Skip to content

Agent! Release Notes

AgentiLoop ยท A macOS app that uses AI providers to carry out coding tasks and automate desktop actions.

CaskAI Toolsย 657
Install command
brew install --cask agentiloop-agent
30-day installs
32
#1457 in App
1-year installs
73
Last 365 days
Latest version
1.1.99.299
Download size
19MB
dmg
Release cadence
Weekly
7 releases in 30 days
  1. 1.1.100.300-prerelease

    LatestPre-release

    GitHub Release

    Agent! 1.1.100.300

    Agent! 1.1.100 (build 300) โ€” Pre-release

    This is a pre-release from main, tagged v1.1.100.300. It builds on v1.1.99.299 with a new Web Search tool, a new app icon and README updates.

    ๐Ÿ” Web Search

    • Web Search toolbar button: a new toolbar icon opens a Web Search popover with the SpecStream, Exa and Tavily keys (54e9b391)
    • SpecStream search provider: uses the speed=fast tier and parses the answer object from the search API (54e9b391, 92902349)
    • Drag-and-drop Search Tool Order: SpecStream, Exa, Tavily, Ollama Web Search, Z.AI Web Search and DuckDuckGo. The top tool is preferred and the rest are tried in order if it fails. Tools without a key are skipped, and Ollama and Z.AI no longer jump ahead just because they're the current LLM (9bbae4ef)
    • Claude Web Search: shown as a fixed row when the current tab's LLM is Claude (75797e75, d506170c)
    • Stats and history: the popover shows which search tool was last used, per-tool ok/error counts with the last error, and Recent Searches (17cdbaf9, 78aa9bdc)
    • Recent Searches rows: show the top result as a title, a clickable link and a short description (1f2241ee)
    • Activity log: names the search tool that ran, on every tab, e.g. ๐ŸŒ Web search backend: SpecStream (17cdbaf9, c6773f65)

    ๐Ÿ› Fixed

    • Claude uses only its own built-in web search on the Anthropic API. Before, it was also sent the app's search tool and sometimes used that or Safari instead (f87534f2)
    • Claude's own web searches are now recorded in the Web Search stats, history and activity log (a82c53ae)
    • The search tool no longer tells the model it's "via Tavily", and Safari is only used for searches when you ask for a browser (f87534f2)
    • ClaudeService: invalid or overlong tool_use names in history are cleaned up to avoid a 400 error (cb1d24c8)

    ๐ŸŽจ UI

    • New app icon: four light circles inside a dark blue circle, with larger dots and a brighter blue (a99ad7eb, 6d126876)
    • Intro screen: shows the app icon, and the "Agent!" title is thinner (fa4facc9, 01f8722b)
    • The Sponsors menu is hidden (63acc86d)

    ๐Ÿ“– Docs

    • README: 27 LLM providers (adds Fluxion AI, Muse Code and TheRouter) (d0949c88)
    • README: a "What Agent! can really do" section and a games table, in all 8 languages (01b8290f, 3bb9a0a0, 0da03e28)
    • README: CLI v0.0.10, the /auto progress log and --worktree, 35 accessibility actions and about 70 AgentScripts (b7a515d9, 35af870d)
    • README: sponsors removed in all 8 languages (d794e85f)

    ๐Ÿ”ข Version

    ๐Ÿ“ฆ Install

    • Download the signed, notarized .dmg or .zip attached to this release. The Release workflow builds and staples them after publishing.

    Full changelog: https://github.com/AgentiLoop/Agent/compare/v1.1.99.299...v1.1.100.300

  2. 1.1.99.299

    Patch

    OpenNavo editorial Added to Homebrew

    CompatibilitymacOS 14 or later

    Agent! 1.1.99

    Agent! 1.1.99 adds in-app updates for Mac users and keeps the previous chat when the next task starts.

    Added
    Download, verify, swap, and relaunch an update in the app.
    Fixed
    The next task keeps the previous conversation, including the final done turn.

    Agent! 1.1.99 (build 299), released on 2026-11-16, adds in-app updates for people using the Mac agent and keeps the previous conversation when the next task starts.

    It builds on 1.1.98.298.

    • Auto-update downloads the build, checks the Developer ID signature, swaps the bundle, and relaunches.
    • The next task keeps the previous conversation, including the final done turn.
    • TypeSafeKit is now a remote package.

