# oh-my-hermes

给 Hermes Agent 用的操作层，把同事放在并行工作线上，并带桌面应用 omher。

> 这是 OpenNavo 上 oh-my-hermes 页面的 Markdown 版本，内容与网页相同： <https://opennavo.com/zh/apps/oh-my-hermes>

- **最新版本**: `1.1.1`
- **开发者**: hope (rlaope)
- **主页**: <https://omher.app/>
- **分类**: [AI 工具](https://opennavo.com/zh/categories/ai)
- **下载**: <https://github.com/rlaope/oh-my-hermes/releases/download/omher-v0.2.3/omher-0.2.3-macos-arm64.dmg>
- **下载地址**: <https://github.com/rlaope/oh-my-hermes/releases/download/omher-v0.2.3/omher-0.2.3-macos-arm64.dmg>

## 介绍

oh-my-hermes 是 Hermes Agent 的一层操作界面。你仍然和 Hermes 对话。它给每条工作线配一个有自己技能和模型的同事，并且并行运行。它不修改、不替换 Hermes，也不会把还在计划中的工作说成已经完成。

桌面应用 omher 把 Hermes 机器人和编程代理放在同一个窗口里。如果还没安装 Hermes Agent，omher 会在第一次启动时提议安装它，并带上 oh-my-hermes。

每条工作线带一个工作类别。类别决定这条线用哪一组模型和多大的推理力度，这些组合都可以改。某个模型被服务商拒绝时，会换到下一组，而不是悄悄降级。

输入 /dag-ulw 后，一个目标会变成一组小节点。准备好的节点并行运行，各自用自己类别的模型和力度。同时运行的节点不共用写入范围。另有一个节点检查目标是否达成，失败的节点会在报告前修复。

路由器会把专项技能作为工具调用放进这次运行。定义、引用和重命名可以从语言服务器取得，而不是靠文本搜索。这条路径默认关闭。

记忆不会悄悄写入。每条候选记忆都会留下保留、拒绝或推迟的书面理由。下一次会话会拿到按任务排序、并按字数上限裁过的回忆包。Hermes 自己的记忆不会被读取或修改。

直接运行 omh 会在 Hermes 的网关上打开 OMH 自己的终端界面。每条状态都会写明所处阶段。只有测试、复查或持续集成检查真正通过，这次运行才算已验证。

命令行版本是 omh 4.0.7。桌面应用是 omher 0.2.3，提供 Apple 芯片和 Intel 的磁盘映像。这些桌面构建没有公证。在 Mac 上第一次打开时，在“应用程序”里右键点 omher，选择“打开”。

## oh-my-hermes 的安全事实

这里只列可以核验的事实，并注明来源和日期，不是安全认证。要检查你下载的文件，可以用 OpenNavo 在本机扫描。

- <https://github.com/rlaope/oh-my-hermes> · 是否开源: 是 — <https://github.com/rlaope/oh-my-hermes> · 核查于 2026-11-11

## 最新版本说明

### 4.0.9 (2026-11-20)

- **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.

## 资料与来源

- <https://opennavo.com/zh/apps/oh-my-hermes>
- <https://opennavo.com/zh/apps/oh-my-hermes/versions>
- <https://github.com/rlaope/oh-my-hermes> · 核查于 2026-11-11
