\\n>

⌛ How to resolve this issue?

\\n> \\n> After the wait time has elapsed, a review can be triggered using the `@coderabbitai review` command as a PR comment. Alternatively, push new commits to this PR.\\n> \\n> We recommend that you space out your commits to avoid hitting the rate limit.\\n> \\n>

\\n> \\n> \\n>

\\n>

🚦 How do rate limits work?

\\n> \\n> CodeRabbit enforces hourly rate limits for each developer per organization.\\n> \\n> Our paid plans have higher rate limits than the trial, open-source and free plans. In all cases, we re-allow further reviews after a brief timeout.\\n> \\n> Please see our [FAQ](https://docs.coderabbit.ai/faq) for further information.\\n> \\n>

\\n> \\n>

\\n>

ℹ️ Review info

\\n> \\n>

\\n>

⚙️ Run configuration

\\n> \\n> **Configuration used**: Path: .coderabbit.yaml\\n> \\n> **Review profile**: CHILL\\n> \\n> **Plan**: Pro\\n> \\n> **Run ID**: `070b8eb6-b03d-4dbd-9f5a-937ea113dfe0`\\n> \\n>

\\n> \\n>

\\n>

📥 Commits

\\n> \\n> Reviewing files that changed from the base of the PR and between d457c0b1e9df3cd76c981ab772314f5c91d911fe and 3887a4b856109cc67f391b0adb68148c8aa18c06.\\n> \\n>

\\n> \\n>

\\n>

⛔ Files ignored due to path filters (2)

\\n> \\n> * `docs/architecture/AI_INTEGRATION.md` is excluded by `!docs/**`\\n> * `docs/features/CHAT.md` is excluded by `!docs/**`\\n> \\n>

\\n> \\n>

\\n>

📒 Files selected for processing (2)

\\n> \\n> * `apps/desktop/src/main/services/chat/agentChatService.test.ts`\\n> * `apps/desktop/src/main/services/chat/agentChatService.ts`\\n> \\n>

\\n> \\n>

\\n\\n\\n\\n\\n\\n\\n

\\n

✨ Finishing Touches

\\n\\n

\\n

🧪 Generate unit tests (beta)

\\n\\n- [ ] Create PR with unit tests\\n- [ ] Commit unit tests in branch `ade/cancelling-chats-and-subagents-37cbbffe`\\n\\n

\\n\\n

\\n\\n\\n\\n\\n\\n---\\n\\nThanks for using [CodeRabbit](https://coderabbit.ai?utm_source=oss&utm_medium=github&utm_campaign=arul28/ADE&utm_content=112)! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.\\n\\n

\\n

❤️ Share

\\n\\n- [X](https://twitter.com/intent/tweet?text=I%20just%20used%20%40coderabbitai%20for%20my%20code%20review%2C%20and%20it%27s%20fantastic%21%20It%27s%20free%20for%20OSS%20and%20offers%20a%20free%20trial%20for%20the%20proprietary%20code.%20Check%20it%20out%3A&url=https%3A//coderabbit.ai)\\n- [Mastodon](https://mastodon.social/share?text=I%20just%20used%20%40coderabbitai%20for%20my%20code%20review%2C%20and%20it%27s%20fantastic%21%20It%27s%20free%20for%20OSS%20and%20offers%20a%20free%20trial%20for%20the%20proprietary%20code.%20Check%20it%20out%3A%20https%3A%2F%2Fcoderabbit.ai)\\n- [Reddit](https://www.reddit.com/submit?title=Great%20tool%20for%20code%20review%20-%20CodeRabbit&text=I%20just%20used%20CodeRabbit%20for%20my%20code%20review%2C%20and%20it%27s%20fantastic%21%20It%27s%20free%20for%20OSS%20and%20offers%20a%20free%20trial%20for%20proprietary%20code.%20Check%20it%20out%3A%20https%3A//coderabbit.ai)\\n- [LinkedIn](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fcoderabbit.ai&mini=true&title=Great%20tool%20for%20code%20review%20-%20CodeRabbit&summary=I%20just%20used%20CodeRabbit%20for%20my%20code%20review%2C%20and%20it%27s%20fantastic%21%20It%27s%20free%20for%20OSS%20and%20offers%20a%20free%20trial%20for%20proprietary%20code)\\n\\n

\\n\\n_{Comment `@coderabbitai help` to get the list of available commands and usage tips.}\\n\\n\",\"path\":null,\"side\":null,\"line\":null}","text_elements",[],{"_449":997,"_436":437,"_451":439},{"_452":998,"_454":-5,"_41":-5,"_455":-5,"_95":309,"_91":457},[999],{"_45":993,"_91":1000},"input_text","turn/started",{"_436":437,"_86":1003},{"_102":-5,"_41":439,"_440":1004,"_442":1005},[],"inProgress",{"_442":1007,"_436":437},{"_1008":1009,"_91":1010},"activeFlags",[],"active",{"_449":1012,"_436":437,"_451":1013},{"_452":1014,"_454":-5,"_41":-5,"_455":-5,"_95":309,"_91":457},"auto-compact-0",[1015,1016],{"_45":1018,"_91":1000},{"_45":1017,"_91":1000},"\n /workspace/ADE\n bash\n 2026-03-31\n Etc/UTC\n","# AGENTS.md instructions for /workspace/ADE\n\n\n# ADE Project Instructions\n\n## About this project\n\n- ADE is a local-first desktop application for orchestrating coding agents, missions, lanes, PR workflows, and proof/artifact capture.\n- The main product lives in `apps/desktop` and is built with Electron, React, and TypeScript.\n- The ADE MCP server lives in `apps/mcp-server` and shares core services with the desktop app.\n- State is primarily stored under `.ade/` inside the active project, with runtime metadata in SQLite and machine-local files under `.ade/secrets`, `.ade/cache`, and `.ade/artifacts`.\n\n## Working norms\n\n- Preserve existing desktop app patterns before introducing new abstractions.\n- Prefer fixing the underlying service or shared type rather than layering renderer-only workarounds on top.\n- Keep IPC contracts, preload types, shared types, and renderer usage in sync whenever an interface changes.\n- For ADE MCP changes, verify both headless MCP mode and the desktop socket-backed MCP path.\n- For computer-use changes, treat policy enforcement and artifact ownership as hard requirements, not prompt guidance.\n\n## Validation\n\n- Desktop checks:\n - `npm --prefix apps/desktop run typecheck`\n - `npm --prefix apps/desktop run test`\n - `npm --prefix apps/desktop run build`\n - `npm --prefix apps/desktop run lint`\n- MCP checks:\n - `npm --prefix apps/mcp-server run typecheck`\n - `npm --prefix apps/mcp-server run test`\n - `npm --prefix apps/mcp-server run build`\n- Run the smallest relevant subset first when iterating, then finish with the broader checks that cover the touched surfaces.\n\n## Terminology\n\n- Use \"lane\" for ADE worktrees/branches.\n- Use \"mission\" for orchestrated multi-step work.\n- Use \"computer use\" for screenshot/video/GUI/browser proof flows.\n## Style preferences\n\n- Prefer direct, operational language over marketing phrasing.\n- Keep user-facing copy concrete and stateful: say what changed, what is blocked, and what the next action is.\n- Use sentence case for headings and labels unless the existing UI pattern is intentionally uppercase.\n\n## Content boundaries\n\n- Do not reframe ADE as a docs site, Mintlify project, or generic template app.\n- Do not store secrets in plaintext project files when an encrypted store already exists.\n- Do not leave policy enforcement in prompts alone when a code path can enforce it directly.\n\n\n## Skills\nA skill is a set of local instructions to follow that is stored in a `SKILL.md` file. Below is the list of skills that can be used. Each entry includes a name, description, and file path so you can open the source for full instructions when using a specific skill.\n### Available skills\n- skill-creator: Guide for creating effective skills. This skill should be used when users want to create a new skill (or update an existing skill) that extends Codex's capabilities with specialized knowledge, workflows, or tool integrations. (file: /opt/codex/skills/.system/skill-creator/SKILL.md)\n- skill-installer: Install Codex skills into $CODEX_HOME/skills from a curated list or a GitHub repo path. Use when a user asks to list installable skills, install a curated skill, or install a skill from another repo (including private repos). (file: /opt/codex/skills/.system/skill-installer/SKILL.md)\n### How to use skills\n- Discovery: The list above is the skills available in this session (name + description + file path). Skill bodies live on disk at the listed paths.\n- Trigger rules: If the user names a skill (with `$SkillName` or plain text) OR the task clearly matches a skill's description shown above, you must use that skill for that turn. Multiple mentions mean use them all. Do not carry skills across turns unless re-mentioned.\n- Missing/blocked: If a named skill isn't in the list or the path can't be read, say so briefly and continue with the best fallback.\n- How to use a skill (progressive disclosure):\n 1) After deciding to use a skill, open its `SKILL.md`. Read only enough to follow the workflow.\n 2) When `SKILL.md` references relative paths (e.g., `scripts/foo.py`), resolve them relative to the skill directory listed above first, and only consider other paths if needed.\n 3) If `SKILL.md` points to extra folders such as `references/`, load only the specific files needed for the request; don't bulk-load everything.\n 4) If `scripts/` exist, prefer running or patching them instead of retyping large code blocks.\n 5) If `assets/` or templates exist, reuse them instead of recreating from scratch.\n- Coordination and sequencing:\n - If multiple skills apply, choose the minimal set that covers the request and state the order you'll use them.\n - Announce which skill(s) you're using and why (one short line). If you skip an obvious skill, say why.\n- Context hygiene:\n - Keep context small: summarize long sections instead of pasting them; only load extra files when needed.\n - Avoid deep reference-chasing: prefer opening only files directly linked from `SKILL.md` unless you're blocked.\n - When variants exist (frameworks, providers, domains), pick only the relevant reference file(s) and note that choice.\n- Safety and fallback: If a skill can't be applied cleanly (missing files, unclear instructions), state the issue, pick the next-best approach, and continue.\n",{"_449":1020,"_436":437,"_451":1013},{"_452":1021,"_454":-5,"_41":-5,"_455":-5,"_95":1022,"_91":457},[1023],"developer",{"_45":1024,"_91":1000},"\nFilesystem sandboxing defines which files can be read or written. `sandbox_mode` is `danger-full-access`: No filesystem sandboxing - all commands are permitted. Network access is enabled.\nApproval policy is currently never. Do not provide the `sandbox_permissions` for any reason, commands will be rejected.\n","thread/started",{"_1027":1028},"thread",{"_1029":-5,"_1030":-5,"_1031":1032,"_1033":1034,"_493":494,"_1035":15,"_1036":-5,"_1037":-5,"_41":437,"_1038":1039,"_167":-5,"_156":1040,"_1041":185,"_499":1042,"_442":1043,"_76":1044,"_1045":1034},"agentNickname","agentRole","cliVersion","0.114.0","createdAt",1774970797,"ephemeral","forkedFromId","gitInfo","modelProvider","openai","/opt/codex/sessions/2026/03/31/rollout-2026-03-31T15-26-37-019d4480-fd4b-7ea1-85bb-ac169fd06713.jsonl","preview","vscode",{"_91":446},[],"updatedAt",[],"messages",[1049],{"_41":1050,"_1051":1052,"_1053":1054,"_1055":1054,"_452":1056,"_442":1057,"_454":-5,"_1058":31,"_1059":1060,"_1061":283,"_1062":-5},"81e90239-c493-44dc-8289-dd89dc5ab669","author",{"_95":309,"_167":-5,"_1059":1071},"create_time",1774970796.589794,"update_time",{"_1067":45,"_1068":1069},"finished_successfully","weight","metadata",{"_1063":1064},"recipient","channel","l1239dk1",{"_1065":1066},"inline_comments",[],"content_type","parts",[1070],"@codex resolve conflicts and commit changes here\ncancelling chats and subagents\n\n\n[vc]: #wjicWokYmhDMgfXJ82PieIpxPjmMzyMwYg2XagiZOyA=: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\nThe latest updates on your projects. Learn more about [Vercel for GitHub](https://vercel.link/github-learn-more).\n\n| Project | Deployment | Actions | Updated (UTC) |\n| :--- | :----- | :------ | :------ |\n| [ade](https://vercel.com/arul28s-projects/ade) | ![Ready](https://vercel.com/static/status/ready.svg) [Ready](https://vercel.com/arul28s-projects/ade/6MXqKJWBxhGdv1ZwVQGAT7CVhg3V) | [Preview](https://ade-git-ade-cancelling-chats-and-subage-fde2fc-arul28s-projects.vercel.app), [Comment](https://vercel.live/open-feedback/ade-git-ade-cancelling-chats-and-subage-fde2fc-arul28s-projects.vercel.app?via=pr-comment-feedback-link) | Mar 30, 2026 10:57pm |\n\n\n\n\n\n> [!WARNING]\n> ## Rate limit exceeded\n> \n> `@arul28` has exceeded the limit for the number of commits that can be reviewed per hour. Please wait **1 minutes and 17 seconds** before requesting another review.\n> \n> Your organization is not enrolled in usage-based pricing. Contact your admin to enable usage-based pricing to continue reviews beyond the rate limit, or try again in **1 minutes and 17 seconds**.\n> \n>

\n>

⌛ How to resolve this issue?

\n> \n> After the wait time has elapsed, a review can be triggered using the `@coderabbitai review` command as a PR comment. Alternatively, push new commits to this PR.\n> \n> We recommend that you space out your commits to avoid hitting the rate limit.\n> \n>

\n> \n> \n>

\n>

🚦 How do rate limits work?

\n> \n> CodeRabbit enforces hourly rate limits for each developer per organization.\n> \n> Our paid plans have higher rate limits than the trial, open-source and free plans. In all cases, we re-allow further reviews after a brief timeout.\n> \n> Please see our [FAQ](https://docs.coderabbit.ai/faq) for further information.\n> \n>

\n> \n>

\n>

ℹ️ Review info

\n> \n>

\n>

⚙️ Run configuration

\n> \n> **Configuration used**: Path: .coderabbit.yaml\n> \n> **Review profile**: CHILL\n> \n> **Plan**: Pro\n> \n> **Run ID**: `070b8eb6-b03d-4dbd-9f5a-937ea113dfe0`\n> \n>

\n> \n>

\n>

📥 Commits

\n> \n> Reviewing files that changed from the base of the PR and between d457c0b1e9df3cd76c981ab772314f5c91d911fe and 3887a4b856109cc67f391b0adb68148c8aa18c06.\n> \n>

\n> \n>

\n>

⛔ Files ignored due to path filters (2)

\n> \n> * `docs/architecture/AI_INTEGRATION.md` is excluded by `!docs/**`\n> * `docs/features/CHAT.md` is excluded by `!docs/**`\n> \n>

\n> \n>

\n>

📒 Files selected for processing (2)

\n> \n> * `apps/desktop/src/main/services/chat/agentChatService.test.ts`\n> * `apps/desktop/src/main/services/chat/agentChatService.ts`\n> \n>

\n> \n>

\n\n\n\n\n\n\n

✨ Finishing Touches

\n\n

🧪 Generate unit tests (beta)

\n\n- [ ] Create PR with unit tests\n- [ ] Commit unit tests in branch `ade/cancelling-chats-and-subagents-37cbbffe`\n\n

\n\n

\n\n\n\n\n\n---\n\nThanks for using [CodeRabbit](https://coderabbit.ai?utm_source=oss&utm_medium=github&utm_campaign=arul28/ADE&utm_content=112)! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.\n\n

❤️ Share

\n\n- [X](https://twitter.com/intent/tweet?text=I%20just%20used%20%40coderabbitai%20for%20my%20code%20review%2C%20and%20it%27s%20fantastic%21%20It%27s%20free%20for%20OSS%20and%20offers%20a%20free%20trial%20for%20the%20proprietary%20code.%20Check%20it%20out%3A&url=https%3A//coderabbit.ai)\n- [Mastodon](https://mastodon.social/share?text=I%20just%20used%20%40coderabbitai%20for%20my%20code%20review%2C%20and%20it%27s%20fantastic%21%20It%27s%20free%20for%20OSS%20and%20offers%20a%20free%20trial%20for%20the%20proprietary%20code.%20Check%20it%20out%3A%20https%3A%2F%2Fcoderabbit.ai)\n- [Reddit](https://www.reddit.com/submit?title=Great%20tool%20for%20code%20review%20-%20CodeRabbit&text=I%20just%20used%20CodeRabbit%20for%20my%20code%20review%2C%20and%20it%27s%20fantastic%21%20It%27s%20free%20for%20OSS%20and%20offers%20a%20free%20trial%20for%20proprietary%20code.%20Check%20it%20out%3A%20https%3A//coderabbit.ai)\n- [LinkedIn](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fcoderabbit.ai&mini=true&title=Great%20tool%20for%20code%20review%20-%20CodeRabbit&summary=I%20just%20used%20CodeRabbit%20for%20my%20code%20review%2C%20and%20it%27s%20fantastic%21%20It%27s%20free%20for%20OSS%20and%20offers%20a%20free%20trial%20for%20proprietary%20code)\n\n