    Source: https://github.com/AgentiLoop/Agent/releases/tag/v1.1.99.299

  3. 1.1.98.298

    Patch

    GitHub Release Added to Homebrew

    CompatibilitymacOS 14 or later

    Agent! 1.1.98 (build 298)

    Built from main, tagged v1.1.98.298. It builds on v1.1.97.297.

    ๐Ÿ› Fixed

    • A refused turn no longer breaks the tab. When the provider refused a turn (stop reason refusal), the refused turn stayed in the carried-over transcript, was re-sent with every later task and was refused again, so nothing worked until the conversation was cleared by hand. The task now ends at once with "The model refused this request. The conversation was reset โ€” rephrase or start a new task." and the transcript is dropped, in both the main task loop and tab tasks (35bc34bb)

    ๐Ÿ”ง Changed

    • Claude OAuth tokens (sk-ant-oat01-): the client version sent in the User-Agent and billing header is now 2.1.999 (bedf129d)

    ๐Ÿงช Tests

    • Two RouteStopReasonTests cases cover the refusal halt, including after the retry budget is spent (cd9d782b)

    ๐Ÿ“ Docs

    ๐Ÿ”ข Version

    ๐Ÿ“ฆ Install

    • Download the signed, notarized .dmg or .zip attached to this release. The Release workflow builds and staples them after publishing.

    Full changelog: https://github.com/AgentiLoop/Agent/compare/v1.1.97.297...v1.1.98.298

  4. 1.1.97.297

    Expand

    GitHub Release Added to Homebrew

    CompatibilitymacOS 14 or later

    Agent! 1.1.97 (build 297) โ€” Hotfix

    This is a hotfix from main, tagged v1.1.97.297. It builds on v1.1.96.296.

    ๐Ÿ› Fixed

    • Window could not be moved on macOS 27.2. The SwiftUI toolbar's hosting views swallow the titlebar drag events, so the main window was stuck in place. The main window now sets isMovableByWindowBackground, so dragging the window background โ€” including the empty titlebar area โ€” moves the window again (11a2f5ad)
    • Toolbar layout: status items stay on the leading edge, action buttons on the trailing edge, with a native flexible spacer between them (16c31623)

    ๐Ÿ“ฆ Install

    • Download the signed, notarized .dmg or .zip attached to this pre-release. The Release workflow builds and staples them after publishing.

    Full changelog: https://github.com/AgentiLoop/Agent/compare/v1.1.96.296...v1.1.97.297

  5. 1.1.96.296-prerelease

    Expand

    GitHub Release

    Agent! 1.1.96.296

    Agent! 1.1.96 (build 296) โ€” Pre-release

    This is a pre-release from main, tagged v1.1.96.296. It builds on v1.1.95.295 with UI polish and repo cleanup.

    ๐ŸŽจ UI

    • The activity log and LLM output re-render when you switch between light and dark mode. The task banner uses a darker green in light mode (c190d94e)
    • Error History shows the full error message instead of cutting it off at 100 characters (de12e052)

    ๐Ÿ’ผ Sponsors

    • Sidrune AI sponsor placements are commented out (f3f51308)

    ๐Ÿ“– Docs & Repo

    • README and 1.1.95 release notes: added Reverse Engineering with Agent! (CrazyTaxi, MarioKart64JS, F1, GoKart, BvS) and the 1.1.95 screenshot (1e1b9503)
    • READMEs proofread, and all 7 translations synced with the English README (4a3c4859)
    • Tidied the repo root: CLA.md moved to .github/, NOTICE and RELEASE_NOTES_1.1.95.md moved to docs/, and the CLA links updated (ae4019b0, ddbe08a7)

    ๐Ÿ“ฆ Install

    • Download the signed, notarized .dmg or .zip attached to this pre-release. The Release workflow builds and staples them after publishing.

    Full changelog: https://github.com/AgentiLoop/Agent/compare/v1.1.95.295...v1.1.96.296

  6. 1.1.95.295

    Expand

    OpenNavo editorial Added to Homebrew

    CompatibilitymacOS 14 or later

    Agent! 1.1.95.295 rolls up every change since the public 1.1.87.287 release.

    Agent! 1.1.95.295 was released on 2026-11-01 and rolls up every change since the public 1.1.87.287 release.

