alias rows because both move on the next release. It carries no effort ladder, and model_class, question_types, max_choice_options, and rate_limits are optional keys that do not move model_contract/v1, on the same terms served_ids and limits_note joined it. max_output_tokens is 0 rather than absent: output is typed answers billed at zero, which is also why the price row is (0.042, 0.0) — a zero is a published rate, and a missing row would make every run on the model report no cost at all. The bare word jev is recognized as a family, because TypeSafe sends it in the model field and a gateway spells it typesafe/jev, but it inherits no contract: no vendor page describes that spelling. Both classifier surfaces read the bare word the same way, so a dynamic workflow spec that names jev describes a model target rather than an agent.

      The coverage dimensions that stay missing for reasons of this class are effort, category_projection, and provider_eligibility, and a test pins that reading. effort is missing because there is no effort parameter; category_projection and provider_eligibility are missing because the model is in no shipped chain, which is the decision rather than an oversight. The only levers that would turn them green are an invented rung, an invented chain entry, or an intentional_exclusion row that would claim OMH deliberately does not cover a model it does cover. The same row also reports data_handling as missing, which is not about this model — every contracted model in the catalog reads the same way — and reports calibration as covered through the generic fallback, which records that the generic discipline applies rather than claiming a model that takes no prompt was calibrated.

    • The HUD route label drops its parent-equality token. A lane the tool routed to the model the parent session itself runs rendered as category:deep(deepseek-flash:high =parent). The token said the dispatch cost what the parent costs, but it compared the model named on the row against the parent's model, which no row on that screen shows — so the reader could not check the claim against anything in front of them, and on the owner's deep chain, whose head IS the session's model, it rode every child row. The lane now reads category:deep(deepseek-flash:high): the category names the lane, the parentheses name the model that ran, and the state tokens that remain (fallback, inherit) each name an EVENT in the route rather than a comparison. Nothing about which model ran is lost — that was always the model name itself. The read_omh_hud payload still reports same_as_parent for a caller that holds both models; the widget no longer reads it, and the widget-pack gate now pins that absence.

    • A planning run that starts editing now asks the person first, instead of being told not to. Give ralplan a prompt with implementation intent folded in — "ralplan implement the refactor now and open the PR" — and it implements. Its skill has refused that in prose since it shipped: one line routes a full delivery cycle to ultrawork, a worked example is built from that exact sentence and answers "stop at the reviewed plan", and a third says to start a follow-on engine "only on the user's explicit go-ahead". Routing is not what failed either; measured on build_chat_interaction_payload, the sentence dispatches to ralplan at score 12 with next_action: present_plan. The message reaches the right skill and the skill's own rule is then read past, so a fourth sentence was not the fix.

      plan_stage_gate is the enforcement that was missing. A planning run stamps its checklist plan_stage: awaiting_acceptance when it declares it (omh_todo, a new closed-vocabulary field on omh_todo/v1), and while that holds, a write_file or patch in the owning session is escalated from pre_tool_call to the host's human-approval gate. That gate is the only directive the hook offers that a model cannot decline — block becomes a tool result, and one measured session read 185 host refusals and repeated the call anyway, while approve is resolved by tools/approval.py::request_tool_approval and never returns to the model as text at all. It is also, literally, somewhere to put the question, and the question is the feature: this session declared a plan and has not recorded you accepting it, and this call edits a file. Approving implements now; denying stops at the reviewed plan.

      Nothing reads anybody's prose. Two fields decide it — the stamp, and the plan projection's own established verdict, reused rather than re-derived so the gate cannot disagree with the HUD about whose plan a record is or when it went stale. This repository has already paid for the alternative: the plan's stop criterion used to be matched out of item text and read "verify the retry is not blocked on the session limit" as blocked while missing "waiting on the owner's review", wrong in both directions on ordinary input. The guard against its return is a mutation table rather than an assertion about today's wording: the same call is made twice over records whose every word argues the opposite of what the field says, and the verdict follows the field.