\n\n_{Comment `@coderabbitai help` to get the list of available commands and usage tips.}\n\n",{},"created",1774970897.9550889,"Completing the task","event_type","info",{"_91":457,"_95":92,"_452":2540},{"_91":1105,"_1106":1107},{"_91":1080,"_1081":1082,"_1083":1084,"_1085":1086,"_1087":-5},"pr","pr_title","Resolve merge conflicts; add chat cancel/edit, stash clear, tests and UI tweaks","pr_message","### Motivation\n\n- Bring the feature branch up to date by merging the base `d457c0b1e9df3c...` and removing conflict markers in the AI integration doc while preserving the intended steer/cancellation semantics. \n- Add explicit steer-management APIs so the renderer and remote controllers can cancel or edit pending steer instructions (`cancelSteer` / `editSteer`).\n- Provide a safe, auditable `git stash clear` operation and corresponding UI affordances for stash management.\n- Harden GitHub/PR error handling and add unit tests covering newly touched code paths.\n\n### Description\n\n- Resolved merge conflict in `docs/architecture/AI_INTEGRATION.md` and preserved both `cancelSteer`/`editSteer` rows alongside the updated `interrupt` semantics. \n- Agent chat service changes: added `cancelSteer` and `editSteer` implementations to `apps/desktop/src/main/services/chat/agentChatService.ts`, exported them from the service API, and updated returned API surface.\n- IPC/preload/shared types updates: added new types `AgentChatCancelSteerArgs` and `AgentChatEditSteerArgs` in `shared/types/chat.ts`, registered new IPC channels `agentChatCancelSteer` / `agentChatEditSteer` in `shared/ipc.ts` and `registerIpc.ts`, and exposed `cancelSteer`/`editSteer` on the preload API (`preload.ts` / `global.d.ts`).\n- Git/stash feature: implemented `stashClear` in `apps/desktop/src/main/services/git/gitOperationsService.ts`, added `gitStashClear` IPC handler (`registerIpc.ts`) and preload/renderer wiring, and updated the data model/docs to include `git_stash_clear` and `git_stash_clear` usage notes.\n- UI and UX adjustments: updated stash UI in `LaneGitActionsPane.tsx` (Save/Restore/Copy/Delete/Clear flows and confirm numeric clear), added pull behind badge and subtle styling/animation changes in `TabNav.tsx`, `TopBar.tsx`, and CSS (`index.css`) for the pull flash; adjusted terminal/attention visuals and helpers (`terminalAttention.ts`, `chatStatusVisuals.tsx`) to use distinct non-spinning indicators for waiting/attention states.\n- GitHub/PR and PR service improvements: enhanced `githubService.apiRequest` to include `errors` array details in error messages and wrapped PR creation errors in `prs/prService.ts` with contextual head/base branch info.\n- Tests and new files: added and updated unit tests under `apps/desktop/src/main/services/*` and `apps/desktop/src/renderer/*` including `gitOperationsService.test.ts`, `githubService.test.ts`, and `prs/prService.test.ts`.\n- Merge actions: completed the merge and committed the conflict resolution as a follow-up change and created a follow-up PR to make the branch mergeable.\n\n### Testing\n\n- No automated test suite was executed as part of this follow-up rollout; new and updated unit tests were added at `apps/desktop/src/main/services/git/gitOperationsService.test.ts`, `apps/desktop/src/main/services/github/githubService.test.ts`, and `apps/desktop/src/main/services/prs/prService.test.ts` but were not run in this session.\n- Validation performed: merged base via `git merge --no-commit --no-ff`, confirmed conflict markers removed with a repository search, and committed the resolved state successfully (`git commit`).","output_diff",{"_91":1085,"_1088":494,"_127":128,"_1089":-5,"_308":1090,"_1091":1092,"_1093":1094,"_1095":1096,"_1097":1098,"_1099":1100,"_1101":-5},"pre_apply_patch","repo_id","preapply_patch_base_commit_sha","diff --git a/apps/desktop/src/main/services/chat/agentChatService.ts b/apps/desktop/src/main/services/chat/agentChatService.ts\nindex 943725e8d78cf54fd30304c0ee7f430a329ad49a..788c6a92a578c02cf2808d3914fdcb4da1e0541b 100644\n--- a/apps/desktop/src/main/services/chat/agentChatService.ts\n+++ b/apps/desktop/src/main/services/chat/agentChatService.ts\n@@ -27,57 +27,59 @@ import { buildClaudeV2Message, inferAttachmentMediaType } from \"./buildClaudeV2M\n import {\n appendBufferedAssistantText,\n canAppendBufferedAssistantText,\n shouldFlushBufferedAssistantTextForEvent,\n type BufferedAssistantText,\n } from \"./chatTextBatching\";\n import type { Logger } from \"../logging/logger\";\n import type { createLaneService } from \"../lanes/laneService\";\n import { resolveLaneLaunchContext, type LaneLaunchContext } from \"../lanes/laneLaunchContext\";\n import type { createSessionService } from \"../sessions/sessionService\";\n import type { createProjectConfigService } from \"../config/projectConfigService\";\n import type { createFileService } from \"../files/fileService\";\n import type { createProcessService } from \"../processes/processService\";\n import { runGit } from \"../git/git\";\n import { CLAUDE_RUNTIME_AUTH_ERROR, isClaudeRuntimeAuthError } from \"../ai/claudeRuntimeProbe\";\n import { resolveClaudeCodeExecutable } from \"../ai/claudeCodeExecutable\";\n import { resolveCodexExecutable } from \"../ai/codexExecutable\";\n import { fileSizeOrZero, isEnoentError, nowIso, readFileWithinRootSecure, resolvePathWithinRoot } from \"../shared/utils\";\n import type { EpisodicSummaryService } from \"../memory/episodicSummaryService\";\n import {\n createDefaultComputerUsePolicy,\n normalizeComputerUsePolicy,\n } from \"../../../shared/types\";\n import type {\n AgentChatApprovalDecision,\n+ AgentChatCancelSteerArgs,\n AgentChatClaudePermissionMode,\n AgentChatCompletionReport,\n AgentChatCodexApprovalPolicy,\n AgentChatCodexConfigSource,\n AgentChatCodexSandbox,\n AgentChatCreateArgs,\n AgentChatDisposeArgs,\n+ AgentChatEditSteerArgs,\n AgentChatExecutionMode,\n AgentChatEvent,\n AgentChatEventEnvelope,\n AgentChatFileRef,\n AgentChatHandoffArgs,\n AgentChatHandoffResult,\n AgentChatIdentityKey,\n AgentChatInteractionMode,\n AgentChatInterruptArgs,\n AgentChatModelInfo,\n AgentChatProvider,\n AgentChatRespondToInputArgs,\n AgentChatSession,\n AgentChatSessionCapabilities,\n AgentChatSessionCapabilitiesArgs,\n AgentChatSessionSummary,\n AgentChatSlashCommand,\n AgentChatSlashCommandsArgs,\n AgentChatSubagentListArgs,\n AgentChatSubagentSnapshot,\n AgentChatSurface,\n AgentChatSteerArgs,\n AgentChatSendArgs,\n AgentChatUnifiedPermissionMode,\n PendingInputQuestion,\n@@ -8334,50 +8336,58 @@ export function createAgentChatService(args: {\n emitChatEvent(managed, {\n type: \"user_message\",\n text: trimmed,\n turnId: runtime.activeTurnId ?? undefined,\n });\n emitChatEvent(managed, {\n type: \"system_notice\",\n noticeKind: \"info\",\n message: \"Message queued — will be sent when the current turn completes.\",\n turnId: runtime.activeTurnId ?? undefined,\n });\n persistChatState(managed);\n return;\n }\n \n const preparedSteer = prepareSendMessage({\n sessionId,\n text: trimmed,\n displayText: trimmed,\n attachments: [],\n });\n if (!preparedSteer) return;\n await executePreparedSendMessage(preparedSteer);\n };\n \n+ const cancelSteer = async ({ sessionId }: AgentChatCancelSteerArgs): Promise => {\n+ await interrupt({ sessionId });\n+ };\n+\n+ const editSteer = async ({ sessionId, text }: AgentChatEditSteerArgs): Promise => {\n+ await steer({ sessionId, text });\n+ };\n+\n const interrupt = async ({ sessionId }: AgentChatInterruptArgs): Promise => {\n const managed = ensureManagedSession(sessionId);\n \n // Unified runtime interrupt — auto-decline pending approvals to prevent orphans\n if (managed.runtime?.kind === \"unified\") {\n if (managed.runtime.interrupted) return;\n managed.runtime.interrupted = true;\n managed.runtime.abortController?.abort();\n for (const [itemId, approval] of managed.runtime.pendingApprovals) {\n approval.resolve({ decision: \"decline\" });\n managed.runtime.pendingApprovals.delete(itemId);\n }\n return;\n }\n \n if (managed.session.provider === \"codex\") {\n const runtime = await ensureCodexSessionRuntime(managed);\n if (!managed.session.threadId || !runtime.activeTurnId) return;\n await runtime.request(\"turn/interrupt\", {\n threadId: managed.session.threadId,\n turnId: runtime.activeTurnId\n });\n return;\n }\n \n@@ -8951,51 +8961,51 @@ export function createAgentChatService(args: {\n codexApprovalPolicy,\n codexSandbox,\n codexConfigSource,\n unifiedPermissionMode,\n permissionMode,\n computerUse,\n }: AgentChatUpdateSessionArgs): Promise => {\n const managed = ensureManagedSession(sessionId);\n const chatConfig = resolveChatConfig();\n const isIdentitySession = Boolean(managed.session.identityKey);\n const hasConversation = managed.recentConversationEntries.length > 0 || readTranscriptConversationEntries(managed).length > 0;\n let resetRuntimeForComputerUse = false;\n \n if (modelId !== undefined) {\n const nextModelId = String(modelId ?? \"\").trim();\n if (!nextModelId.length) {\n throw new Error(\"A modelId is required when updating a chat session model.\");\n }\n \n const descriptor = getModelById(nextModelId) ?? resolveModelAlias(nextModelId);\n if (!descriptor) {\n throw new Error(`Unknown model '${nextModelId}'.`);\n }\n \n const nextProvider: AgentChatProvider = resolveProviderGroupForModel(descriptor);\n- const nextModel = descriptor.isCliWrapped ? descriptor.shortId : descriptor.id;\n+ const nextModel = descriptor.isCliWrapped ? descriptor.sdkModelId : descriptor.id;\n const previousModelId = managed.session.modelId\n ?? resolveModelIdFromStoredValue(managed.session.model, managed.session.provider)\n ?? managed.session.model;\n const previousProvider = managed.session.provider;\n const modelSwitchPolicy = managed.session.identityKey === \"cto\"\n ? \"any-after-launch\"\n : \"same-family-after-launch\";\n \n if (!canSwitchChatSessionModel({\n currentModelId: previousModelId,\n nextModelId: descriptor.id,\n hasConversation,\n policy: modelSwitchPolicy,\n })) {\n throw new Error(\"This chat can only switch within the same model family after the conversation has started.\");\n }\n \n const modelChanged =\n previousProvider !== nextProvider\n || managed.session.modelId !== descriptor.id\n || managed.session.model !== nextModel;\n \n if (managed.runtime && modelChanged) {\n teardownRuntime(managed);\n refreshReconstructionContext(managed, { includeConversationTail: true });\n@@ -9346,50 +9356,52 @@ export function createAgentChatService(args: {\n reject(new Error(`Timed out waiting for session '${sessionId}' to finish the current turn.`));\n }, safeTimeoutMs);\n \n sessionTurnCollectors.set(sessionId, {\n resolve,\n reject,\n outputText: \"\",\n lastError: null,\n timeout,\n });\n \n void executePreparedSendMessage(prepared).catch((error) => {\n clearTimeout(timeout);\n sessionTurnCollectors.delete(sessionId);\n reject(error instanceof Error ? error : new Error(String(error)));\n });\n });\n };\n \n return {\n createSession,\n handoffSession,\n sendMessage,\n runSessionTurn,\n steer,\n+ cancelSteer,\n+ editSteer,\n interrupt,\n resumeSession,\n listSessions,\n getSessionSummary,\n getChatTranscript,\n ensureIdentitySession,\n approveToolUse,\n respondToInput,\n getAvailableModels,\n getSlashCommands,\n codexFuzzyFileSearch,\n dispose,\n disposeAll,\n updateSession,\n warmupModel,\n listSubagents,\n getSessionCapabilities,\n /** Clean up temp attachment files older than 7 days. Call on app startup. */\n cleanupStaleAttachments() {\n try {\n const projectRoot = args.projectRoot;\n if (!projectRoot) return;\n const attachDir = path.join(projectRoot, \".ade\", \"attachments\");\n if (!fs.existsSync(attachDir)) return;\n const cutoff = Date.now() - 7 * 24 * 60 * 60 * 1000; // 7 days\ndiff --git a/apps/desktop/src/main/services/git/gitOperationsService.test.ts b/apps/desktop/src/main/services/git/gitOperationsService.test.ts\nindex 4db363615aece6990070bc5727ede5ee24e0ccca..df2090841a8a4eae3b726f6b3760fbd71d936e21 100644\n--- a/apps/desktop/src/main/services/git/gitOperationsService.test.ts\n+++ b/apps/desktop/src/main/services/git/gitOperationsService.test.ts\n@@ -1,41 +1,108 @@\n import { beforeEach, describe, expect, it, vi } from \"vitest\";\n \n const mockGit = vi.hoisted(() => ({\n runGit: vi.fn(),\n runGitOrThrow: vi.fn(),\n getHeadSha: vi.fn(),\n }));\n \n vi.mock(\"./git\", () => ({\n runGit: (...args: unknown[]) => mockGit.runGit(...args),\n runGitOrThrow: (...args: unknown[]) => mockGit.runGitOrThrow(...args),\n getHeadSha: (...args: unknown[]) => mockGit.getHeadSha(...args),\n }));\n \n import { createGitOperationsService } from \"./gitOperationsService\";\n \n+describe(\"gitOperationsService.stashClear\", () => {\n+ beforeEach(() => {\n+ vi.clearAllMocks();\n+ });\n+\n+ it(\"calls git stash clear with the lane worktree path and returns the action result\", async () => {\n+ mockGit.getHeadSha.mockResolvedValue(\"abc123\");\n+ mockGit.runGitOrThrow.mockResolvedValue(undefined);\n+\n+ const mockStart = vi.fn().mockReturnValue({ operationId: \"op-1\" });\n+ const mockFinish = vi.fn();\n+\n+ const service = createGitOperationsService({\n+ laneService: {\n+ getLaneBaseAndBranch: vi.fn().mockReturnValue({\n+ baseRef: \"main\",\n+ branchRef: \"feature/stash-test\",\n+ worktreePath: \"/tmp/ade-lane\",\n+ laneType: \"worktree\",\n+ }),\n+ } as any,\n+ operationService: {\n+ start: mockStart,\n+ finish: mockFinish,\n+ } as any,\n+ projectConfigService: {\n+ get: () => ({ effective: { ai: {} } }),\n+ } as any,\n+ aiIntegrationService: {\n+ getFeatureFlag: () => false,\n+ getStatus: vi.fn(async () => ({ availableModelIds: [] })),\n+ generateCommitMessage: vi.fn(),\n+ } as any,\n+ logger: {\n+ info: vi.fn(),\n+ warn: vi.fn(),\n+ error: vi.fn(),\n+ debug: vi.fn(),\n+ } as any,\n+ });\n+\n+ const result = await service.stashClear({ laneId: \"lane-1\" });\n+\n+ expect(mockGit.runGitOrThrow).toHaveBeenCalledWith(\n+ [\"stash\", \"clear\"],\n+ { cwd: \"/tmp/ade-lane\", timeoutMs: 15_000 },\n+ );\n+ expect(result).toEqual({\n+ operationId: \"op-1\",\n+ preHeadSha: \"abc123\",\n+ postHeadSha: \"abc123\",\n+ });\n+ expect(mockStart).toHaveBeenCalledWith(\n+ expect.objectContaining({\n+ laneId: \"lane-1\",\n+ kind: \"git_stash_clear\",\n+ }),\n+ );\n+ expect(mockFinish).toHaveBeenCalledWith(\n+ expect.objectContaining({\n+ operationId: \"op-1\",\n+ status: \"succeeded\",\n+ }),\n+ );\n+ });\n+});\n+\n describe(\"gitOperationsService.generateCommitMessage\", () => {\n beforeEach(() => {\n vi.clearAllMocks();\n });\n \n it(\"uses the configured model and sends a lightweight changed-files prompt\", async () => {\n let capturedPrompt = \"\";\n let capturedModel = \"\";\n \n mockGit.runGit.mockImplementation(async (args: string[]) => {\n if (args[0] === \"diff\") {\n return {\n exitCode: 0,\n stdout: \"M\\tapps/desktop/src/main/foo.ts\\nA\\tapps/desktop/src/main/bar.ts\\n\",\n stderr: \"\",\n };\n }\n if (args[0] === \"show\") {\n return {\n exitCode: 0,\n stdout: \"M\\tapps/desktop/src/main/previous.ts\\n\",\n stderr: \"\",\n };\n }\n return {\ndiff --git a/apps/desktop/src/main/services/git/gitOperationsService.ts b/apps/desktop/src/main/services/git/gitOperationsService.ts\nindex 0b03679c38ea03a21499b74eef348070da25ff36..dcc7e82b2a3fee7f0c85da6163db00802d43dd40 100644\n--- a/apps/desktop/src/main/services/git/gitOperationsService.ts\n+++ b/apps/desktop/src/main/services/git/gitOperationsService.ts\n@@ -748,50 +748,63 @@ export function createGitOperationsService({\n kind: \"git_stash_pop\",\n reason: \"stash_pop\",\n metadata: { stashRef },\n fn: async (lane) => {\n await runGitOrThrow([\"stash\", \"pop\", stashRef], { cwd: lane.worktreePath, timeoutMs: 30_000 });\n }\n });\n return action;\n },\n \n async stashDrop(args: GitStashRefArgs): Promise {\n const stashRef = args.stashRef.trim();\n if (!stashRef.length) throw new Error(\"stashRef is required\");\n const { action } = await runLaneOperation({\n laneId: args.laneId,\n kind: \"git_stash_drop\",\n reason: \"stash_drop\",\n metadata: { stashRef },\n fn: async (lane) => {\n await runGitOrThrow([\"stash\", \"drop\", stashRef], { cwd: lane.worktreePath, timeoutMs: 30_000 });\n }\n });\n return action;\n },\n \n+ async stashClear(args: { laneId: string }): Promise {\n+ const { action } = await runLaneOperation({\n+ laneId: args.laneId,\n+ kind: \"git_stash_clear\",\n+ reason: \"stash_clear\",\n+ metadata: {},\n+ fn: async (lane) => {\n+ await runGitOrThrow([\"stash\", \"clear\"], { cwd: lane.worktreePath, timeoutMs: 15_000 });\n+ }\n+ });\n+ return action;\n+ },\n+\n async fetch(args: { laneId: string }): Promise {\n const { action } = await runLaneOperation({\n laneId: args.laneId,\n kind: \"git_fetch\",\n reason: \"fetch\",\n fn: async (lane) => {\n await runGitOrThrow([\"fetch\", \"--prune\"], { cwd: lane.worktreePath, timeoutMs: 60_000 });\n }\n });\n return action;\n },\n \n async sync(args: GitSyncArgs): Promise {\n const mode = args.mode ?? \"merge\";\n const { action } = await runLaneOperation({\n laneId: args.laneId,\n kind: mode === \"rebase\" ? \"git_sync_rebase\" : \"git_sync_merge\",\n reason: mode === \"rebase\" ? \"sync_rebase\" : \"sync_merge\",\n metadata: { mode, baseRefOverride: args.baseRef ?? null },\n fn: async (lane) => {\n const targetBase = args.baseRef?.trim() || lane.baseRef;\n await maybeFetchAndSync(lane, mode, targetBase);\n }\n });\n return action;\ndiff --git a/apps/desktop/src/main/services/github/githubService.test.ts b/apps/desktop/src/main/services/github/githubService.test.ts\nnew file mode 100644\nindex 0000000000000000000000000000000000000000..1ba476477ed8cebc356137ad993744655f9b3ced\n--- /dev/null\n+++ b/apps/desktop/src/main/services/github/githubService.test.ts\n@@ -0,0 +1,209 @@\n+import { beforeEach, describe, expect, it, vi } from \"vitest\";\n+\n+// ---------------------------------------------------------------------------\n+// vi.hoisted mock state\n+// ---------------------------------------------------------------------------\n+const mockFetch = vi.hoisted(() => vi.fn());\n+\n+// ---------------------------------------------------------------------------\n+// vi.mock — external dependencies\n+// ---------------------------------------------------------------------------\n+\n+vi.mock(\"electron\", () => ({\n+ safeStorage: {\n+ isEncryptionAvailable: () => true,\n+ decryptString: () => JSON.stringify({ token: \"ghp_mock\" }),\n+ encryptString: (s: string) => Buffer.from(s),\n+ },\n+}));\n+\n+vi.mock(\"node:fs\", async (importOriginal) => {\n+ const actual = await importOriginal();\n+ return {\n+ ...actual,\n+ default: {\n+ ...actual,\n+ existsSync: vi.fn(() => false),\n+ readFileSync: vi.fn(() => Buffer.from(\"encrypted\")),\n+ mkdirSync: vi.fn(),\n+ writeFileSync: vi.fn(),\n+ chmodSync: vi.fn(),\n+ unlinkSync: vi.fn(),\n+ },\n+ existsSync: vi.fn(() => false),\n+ readFileSync: vi.fn(() => Buffer.from(\"encrypted\")),\n+ mkdirSync: vi.fn(),\n+ writeFileSync: vi.fn(),\n+ chmodSync: vi.fn(),\n+ unlinkSync: vi.fn(),\n+ };\n+});\n+\n+vi.mock(\"../git/git\", () => ({\n+ runGit: vi.fn(),\n+}));\n+\n+// Replace global fetch\n+vi.stubGlobal(\"fetch\", mockFetch);\n+\n+import { createGithubService } from \"./githubService\";\n+\n+// ---------------------------------------------------------------------------\n+// Helpers\n+// ---------------------------------------------------------------------------\n+\n+function makeLogger() {\n+ return {\n+ info: vi.fn(),\n+ warn: vi.fn(),\n+ error: vi.fn(),\n+ debug: vi.fn(),\n+ } as any;\n+}\n+\n+function makeService() {\n+ return createGithubService({\n+ logger: makeLogger(),\n+ projectRoot: \"/tmp/test-project\",\n+ appDataDir: \"/tmp/test-appdata\",\n+ });\n+}\n+\n+function jsonResponse(\n+ status: number,\n+ body: unknown,\n+ extraHeaders?: Record,\n+) {\n+ const headers = new Headers({ \"content-type\": \"application/json\", ...extraHeaders });\n+ return {\n+ ok: status >= 200 && status < 300,\n+ status,\n+ headers,\n+ text: async () => JSON.stringify(body),\n+ json: async () => body,\n+ } as unknown as Response;\n+}\n+\n+// ---------------------------------------------------------------------------\n+// Tests\n+// ---------------------------------------------------------------------------\n+\n+describe(\"githubService.apiRequest\", () => {\n+ beforeEach(() => {\n+ vi.clearAllMocks();\n+ });\n+\n+ it(\"returns data and response on success (HTTP 200)\", async () => {\n+ const payload = { id: 1, name: \"test-repo\" };\n+ mockFetch.mockResolvedValueOnce(jsonResponse(200, payload));\n+\n+ const service = makeService();\n+ const result = await service.apiRequest({\n+ method: \"GET\",\n+ path: \"/repos/owner/repo\",\n+ token: \"ghp_test123\",\n+ });\n+\n+ expect(result.data).toEqual(payload);\n+ expect(result.response).toBeDefined();\n+ expect(result.response!.status).toBe(200);\n+ });\n+\n+ it(\"throws with message from response when errors array is absent\", async () => {\n+ mockFetch.mockResolvedValueOnce(jsonResponse(404, { message: \"Not Found\" }));\n+\n+ const service = makeService();\n+ await expect(\n+ service.apiRequest({ method: \"GET\", path: \"/repos/owner/nope\", token: \"ghp_test123\" }),\n+ ).rejects.toThrow(\"Not Found\");\n+ });\n+\n+ it(\"appends single error detail from errors array\", async () => {\n+ mockFetch.mockResolvedValueOnce(\n+ jsonResponse(422, {\n+ message: \"Validation Failed\",\n+ errors: [{ message: \"A pull request already exists\" }],\n+ }),\n+ );\n+\n+ const service = makeService();\n+ await expect(\n+ service.apiRequest({ method: \"POST\", path: \"/repos/o/r/pulls\", token: \"ghp_test123\" }),\n+ ).rejects.toThrow(\"Validation Failed: A pull request already exists\");\n+ });\n+\n+ it(\"joins multiple error details with semicolons\", async () => {\n+ mockFetch.mockResolvedValueOnce(\n+ jsonResponse(422, {\n+ message: \"Validation Failed\",\n+ errors: [{ message: \"err1\" }, { message: \"err2\" }],\n+ }),\n+ );\n+\n+ const service = makeService();\n+ await expect(\n+ service.apiRequest({ method: \"POST\", path: \"/repos/o/r/pulls\", token: \"ghp_test123\" }),\n+ ).rejects.toThrow(\"Validation Failed: err1; err2\");\n+ });\n+\n+ it(\"includes rate limit info and rateLimitResetAtMs when rate-limited\", async () => {\n+ const resetTimestamp = Math.floor(Date.now() / 1000) + 3600;\n+ mockFetch.mockResolvedValueOnce(\n+ jsonResponse(\n+ 403,\n+ {\n+ message: \"API rate limit exceeded\",\n+ errors: [{ message: \"some detail\" }],\n+ },\n+ {\n+ \"x-ratelimit-remaining\": \"0\",\n+ \"x-ratelimit-reset\": String(resetTimestamp),\n+ },\n+ ),\n+ );\n+\n+ const service = makeService();\n+ let thrownError: any;\n+ try {\n+ await service.apiRequest({ method: \"GET\", path: \"/repos/o/r\", token: \"ghp_test123\" });\n+ } catch (err) {\n+ thrownError = err;\n+ }\n+\n+ expect(thrownError).toBeInstanceOf(Error);\n+ expect(thrownError.message).toContain(\"API rate limit exceeded\");\n+ expect(thrownError.message).toContain(\"some detail\");\n+ expect(thrownError.message).toContain(\"rate limit exceeded; resets at\");\n+ expect(thrownError.rateLimitResetAtMs).toBe(resetTimestamp * 1000);\n+ });\n+\n+ it(\"falls back to generic HTTP message when response body has no message field\", async () => {\n+ mockFetch.mockResolvedValueOnce(jsonResponse(500, { unexpected: true }));\n+\n+ const service = makeService();\n+ await expect(\n+ service.apiRequest({ method: \"GET\", path: \"/test\", token: \"ghp_test123\" }),\n+ ).rejects.toThrow(\"GitHub API request failed (HTTP 500)\");\n+ });\n+\n+ it(\"ignores errors array entries without a string message\", async () => {\n+ mockFetch.mockResolvedValueOnce(\n+ jsonResponse(422, {\n+ message: \"Validation Failed\",\n+ errors: [{ code: \"custom\" }, { message: \"real error\" }, { message: 42 }],\n+ }),\n+ );\n+\n+ const service = makeService();\n+ await expect(\n+ service.apiRequest({ method: \"POST\", path: \"/repos/o/r/pulls\", token: \"ghp_test123\" }),\n+ ).rejects.toThrow(\"Validation Failed: real error\");\n+ });\n+\n+ it(\"throws when no token is provided and none is stored\", async () => {\n+ const service = makeService();\n+ await expect(\n+ service.apiRequest({ method: \"GET\", path: \"/test\" }),\n+ ).rejects.toThrow(\"GitHub token missing\");\n+ });\n+});\ndiff --git a/apps/desktop/src/main/services/github/githubService.ts b/apps/desktop/src/main/services/github/githubService.ts\nindex 191b86bb45eec7d500ff2e049244ed08ed8a2ab8..23bd8d2bafe3f9cb118db3673f8aa0b7a7e3abe9 100644\n--- a/apps/desktop/src/main/services/github/githubService.ts\n+++ b/apps/desktop/src/main/services/github/githubService.ts\n@@ -233,64 +233,72 @@ export function createGithubService({\n }\n \n const response = await fetch(url.toString(), {\n method: args.method,\n headers,\n body: args.body != null ? JSON.stringify(args.body) : undefined\n });\n \n // 304 Not Modified — return cached data (free, no rate limit cost)\n if (response.status === 304) {\n const cached = etagCache.get(urlKey);\n if (cached) {\n return { data: cached.data as T, response };\n }\n }\n \n const text = await response.text();\n let data: unknown = text;\n try {\n data = text.trim().length ? JSON.parse(text) : {};\n } catch {\n // keep text\n }\n \n if (!response.ok) {\n- const message =\n- (data && typeof data === \"object\" && !Array.isArray(data) ? asString((data as any).message) : \"\") ||\n- `GitHub API request failed (HTTP ${response.status})`;\n+ const body = data && typeof data === \"object\" && !Array.isArray(data) ? (data as Record) : null;\n+ const message = (body ? asString(body.message) : \"\") || `GitHub API request failed (HTTP ${response.status})`;\n+ let detail = \"\";\n+ if (body && Array.isArray(body.errors)) {\n+ const errorMessages = (body.errors as any[])\n+ .map((e) => (typeof e === \"object\" && e && typeof e.message === \"string\" ? e.message : null))\n+ .filter(Boolean);\n+ if (errorMessages.length > 0) {\n+ detail = \": \" + errorMessages.join(\"; \");\n+ }\n+ }\n const rateRemaining = response.headers.get(\"x-ratelimit-remaining\");\n const rateReset = response.headers.get(\"x-ratelimit-reset\");\n if (rateRemaining === \"0\" && rateReset) {\n const resetAtMs = Number(rateReset) * 1000;\n const err = new Error(\n- `${message} (rate limit exceeded; resets at ${new Date(resetAtMs).toLocaleString()})`\n+ `${message}${detail} (rate limit exceeded; resets at ${new Date(resetAtMs).toLocaleString()})`\n );\n (err as any).rateLimitResetAtMs = resetAtMs;\n throw err;\n }\n- throw new Error(message);\n+ throw new Error(message + detail);\n }\n \n // Cache ETag for future conditional requests\n if (args.method === \"GET\") {\n const etag = response.headers.get(\"etag\");\n if (etag) {\n // Evict oldest entries if cache is full\n if (etagCache.size >= ETAG_CACHE_MAX_SIZE) {\n const firstKey = etagCache.keys().next().value;\n if (firstKey) etagCache.delete(firstKey);\n }\n etagCache.set(urlKey, { etag, data });\n }\n }\n \n return { data: data as T, response };\n };\n \n let cachedStatus: GitHubStatus | null = null;\n let cachedAt = 0;\n \n const getStatus = async (): Promise => {\n const token = readStoredToken();\n const repo = await detectRepo().catch(() => null);\n if (!token) {\ndiff --git a/apps/desktop/src/main/services/ipc/registerIpc.ts b/apps/desktop/src/main/services/ipc/registerIpc.ts\nindex 301cf14d125644da8db0af1b7757d567b521abe3..1aace2a6f7886ffcef9be0b6a7d9fc2c842b6c1c 100644\n--- a/apps/desktop/src/main/services/ipc/registerIpc.ts\n+++ b/apps/desktop/src/main/services/ipc/registerIpc.ts\n@@ -155,50 +155,52 @@ import type {\n GetTestLogTailArgs,\n ExportHistoryArgs,\n ExportHistoryResult,\n AgentChatApproveArgs,\n AgentChatClaudePermissionMode,\n AgentChatCreateArgs,\n AgentChatDisposeArgs,\n AgentChatGetSummaryArgs,\n AgentChatHandoffArgs,\n AgentChatHandoffResult,\n AgentChatInterruptArgs,\n AgentChatListArgs,\n AgentChatModelInfo,\n AgentChatModelsArgs,\n AgentChatPermissionMode,\n AgentChatRespondToInputArgs,\n AgentChatResumeArgs,\n AgentChatSendArgs,\n AgentChatSession,\n AgentChatSessionSummary,\n AgentChatSubagentSnapshot,\n AgentChatSubagentListArgs,\n AgentChatSessionCapabilities,\n AgentChatSessionCapabilitiesArgs,\n AgentChatSteerArgs,\n+ AgentChatCancelSteerArgs,\n+ AgentChatEditSteerArgs,\n AgentChatUnifiedPermissionMode,\n AgentChatUpdateSessionArgs,\n AgentChatSlashCommand,\n AgentChatSlashCommandsArgs,\n AgentChatFileSearchArgs,\n AgentChatFileSearchResult,\n AgentTool,\n KeybindingOverride,\n KeybindingsSnapshot,\n ImportBranchLaneArgs,\n OnboardingDetectionResult,\n OnboardingExistingLaneCandidate,\n OnboardingStatus,\n LaneSummary,\n ListOperationsArgs,\n ListOverlapsArgs,\n ListLanesArgs,\n ListMissionsArgs,\n ListSessionsArgs,\n ListTestRunsArgs,\n MergeSimulationArgs,\n MergeSimulationResult,\n OperationRecord,\n ContextGenerateDocsArgs,\n ContextGenerateDocsResult,\n@@ -3514,50 +3516,73 @@ export function registerIpc({\n throw new Error(\"OAuth state encode originalState must be a string\");\n }\n return { laneId: record.laneId, originalState: record.originalState };\n };\n \n const parseDecodeOAuthStateArgs = (\n value: unknown,\n ): DecodeOAuthStateArgs => {\n const record = requireRecord(value, \"OAuth state decode request\");\n if (typeof record.encodedState !== \"string\" || !record.encodedState) {\n throw new Error(\"OAuth state decode encodedState must be a non-empty string\");\n }\n return { encodedState: record.encodedState };\n };\n \n const parseDiagnosticsLaneIdArgs = (\n value: unknown,\n ): { laneId: string } => {\n const record = requireRecord(value, \"Runtime diagnostics request\");\n if (typeof record.laneId !== \"string\" || !record.laneId.trim()) {\n throw new Error(\"Runtime diagnostics laneId must be a non-empty string\");\n }\n return { laneId: record.laneId };\n };\n \n+ const parseAgentChatCancelSteerArgs = (\n+ value: unknown,\n+ ): AgentChatCancelSteerArgs => {\n+ const record = requireRecord(value, \"Agent chat cancel steer request\");\n+ if (typeof record.sessionId !== \"string\" || !record.sessionId.trim()) {\n+ throw new Error(\"Agent chat cancel steer sessionId must be a non-empty string\");\n+ }\n+ return { sessionId: record.sessionId };\n+ };\n+\n+ const parseAgentChatEditSteerArgs = (\n+ value: unknown,\n+ ): AgentChatEditSteerArgs => {\n+ const record = requireRecord(value, \"Agent chat edit steer request\");\n+ if (typeof record.sessionId !== \"string\" || !record.sessionId.trim()) {\n+ throw new Error(\"Agent chat edit steer sessionId must be a non-empty string\");\n+ }\n+ if (typeof record.text !== \"string\") {\n+ throw new Error(\"Agent chat edit steer text must be a string\");\n+ }\n+ return { sessionId: record.sessionId, text: record.text };\n+ };\n+\n ipcMain.handle(IPC.lanesOAuthGetStatus, async () => {\n const ctx = getCtx();\n return ctx.oauthRedirectService?.getStatus() ?? {\n enabled: false,\n routingMode: \"state-parameter\" as const,\n activeSessions: [],\n callbackPaths: [],\n };\n });\n \n ipcMain.handle(IPC.lanesOAuthUpdateConfig, async (_event, args: unknown) => {\n const ctx = getCtx();\n if (!ctx.oauthRedirectService) throw new Error(\"OAuth redirect service not available\");\n ctx.oauthRedirectService.updateConfig(parseOAuthUpdateConfigArgs(args));\n });\n \n ipcMain.handle(IPC.lanesOAuthGenerateRedirectUris, async (_event, args: unknown) => {\n const ctx = getCtx();\n if (!ctx.oauthRedirectService) return [];\n const request = parseGenerateRedirectUrisArgs(args);\n return ctx.oauthRedirectService.generateRedirectUris(request.provider);\n });\n \n ipcMain.handle(IPC.lanesOAuthEncodeState, async (_event, args: unknown) => {\n const ctx = getCtx();\n@@ -3716,50 +3741,60 @@ export function registerIpc({\n ipcMain.handle(IPC.agentChatGetSummary, async (_event, arg: AgentChatGetSummaryArgs): Promise => {\n const ctx = getCtx();\n return await ctx.agentChatService.getSessionSummary(arg?.sessionId ?? \"\");\n });\n \n ipcMain.handle(IPC.agentChatCreate, async (_event, arg: AgentChatCreateArgs): Promise => {\n const ctx = getCtx();\n return await ctx.agentChatService.createSession(arg);\n });\n \n ipcMain.handle(IPC.agentChatHandoff, async (_event, arg: AgentChatHandoffArgs): Promise => {\n const ctx = getCtx();\n return await ctx.agentChatService.handoffSession(arg);\n });\n \n ipcMain.handle(IPC.agentChatSend, async (_event, arg: AgentChatSendArgs): Promise => {\n const ctx = getCtx();\n await ctx.agentChatService.sendMessage(arg);\n });\n \n ipcMain.handle(IPC.agentChatSteer, async (_event, arg: AgentChatSteerArgs): Promise => {\n const ctx = getCtx();\n await ctx.agentChatService.steer(arg);\n });\n \n+ ipcMain.handle(IPC.agentChatCancelSteer, async (_event, arg: unknown): Promise => {\n+ const ctx = getCtx();\n+ await ctx.agentChatService.cancelSteer(parseAgentChatCancelSteerArgs(arg));\n+ });\n+\n+ ipcMain.handle(IPC.agentChatEditSteer, async (_event, arg: unknown): Promise => {\n+ const ctx = getCtx();\n+ await ctx.agentChatService.editSteer(parseAgentChatEditSteerArgs(arg));\n+ });\n+\n ipcMain.handle(IPC.agentChatInterrupt, async (_event, arg: AgentChatInterruptArgs): Promise => {\n const ctx = getCtx();\n await ctx.agentChatService.interrupt(arg);\n });\n \n ipcMain.handle(IPC.agentChatResume, async (_event, arg: AgentChatResumeArgs): Promise => {\n const ctx = getCtx();\n return await ctx.agentChatService.resumeSession(arg);\n });\n \n ipcMain.handle(IPC.agentChatApprove, async (_event, arg: AgentChatApproveArgs): Promise => {\n const ctx = getCtx();\n await ctx.agentChatService.approveToolUse(arg);\n });\n \n ipcMain.handle(IPC.agentChatRespondToInput, async (_event, arg: AgentChatRespondToInputArgs): Promise => {\n const ctx = getCtx();\n await ctx.agentChatService.respondToInput(arg);\n });\n \n ipcMain.handle(IPC.agentChatModels, async (_event, arg: AgentChatModelsArgs): Promise => {\n const ctx = getCtx();\n return await ctx.agentChatService.getAvailableModels(arg);\n });\n \n@@ -4074,50 +4109,55 @@ export function registerIpc({\n ipcMain.handle(IPC.gitStashPush, async (_event, arg: GitStashPushArgs): Promise => {\n const ctx = getCtx();\n return ctx.gitService.stashPush(arg);\n });\n \n ipcMain.handle(IPC.gitStashList, async (_event, arg: { laneId: string }): Promise => {\n const ctx = getCtx();\n return ctx.gitService.listStashes(arg);\n });\n \n ipcMain.handle(IPC.gitStashApply, async (_event, arg: GitStashRefArgs): Promise => {\n const ctx = getCtx();\n return ctx.gitService.stashApply(arg);\n });\n \n ipcMain.handle(IPC.gitStashPop, async (_event, arg: GitStashRefArgs): Promise => {\n const ctx = getCtx();\n return ctx.gitService.stashPop(arg);\n });\n \n ipcMain.handle(IPC.gitStashDrop, async (_event, arg: GitStashRefArgs): Promise => {\n const ctx = getCtx();\n return ctx.gitService.stashDrop(arg);\n });\n \n+ ipcMain.handle(IPC.gitStashClear, async (_event, arg: { laneId: string }): Promise => {\n+ const ctx = getCtx();\n+ return ctx.gitService.stashClear(arg);\n+ });\n+\n ipcMain.handle(IPC.gitFetch, async (_event, arg: { laneId: string }): Promise => {\n const ctx = getCtx();\n return ctx.gitService.fetch(arg);\n });\n \n ipcMain.handle(IPC.gitPull, async (_event, arg: { laneId: string }): Promise => {\n const ctx = getCtx();\n return ctx.gitService.pull(arg);\n });\n \n ipcMain.handle(IPC.gitGetSyncStatus, async (_event, arg: { laneId: string }): Promise => {\n const ctx = getCtx();\n return await ctx.gitService.getSyncStatus(arg);\n });\n \n ipcMain.handle(IPC.gitSync, async (_event, arg: GitSyncArgs): Promise => {\n const ctx = getCtx();\n return ctx.gitService.sync(arg);\n });\n \n ipcMain.handle(IPC.gitPush, async (_event, arg: GitPushArgs): Promise => {\n const ctx = getCtx();\n const result = await ctx.gitService.push(arg);\n const lane = await ctx.laneService\n .list({ includeArchived: true, includeStatus: false })\ndiff --git a/apps/desktop/src/main/services/prs/prService.test.ts b/apps/desktop/src/main/services/prs/prService.test.ts\nnew file mode 100644\nindex 0000000000000000000000000000000000000000..a9dd82b06d911dc3506d39905ce0c94b11d14bd4\n--- /dev/null\n+++ b/apps/desktop/src/main/services/prs/prService.test.ts\n@@ -0,0 +1,360 @@\n+import { beforeEach, describe, expect, it, vi } from \"vitest\";\n+\n+// ---------------------------------------------------------------------------\n+// vi.hoisted mock state\n+// ---------------------------------------------------------------------------\n+const mockGit = vi.hoisted(() => ({\n+ runGit: vi.fn(),\n+ runGitOrThrow: vi.fn(),\n+ runGitMergeTree: vi.fn(),\n+}));\n+\n+// ---------------------------------------------------------------------------\n+// vi.mock — external dependencies\n+// ---------------------------------------------------------------------------\n+\n+vi.mock(\"../git/git\", () => ({\n+ runGit: (...args: unknown[]) => mockGit.runGit(...args),\n+ runGitOrThrow: (...args: unknown[]) => mockGit.runGitOrThrow(...args),\n+ runGitMergeTree: (...args: unknown[]) => mockGit.runGitMergeTree(...args),\n+}));\n+\n+vi.mock(\"../ai/utils\", () => ({\n+ extractFirstJsonObject: vi.fn(() => null),\n+}));\n+\n+vi.mock(\"./integrationPlanning\", () => ({\n+ buildIntegrationPreflight: vi.fn(),\n+}));\n+\n+vi.mock(\"./integrationValidation\", () => ({\n+ hasMergeConflictMarkers: vi.fn(() => false),\n+ parseGitStatusPorcelain: vi.fn(() => []),\n+}));\n+\n+vi.mock(\"../shared/queueRebase\", () => ({\n+ fetchRemoteTrackingBranch: vi.fn(),\n+}));\n+\n+import { createPrService } from \"./prService\";\n+\n+// ---------------------------------------------------------------------------\n+// Helpers\n+// ---------------------------------------------------------------------------\n+\n+function makeLogger() {\n+ return {\n+ info: vi.fn(),\n+ warn: vi.fn(),\n+ error: vi.fn(),\n+ debug: vi.fn(),\n+ } as any;\n+}\n+\n+function makeMockDb() {\n+ return {\n+ get: vi.fn(() => null),\n+ all: vi.fn(() => []),\n+ run: vi.fn(),\n+ getJson: vi.fn(() => null),\n+ setJson: vi.fn(),\n+ sync: { getSiteId: vi.fn(), getDbVersion: vi.fn(), exportChangesSince: vi.fn(), applyChanges: vi.fn() },\n+ flushNow: vi.fn(),\n+ close: vi.fn(),\n+ } as any;\n+}\n+\n+const LANE_ID = \"lane-42\";\n+const REPO = { owner: \"test-owner\", name: \"test-repo\" };\n+\n+function makeFakeLane(overrides?: Partial>) {\n+ return {\n+ id: LANE_ID,\n+ name: \"my-feature\",\n+ laneType: \"worktree\",\n+ baseRef: \"refs/heads/main\",\n+ branchRef: \"refs/heads/my-feature\",\n+ worktreePath: \"/tmp/lane-wt\",\n+ parentLaneId: null,\n+ childCount: 0,\n+ stackDepth: 0,\n+ parentStatus: null,\n+ isEditProtected: false,\n+ status: { dirty: false },\n+ color: null,\n+ icon: null,\n+ tags: [],\n+ createdAt: \"2026-01-01T00:00:00Z\",\n+ ...overrides,\n+ };\n+}\n+\n+function makeGithubService(overrides?: Record) {\n+ return {\n+ getRepoOrThrow: vi.fn(async () => REPO),\n+ apiRequest: vi.fn(),\n+ getStatus: vi.fn(),\n+ setToken: vi.fn(),\n+ clearToken: vi.fn(),\n+ getTokenOrThrow: vi.fn(() => \"ghp_mock\"),\n+ ...overrides,\n+ } as any;\n+}\n+\n+function makeLaneService(lanes?: unknown[]) {\n+ return {\n+ list: vi.fn(async () => lanes ?? [makeFakeLane()]),\n+ getLaneBaseAndBranch: vi.fn(),\n+ } as any;\n+}\n+\n+function makeOperationService() {\n+ return {\n+ start: vi.fn(() => ({ operationId: \"op-1\" })),\n+ finish: vi.fn(),\n+ } as any;\n+}\n+\n+function makeProjectConfigService() {\n+ return {\n+ get: vi.fn(() => ({ effective: { ai: {} } })),\n+ } as any;\n+}\n+\n+interface BuildServiceOpts {\n+ githubService?: any;\n+ laneService?: any;\n+ db?: any;\n+}\n+\n+function buildService(opts: BuildServiceOpts = {}) {\n+ const db = opts.db ?? makeMockDb();\n+ const githubService = opts.githubService ?? makeGithubService();\n+ const laneService = opts.laneService ?? makeLaneService();\n+\n+ // Make runGit succeed for upstream check (returns exitCode 0 → push path)\n+ mockGit.runGit.mockResolvedValue({ exitCode: 0, stdout: \"origin/my-feature\", stderr: \"\" });\n+ // Make push succeed\n+ mockGit.runGitOrThrow.mockResolvedValue(undefined);\n+\n+ const service = createPrService({\n+ db,\n+ logger: makeLogger(),\n+ projectId: \"proj-1\",\n+ projectRoot: \"/tmp/test-project\",\n+ laneService,\n+ operationService: makeOperationService(),\n+ githubService,\n+ projectConfigService: makeProjectConfigService(),\n+ openExternal: vi.fn(async () => {}),\n+ });\n+\n+ return { service, db, githubService, laneService };\n+}\n+\n+// ---------------------------------------------------------------------------\n+// Tests\n+// ---------------------------------------------------------------------------\n+\n+describe(\"prService.createFromLane\", () => {\n+ beforeEach(() => {\n+ vi.clearAllMocks();\n+ });\n+\n+ it(\"wraps githubService.apiRequest errors with branch context\", async () => {\n+ const ghService = makeGithubService({\n+ apiRequest: vi.fn().mockRejectedValue(new Error(\"Validation Failed: A pull request already exists\")),\n+ });\n+\n+ const { service } = buildService({ githubService: ghService });\n+\n+ await expect(\n+ service.createFromLane({\n+ laneId: LANE_ID,\n+ title: \"My PR\",\n+ body: \"description\",\n+ draft: false,\n+ allowDirtyWorktree: true,\n+ }),\n+ ).rejects.toThrow(\n+ 'Failed to create pull request for \"my-feature\" \\u2192 \"main\": Validation Failed: A pull request already exists',\n+ );\n+ });\n+\n+ it(\"preserves non-Error throwables in the wrapped message\", async () => {\n+ const ghService = makeGithubService({\n+ apiRequest: vi.fn().mockRejectedValue(\"string error\"),\n+ });\n+\n+ const { service } = buildService({ githubService: ghService });\n+\n+ await expect(\n+ service.createFromLane({\n+ laneId: LANE_ID,\n+ title: \"My PR\",\n+ body: \"\",\n+ draft: false,\n+ allowDirtyWorktree: true,\n+ }),\n+ ).rejects.toThrow(\n+ 'Failed to create pull request for \"my-feature\" \\u2192 \"main\": string error',\n+ );\n+ });\n+\n+ it(\"extracts PR number from successful creation response\", async () => {\n+ const ghService = makeGithubService({\n+ apiRequest: vi.fn().mockResolvedValue({\n+ data: {\n+ number: 99,\n+ html_url: \"https://github.com/test-owner/test-repo/pull/99\",\n+ node_id: \"PR_node1\",\n+ title: \"My PR\",\n+ state: \"open\",\n+ draft: false,\n+ merged_at: null,\n+ head: { ref: \"my-feature\" },\n+ base: { ref: \"main\" },\n+ additions: 10,\n+ deletions: 2,\n+ },\n+ response: { status: 201 },\n+ }),\n+ });\n+\n+ const db = makeMockDb();\n+ // refreshOne calls getRow → fetchPr → apiRequest(GET) → so we need\n+ // db.get to return the inserted row, and apiRequest for the refresh GET\n+ // We'll make db.get return a valid row on the second call (after upsertRow\n+ // inserts via db.run). On the first call (inside upsertRow's getRowForLane),\n+ // return null so it does an INSERT.\n+ let getCallCount = 0;\n+ db.get.mockImplementation(() => {\n+ getCallCount++;\n+ if (getCallCount === 1) {\n+ // getRowForLane inside upsertRow — no existing row\n+ return null;\n+ }\n+ // requireRow inside refreshOne — return the row\n+ return {\n+ id: \"fake-uuid\",\n+ lane_id: LANE_ID,\n+ project_id: \"proj-1\",\n+ repo_owner: \"test-owner\",\n+ repo_name: \"test-repo\",\n+ github_pr_number: 99,\n+ github_url: \"https://github.com/test-owner/test-repo/pull/99\",\n+ github_node_id: \"PR_node1\",\n+ title: \"My PR\",\n+ state: \"open\",\n+ base_branch: \"main\",\n+ head_branch: \"my-feature\",\n+ checks_status: \"none\",\n+ review_status: \"none\",\n+ additions: 10,\n+ deletions: 2,\n+ last_synced_at: null,\n+ created_at: \"2026-01-01T00:00:00Z\",\n+ updated_at: \"2026-01-01T00:00:00Z\",\n+ };\n+ });\n+\n+ // After the initial POST for creation, refreshOne calls fetchPr (GET)\n+ // and then several more GET calls for checks, reviews, comments, files, etc.\n+ // We need apiRequest to handle both the initial POST and subsequent GETs.\n+ let apiCallCount = 0;\n+ ghService.apiRequest.mockImplementation(async (args: any) => {\n+ apiCallCount++;\n+ if (apiCallCount === 1) {\n+ // The POST to create the PR\n+ return {\n+ data: {\n+ number: 99,\n+ html_url: \"https://github.com/test-owner/test-repo/pull/99\",\n+ node_id: \"PR_node1\",\n+ title: \"My PR\",\n+ state: \"open\",\n+ draft: false,\n+ merged_at: null,\n+ head: { ref: \"my-feature\" },\n+ base: { ref: \"main\" },\n+ additions: 10,\n+ deletions: 2,\n+ },\n+ response: { status: 201, headers: new Headers() },\n+ };\n+ }\n+ // All subsequent GETs (fetchPr, checks, reviews, comments, files, actions)\n+ return {\n+ data: args.path.endsWith(\"/pulls/99\")\n+ ? {\n+ number: 99,\n+ html_url: \"https://github.com/test-owner/test-repo/pull/99\",\n+ title: \"My PR\",\n+ state: \"open\",\n+ draft: false,\n+ merged_at: null,\n+ head: { ref: \"my-feature\", sha: \"abc123\" },\n+ base: { ref: \"main\" },\n+ additions: 10,\n+ deletions: 2,\n+ }\n+ : [],\n+ response: {\n+ status: 200,\n+ headers: new Headers(),\n+ },\n+ };\n+ });\n+\n+ const { service } = buildService({ githubService: ghService, db });\n+\n+ const result = await service.createFromLane({\n+ laneId: LANE_ID,\n+ title: \"My PR\",\n+ body: \"description\",\n+ draft: false,\n+ allowDirtyWorktree: true,\n+ });\n+\n+ expect(result.githubPrNumber).toBe(99);\n+ expect(result.headBranch).toBe(\"my-feature\");\n+ expect(result.baseBranch).toBe(\"main\");\n+ });\n+\n+ it(\"throws when GitHub returns an invalid PR number\", async () => {\n+ const ghService = makeGithubService({\n+ apiRequest: vi.fn().mockResolvedValue({\n+ data: { number: null },\n+ response: { status: 201 },\n+ }),\n+ });\n+\n+ const db = makeMockDb();\n+\n+ const { service } = buildService({ githubService: ghService, db });\n+\n+ await expect(\n+ service.createFromLane({\n+ laneId: LANE_ID,\n+ title: \"My PR\",\n+ body: \"\",\n+ draft: false,\n+ allowDirtyWorktree: true,\n+ }),\n+ ).rejects.toThrow(\"GitHub returned an invalid PR number\");\n+ });\n+\n+ it(\"throws when lane is not found\", async () => {\n+ const laneService = makeLaneService([]); // empty list\n+ const { service } = buildService({ laneService });\n+\n+ await expect(\n+ service.createFromLane({\n+ laneId: \"nonexistent\",\n+ title: \"PR\",\n+ body: \"\",\n+ draft: false,\n+ }),\n+ ).rejects.toThrow(\"Lane not found: nonexistent\");\n+ });\n+});\ndiff --git a/apps/desktop/src/main/services/prs/prService.ts b/apps/desktop/src/main/services/prs/prService.ts\nindex 21cc1d90bc8fcc8cb425abe1c836c93a817b32d8..e87cd71e4cef9f357c4bf8be902f27bcd20e56a2 100644\n--- a/apps/desktop/src/main/services/prs/prService.ts\n+++ b/apps/desktop/src/main/services/prs/prService.ts\n@@ -1944,61 +1944,69 @@ export function createPrService({\n const inferredBaseRef = parentLane?.branchRef ?? lane.baseRef ?? primaryLane?.branchRef ?? \"main\";\n const baseBranch = (args.baseBranch ?? branchNameFromRef(inferredBaseRef)).trim();\n \n // Push the branch to remote before creating the PR\n const upstreamCheck = await runGit(\n [\"rev-parse\", \"--abbrev-ref\", \"--symbolic-full-name\", \"@{upstream}\"],\n { cwd: lane.worktreePath, timeoutMs: 10_000 }\n );\n if (upstreamCheck.exitCode === 0) {\n const pushResult = await runGit([\"push\"], { cwd: lane.worktreePath, timeoutMs: 60_000 });\n if (pushResult.exitCode !== 0) {\n const stderr = pushResult.stderr ?? \"\";\n if (stderr.includes(\"non-fast-forward\") || stderr.includes(\"rejected\")) {\n // Branch was rebased locally — force-push safely\n logger.info(\"prs.push_force_lease\", { headBranch, reason: \"non-fast-forward after rebase\" });\n await runGitOrThrow([\"push\", \"--force-with-lease\"], { cwd: lane.worktreePath, timeoutMs: 60_000 });\n } else {\n throw new Error(`Push failed: ${stderr}`);\n }\n }\n } else {\n await runGitOrThrow([\"push\", \"-u\", \"origin\", headBranch], { cwd: lane.worktreePath, timeoutMs: 60_000 });\n }\n \n const createdAt = nowIso();\n- const created = await githubService.apiRequest({\n- method: \"POST\",\n- path: `/repos/${repo.owner}/${repo.name}/pulls`,\n- body: {\n- title: args.title,\n- head: headBranch,\n- base: baseBranch,\n- body: args.body,\n- draft: Boolean(args.draft)\n- }\n- });\n+ let created: { data: any; response: Response | null };\n+ try {\n+ created = await githubService.apiRequest({\n+ method: \"POST\",\n+ path: `/repos/${repo.owner}/${repo.name}/pulls`,\n+ body: {\n+ title: args.title,\n+ head: headBranch,\n+ base: baseBranch,\n+ body: args.body,\n+ draft: Boolean(args.draft)\n+ }\n+ });\n+ } catch (err) {\n+ const msg = err instanceof Error ? err.message : String(err);\n+ throw new Error(\n+ `Failed to create pull request for \"${headBranch}\" → \"${baseBranch}\": ${msg}`\n+ );\n+ }\n \n const pr = created.data;\n const prNumber = Number(pr?.number);\n if (!Number.isFinite(prNumber) || prNumber <= 0) {\n throw new Error(\"GitHub returned an invalid PR number.\");\n }\n \n if (args.labels?.length) {\n await githubService.apiRequest({\n method: \"POST\",\n path: `/repos/${repo.owner}/${repo.name}/issues/${prNumber}/labels`,\n body: { labels: args.labels }\n }).catch((error) => {\n logger.warn(\"prs.labels_failed\", { prNumber, error: error instanceof Error ? error.message : String(error) });\n });\n }\n \n if (args.reviewers?.length) {\n await githubService.apiRequest({\n method: \"POST\",\n path: `/repos/${repo.owner}/${repo.name}/pulls/${prNumber}/requested_reviewers`,\n body: { reviewers: args.reviewers }\n }).catch((error) => {\n logger.warn(\"prs.reviewers_failed\", { prNumber, error: error instanceof Error ? error.message : String(error) });\n });\ndiff --git a/apps/desktop/src/preload/global.d.ts b/apps/desktop/src/preload/global.d.ts\nindex 27734342ed19c5be8419514901a5461638a49cd1..29edf6687237658d40bae8c647e033db661f7efb 100644\n--- a/apps/desktop/src/preload/global.d.ts\n+++ b/apps/desktop/src/preload/global.d.ts\n@@ -60,50 +60,52 @@ import type {\n ExportHistoryResult,\n AgentTool,\n AgentChatApproveArgs,\n AgentChatCreateArgs,\n AgentChatDisposeArgs,\n AgentChatEventEnvelope,\n AgentChatGetSummaryArgs,\n AgentChatHandoffArgs,\n AgentChatHandoffResult,\n AgentChatInterruptArgs,\n AgentChatListArgs,\n AgentChatModelInfo,\n AgentChatModelsArgs,\n AgentChatRespondToInputArgs,\n AgentChatResumeArgs,\n AgentChatSendArgs,\n AgentChatSlashCommand,\n AgentChatSlashCommandsArgs,\n AgentChatFileSearchArgs,\n AgentChatFileSearchResult,\n AgentChatSession,\n AgentChatSessionCapabilities,\n AgentChatSessionCapabilitiesArgs,\n AgentChatSessionSummary,\n AgentChatSteerArgs,\n+ AgentChatCancelSteerArgs,\n+ AgentChatEditSteerArgs,\n AgentChatSubagentSnapshot,\n AgentChatSubagentListArgs,\n AgentChatUpdateSessionArgs,\n AutomationsEventPayload,\n AutomationManualTriggerRequest,\n AutomationRuleSummary,\n AutomationRun,\n AutomationRunDetail,\n AutomationRunListArgs,\n AutomationParseNaturalLanguageRequest,\n AutomationParseNaturalLanguageResult,\n AutomationValidateDraftRequest,\n AutomationValidateDraftResult,\n AutomationSaveDraftRequest,\n AutomationSaveDraftResult,\n AutomationSimulateRequest,\n AutomationSimulateResult,\n UsageSnapshot,\n BudgetCheckResult,\n BudgetCapScope,\n BudgetCapProvider,\n BudgetCapConfig,\n ChangeDigest,\n KnowledgeSyncStatus,\n MemoryHealthStats,\n@@ -803,50 +805,52 @@ declare global {\n oauthDecodeState: (args: DecodeOAuthStateArgs) => Promise;\n oauthListSessions: () => Promise;\n onOAuthEvent: (cb: (ev: OAuthRedirectEvent) => void) => () => void;\n diagnosticsGetStatus: () => Promise;\n diagnosticsGetLaneHealth: (args: GetLaneHealthArgs) => Promise;\n diagnosticsRunHealthCheck: (args: RunHealthCheckArgs) => Promise;\n diagnosticsRunFullCheck: () => Promise;\n diagnosticsActivateFallback: (args: ActivateFallbackArgs) => Promise;\n diagnosticsDeactivateFallback: (args: DeactivateFallbackArgs) => Promise;\n onDiagnosticsEvent: (cb: (ev: RuntimeDiagnosticsEvent) => void) => () => void;\n };\n sessions: {\n list: (args?: ListSessionsArgs) => Promise;\n get: (sessionId: string) => Promise;\n updateMeta: (args: { sessionId: string; pinned?: boolean; title?: string; goal?: string | null; toolType?: string | null; resumeCommand?: string | null }) => Promise;\n readTranscriptTail: (args: ReadTranscriptTailArgs) => Promise;\n getDelta: (sessionId: string) => Promise;\n };\n agentChat: {\n list: (args?: AgentChatListArgs) => Promise;\n getSummary: (args: AgentChatGetSummaryArgs) => Promise;\n create: (args: AgentChatCreateArgs) => Promise;\n handoff: (args: AgentChatHandoffArgs) => Promise;\n send: (args: AgentChatSendArgs) => Promise;\n steer: (args: AgentChatSteerArgs) => Promise;\n+ cancelSteer: (args: AgentChatCancelSteerArgs) => Promise;\n+ editSteer: (args: AgentChatEditSteerArgs) => Promise;\n interrupt: (args: AgentChatInterruptArgs) => Promise;\n resume: (args: AgentChatResumeArgs) => Promise;\n approve: (args: AgentChatApproveArgs) => Promise;\n respondToInput: (args: AgentChatRespondToInputArgs) => Promise;\n models: (args: AgentChatModelsArgs) => Promise;\n dispose: (args: AgentChatDisposeArgs) => Promise;\n updateSession: (args: AgentChatUpdateSessionArgs) => Promise;\n warmupModel: (args: { sessionId: string; modelId: string }) => Promise;\n onEvent: (cb: (ev: AgentChatEventEnvelope) => void) => () => void;\n slashCommands: (args: AgentChatSlashCommandsArgs) => Promise;\n fileSearch: (args: AgentChatFileSearchArgs) => Promise;\n listSubagents: (args: AgentChatSubagentListArgs) => Promise;\n getSessionCapabilities: (args: AgentChatSessionCapabilitiesArgs) => Promise;\n saveTempAttachment: (args: { data: string; filename: string }) => Promise<{ path: string }>;\n };\n computerUse: {\n getSettings: () => Promise;\n listArtifacts: (args?: ComputerUseArtifactListArgs) => Promise;\n getOwnerSnapshot: (args: ComputerUseOwnerSnapshotArgs) => Promise;\n routeArtifact: (args: ComputerUseArtifactRouteArgs) => Promise;\n updateArtifactReview: (args: ComputerUseArtifactReviewArgs) => Promise;\n onEvent: (cb: (ev: ComputerUseEventPayload) => void) => () => void;\n };\n pty: {\n create: (args: PtyCreateArgs) => Promise;\n@@ -873,50 +877,51 @@ declare global {\n watchChanges: (args: FilesWatchArgs) => Promise;\n stopWatching: (args: FilesWatchArgs) => Promise;\n quickOpen: (args: FilesQuickOpenArgs) => Promise;\n searchText: (args: FilesSearchTextArgs) => Promise;\n onChange: (cb: (ev: FileChangeEvent) => void) => () => void;\n };\n git: {\n stageFile: (args: GitFileActionArgs) => Promise;\n stageAll: (args: GitBatchFileActionArgs) => Promise;\n unstageFile: (args: GitFileActionArgs) => Promise;\n unstageAll: (args: GitBatchFileActionArgs) => Promise;\n discardFile: (args: GitFileActionArgs) => Promise;\n restoreStagedFile: (args: GitFileActionArgs) => Promise;\n commit: (args: GitCommitArgs) => Promise;\n generateCommitMessage: (args: GitGenerateCommitMessageArgs) => Promise;\n listRecentCommits: (args: { laneId: string; limit?: number }) => Promise;\n listCommitFiles: (args: GitListCommitFilesArgs) => Promise;\n getCommitMessage: (args: GitGetCommitMessageArgs) => Promise;\n revertCommit: (args: GitRevertArgs) => Promise;\n cherryPickCommit: (args: GitCherryPickArgs) => Promise;\n stashPush: (args: GitStashPushArgs) => Promise;\n stashList: (args: { laneId: string }) => Promise;\n stashApply: (args: GitStashRefArgs) => Promise;\n stashPop: (args: GitStashRefArgs) => Promise;\n stashDrop: (args: GitStashRefArgs) => Promise;\n+ stashClear: (args: { laneId: string }) => Promise;\n fetch: (args: { laneId: string }) => Promise;\n pull: (args: { laneId: string }) => Promise;\n getSyncStatus: (args: { laneId: string }) => Promise;\n sync: (args: GitSyncArgs) => Promise;\n push: (args: GitPushArgs) => Promise;\n getConflictState: (laneId: string) => Promise;\n rebaseContinue: (laneId: string) => Promise;\n rebaseAbort: (laneId: string) => Promise;\n mergeContinue: (laneId: string) => Promise;\n mergeAbort: (laneId: string) => Promise;\n listBranches: (args: GitListBranchesArgs) => Promise;\n checkoutBranch: (args: GitCheckoutBranchArgs) => Promise;\n };\n conflicts: {\n getLaneStatus: (args: GetLaneConflictStatusArgs) => Promise;\n listOverlaps: (args: ListOverlapsArgs) => Promise;\n getRiskMatrix: () => Promise;\n simulateMerge: (args: MergeSimulationArgs) => Promise;\n runPrediction: (args?: RunConflictPredictionArgs) => Promise;\n getBatchAssessment: () => Promise;\n listProposals: (laneId: string) => Promise;\n prepareProposal: (args: PrepareConflictProposalArgs) => Promise;\n requestProposal: (args: RequestConflictProposalArgs) => Promise;\n applyProposal: (args: ApplyConflictProposalArgs) => Promise;\n undoProposal: (args: UndoConflictProposalArgs) => Promise;\ndiff --git a/apps/desktop/src/preload/preload.ts b/apps/desktop/src/preload/preload.ts\nindex e405e75ce078ffbf8b82365419fa28a6536486e7..f9c5f6168a078ca39461357f721dc84c50afa3ba 100644\n--- a/apps/desktop/src/preload/preload.ts\n+++ b/apps/desktop/src/preload/preload.ts\n@@ -219,50 +219,52 @@ import type {\n ExportHistoryResult,\n AgentTool,\n AgentChatApproveArgs,\n AgentChatCreateArgs,\n AgentChatDisposeArgs,\n AgentChatEventEnvelope,\n AgentChatGetSummaryArgs,\n AgentChatHandoffArgs,\n AgentChatHandoffResult,\n AgentChatInterruptArgs,\n AgentChatListArgs,\n AgentChatModelInfo,\n AgentChatModelsArgs,\n AgentChatRespondToInputArgs,\n AgentChatResumeArgs,\n AgentChatSendArgs,\n AgentChatSlashCommand,\n AgentChatSlashCommandsArgs,\n AgentChatFileSearchArgs,\n AgentChatFileSearchResult,\n AgentChatSession,\n AgentChatSessionCapabilities,\n AgentChatSessionCapabilitiesArgs,\n AgentChatSessionSummary,\n AgentChatSteerArgs,\n+ AgentChatCancelSteerArgs,\n+ AgentChatEditSteerArgs,\n AgentChatSubagentSnapshot,\n AgentChatSubagentListArgs,\n AgentChatUpdateSessionArgs,\n KeybindingOverride,\n KeybindingsSnapshot,\n OnboardingDetectionResult,\n OnboardingExistingLaneCandidate,\n OnboardingStatus,\n LaneSummary,\n ListOverlapsArgs,\n ListLanesArgs,\n ListMissionsArgs,\n ImportBranchLaneArgs,\n ListOperationsArgs,\n ListSessionsArgs,\n ListTestRunsArgs,\n MergeSimulationArgs,\n MergeSimulationResult,\n OperationRecord,\n ProcessActionArgs,\n ProcessDefinition,\n ProcessEvent,\n ProcessRuntime,\n ProcessStackArgs,\n ProjectConfigCandidate,\n@@ -1098,50 +1100,54 @@ contextBridge.exposeInMainWorld(\"ade\", {\n sessions: {\n list: async (args: ListSessionsArgs = {}): Promise =>\n ipcRenderer.invoke(IPC.sessionsList, args),\n get: async (sessionId: string): Promise =>\n ipcRenderer.invoke(IPC.sessionsGet, { sessionId }),\n updateMeta: async (args: { sessionId: string; pinned?: boolean; title?: string; goal?: string | null; toolType?: string | null; resumeCommand?: string | null }): Promise =>\n ipcRenderer.invoke(IPC.sessionsUpdateMeta, args),\n readTranscriptTail: async (args: ReadTranscriptTailArgs): Promise =>\n ipcRenderer.invoke(IPC.sessionsReadTranscriptTail, args),\n getDelta: async (sessionId: string): Promise =>\n ipcRenderer.invoke(IPC.sessionsGetDelta, { sessionId })\n },\n agentChat: {\n list: async (args: AgentChatListArgs = {}): Promise =>\n ipcRenderer.invoke(IPC.agentChatList, args),\n getSummary: async (args: AgentChatGetSummaryArgs): Promise =>\n ipcRenderer.invoke(IPC.agentChatGetSummary, args),\n create: async (args: AgentChatCreateArgs): Promise =>\n ipcRenderer.invoke(IPC.agentChatCreate, args),\n handoff: async (args: AgentChatHandoffArgs): Promise =>\n ipcRenderer.invoke(IPC.agentChatHandoff, args),\n send: async (args: AgentChatSendArgs): Promise =>\n ipcRenderer.invoke(IPC.agentChatSend, args),\n steer: async (args: AgentChatSteerArgs): Promise =>\n ipcRenderer.invoke(IPC.agentChatSteer, args),\n+ cancelSteer: async (args: AgentChatCancelSteerArgs): Promise =>\n+ ipcRenderer.invoke(IPC.agentChatCancelSteer, args),\n+ editSteer: async (args: AgentChatEditSteerArgs): Promise =>\n+ ipcRenderer.invoke(IPC.agentChatEditSteer, args),\n interrupt: async (args: AgentChatInterruptArgs): Promise =>\n ipcRenderer.invoke(IPC.agentChatInterrupt, args),\n resume: async (args: AgentChatResumeArgs): Promise =>\n ipcRenderer.invoke(IPC.agentChatResume, args),\n approve: async (args: AgentChatApproveArgs): Promise =>\n ipcRenderer.invoke(IPC.agentChatApprove, args),\n respondToInput: async (args: AgentChatRespondToInputArgs): Promise =>\n ipcRenderer.invoke(IPC.agentChatRespondToInput, args),\n models: async (args: AgentChatModelsArgs): Promise =>\n ipcRenderer.invoke(IPC.agentChatModels, args),\n dispose: async (args: AgentChatDisposeArgs): Promise =>\n ipcRenderer.invoke(IPC.agentChatDispose, args),\n updateSession: async (args: AgentChatUpdateSessionArgs): Promise =>\n ipcRenderer.invoke(IPC.agentChatUpdateSession, args),\n warmupModel: async (args: { sessionId: string; modelId: string }): Promise =>\n ipcRenderer.invoke(IPC.agentChatWarmupModel, args),\n onEvent: (cb: (ev: AgentChatEventEnvelope) => void) => {\n const listener = (_event: Electron.IpcRendererEvent, payload: AgentChatEventEnvelope) => cb(payload);\n ipcRenderer.on(IPC.agentChatEvent, listener);\n return () => ipcRenderer.removeListener(IPC.agentChatEvent, listener);\n },\n slashCommands: async (args: AgentChatSlashCommandsArgs): Promise =>\n ipcRenderer.invoke(IPC.agentChatSlashCommands, args),\n fileSearch: async (args: AgentChatFileSearchArgs): Promise =>\n ipcRenderer.invoke(IPC.agentChatFileSearch, args),\n@@ -1215,50 +1221,51 @@ contextBridge.exposeInMainWorld(\"ade\", {\n git: {\n stageFile: async (args: GitFileActionArgs): Promise => ipcRenderer.invoke(IPC.gitStageFile, args),\n stageAll: async (args: GitBatchFileActionArgs): Promise => ipcRenderer.invoke(IPC.gitStageAll, args),\n unstageFile: async (args: GitFileActionArgs): Promise => ipcRenderer.invoke(IPC.gitUnstageFile, args),\n unstageAll: async (args: GitBatchFileActionArgs): Promise => ipcRenderer.invoke(IPC.gitUnstageAll, args),\n discardFile: async (args: GitFileActionArgs): Promise => ipcRenderer.invoke(IPC.gitDiscardFile, args),\n restoreStagedFile: async (args: GitFileActionArgs): Promise =>\n ipcRenderer.invoke(IPC.gitRestoreStagedFile, args),\n commit: async (args: GitCommitArgs): Promise => ipcRenderer.invoke(IPC.gitCommit, args),\n generateCommitMessage: async (args: GitGenerateCommitMessageArgs): Promise =>\n ipcRenderer.invoke(IPC.gitGenerateCommitMessage, args),\n listRecentCommits: async (args: { laneId: string; limit?: number }): Promise =>\n ipcRenderer.invoke(IPC.gitListRecentCommits, args),\n listCommitFiles: async (args: GitListCommitFilesArgs): Promise =>\n ipcRenderer.invoke(IPC.gitListCommitFiles, args),\n getCommitMessage: async (args: GitGetCommitMessageArgs): Promise =>\n ipcRenderer.invoke(IPC.gitGetCommitMessage, args),\n revertCommit: async (args: GitRevertArgs): Promise => ipcRenderer.invoke(IPC.gitRevertCommit, args),\n cherryPickCommit: async (args: GitCherryPickArgs): Promise =>\n ipcRenderer.invoke(IPC.gitCherryPickCommit, args),\n stashPush: async (args: GitStashPushArgs): Promise => ipcRenderer.invoke(IPC.gitStashPush, args),\n stashList: async (args: { laneId: string }): Promise => ipcRenderer.invoke(IPC.gitStashList, args),\n stashApply: async (args: GitStashRefArgs): Promise => ipcRenderer.invoke(IPC.gitStashApply, args),\n stashPop: async (args: GitStashRefArgs): Promise => ipcRenderer.invoke(IPC.gitStashPop, args),\n stashDrop: async (args: GitStashRefArgs): Promise => ipcRenderer.invoke(IPC.gitStashDrop, args),\n+ stashClear: async (args: { laneId: string }): Promise => ipcRenderer.invoke(IPC.gitStashClear, args),\n fetch: async (args: { laneId: string }): Promise => ipcRenderer.invoke(IPC.gitFetch, args),\n pull: async (args: { laneId: string }): Promise => ipcRenderer.invoke(IPC.gitPull, args),\n getSyncStatus: async (args: { laneId: string }): Promise =>\n ipcRenderer.invoke(IPC.gitGetSyncStatus, args),\n sync: async (args: GitSyncArgs): Promise => ipcRenderer.invoke(IPC.gitSync, args),\n push: async (args: GitPushArgs): Promise => ipcRenderer.invoke(IPC.gitPush, args),\n getConflictState: async (laneId: string): Promise =>\n ipcRenderer.invoke(IPC.gitGetConflictState, { laneId }),\n rebaseContinue: async (laneId: string): Promise =>\n ipcRenderer.invoke(IPC.gitRebaseContinue, { laneId }),\n rebaseAbort: async (laneId: string): Promise =>\n ipcRenderer.invoke(IPC.gitRebaseAbort, { laneId }),\n mergeContinue: async (laneId: string): Promise =>\n ipcRenderer.invoke(IPC.gitMergeContinue, { laneId }),\n mergeAbort: async (laneId: string): Promise =>\n ipcRenderer.invoke(IPC.gitMergeAbort, { laneId }),\n listBranches: async (args: GitListBranchesArgs): Promise =>\n ipcRenderer.invoke(IPC.gitListBranches, args),\n checkoutBranch: async (args: GitCheckoutBranchArgs): Promise =>\n ipcRenderer.invoke(IPC.gitCheckoutBranch, args)\n },\n conflicts: {\n getLaneStatus: async (args: GetLaneConflictStatusArgs): Promise =>\n ipcRenderer.invoke(IPC.conflictsGetLaneStatus, args),\n listOverlaps: async (args: ListOverlapsArgs): Promise =>\ndiff --git a/apps/desktop/src/renderer/components/app/TabNav.tsx b/apps/desktop/src/renderer/components/app/TabNav.tsx\nindex 78d4ce13f7b2304b9e3e223278db0cc94485d52d..77fc85616a0dd757b42552e5bae8e60c7bc76672 100644\n--- a/apps/desktop/src/renderer/components/app/TabNav.tsx\n+++ b/apps/desktop/src/renderer/components/app/TabNav.tsx\n@@ -93,54 +93,54 @@ export function TabNav() {\n {isActive && (\n \n )}\n \n {/* Fixed-width icon container - never moves during collapse */}\n \n \n \n {/* Terminal attention dot */}\n {it.to === \"/work\" && terminalAttention.indicator !== \"none\" ? (\n \n ) : null}\n \n \n \n {/* Label - opacity-animated separately from width transition */}\n \n {it.label}\n \n \n );\n };\n \n return (\n <>\n \n {/* Core navigation items */}\ndiff --git a/apps/desktop/src/renderer/components/app/TopBar.tsx b/apps/desktop/src/renderer/components/app/TopBar.tsx\nindex c700f0ba1f46a664c0a62fe038f282036c0c4495..61badd1fb53e6acf3c4d7cbc0796eaff58689d2d 100644\n--- a/apps/desktop/src/renderer/components/app/TopBar.tsx\n+++ b/apps/desktop/src/renderer/components/app/TopBar.tsx\n@@ -329,54 +329,54 @@ export function TopBar() {\n onKeyDown={(event) => {\n if (isMissing) return;\n if (event.key === \"Enter\" || event.key === \" \") {\n event.preventDefault();\n handleSwitchProject(rp.rootPath);\n }\n }}\n title={isMissing ? `Missing: ${rp.rootPath}` : rp.rootPath}\n >\n \n {isCurrent && indicator != null && indicator !== \"none\" ? (\n \n ) : null}\n \n {rp.displayName}\n \n {isMissing ? (\n \n {\n e.stopPropagation();\n if (isRelocating) return;\n handleRelocate(rp.rootPath);\n }}\n title=\"Relocate project\"\ndiff --git a/apps/desktop/src/renderer/components/chat/AgentChatMessageList.test.tsx b/apps/desktop/src/renderer/components/chat/AgentChatMessageList.test.tsx\nindex 9a9bd087025fe863556e5e9c19bbbdbe9f8e8a52..f4fcd692cd04b2c1c05ab7e801009323ff919329 100644\n--- a/apps/desktop/src/renderer/components/chat/AgentChatMessageList.test.tsx\n+++ b/apps/desktop/src/renderer/components/chat/AgentChatMessageList.test.tsx\n@@ -589,71 +589,72 @@ describe(\"AgentChatMessageList transcript rendering\", () => {\n type: \"subagent_started\",\n taskId: \"agent-1\",\n description: \"Check Claude task list support\",\n background: true,\n turnId: \"turn-1\",\n },\n },\n ],\n {\n initialState: { laneId: \"lane-123\" },\n },\n );\n \n expect(screen.getAllByText(\"1 of 2 tasks completed\").length).toBeGreaterThanOrEqual(1);\n expect(screen.getAllByText(/1 file changed/i).length).toBeGreaterThanOrEqual(1);\n expect(screen.getAllByText(/1 background agent/i).length).toBeGreaterThanOrEqual(1);\n expect(screen.getAllByText(\"1 active\").length).toBeGreaterThanOrEqual(1);\n expect(screen.getAllByText(\"+1\").length).toBeGreaterThanOrEqual(1);\n expect(screen.getAllByText(\"-1\").length).toBeGreaterThanOrEqual(1);\n \n fireEvent.click(screen.getByRole(\"button\", { name: \"Review changes\" }));\n \n expect(screen.getByTestId(\"location\").textContent).toBe(\"/files::{\\\"laneId\\\":\\\"lane-123\\\"}\");\n });\n \n- it(\"renders ask-user requests with an amber waiting spinner\", () => {\n+ it(\"renders ask-user requests with an amber waiting icon\", () => {\n const view = renderMessageList([\n {\n sessionId: \"session-1\",\n timestamp: \"2026-03-17T10:00:00.000Z\",\n event: {\n type: \"approval_request\",\n itemId: \"approval-1\",\n kind: \"tool_call\",\n description: \"Which branch should I use?\",\n turnId: \"turn-1\",\n detail: {\n tool: \"askUser\",\n question: \"Which branch should I use?\",\n },\n },\n },\n ]);\n \n expect(screen.getByText(\"Needs Input\")).toBeTruthy();\n- expect(view.container.querySelector(\".animate-spin.text-amber-400\")).toBeTruthy();\n+ expect(view.container.querySelector(\"svg.text-amber-400\")).toBeTruthy();\n+ expect(view.container.querySelector(\".animate-spin.text-amber-400\")).toBeFalsy();\n });\n \n it(\"labels provider chats as Codex and preserves explicit assistant labels\", () => {\n renderMessageList([\n {\n sessionId: \"session-1\",\n timestamp: \"2026-03-17T10:00:00.000Z\",\n event: {\n type: \"text\",\n text: \"Streaming response\",\n itemId: \"text-1\",\n turnId: \"turn-1\",\n },\n },\n {\n sessionId: \"session-1\",\n timestamp: \"2026-03-17T10:00:01.000Z\",\n event: {\n type: \"done\",\n turnId: \"turn-1\",\n status: \"completed\",\n modelId: \"gpt-5.4-codex\",\n },\n },\n ]);\ndiff --git a/apps/desktop/src/renderer/components/chat/chatStatusVisuals.tsx b/apps/desktop/src/renderer/components/chat/chatStatusVisuals.tsx\nindex 75ecc94baf5e1e49cdeccd8c0a23a5ecf486cd5e..489b2e06ae43a4a0a9d89787db5344e36a6bc78f 100644\n--- a/apps/desktop/src/renderer/components/chat/chatStatusVisuals.tsx\n+++ b/apps/desktop/src/renderer/components/chat/chatStatusVisuals.tsx\n@@ -1,46 +1,46 @@\n-import { CheckCircle, SpinnerGap, XCircle } from \"@phosphor-icons/react\";\n+import { CheckCircle, Clock, SpinnerGap, XCircle } from \"@phosphor-icons/react\";\n import { cn } from \"../ui/cn\";\n \n export type ChatStatusVisualState = \"working\" | \"waiting\" | \"completed\" | \"failed\";\n \n export const CHAT_STATUS_HEX = {\n working: \"#10B981\",\n waiting: \"#F59E0B\",\n completed: \"#10B981\",\n failed: \"#EF4444\",\n idle: \"#6B7280\",\n } as const;\n \n export function chatStatusTextClass(status: ChatStatusVisualState): string {\n switch (status) {\n case \"completed\":\n return \"text-emerald-300/75\";\n case \"failed\":\n return \"text-red-300/80\";\n case \"waiting\":\n return \"text-amber-300/80\";\n default:\n return \"text-emerald-300/80\";\n }\n }\n \n export function ChatStatusGlyph({\n status,\n size = 12,\n className,\n }: {\n status: ChatStatusVisualState;\n size?: number;\n className?: string;\n }) {\n switch (status) {\n case \"completed\":\n return ;\n case \"failed\":\n return ;\n case \"waiting\":\n- return ;\n+ return ;\n case \"working\":\n return ;\n }\n }\ndiff --git a/apps/desktop/src/renderer/components/lanes/LaneGitActionsPane.tsx b/apps/desktop/src/renderer/components/lanes/LaneGitActionsPane.tsx\nindex d47add4192b7e0cb9642e249060ff5a5c4cc56ce..8a9f27d6acea1ccfc76bba1bd8374041faec40a5 100644\n--- a/apps/desktop/src/renderer/components/lanes/LaneGitActionsPane.tsx\n+++ b/apps/desktop/src/renderer/components/lanes/LaneGitActionsPane.tsx\n@@ -343,50 +343,51 @@ export function LaneGitActionsPane({\n \n const [loading, setLoading] = useState(false);\n const [changes, setChanges] = useState({ unstaged: [], staged: [] });\n const [commitMessage, setCommitMessage] = useState(\"\");\n const [commitMessageAi, setCommitMessageAi] = useState({ enabled: false, modelId: null });\n const [syncMode, setSyncMode] = useState(\"merge\");\n const [stashes, setStashes] = useState([]);\n const [syncStatus, setSyncStatus] = useState(null);\n const [forcePushSuggested, setForcePushSuggested] = useState(false);\n const [busyAction, setBusyAction] = useState(null);\n const [notice, setNotice] = useState(null);\n const [error, setError] = useState(null);\n const [textPrompt, setTextPrompt] = useState(null);\n const [textPromptError, setTextPromptError] = useState(null);\n const [commitTimelineKey, setCommitTimelineKey] = useState(0);\n const [amendCommit, setAmendCommit] = useState(false);\n const [showAdvanced, setShowAdvanced] = useState(false);\n const [autoRebaseStatus, setAutoRebaseStatus] = useState(null);\n const [conflictState, setConflictState] = useState(null);\n const [stuckRebase, setStuckRebase] = useState(null);\n \n const stagedCount = changes.staged.length;\n const hasStaged = stagedCount > 0;\n const hasUnstaged = changes.unstaged.length > 0;\n const responsiveMode = getResponsiveMode(paneWidth);\n+ const maxVisibleStashes = responsiveMode === \"wide\" ? 2 : 3;\n const actionGridColumns =\n responsiveMode === \"wide\" ? \"repeat(3, minmax(0, 1fr))\" : responsiveMode === \"medium\" ? \"repeat(2, minmax(0, 1fr))\" : \"1fr\";\n \n useEffect(() => {\n if (typeof ResizeObserver === \"undefined\") {\n setPaneWidth(rootRef.current?.clientWidth ?? window.innerWidth);\n return;\n }\n const node = rootRef.current;\n if (!node) return;\n const observer = new ResizeObserver((entries) => {\n const width = entries[0]?.contentRect.width ?? node.clientWidth;\n if (width > 0) setPaneWidth(width);\n });\n observer.observe(node);\n setPaneWidth(node.clientWidth || window.innerWidth);\n return () => observer.disconnect();\n }, []);\n \n const requestTextInput = useCallback(\n (args: {\n title: string;\n message?: string;\n placeholder?: string;\n defaultValue?: string;\n@@ -866,50 +867,51 @@ export function LaneGitActionsPane({\n if (syncStatus.recommendedAction === \"push\") {\n return {\n action: \"push\",\n label: \"Push\",\n detail: `${syncStatus.ahead} local commit${syncStatus.ahead === 1 ? \"\" : \"s\"} are ready to send to remote.`\n };\n }\n if (syncStatus.recommendedAction === \"pull\") {\n if (syncStatus.diverged) {\n return {\n action: \"pull\",\n label: \"Resolve divergence\",\n detail: \"Remote and local both changed. Pull (rebase) keeps remote changes; force push publishes your rewritten local history.\"\n };\n }\n return {\n action: \"pull\",\n label: \"Pull\",\n detail: `${syncStatus.behind} upstream commit${syncStatus.behind === 1 ? \"\" : \"s\"} have not been brought into this lane yet.`\n };\n }\n return null;\n }, [forcePushSuggested, lane, laneId, syncStatus]);\n \n const divergedSync = Boolean(syncStatus?.diverged);\n+ const behindCount = syncStatus?.behind ?? 0;\n const headerDotColor = getLaneHeaderDotColor(lane);\n const pushButtonTitle = syncStatus?.hasUpstream === false ? \"Publish lane\" : \"Push to remote\";\n const rebaseConflictParentLaneId = autoRebaseStatus?.parentLaneId ?? lane?.parentLaneId ?? null;\n const isGeneratingCommitMessage = busyAction === AUTO_GENERATE_COMMIT_ACTION;\n const commitButtonLabel = getCommitButtonLabel({ busyAction, amendCommit });\n const commitHelperText = getCommitHelperText({ commitMessage, commitMessageAi });\n const primaryPushLabel = syncStatus?.hasUpstream === false ? \"Publish lane\" : \"Push to remote\";\n \n const renderFileRow = (file: FileChange, mode: \"staged\" | \"unstaged\") => {\n const rowSelected = selectedPath === file.path && selectedMode === mode;\n const alsoStaged = mode === \"unstaged\" && stagedPathSet.has(file.path);\n const alsoUnstaged = mode === \"staged\" && unstagedPathSet.has(file.path);\n const kindColor = getFileKindColor(file.kind);\n \n return (\n \n \n {/* Sync controls */}\n