    • MCP servers can be installed in one step with the mcp tool, and their secrets are stored only in the Keychain.
    • API keys pasted into chat are moved to the Keychain by SecretRedactor instead of reaching the model, and the MCP prompt fills them in automatically.
    • Sentry, Linear and Notion support built-in Sign in with Browser through OAuth 2.1 with PKCE, and tokens refresh automatically.
    • Browser sign-in is detected automatically for any remote server that supports OAuth, without a preset.
    • Servers named like 127.0.0.1 or my_server can be reached again.
    • New MCP presets: Cohesivity, Keenable Search and deed.page.
    • TheRouter.ai is available as a new OpenAI-compatible LLM provider.
    • When an Ollama model is missing, Agent! lists the installed models and suggests ollama pull instead of showing the raw error.
    • The new Avatar tab shows a talking face, speaks replies sentence by sentence, and adds a voice picker, mute and dictation.
    • Auto-Pilot supports several tabs per project, each with its own goals and plans, and can isolate a tab in a git worktree.
    • Accessibility automation gains the read_text, select_option, select_row and press_key actions, and click_menu_item can reach menu-bar extras such as Wi-Fi and Control Center.
    • VoiceOver labels, Increase Contrast, Reduce Motion, Reduce Transparency and Bigger Text now cover the whole app.
    • Every shell command has a hard timeout, 10 minutes by default, configurable under Options > Shell timeout.
    • A tool marked as chronically failing is cleared after an app update, so fixed tools are no longer avoided.
  7. 1.1.94.294-prerelease

    Expand

    GitHub Release

    Agent! 1.1.94.294

    Agent! 1.1.94 (build 294) โ€” Pre-release

    Image: Agent! 1.1.94

    This is a pre-release from main, tagged v1.1.94.294. It builds on v1.1.93.293 with MCP sign-in, Keychain-only secrets, and a new LLM provider.

    ๐Ÿ”Œ MCP

    • One-step install through the mcp tool (e.g. "add sentry mcp server"). Secrets go only to the Keychain (146c9a3a)
    • Built-in Sign in with Browser (OAuth 2.1 + PKCE, discovery, dynamic client registration, 127.0.0.1 loopback) for Sentry, Linear and Notion. It doesn't need npx/mcp-remote, and tokens stay in the Keychain with auto-refresh (efc3724c)
    • Browser sign-in is detected automatically for any remote server that returns 401 with OAuth metadata (ac24351b)
    • Fixed: tools are now dispatched by their sanitized wire name, so servers named like 127.0.0.1 or my_server can be reached again (fb110a37)
    • New presets: deed.page (#70), Keenable Search (#69) and Cohesivity (#68). Thanks to the contributors!

    ๐Ÿ” Privacy

    • Pasted API keys never reach the LLM. SecretRedactor moves the tokens into the Keychain, and the MCP prompt fills them in for you (c0217be5)

    ๐Ÿง  LLM Providers

    • New TheRouter.ai provider, an OpenAI-compatible gateway (e3b41695), with 10 Swift Testing cases (56d7e654)

    โœˆ๏ธ Auto-Pilot & Stability

    • Auto-Pilot now carries attached images into every cycle (bbbb76c9)
    • Fixed stalls (beach ball) in script execution and DaemonCore (48c884aa)

    ๐Ÿงน Repo

    • Tidied the repo root: CONTRIBUTING/CODE_OF_CONDUCT moved to .github, and the dev notes moved to docs/DEV_NOTES.md (453c8488). The README gained a star call-to-action.

    ๐Ÿ“ฆ Install

    • Download the signed, notarized .dmg or .zip attached to this pre-release. The Release workflow builds and staples them after publishing.

    Full changelog: https://github.com/AgentiLoop/Agent/compare/v1.1.93.293...v1.1.94.294

  8. 1.1.93.293-prerelease

    Expand

    GitHub Release

    Agent! 1.1.93.293

    Agent! 1.1.93 (build 293) โ€” Pre-release

    This is a pre-release from main, tagged v1.1.93.293. It carries the same feature set as v1.1.90.292 (accessibility tool actions, tool reliability, Auto-Pilot, VoiceOver and keyboard work, new LLM providers and more). The project source is still at 1.1.90 (build 290), so this build's binary reports 1.1.90 (290).