      The approval gate is the strongest thing OMH can do to a person's turn, so every state the records cannot settle is silence. No record, a record this session does not own, a stale one, a finished plan, an unreadable home, a stamp this build cannot classify, a call the host named no session for, and every tool that is not one of the host's two file-mutating ones all return nothing — the discipline #1738 landed for the unarmed-wait directive, applied to a stronger action.

      Ownership is asked separately from freshness, and the separation is a correction rather than a flourish. The renderer's established verdict does not answer "whose plan is this": its identity rule reads an unattributable plan as BELONGING, deliberately, so a real checklist is never hidden from its owner on missing evidence. A gate that stops a person's turn needs the opposite default — and a home-wide stamped record, which the chat writer produces whenever the host names no session on the same call that carries the stamp, projects as established for every session id that asks. Reusing the renderer's verdict as ownership armed this gate for three unrelated sessions under review. The projection now states the raw fact, own_record, decided after transport ids are translated to durable keys so a comparison of raw ids cannot refuse a record the session does own; the gate reads that first. A stamped record with no owner is unattributable by construction, and refusing it loses nothing. Absence of the field is unknown rather than "not accepted", so a delivery checklist, a CLI write, and a record predating the field are all ungated; reading absence the other way would put this in front of every edit anyone makes with a todo list open.

      An unattended run is not gated at all, and that is a decision rather than an oversight. request_tool_approval fails CLOSED with no person present, so escalating on hermes chat -q, a webhook, or a delegated child would not ask anything — it would refuse the edit with the host's wording for a rule nobody can answer, which is how a gate meant to protect a person becomes the reason their automation stopped. The predicate is the one the repeat guard's stage two already uses, and the fallback is deliberately not the repeat guard's block: a loop that keeps running is worse than one that is stopped, while a session doing ordinary work is not. Cron is the stated exception and cannot be closed from a plugin — HERMES_CRON_SESSION is a ContextVar, never a process variable — so a cron turn reads as attended; at the approvals.cron_mode default of deny the edit is refused with the host's wording, and only a cron job that both stamps the field and then edits files reaches it at all.

      The [a]lways allowlist grain is set explicitly to the constant omh_plan_stage_edit, and the reason is not the one request_tool_approval documents for itself. That function hashes the reason into a key when it is given none, but a plugin directive never reaches that branch — _resolve_block_from_details calls it as rule_key or tool_name — so an omitted key is the TOOL name, and what a person's [a]lways would store is plugin_rule:write_file. That grain is wrong twice: it re-asks at the next patch what a write_file prompt already settled, and being named after the tool rather than the rule it would permanently allowlist that tool for every OMH rule that ever escalates it, including ones that do not exist yet. A key carrying the session, the file, or a digest of the plan would instead leave a dead entry in their own command_allowlist per session, or re-prompt somebody who had already answered. One line in their config, visible and removable, covering both file-mutating tools with one answer.

      Two honest limits. The stamp is written by the run, at declaration time, so a planning run that never stamps is not gated — the enforcement binds once the record exists, and what makes that worth building is that stamping happens at the compliant start of the run rather than at the tempted middle of it. And a run can record accepted without being told to; that is a deliberate, recorded declaration in a record a reviewer can read, which is what the prose rule never produced.

  15. 2.0.3

    Expandir

    GitHub Release

    oh-my-hermes 2.0.3 · stable

    Everything merged since 2.0.2 (2026-09-07), 75 pull requests. The grouped summary is in CHANGELOG.md; the full list is below.

    Why this release

    A dispatched unit could stop working and report nothing. A provider's session limit classified as crash; a batch whose every unit failed exited 0, so a caller reading the status was told the work succeeded; a worker alive for 36 minutes retrying one git workaround read as progress because its process existed; and a plan with open items sat unchanged while every turn ended on a status report. Each of those was a separate defect and together they made a stalled run indistinguishable from a working one. This release closes that chain, and adds the refusal that would have prevented the 36 minutes: a unit whose workspace cannot take the work never spawns.