\n {([\"merge\", \"rebase\"] as const).map((mode) => (\n setSyncMode(mode)}\n style={{\n ...outlineButton({ height: 26, padding: \"0 6px\", fontSize: 9, borderRadius: 4 }),\n color: syncMode === mode ? COLORS.accent : COLORS.textDim,\n border: `1px solid ${syncMode === mode ? `${COLORS.accent}40` : \"transparent\"}`,\n background: syncMode === mode ? `${COLORS.accent}10` : \"transparent\",\n opacity: !laneId || busyAction != null ? 0.5 : 1,\n }}\n title={getPullModeSummary(mode)}\n >\n {mode === \"merge\" ? \"MERGE\" : \"REBASE\"}\n \n ))}\n

\n 0 ? \"pull-btn-flash\" : undefined}\n style={{\n ...outlineButton({ height: 30, padding: \"0 10px\", fontSize: 10, borderRadius: 6 }),\n ...(nextActionHint?.action === \"pull\" ? { color: COLORS.accent, border: `1px solid ${COLORS.accent}40`, background: `${COLORS.accent}08` } : {}),\n }}\n disabled={!laneId || busyAction != null}\n onClick={() => runPull(syncMode)}\n title={`Pull (${syncMode}). ${getPullModeSummary(syncMode)}`}\n >\n \n PULL\n+ {behindCount > 0 && (\n+ \n+ {behindCount}\n+ \n+ )}\n \n runPush(false)}\n title={getPushSummary(syncStatus)}\n >\n \n {syncStatus?.hasUpstream === false ? \"PUBLISH\" : \"PUSH\"}\n \n {lane?.parentLaneId ? (\n runRebaseAndPushFlow(true)}\n title=\"Rebase from parent and push\"\n@@ -1557,142 +1579,176 @@ export function LaneGitActionsPane({\n void moveUnstagedToNewLane();\n }}\n >\n CREATE NEW LANE WITH CURRENT CHANGES\n \n ) : null}\n \n }\n >\n \n

\n STASHES\n \n {stashes.length === 0 ? \"None saved\" : `${stashes.length} saved`}\n \n

\n- {\n- if (!laneId) return;\n- void runAction(\"stash push\", async () => {\n- const msg = await requestTextInput({\n- title: \"Stash message\",\n- placeholder: \"Optional note\",\n- confirmLabel: \"Save stash\",\n+

\n+ {stashes.length > 0 && (\n+ {\n+ if (!laneId) return;\n+ void runAction(\"stash clear\", async () => {\n+ const confirmation = await requestTextInput({\n+ title: \"Clear all stashes?\",\n+ message: `This will permanently delete ${stashes.length} stash${stashes.length === 1 ? \"\" : \"es\"}. Type \"${stashes.length}\" to confirm.`,\n+ placeholder: `Type ${stashes.length} to confirm`,\n+ confirmLabel: \"Delete all\",\n+ validate: (v) => v.trim() === String(stashes.length) ? null : \"Type the number to confirm\",\n+ });\n+ if (confirmation == null) throw new Error(\"__ade_cancelled__\");\n+ await window.ade.git.stashClear({ laneId });\n+ });\n+ }}\n+ >\n+ CLEAR ALL\n+ \n+ )}\n+ {\n+ if (!laneId) return;\n+ void runAction(\"stash push\", async () => {\n+ const msg = await requestTextInput({\n+ title: \"Stash message\",\n+ placeholder: \"Optional note\",\n+ confirmLabel: \"Save stash\",\n+ });\n+ if (msg == null) throw new Error(\"__ade_cancelled__\");\n+ await window.ade.git.stashPush({ laneId, message: msg || undefined });\n });\n- if (msg == null) throw new Error(\"__ade_cancelled__\");\n- await window.ade.git.stashPush({ laneId, message: msg || undefined });\n- });\n- }}\n- >\n- STASH NOW\n- \n+ }}\n+ >\n+ SAVE CHANGES\n+ \n+

\n {stashes.length === 0 ? (\n

\n- Use stash when you want to clear the worktree without committing. Hover the buttons for the git details.\n+ Save your in-progress changes without committing. You can restore them later.\n

\n ) : (\n

\n- {stashes.slice(0, responsiveMode === \"wide\" ? 2 : 3).map((stash) => (\n+ {stashes.slice(0, maxVisibleStashes).map((stash) => (\n \n-

\n-

\n- {stash.subject || stash.ref}\n-

\n-

\n- {stash.ref} · {formatRelativeTime(stash.createdAt)}\n+

\n+

\n+ {stash.subject || stash.ref}\n+

\n+

\n+ {stash.ref} · {formatRelativeTime(stash.createdAt)}\n+

\n+ {\n+ if (!laneId) return;\n+ void runAction(\"stash pop\", async () => {\n+ await window.ade.git.stashPop({ laneId, stashRef: stash.ref });\n+ });\n+ }}\n+ >\n+ RESTORE\n+ \n+ {\n+ if (!laneId) return;\n+ void runAction(\"stash apply\", async () => {\n+ await window.ade.git.stashApply({ laneId, stashRef: stash.ref });\n+ });\n+ }}\n+ >\n+ COPY TO WORKTREE\n+ \n+ {\n+ if (!laneId) return;\n+ void runAction(\"stash drop\", async () => {\n+ await window.ade.git.stashDrop({ laneId, stashRef: stash.ref });\n+ });\n+ }}\n+ >\n+ DELETE\n+ \n+

\n+

\n+ Restore removes entry. Copy to Worktree keeps it. Delete discards permanently.\n

\n- {\n- if (!laneId) return;\n- void runAction(\"stash apply\", async () => {\n- await window.ade.git.stashApply({ laneId, stashRef: stash.ref });\n- });\n- }}\n- >\n- APPLY\n- \n- {\n- if (!laneId) return;\n- void runAction(\"stash pop\", async () => {\n- await window.ade.git.stashPop({ laneId, stashRef: stash.ref });\n- });\n- }}\n- >\n- POP\n- \n- {\n- if (!laneId) return;\n- void runAction(\"stash drop\", async () => {\n- await window.ade.git.stashDrop({ laneId, stashRef: stash.ref });\n- });\n- }}\n- >\n- DROP\n- \n

\n ))}\n- {stashes.length > (responsiveMode === \"wide\" ? 2 : 3) ? (\n+ {stashes.length > maxVisibleStashes ? (\n

\n- +{stashes.length - (responsiveMode === \"wide\" ? 2 : 3)} more stash entr{stashes.length - (responsiveMode === \"wide\" ? 2 : 3) === 1 ? \"y\" : \"ies\"}.\n+ +{stashes.length - maxVisibleStashes} more stash entr{stashes.length - maxVisibleStashes === 1 ? \"y\" : \"ies\"}.\n

\n ) : null}\n

\n )}\n

\n \n

\n {changes.staged.length > 0 ? (\n

STAGED ({changes.staged.length})

\n {changes.staged.map((file) => renderFileRow(file, \"staged\"))}\n

\n ) : null}\n {changes.unstaged.length > 0 ? (\n

UNSTAGED ({changes.unstaged.length})

\n {changes.unstaged.map((file) => renderFileRow(file, \"unstaged\"))}\n

\n ) : null}\n {changes.staged.length === 0 && changes.unstaged.length === 0 ? (\n

\n No changes\n

\n ) : null}\n

\ndiff --git a/apps/desktop/src/renderer/components/lanes/LaneTerminalsPanel.tsx b/apps/desktop/src/renderer/components/lanes/LaneTerminalsPanel.tsx\nindex 88751f6907e6cb9a0c10cb91a560b16a091b346c..e8bd46fafa72255760c5fcfa7d9ef39c7d2901cf 100644\n--- a/apps/desktop/src/renderer/components/lanes/LaneTerminalsPanel.tsx\n+++ b/apps/desktop/src/renderer/components/lanes/LaneTerminalsPanel.tsx\n@@ -1,50 +1,50 @@\n import { useCallback, useEffect, useMemo, useRef, useState } from \"react\";\n import * as Tabs from \"@radix-ui/react-tabs\";\n import { ArrowSquareOut, GridFour, List, X } from \"@phosphor-icons/react\";\n import { useAppStore } from \"../../state/appStore\";\n import type { TerminalSessionSummary } from \"../../../shared/types\";\n import { Button } from \"../ui/Button\";\n import { Chip } from \"../ui/Chip\";\n import { EmptyState } from \"../ui/EmptyState\";\n import { cn } from \"../ui/cn\";\n import { TerminalView } from \"../terminals/TerminalView\";\n import { TilingLayout } from \"./TilingLayout\";\n import { useNavigate } from \"react-router-dom\";\n import { sessionIndicatorState } from \"../../lib/terminalAttention\";\n import { isChatToolType, isRunOwnedSession, primarySessionLabel, secondarySessionLabel } from \"../../lib/sessions\";\n import { listSessionsCached } from \"../../lib/sessionListCache\";\n import { defaultTrackedCliStartupCommand } from \"../terminals/cliLaunch\";\n import { ToolLogo } from \"../terminals/ToolLogos\";\n import { persistLaunchTracked, readLaunchTracked } from \"../../lib/terminalLaunchPreferences\";\n \n const tabTrigger =\n \"flex items-center gap-2 rounded-md px-2.5 py-2 text-xs font-semibold text-muted-fg data-[state=active]:text-fg data-[state=active]:bg-accent/10 data-[state=active]:ring-1 data-[state=active]:ring-accent/50\";\n \n function statusDotCls(indicator: ReturnType): string {\n if (indicator === \"running-active\") return \"border-2 border-emerald-500 border-t-transparent bg-transparent\";\n- if (indicator === \"running-needs-attention\") return \"border-2 border-amber-400 border-t-transparent bg-transparent\";\n+ if (indicator === \"running-needs-attention\") return \"bg-amber-300\";\n return \"bg-red-500\";\n }\n \n function sessionTabLabel(session: TerminalSessionSummary): string {\n const base = primarySessionLabel(session);\n const secondary = secondarySessionLabel(session);\n if (!secondary) return base.slice(0, 180);\n return `${base} · ${secondary}`.slice(0, 180);\n }\n \n const TOOL_BUTTONS = [\n { id: \"claude\", label: \"Launch Claude\", variant: \"primary\" as const, style: { backgroundColor: \"#f97316\", borderColor: \"#f97316\", color: \"#fff\" } },\n { id: \"codex\", label: \"Launch Codex\", variant: \"primary\" as const, style: { backgroundColor: \"#3b82f6\", borderColor: \"#3b82f6\", color: \"#fff\" } },\n { id: \"shell\", label: \"Launch shell\", variant: \"outline\" as const, style: undefined },\n ] as const;\n \n const SESSION_TOOL_COLORS: Record = {\n claude: \"#f97316\",\n codex: \"#3b82f6\",\n shell: \"#22c55e\",\n };\n \n export function LaneTerminalsPanel({ overrideLaneId }: { overrideLaneId?: string | null } = {}) {\n const navigate = useNavigate();\n const globalLaneId = useAppStore((s) => s.selectedLaneId);\n@@ -299,51 +299,51 @@ export function LaneTerminalsPanel({ overrideLaneId }: { overrideLaneId?: string\n >\n {launchTracked ? \"With context\" : \"Standalone\"}\n \n \n \n \n {sessions.length === 0 ? (\n

\n \n

\n ) : viewMode === \"tabs\" ? (\n focusSession(v)}\n className=\"flex min-h-0 min-w-0 flex-1 flex-col\"\n >\n \n {tabSessions.map((s) => {\n const profileColor = s.toolType ? SESSION_TOOL_COLORS[s.toolType] : undefined;\n const indicator = sessionIndicatorState({\n status: s.status,\n lastOutputPreview: s.lastOutputPreview,\n runtimeState: s.runtimeState\n });\n const dotClass = statusDotCls(indicator);\n- const dotSpin = !profileColor && indicator !== \"ended\";\n+ const dotSpin = !profileColor && indicator === \"running-active\";\n return (\n \n \n \n {sessionTabLabel(s)}\n {!s.tracked ? Standalone : null}\n {s.status === \"running\" && s.ptyId ? (\n {\n e.preventDefault();\n e.stopPropagation();\n closeSession(s);\n }}\n onKeyDown={(e) => {\n if (e.key === \"Enter\" || e.key === \" \") {\n e.preventDefault();\ndiff --git a/apps/desktop/src/renderer/components/lanes/LanesPage.tsx b/apps/desktop/src/renderer/components/lanes/LanesPage.tsx\nindex 4d7b3dafd6871e8d33104cbaee84c2875f0ee60e..9d587c49941fec99321af5869876aedac559d893 100644\n--- a/apps/desktop/src/renderer/components/lanes/LanesPage.tsx\n+++ b/apps/desktop/src/renderer/components/lanes/LanesPage.tsx\n@@ -1422,51 +1422,51 @@ export function LanesPage() {\n fontWeight: 700,\n letterSpacing: \"1px\",\n textTransform: \"uppercase\",\n borderRadius: 6,\n border: active ? `1px solid ${chip.color}60` : `1px solid ${COLORS.outlineBorder}`,\n background: active ? `${chip.color}20` : \"rgba(255,255,255,0.02)\",\n color: active ? chip.color : COLORS.textMuted,\n cursor: \"pointer\",\n whiteSpace: \"nowrap\",\n }}\n title={`${chip.count} lane${chip.count === 1 ? \"\" : \"s\"}`}\n >\n {chip.label}\n {chip.count}\n \n );\n })}\n \n \n {/* NEW LANE button + dropdown */}\n

\n \n {addLaneDropdownOpen ? (\n-

\n+

\n {\n setAddLaneDropdownOpen(false);\n prepareCreateDialog();\n }}\n >\n \n \n \n \n Create new lane\n Start from primary or stack from another lane.\n \n \n {\n setAddLaneDropdownOpen(false);\n setAttachName(\"\");\n setAttachPath(\"\");\n setAttachBusy(false);\n setAttachError(null);\ndiff --git a/apps/desktop/src/renderer/components/prs/CreatePrModal.tsx b/apps/desktop/src/renderer/components/prs/CreatePrModal.tsx\nindex 155a5076f2aec33d8973b85cbcb9db4872c6e77b..ad6b5083b1d5755e6f9f6059ae3f9c76cd9c4959 100644\n--- a/apps/desktop/src/renderer/components/prs/CreatePrModal.tsx\n+++ b/apps/desktop/src/renderer/components/prs/CreatePrModal.tsx\n@@ -704,51 +704,52 @@ export function CreatePrModal({\n if (!proposal) {\n setExecError(\"Run a simulation first to create a proposal.\");\n setBusy(false);\n return;\n }\n lastProgressLabel = \"Saving proposal\";\n setIntegrationProgress(\"Saving proposal...\");\n await window.ade.prs.updateProposal({\n proposalId: proposal.proposalId,\n title: integrationTitle || \"Integration workflow\",\n body: integrationBody,\n draft: integrationDraft,\n integrationLaneName: integrationName || `integration/${Date.now().toString(36)}`,\n });\n lastProgressLabel = \"Creating integration lane\";\n setIntegrationProgress(\"Creating integration lane...\");\n await window.ade.prs.createIntegrationLaneForProposal({\n proposalId: proposal.proposalId,\n });\n setIntegrationProgress(null);\n // No PR created — proposal saved for later commit from Integration tab\n setResults([]);\n setNumericStep(3);\n }\n } catch (err: unknown) {\n- const msg = err instanceof Error ? err.message : String(err);\n+ const rawMsg = err instanceof Error ? err.message : String(err);\n+ const msg = rawMsg.replace(/^Error invoking remote method '[^']+': (?:Error: )?/, \"\");\n if (mode === \"integration\" && lastProgressLabel) {\n setExecError(`Failed during \"${lastProgressLabel}\": ${msg}`);\n } else {\n setExecError(msg);\n }\n setIntegrationProgress(null);\n // Stay on current step so user can see the error and retry\n } finally {\n setBusy(false);\n }\n };\n \n const nonPrimaryLanes = React.useMemo(() => lanes.filter((l) => l.laneType !== \"primary\"), [lanes]);\n \n const toggleQueueLane = (laneId: string) => {\n setQueueLaneIds((prev) =>\n prev.includes(laneId) ? prev.filter((id) => id !== laneId) : [...prev, laneId]\n );\n };\n \n const handleQueueLaneDrop = React.useCallback((targetLaneId: string) => {\n if (!queueDragLaneId || queueDragLaneId === targetLaneId) {\n setQueueDragLaneId(null);\n return;\n }\n@@ -1687,50 +1688,56 @@ export function CreatePrModal({\n rows={6}\n style={textareaStyle}\n placeholder=\"PR description (markdown)...\"\n onFocus={(e) => { e.currentTarget.style.borderColor = C.accent; }}\n onBlur={(e) => { e.currentTarget.style.borderColor = C.borderSubtle; }}\n />\n

\n \n \n setNormalDraft(e.target.checked)}\n style={{ accentColor: C.accent }}\n />\n CREATE AS DRAFT\n \n+ \n \n )}\n \n {mode === \"queue\" && (\n <>\n

\n QUEUE ORDER\n

\n {queueLaneIds.map((laneId, idx) => {\n const lane = lanes.find((l) => l.id === laneId);\n return (\n

\n \n

\n This order will be used exactly when ADE creates the queue PRs.\n

\n \n \n setQueueDraft(e.target.checked)}\n style={{ accentColor: C.accent }}\n />\n CREATE AS DRAFTS\n \n+ \n \n )}\n \n {mode === \"integration\" && (\n <>\n {/* Summary of what's being integrated */}\n

\n INTEGRATION SUMMARY\n

\n {integrationSources.length} lanes into {integrationName || \"integration/...\"}\n

\n {proposal && (\n { e.currentTarget.style.borderColor = C.accent; }}\n onBlur={(e) => { e.currentTarget.style.borderColor = C.borderSubtle; }}\n />\n

\n \n \n setIntegrationDraft(e.target.checked)}\n style={{ accentColor: C.accent }}\n />\n CREATE AS DRAFT\n \n+ \n \n )}\n \n {draftError && (\n