    ๐Ÿค Sponsors & README

    • The LLM provider sponsor is now Sidrune AI (promo code SDRAGENTLOOP), replacing Fluxion AI in the READMEs and the in-app sponsor list (cbaa3e49)
    • New multilingual Sidrune ad artwork, with the promo-code pill refined across several passes (7826ac7c, 41facad0)

    ๐Ÿ“ฆ Install

    • Download the signed, notarized .dmg or .zip attached to this pre-release. The Release workflow builds and staples them after publishing.
  9. 1.1.90.292-prerelease

    Expand

    GitHub Release

    Agent! 1.1.90.292

    Agent! 1.1.90 โ€” Pre-release

    โ™ฟ Accessibility tool (AX automation)

    • New read_text action: reads every visible text of an app's front window in one call (table rows as Row: a | b | c, toggles as Name [on]). Browser pages read like a document, one Text: line per paragraph (226711a6, 76ebcf40)
    • New select_option action: picks popup, menu-button and combo-box options, radio/tab buttons, checkbox on/off and slider values, then checks the result. It also handles web page <select> menus in Safari and Chrome (05471d0f, 242d2dd2)
    • New select_row action: selects or opens table/list/sidebar rows by any cell's text, plus Finder desktop icons and icon-view items (fee21b3a, e8632ec8)
    • press_key is back: keys and shortcuts in any app (cmd+s, down*3 return) (5d09ba12)
    • click_menu_item clicks without opening menus on screen and lists a menu's items. It also reaches menu-bar extras (Wiโ€‘Fi, Focus, Clock, Control Center) and returns a panel's contents as text (ede6bedc, 6e28a923, 4d9d8cdc)
    • click_element uses AXPress first, like AppleScript click, and falls back to a hit-tested mouse click only when needed (ce33824c)
    • type_into_element fills web page forms in Safari and Chrome. New submit:true presses Return after typing (76d8a790, 3ddf8f36)
    • set_properties accepts AppleScript property names and AppleScript-style element references (button 2 of toolbar 1 of window 1). It presses checkboxes/switches when their state differs, picks pop-up menu items and sets sliders (4a3970d1, 159854f9, 163724c0, b834af9a)
    • Element search now ranks matches (focused window > other windows > hidden > menus) (7a4913fc)
    • Removed the blanket Safari veto on accessibility calls from tab tasks (cbe824e8)
    • Updated tool docs (AgentTools 2.53.34) (089bdfae)

    ๐Ÿ› ๏ธ Tool reliability

    • web navigate, web switch_tab and agent_script update work on the first try. Each one used to fail every call (424b34b0, b94ebbc2, 2a619ca7, 07738648)
    • After 2 failures, UI-scripting tools steer the model to accessibility. Tab tasks get the same advisory (3a3f4144, e330ae39)
    • Chronic tool-failure flags expire with the app build, so fixed tools are no longer flagged (577a767e)
    • xcode select_project with a single open project selects it automatically (ad101ec6)
    • Ollama "model not found" lists the installed models and suggests ollama pull, then tries the fallback chain (c24f8a1c)
    • Fixes in project.pbxproj, ErrorHandler and Localizable (6dbe01c5)

    ๐Ÿงญ Auto-Pilot

    ๐Ÿ—ฃ๏ธ App accessibility (VoiceOver, keyboard, display settings)

    Full Changelog: https://github.com/AgentiLoop/Agent/compare/v1.1.89.289...v1.1.90.292

  10. 1.1.89.289-prerelease

    Expand

    GitHub Release

    Agent! 1.1.89.289

    Agent! 1.1.89 โ€” Pre-release

    ๐Ÿš Shell

    • Fixed an in-process shell hang when an orphaned background job kept the output pipe open. Pipes are now drained on a reader thread and the shell waits only 2s for EOF after exit, so the timeout always fires (7f0c1e9b)

    ๐Ÿ”‘ Claude OAuth

    • OAuth (sk-ant-oat01-) requests now send a Claude Code client profile (User-Agent, x-app, claude-code-20250219 beta flag) and a billing-context system block for compatibility testing (a7e0a530, 89e6f2b9, 9f8bb5a3)
    • Opening a new tab clears every provider's Retry-After backoff (a7e0a530)

    ๐Ÿงน Misc

    • Auto-Pilot stop string updated to per-tab wording (d41108a9)
    • README: LaunchNest featured badge (2d923f8b)
    • Version bump to 1.1.89 (build 289) (9373aab3)

    Full Changelog: https://github.com/AgentiLoop/Agent/compare/v1.1.88.288...v1.1.89.289