    Highlights

    • A failed dispatch stops reporting success. A provider's session limit classifies as limit-shaped, not a crash, and any unit carrying a failure_kind makes the command exit non-zero. A derived gate re-discovers every *_exit_code mapper from source and fails one that maps failed work to 0. (#1486, #1492)
    • Process alive and work progressing are different states, and never render the same. running requires new evidence; a live process without it is progress_stalled, beside awaiting_input, permission_blocked, account_limit, data_missing, failed, and verified — the only success state, which requires a validated result record, so exit 0 with no record is failed. Progress is read from the unit's own output, including the same failure line repeating while bytes still grow. The status board, plugin reader, TUI widget and omh coding fanout status --json all carry the state, its reason, and how long output has been still; the roster answers the poll loop's own question with all_units_terminal and stuck_units. (#1488, #1491)
    • The workspace is probed before the spawn, not after the failure. File write, git index write, the objects the work needs actually present (including a partial clone whose promisor fetch is forbidden), and case-only filename collisions on a case-insensitive filesystem — all in the unit's own isolation, with no network call. A failing check ends the unit as workspace_blocked before any CLI starts. Executor readiness stops claiming more than it saw: a passing --version is recorded as binary_version_only. (#1489)
    • Recovery matches the cause. A permission denial re-probes the workspace; a limit checks the account and its reset before any retry; missing objects are supplemented and verified before one unit resumes; a stall is re-run only with changed conditions or from a checkpoint. A live in-flight marker on a scope outranks every cause — no second worker is added to work already running. The account a CLI is configured with is recorded as a redacted one-way tag, so a limit can be told apart from the account that hit it. (#1497)
    • The plan drives the work. Open items mean the plan is not finished: the per-turn reminder says to advance the next item rather than end on a status report, and it stops when every item is done or one is recorded blocked with its reason. A plan that stops moving is a finding even while tool calls keep running. A finished dispatch arrives as an event with the verb it owes, and a continuation announced with nothing started is reported back. (#1487, #1490, #1498)
    • Fixed: a grown --help page silently disabled the structured-session lane. The help probe was capped at 16 KiB; claude --help reached 21,401 bytes with --verbose past the cut, so the negotiation refused a protocol the CLI does support, the spawn omitted --output-format stream-json, and every Claude unit ran with no token counts and no session id to steer with — visible only as an empty column. The help probe now gets a document-sized budget, a truncated read still answers the question when every flag was found in what was read, and a refusal states its reason. (#1499)
    • Setup, routing and models. Provider entitlements are asked as one plain ticked list with detected providers pre-ticked; the native-lane chain interview is offered during setup; capable, simple-work and deep-work join the model categories. DeepSeek V4.1 Flash onboarded with superseded generations retired from the shipped chains, and a vendor's dated snapshot id is recognized as the base alias it is a snapshot of. (#1465, #1468, #1469, #1479, #1493)
    • Windows fixes found through "flaky" tests. A sharing-denial retry for the store write that silently dropped a record, and store containment decided lexically instead of by two resolves. Both Windows shards run to completion, and a process-global interrupt flag is restored after every test that sets it — it had been showing up as a CI-only failure in whichever test ran next. (#1448, #1467, #1481, #1483)

    Updating from 2.0.2

    omh update, then restart any running Hermes session — the plugin and TUI widgets are loaded at session start, so a running session keeps the old ones. omh doctor confirms the active generation.

    Install

    curl -fsSL https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.sh | sh      # macOS / Linux
    irm https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.ps1 | iex           # Windows
    omh setup
    

    Hermes skill tap: hermes skills tap add rlaope/oh-my-hermes then hermes skills install rlaope/oh-my-hermes/skills/omh-routing --yes.