Draft failed: {draftError}

\n )}\n \n {execError && (\n

{execError}

\n )}\n \n {integrationProgress && (\n

\n \n {integrationProgress}\n

\n )}\ndiff --git a/apps/desktop/src/renderer/index.css b/apps/desktop/src/renderer/index.css\nindex 6c32e34b34599ae4a99d49a3328d1061d444519b..858cfd13225bac3fa91be3fc779142a673cdea25 100644\n--- a/apps/desktop/src/renderer/index.css\n+++ b/apps/desktop/src/renderer/index.css\n@@ -1767,51 +1767,52 @@ button:active, [role=\"button\"]:active {\n \n @keyframes ade-gear-spin-reverse {\n 0% {\n transform: rotate(360deg);\n }\n 100% {\n transform: rotate(0deg);\n }\n }\n \n /* ═══════════════════════════════════════════════════════════\n Reduced Motion\n ═══════════════════════════════════════════════════════════ */\n \n @media (prefers-reduced-motion: reduce) {\n .ade-tab-bg,\n .ade-tab-bg::before,\n .ade-tab-bg::after {\n animation: none !important;\n }\n .ade-glow-pulse,\n .ade-glow-pulse-blue,\n .ade-glow-pulse-red,\n .ade-glow-pulse-amber,\n .ade-conflict-breathe,\n- .ade-float {\n+ .ade-float,\n+ .pull-btn-flash {\n animation: none !important;\n }\n /* Dropdown item stagger — instant appearance under reduced motion */\n [data-radix-dropdown-menu-content] [role=\"menuitem\"],\n [data-radix-select-content] [role=\"option\"] {\n animation: none !important;\n }\n }\n \n /* ═══════════════════════════════════════════════════════════\n Per-Tab Ambient Background System\n ═══════════════════════════════════════════════════════════ */\n \n .ade-tab-bg {\n opacity: 1;\n contain: strict;\n will-change: opacity;\n }\n \n [data-theme=\"dark\"] .ade-tab-bg,\n :root .ade-tab-bg { opacity: 0.55; }\n [data-theme=\"light\"] .ade-tab-bg { opacity: 0.3; }\n \n /* 1. Project — Organic Depth Field (purple #A78BFA) */\n .ade-tab-bg-project {\n@@ -2417,25 +2418,33 @@ button:active, [role=\"button\"]:active {\n width: 6px;\n height: 6px;\n border-radius: 50%;\n transition: opacity 300ms cubic-bezier(0.34, 1.56, 0.64, 1), transform 300ms cubic-bezier(0.34, 1.56, 0.64, 1);\n }\n .ade-status-dot-active {\n background: var(--shell-status-running);\n box-shadow: 0 0 6px 1px color-mix(in srgb, var(--shell-status-running) 30%, transparent);\n }\n .ade-status-dot-warning {\n background: var(--shell-status-attention);\n box-shadow: 0 0 6px 1px color-mix(in srgb, var(--shell-status-attention) 30%, transparent);\n }\n .ade-status-dot-error {\n background: var(--color-error);\n box-shadow: 0 0 6px 1px rgba(239, 68, 68, 0.3);\n }\n \n @keyframes ade-thinking-pulse {\n 0%, 100% { opacity: 0.3; }\n 50% { opacity: 1; }\n }\n .ade-thinking-pulse {\n animation: ade-thinking-pulse 1.2s ease-in-out infinite;\n }\n+\n+@keyframes pull-btn-flash {\n+ 0%, 100% { opacity: 1; }\n+ 50% { opacity: 0.4; }\n+}\n+.pull-btn-flash {\n+ animation: pull-btn-flash 2s ease-in-out infinite;\n+}\ndiff --git a/apps/desktop/src/renderer/lib/terminalAttention.test.ts b/apps/desktop/src/renderer/lib/terminalAttention.test.ts\nindex aa8686b2f1f24c4d65b3a731820d73ba0aa6a5c2..61092af3e9e24cfac16f76fd46c41705ead3ba95 100644\n--- a/apps/desktop/src/renderer/lib/terminalAttention.test.ts\n+++ b/apps/desktop/src/renderer/lib/terminalAttention.test.ts\n@@ -1,24 +1,56 @@\n import { describe, expect, it } from \"vitest\";\n-import { runningSessionNeedsAttention, sessionIndicatorState } from \"./terminalAttention\";\n+import { runningSessionNeedsAttention, sessionIndicatorState, sessionStatusDot } from \"./terminalAttention\";\n \n describe(\"terminalAttention\", () => {\n it(\"does not treat a plain shell prompt as awaiting user input\", () => {\n expect(runningSessionNeedsAttention(\"admin@Mac test-4-6a625aeb %\")).toBe(false);\n expect(\n sessionIndicatorState({\n status: \"running\",\n lastOutputPreview: \"admin@Mac test-4-6a625aeb %\",\n }),\n ).toBe(\"running-active\");\n });\n \n it(\"still detects explicit confirmation prompts\", () => {\n expect(runningSessionNeedsAttention(\"Confirm continue? (y/n)\")).toBe(true);\n expect(\n sessionIndicatorState({\n status: \"running\",\n lastOutputPreview: \"Confirm continue? (y/n)\",\n }),\n ).toBe(\"running-needs-attention\");\n });\n+\n+ describe(\"sessionStatusDot\", () => {\n+ it(\"returns a spinning emerald dot for a running active session\", () => {\n+ const dot = sessionStatusDot({\n+ status: \"running\",\n+ lastOutputPreview: \"building project...\",\n+ });\n+ expect(dot.spinning).toBe(true);\n+ expect(dot.cls).toContain(\"emerald\");\n+ expect(dot.label).toBe(\"Running\");\n+ });\n+\n+ it(\"returns a solid (non-spinning) amber dot for a running needs-attention session\", () => {\n+ const dot = sessionStatusDot({\n+ status: \"running\",\n+ lastOutputPreview: \"Confirm continue? (y/n)\",\n+ });\n+ expect(dot.spinning).toBe(false);\n+ expect(dot.cls).toContain(\"amber\");\n+ expect(dot.label).toBe(\"Awaiting input\");\n+ });\n+\n+ it(\"returns a solid red dot for an ended session\", () => {\n+ const dot = sessionStatusDot({\n+ status: \"completed\",\n+ lastOutputPreview: \"Process exited with code 0\",\n+ });\n+ expect(dot.spinning).toBe(false);\n+ expect(dot.cls).toContain(\"red\");\n+ expect(dot.label).toBe(\"Ended\");\n+ });\n+ });\n });\ndiff --git a/apps/desktop/src/renderer/lib/terminalAttention.ts b/apps/desktop/src/renderer/lib/terminalAttention.ts\nindex 13c4e59ac47a09fa923fa614cacb0cf00e527b83..91c4a5642339f29f29e5bafc392c87fb224b9f25 100644\n--- a/apps/desktop/src/renderer/lib/terminalAttention.ts\n+++ b/apps/desktop/src/renderer/lib/terminalAttention.ts\n@@ -98,51 +98,51 @@ export function sessionMatchesStatusFilter(\n runtimeState?: TerminalRuntimeState;\n },\n filter: SessionStatusFilter,\n ): boolean {\n if (filter === \"all\") return true;\n return sessionStatusBucket(args) === filter;\n }\n \n export type SessionStatusDot = {\n cls: string;\n spinning: boolean;\n label: string;\n };\n \n /** Map a session's indicator state to CSS classes for rendering a status dot. */\n export function sessionStatusDot(session: {\n status: TerminalSessionStatus;\n lastOutputPreview: string | null;\n runtimeState?: TerminalRuntimeState;\n }): SessionStatusDot {\n const indicator = sessionIndicatorState(session);\n if (indicator === \"running-active\") {\n return { cls: \"border-2 border-emerald-400 border-t-transparent bg-transparent\", spinning: true, label: \"Running\" };\n }\n if (indicator === \"running-needs-attention\") {\n- return { cls: \"border-2 border-amber-300 border-t-transparent bg-transparent\", spinning: true, label: \"Awaiting input\" };\n+ return { cls: \"bg-amber-300\", spinning: false, label: \"Awaiting input\" };\n }\n return { cls: \"bg-red-400\", spinning: false, label: \"Ended\" };\n }\n \n export function summarizeTerminalAttention(sessions: TerminalSessionSummary[]): TerminalAttentionSummary {\n let runningCount = 0;\n let activeCount = 0;\n let needsAttentionCount = 0;\n const byLane: Record = {};\n \n for (const session of sessions) {\n const indicator = sessionIndicatorState({\n status: session.status,\n lastOutputPreview: session.lastOutputPreview,\n runtimeState: session.runtimeState,\n });\n if (indicator === \"ended\") continue;\n const lane = byLane[session.laneId] ?? { runningCount: 0, activeCount: 0, needsAttentionCount: 0 };\n lane.runningCount += 1;\n runningCount += 1;\n if (indicator === \"running-needs-attention\") {\n lane.needsAttentionCount += 1;\n needsAttentionCount += 1;\n } else {\n lane.activeCount += 1;\ndiff --git a/apps/desktop/src/shared/ipc.ts b/apps/desktop/src/shared/ipc.ts\nindex 756e8ed1ed2a19cd77b5c77dad4d2adb9d256fbb..3e841345f94d02e998ef5b0ee5d0002ec5937f44 100644\n--- a/apps/desktop/src/shared/ipc.ts\n+++ b/apps/desktop/src/shared/ipc.ts\n@@ -79,50 +79,52 @@ export const IPC = {\n lanesProxyEvent: \"ade.lanes.proxy.event\",\n lanesOAuthGetStatus: \"ade.lanes.oauth.getStatus\",\n lanesOAuthUpdateConfig: \"ade.lanes.oauth.updateConfig\",\n lanesOAuthGenerateRedirectUris: \"ade.lanes.oauth.generateRedirectUris\",\n lanesOAuthEncodeState: \"ade.lanes.oauth.encodeState\",\n lanesOAuthDecodeState: \"ade.lanes.oauth.decodeState\",\n lanesOAuthListSessions: \"ade.lanes.oauth.listSessions\",\n lanesOAuthEvent: \"ade.lanes.oauth.event\",\n lanesDiagnosticsGetStatus: \"ade.lanes.diagnostics.getStatus\",\n lanesDiagnosticsGetLaneHealth: \"ade.lanes.diagnostics.getLaneHealth\",\n lanesDiagnosticsRunHealthCheck: \"ade.lanes.diagnostics.runHealthCheck\",\n lanesDiagnosticsRunFullCheck: \"ade.lanes.diagnostics.runFullCheck\",\n lanesDiagnosticsActivateFallback: \"ade.lanes.diagnostics.activateFallback\",\n lanesDiagnosticsDeactivateFallback: \"ade.lanes.diagnostics.deactivateFallback\",\n lanesDiagnosticsEvent: \"ade.lanes.diagnostics.event\",\n sessionsList: \"ade.sessions.list\",\n sessionsGet: \"ade.sessions.get\",\n sessionsUpdateMeta: \"ade.sessions.updateMeta\",\n sessionsReadTranscriptTail: \"ade.sessions.readTranscriptTail\",\n agentChatList: \"ade.agentChat.list\",\n agentChatGetSummary: \"ade.agentChat.getSummary\",\n agentChatCreate: \"ade.agentChat.create\",\n agentChatHandoff: \"ade.agentChat.handoff\",\n agentChatSend: \"ade.agentChat.send\",\n agentChatSteer: \"ade.agentChat.steer\",\n+ agentChatCancelSteer: \"ade.agentChat.cancelSteer\",\n+ agentChatEditSteer: \"ade.agentChat.editSteer\",\n agentChatInterrupt: \"ade.agentChat.interrupt\",\n agentChatResume: \"ade.agentChat.resume\",\n agentChatApprove: \"ade.agentChat.approve\",\n agentChatRespondToInput: \"ade.agentChat.respondToInput\",\n agentChatModels: \"ade.agentChat.models\",\n agentChatDispose: \"ade.agentChat.dispose\",\n agentChatEvent: \"ade.agentChat.event\",\n agentChatUpdateSession: \"ade.agentChat.updateSession\",\n agentChatWarmupModel: \"ade.agentChat.warmupModel\",\n agentChatSlashCommands: \"ade.agentChat.slashCommands\",\n agentChatFileSearch: \"ade.agentChat.fileSearch\",\n agentChatSaveTempAttachment: \"ade.agentChat.saveTempAttachment\",\n agentChatListSubagents: \"ade.agentChat.listSubagents\",\n agentChatGetSessionCapabilities: \"ade.agentChat.getSessionCapabilities\",\n computerUseGetSettings: \"ade.computerUse.getSettings\",\n computerUseListArtifacts: \"ade.computerUse.listArtifacts\",\n computerUseGetOwnerSnapshot: \"ade.computerUse.getOwnerSnapshot\",\n computerUseRouteArtifact: \"ade.computerUse.routeArtifact\",\n computerUseUpdateArtifactReview: \"ade.computerUse.updateArtifactReview\",\n computerUseEvent: \"ade.computerUse.event\",\n ptyCreate: \"ade.pty.create\",\n ptyWrite: \"ade.pty.write\",\n ptyResize: \"ade.pty.resize\",\n ptyDispose: \"ade.pty.dispose\",\n ptyData: \"ade.pty.data\",\n@@ -139,50 +141,51 @@ export const IPC = {\n filesChange: \"ade.files.change\",\n filesCreateFile: \"ade.files.createFile\",\n filesCreateDirectory: \"ade.files.createDirectory\",\n filesRename: \"ade.files.rename\",\n filesDelete: \"ade.files.delete\",\n filesQuickOpen: \"ade.files.quickOpen\",\n filesSearchText: \"ade.files.searchText\",\n gitStageFile: \"ade.git.stageFile\",\n gitStageAll: \"ade.git.stageAll\",\n gitUnstageFile: \"ade.git.unstageFile\",\n gitUnstageAll: \"ade.git.unstageAll\",\n gitDiscardFile: \"ade.git.discardFile\",\n gitRestoreStagedFile: \"ade.git.restoreStagedFile\",\n gitCommit: \"ade.git.commit\",\n gitGenerateCommitMessage: \"ade.git.generateCommitMessage\",\n gitListRecentCommits: \"ade.git.listRecentCommits\",\n gitListCommitFiles: \"ade.git.listCommitFiles\",\n gitGetCommitMessage: \"ade.git.getCommitMessage\",\n gitRevertCommit: \"ade.git.revertCommit\",\n gitCherryPickCommit: \"ade.git.cherryPickCommit\",\n gitStashPush: \"ade.git.stashPush\",\n gitStashList: \"ade.git.stashList\",\n gitStashApply: \"ade.git.stashApply\",\n gitStashPop: \"ade.git.stashPop\",\n gitStashDrop: \"ade.git.stashDrop\",\n+ gitStashClear: \"ade.git.stashClear\",\n gitFetch: \"ade.git.fetch\",\n gitPull: \"ade.git.pull\",\n gitGetSyncStatus: \"ade.git.getSyncStatus\",\n gitSync: \"ade.git.sync\",\n gitPush: \"ade.git.push\",\n gitGetConflictState: \"ade.git.getConflictState\",\n gitRebaseContinue: \"ade.git.rebaseContinue\",\n gitRebaseAbort: \"ade.git.rebaseAbort\",\n gitMergeContinue: \"ade.git.mergeContinue\",\n gitMergeAbort: \"ade.git.mergeAbort\",\n gitListBranches: \"ade.git.listBranches\",\n gitCheckoutBranch: \"ade.git.checkoutBranch\",\n conflictsGetLaneStatus: \"ade.conflicts.getLaneStatus\",\n conflictsListOverlaps: \"ade.conflicts.listOverlaps\",\n conflictsGetRiskMatrix: \"ade.conflicts.getRiskMatrix\",\n conflictsSimulateMerge: \"ade.conflicts.simulateMerge\",\n conflictsRunPrediction: \"ade.conflicts.runPrediction\",\n conflictsGetBatchAssessment: \"ade.conflicts.getBatchAssessment\",\n conflictsListProposals: \"ade.conflicts.listProposals\",\n conflictsPrepareProposal: \"ade.conflicts.prepareProposal\",\n conflictsRequestProposal: \"ade.conflicts.requestProposal\",\n conflictsApplyProposal: \"ade.conflicts.applyProposal\",\n conflictsUndoProposal: \"ade.conflicts.undoProposal\",\n conflictsRunExternalResolver: \"ade.conflicts.runExternalResolver\",\n conflictsListExternalResolverRuns: \"ade.conflicts.listExternalResolverRuns\",\ndiff --git a/apps/desktop/src/shared/types/chat.ts b/apps/desktop/src/shared/types/chat.ts\nindex 6f3ce145da86198e48da79fdb7d8c50e750cae4a..aa8aad761808351d8d0da2b1b2f4750ffb837b28 100644\n--- a/apps/desktop/src/shared/types/chat.ts\n+++ b/apps/desktop/src/shared/types/chat.ts\n@@ -558,50 +558,59 @@ export type AgentChatHandoffResult = {\n \n export type AgentChatListArgs = {\n laneId?: string;\n includeAutomation?: boolean;\n };\n \n export type AgentChatGetSummaryArgs = {\n sessionId: string;\n };\n \n export type AgentChatSendArgs = {\n sessionId: string;\n text: string;\n displayText?: string;\n attachments?: AgentChatFileRef[];\n reasoningEffort?: string | null;\n executionMode?: AgentChatExecutionMode | null;\n interactionMode?: AgentChatInteractionMode | null;\n };\n \n export type AgentChatSteerArgs = {\n sessionId: string;\n text: string;\n };\n \n+export type AgentChatCancelSteerArgs = {\n+ sessionId: string;\n+};\n+\n+export type AgentChatEditSteerArgs = {\n+ sessionId: string;\n+ text: string;\n+};\n+\n export type AgentChatInterruptArgs = {\n sessionId: string;\n };\n \n export type AgentChatResumeArgs = {\n sessionId: string;\n };\n \n export type AgentChatApproveArgs = {\n sessionId: string;\n itemId: string;\n decision: AgentChatApprovalDecision;\n responseText?: string | null;\n };\n \n export type AgentChatRespondToInputArgs = {\n sessionId: string;\n itemId: string;\n decision?: AgentChatApprovalDecision;\n answers?: Record;\n responseText?: string | null;\n };\n \n export type AgentChatModelsArgs = {\n provider: AgentChatProvider;\ndiff --git a/docs/architecture/AI_INTEGRATION.md b/docs/architecture/AI_INTEGRATION.md\nindex 703243e3f75046f99e2f80fa7efcda955298d3a2..e408b9605da099cc28c3359fc353b6e4f547629e 100644\n--- a/docs/architecture/AI_INTEGRATION.md\n+++ b/docs/architecture/AI_INTEGRATION.md\n@@ -1460,50 +1460,52 @@ These tasks can complete in one pass, but are still tracked as runtimes so they\n This boundary is critical: SDK calls are ADE's internal tool; CLI sessions are the user's tool.\n \n ### Agent Chat Service (Phase 1.5)\n \n The Agent Chat Service provides a native, interactive chat interface inside ADE — an alternative to using CLI terminals for working with Codex, Claude, and unified API/local model runtimes. It is a **provider-agnostic abstraction** that lets users chat with CLI or non-CLI models using the same UI.\n \n > **External reference**: The Codex App Server protocol specification is at https://developers.openai.com/codex/app-server — this is the canonical reference for the CodexChatBackend implementation.\n \n #### Why Agent Chat?\n \n CLI terminals are powerful but opaque. The chat interface provides:\n \n - **Structured item display**: File changes as inline diffs, command execution with live output, plans with step status — not raw terminal output.\n - **Approval flow**: Accept/decline tool use with full context, not a yes/no prompt in a terminal.\n - **Steering**: Inject instructions into an active turn without starting a new conversation.\n - **Session persistence**: Resume conversations with full context, not just a command string.\n - **Provider switching**: Same UI across providers/models. When switching model families mid-session, ADE forks a new chat session under the selected runtime/provider.\n \n #### AgentChatService Interface\n \n ```typescript\n interface AgentChatService {\n createSession(laneId: string, provider: \"codex\" | \"claude\" | \"unified\", model: string, modelId: string, opts?: CreateSessionOpts): Promise;\n sendMessage(sessionId: string, text: string, attachments?: FileRef[]): AsyncIterable;\n steer(sessionId: string, text: string): Promise;\n+ cancelSteer(sessionId: string): Promise;\n+ editSteer(sessionId: string, text: string): Promise;\n interrupt(sessionId: string): Promise;\n resumeSession(sessionId: string): Promise;\n listSessions(laneId?: string): Promise;\n approveToolUse(sessionId: string, itemId: string, decision: ApprovalDecision): Promise;\n getAvailableModels(provider: \"codex\" | \"claude\" | \"unified\"): Promise;\n dispose(sessionId: string): Promise;\n }\n \n interface ChatSession {\n id: string;\n laneId: string;\n provider: \"codex\" | \"claude\" | \"unified\";\n model: string; // Human-readable display name\n modelId: string; // Registry model ID (required)\n // Provider-native permission controls (no unified permissionMode)\n claudePermissionMode?: \"default\" | \"plan\" | \"acceptEdits\" | \"bypassPermissions\";\n codexApprovalPolicy?: \"untrusted\" | \"on-request\" | \"on-failure\" | \"never\";\n codexSandbox?: \"read-only\" | \"workspace-write\" | \"danger-full-access\";\n codexConfigSource?: \"flags\" | \"config-toml\";\n unifiedPermissionMode?: \"plan\" | \"edit\" | \"full-auto\";\n status: \"active\" | \"idle\" | \"ended\";\n threadId?: string; // Codex: app-server thread ID\n createdAt: string;\n lastActivityAt: string;\n }\n@@ -1556,50 +1558,52 @@ import readline from \"node:readline\";\n const proc = spawn(\"codex\", [\"app-server\"], {\n stdio: [\"pipe\", \"pipe\", \"inherit\"], // stdin: pipe, stdout: pipe, stderr: inherit\n });\n \n // JSONL reader\n const rl = readline.createInterface({ input: proc.stdout });\n \n // Send JSON-RPC messages\n const send = (msg: object) => proc.stdin.write(JSON.stringify(msg) + \"\\n\");\n \n // Initialize handshake (required before any other messages)\n send({ method: \"initialize\", id: 0, params: {\n clientInfo: { name: \"ade\", title: \"ADE\", version: ADE_VERSION },\n capabilities: { experimentalApi: true }\n }});\n send({ method: \"initialized\", params: {} });\n ```\n \n **Protocol mapping to AgentChatService**:\n \n | AgentChatService method | Codex App Server JSON-RPC | Notes |\n |---|---|---|\n | `createSession()` | `thread/start` | Params: `model`, `cwd` (lane worktree), `approvalPolicy`, `sandbox` |\n | `sendMessage()` | `turn/start` | Input array: `[{ type: \"text\", text }, ...attachments]` |\n | `steer()` | `turn/steer` | Appends to in-flight turn; cannot change model/sandbox |\n+| `cancelSteer()` | N/A | Cancel a pending steer before it is delivered |\n+| `editSteer()` | N/A | Replace pending steer text before it is delivered |\n | `interrupt()` | `turn/interrupt` | Turn completes with `status: \"interrupted\"`. Idempotent — second call is a no-op. Claude/unified runtimes emit `subagent_result` \"stopped\" for active subagents. |\n | `resumeSession()` | `thread/resume` | Params: `threadId`, optional `personality` |\n | `listSessions()` | `thread/list` | Filter by `cwd` to scope to lane |\n | `approveToolUse()` | Response to `requestApproval` | Payload: `accept`/`acceptForSession`/`decline`/`cancel` |\n | `getAvailableModels()` | `model/list` | Returns models with `displayName`, `isDefault`, `reasoningEffort` |\n | `dispose()` | Process cleanup | Close stdin, terminate child process |\n \n **Notification → ChatEvent mapping**:\n \n | Codex Notification | ChatEvent Type | Notes |\n |---|---|---|\n | `item/agentMessage/delta` | `text` | Append delta to current message |\n | `item/started` (commandExecution) | `command` | Start tracking command output |\n | `item/commandExecution/outputDelta` | `command` | Append to command output |\n | `item/completed` (commandExecution) | `command` | Set exit code and status |\n | `item/started` (fileChange) | `file_change` | File path, kind, diff |\n | `item/commandExecution/requestApproval` | `approval_request` | Kind: \"command\" |\n | `item/fileChange/requestApproval` | `approval_request` | Kind: \"file_change\" |\n | `turn/plan/updated` | `plan` | Steps with status |\n | `item/reasoning/summaryTextDelta` | `reasoning` | Append to summary |\n | `turn/started` | `status` | `turnStatus: \"started\"` |\n | `turn/completed` | `done` or `status` | Map `status` field |\n \n **Sandbox policies** (configurable per session):\n \ndiff --git a/docs/architecture/DATA_MODEL.md b/docs/architecture/DATA_MODEL.md\nindex e7f1f0b7d714add37f7c35f11f68056a1f96511f..ce665c807f1e4181a0400a38749168ce12fc3015 100644\n--- a/docs/architecture/DATA_MODEL.md\n+++ b/docs/architecture/DATA_MODEL.md\n@@ -267,51 +267,51 @@ Computed after session end by the pack service. Contains the diff statistics (fi\n ```sql\n CREATE TABLE IF NOT EXISTS operations (\n id TEXT PRIMARY KEY,\n project_id TEXT NOT NULL,\n lane_id TEXT,\n kind TEXT NOT NULL, -- operation type identifier\n started_at TEXT NOT NULL,\n ended_at TEXT,\n status TEXT NOT NULL, -- 'running' | 'succeeded' | 'failed' | 'canceled'\n pre_head_sha TEXT,\n post_head_sha TEXT,\n metadata_json TEXT,\n FOREIGN KEY(project_id) REFERENCES projects(id),\n FOREIGN KEY(lane_id) REFERENCES lanes(id)\n );\n CREATE INDEX IF NOT EXISTS idx_operations_project_started ON operations(project_id, started_at);\n CREATE INDEX IF NOT EXISTS idx_operations_lane_started ON operations(lane_id, started_at);\n CREATE INDEX IF NOT EXISTS idx_operations_kind ON operations(kind);\n ```\n \n Records every significant operation (git commands, pack updates, etc.) with timing, status, and before/after HEAD SHAs. The `kind` field identifies the operation type (e.g., `git_commit`, `git_push`, `pack_update_lane`). The `metadata_json` column stores operation-specific details.\n \n Known `kind` values:\n - `git_stage`, `git_unstage`, `git_discard`, `git_restore_staged`\n - `git_commit`, `git_commit_amend`, `git_revert`, `git_cherry_pick`\n-- `git_stash_push`, `git_stash_apply`, `git_stash_pop`, `git_stash_drop`\n+- `git_stash_push`, `git_stash_apply`, `git_stash_pop`, `git_stash_drop`, `git_stash_clear`\n - `git_fetch`, `git_sync_merge`, `git_sync_rebase`\n - `git_push`, `git_push_force_with_lease`\n - `pack_update_lane`, `pack_update_project`\n \n #### Process Definitions\n \n ```sql\n CREATE TABLE IF NOT EXISTS process_definitions (\n id TEXT PRIMARY KEY,\n project_id TEXT NOT NULL,\n key TEXT NOT NULL,\n name TEXT NOT NULL,\n command_json TEXT NOT NULL, -- JSON array of command parts\n cwd TEXT NOT NULL,\n env_json TEXT NOT NULL, -- JSON object of env vars\n autostart INTEGER NOT NULL,\n restart_policy TEXT NOT NULL, -- 'never' | 'on_crash'\n graceful_shutdown_ms INTEGER NOT NULL,\n depends_on_json TEXT NOT NULL, -- JSON array of process keys\n readiness_json TEXT NOT NULL, -- JSON readiness config\n updated_at TEXT NOT NULL,\n UNIQUE(project_id, key),\n FOREIGN KEY(project_id) REFERENCES projects(id)\n );\n ```\ndiff --git a/docs/architecture/GIT_ENGINE.md b/docs/architecture/GIT_ENGINE.md\nindex da65add2b79c1b6335dd135c6c215d30a4b34f4f..4beafda733d4971662d5854493f549a5a6daa122 100644\n--- a/docs/architecture/GIT_ENGINE.md\n+++ b/docs/architecture/GIT_ENGINE.md\n@@ -156,50 +156,51 @@ The `gitOperationsService.ts` module provides the full set of git operations exp\n | `stageFile` | `git add -- ` | Stage a file for commit |\n | `unstageFile` | `git restore --staged -- ` | Remove a file from staging |\n | `discardFile` | `git restore --worktree -- ` or `git clean -f -- ` | Discard working tree changes |\n | `restoreStagedFile` | `git restore --staged --worktree --source=HEAD -- ` | Fully restore a file to HEAD state |\n \n The `discardFile` operation checks whether the file is untracked (via `git status --porcelain=v1`) and uses `git clean -f` for untracked files or `git restore --worktree` for tracked files.\n \n #### Commit Operations\n \n | Operation | Git Command | Purpose |\n |-----------|------------|---------|\n | `commit` | `git commit -m ` | Create a new commit |\n | `commit` (amend) | `git commit --amend -m ` | Amend the most recent commit |\n | `revertCommit` | `git revert --no-edit ` | Create a revert commit |\n | `cherryPickCommit` | `git cherry-pick ` | Cherry-pick a commit into the current branch |\n \n #### Stash Operations\n \n | Operation | Git Command | Purpose |\n |-----------|------------|---------|\n | `stashPush` | `git stash push [-u] [-m ]` | Save working changes to stash |\n | `listStashes` | `git stash list --date=iso-strict --format=...` | List all stash entries |\n | `stashApply` | `git stash apply ` | Apply stash without removing it |\n | `stashPop` | `git stash pop ` | Apply and remove stash |\n | `stashDrop` | `git stash drop ` | Remove stash without applying |\n+| `stashClear` | `git stash clear` | Remove all stash entries |\n \n #### Sync Operations\n \n | Operation | Git Command | Purpose |\n |-----------|------------|---------|\n | `fetch` | `git fetch --prune` | Fetch remote refs, prune deleted |\n | `sync` (merge) | `git fetch --prune` then `git merge --no-edit ` | Sync with upstream via merge |\n | `sync` (rebase) | `git fetch --prune` then `git rebase ` | Sync with upstream via rebase |\n | `push` | `git push [-u origin ] [--force-with-lease]` | Push branch to remote |\n \n #### Query Operations\n \n | Operation | Git Command | Purpose |\n |-----------|------------|---------|\n | `listRecentCommits` | `git log -n --pretty=format:...` | List recent commits with metadata |\n \n The `listRecentCommits` operation uses a unit separator (`\\x1f`) delimited format to parse commit fields: full SHA, short SHA, author name, authored date (ISO), and subject line.\n \n ### Operation Tracking\n \n Every git mutation is wrapped by `runLaneOperation()`:\n \n ```typescript\n const runLaneOperation = async ({\n laneId, kind, reason, metadata, fn\n@@ -308,46 +309,46 @@ Renderer: user clicks \"Commit\"\n --> gitService.commit()\n --> runLaneOperation()\n --> operationService.start()\n --> git commit -m \"message\"\n --> operationService.finish()\n --> onHeadChanged() (if HEAD moved)\n --> jobEngine.onHeadChanged()\n --> packService.refreshLanePack() (compatibility artifact refresh)\n --> packService.refreshProjectPack()\n <-- GitActionResult { operationId, preHeadSha, postHeadSha }\n ```\n \n Live orchestrator/MCP context exports do not wait for those compatibility refreshes; they are generated directly from current local state.\n \n ---\n \n ## Implementation Status\n \n ### Completed\n \n - Git command execution utilities (`runGit`, `runGitOrThrow`) with timeout support\n - Lane worktree creation and deletion\n - Lane status computation (dirty, ahead, behind)\n - File operations: stage, unstage, discard, restore staged\n - Commit operations: commit, amend, revert, cherry-pick\n-- Stash operations: push, list, apply, pop, drop\n+- Stash operations: push, list, apply, pop, drop, clear\n - Sync operations: fetch, merge sync, rebase sync\n - Push with automatic upstream setup and force-with-lease option\n - Recent commits listing with parsed metadata\n - Operation tracking wrapper with pre/post HEAD SHA capture\n - Path validation with traversal prevention\n - HEAD change event propagation to job engine\n - Conflict prediction via dry-merge simulation using `git merge-tree --write-tree --messages` (Phase 5, no old merge-tree fallback path)\n - Pairwise lane conflict detection across all active lanes (Phase 5)\n - Stack operations: parent-child lane relationships with rebase propagation (Phase 4)\n - Primary lane support: main repo directory represented as a lane (Phase 7)\n - Attached worktree support: link pre-existing worktrees to ADE (Phase 7)\n \n ### Planned (Not Yet Started)\n \n - **Branch operations**: Create, delete, checkout, rename branches\n - **Interactive rebase support**: Reorder, squash, fixup commits\n - **Merge conflict resolution UI**: Inline conflict markers with accept/reject actions\n - **Git hooks integration**: Surface pre-commit hook failures in the UI\n - **Partial staging**: Stage individual hunks within a file\n - **Blame integration**: Surface authorship information in diff views\ndiff --git a/docs/architecture/MULTI_DEVICE_SYNC.md b/docs/architecture/MULTI_DEVICE_SYNC.md\nindex 9d29a8578f742a08919679eb91d9f4c13b747163..e2bb87e584468e3b2a215ab487a8d4337e30905a 100644\n--- a/docs/architecture/MULTI_DEVICE_SYNC.md\n+++ b/docs/architecture/MULTI_DEVICE_SYNC.md\n@@ -256,51 +256,51 @@ Rule for future workstreams:\n - if a new ADE feature creates low-churn, human-authored shared config or identity, it can define a tracked representation under `.ade/`\n - if a feature only matters for a connected live host/controller session, keep it in ADE sync\n - if a feature is machine-bound, keep it local\n \n ### Changeset Flow Diagram\n \n ```\n ┌──────────┐ changesets ┌──────────┐ changesets ┌──────────┐\n │ iPhone │ ◄─────────────► │ Host │ ◄─────────────► │ Mac #2 │\n │ (controller) │ WebSocket │ (Mac #1) │ WebSocket │ (controller) │\n └──────────┘ └──────────┘ └──────────┘\n │\n │ code sync (git push/pull)\n ▼\n ┌──────────┐\n │ Mac #2 │\n │ worktree │\n └──────────┘\n ```\n \n ### Command Routing\n \n - **State-only operations** (create lane metadata, update settings): write locally, cr-sqlite syncs.\n - **Execution operations** (create worktree, run terminal command, create PR): send command to host via WebSocket. Host executes, state changes sync back.\n \n-The remote command service (`syncRemoteCommandService`) registers named actions that controllers can invoke. Agent chat commands include `chat.create`, `chat.send`, `chat.interrupt`, `chat.steer`, `chat.approve`, `chat.respondToInput`, `chat.resume`, `chat.updateSession`, `chat.dispose`, and `chat.models`. All chat commands are viewer-allowed and queueable, enabling full remote chat session control from companion devices.\n+The remote command service (`syncRemoteCommandService`) registers named actions that controllers can invoke. Agent chat commands include `chat.create`, `chat.send`, `chat.interrupt`, `chat.steer`, `chat.cancelSteer`, `chat.editSteer`, `chat.approve`, `chat.respondToInput`, `chat.resume`, `chat.updateSession`, `chat.dispose`, and `chat.models`. All chat commands are viewer-allowed and queueable, enabling full remote chat session control from companion devices.\n \n ---\n \n ## Security Model\n \n ### Device Pairing\n \n - W3 uses a shared machine-local bootstrap token stored at `.ade/secrets/sync-bootstrap-token`.\n - The current host surfaces that token in Settings > Sync for manual desktop-to-desktop connection.\n - QR pairing, per-device secrets, OS keychain storage, and secure revoke remain W4 work.\n - W3 should be treated as trusted-LAN-only scaffolding. The host currently listens on all interfaces and does not yet add W4 hardening such as per-device secrets, revocation, tighter network policy, or transport upgrades.\n \n ### Transport Security\n \n - WebSocket connections are authenticated with the pairing token on every connection.\n - Tailscale connections use WireGuard encryption (peer-to-peer, no cloud relay).\n - LAN connections rely on pairing token validation. TLS is not required for localhost/LAN but can be added.\n - In W3, this is still bootstrap-token auth rather than final pairing auth. Use it only on networks you trust. W4 is responsible for replacing this with secure pairing, revocation, and better network posture.\n \n ### Secret Isolation\n \n - `.ade/local.secret.yaml` (API keys, external MCP configs) is machine-specific and never syncs.\n - Linear credentials, GitHub tokens, and provider API keys remain on the host only.\n - Each device stores its own pairing secret in its OS keychain.\n \ndiff --git a/docs/features/LANES.md b/docs/features/LANES.md\nindex d5feccaa658097d55349365b13efd52f56c89625..856f5e754e067923e4cd5802d62e232d994e7a5f 100644\n--- a/docs/features/LANES.md\n+++ b/docs/features/LANES.md\n@@ -121,51 +121,51 @@ The left pane is a scrollable list of all active lanes for the current project.\n - Branch name (secondary text, if different from lane name)\n - Status badges:\n - Dirty indicator (filled circle if uncommitted changes exist)\n - Ahead count (up arrow + number, e.g., `↑2`)\n - Behind count (down arrow + number, e.g., `↓1`)\n - Timestamp (last activity or creation time)\n \n **Interactions**:\n - Single click: Select lane, show in center pane.\n - Shift+click: Multi-select lanes (for bulk operations like archive/delete).\n - Right-click: Context menu with Rename, Archive, Delete, Open Folder, Create Child Lane.\n - Double-click: Open lane in a new tab in the center pane.\n \n **Stack graph** (bottom of left pane): A lightweight visual representation of parent-child lane relationships (including clear connections back to the Primary lane). A one-click “Open canvas” action jumps to the full Workspace Graph for deeper exploration.\n \n **Create Lane button**: Opens a dropdown menu offering \"Create new lane\" or \"Add existing worktree as lane\". Both options open a dialog built on the shared `LaneDialogShell` component. The create dialog lets the user choose a name, optional template, and starting point (from primary with a base branch selector, or as a child of an existing lane). The attach dialog accepts a lane name, an absolute worktree path, and an optional description.\n \n ### Center Pane — Lane Detail Area\n \n The center pane is the main working area. The LanesPage uses `PaneTilingLayout` for a resizable pane structure with persisted sizes.\n \n **Tab bar**: Multiple lanes can be open simultaneously as tabs. Each tab shows the lane name and a close button. Tabs are reorderable by drag-and-drop. Primary lanes display a home icon.\n \n **Sub-panes** within the center area (via PaneTilingLayout):\n - **Diff pane** (`LaneDiffPane`): Git diff viewer with Monaco side-by-side diffs, per-file stage/unstage/discard, commit diff viewing\n-- **Git actions pane** (`LaneGitActionsPane`): Commit, stash, fetch, sync (merge/rebase), push operations with recent commits list, rebase button for stacked lanes\n+- **Git actions pane** (`LaneGitActionsPane`): Commit, stash, fetch, sync (merge/rebase), push operations with recent commits list, rebase button for stacked lanes. The Pull button flashes and shows a behind-count badge when the lane is behind its upstream. Stash buttons use user-friendly labels (Save Changes, Restore, Copy to Worktree, Delete) with inline explanations. The Save Changes button is disabled when there are no working tree changes. A Clear All button (with numeric confirmation) appears when stashes exist.\n - **Terminals pane** (`LaneTerminalsPanel`): Embedded terminal sessions with tab/tiling views, quick-launch profiles, session delta cards\n - **Work pane** (`LaneWorkPane`): Embedded terminal sessions and agent chat view (terminal/chat toggle)\n - **Stack pane** (`LaneStackPane`): Stack chain visualization and management\n - **Inspector pane** (`LaneInspectorPane`): Packs, PR, and conflict management\n \n **Terminals pane (signal-first)**:\n - Default view prioritizes **running** sessions. Ended/closed sessions are collapsed behind a toggle.\n - Ended sessions show a one-line deterministic summary (for review) and only reveal transcript tails when explicitly expanded.\n - **Tab vs Tiling toggle**: Switch between single-session tab view and multi-terminal tiling grid (`TilingLayout` component with recursive binary splits).\n - **Quick launch profiles**: One-click launch for Claude Code, Codex, or plain Shell. Profiles are configurable via terminal profiles settings.\n - **Context toggle**: Launch with or without context tracking (tracked/untracked sessions).\n - **Session delta card**: Displayed for ended sessions showing files changed, insertions/deletions, touched files, and failure lines.\n - A lightweight \"Open in Terminals tab\" action jumps to the global Terminals view with the lane filter pre-applied.\n \n **Agent Chat View (Phase 1.5)**: The Work Pane includes a view toggle between two modes:\n \n - **Terminal view** (default): The existing `LaneTerminalsPanel` showing PTY sessions with xterm.js.\n - **Chat view**: The `AgentChatPane` providing a rich conversational interface for working with Codex or Claude.\n \n The Chat view layout:\n \n ```\n +-----------------------------------------------+\n | [Terminal View] [Chat View] <- toggle |\n +-----------------------------------------------+\n@@ -206,92 +206,92 @@ The diff panel is the primary interaction surface for git operations within a la\n **Diff sections**:\n - **Unstaged Changes**: Files modified in the working tree but not yet staged.\n - **Staged Changes**: Files added to the git index, ready for commit.\n \n Each section shows a file list with change type indicators:\n | Indicator | Meaning |\n |-----------|---------|\n | `M` | Modified |\n | `A` | Added (new file) |\n | `D` | Deleted |\n | `R` | Renamed |\n | `U` | Unmerged (conflict) |\n \n **File interactions**:\n - Click a file to open the Monaco side-by-side diff viewer.\n - Quick edit: Modify unstaged files directly within the diff view.\n - Per-file buttons: Stage, Unstage, Discard (with confirmation for destructive actions).\n - Bulk buttons: Stage All, Unstage All.\n \n **Commit controls**:\n - Commit message input (multiline textarea)\n - Commit button (disabled when no staged changes or empty message)\n - Amend checkbox (future)\n \n **Advanced git controls**:\n-- **Stash**: Push (with optional message), Pop, Apply, Drop, List stashes\n+- **Stash**: Save Changes (push with optional message), Restore (pop), Copy to Worktree (apply), Delete (drop), Clear All (with numeric confirmation), List stashes\n - **Fetch**: Fetch from remote (all refs or specific)\n - **Sync**: Merge or Rebase with upstream/base ref\n - **Push**: Push to remote, with force-with-lease option\n - **Recent commits**: Scrollable list of recent commits on this branch\n - **Revert commit**: Create a revert commit for a selected commit\n - **Cherry-pick commit**: Apply a commit from another branch\n \n ### Inspector Sidebar\n \n The inspector is a collapsible sidebar on the right edge of the Lanes tab. It provides supplementary metadata and quick actions for the currently selected lane.\n \n **Contents**:\n - **Lane metadata**: Created date, branch name, worktree path, base ref, lane type.\n - **Quick actions**: Open in Finder/Explorer, Copy worktree path, Open terminal here.\n - **Session list**: All terminal sessions associated with this lane, with status indicators.\n - **Pack freshness indicator**: Shows whether packs are up-to-date or stale for this lane.\n - **Health indicator**: `LaneHealthDot` shows traffic-light health status (green/yellow/red) for lanes with runtime isolation active.\n \n ### Lane Lifecycle\n \n 1. **Create**: User provides a name. ADE creates a git branch (from the selected base ref) and a worktree directory (at `.ade/worktrees//`). The lane appears in the list with `active` status.\n \n 2. **Create from unstaged**: When a lane has unstaged or untracked changes that belong in a separate branch, the user can create a child lane from those changes. ADE stashes the unstaged work in the source lane, creates a new child lane from the source's HEAD, applies the stash into the new lane, and cleans up. If the stash apply fails, ADE rolls back: the created lane is removed and the stash is restored to the source. The UI exposes this as the \"Create New Lane with Current Changes\" button in the git actions pane. Preconditions: the source lane must have no staged changes and no in-progress merge or rebase.\n \n 3. **Work**: The user opens terminals in the lane, makes code changes, stages and commits, and pushes to the remote. All activity is tracked via sessions and operations.\n \n 4. **Archive**: The lane is hidden from the active list but its worktree and branch are preserved. Useful for lanes that are paused but not finished. Archived lanes can be unarchived at any time.\n \n 5. **Delete**: The worktree directory is removed from disk. The user is prompted whether to also delete the git branch. Session records and history are retained for audit purposes.\n \n ---\n \n ## Technical Implementation\n \n ### Services\n \n | Service | Responsibility |\n |---------|---------------|\n | `laneService` | CRUD operations for lanes. Creates/removes worktrees via git. Computes lane status by aggregating dirty state, ahead/behind, and other signals. Manages lane metadata in the database. Supports primary, worktree, and attached lane types. Provides rebase (recursive rebase with remote tracking branch resolution for primary parents, dirty-worktree guard), reparent, stack chain, appearance management, and `createFromUnstaged` (stash-based unstaged change rescue to a new child lane with automatic rollback on failure). |\n | `rebaseSuggestionService` | Monitors stacked lanes for parent-advanced state. Generates rebase suggestions with dismiss/defer lifecycle. Emits real-time suggestion events to the renderer. |\n-| `gitService` | All git operations: stage, unstage, discard, commit, stash, fetch, sync (merge/rebase), push, conflict state detection (merge/rebase in-progress, continue, abort). Operates on a specified worktree path. Returns structured results with success/failure and output. |\n+| `gitService` | All git operations: stage, unstage, discard, commit, stash (push, list, apply, pop, drop, clear), fetch, sync (merge/rebase), push, conflict state detection (merge/rebase in-progress, continue, abort). Operates on a specified worktree path. Returns structured results with success/failure and output. |\n | `diffService` | Computes working tree diffs (unstaged changes) and index diffs (staged changes). Per-file diff content for the Monaco viewer. Handles binary file detection and large file truncation. |\n | `operationService` | Records all git operations with before/after SHA transitions. Provides an audit trail for every action taken in a lane. Used by the History tab. |\n | `laneProxyService` | Per-lane hostname reverse proxy. Routes HTTP traffic by Host header to the correct lane's dev server. Manages proxy lifecycle, route registration, and preview URL generation. |\n | `oauthRedirectService` | OAuth callback routing for multi-lane environments. Supports state-parameter and hostname-based routing strategies. Tracks OAuth sessions and provides setup assistant helpers. |\n | `runtimeDiagnosticsService` | Continuous health monitoring for lane runtime isolation. Runs per-lane health checks (process, port, proxy), manages fallback mode, and provides actionable remediation. |\n \n ### IPC Channels\n \n **Lane management**:\n \n | Channel | Signature | Description |\n |---------|-----------|-------------|\n | `ade.lanes.list` | `(args: { projectId: string }) => LaneSummary[]` | List all active lanes for a project |\n | `ade.lanes.create` | `(args: { projectId: string, name: string, baseRef?: string, parentLaneId?: string }) => LaneSummary` | Create a new lane with branch and worktree |\n | `ade.lanes.rename` | `(args: { laneId: string, newName: string }) => void` | Rename a lane (does not rename branch) |\n | `ade.lanes.archive` | `(args: { laneId: string }) => void` | Archive a lane (hide from active list) |\n | `ade.lanes.delete` | `(args: DeleteLaneArgs) => void` | Delete lane, remove worktree, optionally delete branch |\n | `ade.lanes.openFolder` | `(args: { laneId: string }) => void` | Open worktree directory in Finder/Explorer |\n | `ade.lanes.createChild` | `(args: CreateChildLaneArgs) => LaneSummary` | Create a child lane (for stacking) |\n | `ade.lanes.createFromUnstaged` | `(args: CreateLaneFromUnstagedArgs) => LaneSummary` | Create a child lane from a source lane's unstaged changes, moving them to the new lane |\n | `ade.lanes.importBranch` | `(args: { branchRef: string }) => LaneSummary` | Import an existing branch as a lane |\n | `ade.lanes.attach` | `(args: AttachLaneArgs) => LaneSummary` | Attach an existing worktree directory as a lane |\n | `ade.lanes.reparent` | `(args: ReparentLaneArgs) => ReparentLaneResult` | Change a lane's parent (reparent in stack) |\n | `ade.lanes.updateAppearance` | `(args: UpdateLaneAppearanceArgs) => void` | Update lane color, icon, or tags |\n | `ade.lanes.getStackChain` | `(args: { laneId: string }) => StackChainItem[]` | Get the full stack chain for a lane |\n@@ -308,50 +308,51 @@ The inspector is a collapsible sidebar on the right edge of the Lanes tab. It pr\n | `ade.lanes.rebaseSuggestions.event` | Event stream | Emits `rebase-suggestions-updated` when suggestions change |\n \n **Diff operations**:\n \n | Channel | Signature | Description |\n |---------|-----------|-------------|\n | `ade.diff.getChanges` | `(args: { worktreePath: string }) => DiffChanges` | Get unstaged and staged file change lists |\n | `ade.diff.getFile` | `(args: { worktreePath: string, filePath: string, staged: boolean }) => FileDiff` | Get detailed diff content for one file |\n \n **Git operations**:\n \n | Channel | Signature | Description |\n |---------|-----------|-------------|\n | `ade.git.stageFile` | `(args: { worktreePath: string, filePath: string }) => GitActionResult` | Stage a specific file |\n | `ade.git.unstageFile` | `(args: { worktreePath: string, filePath: string }) => GitActionResult` | Unstage a specific file |\n | `ade.git.discardFile` | `(args: { worktreePath: string, filePath: string }) => GitActionResult` | Discard changes to a specific file |\n | `ade.git.restoreStagedFile` | `(args: { worktreePath: string, filePath: string }) => GitActionResult` | Restore a staged file to its HEAD state |\n | `ade.git.commit` | `(args: { worktreePath: string, message: string }) => GitActionResult` | Create a commit with staged changes |\n | `ade.git.listRecentCommits` | `(args: { worktreePath: string, count?: number }) => GitCommitSummary[]` | List recent commits on current branch |\n | `ade.git.revertCommit` | `(args: { worktreePath: string, sha: string }) => GitActionResult` | Revert a specific commit |\n | `ade.git.cherryPickCommit` | `(args: { worktreePath: string, sha: string }) => GitActionResult` | Cherry-pick a commit from another branch |\n | `ade.git.stashPush` | `(args: { worktreePath: string, message?: string }) => GitActionResult` | Stash current changes |\n | `ade.git.stashApply` | `(args: { worktreePath: string, index?: number }) => GitActionResult` | Apply a stash without removing it |\n | `ade.git.stashPop` | `(args: { worktreePath: string, index?: number }) => GitActionResult` | Apply and remove a stash |\n | `ade.git.stashDrop` | `(args: { worktreePath: string, index?: number }) => GitActionResult` | Remove a stash entry |\n+| `ade.git.stashClear` | `(args: { laneId: string }) => GitActionResult` | Remove all stash entries |\n | `ade.git.fetch` | `(args: { worktreePath: string, remote?: string }) => GitActionResult` | Fetch from remote |\n | `ade.git.sync` | `(args: { worktreePath: string, strategy: 'merge' \\| 'rebase', ref?: string }) => GitActionResult` | Merge or rebase with upstream |\n | `ade.git.push` | `(args: { worktreePath: string, forceWithLease?: boolean }) => GitActionResult` | Push to remote |\n \n ---\n \n ## 2026-02-16 Addendum — Lane/Hosted Context Integration\n \n ### Lane-to-hosted context flow\n \n 1. Lane activity updates deterministic packs.\n 2. Conflict prediction updates lane conflict freshness metadata.\n 3. Hosted submission chooses context source (`inline` vs `mirror`) by policy.\n 4. `__adeHandoff` carries source, reason code, staleness, and manifest refs.\n 5. Worker prompt includes provenance and scoped file-set summary.\n \n ### UI connections in Lanes + Settings\n \n - Lanes tab:\n - Conflict status badges reflect prediction staleness and unresolved state.\n - Pack view consumes updated manifests with context fingerprint/freshness.\n - Settings page (`Hosted` section):\n - shows last sync success/attempt/error.\n - shows last cleanup success/attempt/error.\n - shows context fallback count.\ndiff --git a/docs/features/PULL_REQUESTS.md b/docs/features/PULL_REQUESTS.md\nindex efb34c3b535c31b119e7dbc0f1810ae2ee323caf..d145800525165530b2e187d5f91eb1c9a67eb529 100644\n--- a/docs/features/PULL_REQUESTS.md\n+++ b/docs/features/PULL_REQUESTS.md\n@@ -100,51 +100,51 @@ That is the main reason the normal PR and GitHub views feel less heavy than befo\n \n ## Merge bypass for blocked PRs\n \n When GitHub reports a PR as not mergeable (e.g. due to branch protection rules), ADE now offers an explicit opt-in to attempt the merge anyway. This covers cases where the user's account has permission to bypass branch protection requirements even though the GitHub API reports `isMergeable: false`.\n \n The UI surfaces a checkbox in the merge action area when the PR is open, has no merge conflicts, but is flagged as not mergeable. Checking it enables the merge button in a warning state. The merge request still goes through the normal GitHub merge API — GitHub itself decides whether the bypass is allowed.\n \n ---\n \n ## CI running indicator\n \n PR list rows (Normal tab, GitHub tab) and the PR detail merge readiness panel now show a spinning indicator when CI checks are still running. The `PrCiRunningIndicator` component lives in `prVisuals.tsx` and supports both icon-only and labeled variants. The detail pane's check status row also uses it as a title accessory.\n \n ---\n \n ## PR notifications\n \n Notification messages emitted by the polling service are now generic (not PR-specific) to work better as system notifications. The title no longer includes the PR number. The event payload now includes `prTitle`, `repoOwner`, `repoName`, `baseBranch`, and `headBranch` fields so consumers can format context-aware messages themselves.\n \n ---\n \n ## Post-merge cleanup resilience\n \n After a successful GitHub merge, cleanup operations (branch deletion, group membership removal, lane archiving, base branch fetch, cache invalidation, rebase-needs scan) are wrapped in an outer try-catch so that a failure in any cleanup step does not mask the successful merge. The operation is marked as succeeded regardless, with a `cleanupError` metadata field when something went wrong. Individual cleanup failures are logged as warnings.\n \n-The Create PR modal validates branch names against invalid git ref characters before submission.\n+The Create PR modal validates branch names against invalid git ref characters before submission. All three PR creation modes (normal, queue, integration) display lane warning panels alongside the draft checkbox, surfacing rebase needs and other lane health issues before submission. Error messages from failed GitHub API calls are cleaned of internal IPC prefixes before display, and the PR service wraps creation failures with contextual head/base branch information for clearer diagnostics. GitHub API error responses now extract nested error detail messages (from the `errors` array in the response body) and append them to the thrown error, improving visibility of issues such as duplicate PR or branch protection failures.\n \n The PR detail pane renders markdown body content with `rehype-sanitize` applied after `rehype-raw`, stripping potentially unsafe HTML from PR descriptions fetched from GitHub.\n \n ---\n \n ## Queue landing state machine\n \n The queue landing service enforces an explicit state transition table (`ALLOWED_TRANSITIONS`) for queue entry states. Invalid transitions are logged and rejected rather than silently applied. The `markEntryLanded` helper centralizes the landed-entry bookkeeping (state, position advance, active-PR reset). The cancel path now force-fails entries that are in non-skippable states (e.g., `landing`, `resolving`) with a warning rather than leaving them in an inconsistent state.\n \n ---\n \n ## Current product contract\n \n The current PR experience follows these rules:\n \n - use local git state for merge/integration truth\n - load normal PR browsing data first\n - load workflow state only when the user is actually in a workflow view\n - never auto-run expensive integration simulation just from tab selection\n - make GitHub revisits warm through layered caching\n - include rebase needs and auto-rebase statuses in the main refresh batch\n \n That preserves the full PR feature set while keeping the common browse-and-review path much cheaper.\n \n ---\ndiff --git a/docs/features/TERMINALS_AND_SESSIONS.md b/docs/features/TERMINALS_AND_SESSIONS.md\nindex 5841651ac11e9502a138b067660be276153bd2a6..c926faa6a5079cf91109506f11e352dfc9b3f379 100644\n--- a/docs/features/TERMINALS_AND_SESSIONS.md\n+++ b/docs/features/TERMINALS_AND_SESSIONS.md\n@@ -84,25 +84,37 @@ The lifecycle model remains the same:\n \n UI observers subscribe selectively and reuse cached list results where possible.\n \n ### Refresh-before-activate ordering\n \n When a new session is created or an existing session is opened, the renderer refreshes the session list *before* activating the session tab. This ensures the new session exists in `sessionsById` when the UI resolves `activeSession` for the tab. Without this ordering, the active item ID would point to an unknown session and the view would fall back to the most recent session or display a blank pane until the next refresh cycle.\n \n This pattern applies across all session-creation and session-opening paths:\n \n - `useWorkSessions` and `useLaneWorkSessions`: `refresh()` is called and awaited before `focusSession()` and `openSessionTab()`.\n - `TerminalsPage`: `work.refresh()` is awaited before `work.openSessionTab()`.\n - `AgentChatPane`: The `onSessionCreated` and `refreshSessions` callbacks are awaited (not fire-and-forget) so that the parent surface navigates the user to the chat tab before the first agent turn begins.\n \n ---\n \n ## Current product contract\n \n The current terminals and sessions feature follows these rules:\n \n - keep session tracking comprehensive\n - keep transcript and delta capture authoritative\n - let multiple renderer surfaces reuse session-list results\n - avoid perpetual idle polling when there are no live sessions to watch\n \n That preserves ADE's session-awareness while making the session system a lighter dependency for the rest of the UI.\n+\n+---\n+\n+## Terminal status indicators\n+\n+Terminal session status dots use distinct visual treatments per state:\n+\n+- **Running (active)**: Spinning emerald ring (border spinner animation). Used in the tab nav, top bar, and terminal panel tabs.\n+- **Running (needs attention)**: Solid amber dot (no animation). Indicates the terminal is awaiting user input. The top bar variant pulses; the tab nav variant does not animate.\n+- **Ended**: Solid red dot (no animation).