  11. 1.1.88.288-prerelease

    Expand

    GitHub Release

    Agent! 1.1.88.288

    Agent! 1.1.88 โ€” Pre-release

    ๐Ÿ—ฃ๏ธ Avatar Tabs

    • New talking-face avatar pane with per-sentence expressions; the New Tab sheet gets Avatar + Default to Auto-Pilot toggles (bef43d3b, 5035a31e)
    • Jev picks the facial expression per sentence, with a keyword fallback that matches whole words only (847c028c, 45a0a073)
    • Replies and done summaries are spoken sentence by sentence as they stream, and the HUD text stays in sync word by word (26aeffc0, dbd8f419, a0f2c776)
    • A new avatar tab says a greeting aloud. Pick the voice in the pane (Daniel is the default), or use mute for expressions only (65a9bb67, 04069c45, 5cf1f1a9)
    • Dictation auto-sends after a 2.5s pause; replying or closing the tab stops the avatar talking (030ac20c, 26d8cca0, ba202cf5)
    • Sentences are no longer spoken twice, and file paths and links are read in a natural way (9eb74dae, 00fae20b, e157e931)

    ๐Ÿค– Auto-Pilot

    • Multiple tabs can now work on one project, with per-tab progress, a shared registry, git worktree isolation (automatic or forced) and memory shared from worktrees (566014f6)
    • Goal state and plans are kept per tab, so parallel tabs on one project no longer share them (beb30ab6)

    ๐Ÿš Shell

    • Every shell command now has a hard timeout: 10 min by default, set in Options > Shell timeout, 0 = off (965e68f3)

    ๐Ÿ›  Fixes

    • Red/green stripes on tab-indented edit diffs (AgentColorSyntax 1.2.13) (76b8df7e)
    • The debug build has its own SwiftData store, so release and debug builds stop wiping each other's tabs (1e28b209)
    • Swift packages updated (a835a094)

    Full Changelog: https://github.com/AgentiLoop/Agent/compare/v1.1.87.287...v1.1.88.288

  12. 1.1.87.287

    Expand

    OpenNavo editorial Added to Homebrew

    CompatibilitymacOS 14 or later

    Adds goal-driven unattended /auto sessions, new LLM providers and decision tools, and updates context handling, platform support, accessibility, safety, and licensing.

    Automation
    Adds timed /auto sessions with goal history and restart recovery.Adds Fluxion AI, Muse Code, Requesty, A2Agent, OrcaRouter, and Qwen Code.Adds experimental fm serve and the Jev decision layer.
    Core & UI
    Changes critic enforcement, context compaction, and token reporting.Adds accessibility actions and fixes tab-specific state and output.Supports macOS 14.6+; FoundationModels requires macOS 26.
    Safety
    Blocks recursive deletion of the current project folder.Fixes a ShellSafety and read-only bypass, plus streaming and task issues.Changes the license to PolyForm Noncommercial 1.0.0.

    Agent! 1.1.87.287 adds /auto <goal> for unattended, goal-driven cycles with a time budget, goal history, and session recovery after app restarts. Esc stops only the current cycle; Stop All ends the session. New provider options include Fluxion AI, Muse Code, Requesty, A2Agent, OrcaRouter, and Qwen Code. Experimental fm serve exposes Apple Foundation Models through a local Chat Completions API, while Jev advises the tool loop.

    The release changes critic-review enforcement, context sizing and compaction, overflow and max_tokens detection, and token reporting. It adds accessibility actions and fixes tab-specific output and state, blocks recursive deletion of the current project folder, and fixes a ShellSafety and read-only bypass. Other fixes address streaming, task completion, parallel tool-call IDs, and local-model out-of-memory handling.

    macOS 14.6+ is supported on Apple Silicon and Intel; FoundationModels requires macOS 26 or later. The project license changes to PolyForm Noncommercial 1.0.0. Official release notes

  13. 1.1.33.233

    Expand

    OpenNavo editorial Added to Homebrew

    Rebuilds context compaction, adds read-before-edit validation, strengthens shell/XPC and completion checks, and fixes streaming, tool routing and tab logs.

    Context compaction uses actual provider input-token usage and a threshold based on the context window minus reserved output and a buffer. Provider-generated summaries preserve goals, plans and edited-file context; oversized tool output can be restored with provenance and stale-file warnings. Overflow and max_tokens recovery and chat-history size limits are improved.

    The read-before-edit gate refuses edits to unread files or files changed since reading, then reads the file to enable a subsequent edit. Adds daemon-side shell safety and release-build XPC client trust checks; fixes process-pipe deadlocks, streaming errors and cancellation, rate-limit retries, tool routing and completion gates. Improves large activity-log rendering and persistent per-tab UI state. Upstream reports 273/273 tests passing for this release; this is an upstream result, not a separate local validation.

    https://github.com/AgentiLoop/Agent/releases/tag/v1.1.33.233

  14. 1.0.89.183

    Expand

    GitHub Release

    Agent! 1.0.89 (Build 183)

    Autonomous macOS agent with a privileged Launch Daemon for root-level command execution via XPC.