    Release evidence

    Check Status
    Full unit suite + compileall on the bumped tree passed inside the Cut Release run (34666287443)
    CI on the tagged tree's parent (cac23667): sharded suite, generated-docs byte checks, dry-run install and stable-channel smokes, capability probe, release smoke plan, readiness checklist, installed-command smoke green
    Wheel built, verified immutable, attached oh_my_hermes-2.0.3-py3-none-any.whl + omh-release-evidence-2.0.3.json (34667207561)
    npm publish step passed, provenance signed; registry latest observed at 2.0.3
    Homebrew tap rlaope/tap/omh formula at 2.0.3, wheel sha256 pinned
    Workflow docs, roles, capability families, ULW inventory and site regions byte-exact, checked in CI
    Workflow-learning review queue not applicable, no learning contract changed
    Plugin bundle changed: omh setup gained the provider-entitlement list and the chain interview (#1468, #1479); the bundle also gained dispatch_outcomes and the continuation rules
    GitHub Pages site copy changed only by the version badge
    Machine sync after the cut the owner machine ran omh update on the merged tree before the cut and reports the plugin rules and dispatch_outcomes installed; no machine has yet run omh update against the 2.0.3 tag

    Not observed in this release run (manual checks remain)

    • No live fanout unit was dispatched under the repaired session protocol, so the token counts and session ids it should now produce are not observed on a real run. The repair itself was measured against the installed CLI: claude 2.1.268 negotiated no protocol before the change and claude_stream_json after.
    • No live Hermes turn consumed the continuation rule or the unacknowledged-outcome lines; both are proven by hook payloads and synthetic fanout_dispatch_summary/v1 fixtures.
    • omh release hermes-smoke --live against a real Hermes profile.
    • Every runtime claim above the plugin install is prompt-level or wrapper-observed evidence, per docs/RELEASE.md; a routed model is not proof that a provider serves it.

    What's Changed

    New Contributors

    Full Changelog: https://github.com/rlaope/oh-my-hermes/compare/v2.0.2...v2.0.3

  16. 2.0.2

    Expandir

    GitHub Release

    oh-my-hermes 2.0.2 · stable

    Everything merged since 2.0.1 (2026-09-05), 26 pull requests. The grouped summary is in CHANGELOG.md; the full list is below.

    Why this release

    omh update on an installer-managed 2.0.1 machine could stop with OMH update stopped during post_activation; the known-good generation remains active. and nothing else to act on. The staged self-update rendered the candidate workflow pack into its own generation, then re-entered the update against the real home, whose manifest still recorded the active generation's pack; every skill the catalog had changed since the last install was judged a "local modification", the re-entry refused without --force, and the whole update rolled back. Any machine whose recorded pack predated the candidate render hit it, which would have been every machine on the next version bump.

    Highlights

    • The installer applies its local-modification and orphan-prune guards only to the directory the home manifest actually records, so a candidate generation is judged on its own content. A stopped update now prints the failing phase's reason, the generation it rolled back to, and the rerun; the re-entered update's stderr is captured for that reason and replayed. (#1377)
    • A transaction-level regression test drives the real CLI through the pack smoke and the post-activation re-entry against an older-manifest home, and a companion test proves the guard is what prevents the reported refusal. (#1378)
    • Release and update: a cut ends with the machine synced and a bare omh announces its Auto Update; omh --version and omh doctor expose a verifiable build identity; the cut stages the site badge it rewrites.
    • Coding and fanout: dispatch runs under an OS write fence proven per run; workspace targets by path shape; decode-to-fixpoint before a containment slot is charged; waits bind to the host's completion signal; cancelled is a first-class terminal state; an explicit user merge directive survives the delegation boundary.
    • Routing and models: Claude Mythos 5.1 leaves the shipped recommendation chains; the Codex Maestro defaults name the GPT generation the CLI serves.
    • Install and memory: the PowerShell installer reads the redirect Location header through a capability probe, proven on both PowerShell versions; memory credential boundaries hardened.

    Updating from 2.0.1

    omh update. The re-entry runs the 2.0.2 installer from the new generation, so a machine that hit the 2.0.1 rollback needs no --force and no manual repair. omh doctor confirms the active generation afterwards.