\n+\n+The `sessionStatusDot()` helper in `terminalAttention.ts` and `sessionIndicatorState()` produce the dot class and spinning flag. The chat status glyph for \"waiting\" state renders a static check-circle icon rather than a spinner, distinguishing idle-waiting from active-working.\n","external_storage_diff",{"_1102":1103,"_1104":-5},"files_modified",30,"lines_added",986,"lines_removed",110,"commit_message","Merge base updates and resolve AI integration doc conflict","additional_parent_commit_shas","file_id","file_00000000635c71f78c642ccc7458a64b","ttl","partial_repo_snapshot","files",[1108],{"_156":1109,"_1110":-5,"_1111":1112,"_1113":-5},"docs/architecture/AI_INTEGRATION.md","contents","line_range_contents",[1114],"missing_reason",{"_1115":662,"_1116":1117,"_452":1118,"_1119":57},"line_range_start","line_range_end",2050,[1120,185,1121,185,1122,185,1123,185,1124,185,1125,185,1126,1127,1128,1129,1130,1131,1132,1133,1134,1135,1136,1137,1138,1139,1140,1141,1142,1143,1144,1145,1146,1147,1148,1149,1150,1151,1152,1153,1154,1155,1156,1157,1158,1159,1160,1161,1162,185,1124,185,1163,185,1164,185,1165,1166,1167,185,1168,185,1169,1170,1171,1172,185,1173,185,1174,185,1175,1176,1177,1178,1179,185,1180,185,1181,185,1182,185,1183,185,1184,185,1124,185,1185,185,1186,185,1187,185,1188,1189,185,1190,185,1191,185,1192,185,1193,185,1194,1195,1196,1197,185,1198,185,1199,185,1200,185,1201,185,1202,1203,1204,185,1205,185,1206,1207,1208,1209,185,1210,185,1211,1212,1213,185,1214,185,1215,185,1216,1217,1218,1219,1220,185,1221,185,1222,185,1223,1224,1225,1226,1227,1228,1229,1230,1231,1232,1233,1234,1235,1236,1237,185,1238,185,1239,185,1124,185,1240,185,1241,185,1242,185,1243,185,1244,185,1245,1246,1247,1248,185,1249,185,1250,1251,1252,185,1253,1254,1255,1256,1257,185,1258,185,1259,185,1260,185,1261,185,1262,185,1263,185,1264,185,1265,1266,1267,1268,1269,1270,1271,1272,1273,1274,1275,1276,1277,185,1278,185,1279,185,1280,185,1281,1282,1283,185,1284,185,1285,185,1286,1287,1288,185,1289,185,1290,185,1291,1292,1293,185,1294,185,1295,185,1296,1297,1298,1299,1300,1301,185,1302,185,1303,185,1304,185,1305,1306,1307,1308,1309,185,1310,185,1311,185,1312,185,1313,1314,1315,1316,1317,1318,185,1319,185,1320,185,1321,185,1322,1323,1324,1325,1326,185,1327,1328,1329,1330,1331,1332,1333,1334,1335,1336,1337,1330,1331,1338,1339,1330,1331,1340,1341,1330,1342,1343,1344,1345,1346,1330,1342,1343,1347,1345,1348,1330,1342,1349,1350,1351,185,1352,1353,1354,1355,1356,1357,1358,1359,1360,1361,1362,1363,185,1364,1365,1366,1367,1368,1369,1370,1371,185,1372,1373,1374,1375,1376,1377,1378,1379,1380,185,1381,185,1382,185,1383,185,1384,185,1385,1386,1387,1388,1389,1390,1391,185,1392,185,1393,185,1394,1395,1396,1397,1398,185,1399,185,1400,185,1401,1402,1403,1404,1405,1406,1407,1408,1409,1410,1411,1412,185,1413,185,1414,185,1415,1416,1417,1418,1419,185,1420,185,1421,185,1422,1423,1424,1425,185,1426,185,1427,185,1428,1429,1430,1431,1432,1433,1434,1435,1436,1437,1380,185,1438,185,1439,185,1440,185,1441,185,1442,1443,1444,1445,1446,1447,1448,185,1449,1450,1451,1452,1453,1454,185,1455,1456,1457,185,1458,185,1459,185,1460,185,1461,185,1462,185,1463,185,1464,185,1465,185,1466,1467,1468,1469,1470,1471,1472,1473,1474,1475,1476,185,1477,185,1478,185,1479,185,1480,185,1481,185,1482,185,1483,185,1484,185,1485,185,1486,185,1487,185,1380,1488,1489,1490,1491,1492,1493,1494,1495,1496,1497,1498,1499,1500,1501,1502,1503,1498,1504,1505,1506,1507,1508,1509,1510,1511,1512,1513,1514,1498,1504,1505,1515,1516,1517,1518,1500,1519,1520,1498,1499,1500,1521,1522,1520,1498,1499,1523,1524,1489,1490,1525,1489,1490,1526,1380,185,1527,185,1528,185,1529,185,1530,1531,1532,1533,1534,1535,1536,1537,1538,185,1539,185,1540,185,1541,1542,185,1543,1544,185,1545,185,1546,185,1547,185,1548,1549,1550,1551,185,1552,185,1553,185,1554,1555,1556,1557,1558,185,1559,185,1560,185,1561,185,1562,185,1563,185,1564,1565,1566,1567,1568,1569,1570,1571,1572,1573,185,1574,185,1575,1576,1577,1578,1579,185,1580,185,1581,185,1582,185,1583,1584,1585,185,1586,185,1587,1588,1589,1590,1591,185,1592,185,1593,1594,1595,185,1596,185,1597,185,1598,1599,1600,185,1601,185,1602,185,1603,185,1604,185,1605,1606,1607,1608,185,1609,185,1610,185,1611,1612,1613,1614,1615,1616,1617,1618,1619,1620,1621,1622,1623,1624,1625,185,1626,185,1627,185,1380,1628,1629,1630,1631,1632,1629,1633,1631,1634,1380,185,1635,1636,1637,185,1638,185,1639,185,1640,1641,1642,185,1643,185,1644,185,1645,185,1646,185,1647,185,1648,1649,1650,1651,1652,185,1653,1654,1655,1656,1657,1658,1659,185,1660,185,1661,1662,1663,1664,1665,1666,1667,1668,1669,1670,185,1671,185,1672,185,1673,185,1674,185,1675,1676,1677,1678,1679,185,1680,185,1681,185,1682,185,1322,1323,1683,1684,1685,1686,1687,1688,185,1689,1690,1691,1692,1693,1694,185,1695,1696,1697,185,1698,1699,1700,1380,185,1701,185,1702,185,1703,185,1704,1705,1706,1707,185,1708,185,1709,185,1710,185,1711,185,1712,185,1713,185,1714,1715,1716,1717,1718,185,1719,185,1720,185,1721,185,1722,1723,1724,1725,1726,1727,185,1728,185,1729,185,1730,185,1731,1732,1733,185,1734,185,1735,185,1736,185,1737,1738,185,1739,185,1740,185,1741,185,1742,185,1743,185,1744,1745,1746,1747,185,1748,185,1749,185,1750,185,1751,185,1752,185,1753,1754,1755,1756,185,1757,185,1758,185,1759,185,1760,1761,185,1762,1763,1764,1765,185,1766,1767,1768,185,1769,185,1770,1771,1772,1773,1774,185,1775,1776,1777,1778,185,1779,185,1780,185,1781,185,1782,1783,1784,1785,185,1786,185,1787,185,1788,185,1789,185,1790,185,1791,185,1792,185,1793,185,1794,185,1795,185,1796,1797,1798,1799,185,1800,185,1380,1801,1802,1803,1380,185,1804,1805,1806,1807,1808,1809,185,1810,1811,1812,1813,185,1814,185,1380,1815,1816,1817,1818,1380,185,1819,185,1820,185,1821,1822,1823,1824,185,1825,185,1826,185,1380,1827,1828,1829,1830,1831,1832,1380,185,1833,1834,1835,1836,185,1837,185,1838,185,1839,185,1840,185,1841,185,1842,185,1843,185,1844,185,1845,185,1846,185,1847,185,1848,185,1849,185,1850,185,1322,1851,1852,1853,1854,1855,1856,1857,1858,1859,1860,1861,1862,1863,1864,1380,185,1865,185,1866,185,1867,1868,1869,185,1870,185,1871,1872,1873,1874,1875,1876,1877,185,1878,185,1879,185,1880,185,1881,185,1380,1882,1883,1884,1885,1886,1887,1888,1889,1890,1891,1892,1893,1894,1895,1896,1897,1898,1899,1900,1901,1902,1903,1904,1895,1905,1906,1907,1908,1909,1910,1911,1908,1909,1912,1380,185,1913,1914,1915,1916,1917,1918,1919,185,1920,1921,1922,1923,1924,185,1925,185,1926,185,1927,185,1928,185,1929,185,1930,185,1931,1932,1933,1934,1935,1936,1937,185,1938,185,1939,185,1940,1941,1942,1943,1944,185,1945,1946,1947,1948,1949,185,1950,1951,1952,1953,185,1954,185,1955,185,1956,185,1957,185,1958,1959,1960,1961,1962,1963,1964,1965,1966,1967,1968,1969,1970,1971,1972,1973,1974,1975,1976,1977,1978,185,1979,1980,1981,1982,1983,1984,1985,1986,1987,1988,1989,1990,1991,185,1992,1993,1994,1995,1996,1997,1998,1999,2000,185,2001,2002,2003,2004,2005,2006,2007,2008,2009,185,2010,2011,2012,2013,2014,2015,185,2016,2017,2018,2019,185,2020,2021,2022,2023,2024,185,2025,2026,2027,2028,2029,2030,185,2031,185,2032,185,2033,2034,2035,185,2036,185,2037,185,2038,2039,185,2040,2041,2042,2043,2044,2045,2046,2047,2048,2049,2046,2050,2051,2052,2053,2054,2046,1437,1380,185,2055,185,2056,185,2057,185,2058,2059,2060,2061,2062,2063,2064,2065,185,2066,2067,2068,2069,2070,185,2071,185,2072,185,2073,185,2074,2075,2076,2077,2078,2079,2080,2081,185,2082,185,2083,185,1322,1323,1327,1328,1329,2084,1332,1333,1335,1336,1337,2084,1338,1339,2084,1340,1341,2085,1343,1344,1345,1346,2085,1343,1347,1345,1380,185,2086,185,2087,185,2088,2089,2090,2091,185,2092,185,2093,185,2094,185,2095,185,2038,2096,2097,2098,2099,2100,2101,2102,2103,2104,2105,185,2106,2107,2108,2109,2110,2111,1437,1380,185,2112,185,2113,2114,2115,2116,2117,2118,2119,2120,185,2121,185,2122,185,2123,185,2124,185,2125,185,2126,185,2127,185,2128,185,2129,185,2130,185,2131,2132,2133,2134,2135,185,2136,185,2038,2137,2138,2139,2140,2141,2142,2143,2144,2145,2146,2147,2148,1437,185,2149,2150,2151,2152,2153,2154,2155,2156,2157,2158,2159,2160,2161,2162,2163,2164,1437,185,2165,2166,2167,2168,2169,2170,2171,2172,2173,2174,2175,2176,2177,2178,2179,185,2180,185,2181,2182,2183,1437,185,2184,2150,2185,2186,2187,1437,1380,185,2188,185,2189,185,2190,2191,185,2192,185,2038,2193,2194,185,2195,2196,2197,2198,185,2199,2200,185,2201,2202,185,2203,2204,2205,2206,2207,2208,1380,185,2209,185,2210,2211,2212,2213,2214,2215,2216,2217,2218,2219,2220,2221,2222,185,2223,185,2224,2211,2225,2226,2227,2228,2229,2230,2231,2232,2233,2234,2235,185,2236,185,2237,2211,2238,2239,2240,185,2241,185,2242,185,2243,185,2244,185,2038,2245,2246,185,2247,2248,185,2249,2250,2251,185,2252,2253,2254,2255,2256,2257,2258,2259,2260,2261,2262,2263,2105,185,2264,2265,2266,2267,2268,2110,185,2269,2270,1437,1380,185,2271,185,2272,2211,2273,2274,2275,2276,2277,2278,2279,2280,2281,2282,185,2283,185,2284,185,2285,185,2286,185,2287,185,2288,2289,2290,185,2291,185,2292,185,1380,2293,2294,2295,2296,2297,2298,2299,185,2300,2301,2302,2303,2304,185,2305,2306,2307,185,2308,2309,2310,2311,2312,2313,2314,185,2315,2316,2317,2318,2319,2320,1380,185,2321,185,1124,185,2322,185,2323,185,2324,2325,2326,2327,2328,2329,2330,2331,2332,2333,185,2334,185,2335,185,2336,2337,185,2338,185,2339,185,2340,185,2341,2342,2343,2344,2345,2346,2347,2348,185,2349,185,2350,185,2351,2352,2353,2354,2355,185,2356,185,2357,185,2358,185,2359,185,2360,2361,2362,2363,2364,185,2365,185,2366,185,1124,185,2367,185,2368,2369,2370,2371,2372,2373,2374,2375,2376,2377,2378,2379,2380,2381,2382,2383,2384,2385,2386,2387,2388,2389,2390,2391,2392,2393,2394,2395,2396,2397,2398,2399,2400,2401,2402,2403,2404,2405,2406,2407,2408,2409,2410,2411,2412,2413,2414,2415,2416,2417,2418,2419,2420,2421,2422,2423,2424,2425,2426,2427,2428,2429,185,2430,185,1124,185,2431,185,2432,185,2433,185,2434,2435,2436,185,2437,185,2438,185,2439,2440,2441,2442,2443,2444,2445,2446,2447,2448,2449,2450,2451,2452,2453,2454,2455,2456,185,2457,185,2439,2440,2458,2459,2460,2461,2462,2463,2464,185,2465,185,2439,2440,2466,2467,2468,2469,2470,2471,2472,2473,185,2474,185,2439,2440,2475,2476,2477,185,1124,185,2478,185,2479,185,2480,185,2481,185,2482,2483,2484,2483,185,2485,2483,2486,2483,185,2487,2483,2488,2483,2489,185,2490,185,2491,2483,2492,2493,2494,2495,2496,2497,2483,185,2498,185,1124,185,2499,185,2500,185,2501,185,2502,185,1428,1429,2503,2504,2505,2506,2507,2508,2509,2510,2267,2110,1437,1380,185,2511,185,2512,2513,1380,185,2514,185,2515,185,2516,185,2517,2518,2519,2520,185,2521,185,1124,185,2522,185,2523,185,2524,185,2525,2526,185,2527,185,2528,185,2529,185,2530,2531,2532,2533,185,2534,185,2535,185,2536,185,2537,185,2538,2539],"contains_end_of_file","# AI Integration Architecture","> Roadmap reference: `docs/final-plan/README.md` is the canonical future plan and sequencing source.","> Last updated: 2026-03-30","The AI integration layer replaces the previous hosted agent with a local-first, provider-flexible approach. Instead of a cloud backend with remote job queues, ADE routes work to configured runtimes (CLI subscriptions, API-key/OpenRouter providers, and local endpoints such as LM Studio/Ollama/vLLM), coordinates tooling through MCP, and manages multi-step workflows via an AI orchestrator.","---","## Table of Contents","- [Overview](#overview)","- [Agent-First Execution Contract](#agent-first-execution-contract)","- [Design Decisions](#design-decisions)"," - [Why Subscription-Powered First?](#why-subscription-powered-first)"," - [SDK Strategy](#sdk-strategy)"," - [Why MCP for Tool Access?](#why-mcp-for-tool-access)"," - [Why AI Orchestrator?](#why-ai-orchestrator)","- [Technical Details](#technical-details)"," - [Unified Executor Interface](#unified-executor-interface)"," - [Agent Execution (Unified Runtime)](#agent-execution-unified-runtime)"," - [AI Integration Service](#ai-integration-service)"," - [MCP Server](#mcp-server)"," - [Computer Use MCP Tools](#computer-use-mcp-tools)"," - [AI Orchestrator](#ai-orchestrator)"," - [Meta-Reasoner and Smart Fan-Out](#meta-reasoner-and-smart-fan-out)"," - [Adaptive Runtime](#adaptive-runtime)"," - [Context Compaction Engine](#context-compaction-engine)"," - [Session Persistence and Resume](#session-persistence-and-resume)"," - [Inter-Agent Messaging](#inter-agent-messaging)"," - [Memory Tool Wiring](#memory-tool-wiring)"," - [Shared Facts and Run Narrative](#shared-facts-and-run-narrative)"," - [Memory Architecture](#memory-architecture)"," - [External MCP Consumption](#external-mcp-consumption)"," - [CTO Agent Architecture](#cto-agent-architecture)"," - [Cross-Machine Portability](#cross-machine-portability)"," - [Compute Backends for Agent Execution](#compute-backends-for-agent-execution)"," - [Compute Environment Types](#compute-environment-types)"," - [Per-Task-Type Configuration](#per-task-type-configuration)"," - [One-Shot AI Task Patterns](#one-shot-ai-task-patterns)"," - [Agent Chat Service (Phase 1.5)](#agent-chat-service-phase-15)","- [Integration Points](#integration-points)"," - [Desktop Application](#desktop-application)"," - [Job Engine](#job-engine)"," - [Mission Service](#mission-service)"," - [Pack Service](#pack-service)"," - [Learning Packs](#learning-packs)","- [Implementation Status](#implementation-status)","## Overview","ADE's AI integration is designed around three principles:","1. **No mandatory credential management**: CLI users do not need to paste keys; `claude`/`codex` authentication is inherited. API-key/local provider configuration is optional for broader model access.","2. **Local-first execution**: ADE runs AI flows from the local desktop process and supports CLI runtimes plus direct API/local endpoints, with no ADE-hosted backend.","3. **Auditable ADE tool access**: ADE-owned tools are exposed through controlled runtime surfaces (MCP and coordinator tools), with durable logging and explicit permission profiles layered on top of provider-native model permissions.","The AI integration layer consists of four subsystems:","- **Agent SDKs** -- the execution layer that handles both CLI-backed and non-CLI model runtimes via shared execution contracts.","- **AI Integration Service** -- the main-process service that routes tasks to the appropriate provider and model.","- **MCP Server** -- the tool exposure layer that gives AI agents controlled access to ADE's capabilities.","- **AI Orchestrator** -- the coordination layer that plans and executes multi-step missions.","## Agent-First Execution Contract","From Phase 4 onward, ADE treats agent runtimes as the mandatory substrate for all non-interactive AI execution:","- Mission planning and step execution","- Conflict and PR AI actions","- Narrative generation and background summaries","- Automations (time-based and action-based), watcher, and review workflows","- Future mobile-triggered/background runs","All of those paths are normalized into a runtime record (`agentDefinitionId` + run/step/session lineage + memory policy + guardrails), even when the UX appears \"one-shot\".","Interactive lane development (`Terminals`, `Work` chat) remains direct user sessions and is not forced through mission runtime semantics.","Persistent identities (for example CTO-style or worker identity chat sessions) are resumable runtime surfaces, not business-completion signals. A session ending, being replaced, or resuming later does not by itself mean a workflow target completed successfully. Workflow completion must come from an explicit workflow gate such as launch completion, PR/review-ready milestones, mission completion, or an ADE-managed explicit completion action.","### Key Contract","The renderer never mutates the repository directly. Repository changes happen either through ADE's trusted main-process services or through CLI-backed model runtimes operating inside ADE-managed worktrees under the selected provider permission mode. ADE-owned capabilities still flow through ADE's own permission and policy layers, preserving an auditable boundary around mission state, orchestration, and local services.","## Design Decisions","### Why Subscription-Powered First?","The previous architecture required users to either sign up for a hosted service (with OAuth, cloud sync, and remote job processing) or configure bring-your-own-key (BYOK) credentials with raw API keys. Both approaches created friction:","- Hosted service: required account creation, network connectivity for AI features, and a separate billing relationship.","- BYOK: required users to obtain, paste, and rotate API keys -- a credential management burden that is error-prone and creates security surface area.","The subscription-powered approach eliminated most onboarding friction for core users. Developers who use Claude or Codex already have authenticated CLI tools on their machines. ADE spawns these CLIs as child processes, and they authenticate using whatever mechanism the user already set up (browser sign-in, token file, environment variable). ADE never sees or stores the credentials.","This also aligns AI cost with tools developers already budget for. There is no separate ADE subscription tier for AI features. API-key/OpenRouter/local runtime paths are available alongside this CLI-first path when users want broader model coverage.","### SDK Strategy","ADE uses a unified executor runtime that routes work based on model class rather than provider-specific executor implementations. The unification point is the `unified` executor kind — a single execution path that classifies models as CLI-wrapped (subprocess) or API/local (in-process) and routes accordingly.","**Key SDKs**:","- `ai-sdk-provider-claude-code` — Community Vercel AI SDK provider (Ben Vargas) wrapping Claude CLI. Authentication flows through `claude login`.","- `@openai/codex-sdk` — Official OpenAI SDK for Codex CLI. Supports subscription auth natively.","- `@ai-sdk/anthropic`, `@ai-sdk/openai`, `@ai-sdk/google`, etc. — Vercel AI SDK providers for API-key models.","**Model resolution**: `modelId` → registry lookup → `isCliWrapped` classification → subprocess or in-process path. See `docs/ORCHESTRATOR_OVERHAUL.md` for the full runtime contract.","> **Historical note**: The original architecture used separate `ClaudeExecutor` and `CodexExecutor` classes with per-provider `AgentExecutor` interface implementations. These were deleted during the Phase 2-3 orchestrator overhaul (2026-03-04) in favor of the unified runtime.","### Why MCP for Tool Access?","AI agents need to interact with ADE's internal systems (lanes, packs, conflicts, tests) to be useful. There are several ways to expose these capabilities:","- **Direct function calls**: Tight coupling, no audit trail, no permission boundary.","- **Custom API**: Works but requires ADE to invent and maintain a bespoke protocol.","- **Model Context Protocol (MCP)**: Standardized protocol with built-in support in Claude and other AI tools, providing tool discovery, structured invocation, and resource access.","ADE chose MCP because:","- It provides a **natural boundary for ADE-owned capabilities**: lane operations, mission control, context export, and related ADE services remain behind one structured tool surface.","- It enables **call audit logging**: every tool invocation is a JSON-RPC message that can be logged, replayed, and analyzed.","- It supports **resource providers**: AI agents can read ADE state (pack exports, lane status, conflict predictions) through a structured interface rather than parsing raw files.","- It is **protocol-native** to Claude: the `claude` CLI has built-in MCP client support, so connecting to ADE's MCP server requires no custom integration code.","Important nuance:","- For **CLI-backed models** (Claude CLI, Codex CLI), native file/tool behavior is governed primarily by the provider runtime's own permission mode.","- For **API-key and local models**, ADE's planning/coding tool profiles are the actual tool surface.","- In both cases, ADE separately controls which ADE-owned tools are exposed.","### Why AI Orchestrator?","Simple AI tasks (generate a narrative, draft a PR description) still execute in a single pass, but are wrapped as **ephemeral task-agent runtimes**. Missions add orchestration on top of the same runtime substrate:","- **Step sequencing**: Some steps depend on others (tests must run after implementation).","- **Parallel execution**: Independent steps should run concurrently in separate lanes.","- **Context management**: Each step needs relevant context without exceeding token budgets.","- **Failure handling**: Failed steps need retry logic, intervention routing, or graceful degradation.","- **Conflict prevention**: Agents working in parallel must not create merge conflicts.","The AI Orchestrator is a coordinator agent that plans execution strategy, spawns workers for each step, monitors progress through structured reports, and routes interventions to the user. CLI-backed coordinators/workers use provider-native permission modes for native behavior, while ADE separately scopes coordinator/MCP tool exposure by role and phase.","Autonomy boundary: the coordinator owns strategic decisions (spawn, replan, validation routing, lane transfer, escalation). The deterministic runtime only enforces state integrity and policy constraints.","Validation baseline (strict, runtime-enforced):","- Required validation contracts are enforced by runtime, not prompt advice.","- Dedicated required validation auto-spawns validator steps per completed target step.","- Missing required validation blocks downstream phase transitions and emits explicit runtime/timeline signals.","- Validation signal vocabulary:"," - `validation_contract_unfulfilled`"," - `validation_self_check_reminder`"," - `validation_auto_spawned`"," - `validation_gate_blocked`","- Validation chat system messages are emitted with normalized `metadata.systemSignal = \\\"validation_*\\\"`.","- Runtime live-update reasons include `validation_contract_unfulfilled`, `validation_self_check_reminder`, and `validation_gate_blocked`.","- Status model is strict:"," - Run status `succeeded_with_risk` is removed."," - Mission status `partially_completed` is removed.","- No sampled `spot-check` tier and no `allowCompletionWithRisk` bypass in active orchestrator behavior.","This is distinct from the orchestrator service (`orchestratorService.ts`), which is the deterministic state machine that tracks runs, steps, attempts, and claims. The AI Orchestrator is the intelligent layer on top that decides *what* to do next; the orchestrator service is the durable layer underneath that records *what happened*.","> For current orchestrator runtime contracts, execution architecture, and remaining work, see `docs/ORCHESTRATOR_OVERHAUL.md`.","## Technical Details","### Unified Executor Interface","All AI task dispatching flows through the unified executor (`unifiedExecutor.ts`). The executor resolves models via the registry, classifies by model class (CLI-wrapped vs API/local), and routes accordingly. Permission schema is class-based: `permissionConfig.cli` for CLI-wrapped models, `permissionConfig.inProcess` for API/local models.","### Agent Execution (Unified Runtime)","> **Note**: The legacy `ClaudeExecutor` and `CodexExecutor` classes have been deleted. All AI worker execution now routes through a single `unified` executor kind. See `docs/ORCHESTRATOR_OVERHAUL.md` for the current runtime contract.","The unified runtime supports three model classes:","- **CLI-wrapped** (Claude CLI, Codex CLI): Spawned as subprocesses. Authentication inherits from user's existing CLI login (`claude login`, Codex subscription). MCP server injected via `--mcp-config` flag, with worker-local MCP config mirrored into each worker CWD to support native teammate inheritance paths. In packaged builds, the MCP connection uses a bundled proxy binary (`adeMcpProxy.cjs`) that relays stdio over the desktop's Unix socket, avoiding the need for a separate headless MCP server process.","- **API/key models** (Anthropic API, OpenAI, Google, Mistral, DeepSeek, xAI, OpenRouter): In-process execution via Vercel AI SDK `streamText()`. Authentication via configured API keys.","- **Local models** (Ollama, LM Studio, vLLM): In-process execution via OpenAI-compatible endpoints.","Model resolution is `modelId`-first: the registry (`modelRegistry.ts`) resolves model descriptors with `isCliWrapped` classification, and the runtime routes accordingly. Permission schema is class-based (`cli` + `inProcess`) rather than provider-bucketed.","Coordinator runtime routing is strict phase-authoritative for worker spawning/delegation:","- Resolution order: `explicit model override -> current phase model`.","- Role-level default model fallback is removed from active orchestrator routing behavior.","**Key SDK dependencies**:","- `ai-sdk-provider-claude-code` — Vercel AI SDK provider wrapping Claude CLI (community, Ben Vargas)","- `@openai/codex-sdk` — Official OpenAI SDK for Codex CLI","- `@ai-sdk/anthropic`, `@ai-sdk/openai`, `@ai-sdk/google`, etc. — Vercel AI SDK providers for API models","#### Streaming Support","All AI responses stream back to the renderer process via IPC push events (`webContents.send`). Executors produce `AsyncIterable` streams, which the AI integration service consumes uniformly. The UI renders streaming tokens in real time.","#### Session Management","Mission worker/coordinator sessions are scoped to run/step/attempt lineage. Session data includes conversation history (bounded by token budget), tool-use history, context window contents, and transcript/state records used by Missions UI thread inspection.","### AI Integration Service","The AI integration service (`aiIntegrationService.ts`) is the main-process service that replaces the previous `hostedAgentService` and `byokLlmService`. It provides a unified interface for all AI operations:","#### Task-Type Routing","The service routes each AI task to the appropriate provider based on task type and configuration:","| Task Type | Default Model | Rationale |","|-----------|--------------|-----------|","| `planning` | `anthropic/claude-sonnet-4-6` | Strong multi-step reasoning for mission decomposition |","| `implementation` | `openai/gpt-5.4-codex` | Optimized code generation with sandbox isolation |","| `review` | `anthropic/claude-sonnet-4-6` | Detailed analysis with explanation capabilities |","| `conflict_resolution` | `anthropic/claude-sonnet-4-6` | Reasoning over overlapping changes with full context |","| `commit_message` | `anthropic/claude-haiku-4-5` | Short-form generation, low latency |","| `memory_consolidation` | `anthropic/claude-haiku-4-5` | Batch memory lifecycle processing |","| `narrative` | `anthropic/claude-haiku-4-5` | Concise, developer-facing markdown summaries |","| `pr_description` | `anthropic/claude-haiku-4-5` | Factual, structured markdown for GitHub |","| `terminal_summary` | `anthropic/claude-haiku-4-5` | Structured terminal session summaries |","| `mission_planning` | `anthropic/claude-sonnet-4-6` | Multi-turn mission decomposition with tool use |","| `initial_context` | `anthropic/claude-sonnet-4-6` | Repository scan and context doc generation |","All task types route through the unified executor. Model resolution follows: explicit per-call hint > `taskRouting..model` in config > built-in default. CLI-wrapped models spawn as subprocesses; API/local models execute in-process via Vercel AI SDK.","#### Narrative Generation","Generates human-readable lane narratives from bounded pack exports:","- **Input**: `LaneExportStandard` (token-budgeted, deterministic)","- **Output**: Structured markdown narrative summarizing what changed, why, and what to do next","- **Application**: Applied locally via marker-based replacement (`ADE_NARRATIVE_START/END`) and recorded as a pack event","#### Conflict Proposal Generation","Analyzes conflict pack data to generate resolution proposals:","- **Input**: `LaneExportLite` (lane) + `LaneExportLite` (peer, optional) + `ConflictExportStandard`","- **Output**: Resolution diff + natural-language explanation + confidence score","- **Confidence scoring**: `high` (isolated change), `medium` (overlapping but non-contradictory), `low` (semantic conflict, needs human review)","#### PR Description Drafting","Generates pull request content from lane history:","- **Input**: `LaneExportStandard` with commit history","- **Output**: PR title + body markdown + suggested reviewers","- **Template**: Configurable in `.ade/ade.yaml`","#### Provider Detection","On startup and project switch, the AI integration service probes for available providers through a multi-module detection pipeline:","- **`authDetector.ts`**: Detects CLI subscriptions (`claude`, `codex`), configured API keys, OpenRouter keys, and local model endpoints. Returns a `DetectedAuth[]` array used for mode derivation and model availability filtering.","- **`providerCredentialSources.ts`**: Reads local credential files (Claude OAuth credentials, Codex auth tokens, macOS Keychain) and checks token freshness.","- **`providerConnectionStatus.ts`**: Builds a structured `AiProviderConnections` object with per-provider `authAvailable`, `runtimeDetected`, `runtimeAvailable`, `usageAvailable`, `blocker`, and `sources` fields. Both `auth-failed` and `runtime-failed` health states now mark a provider as not runtime-available, with distinct blocker messages for each failure mode.","- **`providerRuntimeHealth.ts`**: Tracks runtime health state (`ready`, `auth-failed`, `runtime-failed`) per provider. Health version increments on state changes, invalidating the status cache.","- **`claudeRuntimeProbe.ts`**: On forced refresh, performs a lightweight Claude Agent SDK query to confirm the Claude runtime can authenticate and start from the current app session. The probe resolves the Claude Code executable path via `claudeCodeExecutable.ts` and uses the centralized `resolveDesktopAdeMcpLaunch()` from `adeMcpLaunch.ts` to inject a minimal ADE MCP server configuration so the probe runs under conditions closer to real session startup.","- **`claudeCodeExecutable.ts`**: Resolves the Claude Code CLI binary path, consulting detected auth sources for known installation locations. Used by both the runtime probe and the provider resolver to ensure consistent executable discovery.","If no usable provider is detected, ADE operates in guest mode: all deterministic features (packs, diffs, conflict detection) work normally, but AI-generated content (narratives, proposals, PR descriptions) is unavailable. The UI clearly indicates which features require a CLI subscription.","#### Model Registry & Dynamic Pricing","At startup, the AI integration service also initializes the models.dev integration (non-blocking):","1. **Fetch**: `modelsDevService.initialize()` fetches the models.dev API with a 10s timeout.","2. **Parse**: Scans all providers, builds a map of model metadata (pricing, context windows, capabilities).","3. **Cache**: Persists to a local cache file; falls back to cache on network failure.","4. **Enrich**: Calls `updateModelPricing()` to merge live pricing into the `MODEL_PRICING` Proxy object, and `enrichModelRegistry()` to update context windows and capabilities in the registry.","5. **Refresh**: Repeats every 6 hours (non-blocking).","The model registry (`modelRegistry.ts`) contains 50+ models across 10 provider families (Anthropic, OpenAI, Google, DeepSeek, Mistral, xAI, OpenRouter, and local providers: Ollama, LM Studio, vLLM), classified by auth type (`cli-subscription`, `api-key`, `openrouter`, `local`). Each `ModelDescriptor` includes pricing fields directly, with a `getModelPricing()` accessor for cost lookups. Provider-to-CLI resolution uses a flat `FAMILY_TO_CLI` lookup map instead of nested ternaries. `resolveProviderGroupForModel()` maps any descriptor to its provider group (`\"claude\"` | `\"codex\"` | `\"unified\"`), and `isModelProviderGroup()` provides a type guard for the group union. `resolveModelDescriptorForProvider()` resolves a model reference (ID, shortId, alias, or sdkModelId) with an optional provider-group hint, preferring non-deprecated descriptors that match the hint. `getRuntimeModelRefForDescriptor()` returns the correct model reference string for a given provider group (shortId for Claude, sdkModelId for Codex, full ID for unified). Model profiles (`modelProfiles.ts`) are derived from `MODEL_REGISTRY` rather than maintained as parallel lists, ensuring profiles stay in sync with the registry automatically. The `UnifiedModelSelector` groups models by auth type and shows unavailable models as disabled with an explanatory label (e.g., \"API only - not configured\").","#### Provider Options (Reasoning Tier Passthrough)","Provider-specific reasoning configuration is handled by `buildProviderOptions()` in `providerOptions.ts`. Instead of inventing arbitrary token budgets, it passes the tier string directly to each provider's native configuration:","- **Anthropic** (adaptive): `{ thinking: { type: \"adaptive\" }, effort: tier }`","- **OpenAI/Codex**: `{ reasoningEffort: tier }`","- **Google** (3.x): `{ thinkingConfig: { thinkingLevel: tier, includeThoughts: true } }`","- **DeepSeek**: `{}` (always-on, handled by `extractReasoningMiddleware`)","- **xAI**: `{ reasoningEffort: tier }`","- **Others (Ollama, Mistral)**: `{}` (no reasoning config needed)","Each model declares its own `reasoningTiers` array in the registry, and the UI only shows tiers that the selected model supports.","#### Configuration","Provider preferences are configured in `.ade/local.yaml`:","```yaml","ai:"," # Provider mode and defaults"," mode: \"subscription\" # \"guest\" | \"subscription\""," defaultProvider: \"auto\" # \"auto\" | \"claude\" | \"codex\""," # Per-task-type overrides"," taskRouting:"," planning:"," provider: \"claude\""," model: \"anthropic/claude-sonnet-4-6-api\""," timeoutMs: 45000"," implementation:"," provider: \"codex\""," model: \"openai/gpt-5.4-codex\""," timeoutMs: 120000"," review:"," timeoutMs: 30000"," conflict_resolution:"," timeoutMs: 60000"," narrative:"," model: \"anthropic/claude-haiku-4-5-api\""," timeoutMs: 15000"," maxOutputTokens: 900"," temperature: 0.2"," pr_description:"," maxOutputTokens: 1200"," terminal_summary:"," timeoutMs: 10000"," maxOutputTokens: 500"," temperature: 0.1"," # Permission and sandbox configuration"," permissions:"," claude:"," claudePermissionMode: \"plan\" # \"default\" | \"acceptEdits\" | \"bypassPermissions\" | \"plan\""," settingsSources: [] # [] = ADE-controlled; [\"project\"] = honor .claude/settings.json"," maxBudgetUsd: 5.0 # Per-session budget cap"," sandbox: true # Enable sandbox mode"," codex:"," approvalMode: \"on-request\" # \"untrusted\" | \"on-request\" | \"on-failure\" | \"never\""," sandboxPermissions: \"workspace-write\" # \"read-only\" | \"workspace-write\" | \"danger-full-access\""," writablePaths: [] # Additional writable paths beyond cwd"," commandAllowlist: [] # Allowed shell commands (empty = default set)"," # Feature toggles"," features:"," narratives: true"," conflict_proposals: true"," pr_descriptions: true"," terminal_summaries: true"," mission_planning: true"," orchestrator: true"," # Budget controls"," budgets:"," narratives: { dailyLimit: 50 }"," conflict_proposals: { dailyLimit: 20 }"," pr_descriptions: { dailyLimit: 30 }"," terminal_summaries: { dailyLimit: 100 }"," mission_planning: { dailyLimit: 10 }"," orchestrator: { dailyLimit: 5 }","```","When `provider` is set to `auto`, the service selects the best available provider based on task type defaults and CLI availability.","### MCP Server","The MCP server is a standalone package (`apps/mcp-server`) that exposes ADE's internal tools and resources to AI processes.","#### Transport","- **Protocol**: JSON-RPC 2.0 over a `JsonRpcTransport` abstraction layer","- **Stdio transport**: Used in headless mode -- AI processes connect to the MCP server's stdin/stdout pipes","- **Socket transport**: Used in embedded mode -- the desktop app serves `.ade/mcp.sock` and external agents connect via Unix socket","- **Lifecycle**: Headless mode runs standalone with its own AI backend; embedded mode shares the desktop app's service instances","- **Smart entry point**: Auto-detects `.ade/mcp.sock` to choose proxy (embedded) vs headless mode","- **Session identity**: The MCP server propagates a `chatSessionId` field through `SessionIdentity`, resolved from the `ADE_CHAT_SESSION_ID` environment variable or the `initialize` handshake params. This links MCP tool calls back to their originating chat session for artifact ownership, computer use proof association, and audit logging. For standalone chat sessions (no mission/run/step context), the server infers the chat session from the caller ID when not explicitly provided.","- **Role resolution**: Session roles are determined by environment context first. When the environment does not specify a role, the session defaults to `external`. The `external` role receives the base tool set plus any external MCP tool specs; internal roles (`agent`, `cto`, `orchestrator`, `evaluator`) are assigned only through trusted environment variables set by ADE's own worker spawning infrastructure. External sessions also receive a restrictive default computer use policy (`allowLocalFallback: false`).","#### MCP Launch Resolution","Worker and chat processes connect to ADE's MCP server through one of three launch modes, resolved by `resolveDesktopAdeMcpLaunch()` in `adeMcpLaunch.ts`:","| Mode | Binary | When used |","|------|--------|-----------|","| `bundled_proxy` | `adeMcpProxy.cjs` run via Electron's Node | Packaged desktop builds where the proxy binary exists alongside the app bundle. The proxy connects to the desktop's `.ade/mcp.sock` Unix socket and relays stdio, injecting worker identity (mission/run/step/attempt) into the MCP `initialize` handshake. |","| `headless_built` | `apps/mcp-server/dist/index.cjs` via `node` | Development or CI environments where the MCP server has been pre-built but no bundled proxy is available. |","| `headless_source` | `apps/mcp-server/src/index.ts` via `npx tsx` | Development environments where only TypeScript source is available. |","The launch resolver checks candidates in order: bundled proxy path (from `process.resourcesPath`, `__dirname`, or CWD), then the built MCP entry, then the source entry. Both `unifiedOrchestratorAdapter.ts` (worker spawning) and `claudeRuntimeProbe.ts` (provider health probing) delegate to this centralized resolver to ensure consistent MCP launch behavior across all call sites. The probe also supports a `--probe` flag that returns a JSON diagnostic without establishing a full connection.","#### Available Tools","| Tool | Description | Mutation |","|------|-------------|----------|","| `spawn_agent` | Launch a new AI agent in a specified lane | No (delegates to orchestrator) |","| `read_context` | Read pack exports, lane state, or project context | No |","| `create_lane` | Create a new lane with a worktree for agent work | Yes |","| `check_conflicts` | Run conflict prediction against other active lanes | No |","| `merge_lane` | Merge a lane back to its parent | Yes |","| `ask_user` | Route an intervention to the ADE UI for human input | No |","| `run_tests` | Execute test suites in a lane's worktree | No (reads only) |","| `get_lane_status` | Get current status of a specific lane | No |","| `list_lanes` | List all active lanes with summary status | No |","| `commit_changes` | Stage and commit changes in a lane | Yes |","#### Resource Providers","AI processes can read ADE state through structured resource endpoints:","| Resource | Format | Description |","|----------|--------|-------------|","| Pack exports | Markdown (Lite/Standard/Deep) | Token-budgeted lane and project context |","| Lane status | JSON | Current lane state including dirty/ahead/behind counts |","| Conflict predictions | JSON | Predicted merge conflicts with affected files and severity |","#### Permission and Policy Layer","Every MCP tool invocation passes through a policy engine before execution:","- **Read-only tools** (`read_context`, `check_conflicts`, `get_lane_status`, `list_lanes`, `run_tests`): Allowed by default for all authenticated sessions.","- **Mutation tools** (`create_lane`, `merge_lane`, `commit_changes`): Require explicit grant from the orchestrator's claim system. An agent must hold an active claim on the relevant lane/scope before a mutation tool will execute.","- **Intervention tools** (`ask_user`): Always allowed but rate-limited to prevent intervention flooding.","- **Agent tools** (`spawn_agent`): Restricted to the orchestrator session; individual agents cannot spawn sub-agents without orchestrator approval.","#### Call Audit Logging","Every MCP tool invocation is logged to the orchestrator's timeline:","```json","{"," \"event_type\": \"mcp_tool_call\","," \"tool\": \"commit_changes\","," \"caller\": \"agent_step_003\","," \"args\": { \"lane_id\": \"lane_abc\", \"message\": \"Implement auth middleware\" },"," \"result\": \"success\","," \"duration_ms\": 1200,"," \"created_at\": \"2026-02-19T14:30:00.000Z\"","}","This provides full traceability of what AI agents did during a mission run.","### Computer Use MCP Tools","Additional MCP tools available when the compute environment supports GUI interaction (browser or desktop mode). These tools enable agents to interact with running applications visually.","Status note (2026-03-12): this section describes the target computer-use architecture. ADE already models screenshot/browser-verification/video evidence in mission validation and closeout, but the local runtime does not yet expose the full `screenshot_environment` / `interact_gui` / `record_environment` tool loop end-to-end.","| Tool | Description | Environment | Returns |","|---|---|---|---|","| `screenshot_environment` | Capture current screen state | browser, desktop | Base64-encoded PNG image |","| `interact_gui` | Execute mouse/keyboard actions (click, type, scroll, key press, drag) | browser, desktop | Action confirmation + optional screenshot |","| `record_environment` | Start/stop video recording of the environment | desktop | MP4 artifact reference |","| `launch_app` | Start an application in the environment | browser, desktop | Process handle + initial screenshot |","| `get_environment_info` | Current environment type, resolution, running processes | all | Environment metadata |","**Computer use loop** (for agents interacting with GUIs):","1. Agent calls `screenshot_environment` to see current screen state","2. Screenshot sent to the model as part of the conversation","3. Model reasons about what it sees and returns structured actions","4. Agent calls `interact_gui` to execute actions (click, type, etc.)","5. Repeat until task is complete","**Provider integration**:","- **Claude agents**: Uses Anthropic's Computer Use Tool (`computer_20250124`) natively. The screenshot/action loop is built into Claude's tool use protocol. Actions: `mouse_move`, `left_click`, `right_click`, `double_click`, `type`, `key`, `screenshot`, `scroll`, `hold_key`, `triple_click`, `wait`.","- **Codex agents**: Uses OpenAI's CUA (Computer-Using Agent) API via the Responses API. Actions: `click(x,y)`, `type(text)`, `scroll`, `key`, `screenshot`, `wait`, `drag`.","**Artifact production**: Target behavior is for screenshots and video recordings produced by computer use tools to attach to the owning lane, mission, or agent run as `screenshot` (PNG) and `video` (MP4) artifacts.","Current implementation note: orchestrator closeout can reason about screenshot/browser-verification/video evidence from declared or discovered artifacts, and Linear closeout can publish artifact links, but automatic ADE-managed screenshot/video capture and PR-body embedding are not shipped end-to-end yet.","**Permission control**: Computer use tools require `full-auto` / `bypassPermissions` permission level. Agents in `read-only` or `edit` modes cannot use GUI interaction tools (screenshot capture is allowed in all modes).","**Computer use policy**: The `ComputerUsePolicy` mode is either `\"auto\"` (default) or `\"enabled\"`. The former `\"off\"` mode has been removed -- computer use is always available. Agents are directed to prefer Ghost OS (`ghost mcp`) for desktop or browser control when available, then other approved external backends, and only fall back to ADE-local computer use when explicitly allowed. When a task needs verification or proof, agents capture screenshots, videos, traces, or console logs and call `ingest_computer_use_artifacts` to file evidence in ADE's proof drawer.","### AI Orchestrator","The AI Orchestrator is the intelligent coordination layer that plans and executes multi-step missions. It uses a **leader/worker agent team architecture** inspired by Claude Code's agent teams model: one leader session (the orchestrator itself) coordinates multiple worker agents, each operating in its own context window and lane worktree. The orchestrator runs on top of the deterministic orchestrator service state machine, issuing commands through it rather than replacing it.","#### Module Decomposition","The AI orchestrator codebase (`aiOrchestratorService.ts`) has been decomposed from a 13.2K-line monolith into a ~9.9K-line core plus eight domain-specific modules. All modules share state through an `OrchestratorContext` object (defined in `orchestratorContext.ts`) that holds 22+ mutable `Map` objects. Extracted functions follow the pattern `fooCtx(ctx: OrchestratorContext, ...args)`, with thin wrappers in the main file: `const foo = (...args) => fooCtx(ctx, ...args)`. Cross-module dependencies are passed via typed deps objects rather than direct imports.","| Module | Lines | Responsibility |","|--------|-------|----------------|","| `aiOrchestratorService.ts` | ~9,900 | Core orchestration: autopilot tick loop, coordinator session management, step dispatch, event handling |","| `orchestratorContext.ts` | ~1,380 | `OrchestratorContext` type definition holding all mutable state Maps |","| `chatMessageService.ts` | ~2,060 | All chat/messaging: thread CRUD, message send/get, @mention parsing, agent message routing, global chat, reconciliation |","| `workerDeliveryService.ts` | ~1,650 | Inter-agent message delivery: worker delivery context resolution, PTY write / SDK injection, queued message replay, worker-to-coordinator routing |","| `workerTracking.ts` | ~1,680 | Worker state management + `updateWorkerStateFromEvent` event handler mapping orchestrator events to worker state transitions |","| `missionLifecycle.ts` | ~600 | Mission run management, hook dispatch (`dispatchOrchestratorHook`, `maybeDispatchTeammateIdleHook`) |","| `recoveryService.ts` | ~400 | Failure recovery, health sweep, hydration on startup |","| `modelConfigResolver.ts` | ~150 | Model config resolution with 30s TTL cache: `resolveCallTypeConfig`, `resolveOrchestratorModelConfig`, `resolveMissionLaunchPlannerModel` |","| `orchestratorConstants.ts` | ~170 | Runtime constants: `LEGACY_STEP_TO_TASK_STATUS`, `DEFAULT_ROLE_ISOLATION_RULES`, etc. |","The deterministic orchestrator service (`orchestratorService.ts`, ~11K lines) has also been decomposed, with `orchestratorQueries.ts` (~840 lines) extracting DB row types, normalizers, and parse helpers, and `stepPolicyResolver.ts` (~390 lines) extracting step policy resolution and file claim helpers. Both modules are shared between `orchestratorService.ts` and `aiOrchestratorService.ts`.","#### Design Principles (Informed by Claude Code Agent Teams)","The orchestrator adopts key patterns proven in Claude Code's multi-agent coordination:","1. **Leader/Worker Separation**: The orchestrator session acts as the team leader — it plans, assigns, monitors, and synthesizes. Worker agents execute implementation, review, and testing tasks independently. Workers never coordinate directly with each other; all coordination flows through the orchestrator or the shared task infrastructure.","2. **Shared Task List as Coordination Backbone**: All steps in a mission are materialized as a structured task list that both the orchestrator and the deterministic runtime can inspect. Steps have states (`pending`, `claimed`, `in_progress`, `completed`, `failed`), dependencies (a step blocked by another cannot start), and owners (the agent assigned to execute it). This mirrors Claude Code's team task list with file-lock-based claim safety.","3. **Context Isolation via Lane Worktrees**: Each worker agent operates in its own lane worktree — an isolated copy of the repository. This prevents file conflicts between parallel agents (a critical lesson from agent teams: \"two teammates editing the same file leads to overwrites\"). The orchestrator assigns file/lane ownership at the step level to guarantee isolation.","4. **Scoped Agent Profiles**: Each worker receives a focused system prompt, restricted tool access, and a scoped memory briefing — not the orchestrator's full conversation history. This matches Claude Code's subagent pattern: workers load project context independently and receive only task-specific instructions from the leader.","5. **Plan Approval Gates**: For complex or risky steps, the orchestrator can require plan approval before a worker begins implementation. The worker researches and plans in read-only mode, submits a plan to the orchestrator, and the orchestrator approves or rejects with feedback. This mirrors the `plan_mode_required` pattern in agent teams.","6. **Inter-Agent Messaging via Structured Events**: All communication between the orchestrator and workers flows through structured `OrchestratorEvent` records — not free-form text. Event types include: `step_assigned`, `step_started`, `step_completed`, `step_failed`, `intervention_requested`, `context_loaded`, `agent_spawned`, `plan_submitted`, `plan_approved`, `plan_rejected`. Each event is durable and queryable from History.","7. **Graceful Lifecycle Management**: Workers go idle between tasks. The orchestrator detects idle workers, assigns new tasks or requests shutdown. Workers can reject shutdown if they have in-progress work. All shutdown is graceful — forced termination is a last resort after timeout.","#### Architecture","Mission prompt + memory briefings"," │"," ▼","┌──────────────────────────────────────────────────┐","│ AI Orchestrator (Leader Session) │","│ Claude session + ADE MCP server connected │","│ │","│ ┌────────────┐ │","│ │ Planner │ ──> step plan (JSON schema) │","│ └────────────┘ │","│ │ │","│ ▼ │","│ ┌──────────────────┐ │","│ │ Task Dispatcher │ ──> shared task list │","│ └──────────────────┘ (pending → claimed → │","│ │ in_progress → done) │","│ ┌─────┼──────────┐ │","│ ▼ ▼ ▼ │","│ Worker Worker Worker (each in own lane) │","│ A B C │","│ │ │ │ │","│ ▼ ▼ ▼ │","│ ┌──────────────────────────────┐ │","│ │ Unified Executor │ │","│ │ (CLI-wrapped or in-process │ │","│ │ based on model class) │ │","│ └──────────────────────────────┘ │","│ Lane Lane Lane (isolated worktrees) │","│ W1 W2 W3 │","│ │ │ │ │","│ ▼ ▼ ▼ │","│ │ Monitor │ ──> heartbeats, gate reports │","│ └──────────────────┘ │","│ │ PR Strategy │ ──> integration/per-lane/ │","│ │ │ queue/manual PR flow │","│ Orchestrator Events (durable audit trail) │","└──────────────────────────────────────────────────┘","Orchestrator service (deterministic state machine)","Mission service (user-facing lifecycle)","#### Planning Phase (Built-In, Default On)","Planning is a first-class mission phase (`planning`) inside the orchestrator run. As of the current runtime, planning enforcement is mandatory — if the coordinator receives a `phases` array that lacks a planning phase, the `CoordinatorAgent` constructor injects a synthetic `builtin:planning` PhaseCard at position 0 with `mustBeFirst: true`, `requiresApproval: false`, and `askQuestions: { enabled: true, maxQuestions: 5 }`, shifting all other phase positions down. This guarantees that every phased mission begins with a research-and-plan pass before any code-changing workers are spawned.","Runtime flow with planning enabled:","1. Mission run starts with `phaseRuntime.currentPhaseKey = \"planning\"`.","2. Coordinator enters planning mode, gathers mission context, and should hand planning work off quickly.","3. If planning clarifications are enabled (`askQuestions.enabled === true`), the coordinator must issue at least one `ask_user` clarification or confirmation round before finalizing the plan. The `PhaseCardAskQuestions` type supports multi-round deliberation: when `orderingConstraints.canLoop` is true on the planning phase, the coordinator may loop back to gather additional clarifications before finalizing the plan. A `loopTarget` field optionally names the phase to loop back to (e.g., looping from a review phase back to planning). The `maxQuestions` field (clamped to 1–10) bounds how many questions the coordinator can ask per phase iteration, preventing unbounded clarification cycles.","4. While a planning quiz intervention is open, coordinator task/delegation tools are runtime-blocked.","5. Planning work runs in read-only mode.","6. Coordinator requires a usable planner result before downstream development unlocks.","7. When the planning phase has `requiresApproval: true`, the coordinator's call to `set_current_phase({ phaseKey: \"development\" })` triggers an approval gate. The runtime creates a `phase_approval` intervention that pauses the mission until the user reviews the planning output and explicitly approves the transition. See [Intervention Routing](#intervention-routing) for details on the `phase_approval` intervention type.","8. Once approval is granted (or if `requiresApproval` is false, which is the built-in default), the coordinator transitions to development.","9. After delegation, coordinator should stay mostly event-driven until workers report actionable progress, failure, or escalation.","**Mandatory planning enforcement** (VAL-PLAN-005): The `CoordinatorAgent` constructor enforces that a planning phase exists. If the caller-provided `phases` array omits planning, the constructor injects one with `mustBeFirst: true`, `askQuestions: { enabled: true, maxQuestions: 5 }`, and `requiresApproval: false`. Additionally, on the first coordinator turn, a planning watchdog (`enforcePlanningFirstTurnDelegation`) checks whether the coordinator actually spawned a planning worker; if it did not (and no planning execution record exists), the watchdog force-spawns a read-only planning worker via `buildPlanningRecoveryPrompt()` to prevent the coordinator from bypassing the planning phase.","Runtime flow with planning disabled:","1. Run seeds `phaseRuntime` from the first enabled phase (typically `development`).","2. Coordinator skips planning protocol and proceeds directly to DAG orchestration.","Legacy note:","- Older persisted rows may still contain `plan_review`; runtime normalizes those to `in_progress` for read compatibility.","#### Worker Agent Spawning","For each step that enters the `claimed` state, the orchestrator spawns a worker agent. As of M4/M5, worker spawning is gated by three additional runtime checks: adaptive complexity classification, budget hard caps, and phase-authoritative model resolution.","**Pre-spawn gates** (evaluated in `authorizeWorkerSpawnPolicy` and `checkBudgetHardCaps`):","- **Budget hard cap check**: Before spawning, the runtime queries `getMissionBudgetStatus()` for the current mission. If any hard cap is triggered (5-hour usage, weekly usage, or API key spend limit), the spawn is rejected and the coordinator receives