    What's New

    • Requesty LLM provider โ€” Requesty added as a selectable LLM provider alongside Claude, Ollama, OpenAI-compatible endpoints and Apple Intelligence (Foundation Models). Thanks to @Thibaultjaigu (#25).
    • Tailscale support โ€” App Transport Security settings updated so Agent! can reach LLM endpoints served over a Tailscale tailnet (self-hosted Ollama / OpenAI-compatible servers on your private network).
    • Version bump to 1.0.89 / build 183.

    Install

    1. Download Agent_1.0.89.dmg below.
    2. Open the DMG and drag Agent! into the Applications shortcut.
    3. Launch Agent! and approve the privileged helper daemon once when prompted (System Settings โ†’ General โ†’ Login Items & Extensions).

    Requirements

    • macOS 26.0 or later
    • Apple Silicon or Intel Mac

    Full Changelog: https://github.com/macOS26/Agent/compare/v1.0.88.182...v1.0.89.183

  15. 1.0.88.182

    Expand

    GitHub Release

    ๐Ÿค– Agent! 1.0.88.182 Release Notes

    ๐Ÿ”ง DeepSeek Re-Fix

    • Fixed DeepSeek provider integration issues that regressed in 1.0.87
    • Re-tuned tool consolidation for DeepSeek models to prevent malformed tool calls
    • Improved error handling for DeepSeek API responses
    • Fixed streaming mode compatibility with DeepSeek v3 models

    ๐Ÿ“ฆ Install

    • Download Agent_1.0.88.181.dmg
    • Drag Agent! to /Applications
    • Launch and approve the privileged daemon when prompted
  16. 1.0.80.170

    Expand

    GitHub Release

    ๐ŸŒ New Provider: OpenRouter

    • OpenRouter as a first-class LLM provider โ€” access hundreds of models through a single API key
    • Protocol toggle โ€” switch between OpenAI-compatible and Anthropic-compatible endpoints per model
    • Smart model catalog โ€” fetches and filters OpenRouter's /models to only show tool-capable models with real context windows (strips embedding, image-only, and preview entries)
    • ClaudeService integration โ€” Anthropic-compatible OpenRouter models route through ClaudeService with proper Bearer auth (sk-or- key detection) and tool forwarding
    • Full UI support โ€” settings panel, model picker, fallback chain, and tab configuration all wired up

    ๐Ÿ›ก๏ธ Error Handling & Reliability

    • Stop infinite-loop on misdetected context overflow โ€” narrowed detection to require "exceed/too long/too many" alongside keywords; plain max_tokens in unrelated parameter errors no longer triggers aggressive pruning
    • Short-circuit OpenRouter free-tier 429s โ€” detects shared free-tier rate limits and bails immediately instead of churning through futile retries
    • Prune loop cap โ€” if pruning messages doesn't actually shrink history, stops retrying (max 3 consecutive prune attempts)

    โœ๏ธ Code Editing Guardrails

    • Duplicate edit rejection โ€” code-level detection blocks the exact same (file, old_string) pair on consecutive turns, preventing stuck loops
    • Edit cycle detection โ€” tracks files edited across recent turns and nudges when the same 2โ€“3 files keep alternating without progress
    • Stuck threshold lowered โ€” nudge at 2 failures (was 3), give-up at 4โ€“6 failures (was 6)
    • Edit size safety โ€” refuses edits with old_string over 50 lines; requires minimum 10 chars for replace_all
    • Source verification for diff_apply โ€” validates that the provided source parameter matches actual file content at the specified lines, preventing silent corruption from stale line numbers
    • Truncation warning on write_file โ€” warns when new content is less than 50% of the original file size
    • Post-edit diagnostic gating โ€” swiftc -parse checks only run after 2+ consecutive edit failures, avoiding noisy false positives

    ๐Ÿ“ System Prompt Improvements

    • Commitment Rule โ€” new hard contract: claiming "I found the problem" MUST be followed by an edit tool call, not another read. Repeated certainty claims without shipping edits are flagged as a failure mode
    • Anti-hallucination reinforcement โ€” stronger guidance against fabricating findings from incomplete reads

    ๐Ÿ”ง Other

    • AgentTools bumped to v2.52.0
    • Version bumped to 1.0.80 (build 170)