    Install

    curl -fsSL https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.sh | sh      # macOS / Linux
    irm https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.ps1 | iex           # Windows
    omh setup
    

    Hermes skill tap: hermes skills tap add rlaope/oh-my-hermes then hermes skills install rlaope/oh-my-hermes/skills/omh-routing --yes.

    Release evidence

    Check Status
    Full unit suite + compileall on the bumped tree passed inside the Cut Release run (34097955074)
    CI on the tagged tree's parent (5183a015): sharded suite, generated-docs byte checks, dry-run install and stable-channel smokes, capability probe, release smoke plan, readiness checklist, installed-command smoke green
    Wheel built, verified immutable, attached oh_my_hermes-2.0.2-py3-none-any.whl + omh-release-evidence-2.0.2.json (34099335751)
    npm publish step passed; registry latest observed at 2.0.2
    Homebrew tap rlaope/tap/omh formula at 2.0.2, wheel sha256 pinned
    Workflow docs, roles, capability families, ULW inventory and site regions byte-exact, checked in CI
    Workflow-learning review queue not applicable, no learning contract changed
    Plugin bundle omh setup unchanged in this release beyond the version surface
    GitHub Pages site copy changed only by the version badge

    Not observed in this release run (manual checks remain)

    • omh update from 2.0.1 to 2.0.2 on a live installer-managed machine, the exact path this release fixes; the fix is proven by the in-process transaction test, not by a live venv crossing the bump.
    • omh release hermes-smoke --live against a real Hermes profile.
    • Every runtime claim above the plugin install is prompt-level or wrapper-observed evidence, per docs/RELEASE.md; a routed model is not proof that a provider serves it.

    What's Changed

    Full Changelog: https://github.com/rlaope/oh-my-hermes/compare/v2.0.1...v2.0.2

  17. 2.0.1

    Expandir

    GitHub Release

    oh-my-hermes 2.0.1 · stable

    Everything merged since 2.0.0 (2026-08-29), 118 pull requests. The grouped summary is in CHANGELOG.md; the full list is below.

    Highlights

    • GPT-6 Astra onboarded with an exact-model contract, both-lane routing, and a calibration block that was measured against the block it replaced and revised on the number; the six Astra ids inherit the contract and omh coding model-contract-audit reconciles a host catalog against OMH coverage.
    • Request scoring gains an exhaustive-search signal; Claude Fable 5.1 / Mythos 5.1, GLM 5.3, and MiniMax recognized; token prices are a cited config document.
    • Revision-bound verification plans, resumable fan-out, provider-pressure admission, and a completion-integrity gate on the coding side; native cost provenance and session-scoped todos in the HUD.
    • Twenty-plus new omh-* workflows (codebase-uml, refactor-plan, inference-serving, tech-debt-audit, decision records, omh-docs, GitHub issue intake, and more).
    • The unit-test suite is sharded with a serial quarantine and exact-once aggregation; the site version badge is parity-gated and moves with the other version surfaces.
    • Fixes: plugins.enabled items level with their key are read and written correctly; two CI flakes retired.

    Install

    curl -fsSL https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.sh | sh      # macOS / Linux
    irm https://raw.githubusercontent.com/rlaope/oh-my-hermes/main/install.ps1 | iex           # Windows
    omh setup
    

    Hermes skill tap: hermes skills tap add rlaope/oh-my-hermes then hermes skills install rlaope/oh-my-hermes/skills/omh-routing --yes. Existing installs: omh update.