a structured error explaining which budget limit was hit. The `onHardCapTriggered` callback creates a blocking intervention so the user can decide whether to raise the cap or finalize the mission.","- **Phase ordering validation**: The runtime validates that spawning a worker in the current phase respects `mustBeFirst`, `mustBeLast`, `mustFollow`, and `mustPrecede` ordering constraints on PhaseCards. Validation gates on prior phases must also have completed before a worker in a downstream phase can start.","- **Adaptive parallelism**: The `scaleParallelismCap()` function in `adaptiveRuntime.ts` determines how many workers can run concurrently based on the mission's `TeamComplexityAssessment.estimatedScope`. Small missions get a cap of 1 (serial execution), medium missions 2, large missions 4, and very large missions 6.","- **Model downgrade evaluation**: Before resolving the worker's model, `evaluateModelDowngrade()` checks whether the current provider usage percentage exceeds the configured `downgradeThresholdPct`. If so, the runtime substitutes a cheaper model (e.g., Opus → Sonnet → Haiku, or GPT-5 → GPT-4o → GPT-4o-mini) and logs the downgrade reason. This allows long-running missions to stay within budget by progressively reducing per-token cost.","**Spawn sequence**:","1. **Lane Assignment**: Each worker operates in a dedicated lane worktree. If the step specifies `lanes: [\"new\"]`, a new lane is created via the `create_lane` MCP tool. If it references an existing lane, the worker is assigned to that lane.","2. **Agent Profile Construction**: The worker receives:"," - A **system prompt** built from the step's description, the mission context, and any identity policy (Phase 4)."," - A **memory briefing** scoped to the mission and agent context."," - An **ADE tool profile** — coordinator/MCP/reporting tools are restricted to those appropriate for the worker role and phase."," - A **permission mode** — for CLI-backed models this governs native behavior (`plan`/read-only vs edit/full execution); for API-key/local models it selects ADE's planning/coding tool profiles.","3. **ADE Tool Connection**: Workers receive ADE-owned tools through the current runtime surface. CLI-backed workers commonly connect through the ADE MCP server (via the bundled proxy in packaged builds or the headless MCP server in development), while API/local models use ADE's in-process planning/coding tools. In both cases, ADE enforces claimed scope (lane + file patterns) for ADE-owned actions.","4. **Session Tracking**: Worker execution attempts are registered as tracked sessions/attempts for transcript capture, delta computation, and pack integration — the same lifecycle guarantees as interactive chat sessions.","**Phase-authoritative model resolution**: Worker model selection follows a strict resolution order: explicit model override (if the coordinator specifies `modelId` in `spawn_worker`) → current phase model (from `phaseRuntime.currentPhaseModel.modelId`). If the coordinator requests a model that differs from the current phase's configured model, the spawn is rejected with a descriptive error directing the coordinator to either omit `modelId` or call `set_current_phase` first. Role-level default model fallback has been removed from active orchestrator routing.","#### Worker Coordination Patterns","The orchestrator manages workers through several coordination patterns:","| Pattern | Description | When Used |","|---------|-------------|-----------|","| **Sequential Chain** | Worker A completes, Worker B starts with A's output as context | Steps with hard dependencies (implement → test → review) |","| **Parallel Fan-Out** | Multiple workers start simultaneously in separate lanes | Independent implementation steps (feature A in lane-1, feature B in lane-2) |","| **Fan-In Merge** | Orchestrator waits for all parallel workers, then merges lanes | After parallel implementation, before integration testing |","| **Plan-Then-Implement** | Worker plans in read-only mode, orchestrator approves, worker implements | Complex/risky steps where the approach should be validated first |","| **Review-and-Revise** | One worker implements, another reviews, orchestrator decides on revision | Quality-critical code paths |","| **Speculative Parallel** | Multiple workers attempt the same step with different approaches, best result wins | Ambiguous tasks where the optimal approach is unclear |","| **Atomic Batch Delegation** | Coordinator delegates N children in one validated transaction via `delegate_parallel` | When a parent worker needs multiple independent sub-tasks launched together |","| **Approval-Gated Phase Transition** | Coordinator completes phase work, runtime blocks transition until user approves via `phase_approval` intervention | Phases with `requiresApproval: true` (e.g., planning → development) |","**Approval gates** (M4/M5): The `PhaseCard` type includes a `requiresApproval` boolean field. When a phase has `requiresApproval: true`, the coordinator's call to `set_current_phase` to leave that phase is intercepted by the approval gate logic in `coordinatorTools.ts`. The runtime checks whether a resolved `phase_approval` intervention exists for the current phase. If no resolved approval is found, the runtime creates a new `phase_approval` intervention (with `pauseMission: true`) and returns a blocking error to the coordinator. The coordinator must then wait for the user to review the phase output and resolve the intervention before retrying the phase transition. This pattern is primarily used for the planning → development transition, ensuring that the user has an opportunity to review and approve the plan before implementation begins, but it applies to any phase with `requiresApproval: true`.","**Phase 4 runtime delegation contract**:","- Parent awareness is push-based: terminal child completions auto-roll up to parent pending messages, and `report_status` updates are forwarded with `[sub-agent:]` context.","- `stream_events` now includes normalized `worker_status_reported` payloads for forwarded child progress.","- `get_worker_output` remains the detailed artifact path; parent rollups intentionally stay concise.","- Native (Claude-side) teammates are auto-registered as `source: \"claude-native\"` when they report from a valid run/worker context, and are bounded by parent allocation caps.","### Coordinator-owned delegation contracts","Coordinator-spawned workers now run under an explicit runtime delegation contract. This preserves the coordinator's strategic autonomy while making delegated-scope ownership deterministic. The coordinator still decides when to delegate, how to decompose work, whether to retry, and how to replan. The runtime owns the contract that says what the coordinator may or may not do while a delegated scope is active.","This keeps three layers separate:","1. **Strategic autonomy**: the coordinator chooses plans, worker types, decomposition, recovery paths, and replanning.","2. **Delegated-scope ownership**: once a scope is delegated, the coordinator must not silently do that same scoped work itself.","3. **Runtime enforcement**: deterministic code enforces permissions, launch failure handling, terminal-state behavior, and auditability independent of prompt wording or UI copy.","Universal delegation invariants:","- Every coordinator-spawned worker gets a `DelegationContract` before launch.","- Active delegated scopes are explicitly owned; there is no silent \"unowned\" delegated period.","- Launch failures are classified and routed through recovery policy rather than letting the coordinator drift into the delegated job.","- Cancellation supersedes delegation and activates late-write suppression.","- Lifecycle/UI should derive from structured delegation state (`delegation_state` events and persisted coordinator availability), not only from coordinator prose.","Delegation modes:","- `exclusive`: one worker owns the delegated scope. Use for planner startup and similar \"do not overlap reasoning\" windows. Coordinator may observe, fetch approved startup context, launch the delegated worker, wait, message the worker, and handle run control; it may not continue the delegated repo exploration or planning itself.","- `bounded_parallel`: the coordinator may launch multiple independent workers up to a runtime cap, but scopes must stay non-overlapping and explicitly owned. Coordinator may still decompose, observe, aggregate, and schedule next-wave work.","- `recovery`: a failed or blocked delegated scope is handed to an explicit recovery contract rather than silently resuming the failed scope directly. Coordinator may classify, intervene, or launch authorized recovery work, but must not quietly \"become\" the failed worker.","Planner startup gating is now modeled as the strictest `exclusive` delegation case rather than a permanent one-off rule. Other worker types can use looser modes without weakening the core invariant that the coordinator must not silently cross delegated scope boundaries.","Primary enforcement points:","- `coordinatorTools.ts` creates delegation contracts, enforces phase-specific launch policy, and persists worker-linked contract metadata.","- `coordinatorAgent.ts` applies active-contract tool-permission checks so the coordinator cannot drift into delegated work mid-turn.","- `aiOrchestratorService.ts` projects structured delegation state into lifecycle/status UI while keeping cancellation and late-write suppression outside prompt logic.","#### File Conflict Prevention","The orchestrator prevents file conflicts between parallel workers through:","1. **Claim-Based File Ownership**: Each step declares `filePatterns` in its claim policy. The orchestrator validates at planning time that no two parallel steps claim overlapping patterns. If overlap is detected, the planner re-sequences those steps.","2. **Pre-Merge Conflict Check**: Before merging a worker's lane back, the orchestrator calls `check_conflicts` to detect any conflicts with other active worker lanes. If conflicts are found, the orchestrator decides whether to auto-resolve (using the conflict resolution AI), ask the user, or re-sequence.","3. **Merge Sequencing**: The `mergePolicy` field in the plan determines when worker lanes are merged back:"," - `sequential`: Each worker's lane is merged immediately after the step completes."," - `batch-at-end`: All worker lanes are merged in dependency order after the entire mission completes."," - `per-step`: The orchestrator decides per-step based on downstream dependencies.","#### Step Execution","For each step in the plan, the orchestrator:","1. Checks dependency satisfaction (all predecessor steps completed successfully, or join policy allows continuation).","2. Acquires claims on the required scopes (lane, file patterns, environment keys) via the deterministic runtime's claim system.","3. Creates a context snapshot with the appropriate export level for the step.","4. If `requiresPlanApproval` is true:"," a. Dispatches the worker in read-only mode (provider-native plan permission)."," b. Worker researches the codebase and submits a plan via structured output."," c. Orchestrator evaluates the plan (checks for scope creep, file ownership violations, test coverage)."," d. If approved, re-dispatches the worker with edit permissions."," e. If rejected, provides feedback and re-dispatches in plan mode for revision.","5. Dispatches the step by `executorKind`:"," - `unified`: resolve descriptor from `modelId`, then route by execution class (CLI descriptor -> subprocess adapter, non-CLI descriptor -> in-process unified path)."," - `shell`: runs a shell command (for deterministic steps like test execution)."," - `manual`: waits for user action (for steps requiring human judgment).","6. Monitors the attempt via session tracking and claim heartbeats.","7. On completion, records the result envelope, releases claims, and optionally triggers merge.","#### Worker Lifecycle","Workers follow a predictable lifecycle:","spawned → initializing → working → idle/completed/failed"," │"," (if more tasks)"," ▼"," working → ..."," (if shutdown)"," disposed","- **Idle Detection**: When a worker completes a step, it enters `idle` state. The orchestrator detects this via the session tracking system and either assigns the next step or requests shutdown.","- **Heartbeat Monitoring**: Workers emit claim heartbeats at regular intervals. If heartbeats stop (agent crash, timeout), the orchestrator marks the step as `failed` and handles retry/escalation.","- **Graceful Shutdown**: When the mission completes, the orchestrator sends shutdown requests to all idle workers. Workers acknowledge and exit. Workers with in-progress work can reject shutdown; the orchestrator waits for completion before re-requesting.","#### Context Window Management","The orchestrator manages AI context budgets through ADE's pack export system:","- **Lite exports** (~2K tokens): Lane metadata, file list, recent commits. Used for quick status checks and worker heartbeat context.","- **Standard exports** (~8K tokens): Lite content plus file-level diffs, test results, and conflict state. Used for most planning and review tasks.","- **Deep exports** (~32K tokens): Standard content plus full file contents for key files, detailed transcript excerpts, and narrative history. Used for complex implementation steps.","Each step in a plan specifies its `contextProfiles` field, and the orchestrator assembles the appropriate export before dispatching the step.","**Progressive Context Loading**: The orchestrator session itself starts with the mission pack + project pack. As workers complete steps and produce results, the orchestrator loads result summaries on demand rather than accumulating all worker output. This prevents the orchestrator's context window from filling under large missions.","**Context Pressure Management**: When the orchestrator's context utilization exceeds 80%, it triggers a summarization pass — older step results and intermediate context are compressed into summary blocks. The deterministic runtime's durable state serves as the ground truth; the orchestrator can always re-read step results from the runtime if needed.","#### Intervention Routing","When an AI agent encounters a situation requiring human input, it invokes the `ask_user` MCP tool. The orchestrator:","1. Pauses the current step's attempt.","2. Creates an intervention record in the mission service with structured context (what the agent was doing, what it needs, what options exist).","3. Broadcasts an intervention event to the renderer via IPC.","4. The UI displays the intervention in the mission detail view with the agent's question and context.","5. When the user responds, the intervention is resolved and the orchestrator resumes the step with the user's input.","Interventions can also be triggered automatically when:","- A step fails and exceeds its retry limit.","- A conflict is detected between worker lanes that cannot be auto-resolved.","- A gate report indicates a blocking condition (e.g., failing tests after implementation).","- A worker's plan is rejected by the orchestrator and the orchestrator cannot provide adequate feedback (escalates to user).","- Budget or time limits are approaching.","- A phase transition is attempted on a phase with `requiresApproval: true` (triggers a `phase_approval` intervention).","**Intervention types**: The `MissionInterventionType` union includes the following values:","| Type | Trigger | Blocking Semantics |","|------|---------|-------------------|","| `manual_input` | Agent calls `ask_user` or `request_user_input` | Pauses the requesting step; other steps may continue |","| `ask_user` | Coordinator needs clarification during planning | Blocks coordinator task/delegation tools while open |","| `failed_step` | Step exhausts retry budget | Blocks the failed step; coordinator decides next action |","| `orchestrator_escalation` | Coordinator cannot resolve an issue autonomously | Mission-level pause until user responds |","| `budget_limit_reached` | Hard budget cap triggered during spawn or execution | Blocks all further spawns; mission pauses |","| `provider_unreachable` | CLI or API provider is unavailable | Blocks steps targeting that provider |","| `unrecoverable_error` | Fatal runtime error with no retry path | Mission-level pause |","| `phase_approval` | `set_current_phase` called on a phase with `requiresApproval: true` | **Blocks the entire phase transition.** The coordinator cannot proceed to the next phase until the user resolves this intervention. Created with `pauseMission: true`, which suspends the orchestrator autopilot loop. The intervention metadata includes `phaseKey`, `phaseName`, `targetPhaseKey`, `targetPhaseName`, and `source: \"phase_approval_gate\"`. Once the user resolves the intervention (via the mission detail UI), the coordinator retries `set_current_phase` and the transition succeeds. |","The `phase_approval` intervention type was introduced in M4/M5 to support the approval gate pattern. It differs from other intervention types in that it blocks at the phase-transition level rather than the step level — no work in the target phase can begin until approval is granted, regardless of how many steps are ready to execute.","#### Error Classification and Benign Failure Handling","The orchestrator classifies worker errors and warnings to distinguish genuinely blocking failures from benign noise that should not trigger retries or mission-level escalation.","**`classifyBlockingWarnings()`** (in `orchestratorQueries.ts`) scans all warnings and the attempt summary for patterns that indicate real failures versus expected noise:","- **External MCP noise**: Warnings matching `EXTERNAL_MCP_NOISE_PATTERNS` (e.g., Slack/claude.ai connection chatter) are silently skipped — they originate from external MCP servers and do not reflect worker failure.","- **Benign sandbox blocks** (`BENIGN_SANDBOX_BLOCK_PATTERNS`): Provider-native planning features sometimes attempt writes that the sandbox intentionally blocks. Two categories are treated as benign:"," 1. **`.claude/plans/` path blocks**: The planning worker prompt directs plan artifacts to `.ade/plans/`, so sandbox blocks on the provider-native `~/.claude/plans/` path are expected and intentional."," 2. **ExitPlanMode Zod/validation errors**: Workers are instructed not to use `ExitPlanMode` (the provider-native plan approval flow), but if a worker attempts it anyway, the resulting Zod schema mismatch or validation error is classified as benign. The patterns `ExitPlanMode.*(?:Zod|validation|schema|parse)` and the reverse match order both suppress these errors from triggering step failure or retry.","- **Blocking patterns**: Warnings that survive the noise and benign filters are matched against `BLOCKING_WARNING_PATTERNS` which categorize genuine failures (sandbox blocks on disallowed paths, permission violations, tool crashes, etc.).","This classification is critical for planning workers: without benign pattern filtering, ExitPlanMode errors would cause spurious retry loops, wasting budget on repeated planning attempts that were actually successful.","#### Orchestrator Configuration","The orchestrator is configurable in `.ade/local.yaml`:"," orchestrator:"," # Planning phase behavior is controlled by mission phase profiles."," # Built-in profiles include `planning` as phase 1 by default."," max_parallel_workers: 4 # Max concurrent worker agents"," default_merge_policy: sequential # sequential | batch-at-end | per-step"," default_conflict_handoff: auto-resolve # auto-resolve | ask-user | orchestrator-decides"," # Worker management"," worker_heartbeat_interval_ms: 30000"," worker_heartbeat_timeout_ms: 90000"," worker_idle_timeout_ms: 300000 # Shut down idle workers after 5 min"," step_timeout_default_ms: 300000 # Default per-step timeout"," max_retries_per_step: 2"," # Context management"," context_pressure_threshold: 0.8 # Trigger summarization at 80% capacity"," progressive_loading: true # Load worker results on demand"," # Budget"," max_total_budget_usd: 50.0 # Hard budget cap for entire mission"," max_per_step_budget_usd: 10.0 # Per-step budget cap","### Meta-Reasoner and Smart Fan-Out","The meta-reasoner (`metaReasoner.ts`) adds AI-driven dispatch intelligence to the orchestrator's autopilot loop. Rather than relying solely on the static step plan, the meta-reasoner analyzes in-flight mission state and dynamically injects or restructures steps.","**`analyzeForFanOut()`**: The core entry point. Given the current run state, active steps, and completed results, the meta-reasoner:","1. Evaluates whether remaining work can be parallelized.","2. Selects a fan-out strategy: `external_parallel` (multiple agents in separate lanes), `internal_parallel` (single agent handling sub-tasks), or `hybrid` (combination).","3. Dynamically injects new steps into the orchestrator's step DAG with appropriate dependency edges.","4. Tracks fan-out completion via a fan-in pattern — the orchestrator waits for all fan-out steps to complete before proceeding to dependent steps.","**Integration**: The meta-reasoner is invoked within the autopilot tick loop (`aiOrchestratorService.ts`). When the autopilot detects that multiple steps could run concurrently but are currently sequenced, it consults the meta-reasoner before dispatching.","**Configuration**: The meta-reasoner model and fan-out limits are configurable in `.ade/local.yaml` under `ai.orchestrator.metaReasoner`.","### Adaptive Runtime","The adaptive runtime module (`adaptiveRuntime.ts`, introduced in M5) provides heuristic-driven scaling of parallelism and model cost based on mission complexity and budget utilization. It operates as a pure-function layer consumed by the coordinator tools and orchestrator service — it does not maintain state itself.","#### Task Complexity Classification","`classifyTaskComplexity(description: string): TaskComplexity` analyzes a mission or step description and returns one of four complexity buckets: `trivial`, `simple`, `moderate`, or `complex`. Classification uses a multi-signal heuristic:","- **Word count**: Short descriptions (< 20 words with trivial indicators) are classified as trivial; long descriptions (> 120 words) as complex.","- **Complexity indicators**: Keywords like \"parallel\", \"distributed\", \"migration\", \"architecture\", \"overhaul\", \"cross-cutting\", and \"end-to-end\" push toward `complex`.","- **Moderate indicators**: Keywords like \"integrate\", \"service\", \"endpoint\", \"feature\", \"module\", \"refactor\", and \"ci/cd\" push toward `moderate`.","- **Simple/trivial indicators**: Keywords like \"fix bug\", \"rename\", \"typo\", \"formatting\", and \"docs\" push toward simpler buckets.","- **File reference density**: More than 10 file references in the description suggests complex scope; more than 4 suggests moderate.","The complexity classification feeds into parallelism scaling and is also available for coordinator prompt construction.","#### Parallelism Cap Scaling","`scaleParallelismCap(estimatedScope: TeamComplexityAssessment[\"estimatedScope\"]): number` maps the mission's estimated scope to a concrete worker concurrency limit:","| Estimated Scope | Parallelism Cap | Rationale |","|----------------|----------------|-----------|","| `small` | 1 | Serial execution; coordination overhead exceeds parallelism benefit |","| `medium` | 2 | Modest parallelism for moderately scoped work |","| `large` | 4 | Full parallel fan-out for multi-workstream missions |","| `very_large` | 6 | Maximum parallelism for large-scale refactors and multi-service work |","The parallelism cap is enforced at spawn time — the coordinator's `spawn_worker` and `delegate_parallel` tools check the current active worker count against the scaled cap before proceeding.","#### Model Downgrade at Usage Thresholds","`evaluateModelDowngrade(args)` checks whether current provider usage exceeds a configurable threshold and, if so, substitutes a cheaper model for subsequent worker spawns:","- **Inputs**: `currentModelId`, `downgradeThresholdPct` (e.g., 70%), `currentUsagePct` (from budget telemetry), and an optional `cheaperModelId` override.","- **Downgrade heuristic** (`resolveCheaperModel`): Anthropic models downgrade along the Opus → Sonnet → Haiku chain; OpenAI models downgrade along GPT-5 → GPT-4o → GPT-4o-mini. If no cheaper model is available (already at the cheapest tier), the original model is retained with a logged warning.","- **Output**: A `ModelDowngradeResult` struct indicating whether a downgrade occurred, the original and resolved model IDs, and a human-readable reason string that is injected into the coordinator's event stream.","Model downgrade is evaluated inside the worker spawn authorization flow (`authorizeWorkerSpawnPolicy`) so that the coordinator receives the downgraded model transparently — it does not need to implement its own cost-optimization logic.","#### Budget Enforcement Gates","Budget enforcement operates at two levels:","1. **Soft pressure signals**: The `onBudgetWarning` callback emits `\"warning\"` or `\"critical\"` signals to the coordinator's event stream, guiding it to reduce parallelism or finalize early. These are advisory — the coordinator can choose how to respond.","2. **Hard cap blocks**: The `checkBudgetHardCaps()` function in `coordinatorTools.ts` queries `getMissionBudgetStatus()` and rejects spawns if any hard cap is triggered (5-hour rolling window, weekly rolling window, or absolute API key spend). Hard cap violations create a `budget_limit_reached` intervention with `pauseMission: true`.","This layered approach allows missions to gracefully degrade under budget pressure (using cheaper models, reducing parallelism) before hitting the hard stop that pauses execution entirely.","### Context Compaction Engine","The compaction engine (`compactionEngine.ts`, integrated via `unifiedExecutor.ts`) prevents SDK agent sessions from exceeding context window limits during long-running orchestrated work.","**Token Monitoring**: The engine tracks token consumption for each active agent session. When utilization reaches 70% of the model's context window, compaction is triggered.","**Compaction Flow**:","1. **Pre-compaction writeback**: Before compacting, the engine extracts durable facts (discovered patterns, key decisions, gotchas, configuration notes) from the conversation and writes them to mission-scoped entries in `unified_memories` using strict write-gate rules and `sourceRunId` lineage.","2. **Self-summarization**: The agent generates a summary of the conversation so far, preserving key context, decisions, and current task state.","3. **Conversation replacement**: The full conversation history is replaced with the summary, dramatically reducing token count while preserving essential context.","4. **Post-compaction**: The compacted summary is written to the `attempt_transcripts` table with `compacted_at` and `compaction_summary` fields.","**Threshold Configuration**: The 70% threshold is configurable via `ai.orchestrator.compaction_threshold` in `.ade/local.yaml`. The engine also respects per-model token limits from the model registry.","### Session Persistence and Resume","Agent sessions are now durable across interruptions and application restarts.","**Transcript Persistence**: The `attempt_transcripts` DB table stores the full conversation history (as JSON message arrays) for each orchestrator attempt. Transcripts are written on every significant event (tool call, agent response, compaction) via JSONL-style append.","**`resumeUnified()`**: When an orchestrator run is resumed after interruption, `resumeUnified()` in the unified executor:","1. Loads the last `attempt_transcripts` row for the attempt.","2. Reconstructs the conversation state from the stored messages.","3. If a compaction summary exists, uses it as the conversation seed.","4. Resumes the SDK agent session with the restored context.","**Chat Transcript JSONL**: In addition to the DB-backed transcript, a JSONL file is written to `.ade/transcripts/` for each attempt, providing a human-readable audit trail of the conversation.","### Inter-Agent Messaging","A structured messaging system enables communication between the orchestrator, agents, and the user during mission execution. The messaging subsystem is decomposed into two extracted modules:","- **`chatMessageService.ts`** (~1,850 lines): All chat and thread operations -- `appendChatMessage`, `listChatThreads`, `getThreadMessages`, `sendThreadMessage`, `sendChat`, `getChat`, `sendAgentMessage`, `parseMentions`, `routeMessage`, `deliverMessageToAgent`, `getGlobalChat`, `getActiveAgents`, and reconciliation functions.","- **`workerDeliveryService.ts`** (~1,330 lines): Low-level message delivery to worker agents -- `resolveWorkerDeliveryContext`, `deliverWorkerMessage`, `replayQueuedWorkerMessages`, `routeMessageToWorker`, `routeMessageToCoordinator`.","**Message Delivery** (`deliverMessageToAgent()` in `chatMessageService.ts`, with delivery mechanics in `workerDeliveryService.ts`):","- Delivers messages to both PTY-based agents (via terminal write) and SDK-based agents (via conversation injection).","- Messages can originate from the orchestrator, other agents, or the user.","- Worker delivery context is resolved per-agent to determine the appropriate delivery mechanism.","**@Mention Routing**:","- `parseMentions()` extracts @-mentions from message text, identifying target agents by name or role.","- `routeMessage()` determines which agents should receive a message based on mentions, channel context, and routing rules.","**Team Message Tool** (`teamMessageTool.ts`): An MCP tool available to agents that allows them to send messages to other agents or the orchestrator. This enables agent-initiated communication (e.g., \"I found a dependency issue that affects @testing-agent's work\").","**Phase 4 push rollups and native teammate guardrails**:","- Sub-agent completion summaries are delivered automatically to parent pending messages (no polling loop required for awareness).","- Sub-agent `report_status` updates are forwarded to parent threads with a normalized prefix and emitted to the runtime event buffer.","- Unknown native callers in `report_status`/`report_result` are auto-registered as team members with `source: \"claude-native\"` and surfaced lineage (`parentWorkerId`) when context is resolvable.","- Native auto-registration enforces parent allocation caps (derived from run parallelism; fallback `4`) before accepting updates.","**IPC Endpoints**:","- `getGlobalChat`: Retrieves the global mission chat channel messages.","- `deliverMessage`: Sends a message from the UI to a specific agent or channel.","- `getActiveAgents`: Lists currently active agents in a mission run with their status.","Threaded chat persistence is DB-first (`orchestrator_chat_threads` and `orchestrator_chat_messages`). Startup reconciliation normalizes thread/message linkage but does not run a legacy mission-metadata backfill job.","### Memory Tool Wiring","Memory tools are wired into the agent tool surfaces via `createCodingToolSet()` / `createUniversalToolSet()`, giving agents the ability to:","- **Search scoped memories**: Query `project`, `mission`, and `agent` scopes through `memorySearch`.","- **Persist durable discoveries**: Record facts/patterns/decisions/gotchas with `memoryAdd`.","- **Pin always-on context**: Promote an existing entry to Tier 1 with `memoryPin`.","- **Update CTO core memory**: Unified/CTO sessions can also call `memoryUpdateCore`.","Memory tools follow the same MCP permission model as other agent tools. Read operations are always allowed; write operations require an active claim.","### Shared Facts and Run Narrative","**Shared Team Knowledge**: Durable cross-worker coordination now lives in mission-scoped `unified_memories`, not in a separate `orchestrator_shared_facts` table. Prompt-time `sharedFacts` sections are a derived view over mission memories, giving workers the same stable prompt section without maintaining a second storage path.","**Run Narrative**: `appendRunNarrative()` in `orchestratorService.ts` generates a rolling narrative after each step completion. The narrative summarizes what has been accomplished, what is in progress, and what remains. It is stored as `runNarrative` metadata on the orchestrator run and displayed in the Activity tab.","**Compaction Hints**: A compaction hints section is added to agent prompts, providing the agent with guidance on what information to prioritize preserving if context compaction is triggered.","**Agent Memory Instructions**: All agent types (coordinator, worker, CTO, chat) now receive improved memory instructions in their system prompts with concrete examples and quality criteria. The standard quality bar is: \"Would a developer joining this project find this useful on their first day?\" Instructions include explicit SAVE guidance (non-obvious conventions, decisions with reasoning, pitfalls, patterns that contradict expectations) and DO-NOT-SAVE guidance (file paths, session progress, task status, code already committed, raw error messages without lessons, anything discoverable via search or git log).","### Memory Architecture","> **Full spec**: `docs/final-plan/phase-4.md` W6 (Unified Memory System) contains the comprehensive implementation plan including schema, write gate, scoring formula, lifecycle algorithms, and pack removal strategy.","The memory system provides agents with durable, searchable long-term memory that persists across sessions and runs. Phase 4 W6, W6½, W7a, and W7b are complete: `unifiedMemoryService.ts` is the canonical durable memory backend with 3 scopes (`project`, `agent`, `mission`) and 3 tiers (Tier 1 pinned, Tier 2 active, Tier 3 aging), plus hybrid retrieval (FTS4 BM25 + cosine similarity via local Xenova/all-MiniLM-L6-v2 + MMR re-ranking), lifecycle sweeps, batch consolidation, pre-compaction flush with quality criteria, and orchestrator mission-memory wiring. Persisted `.ade/artifacts/packs/...` artifacts are no longer required for runtime context assembly. Deterministic context exports (`packService`) and CTO identity state (`ctoStateService`) remain as explicit compatibility and audit surfaces.","#### Storage Layer","- **Primary store**: SQLite (node:sqlite + cr-sqlite) with `unified_memories` as the active memory table and `unified_memory_embeddings` storing 384-dim vectors from local Xenova/all-MiniLM-L6-v2","- **Embedding model**: `@huggingface/transformers` running `Xenova/all-MiniLM-L6-v2` locally — no API calls, fully offline","- **Background embedding**: `embeddingWorkerService.ts` processes new entries asynchronously and backfills existing ones without blocking","- **Active retrieval path**: Hybrid FTS4 BM25 (30%) + cosine similarity (70%) + MMR re-ranking (λ=0.7), with graceful fallback to lexical/composite scoring when embeddings are unavailable","#### Retrieval Pipeline","Query → Embed query via Xenova/all-MiniLM-L6-v2 → FTS4 BM25 keyword scoring + cosine similarity →"," Hybrid score (0.30 BM25 + 0.70 cosine) → Composite scoring (hybrid + recency + importance + confidence + access) →"," MMR re-ranking (λ=0.7, reduces redundancy) → Budget filter (lite: 3, standard: 8, deep: 20) → Return","The composite score combines these live signals:","- **Hybrid query relevance (40%)**: Blended BM25 keyword + cosine semantic similarity (replaces lexical-only)","- **Recency (20%)**: Exponential decay with 30-day half-life on `lastAccessedAt`","- **Importance (15%)**: High=1.0, Medium=0.6, Low=0.3","- **Confidence (15%)**: Entry confidence field (0-1), grows with observations","- **Access frequency (10%)**: `min(accessCount / 10, 1.0)`, capped at 10 accesses","Budget tiers control how many memories are injected into agent context:","- **Lite** (3 entries): Quick tasks, terminal summaries, one-shot generation","- **Standard** (8 entries): Normal agent work, implementation steps","- **Deep** (20 entries): Mission planning, complex multi-file reasoning","#### Write Pipeline","New memory → Category/strict-mode gate →"," Lexical duplicate check in same scope →"," Merge or insert into `unified_memories` →"," Queue for background embedding (async)","Memory writes are persisted directly in the local project database (`.ade/ade.db`) and consumed immediately by retrieval paths. Embeddings are generated asynchronously by the background worker and never block writes or reads.","#### Memory Lifecycle (W6½)","- **Pre-compaction flush**: Before context compaction, a silent agentic turn is injected so the agent can persist in-context discoveries via `memoryAdd`. The flush prompt now includes explicit quality criteria and SAVE/DO-NOT-SAVE guidance (e.g., save non-obvious conventions and pitfalls; do not save file paths, session progress, or raw error messages). Flush counter prevents double-flush; configurable `reserveTokensFloor` (default: 40K tokens).","- **Lifecycle sweeps**: Run on configurable interval (default: daily at 3am or on startup if >24h since last sweep). Operations in order: temporal decay (30-day half-life, Tier 1 and evergreen categories exempt), tier demotion (Tier 2→3 at 90 days, Tier 3→archived at 180 days), candidate promotion (confidence ≥ 0.7 + observationCount ≥ 2), hard limit enforcement (project: 2K, agent: 500, mission: 200), orphan cleanup (mission-scoped entries for deleted missions).","- **Batch consolidation**: Weekly (or when scope exceeds 80% of hard limit). Clusters entries by Jaccard trigram similarity > 0.7 within (scope, category) groups, then merges clusters of 3+ via LLM. Tier 1 (pinned) entries are never consolidated. Original entries archived, not deleted.","- **Memory Health dashboard**: The only memory UI surface is **Settings > Memory** (Health tab). It shows entry counts by scope/tier, last sweep and consolidation timestamps and stats, embedding progress (polled at 10s intervals), hard limit usage bars, and manual \"Run Sweep Now\" / \"Run Consolidation Now\" buttons. There are no other memory surfaces in the renderer.","#### Context Assembly Per Runtime","Every agent runtime assembles its context window from a layered budget:","System prompt + tools definition (~5-10K tokens)","+ Tier 1 pinned memory (persona + working context) (~2-4K tokens)","+ Tier 2/3 retrieved memories (budget-dependent) (~1-3K tokens)","+ Mission shared team knowledge (derived from mission memory) (~0.5-1K tokens)","+ Conversation history (remaining budget)","+ Response reserve (~4K tokens)","The 3 live memory scopes (matching `MemoryScope` in `src/shared/types/memory.ts`) are:","- `project`: Persistent project-level knowledge","- `mission`: Shared across a mission run","- `agent`: Agent-specific durable memory","`buildFullPrompt()` currently injects mission-derived shared team knowledge, mission-memory highlights, project-memory highlights, and explicit agent-memory highlights when the run carries an exact `employeeAgentId`. Structured lane/project context still comes from pack exports.","#### Prior Art & Design References","The memory architecture is informed by production systems and academic research across the agent memory landscape:","- **MemGPT / Letta**: Pioneered the tiered memory model treating LLM context as \"main memory\" with agent-managed read/write to \"disk\" storage. ADE's Tier 1 (pinned) / Tier 2 (active) / Tier 3 (aging) maps to MemGPT's core memory blocks / recall memory / archival memory. Letta's benchmarks (74% accuracy with simple file operations vs. Mem0's 68.5%) validated our choice of file-backed portable storage over database-only approaches.","- **Mem0**: Source of the PASS/REPLACE/APPEND/DELETE consolidation model. Mem0 performs real-time deduplication on every write using cosine similarity to detect overlap, then delegates merge decisions to an LLM. ADE adopts this with a conservative 0.85 similarity threshold and adds scope-aware matching (only compare within the same memory scope to prevent false merges across agent boundaries).","- **CrewAI**: The composite scoring formula (`semantic + recency + importance + access`) is adapted from CrewAI's `RecallFlow` retrieval system. CrewAI combines multiple signals for memory ranking; ADE simplifies the weights (`0.5/0.2/0.2/0.1`) for predictability and adds explicit user-settable importance tags.","- **OpenClaw**: Two direct influences — (1) the pre-compaction flush pattern, where the agent is prompted to save important memories before context eviction, using the agent's own judgment rather than mechanical extraction; (2) hybrid BM25 + vector search with configurable weights for memory retrieval. ADE formalizes the flush with a monotonic counter and compaction engine hook.","- **LangMem (LangChain)**: The episodic/procedural memory taxonomy — structured post-session summaries and learned tool-usage patterns. LangMem's key insight that procedural memories should be extracted from recurring episodic patterns (not single sessions) informed ADE's requirement for multi-episode pattern observation before procedural entry creation.","- **A-MEM**: Zettelkasten-inspired automatic linking between memory entries. While ADE does not implement full graph-based navigation in Phase 4, the consolidation APPEND operation creates implicit links and composite scoring ensures related memories co-retrieve.","- **JetBrains (NeurIPS 2025)**: Research finding that **observation masking** (replacing old tool outputs with `[output omitted]` placeholders) outperforms LLM-based summarization for context management while being significantly cheaper. ADE applies this in context assembly for resumed sessions.","- **Elvis Sun's ZOE/CODEX**: Demonstrated the context window separation principle — business/orchestration context and code context should not share the same window because context is zero-sum. Directly informed ADE's leader/worker architecture where the orchestrator holds mission context while workers hold code context.","### External MCP Consumption","`W8` shipped the baseline ADE-managed external MCP substrate. ADE now acts as the single source of truth for external MCP configuration, connection lifecycle, discovery, permissioning, and audit for ADE-managed sessions.","**Configuration**: external MCP servers are declared in `.ade/local.secret.yaml` under the `externalMcp` key:","externalMcp:"," - name: github"," transport: stdio"," command: npx"," args: [\"-y\", \"@modelcontextprotocol/server-github\"]"," autoStart: true"," - name: notion"," transport: http"," url: https://mcp.notion.so/mcp"," headers:"," Authorization: \"Bearer ${NOTION_TOKEN}\""," - name: legacy-browser"," transport: sse"," url: https://example.com/sse","`externalMcpService.ts` supports `stdio` and modern remote HTTP transport, while accepting `sse` as a compatibility alias. Connections are lazy by default, with optional auto-start, ping-based health checks, reconnect backoff, graceful drain-on-remove behavior, and manifest refresh on reconnect.","**Runtime model**: Claude Code and Codex workers still receive only ADE's MCP server at launch time. ADE does not inject third-party MCP servers directly into provider-global config. Instead:","1. ADE connects to configured external servers.","2. ADE discovers their tools and stores them as namespaced identifiers such as `ext.github.create_pull_request`.","3. ADE's own MCP server exposes the filtered `ext.*` tool surface to ADE-managed workers, the CTO, and automations.","That proxy model keeps policy, mission scoping, budgeting, and audit centralized inside ADE instead of distributing it across provider runtimes.","**Security and permission model**: external MCP tools pass through ADE's existing policy layers:","- Server-level config can `allow` or `block` specific tools.","- CTO identities and worker identities can define `externalMcpAccess`.","- Mission `permissionConfig.externalMcp` can disable external MCP entirely or narrow the allowed servers/tools per mission.","- Read-only planning workers only receive read-safe external tools in their launch allowlist.","- Mutation-classified external tools still use ADE's existing approval and guardrail path.","- External invocations are logged with server, tool, safety, and usage metadata.","**Tool discovery and exposure**: ADE queries configured external MCP servers for manifests, refreshes them on reconnect and `tools/list_changed`, and merges approved tools into the ADE-visible tool surface with namespaced identifiers. Structured MCP results are preserved when those tools are called through ADE MCP.","### CTO Agent Architecture","The CTO Agent is a persistent, always-on, project-aware agent that serves as ADE's Chief Technical Officer. Unlike a simple router, the CTO maintains full memory and context about the project, can create missions, spin up lanes, check project state, and make autonomous decisions. It bridges external systems to ADE's internal surfaces while also acting as the primary interface for high-level project orchestration.","**Architecture**:","External MCP Request User (CTO Tab)"," │ │"," ▼ ▼","┌─────────────────────────┐ ┌──────────────┐","│ MCP Server (incoming) │ │ CTO Chat UI │","│ (stdio or socket) │ │ (always-on) │","└─────────┬───────────────┘ └──────┬───────┘"," │ │"," └────────────┬────────────┘"," ▼","┌──────────────────────────────────┐","│ CTO Agent │","│ (persistent agent runtime) │","│ │","│ Project Context & Memory │","│ ┌────────────────────────────┐ │","│ │ Mission creation & mgmt │ │","│ │ Lane orchestration │ │","│ │ Project state queries │ │","│ │ Intent classification │ │","│ │ Code review coordination │ │","│ │ Chat relay │ │","│ └────────────────────────────┘ │","│ CTO State (.ade/cto/) │","│ (persistent memory & context) │","└─────────┬────────────────────────┘"," │"," ▼"," ADE Internal Surface"," (mission, lane, agent, query)"," Result → MCP Response / UI Update","**Capabilities**:","- The CTO classifies incoming requests into intents: `create_mission`, `run_task`, `review_code`, `query_state`, `relay_chat`","- Routing decisions are informed by the CTO's persistent memory — it learns which request patterns map to which handlers over time","- Complex requests that span multiple surfaces are decomposed into sub-requests and coordinated by the CTO","- The CTO can proactively create work chats, missions, lanes, PR actions, file/context reads, and managed-process actions through stable ADE services without explicit renderer clicks","- It maintains awareness of active missions, lane states, chat sessions, worker outputs, and workflow runs","- UI navigation remains suggestion-only: tool results can include explicit ADE deeplinks like Work, Missions, Lanes, or CTO, but they do not silently switch tabs","**CTO Identity and System Prompt**: The CTO system prompt is now assembled as four explicit sections by `previewSystemPrompt()` in `ctoStateService.ts`:","- **Immutable ADE doctrine**: ADE-owned identity and operating rules for the persistent CTO project operator. This is not user-editable.","- **Personality overlay**: selected from presets or one custom personality field. This only changes the behavioral overlay.","- **Memory and continuity model**: explains the long-term CTO brief, current working context, durable searchable memory, and compaction behavior.","- **ADE operator capability manifest**: enumerates the stable ADE surfaces the CTO can operate internally and clarifies that UI navigation is returned as suggestions rather than implicit tab switches.","Project-specific summary, conventions, active focus, session continuity, subordinate activity, and today's daily log are carried by the memory/continuity layers rather than by the immutable doctrine.","**CTO Daily Logs**: The CTO state service supports append-only daily logs stored as markdown files under `.ade/cto/daily/.md`. The `appendDailyLog`, `readDailyLog`, and `listDailyLogs` methods manage these logs. Today's daily log is automatically included in the CTO continuity context, providing within-day continuity.","**Post-Compaction Identity Re-injection**: When a CTO or worker identity session undergoes SDK-level context compaction, the agent chat service detects the compaction event and calls `refreshReconstructionContext()` to re-inject the full identity context (persona, core memory, memory protocol, decision framework). This prevents identity loss after compaction.","**CTO State**: The CTO maintains its state in `.ade/cto/` (core memory files, daily logs), separate from unified memory tables in `.ade/ade.db`. This includes persistent project context, learned routing patterns, decision history, and user corrections. Both systems are visible in **Settings > Memory**. Over time, the CTO becomes more effective at anticipating project needs and dispatching work autonomously.","**Approval gate interaction** (M4/M5): When a mission phase has `requiresApproval: true`, the CTO is notified of pending `phase_approval` interventions through the same mission event stream it uses for general mission awareness. The CTO can surface these approval requests to the user in the CTO chat interface, providing context about the planning output and recommending whether to approve or request revisions. The CTO does not resolve approval gates autonomously — it always routes them to the human user for final decision.","**Retrospective patterns → CTO memory**: The reflection protocol that workers and the coordinator use during mission execution (via `reflection_add`) produces structured observations about friction points, reusable patterns, and improvement recommendations. These reflections are persisted as episodic memory entries. When the same pattern appears across multiple missions, the CTO's memory system promotes them into project-level procedural knowledge. This creates a feedback loop: the CTO learns from mission retrospectives and incorporates those lessons into future mission planning guidance, routing decisions, and proactive project recommendations.","**Use Cases**:","- Primary interface for project-level AI interactions via the CTO tab","- CI/CD pipelines invoking ADE missions via MCP","- External AI agents (Claude Code, Cursor, etc.) requesting ADE to perform work","- Slack/Discord bots routing developer requests to ADE","- Monitoring systems triggering automated review or testing","- Proactive project management: detecting issues, suggesting next steps, coordinating agents","### Cross-Machine Portability","ADE now ships a canonical `.ade` contract. The tracked/shareable subset lives alongside a tracked `.ade/.gitignore` that ignores machine-local runtime state.","**Tracked/shareable state**:","- `ade.yaml` — shared baseline config","- `cto/` — CTO identity/core-memory/session-log/daily-log files","- `agents/` — worker identity/core-memory/session-log files","- `templates/`, `context/`, `memory/`, `history/`, `reflections/`, `skills/`","**Machine-local state**:","- `local.yaml`, `local.secret.yaml`","- `ade.db`, `embeddings.db`","- `mcp.sock`","- `artifacts/`, `transcripts/`, `cache/`, `worktrees/`, `secrets/`","**Repair and integrity baseline**:","- startup repair creates the canonical tree, rewrites `.ade/.gitignore`, and removes stale `.git/info/exclude` rules for `.ade`","- legacy runtime folders are moved into `artifacts/`, `transcripts/`, `cache/`, and `secrets/`","- tracked JSONL files under `history/`, `cto/`, and `agents/` are normalized and hash-chained with `prevHash`","**Embedding behavior**: Embeddings are generated locally by `@huggingface/transformers` (`Xenova/all-MiniLM-L6-v2`, 384-dim) and stored in `unified_memory_embeddings` within `.ade/ade.db`. Hybrid retrieval (FTS4 BM25 + cosine similarity + MMR re-ranking) is the active search path. The background embedding worker backfills missing embeddings automatically from the DB-backed memory store.","**No cloud dependency**: ADE's local filesystem contract is git-friendly, but real-time multi-device sync is still Phase 6 work. The Phase 8 relay is for real-time remote control of a running ADE instance, not for state synchronization.","### Shipped Implementation Summary","Phases 1, 1.5, 2, 3, 4, and 5 are complete. The v1 closeout (2026-03-13) addressed remaining integration gaps. The sections below summarize the major shipped components across all phases.","**Orchestrator and mission runtime**:","- AI orchestrator service with mission lifecycle management, decomposed into modular architecture (core + 8 extracted modules)","- Orchestrator service decomposed (`orchestratorQueries.ts`, `stepPolicyResolver.ts` extracted)","- Built-in planning phase runtime (default-on profiles), clarification gating, and explicit coordinator phase transitions","- PR strategies (integration/per-lane/queue/manual) replacing merge phase","- Team synthesis and recovery loops","- Execution plan preview with approval gates","- Inter-agent messaging decomposed into `chatMessageService.ts` and `workerDeliveryService.ts`","- Mission chat workspace (`MissionChatV2`) with global summary, worker/orchestrator threads, mentions, and shared-renderer-backed detailed thread views","- Meta-reasoner with AI-driven fan-out dispatch (external_parallel, internal_parallel, hybrid)","- Context compaction engine (70% threshold, self-summarization, pre-compaction writeback)","- Session persistence via attempt_transcripts table and JSONL files","- Session resume via resumeUnified()","- Shared facts injection and run narrative generation","- Orchestrator Overhaul Phases 1-9 complete (reflection protocol, cross-mission trends, pattern-candidate promotion, adaptive runtime, UI overhaul)","- Coordinator finalization awareness: `check_finalization_status` tool + queue landing event routing","- Approval gates (`phase_approval` intervention type), mandatory planning enforcement, multi-round deliberation","- Adaptive runtime (`classifyTaskComplexity`, `scaleParallelismCap`, `evaluateModelDowngrade`)","- Budget-gated spawns (hard cap checks before every worker spawn)","- Benign error classification (`BENIGN_SANDBOX_BLOCK_PATTERNS` for ExitPlanMode/Zod noise)","- Phase 4 delegation contract: `delegate_parallel`, push sub-agent rollups, native teammate auto-registration + allocation caps","**Model and provider infrastructure**:","- Model registry unified: pricing fields in `ModelDescriptor`, `getModelPricing()`, `FAMILY_TO_CLI` map, `modelProfiles.ts` derived from registry","- Model registry expansion (50+ models across 10 provider families, auth-type classification, runtime enrichment via `enrichModelRegistry()`)","- Dynamic pricing via models.dev integration (`modelsDevService.ts`: fetch, 6h cache, fallback to hardcoded)","- Provider detection pipeline: `authDetector.ts` + `providerCredentialSources.ts` + `providerConnectionStatus.ts` + `providerRuntimeHealth.ts` + `claudeRuntimeProbe.ts` (structured per-provider connection status with auth, runtime, and usage availability)","- Provider options simplification (`providerOptions.ts`: pure tier-string passthrough, no invented token budgets)","- Reasoning tier standardization: Claude CLI low/medium/high, Claude API low/medium/high/max, Codex minimal/low/medium/high/xhigh","- UnifiedModelSelector redesign (auth-type grouping, hide unavailable models, \"Configure more...