    Release evidence

    Check Status
    Full unit suite + compileall on the bumped tree passed inside the Cut Release run (33945214557)
    CI on the tagged tree's parent (94928e89): sharded suite, generated-docs byte checks, dry-run install and stable-channel smokes, capability probe, release smoke plan, readiness checklist, installed-command smoke green
    Wheel built, verified immutable, attached oh_my_hermes-2.0.1-py3-none-any.whl + omh-release-evidence-2.0.1.json (33945782041)
    npm publish step passed; registry propagation may lag a few minutes
    Homebrew tap rlaope/tap/omh formula at 2.0.1, wheel sha256 pinned
    Workflow docs, roles, capability families, ULW inventory and site regions byte-exact, checked in CI
    Workflow-learning review queue not applicable, no learning contract changed
    Plugin bundle omh setup unchanged in this release beyond the version surface
    GitHub Pages site copy changed only by the version badge

    Not observed in this release run (manual checks remain)

    • omh update on a live machine from 2.0.0, and omh release hermes-smoke --live against a real Hermes profile.
    • Every runtime claim above the plugin install is prompt-level or wrapper-observed evidence, per docs/RELEASE.md; a routed model is not proof that a provider serves it.

    What's Changed

    Full Changelog: https://github.com/rlaope/oh-my-hermes/compare/v2.0.0...v2.0.1

  18. 1.1.1

    Expandir

    Editado por OpenNavo

    omher 1.1.1 da a quien usa bots de Hermes teclas, divisiones y copiar y pegar que funcionan en Claude Code.

    Pegar
    ⌘V y ⌘C vuelven a pegar y copiar en Claude Code o Codex (#3031).Una captura pegada llega a Claude Code como [Image #1] (#3031).
    Teclas
    ⌘= / ⌘- / ⌘0 cambian el tamaño de letra del terminal (#3031).⌘← / ⌘→ y ⌥← / ⌥→ editan la línea; ⌘K la limpia (#3031).Fuera de macOS, Ctrl+Shift+C/V copian y pegan (#3031).
    Paneles
    ⌘D y ⌘⇧D dividen al momento con un Shell nuevo (#3031).Arrastra una fila de TERMINALS a un terminal para colocarla (#3031).⌘W cierra un terminal sin dividir, y se puede deshacer (#3031).
    Menú
    Clic derecho: copiar, pegar, seleccionar, buscar, limpiar, reset (#3031).El mismo menú divide el panel a la derecha o abajo (#3031).
    Fallos
    El coreano ya no parpadea ni oculta la sílaba que se escribe (#3031).Desaparece una pausa de 50–100 ms cada 30 segundos (#3031).
  19. 1.1.0

    Expandir

    Editado por OpenNavo

    omher 1.1.0 deja a quien usa bots Hermes dividir terminales, restaurarlos y seguirlos por enlace.

    Dividir
    ⌘D y ⌘⇧D dividen una sesión en hasta cuatro paneles (#2993).Los paneles sin foco se atenúan y vuelven tras reiniciar (#2993).En Windows y Linux las teclas de panel usan Ctrl (#2993).
    Volver
    Los terminales de Claude Code y Codex retoman su charla (#2993).La salida vieja se ve atenuada sobre la sesión restaurada (#2993).
    Enlaces
    Cada sesión tiene un enlace omher://terminal/<id>, con ⌘⇧C (#3000).Un bot lee título, estado y última línea de ese enlace (#3000).No se escribe nada hasta pulsar Permitir. Off en Windows (#3000).
    Motor
    xterm 6.1 manda Shift+Enter como tecla propia a Claude Code (#2983).El arte en bloques se ve sólido, sin barra oscura abajo (#3000).
    Chat
    Cada mensaje muestra Leído cuando el bot lo tiene (#2992).Nuevo hilo vuelve a funcionar en la pasarela actual de Hermes (#2987).Desaparece la línea vacía + New thread bajo cada bot (#2990).
  20. 1.0.10

    Expandir

    GitHub Release

    What's Changed

    New Contributors

    Full Changelog: https://github.com/rlaope/oh-my-hermes/compare/v1.0.9...v1.0.10

  21. 1.0.7

    Expandir

    GitHub Release

    v1.0.7-beta

    What's Changed

    Full Changelog: https://github.com/rlaope/oh-my-hermes/compare/v1.0.6...v1.0.7

  22. 1.0.6

    Expandir

    GitHub Release

    What's Changed

    New Contributors

    Full Changelog: https://github.com/rlaope/oh-my-hermes/compare/v1.0.5...v1.0.6