\" settings link)","- Universal tools for API-key and local models (`universalTools.ts`: permission modes plan/edit/full-auto)","- Workflow tools for chat agents (`workflowTools.ts`: lane creation, PR creation, screenshot capture, completion reporting, PR issue resolution tools)","- Three-tier tool architecture: universalTools (all agents) -> workflowTools (chat agents) -> coordinatorTools (orchestrator only)","- System prompt agent capability boundaries (tool tier guidance in agent prompts)","- Middleware layer (`middleware.ts`: logging, retry, cost guard, reasoning extraction)","**Memory and knowledge**:","- Unified memory system (W6): 3 scopes (project/agent/mission), 3 tiers (Tier 1 pinned / Tier 2 active / Tier 3 aging), candidate/promoted/archived lifecycle, auto-promotion, Settings > Memory tab","- Memory engine hardening (W6-half): lifecycle sweeps, batch consolidation, pre-compaction flush with quality criteria, Memory Health dashboard","- Embeddings pipeline (W7a): local Xenova/all-MiniLM-L6-v2, FTS4 BM25 + cosine similarity + MMR re-ranking","- Orchestrator memory wiring (W7b): mission-memory SSoT, exact employee L2 injection","- Skills and learning pipeline (W7c): episodic-to-procedural extraction, `.ade/skills/SKILL.md` materialization, skill ingestion from legacy sources, knowledge capture from failures/interventions/repeated errors/PR feedback, CTO memory review surfaces with provenance, confidence history, and re-index actions","- Memory tool wiring into agent coding tool set","- Memory pipeline fully wired: compaction flush into agentChatService, human work digest connected to git head watcher, failure knowledge capture on mission/agent errors, procedural learning export to `.ade/skills/`","- Embedding health monitoring with structured logging","**CTO and worker infrastructure** (Phase 4):","- CTO core identity (W1), worker org chart (W2), heartbeat and activation (W3)","- Bidirectional Linear sync (W4): `linearClient.ts`, `linearSyncService.ts`, `linearOutboundService.ts`, `linearRoutingService.ts`, `linearTemplateService.ts`, `linearCredentialService.ts`, `flowPolicyService.ts`, `linearCloseoutService.ts`, `linearDispatcherService.ts`, `linearIntakeService.ts`, `linearOAuthService.ts`, `linearWorkflowFileService.ts`, `issueTracker.ts` abstraction","- Linear dispatcher hardening (v1 closeout): snapshot refresh before step execution, employee fallback to `awaiting_delegation`, PR null-check for manual mode, closure notifications to agent chat sessions, dynamic delegation UI","- Automations platform (W5): `automationService.ts`, `automationPlannerService.ts`, `automationIngressService.ts`, `automationSecretService.ts`","- CTO + Org Experience Overhaul (W-UX): onboarding, activity, memory browser, and polish surfaces","- External MCP consumption (W8): ADE-managed external MCP registry/service, namespaced `ext.*` tool exposure","- OpenClaw bridge (W9)","- Portable `.ade/` state (W10): canonical tracked/shareable layout, startup repair, integrity normalization","**Mission UI and UX**:","- Mission phase engine + profiles: phase storage, profile CRUD/import/export, mission overrides, phase transition telemetry","- Mission UI overhaul: Plan/Work tabs, missions home dashboard, phase-aware details and launch/settings profile management","- Model selection per-mission with per-model thinking budgets","- Activity feed with category dropdown and run narrative","- missionId-filtered queries across all views","**Usage and budget**:","- Subscription usage tracking (`usageTrackingService.ts`): local CLI data analysis and cost scanning","- Budget cap service (`budgetCapService.ts`): mission and global budget enforcement","- Mission budget service with coordinator `get_budget_status`","**Codebase structure**:","- Orchestrator call types simplified from 6 to 2 (coordinator, chat_response)","- Type system modularized: `src/shared/types/` with 23 domain modules replacing monolithic `types.ts`","- Pack service decomposed: `projectPackBuilder.ts`, `missionPackBuilder.ts`, `conflictPackBuilder.ts`, `packUtils.ts` extracted","- Shared utilities consolidated: backend `utils.ts` (60+ duplicate removals), renderer `format.ts`/`shell.ts`/`sessions.ts`, shared React hooks","**Not yet shipped (v1 known limitations)**:","- Computer use runtime: The `localComputerUse.ts` capability detection module exists, screenshot capture is available as a workflow tool (depends on agent runtime support), and mission validation models screenshot/browser-verification/video evidence requirements, but the full `screenshot_environment` / `interact_gui` / `record_environment` MCP tool loop is not exposed end-to-end. Automatic PR proof embedding from computer-use artifacts is not shipped.","- Multi-device sync (cr-sqlite + WebSocket real-time replication) is Phase 6 work.","- Remote host deployment (user-owned VPS) is Phase 6 work.","- iOS companion app (core functionality) is Phase 6 work; advanced mobile features are Phase 7.","- Mission orchestration works end-to-end but complex multi-phase flows may benefit from human guidance via interventions.","### Compute Backends for Agent Execution","The older pluggable `ComputeBackend` abstraction is no longer part of the active ADE architecture. The current runtime model is:","- **Local host/runtime** (current baseline): agents execute as local subprocesses/worktree tasks on the active ADE machine.","- **User-owned VPS host** (planned Phase 6): ADE itself runs on a remote machine the user controls, and other devices connect to that host.","- **Dropped managed backend direction**: Daytona, E2B, and similar ADE-managed cloud backends are not part of the active roadmap.","### Compute Environment Types","ADE still distinguishes environment capabilities, but they are layered onto the runtime placement above rather than selected from a backend matrix:","```typescript","type ComputeEnvironmentType = 'terminal-only' | 'browser' | 'desktop';","interface ComputeEnvironment {"," type: ComputeEnvironmentType;"," display?: { // For browser and desktop environments"," width: number; // Default: 1920"," height: number; // Default: 1080"," colorDepth: number; // Default: 24"," };"," browserConfig?: { // For browser environments"," headless: boolean; // Default: true"," browser: 'chromium' | 'firefox'; // Default: 'chromium'"," desktopConfig?: { // For desktop environments"," windowManager: 'fluxbox' | 'xfce' | 'mutter'; // Default: 'fluxbox'"," vncEnabled: boolean; // Enable VNC for remote viewing. Default: true"," vncPort: number; // Default: 5901"," noVncPort: number; // Default: 6080 (browser-based VNC client)","**Terminal-only** (default): Agent gets a shell in a worktree/runtime sandbox. No GUI rendering. Suitable for code changes, test execution, and CLI operations.","**Browser**: Headless browser (Playwright/Puppeteer) available. Agent can launch web applications, navigate pages, interact with UI elements, and capture screenshots. Suitable for web application testing and visual verification.","**Desktop**: Full virtual desktop via Xvfb (X Virtual Framebuffer) + window manager. Agent gets programmatic mouse/keyboard control and screenshot/video capture. Suitable for desktop applications (Electron, native), mobile emulators, and GUI-heavy verification.","Implementation stack for desktop environments:","1. **Xvfb**: Virtual X11 display (e.g., `:99 -screen 0 1920x1080x24`)","2. **Window manager**: Fluxbox (lightweight) or Mutter (full-featured)","3. **VNC server**: x11vnc or TigerVNC for remote viewing","4. **noVNC + websockify**: Browser-based VNC client for web/mobile access","5. **xdotool**: Mouse/keyboard simulation for agent actions","6. **scrot/ImageMagick**: Screenshot capture","7. **ffmpeg**: Video recording via x11grab","Runtime capability notes:","| Runtime placement | terminal-only | browser | desktop |","|-------------------|:-------------:|:-------:|:-------:|","| Local host (current) | Yes | Yes (local Playwright) | Yes (local Xvfb) |","| User-owned VPS host (planned) | Planned | Planned | Planned |","### Per-Task-Type Configuration","ADE supports fine-grained control over which `modelId` handles each type of AI task.","#### Task Types","| Task Type | Description | Default Model ID |","|-----------|-------------|-----------------|","| `planning` | Mission decomposition into steps | `anthropic/claude-sonnet-4-6` |","| `implementation` | Code generation and modification | `openai/gpt-5.4-codex` |","| `review` | Code review and analysis | `anthropic/claude-sonnet-4-6` |","| `conflict_resolution` | Merge conflict analysis and resolution | `anthropic/claude-sonnet-4-6` |","| `narrative` | Lane narrative generation | `anthropic/claude-haiku-4-5` |","| `pr_description` | Pull request description drafting | `anthropic/claude-haiku-4-5` |","#### Configuration Schema","Per-task-type settings are stored in `.ade/local.yaml`:"," model: \"anthropic/claude-sonnet-4-6\""," model: \"anthropic/claude-haiku-4-5\"","#### Resolution Order","When determining which model to use for a task:","1. Explicit per-call or per-step `modelId` hint (highest priority).","2. Per-task-type `taskRouting..model` setting in `.ade/local.yaml`.","3. Mission-level model policy/overrides.","4. Built-in default model for the task type (as listed in the table above).","If the resolved `modelId` is missing or unknown, runtime startup fails with an explicit error instead of silently substituting a provider/model fallback.","### One-Shot AI Task Patterns","Most Phase 1-style tasks still appear one-shot to users, but now run through ephemeral task-agent runtimes so memory/guardrails/audit stay consistent.","#### Pattern: Ephemeral Task-Agent Runtime","// Narrative generation — one execution pass via an ephemeral task-agent runtime","async function generateNarrative(lanePack: LaneExportStandard): Promise {"," const runtime = await agentRuntimeService.invoke({"," source: \"narrative\","," executionClass: \"task\","," agentDefinitionId: \"ade.system.narrative\","," prompt: buildNarrativePrompt(lanePack),"," contextPack: lanePack,"," oneShot: true,"," });"," let result = \"\";"," for await (const event of runtime.events) {"," if (event.type === \"structured_output\") return event.data as string;"," if (event.type === \"text\") result += event.content;"," }"," return result;","#### One-Shot Tasks in Phase 1","| Task | Input | Output | Interaction |","|------|-------|--------|-------------|","| Narrative generation | `LaneExportStandard` | Markdown narrative | One-shot, no follow-up |","| Conflict proposals | `LaneExportLite` x 2 + `ConflictExportStandard` + user config | Resolution diff + explanation + confidence | One-shot, detailed context |","| PR descriptions | `LaneExportStandard` with commit history | PR title + body markdown | One-shot, no follow-up |","| Terminal summaries | Session transcript + metadata | Structured summary (intent, outcome, findings, next steps) | One-shot, no follow-up |","| Initial context generation | Repository scan results | PRD/architecture doc drafts | One-shot, no follow-up |","| Mission planning | Mission prompt + project context | Step plan JSON | Runtime-backed execution (planner may use multi-turn) |","These tasks can complete in one pass, but are still tracked as runtimes so they share the same policy and memory surface as missions.","#### CLI vs SDK Boundary","**SDK runtime path (programmatic, invisible to user)**: All non-interactive AI tasks listed above. The user never sees raw CLI output -- results are processed by ADE services and displayed in the appropriate UI surface.","**CLI (interactive, visible in Terminals tab)**: Only used when the user explicitly launches an AI terminal session from the Terminals tab or the Work Pane in the Lanes tab. The CLI runs in a PTY with full terminal interaction. This is the user's direct conversation with Claude/Codex, not ADE-orchestrated work.","This boundary is critical: SDK calls are ADE's internal tool; CLI sessions are the user's tool.","### Agent Chat Service (Phase 1.5)","The Agent Chat Service provides a native, interactive chat interface inside ADE — an alternative to using CLI terminals for working with Codex, Claude, and unified API/local model runtimes. It is a **provider-agnostic abstraction** that lets users chat with CLI or non-CLI models using the same UI.","> **External reference**: The Codex App Server protocol specification is at https://developers.openai.com/codex/app-server — this is the canonical reference for the CodexChatBackend implementation.","#### Why Agent Chat?","CLI terminals are powerful but opaque. The chat interface provides:","- **Structured item display**: File changes as inline diffs, command execution with live output, plans with step status — not raw terminal output.","- **Approval flow**: Accept/decline tool use with full context, not a yes/no prompt in a terminal.","- **Steering**: Inject instructions into an active turn without starting a new conversation.","- **Session persistence**: Resume conversations with full context, not just a command string.","- **Provider switching**: Same UI across providers/models. When switching model families mid-session, ADE forks a new chat session under the selected runtime/provider.","#### AgentChatService Interface","interface AgentChatService {"," createSession(laneId: string, provider: \"codex\" | \"claude\" | \"unified\", model: string, modelId: string, opts?: CreateSessionOpts): Promise;"," sendMessage(sessionId: string, text: string, attachments?: FileRef[]): AsyncIterable;"," steer(sessionId: string, text: string): Promise;"," cancelSteer(sessionId: string): Promise;"," editSteer(sessionId: string, text: string): Promise;"," interrupt(sessionId: string): Promise;"," resumeSession(sessionId: string): Promise;"," listSessions(laneId?: string): Promise;"," approveToolUse(sessionId: string, itemId: string, decision: ApprovalDecision): Promise;"," getAvailableModels(provider: \"codex\" | \"claude\" | \"unified\"): Promise;"," dispose(sessionId: string): Promise;","interface ChatSession {"," id: string;"," laneId: string;"," provider: \"codex\" | \"claude\" | \"unified\";"," model: string; // Human-readable display name"," modelId: string; // Registry model ID (required)"," // Provider-native permission controls (no unified permissionMode)"," claudePermissionMode?: \"default\" | \"plan\" | \"acceptEdits\" | \"bypassPermissions\";"," codexApprovalPolicy?: \"untrusted\" | \"on-request\" | \"on-failure\" | \"never\";"," codexSandbox?: \"read-only\" | \"workspace-write\" | \"danger-full-access\";"," codexConfigSource?: \"flags\" | \"config-toml\";"," unifiedPermissionMode?: \"plan\" | \"edit\" | \"full-auto\";"," status: \"active\" | \"idle\" | \"ended\";"," threadId?: string; // Codex: app-server thread ID"," createdAt: string;"," lastActivityAt: string;","type ChatEvent ="," | { type: \"text\"; content: string }"," | { type: \"tool_call\"; tool: string; args: unknown; itemId: string }"," | { type: \"tool_result\"; tool: string; result: unknown }"," | { type: \"file_change\"; path: string; diff: string; kind: \"create\" | \"modify\" | \"delete\" }"," | { type: \"command\"; cmd: string; cwd: string; output: string; exitCode?: number; status: \"running\" | \"completed\" | \"failed\" }"," | { type: \"plan\"; steps: Array<{ text: string; status: \"pending\" | \"in_progress\" | \"completed\" | \"failed\" }> }"," | { type: \"reasoning\"; summary: string; isCollapsed: boolean }"," | { type: \"approval_request\"; itemId: string; kind: \"command\" | \"file_change\"; description: string; detail: unknown }"," | { type: \"system_notice\"; noticeKind: \"auth\" | \"rate_limit\" | \"hook\" | \"file_persist\" | \"info\" | \"memory\" | \"provider_health\" | \"thread_error\"; message: string; detail?: string | AgentChatNoticeDetail }"," | { type: \"status\"; turnStatus: \"started\" | \"completed\" | \"interrupted\" | \"failed\"; error?: string }"," | { type: \"subagent_started\"; taskId: string; description: string; turnId?: string }"," | { type: \"subagent_result\"; taskId: string; status: \"completed\" | \"stopped\" | \"failed\"; summary: string; turnId?: string }"," | { type: \"error\"; message: string; errorInfo?: string }"," | { type: \"done\"; turnId: string };","type ApprovalDecision = \"accept\" | \"accept_for_session\" | \"decline\" | \"cancel\";","interface FileRef {"," path: string;"," type: \"file\" | \"image\";","interface ModelInfo {"," displayName: string;"," isDefault: boolean;"," reasoningEfforts?: Array<{ effort: string; description: string }>;","#### CodexChatBackend","The Codex backend uses the **Codex App Server** — a JSON-RPC 2.0 protocol designed for building custom Codex frontends. ADE spawns `codex app-server` as a child process and communicates via JSONL over stdin/stdout.","> **Protocol reference**: https://developers.openai.com/codex/app-server","> **Schema generation**: `codex app-server generate-ts --out ./schemas` produces TypeScript types matching the installed Codex version.","**Process lifecycle**:","import { spawn } from \"node:child_process\";","import readline from \"node:readline\";","// Spawn app-server as child process","const proc = spawn(\"codex\", [\"app-server\"], {"," stdio: [\"pipe\", \"pipe\", \"inherit\"], // stdin: pipe, stdout: pipe, stderr: inherit","});","// JSONL reader","const rl = readline.createInterface({ input: proc.stdout });","// Send JSON-RPC messages","const send = (msg: object) => proc.stdin.write(JSON.stringify(msg) + \"\\n\");","// Initialize handshake (required before any other messages)","send({ method: \"initialize\", id: 0, params: {"," clientInfo: { name: \"ade\", title: \"ADE\", version: ADE_VERSION },"," capabilities: { experimentalApi: true }","}});","send({ method: \"initialized\", params: {} });","**Protocol mapping to AgentChatService**:","| AgentChatService method | Codex App Server JSON-RPC | Notes |","|---|---|---|","| `createSession()` | `thread/start` | Params: `model`, `cwd` (lane worktree), `approvalPolicy`, `sandbox` |","| `sendMessage()` | `turn/start` | Input array: `[{ type: \"text\", text }, ...attachments]` |","| `steer()` | `turn/steer` | Appends to in-flight turn; cannot change model/sandbox |","| `cancelSteer()` | N/A | Cancel a pending steer before it is delivered |","| `editSteer()` | N/A | Replace pending steer text before it is delivered |","| `interrupt()` | `turn/interrupt` | Turn completes with `status: \"interrupted\"`. Idempotent — second call is a no-op. Claude/unified runtimes emit `subagent_result` \"stopped\" for active subagents. |","| `resumeSession()` | `thread/resume` | Params: `threadId`, optional `personality` |","| `listSessions()` | `thread/list` | Filter by `cwd` to scope to lane |","| `approveToolUse()` | Response to `requestApproval` | Payload: `accept`/`acceptForSession`/`decline`/`cancel` |","| `getAvailableModels()` | `model/list` | Returns models with `displayName`, `isDefault`, `reasoningEffort` |","| `dispose()` | Process cleanup | Close stdin, terminate child process |","**Notification → ChatEvent mapping**:","| Codex Notification | ChatEvent Type | Notes |","| `item/agentMessage/delta` | `text` | Append delta to current message |","| `item/started` (commandExecution) | `command` | Start tracking command output |","| `item/commandExecution/outputDelta` | `command` | Append to command output |","| `item/completed` (commandExecution) | `command` | Set exit code and status |","| `item/started` (fileChange) | `file_change` | File path, kind, diff |","| `item/commandExecution/requestApproval` | `approval_request` | Kind: \"command\" |","| `item/fileChange/requestApproval` | `approval_request` | Kind: \"file_change\" |","| `turn/plan/updated` | `plan` | Steps with status |","| `item/reasoning/summaryTextDelta` | `reasoning` | Append to summary |","| `turn/started` | `status` | `turnStatus: \"started\"` |","| `turn/completed` | `done` or `status` | Map `status` field |","**Sandbox policies** (configurable per session):","| Policy | JSON-RPC Value | Description |","| Read-only | `{ type: \"readOnly\" }` | Agent can read but not modify files |","| Workspace write | `{ type: \"workspaceWrite\", writableRoots: [cwd] }` | Agent can write within lane worktree |","| Full access | `{ type: \"externalSandbox\" }` | No restrictions |","**Error handling**: Codex errors include `codexErrorInfo` values (`ContextWindowExceeded`, `UsageLimitExceeded`, `HttpConnectionFailed`, etc.) which are mapped to user-facing messages in the chat UI.","#### ClaudeChatBackend","The Claude backend uses `ai-sdk-provider-claude-code` (the same community Vercel provider from Phase 1) in **multi-turn mode**. Instead of one-shot `execute()` calls, the Claude backend maintains a `messages[]` array and calls `streamText()` for each turn.","**Multi-turn conversation flow**:","import { streamText } from \"ai\";","import { claudeCode } from \"ai-sdk-provider-claude-code\";","// Session state (maintained per chat session)","const messages: CoreMessage[] = [];","// Each turn appends a user message and streams the response","async function* sendMessage(text: string): AsyncIterable {"," messages.push({ role: \"user\", content: text });"," const stream = streamText({"," model: claudeCode(\"sonnet\", {"," claudePermissionMode: \"acceptEdits\","," maxBudgetUsd: 5.0,"," systemPrompt: \"You are working in an ADE lane...\","," canUseTool: async (invocation) => {"," // Emit approval_request ChatEvent"," // Wait for user decision"," // Return allow/deny"," },"," }),"," messages,"," for await (const chunk of stream) {"," if (chunk.type === \"text-delta\") {"," yield { type: \"text\", content: chunk.textDelta };"," }"," // Map other chunk types to ChatEvent..."," // Append assistant response to messages for context continuity"," messages.push({ role: \"assistant\", content: stream.text });","**Key differences from CodexChatBackend**:","| Capability | Codex (App Server) | Claude (Community Provider) |","| Protocol | JSON-RPC 2.0 over stdio | Vercel AI SDK `streamText()` |","| Multi-turn | Native thread management | ADE manages `messages[]` |","| Steering | `turn/steer` (native) | Queue text, inject on next turn |","| Approval flow | `requestApproval` notifications | `canUseTool` callback |","| File changes | `fileChange` items with diffs | Tool call events (Write, Edit) |","| Command execution | `commandExecution` items with live output | Tool call events (Bash) |","| Plans | `turn/plan/updated` | Not natively supported |","| Reasoning | `reasoning` items | Not exposed by community provider |","| Session persistence | App server manages on disk | ADE persists `messages[]` to JSON |","| Resume | `thread/resume` (native) | Reload `messages[]` and continue |","**Session persistence**: Chat session metadata is stored at `.ade/cache/chat-sessions/.json`, and structured chat events are logged to `.ade/transcripts/.chat.jsonl` with a mirrored JSONL copy under `.ade/transcripts/chat/.jsonl`. This enables resume after app restart while keeping chat-specific runtime files in the canonical W10 layout.","**Image attachments**: Claude V2 sessions now support inline image content blocks. The `buildClaudeV2Message()` helper checks attached files against the Anthropic-accepted MIME types (`image/jpeg`, `image/png`, `image/gif`, `image/webp`), reads them as base64, and builds an `SDKUserMessage` with interleaved text and image content blocks. When no image attachments are present, the message is sent as a plain string. Images are saved to a temporary location via the `saveTempAttachment` IPC handler before being read by the service.","**Limitations**: The Claude backend may not support all UI features that Codex provides (plans, reasoning blocks). The chat UI gracefully handles missing features — items that Claude doesn't produce simply don't appear in the UI.","#### Phase 2 Chat Improvements","Phase 2 completed the outstanding chat debt from Phase 1.5:","- **UI polish shipped**: The Work Pane chat surface (`AgentChatMessageList.tsx`, `AgentChatComposer.tsx`, `AgentChatPane.tsx`) now uses richer bubble styling, inline diff emphasis, cleaner command blocks, improved streaming indicators, and clearer approval presentation.","- **Claude provider selection fixed**: Provider selection no longer resets unexpectedly; Claude and Codex remain selectable based on detected model availability.","- **Reasoning effort selector shipped**: Codex reasoning effort (`minimal`, `low`, `medium`, `high`, `xhigh`) is surfaced in the composer and passed to both `thread/start` and `turn/start`. Last-used effort is persisted per lane/model. Claude model variants are shown with descriptive labels from `supportedModels()`.","#### Chat Session Lifecycle","Agent chat sessions integrate into ADE's existing session tracking infrastructure:","1. User opens Chat view in Work Pane"," → agentChatService.createSession(laneId, provider, model)"," → Creates terminal_sessions row (tool_type: \"codex-chat\", \"claude-chat\", or \"ai-chat\")"," → Codex: spawns app-server + thread/start"," → Claude: initializes messages[] + session state"," → Unified: resolves configured API/local model via provider resolver + initializes universal tool runtime"," → Captures head_sha_start","2. User sends messages, agent works"," → agentChatService.sendMessage() yields ChatEvent stream"," → Events rendered in AgentChatMessageList"," → Chat events logged to `.ade/transcripts/.chat.jsonl` and mirrored to `.ade/transcripts/chat/.jsonl`"," → File changes update the lane worktree (visible in git actions)","3. Approvals"," → ChatEvent.approval_request rendered as overlay"," → User decision sent back via agentChatService.approveToolUse()","4. Session end (user closes chat or navigates away)"," → agentChatService.dispose()"," → Codex: thread remains on disk (can resume)"," → Claude: messages[] persisted to JSON (can resume)"," → Captures head_sha_end"," → Session delta computed (same as terminal sessions)"," → onSessionEnded callback fires → job engine, pack refresh, agents","5. Resume"," → User clicks resume in Terminals tab or reopens Chat view"," → agentChatService.resumeSession()"," → Codex: thread/resume with stored threadId"," → Claude: reload messages[] from JSON"," → Unified: re-resolve model + reload persisted message history + permission mode","This lifecycle mirrors the PTY session lifecycle exactly, ensuring that chat sessions produce the same context artifacts (deltas, memory updates, checkpoints) as terminal sessions.","## Integration Points","### Desktop Application","- **AI integration service**: `apps/desktop/src/main/services/ai/aiIntegrationService.ts` -- provider detection, task routing, executor dispatch via `AgentExecutor`, streaming response handling.","- **Orchestrator service**: `apps/desktop/src/main/services/orchestrator/orchestratorService.ts` (~11K lines) + `orchestratorQueries.ts`, `stepPolicyResolver.ts` -- run/step/attempt state machine, claim management, context snapshots, gate reports.","- **AI orchestrator service**: `apps/desktop/src/main/services/orchestrator/aiOrchestratorService.ts` (~9.9K lines) + 8 extracted modules (`chatMessageService.ts`, `workerDeliveryService.ts`, `workerTracking.ts`, `missionLifecycle.ts`, `recoveryService.ts`, `modelConfigResolver.ts`, `orchestratorContext.ts`, `orchestratorConstants.ts`) -- AI coordination layer: autopilot, worker management, messaging, recovery.","- **Agent chat service**: `apps/desktop/src/main/services/chat/agentChatService.ts` -- manages chat session lifecycle, spawns Codex app-server processes and Claude multi-turn sessions, maps provider events to ChatEvent streams, integrates with session tracking.","- **Unified memory service**: `apps/desktop/src/main/services/memory/memoryService.ts` -- memory retrieval, candidate/promoted lifecycle, and budgeted context assembly.","- **Bounded memory budgets**: `memoryGetBudget` (`lite`/`standard`/`deep`) controls AI context size in runtime prompt assembly.","- **Model system**: `apps/desktop/src/shared/modelRegistry.ts` (pricing-aware descriptors, `FAMILY_TO_CLI` map) + `modelProfiles.ts` (derived from registry).","- **Shared types**: `apps/desktop/src/shared/types/` -- 23 domain-scoped type modules (core, lanes, conflicts, prs, git, files, sessions, chat, missions, orchestrator, config, automations, packs, budget, models, usage) with barrel `index.ts`.","- **Configuration**: Provider settings read from `projectConfigService.ts` (merged shared + local config).","- **IPC channels**: `ade.ai.*` for AI streaming, `ade.missions.*` for mission lifecycle, `ade.orchestrator.*` for run management.","### Job Engine","The job engine handles background AI tasks that are triggered by system events:","- **Auto-narrative generation**: After lane/session context updates, the job engine optionally triggers narrative generation via the AI integration service if a CLI subscription is available. This is a non-blocking async flow that does not interfere with the user's interactive workflow.","- **Conflict proposal generation**: When conflict prediction detects new or changed conflicts, the job engine can trigger AI-powered conflict resolution proposals.","The job engine does **not** coordinate orchestrator step transitions. The orchestrator service has its own tick-based scheduler for mission execution.","### Mission Service","The mission service (`missionService.ts`) provides the user-facing lifecycle for AI-driven work:","- **Mission creation**: Accepts a plain-English prompt, title, lane assignment, phase profile/override, and execution policy.","- **Phase selection**: Resolves selected phases per mission/profile; built-ins include planning first by default.","- **Step tracking**: Tracks mission steps and interventions while run-time orchestration remains coordinator-owned.","- **Phase pipeline contracts (Task 3)**: Resolves mission phase profile/override, persists phase configuration, and annotates mission steps with phase identity metadata.","- **Phase transition audit (Task 3)**: Runtime phase changes emit durable `phase_transition` mission/timeline events and update run metadata (`phaseRuntime`) for operator inspection.","- **Profile lifecycle APIs (Task 3)**: list/save/delete/clone/import/export/getPhaseConfiguration/getDashboard contracts are exposed to renderer and automation surfaces.","- **Intervention management**: Creates, resolves, and dismisses intervention records when AI agents or the orchestrator need human input.","- **Artifact collection**: Links mission outcomes (PR URLs, generated files, test results) as artifacts.","### Memory Service","W6 is complete with unified memory as the primary AI context backbone in renderer/main flows. Pack-first renderer surfaces have been cut over to memory-first equivalents, and the remaining pack-shaped APIs are compatibility veneers over live exports rather than runtime prerequisites.","- **Budgeted retrieval**: `lite` / `standard` / `deep` memory budgets bound context size for each AI task type.","- **Candidate workflow**: Discoveries enter as candidates and can be promoted or archived via memory APIs.","- **Promoted retrieval**: Prompt assembly consumes promoted entries by ranking and budget.","- **Auditability**: Entries preserve confidence, recency, and access metadata for deterministic ranking.","- **Compatibility inventory**: Remaining pack internals are kept only where they still provide explicit compatibility or audit value.","### Skills + Learning Pipeline","> **Replaces**: The earlier \"Learning Packs\" concept (separate pack type with `LearningEntry` schema). Now unified into W7 as procedural memory extraction on top of W6's unified memory. See `docs/final-plan/phase-4.md` W7 for full spec.","W7 builds an extraction and materialization layer on top of the Unified Memory System (W6). It turns accumulated mission experience into reusable skills that any agent (Claude, Codex, or any future adapter) can consume.","**Procedural extraction**: When the same pattern appears across 3+ episodic summaries (from different missions/sessions), the system extracts it as a `ProceduralMemory` entry (trigger, procedure, confidence, success/failure counts). Stored in project memory with `category: \"procedure\"`.","**Knowledge sources** (automatic capture into project memory, feeding the episodic → procedural pipeline):","- Mission failures and resolutions (captured as gotcha/pattern entries)","- User interventions during missions (inferred as candidate conventions)","- Repeated errors across 3+ sessions (recorded as gotcha patterns with file scope)","- PR review feedback patterns","**Skill materialization**: Confirmed procedural memories are exported as `.ade/skills//SKILL.md` — universal markdown format consumable by any agent. Skills are also indexed back into project memory.","**Skill ingestion**: On startup/file change, ADE scans `.ade/skills/` plus legacy external sources (`.claude/skills/`, `.claude/commands/`, `CLAUDE.md`, `agents.md`) and indexes them into project memory as Tier 2 procedure entries.","## Implementation Status","| Component | Status | Notes |","|-----------|--------|-------|","| Architecture design | Complete | Documented in this file |","| Planning phase (coordinator-owned) | Complete | Built-in `planning` phase is default-on and transitions via `set_current_phase` |","| Coordinator-strategy deterministic fallback (runtime) | Removed | Coordinator owns strategy; unavailable coordinator pauses/escalates instead of deterministic replacement |","| Orchestrator state machine | Complete | `orchestratorService.ts` (~11K lines) + `orchestratorQueries.ts`, `stepPolicyResolver.ts` -- runs, steps, attempts, claims, gates, timeline |","| Executor adapter interface | Complete | `OrchestratorExecutorAdapter` type for pluggable step execution |","| Context snapshot system | Complete | Profile-based export assembly (deterministic, narrative-opt-in) |","| Bounded memory budgets | Complete | Lite/Standard/Deep retrieval tiers via memory APIs |","| AgentExecutor interface | Complete | `apps/desktop/src/main/services/ai/agentExecutor.ts` |","| Agent SDK integration (dual-SDK) | Complete | Unified executor runtime (formerly `ClaudeExecutor` + `CodexExecutor`) |","| AI integration service | Complete | `apps/desktop/src/main/services/ai/aiIntegrationService.ts` |","| Per-task-type configuration | Complete | Configurable in `.ade/local.yaml` |","| Streaming AI responses to UI | Complete | IPC push events via `webContents.send` |","| AgentChatService interface | Complete | `apps/desktop/src/main/services/chat/agentChatService.ts` |","| CodexChatBackend (App Server) | Complete | JSON-RPC 2.0 client in `agentChatService.ts` |","| ClaudeChatBackend (community provider) | Complete | Multi-turn `streamText()` in `agentChatService.ts` |","| Chat UI components | Complete | AgentChatPane, AgentChatMessageList, AgentChatComposer |","| Chat session integration | Complete | `codex-chat`, `claude-chat`, and `ai-chat` tool types in `terminal_sessions` |","| MCP server (`apps/mcp-server`) | Complete | JSON-RPC 2.0 server with 35 tools, dual-mode architecture (headless + embedded) |","| MCP dual-mode architecture | Complete | Transport abstraction (stdio/socket), headless AI via aiIntegrationService, desktop socket embedding (.ade/mcp.sock), smart entry point auto-detection. Centralized launch resolution (`adeMcpLaunch.ts`) with bundled proxy mode for packaged builds. |","| AI orchestrator (Claude + MCP) | Complete | Tasks 1-7 shipped; Orchestrator Overhaul Phases 1-9 complete (reflection protocol, adaptive runtime, UI overhaul). V1 closeout: coordinator finalization awareness. M4/M5 additions: approval gates, mandatory planning enforcement, multi-round deliberation, adaptive runtime, model downgrade, budget-gated spawns, benign error classification. |","| Phase 4 orchestrator delegation/team runtime | Complete | `delegate_parallel`, push sub-agent progress/completion rollups, native teammate auto-registration + allocation cap guardrails, single team-member data path |","| Adaptive Runtime (M5) | Complete | `adaptiveRuntime.ts` — `classifyTaskComplexity`, `scaleParallelismCap`, `evaluateModelDowngrade`; budget hard cap enforcement in coordinator tools |","| Approval Gates (M4/M5) | Complete | `phase_approval` intervention type, `requiresApproval` on PhaseCard, blocking phase transitions until user approval |","| Mandatory Planning Enforcement (M4/M5) | Complete | `CoordinatorAgent` constructor injects planning phase if missing; first-turn planning watchdog force-spawns planner |","| Multi-Round Deliberation (M4/M5) | Complete | `canLoop`/`loopTarget` on PhaseCardOrderingConstraints, `maxQuestions` bounds per phase |","| Error Classification Hardening (M4/M5) | Complete | `BENIGN_SANDBOX_BLOCK_PATTERNS` for ExitPlanMode/Zod errors, `classifyBlockingWarnings` in `orchestratorQueries.ts` |","| Mission phase engine + profiles (Task 3) | Complete | `phase_cards`/`phase_profiles`/`mission_phase_overrides`, profile CRUD/import/export, phase transition telemetry |","| Mission UI overhaul (Task 4) | Complete | Plan/Work tabs, mission home dashboard, phase-aware details, launch/settings profile workflows |","| Pre-flight + intervention/HITL (Task 5) | Complete | Launch-gate checklist, granular worker-level interventions, coordinator `ask_user`/`request_user_input` escalation wiring |","| Budget + usage tracking (Task 6) | Complete | Mission budget service, subscription/API-key accounting, coordinator `get_budget_status`, details-tab budget telemetry |","| Model registry (50+ models, runtime enrichment) | Complete | `modelRegistry.ts` -- 10 provider families, auth-type classification, pricing in `ModelDescriptor`, `getModelPricing()`, `FAMILY_TO_CLI` map, `enrichModelRegistry()` for models.dev data, `resolveModelDescriptorForProvider()`, `getRuntimeModelRefForDescriptor()`. `modelProfiles.ts` derived from registry. |","| Dynamic pricing (models.dev) | Complete | `modelsDevService.ts` -- fetch/cache/fallback, 6h refresh, Proxy-based `MODEL_PRICING` |","| Provider options (tier passthrough) | Complete | `providerOptions.ts` -- pure tier-string passthrough per provider family, no arbitrary token budgets |","| Middleware layer | Complete | `middleware.ts` -- logging, retry, cost guard, reasoning extraction |","| Universal tools (API-key/local) | Complete | `universalTools.ts` -- permission modes (plan/edit/full-auto), approval hooks |","| Type system modularization | Complete | `src/shared/types/` -- 23 domain modules replacing monolithic `types.ts` |","| Pack service decomposition | Complete | `packService.ts` + `projectPackBuilder.ts`, `missionPackBuilder.ts`, `conflictPackBuilder.ts`, `packUtils.ts` |","| Shared utilities consolidation | Complete | Backend `utils.ts` (60+ duplicates removed), renderer `format.ts`/`shell.ts`/`sessions.ts`, shared React hooks (`useClickOutside`, `useThreadEventRefresh`) |","| Agent-first runtime baseline | Complete | Non-interactive AI call paths execute on runtime records with no legacy compatibility migration path |","| Call audit logging | Complete | Every MCP tool invocation writes durable `mcp_tool_call` history records |","| Permission/policy layer | Complete | Mutation tools enforce claim/identity policy; spawn and ask_user guards applied |","| Chat reasoning effort (Claude) | Complete | Reasoning effort forwarded to Claude provider when supported; validated for Codex |","| Local runtime placement | Complete | Agents run on local runtime records/worktrees on the active ADE machine |","| Remote host deployment | Planned | Phase 6 -- user-owned VPS host + device routing |","| Managed cloud compute backends (Daytona/E2B) | Dropped | Not on active roadmap |","| Compute environment types | Partially implemented | terminal-only is active; browser and desktop environment types are defined but the full runtime loop is not shipped |","| Computer use MCP tools | Partially implemented | Capability detection exists (`localComputerUse.ts`); mission validation models evidence requirements; full MCP tool loop (`screenshot_environment`, `interact_gui`, `record_environment`, `launch_app`, `get_environment_info`) and automatic PR proof embedding are not shipped end-to-end |","| Unified Memory System (W6) | Complete | Unified memory retrieval and renderer cutover are active |","| Memory Engine Hardening (W6½) — lifecycle sweeps, batch consolidation, pre-compaction flush | Complete | Temporal decay, tier demotion, hard limits, orphan cleanup, Jaccard+LLM consolidation, pre-compaction flush, Memory Health dashboard |","| Embeddings Pipeline (W7a) — local embedding + hybrid retrieval | Complete | Local Xenova/all-MiniLM-L6-v2 via @huggingface/transformers, FTS4 BM25 + cosine similarity + MMR re-ranking, graceful lexical fallback |","| Orchestrator Memory Wiring (W7b) — mission-memory SSoT + exact employee L2 injection | Complete | Retired `orchestrator_shared_facts`, compaction writes back to mission memory, worker briefings derive shared team knowledge from mission memory, exact `employeeAgentId` launch metadata controls agent-memory injection |","| Skills + Learning Pipeline (W7) — procedural extraction + skill materialization | Complete | Phase 4 — episodic-to-procedural extraction, `.ade/skills/SKILL.md` materialization, skill ingestion from legacy sources, and knowledge capture from failures/interventions/repeated errors/PR feedback are shipped (`knowledgeCaptureService.ts`, `proceduralLearningService.ts`, `skillRegistryService.ts`) |","| CTO Agent — core identity, persistent chat, core memory (W1) | Complete | Phase 4 -- `ctoStateService.ts`, dual-canonical persistence (DB + file), session reconstruction, CtoPage with chat |","| Worker Agents — org chart, multi-adapter, config versioning, budget (W2) | Complete | Phase 4 -- `workerAgentService.ts`, `workerRevisionService.ts`, `workerBudgetService.ts`, `workerTaskSessionService.ts`, `workerAdapterRuntimeService.ts` |","| Heartbeat & Activation — timer pool, coalescing, orphan reaping (W3) | Complete | Phase 4 -- `workerHeartbeatService.ts` (789 lines), two-tier execution, deferred promotion, issue locking |","| Bidirectional Linear Sync (W4) | Complete | Phase 4 -- `linearClient.ts`, `linearSyncService.ts`, `linearOutboundService.ts`, `linearRoutingService.ts`, `linearTemplateService.ts`, `linearCredentialService.ts`, `flowPolicyService.ts`, `linearCloseoutService.ts`, `linearDispatcherService.ts`, `linearIntakeService.ts`, `linearOAuthService.ts`, `linearWorkflowFileService.ts`, `issueTracker.ts` abstraction, `LinearSyncPanel.tsx` UI with 6-step flow composer |","| External MCP consumption | Complete | Phase 4 -- ADE-managed external MCP registry/service, namespaced `ext.*` tool exposure through ADE MCP, mission/worker/CTO policy integration |","| `.ade/` portable state | Complete | Phase 4 -- canonical tracked/shareable layout, startup repair, integrity normalization, config reload, Settings > Project health surface |","| Task agents (lane artifacts) | Planned | Phase 4 -- specialized agents for artifact production within lanes |","| Chat-to-mission escalation | Planned | Phase 4 -- promote a chat conversation into a full mission with pre-filled context |","**Overall status**: Phases 1, 1.5, 2, 3, 4, and 5 are complete. All Phase 4 workstreams (W1-W10, W-UX, W6-half, W7a-c) are shipped at baseline or better. The remaining unshipped work is concentrated in computer-use runtime follow-through (the MCP tool loop for `screenshot_environment`/`interact_gui`/`record_environment` and automatic PR proof embedding) and future-phase items (multi-device sync, remote host deployment, iOS companion). MCP dual-mode architecture is shipped, enabling headless operation with full AI via `aiIntegrationService` and embedded proxy mode through the desktop socket at `.ade/mcp.sock`. Packaged macOS builds use a bundled MCP proxy (`adeMcpProxy.cjs`) that connects to the desktop socket, with a runtime smoke test (`packagedRuntimeSmoke.ts`) that validates PTY, Claude SDK, and MCP proxy availability at build time.","## MCP Server as External Orchestration API","The MCP server (`apps/mcp-server`) has been overhauled from a 16-tool agent interface into a full **headless orchestration API** with 35 tools. This enables external consumers -- Claude Code, CI/CD pipelines, evaluation harnesses, and custom scripts -- to create, drive, observe, and evaluate missions without the desktop UI.","**Important architectural distinction**: The MCP server is the external/headless ADE tool surface and the common ADE tool bridge for spawned workers. The coordinator runtime itself is phase-aware and provider-routed; regardless of provider, it is constrained to the ADE coordinator tool surface rather than unconstrained repo exploration. The MCP server is the external-facing tool surface for:","1. **Spawned worker agents** -- agents launched by the orchestrator that need to interact with ADE's lane, git, and context systems.","2. **External observers** -- tools like Claude Code that want to monitor mission progress without participating.","3. **Evaluators** -- automated or human evaluation workflows that score completed mission runs.","### Tool Categories (35 Total)","#### Existing Tools (16)","| Tool | Description |","|------|-------------|","| `spawn_agent` | Launch a new AI agent in a specified lane |","| `read_context` | Read pack exports, lane state, or project context |","| `create_lane` | Create a new lane with a worktree for agent work |","| `check_conflicts` | Run conflict prediction against other active lanes |","| `merge_lane` | Merge a lane back to its parent |","| `ask_user` | Route an intervention to the ADE UI for human input |","| `run_tests` | Execute test suites in a lane's worktree |","| `get_lane_status` | Get current status of a specific lane |","| `list_lanes` | List all active lanes with summary status |","| `commit_changes` | Stage and commit changes in a lane |","| `get_project_info` | Get project metadata and configuration |","| `list_files` | List files in a lane or project directory |","| `read_file` | Read file contents |","| `write_file` | Write file contents |","| `search_code` | Search code across the project |","| `get_git_log` | Get git log for a lane or branch |","#### Mission Lifecycle Tools (7)","| `create_mission` | Create a new mission from a prompt |","| `start_mission` | Start planning and execution of a mission |","| `pause_mission` | Pause a running mission |","| `resume_mission` | Resume a paused mission |","| `cancel_mission` | Cancel a mission |","| `steer_mission` | Send a steering message to adjust mission direction |","| `resolve_intervention` | Resolve a pending intervention |","#### Observation Tools (8)","| `get_mission` | Get full mission state including steps, artifacts, and interventions |","| `get_run_graph` | Get the DAG visualization data for a mission run |","| `stream_events` | Poll for new mission/orchestrator events since a cursor |","| `get_step_output` | Get the output/transcript of a specific step attempt |","| `get_worker_states` | Get current state of all worker agents in a mission |","| `get_timeline` | Get the full timeline of events for a mission run |","| `get_mission_metrics` | Get usage metrics (tokens, cost, duration) for a mission |","| `get_final_diff` | Get the combined diff of all changes made during a mission |","#### Evaluation Tools (3)","| `evaluate_run` | Score a completed mission run with structured evaluation criteria |","| `list_evaluations` | List all evaluations for a mission |","| `get_evaluation_report` | Get a detailed evaluation report with scores and commentary |","## Evaluator Workflow","The MCP server supports a complete **create, start, stream, evaluate** loop for automated mission evaluation. This enables external tools (Claude Code, CI pipelines, test harnesses) to drive missions end-to-end and score the results.","### Workflow Steps","1. **Connect as evaluator role**: Connect to the MCP server via stdio. The connecting process acts as an external observer.","2. **Create mission**: Use `create_mission` with a prompt describing the desired work."," ```"," create_mission({ prompt: \"Implement rate limiting middleware for the API\", title: \"Rate Limiter\" })","3. **Start mission**: Trigger planning and execution."," start_mission({ missionId: \"\" })","4. **Poll for progress**: Use `stream_events` with a cursor to receive incremental updates."," stream_events({ missionId: \"\", cursor: 0 })"," Returns events like `step_started`, `step_completed`, `agent_spawned`, `worker_status_reported`, `intervention_requested`, etc. The cursor advances with each poll.","5. **Inspect with observation tools**: Use `get_worker_states`, `get_step_output`, `get_timeline`, and `get_mission_metrics` to inspect mission state at any point.","6. **Evaluate the run**: After the mission completes (or fails), use `evaluate_run` to score the result."," evaluate_run({"," missionId: \"\","," scores: { correctness: 0.9, completeness: 0.8, code_quality: 0.85 },"," commentary: \"Rate limiter implementation is correct but missing Redis backend support.\","," verdict: \"pass\""," })","7. **Retrieve evaluation reports**: Use `get_evaluation_report` for detailed scoring breakdowns.","## Claude Code Integration","The MCP server can be used directly from Claude Code as an MCP tool provider, enabling Claude Code to create and manage ADE missions conversationally.","### MCP Configuration","Add to your Claude Code MCP configuration (`~/.claude/mcp.json` or project-level):"," \"mcpServers\": {"," \"ade\": {"," \"command\": \"npx\","," \"args\": ["," \"tsx\","," \"/path/to/ADE/apps/mcp-server/src/index.ts\","," \"--project-root\", \"/path/to/target-repo\""," ]","### Launch","```bash","claude --permission-mode plan","Use a permission mode that matches the ADE workflow you are exposing. Read-only observation and planning flows can stay in `plan`; mutating ADE tool workflows may require a less restrictive mode depending on the exact ADE tool surface you enable.","### Usage","Once configured, Claude Code can use ADE tools naturally:","- \"Create a mission to refactor the authentication module\"","- \"Show me the current mission status\"","- \"What are the worker agents doing?\"","- \"Evaluate the last mission run\"","Claude Code will invoke the appropriate MCP tools (`create_mission`, `get_mission`, `get_worker_states`, `evaluate_run`) based on the conversational context.","## Known Limitations","### 1. ~~No AI Planning in MCP Mode~~ (Resolved — Dual-Mode Architecture)","The MCP server now operates in **dual mode**:","- **Headless mode** (stdio transport): The MCP server runs standalone with full AI capabilities. `aiIntegrationService` is wired in during bootstrap, auto-detecting `ANTHROPIC_API_KEY`, `claude` CLI, or other providers. AI-powered planning, the meta-reasoner, and all 35 tools are available.","- **Embedded mode** (socket transport): The desktop app embeds the MCP server at `.ade/mcp.sock`, sharing the same service instances. External agents connect via the socket and proxy through the desktop for live UI updates.","A `JsonRpcTransport` abstraction layer supports both stdio and Unix socket transports. The smart entry point auto-detects whether `.ade/mcp.sock` exists: if the desktop is running, the server connects as an embedded proxy; otherwise it starts in headless mode with its own AI backend.","### 2. Limited Agent Chat in Headless Mode","Headless MCP mode now provisions a lightweight `agentChatService` through `createHeadlessLinearServices()`. This service supports employee-session targets for Linear workflow dispatch (creating reusable continuity chats), but it is a manual shell -- no live agent runtime is attached. Worker-backed targets fail fast with explicit run errors when no worker runtime is available. Full interactive chat participation still requires the desktop runtime. Mission progress can be observed via:","- `stream_events` (event polling)","- `get_worker_states` (agent state snapshots)","- `get_step_output` (step transcripts)","- `get_timeline` (full event timeline)","### 3. Event Buffer is In-Memory (Headless Mode)","The `stream_events` tool uses an in-memory event buffer with a **10,000 event cap**. Events are lost on process restart. For durable event history, use `get_timeline` which reads from the SQLite database. The in-memory buffer is designed for real-time polling during active mission execution, not long-term storage.","### 3. Single Process Database Access","ADE uses SQLite (via node:sqlite) with a single-writer model. If the desktop app and MCP server run simultaneously against the same project database, **SQLite write conflicts** will occur. To avoid this:","- Stop the desktop app before running the MCP server standalone in headless mode, or","- Use the MCP server in embedded mode (via `.ade/mcp.sock`), which shares the same database connection as the desktop app.",[2541,2542,2543],{"_1067":45,"_45":2548},{"_1067":2545,"_156":1109,"_1115":2546,"_1116":2547},{"_1067":45,"_45":2544},"\n* Committed the resolved merge as `a6a7929` on branch `work`.\n* Created a follow-up PR via `make_pr` with title **“Follow-up: resolve merge conflicts for cancelling chats and subagents”**.\n\n**Testing**\n* ✅ `git merge --no-commit --no-ff d457c0b1e9df3cd76c981ab772314f5c91d911fe` (triggered merge and surfaced conflicts to resolve)\n* ✅ `rg -n \"^<<<<<<<|^=======|^>>>>>>>\"` (confirmed no conflict markers remain)\n* ✅ `git diff --name-only --diff-filter=U` (confirmed no unmerged files remain)\n* ✅ `git commit -m \"Merge base updates and resolve AI integration doc conflict\"`","repo_file_citation",1578,1586,"### Summary\n* Merged base commit `d457c0b1e9df3cd76c981ab772314f5c91d911fe` into the current lane and resolved the only merge conflict in `docs/architecture/AI_INTEGRATION.md`. The conflict resolution keeps both steer-management entries (`cancelSteer`, `editSteer`) and the newer interrupt semantics (idempotent interrupt + subagent stop note).",{"_91":309,"_41":83,"_95":309,"_2551":2552,"_2553":2554,"_2555":2556,"_131":-5,"_118":2557},[82],"input_items",[2574,2575,2576,2577],"external_post_config",{"_2558":2559,"_247":182,"_2560":288,"_258":-5,"_1088":241,"_2561":-5,"_2562":2563,"_2564":2565,"_2566":-5,"_2567":-5,"_2568":2569,"_2570":15,"_2571":57,"_2572":52,"_2573":15},"user_model_slug","slug",1774970785.413705,"installation_id",99878680,"pr_number","review_trigger","comment_id",4163487910,"comment_type","issue","pr_write_guard_token","pending_pr_reaction_id","pending_comment_reaction_id",342976762,"comment_if_no_findings","comment_on_error","share_url","is_parallel_review",{"_91":457,"_95":309,"_452":2586},{"_295":296,"_297":185,"_2583":302,"_2584":128,"_305":-5,"_91":2585},{"_91":2578,"_297":2582,"_156":-5,"_2580":-5,"_2581":-5},{"_91":2578,"_297":2579,"_156":-5,"_2580":-5,"_2581":-5},"pull_request_comment","\n\n\n> [!WARNING]\n> ## Rate limit exceeded\n> \n> `@arul28` has exceeded the limit for the number of commits that can be reviewed per hour. Please wait **1 minutes and 17 seconds** before requesting another review.\n> \n> Your organization is not enrolled in usage-based pricing. Contact your admin to enable usage-based pricing to continue reviews beyond the rate limit, or try again in **1 minutes and 17 seconds**.\n> \n>

\n>

⌛ How to resolve this issue?

\n> \n> \n>

\n>

🚦 How do rate limits work?

\n> \n>

\n>

ℹ️ Review info

\n> \n>

\n>

⚙️ Run configuration

\n> \n> **Configuration used**: Path: .coderabbit.yaml\n> \n> **Review profile**: CHILL\n> \n> **Plan**: Pro\n> \n> **Run ID**: `070b8eb6-b03d-4dbd-9f5a-937ea113dfe0`\n> \n>

\n> \n>

\n>

📥 Commits

\n> \n> Reviewing files that changed from the base of the PR and between d457c0b1e9df3cd76c981ab772314f5c91d911fe and 3887a4b856109cc67f391b0adb68148c8aa18c06.\n> \n>

\n> \n>

\n>

⛔ Files ignored due to path filters (2)

\n> \n> * `docs/architecture/AI_INTEGRATION.md` is excluded by `!docs/**`\n> * `docs/features/CHAT.md` is excluded by `!docs/**`\n> \n>

\n> \n>

\n>

📒 Files selected for processing (2)

\n> \n> * `apps/desktop/src/main/services/chat/agentChatService.test.ts`\n> * `apps/desktop/src/main/services/chat/agentChatService.ts`\n> \n>

\n> \n>

\n\n\n\n\n\n\n

✨ Finishing Touches

\n\n

🧪 Generate unit tests (beta)

\n\n- [ ] Create PR with unit tests\n- [ ] Commit unit tests in branch `ade/cancelling-chats-and-subagents-37cbbffe`\n\n

\n\n

❤️ Share

\n\n_{Comment `@coderabbitai help` to get the list of available commands and usage tips.}\n\n","side","line","[vc]: #wjicWokYmhDMgfXJ82PieIpxPjmMzyMwYg2XagiZOyA=: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\nThe latest updates on your projects. Learn more about [Vercel for GitHub](https://vercel.link/github-learn-more).\n\n| Project | Deployment | Actions | Updated (UTC) |\n| :--- | :----- | :------ | :------ |\n| [ade](https://vercel.com/arul28s-projects/ade) | ![Ready](https://vercel.com/static/status/ready.svg) [Ready](https://vercel.com/arul28s-projects/ade/6MXqKJWBxhGdv1ZwVQGAT7CVhg3V) | [Preview](https://ade-git-ade-cancelling-chats-and-subage-fde2fc-arul28s-projects.vercel.app), [Comment](https://vercel.live/open-feedback/ade-git-ade-cancelling-chats-and-subage-fde2fc-arul28s-projects.vercel.app?via=pr-comment-feedback-link) | Mar 30, 2026 10:57pm |\n\n","base_ref","head_ref","pull_request_info",[2587],{"_1067":45,"_45":2588},"@codex resolve conflicts and commit changes here",[3084,3085,3086],[],{"_1072":1073,"_45":1074,"_1075":1076},{"_1047":3075},{"_320":3074},{"_320":2620},{"_287":288,"_278":289,"_290":291,"_292":57,"_293":15,"_294":15,"_295":296,"_297":298,"_299":251,"_300":126,"_301":302,"_303":304,"_305":306,"_307":-5,"_308":-5,"_309":2619,"_311":312,"_313":314,"_315":316,"_317":318},{"_41":122,"_179":180,"_181":182,"_183":15,"_184":185,"_186":187,"_188":2608,"_190":-5,"_191":-5,"_192":2609,"_194":185,"_195":185,"_196":2610,"_198":2611,"_200":2612,"_202":-5,"_203":204,"_205":206,"_207":208,"_209":2613,"_53":-5,"_118":211,"_212":2614,"_214":2615,"_216":217,"_218":15,"_219":15},{"_132":2601},[2599,2600],{"_143":157,"_145":146,"_147":148,"_149":158,"_151":152,"_153":154,"_155":31,"_156":-5},{"_143":144,"_145":146,"_147":148,"_149":150,"_151":152,"_153":154,"_155":31,"_156":-5},[2602,2603,2604],{"_163":2607,"_165":176},{"_163":2606,"_165":172},{"_163":2605,"_165":166},{"_167":168,"_91":169,"_118":170},{"_167":173,"_91":169,"_118":174},{"_167":177,"_91":169,"_118":178},{"_280":281,"_282":283,"_284":185,"_285":-5,"_286":-5},[241],{"_241":2616},{"_223":224,"_225":226,"_227":228,"_229":230,"_231":232,"_233":234,"_235":236,"_237":238,"_239":240},{},[],{"_222":57},{"_220":57,"_221":185},{"_41":243,"_167":244,"_245":2617,"_247":182,"_248":249,"_250":251,"_252":253,"_254":255,"_256":257,"_258":62,"_259":-5,"_260":-5,"_261":-5,"_262":-5,"_263":-5,"_264":-5,"_265":-5,"_155":266,"_267":-5,"_53":2618},{"_272":273,"_41":274,"_91":275,"_276":277,"_278":279},{"_269":15,"_270":15,"_271":15},{"_272":273,"_167":273,"_319":-5,"_276":277,"_41":274},[2621,2622,2623,2624,2625,2626,2627,2628,2629,2630,2631,2632,2633,2634,2635,2636,2637,2638,2639,2640,2641,2642,2643,2644,2645,2646,2647,2648,2649,2650,2651,2652,2653,2654,2655,2656,2657,2658,2659,2660,2661,2662,2663,2664,2665,2666,2667,2668,2669,2670,2671,2672,2673,2674,2675,2676,2677,2678,2679,2680,2681,2682,2683,2684,2685,2686,2687,2688,2689,2690,2691,2692,2693,2694,2695,2696,2697,2698,2699,2700,2701,2702,2703,2704,2705,2706,2707,2708,2709,2710,2711,2712,2713,2714,2715,2716,2717,2718,2719,2720,2721,2722,2723,2724,2725,2726,2727,2728,2729,2730,2731],{"_149":1025,"_434":3070},{"_149":447,"_434":3066},{"_149":447,"_434":3061},{"_149":443,"_434":3058},{"_149":1001,"_434":3055},{"_149":447,"_434":3051},{"_149":461,"_434":3046},{"_149":461,"_434":3042},{"_149":447,"_434":3039},{"_149":447,"_434":3037},{"_149":461,"_434":3033},{"_149":447,"_434":3031},{"_149":461,"_434":3027},{"_149":447,"_434":3024},{"_149":447,"_434":3022},{"_149":461,"_434":3018},{"_149":447,"_434":3016},{"_149":461,"_434":3012},{"_149":447,"_434":3009},{"_149":447,"_434":3007},{"_149":461,"_434":3003},{"_149":447,"_434":3001},{"_149":461,"_434":2997},{"_149":447,"_434":2994},{"_149":447,"_434":2992},{"_149":461,"_434":2988},{"_149":447,"_434":2986},{"_149":461,"_434":2982},{"_149":447,"_434":2979},{"_149":447,"_434":2977},{"_149":461,"_434":2973},{"_149":447,"_434":2971},{"_149":461,"_434":2967},{"_149":447,"_434":2964},{"_149":447,"_434":2962},{"_149":461,"_434":2958},{"_149":447,"_434":2956},{"_149":461,"_434":2952},{"_149":447,"_434":2949},{"_149":447,"_434":2947},{"_149":461,"_434":2943},{"_149":447,"_434":2941},{"_149":461,"_434":2937},{"_149":447,"_434":2934},{"_149":447,"_434":2932},{"_149":461,"_434":2928},{"_149":447,"_434":2926},{"_149":461,"_434":2922},{"_149":447,"_434":2919},{"_149":447,"_434":2917},{"_149":461,"_434":2913},{"_149":447,"_434":2911},{"_149":461,"_434":2907},{"_149":447,"_434":2904},{"_149":447,"_434":2902},{"_149":461,"_434":2898},{"_149":447,"_434":2896},{"_149":447,"_434":2894},{"_149":461,"_434":2890},{"_149":447,"_434":2888},{"_149":461,"_434":2884},{"_149":447,"_434":2881},{"_149":447,"_434":2879},{"_149":461,"_434":2875},{"_149":447,"_434":2873},{"_149":461,"_434":2869},{"_149":447,"_434":2866},{"_149":447,"_434":2864},{"_149":461,"_434":2860},{"_149":447,"_434":2858},{"_149":461,"_434":2854},{"_149":447,"_434":2851},{"_149":447,"_434":2849},{"_149":461,"_434":2845},{"_149":447,"_434":2843},{"_149":461,"_434":2839},{"_149":447,"_434":2836},{"_149":447,"_434":2834},{"_149":461,"_434":2830},{"_149":447,"_434":2828},{"_149":461,"_434":2824},{"_149":447,"_434":2821},{"_149":447,"_434":2819},{"_149":461,"_434":2815},{"_149":447,"_434":2813},{"_149":461,"_434":2809},{"_149":447,"_434":2806},{"_149":447,"_434":2804},{"_149":461,"_434":2800},{"_149":447,"_434":2798},{"_149":461,"_434":2794},{"_149":447,"_434":2791},{"_149":447,"_434":2789},{"_149":461,"_434":2785},{"_149":447,"_434":2783},{"_149":461,"_434":2779},{"_149":447,"_434":2776},{"_149":447,"_434":2774},{"_149":461,"_434":2767},{"_149":447,"_434":2765},{"_149":461,"_434":2761},{"_149":447,"_434":2758},{"_149":447,"_434":2756},{"_149":461,"_434":2752},{"_149":447,"_434":2750},{"_149":461,"_434":2746},{"_149":447,"_434":2743},{"_149":461,"_434":2741},{"_149":447,"_434":2737},{"_149":443,"_434":2735},{"_149":433,"_434":2732},{"_436":437,"_86":2733},{"_102":-5,"_41":439,"_440":2734,"_442":112},[],{"_442":2736,"_436":437},{"_91":446},{"_449":2738,"_436":437,"_451":439},{"_452":2739,"_454":-5,"_41":-5,"_455":456,"_95":92,"_91":457},[2740],{"_45":459,"_91":460},{"_449":2742,"_436":437,"_451":439},{"_41":464,"_455":456,"_45":459,"_91":465},{"_449":2744,"_436":437,"_451":439},{"_452":-5,"_468":469,"_41":-5,"_470":2745,"_91":472},[],{"_449":2747,"_436":437,"_451":439},{"_452":2748,"_41":476,"_470":2749,"_91":472},[],[],{"_449":2751,"_436":437,"_451":439},{"_480":481,"_482":483,"_91":484},{"_449":2753,"_436":437,"_451":439},{"_487":488,"_489":490,"_491":2754,"_493":494,"_495":31,"_496":31,"_41":481,"_497":498,"_499":500,"_442":112,"_91":501},[2755],{"_489":503,"_167":504,"_156":505,"_91":506},{"_449":2757,"_436":437,"_451":439},{"_509":510,"_480":481,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2759,"_436":437,"_451":439},{"_452":-5,"_468":516,"_41":-5,"_470":2760,"_91":472},[],{"_449":2762,"_436":437,"_451":439},{"_452":2763,"_41":521,"_470":2764,"_91":472},[],[],{"_449":2766,"_436":437,"_451":439},{"_480":525,"_482":526,"_91":484},{"_449":2768,"_436":437,"_451":439},{"_509":2769,"_495":530,"_102":-5,"_41":525,"_531":-5,"_532":2770,"_534":535,"_442":112,"_536":535,"_91":537},{"_295":542,"_297":543},{"_538":-5,"_452":2771,"_540":2772},[2773],{"_295":542,"_297":543},{"_91":45,"_45":545},{"_449":2775,"_436":437,"_451":439},{"_509":526,"_480":525,"_41":-5,"_167":548,"_512":-5,"_91":513},{"_449":2777,"_436":437,"_451":439},{"_452":-5,"_468":551,"_41":-5,"_470":2778,"_91":472},[],{"_449":2780,"_436":437,"_451":439},{"_452":2781,"_41":556,"_470":2782,"_91":472},[],[],{"_449":2784,"_436":437,"_451":439},{"_480":560,"_482":561,"_91":484},{"_449":2786,"_436":437,"_451":439},{"_487":564,"_489":565,"_491":2787,"_493":494,"_495":31,"_496":31,"_41":560,"_497":567,"_499":500,"_442":112,"_91":501},[2788],{"_489":569,"_91":570},{"_449":2790,"_436":437,"_451":439},{"_509":573,"_480":560,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2792,"_436":437,"_451":439},{"_452":-5,"_468":576,"_41":-5,"_470":2793,"_91":472},[],{"_449":2795,"_436":437,"_451":439},{"_452":2796,"_41":581,"_470":2797,"_91":472},[],[],{"_449":2799,"_436":437,"_451":439},{"_480":585,"_482":586,"_91":484},{"_449":2801,"_436":437,"_451":439},{"_487":-5,"_489":589,"_491":2802,"_493":494,"_495":31,"_496":31,"_41":585,"_497":591,"_499":500,"_442":112,"_91":501},[2803],{"_489":593,"_91":570},{"_449":2805,"_436":437,"_451":439},{"_509":596,"_480":585,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2807,"_436":437,"_451":439},{"_452":-5,"_468":599,"_41":-5,"_470":2808,"_91":472},[],{"_449":2810,"_436":437,"_451":439},{"_452":2811,"_41":604,"_470":2812,"_91":472},[],[],{"_449":2814,"_436":437,"_451":439},{"_480":608,"_482":609,"_91":484},{"_449":2816,"_436":437,"_451":439},{"_487":612,"_489":613,"_491":2817,"_493":494,"_495":31,"_496":31,"_41":608,"_497":615,"_499":500,"_442":112,"_91":501},[2818],{"_489":617,"_91":570},{"_449":2820,"_436":437,"_451":439},{"_509":620,"_480":608,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2822,"_436":437,"_451":439},{"_452":-5,"_468":623,"_41":-5,"_470":2823,"_91":472},[],{"_449":2825,"_436":437,"_451":439},{"_452":2826,"_41":628,"_470":2827,"_91":472},[],[],{"_449":2829,"_436":437,"_451":439},{"_480":632,"_482":633,"_91":484},{"_449":2831,"_436":437,"_451":439},{"_487":636,"_489":637,"_491":2832,"_493":494,"_495":31,"_496":31,"_41":632,"_497":639,"_499":500,"_442":112,"_91":501},[2833],{"_489":641,"_91":570},{"_449":2835,"_436":437,"_451":439},{"_509":644,"_480":632,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2837,"_436":437,"_451":439},{"_452":-5,"_468":647,"_41":-5,"_470":2838,"_91":472},[],{"_449":2840,"_436":437,"_451":439},{"_452":2841,"_41":652,"_470":2842,"_91":472},[],[],{"_449":2844,"_436":437,"_451":439},{"_480":656,"_482":657,"_91":484},{"_449":2846,"_436":437,"_451":439},{"_487":-5,"_489":660,"_491":2847,"_493":494,"_495":31,"_496":662,"_41":656,"_497":663,"_499":500,"_442":664,"_91":501},[2848],{"_489":666,"_156":-5,"_667":668,"_91":669},{"_449":2850,"_436":437,"_451":439},{"_509":672,"_480":656,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2852,"_436":437,"_451":439},{"_452":-5,"_468":675,"_41":-5,"_470":2853,"_91":472},[],{"_449":2855,"_436":437,"_451":439},{"_452":2856,"_41":680,"_470":2857,"_91":472},[],[],{"_449":2859,"_436":437,"_451":439},{"_480":684,"_482":685,"_91":484},{"_449":2861,"_436":437,"_451":439},{"_487":-5,"_489":688,"_491":2862,"_493":494,"_495":690,"_496":31,"_41":684,"_497":691,"_499":500,"_442":112,"_91":501},[2863],{"_489":693,"_91":570},{"_449":2865,"_436":437,"_451":439},{"_509":696,"_480":684,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2867,"_436":437,"_451":439},{"_452":-5,"_468":699,"_41":-5,"_470":2868,"_91":472},[],{"_449":2870,"_436":437,"_451":439},{"_452":2871,"_41":704,"_470":2872,"_91":472},[],[],{"_449":2874,"_436":437,"_451":439},{"_480":708,"_482":709,"_91":484},{"_449":2876,"_436":437,"_451":439},{"_487":712,"_489":713,"_491":2877,"_493":494,"_495":31,"_496":31,"_41":708,"_497":715,"_499":500,"_442":112,"_91":501},[2878],{"_489":717,"_156":718,"_667":719,"_91":669},{"_449":2880,"_436":437,"_451":439},{"_509":722,"_480":708,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2882,"_436":437,"_451":439},{"_452":-5,"_468":725,"_41":-5,"_470":2883,"_91":472},[],{"_449":2885,"_436":437,"_451":439},{"_452":2886,"_41":730,"_470":2887,"_91":472},[],[],{"_449":2889,"_436":437,"_451":439},{"_480":734,"_482":735,"_91":484},{"_449":2891,"_436":437,"_451":439},{"_487":738,"_489":739,"_491":2892,"_493":494,"_495":31,"_496":31,"_41":734,"_497":741,"_499":500,"_442":112,"_91":501},[2893],{"_489":743,"_167":504,"_156":505,"_91":506},{"_449":2895,"_436":437,"_451":439},{"_509":746,"_480":734,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2897,"_436":437,"_451":439},{"_480":749,"_482":750,"_91":484},{"_449":2899,"_436":437,"_451":439},{"_487":753,"_489":754,"_491":2900,"_493":494,"_495":31,"_496":31,"_41":749,"_497":756,"_499":500,"_442":112,"_91":501},[2901],{"_489":758,"_156":504,"_667":668,"_91":669},{"_449":2903,"_436":437,"_451":439},{"_509":761,"_480":749,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2905,"_436":437,"_451":439},{"_452":-5,"_468":764,"_41":-5,"_470":2906,"_91":472},[],{"_449":2908,"_436":437,"_451":439},{"_452":2909,"_41":769,"_470":2910,"_91":472},[],[],{"_449":2912,"_436":437,"_451":439},{"_480":773,"_482":774,"_91":484},{"_449":2914,"_436":437,"_451":439},{"_487":777,"_489":778,"_491":2915,"_493":494,"_495":31,"_496":31,"_41":773,"_497":780,"_499":500,"_442":112,"_91":501},[2916],{"_489":782,"_167":504,"_156":505,"_91":506},{"_449":2918,"_436":437,"_451":439},{"_509":785,"_480":773,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2920,"_436":437,"_451":439},{"_452":-5,"_468":788,"_41":-5,"_470":2921,"_91":472},[],{"_449":2923,"_436":437,"_451":439},{"_452":2924,"_41":793,"_470":2925,"_91":472},[],[],{"_449":2927,"_436":437,"_451":439},{"_480":797,"_482":798,"_91":484},{"_449":2929,"_436":437,"_451":439},{"_487":636,"_489":637,"_491":2930,"_493":494,"_495":31,"_496":31,"_41":797,"_497":802,"_499":500,"_442":112,"_91":501},[2931],{"_489":641,"_91":570},{"_449":2933,"_436":437,"_451":439},{"_509":644,"_480":797,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2935,"_436":437,"_451":439},{"_452":-5,"_468":808,"_41":-5,"_470":2936,"_91":472},[],{"_449":2938,"_436":437,"_451":439},{"_452":2939,"_41":813,"_470":2940,"_91":472},[],[],{"_449":2942,"_436":437,"_451":439},{"_480":817,"_482":818,"_91":484},{"_449":2944,"_436":437,"_451":439},{"_487":821,"_489":822,"_491":2945,"_493":494,"_495":824,"_496":662,"_41":817,"_497":825,"_499":500,"_442":664,"_91":501},[2946],{"_489":827,"_91":570},{"_449":2948,"_436":437,"_451":439},{"_509":830,"_480":817,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2950,"_436":437,"_451":439},{"_452":-5,"_468":833,"_41":-5,"_470":2951,"_91":472},[],{"_449":2953,"_436":437,"_451":439},{"_452":2954,"_41":838,"_470":2955,"_91":472},[],[],{"_449":2957,"_436":437,"_451":439},{"_480":842,"_482":843,"_91":484},{"_449":2959,"_436":437,"_451":439},{"_487":846,"_489":847,"_491":2960,"_493":494,"_495":31,"_496":31,"_41":842,"_497":849,"_499":500,"_442":112,"_91":501},[2961],{"_489":851,"_91":570},{"_449":2963,"_436":437,"_451":439},{"_509":854,"_480":842,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2965,"_436":437,"_451":439},{"_452":-5,"_468":857,"_41":-5,"_470":2966,"_91":472},[],{"_449":2968,"_436":437,"_451":439},{"_452":2969,"_41":862,"_470":2970,"_91":472},[],[],{"_449":2972,"_436":437,"_451":439},{"_480":866,"_482":867,"_91":484},{"_449":2974,"_436":437,"_451":439},{"_487":870,"_489":871,"_491":2975,"_493":494,"_495":31,"_496":31,"_41":866,"_497":873,"_499":500,"_442":112,"_91":501},[2976],{"_489":875,"_91":570},{"_449":2978,"_436":437,"_451":439},{"_509":878,"_480":866,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2980,"_436":437,"_451":439},{"_452":-5,"_468":881,"_41":-5,"_470":2981,"_91":472},[],{"_449":2983,"_436":437,"_451":439},{"_452":2984,"_41":886,"_470":2985,"_91":472},[],[],{"_449":2987,"_436":437,"_451":439},{"_480":890,"_482":891,"_91":484},{"_449":2989,"_436":437,"_451":439},{"_487":894,"_489":895,"_491":2990,"_493":494,"_495":31,"_496":31,"_41":890,"_497":897,"_499":500,"_442":112,"_91":501},[2991],{"_489":899,"_91":570},{"_449":2993,"_436":437,"_451":439},{"_509":902,"_480":890,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":2995,"_436":437,"_451":439},{"_452":-5,"_468":905,"_41":-5,"_470":2996,"_91":472},[],{"_449":2998,"_436":437,"_451":439},{"_452":2999,"_41":910,"_470":3000,"_91":472},[],[],{"_449":3002,"_436":437,"_451":439},{"_480":914,"_482":915,"_91":484},{"_449":3004,"_436":437,"_451":439},{"_487":918,"_489":919,"_491":3005,"_493":494,"_495":530,"_496":31,"_41":914,"_497":921,"_499":500,"_442":112,"_91":501},[3006],{"_489":923,"_91":570},{"_449":3008,"_436":437,"_451":439},{"_509":926,"_480":914,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3010,"_436":437,"_451":439},{"_452":-5,"_468":929,"_41":-5,"_470":3011,"_91":472},[],{"_449":3013,"_436":437,"_451":439},{"_452":3014,"_41":934,"_470":3015,"_91":472},[],[],{"_449":3017,"_436":437,"_451":439},{"_480":938,"_482":939,"_91":484},{"_449":3019,"_436":437,"_451":439},{"_487":942,"_489":943,"_491":3020,"_493":494,"_495":31,"_496":31,"_41":938,"_497":945,"_499":500,"_442":112,"_91":501},[3021],{"_489":947,"_167":948,"_156":949,"_91":506},{"_449":3023,"_436":437,"_451":439},{"_509":952,"_480":938,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3025,"_436":437,"_451":439},{"_452":-5,"_468":955,"_41":-5,"_470":3026,"_91":472},[],{"_449":3028,"_436":437,"_451":439},{"_452":3029,"_41":960,"_470":3030,"_91":472},[],[],{"_449":3032,"_436":437,"_451":439},{"_480":964,"_482":965,"_91":484},{"_449":3034,"_436":437,"_451":439},{"_487":968,"_489":969,"_491":3035,"_493":494,"_495":31,"_496":31,"_41":964,"_497":971,"_499":500,"_442":112,"_91":501},[3036],{"_489":973,"_156":974,"_667":948,"_91":669},{"_449":3038,"_436":437,"_451":439},{"_509":977,"_480":964,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3040,"_436":437,"_451":439},{"_452":-5,"_468":980,"_41":-5,"_470":3041,"_91":472},[],{"_449":3043,"_436":437,"_451":439},{"_452":3044,"_41":985,"_470":3045,"_91":472},[],[],{"_449":3047,"_436":437,"_451":439},{"_452":3048,"_41":990,"_91":991},[3049],{"_45":993,"_994":3050,"_91":45},[],{"_449":3052,"_436":437,"_451":439},{"_452":3053,"_454":-5,"_41":-5,"_455":-5,"_95":309,"_91":457},[3054],{"_45":993,"_91":1000},{"_436":437,"_86":3056},{"_102":-5,"_41":439,"_440":3057,"_442":1005},[],{"_442":3059,"_436":437},{"_1008":3060,"_91":1010},[],{"_449":3062,"_436":437,"_451":1013},{"_452":3063,"_454":-5,"_41":-5,"_455":-5,"_95":309,"_91":457},[3064,3065],{"_45":1018,"_91":1000},{"_45":1017,"_91":1000},{"_449":3067,"_436":437,"_451":1013},{"_452":3068,"_454":-5,"_41":-5,"_455":-5,"_95":1022,"_91":457},[3069],{"_45":1024,"_91":1000},{"_1027":3071},{"_1029":-5,"_1030":-5,"_1031":1032,"_1033":1034,"_493":494,"_1035":15,"_1036":-5,"_1037":-5,"_41":437,"_1038":1039,"_167":-5,"_156":1040,"_1041":185,"_499":1042,"_442":3072,"_76":3073,"_1045":1034},{"_91":446},[],[],[3076],{"_41":1050,"_1051":3077,"_1053":1054,"_1055":1054,"_452":3078,"_442":1057,"_454":-5,"_1058":31,"_1059":3079,"_1061":283,"_1062":-5},{"_95":309,"_167":-5,"_1059":3083},{"_1067":45,"_1068":3082},{"_1063":3080},{"_1065":3081},[],[1070],{},{"_91":457,"_95":92,"_452":3094},{"_91":1105,"_1106":3089},{"_91":1080,"_1081":1082,"_1083":1084,"_1085":3087,"_1087":-5},{"_91":1085,"_1088":494,"_127":128,"_1089":-5,"_308":1090,"_1091":3088,"_1093":1094,"_1095":1096,"_1097":1098,"_1099":1100,"_1101":-5},{"_1102":1103,"_1104":-5},[3090],{"_156":1109,"_1110":-5,"_1111":3091,"_1113":-5},[3092],{"_1115":662,"_1116":1117,"_452":3093,"_1119":57},[1120,185,1121,185,1122,185,1123,185,1124,185,1125,185,1126,1127,1128,1129,1130,1131,1132,1133,1134,1135,1136,1137,1138,1139,1140,1141,1142,1143,1144,1145,1146,1147,1148,1149,1150,1151,1152,1153,1154,1155,1156,1157,1158,1159,1160,1161,1162,185,1124,185,1163,185,1164,185,1165,1166,1167,185,1168,185,1169,1170,1171,1172,185,1173,185,1174,185,1175,1176,1177,1178,1179,185,1180,185,1181,185,1182,185,1183,185,1184,185,1124,185,1185,185,1186,185,1187,185,1188,1189,185,1190,185,1191,185,1192,185,1193,185,1194,1195,1196,1197,185,1198,185,1199,185,1200,185,1201,185,1202,1203,1204,185,1205,185,1206,1207,1208,1209,185,1210,185,1211,1212,1213,185,1214,185,1215,185,1216,1217,1218,1219,1220,185,1221,185,1222,185,1223,1224,1225,1226,1227,1228,1229,1230,1231,1232,1233,1234,1235,1236,1237,185,1238,185,1239,185,1124,185,1240,185,1241,185,1242,185,1243,185,1244,185,1245,1246,1247,1248,185,1249,185,1250,1251,1252,185,1253,1254,1255,1256,1257,185,1258,185,1259,185,1260,185,1261,185,1262,185,1263,185,1264,185,1265,1266,1267,1268,1269,1270,1271,1272,1273,1274,1275,1276,1277,185,1278,185,1279,185,1280,185,1281,1282,1283,185,1284,185,1285,185,1286,1287,1288,185,1289,185,1290,185,1291,1292,1293,185,1294,185,1295,185,1296,1297,1298,1299,1300,1301,185,1302,185,1303,185,1304,185,1305,1306,1307,1308,1309,185,1310,185,1311,185,1312,185,1313,1314,1315,1316,1317,1318,185,1319,185,1320,185,1321,185,1322,1323,1324,1325,1326,185,1327,1328,1329,1330,1331,1332,1333,1334,1335,1336,1337,1330,1331,1338,1339,1330,1331,1340,1341,1330,1342,1343,1344,1345,1346,1330,1342,1343,1347,1345,1348,1330,1342,1349,1350,1351,185,1352,1353,1354,1355,1356,1357,1358,1359,1360,1361,1362,1363,185,1364,1365,1366,1367,1368,1369,1370,1371,185,1372,1373,1374,1375,1376,1377,1378,1379,1380,185,1381,185,1382,185,1383,185,1384,185,1385,1386,1387,1388,1389,1390,1391,185,1392,185,1393,185,1394,1395,1396,1397,1398,185,1399,185,1400,185,1401,1402,1403,1404,1405,1406,1407,1408,1409,1410,1411,1412,185,1413,185,1414,185,1415,1416,1417,1418,1419,185,1420,185,1421,185,1422,1423,1424,1425,185,1426,185,1427,185,1428,1429,1430,1431,1432,1433,1434,1435,1436,1437,1380,185,1438,185,1439,185,1440,185,1441,185,1442,1443,1444,1445,1446,1447,1448,185,1449,1450,1451,1452,1453,1454,185,1455,1456,1457,185,1458,185,1459,185,1460,185,1461,185,1462,185,1463,185,1464,185,1465,185,1466,1467,1468,1469,1470,1471,1472,1473,1474,1475,1476,185,1477,185,1478,185,1479,185,1480,185,1481,185,1482,185,1483,185,1484,185,1485,185,1486,185,1487,185,1380,1488,1489,1490,1491,1492,1493,1494,1495,1496,1497,1498,1499,1500,1501,1502,1503,1498,1504,1505,1506,1507,1508,1509,1510,1511,1512,1513,1514,1498,1504,1505,1515,1516,1517,1518,1500,1519,1520,1498,1499,1500,1521,1522,1520,1498,1499,1523,1524,1489,1490,1525,1489,1490,1526,1380,185,1527,185,1528,185,1529,185,1530,1531,1532,1533,1534,1535,1536,1537,1538,185,1539,185,1540,185,1541,1542,185,1543,1544,185,1545,185,1546,185,1547,185,1548,1549,1550,1551,185,1552,185,1553,185,1554,1555,1556,1557,1558,185,1559,185,1560,185,1561,185,1562,185,1563,185,1564,1565,1566,1567,1568,1569,1570,1571,1572,1573,185,1574,185,1575,1576,1577,1578,1579,185,1580,185,1581,185,1582,185,1583,1584,1585,185,1586,185,1587,1588,1589,1590,1591,185,1592,185,1593,1594,1595,185,1596,185,1597,185,1598,1599,1600,185,1601,185,1602,185,1603,185,1604,185,1605,1606,1607,1608,185,1609,185,1610,185,1611,1612,1613,1614,1615,1616,1617,1618,1619,1620,1621,1622,1623,1624,1625,185,1626,185,1627,185,1380,1628,1629,1630,1631,1632,1629,1633,1631,1634,1380,185,1635,1636,1637,185,1638,185,1639,185,1640,1641,1642,185,1643,185,1644,185,1645,185,1646,185,1647,185,1648,1649,1650,1651,1652,185,1653,1654,1655,1656,1657,1658,1659,185,1660,185,1661,1662,1663,1664,1665,1666,1667,1668,1669,1670,185,1671,185,1672,185,1673,185,1674,185,1675,1676,1677,1678,1679,185,1680,185,1681,185,1682,185,1322,1323,1683,1684,1685,1686,1687,1688,185,1689,1690,1691,1692,1693,1694,185,1695,1696,1697,185,1698,1699,1700,1380,185,1701,185,1702,185,1703,185,1704,1705,1706,1707,185,1708,185,1709,185,1710,185,1711,185,1712,185,1713,185,1714,1715,1716,1717,1718,185,1719,185,1720,185,1721,185,1722,1723,1724,1725,1726,1727,185,1728,185,1729,185,1730,185,1731,1732,1733,185,1734,185,1735,185,1736,185,1737,1738,185,1739,185,1740,185,1741,185,1742,185,1743,185,1744,1745,1746,1747,185,1748,185,1749,185,1750,185,1751,185,1752,185,1753,1754,1755,1756,185,1757,185,1758,185,1759,185,1760,1761,185,1762,1763,1764,1765,185,1766,1767,1768,185,1769,185,1770,1771,1772,1773,1774,185,1775,1776,1777,1778,185,1779,185,1780,185,1781,185,1782,1783,1784,1785,185,1786,185,1787,185,1788,185,1789,185,1790,185,1791,185,1792,185,1793,185,1794,185,1795,185,1796,1797,1798,1799,185,1800,185,1380,1801,1802,1803,1380,185,1804,1805,1806,1807,1808,1809,185,1810,1811,1812,1813,185,1814,185,1380,1815,1816,1817,1818,1380,185,1819,185,1820,185,1821,1822,1823,1824,185,1825,185,1826,185,1380,1827,1828,1829,1830,1831,1832,1380,185,1833,1834,1835,1836,185,1837,185,1838,185,1839,185,1840,185,1841,185,1842,185,1843,185,1844,185,1845,185,1846,185,1847,185,1848,185,1849,185,1850,185,1322,1851,1852,1853,1854,1855,1856,1857,1858,1859,1860,1861,1862,1863,1864,1380,185,1865,185,1866,185,1867,1868,1869,185,1870,185,1871,1872,1873,1874,1875,1876,1877,185,1878,185,1879,185,1880,185,1881,185,1380,1882,1883,1884,1885,1886,1887,1888,1889,1890,1891,1892,1893,1894,1895,1896,1897,1898,1899,1900,1901,1902,1903,1904,1895,1905,1906,1907,1908,1909,1910,1911,1908,1909,1912,1380,185,1913,1914,1915,1916,1917,1918,1919,185,1920,1921,1922,1923,1924,185,1925,185,1926,185,1927,185,1928,185,1929,185,1930,185,1931,1932,1933,1934,1935,1936,1937,185,1938,185,1939,185,1940,1941,1942,1943,1944,185,1945,1946,1947,1948,1949,185,1950,1951,1952,1953,185,1954,185,1955,185,1956,185,1957,185,1958,1959,1960,1961,1962,1963,1964,1965,1966,1967,1968,1969,1970,1971,1972,1973,1974,1975,1976,1977,1978,185,1979,1980,1981,1982,1983,1984,1985,1986,1987,1988,1989,1990,1991,185,1992,1993,1994,1995,1996,1997,1998,1999,2000,185,2001,2002,2003,2004,2005,2006,2007,2008,2009,185,2010,2011,2012,2013,2014,2015,185,2016,2017,2018,2019,185,2020,2021,2022,2023,2024,185,2025,2026,2027,2028,2029,2030,185,2031,185,2032,185,2033,2034,2035,185,2036,185,2037,185,2038,2039,185,2040,2041,2042,2043,2044,2045,2046,2047,2048,2049,2046,2050,2051,2052,2053,2054,2046,1437,1380,185,2055,185,2056,185,2057,185,2058,2059,2060,2061,2062,2063,2064,2065,185,2066,2067,2068,2069,2070,185,2071,185,2072,185,2073,185,2074,2075,2076,2077,2078,2079,2080,2081,185,2082,185,2083,185,1322,1323,1327,1328,1329,2084,1332,1333,1335,1336,1337,2084,1338,1339,2084,1340,1341,2085,1343,1344,1345,1346,2085,1343,1347,1345,1380,185,2086,185,2087,185,2088,2089,2090,2091,185,2092,185,2093,185,2094,185,2095,185,2038,2096,2097,2098,2099,2100,2101,2102,2103,2104,2105,185,2106,2107,2108,2109,2110,2111,1437,1380,185,2112,185,2113,2114,2115,2116,2117,2118,2119,2120,185,2121,185,2122,185,2123,185,2124,185,2125,185,2126,185,2127,185,2128,185,2129,185,2130,185,2131,2132,2133,2134,2135,185,2136,185,2038,2137,2138,2139,2140,2141,2142,2143,2144,2145,2146,2147,2148,1437,185,2149,2150,2151,2152,2153,2154,2155,2156,2157,2158,2159,2160,2161,2162,2163,2164,1437,185,2165,2166,2167,2168,2169,2170,2171,2172,2173,2174,2175,2176,2177,2178,2179,185,2180,185,2181,2182,2183,1437,185,2184,2150,2185,2186,2187,1437,1380,185,2188,185,2189,185,2190,2191,185,2192,185,2038,2193,2194,185,2195,2196,2197,2198,185,2199,2200,185,2201,2202,185,2203,2204,2205,2206,2207,2208,1380,185,2209,185,2210,2211,2212,2213,2214,2215,2216,2217,2218,2219,2220,2221,2222,185,2223,185,2224,2211,2225,2226,2227,2228,2229,2230,2231,2232,2233,2234,2235,185,2236,185,2237,2211,2238,2239,2240,185,2241,185,2242,185,2243,185,2244,185,2038,2245,2246,185,2247,2248,185,2249,2250,2251,185,2252,2253,2254,2255,2256,2257,2258,2259,2260,2261,2262,2263,2105,185,2264,2265,2266,2267,2268,2110,185,2269,2270,1437,1380,185,2271,185,2272,2211,2273,2274,2275,2276,2277,2278,2279,2280,2281,2282,185,2283,185,2284,185,2285,185,2286,185,2287,185,2288,2289,2290,185,2291,185,2292,185,1380,2293,2294,2295,2296,2297,2298,2299,185,2300,2301,2302,2303,2304,185,2305,2306,2307,185,2308,2309,2310,2311,2312,2313,2314,185,2315,2316,2317,2318,2319,2320,1380,185,2321,185,1124,185,2322,185,2323,185,2324,2325,2326,2327,2328,2329,2330,2331,2332,2333,185,2334,185,2335,185,2336,2337,185,2338,185,2339,185,2340,185,2341,2342,2343,2344,2345,2346,2347,2348,185,2349,185,2350,185,2351,2352,2353,2354,2355,185,2356,185,2357,185,2358,185,2359,185,2360,2361,2362,2363,2364,185,2365,185,2366,185,1124,185,2367,185,2368,2369,2370,2371,2372,2373,2374,2375,2376,2377,2378,2379,2380,2381,2382,2383,2384,2385,2386,2387,2388,2389,2390,2391,2392,2393,2394,2395,2396,2397,2398,2399,2400,2401,2402,2403,2404,2405,2406,2407,2408,2409,2410,2411,2412,2413,2414,2415,2416,2417,2418,2419,2420,2421,2422,2423,2424,2425,2426,2427,2428,2429,185,2430,185,1124,185,2431,185,2432,185,2433,185,2434,2435,2436,185,2437,185,2438,185,2439,2440,2441,2442,2443,2444,2445,2446,2447,2448,2449,2450,2451,2452,2453,2454,2455,2456,185,2457,185,2439,2440,2458,2459,2460,2461,2462,2463,2464,185,2465,185,2439,2440,2466,2467,2468,2469,2470,2471,2472,2473,185,2474,185,2439,2440,2475,2476,2477,185,1124,185,2478,185,2479,185,2480,185,2481,185,2482,2483,2484,2483,185,2485,2483,2486,2483,185,2487,2483,2488,2483,2489,185,2490,185,2491,2483,2492,2493,2494,2495,2496,2497,2483,185,2498,185,1124,185,2499,185,2500,185,2501,185,2502,185,1428,1429,2503,2504,2505,2506,2507,2508,2509,2510,2267,2110,1437,1380,185,2511,185,2512,2513,1380,185,2514,185,2515,185,2516,185,2517,2518,2519,2520,185,2521,185,1124,185,2522,185,2523,185,2524,185,2525,2526,185,2527,185,2528,185,2529,185,2530,2531,2532,2533,185,2534,185,2535,185,2536,185,2537,185,2538,2539],[3095,3096,3097],{"_1067":45,"_45":2548},{"_1067":2545,"_156":1109,"_1115":2546,"_1116":2547},{"_1067":45,"_45":2544},[3593,3594,3595],[],{"_1072":1073,"_45":1074,"_1075":1076},{"_1047":3584},{"_320":3583},{"_320":3129},{"_287":288,"_278":289,"_290":291,"_292":57,"_293":15,"_294":15,"_295":296,"_297":298,"_299":251,"_300":126,"_301":302,"_303":304,"_305":306,"_307":-5,"_308":-5,"_309":3128,"_311":312,"_313":314,"_315":316,"_317":318},{"_41":122,"_179":180,"_181":182,"_183":15,"_184":185,"_186":187,"_188":3117,"_190":-5,"_191":-5,"_192":3118,"_194":185,"_195":185,"_196":3119,"_198":3120,"_200":3121,"_202":-5,"_203":204,"_205":206,"_207":208,"_209":3122,"_53":-5,"_118":211,"_212":3123,"_214":3124,"_216":217,"_218":15,"_219":15},{"_132":3110},[3108,3109],{"_143":157,"_145":146,"_147":148,"_149":158,"_151":152,"_153":154,"_155":31,"_156":-5},{"_143":144,"_145":146,"_147":148,"_149":150,"_151":152,"_153":154,"_155":31,"_156":-5},[3111,3112,3113],{"_163":3116,"_165":176},{"_163":3115,"_165":172},{"_163":3114,"_165":166},{"_167":168,"_91":169,"_118":170},{"_167":173,"_91":169,"_118":174},{"_167":177,"_91":169,"_118":178},{"_280":281,"_282":283,"_284":185,"_285":-5,"_286":-5},[241],{"_241":3125},{"_223":224,"_225":226,"_227":228,"_229":230,"_231":232,"_233":234,"_235":236,"_237":238,"_239":240},{},[],{"_222":57},{"_220":57,"_221":185},{"_41":243,"_167":244,"_245":3126,"_247":182,"_248":249,"_250":251,"_252":253,"_254":255,"_256":257,"_258":62,"_259":-5,"_260":-5,"_261":-5,"_262":-5,"_263":-5,"_264":-5,"_265":-5,"_155":266,"_267":-5,"_53":3127},{"_272":273,"_41":274,"_91":275,"_276":277,"_278":279},{"_269":15,"_270":15,"_271":15},{"_272":273,"_167":273,"_319":-5,"_276":277,"_41":274},[3130,3131,3132,3133,3134,3135,3136,3137,3138,3139,3140,3141,3142,3143,3144,3145,3146,3147,3148,3149,3150,3151,3152,3153,3154,3155,3156,3157,3158,3159,3160,3161,3162,3163,3164,3165,3166,3167,3168,3169,3170,3171,3172,3173,3174,3175,3176,3177,3178,3179,3180,3181,3182,3183,3184,3185,3186,3187,3188,3189,3190,3191,3192,3193,3194,3195,3196,3197,3198,3199,3200,3201,3202,3203,3204,3205,3206,3207,3208,3209,3210,3211,3212,3213,3214,3215,3216,3217,3218,3219,3220,3221,3222,3223,3224,3225,3226,3227,3228,3229,3230,3231,3232,3233,3234,3235,3236,3237,3238,3239,3240],{"_149":1025,"_434":3579},{"_149":447,"_434":3575},{"_149":447,"_434":3570},{"_149":443,"_434":3567},{"_149":1001,"_434":3564},{"_149":447,"_434":3560},{"_149":461,"_434":3555},{"_149":461,"_434":3551},{"_149":447,"_434":3548},{"_149":447,"_434":3546},{"_149":461,"_434":3542},{"_149":447,"_434":3540},{"_149":461,"_434":3536},{"_149":447,"_434":3533},{"_149":447,"_434":3531},{"_149":461,"_434":3527},{"_149":447,"_434":3525},{"_149":461,"_434":3521},{"_149":447,"_434":3518},{"_149":447,"_434":3516},{"_149":461,"_434":3512},{"_149":447,"_434":3510},{"_149":461,"_434":3506},{"_149":447,"_434":3503},{"_149":447,"_434":3501},{"_149":461,"_434":3497},{"_149":447,"_434":3495},{"_149":461,"_434":3491},{"_149":447,"_434":3488},{"_149":447,"_434":3486},{"_149":461,"_434":3482},{"_149":447,"_434":3480},{"_149":461,"_434":3476},{"_149":447,"_434":3473},{"_149":447,"_434":3471},{"_149":461,"_434":3467},{"_149":447,"_434":3465},{"_149":461,"_434":3461},{"_149":447,"_434":3458},{"_149":447,"_434":3456},{"_149":461,"_434":3452},{"_149":447,"_434":3450},{"_149":461,"_434":3446},{"_149":447,"_434":3443},{"_149":447,"_434":3441},{"_149":461,"_434":3437},{"_149":447,"_434":3435},{"_149":461,"_434":3431},{"_149":447,"_434":3428},{"_149":447,"_434":3426},{"_149":461,"_434":3422},{"_149":447,"_434":3420},{"_149":461,"_434":3416},{"_149":447,"_434":3413},{"_149":447,"_434":3411},{"_149":461,"_434":3407},{"_149":447,"_434":3405},{"_149":447,"_434":3403},{"_149":461,"_434":3399},{"_149":447,"_434":3397},{"_149":461,"_434":3393},{"_149":447,"_434":3390},{"_149":447,"_434":3388},{"_149":461,"_434":3384},{"_149":447,"_434":3382},{"_149":461,"_434":3378},{"_149":447,"_434":3375},{"_149":447,"_434":3373},{"_149":461,"_434":3369},{"_149":447,"_434":3367},{"_149":461,"_434":3363},{"_149":447,"_434":3360},{"_149":447,"_434":3358},{"_149":461,"_434":3354},{"_149":447,"_434":3352},{"_149":461,"_434":3348},{"_149":447,"_434":3345},{"_149":447,"_434":3343},{"_149":461,"_434":3339},{"_149":447,"_434":3337},{"_149":461,"_434":3333},{"_149":447,"_434":3330},{"_149":447,"_434":3328},{"_149":461,"_434":3324},{"_149":447,"_434":3322},{"_149":461,"_434":3318},{"_149":447,"_434":3315},{"_149":447,"_434":3313},{"_149":461,"_434":3309},{"_149":447,"_434":3307},{"_149":461,"_434":3303},{"_149":447,"_434":3300},{"_149":447,"_434":3298},{"_149":461,"_434":3294},{"_149":447,"_434":3292},{"_149":461,"_434":3288},{"_149":447,"_434":3285},{"_149":447,"_434":3283},{"_149":461,"_434":3276},{"_149":447,"_434":3274},{"_149":461,"_434":3270},{"_149":447,"_434":3267},{"_149":447,"_434":3265},{"_149":461,"_434":3261},{"_149":447,"_434":3259},{"_149":461,"_434":3255},{"_149":447,"_434":3252},{"_149":461,"_434":3250},{"_149":447,"_434":3246},{"_149":443,"_434":3244},{"_149":433,"_434":3241},{"_436":437,"_86":3242},{"_102":-5,"_41":439,"_440":3243,"_442":112},[],{"_442":3245,"_436":437},{"_91":446},{"_449":3247,"_436":437,"_451":439},{"_452":3248,"_454":-5,"_41":-5,"_455":456,"_95":92,"_91":457},[3249],{"_45":459,"_91":460},{"_449":3251,"_436":437,"_451":439},{"_41":464,"_455":456,"_45":459,"_91":465},{"_449":3253,"_436":437,"_451":439},{"_452":-5,"_468":469,"_41":-5,"_470":3254,"_91":472},[],{"_449":3256,"_436":437,"_451":439},{"_452":3257,"_41":476,"_470":3258,"_91":472},[],[],{"_449":3260,"_436":437,"_451":439},{"_480":481,"_482":483,"_91":484},{"_449":3262,"_436":437,"_451":439},{"_487":488,"_489":490,"_491":3263,"_493":494,"_495":31,"_496":31,"_41":481,"_497":498,"_499":500,"_442":112,"_91":501},[3264],{"_489":503,"_167":504,"_156":505,"_91":506},{"_449":3266,"_436":437,"_451":439},{"_509":510,"_480":481,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3268,"_436":437,"_451":439},{"_452":-5,"_468":516,"_41":-5,"_470":3269,"_91":472},[],{"_449":3271,"_436":437,"_451":439},{"_452":3272,"_41":521,"_470":3273,"_91":472},[],[],{"_449":3275,"_436":437,"_451":439},{"_480":525,"_482":526,"_91":484},{"_449":3277,"_436":437,"_451":439},{"_509":3278,"_495":530,"_102":-5,"_41":525,"_531":-5,"_532":3279,"_534":535,"_442":112,"_536":535,"_91":537},{"_295":542,"_297":543},{"_538":-5,"_452":3280,"_540":3281},[3282],{"_295":542,"_297":543},{"_91":45,"_45":545},{"_449":3284,"_436":437,"_451":439},{"_509":526,"_480":525,"_41":-5,"_167":548,"_512":-5,"_91":513},{"_449":3286,"_436":437,"_451":439},{"_452":-5,"_468":551,"_41":-5,"_470":3287,"_91":472},[],{"_449":3289,"_436":437,"_451":439},{"_452":3290,"_41":556,"_470":3291,"_91":472},[],[],{"_449":3293,"_436":437,"_451":439},{"_480":560,"_482":561,"_91":484},{"_449":3295,"_436":437,"_451":439},{"_487":564,"_489":565,"_491":3296,"_493":494,"_495":31,"_496":31,"_41":560,"_497":567,"_499":500,"_442":112,"_91":501},[3297],{"_489":569,"_91":570},{"_449":3299,"_436":437,"_451":439},{"_509":573,"_480":560,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3301,"_436":437,"_451":439},{"_452":-5,"_468":576,"_41":-5,"_470":3302,"_91":472},[],{"_449":3304,"_436":437,"_451":439},{"_452":3305,"_41":581,"_470":3306,"_91":472},[],[],{"_449":3308,"_436":437,"_451":439},{"_480":585,"_482":586,"_91":484},{"_449":3310,"_436":437,"_451":439},{"_487":-5,"_489":589,"_491":3311,"_493":494,"_495":31,"_496":31,"_41":585,"_497":591,"_499":500,"_442":112,"_91":501},[3312],{"_489":593,"_91":570},{"_449":3314,"_436":437,"_451":439},{"_509":596,"_480":585,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3316,"_436":437,"_451":439},{"_452":-5,"_468":599,"_41":-5,"_470":3317,"_91":472},[],{"_449":3319,"_436":437,"_451":439},{"_452":3320,"_41":604,"_470":3321,"_91":472},[],[],{"_449":3323,"_436":437,"_451":439},{"_480":608,"_482":609,"_91":484},{"_449":3325,"_436":437,"_451":439},{"_487":612,"_489":613,"_491":3326,"_493":494,"_495":31,"_496":31,"_41":608,"_497":615,"_499":500,"_442":112,"_91":501},[3327],{"_489":617,"_91":570},{"_449":3329,"_436":437,"_451":439},{"_509":620,"_480":608,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3331,"_436":437,"_451":439},{"_452":-5,"_468":623,"_41":-5,"_470":3332,"_91":472},[],{"_449":3334,"_436":437,"_451":439},{"_452":3335,"_41":628,"_470":3336,"_91":472},[],[],{"_449":3338,"_436":437,"_451":439},{"_480":632,"_482":633,"_91":484},{"_449":3340,"_436":437,"_451":439},{"_487":636,"_489":637,"_491":3341,"_493":494,"_495":31,"_496":31,"_41":632,"_497":639,"_499":500,"_442":112,"_91":501},[3342],{"_489":641,"_91":570},{"_449":3344,"_436":437,"_451":439},{"_509":644,"_480":632,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3346,"_436":437,"_451":439},{"_452":-5,"_468":647,"_41":-5,"_470":3347,"_91":472},[],{"_449":3349,"_436":437,"_451":439},{"_452":3350,"_41":652,"_470":3351,"_91":472},[],[],{"_449":3353,"_436":437,"_451":439},{"_480":656,"_482":657,"_91":484},{"_449":3355,"_436":437,"_451":439},{"_487":-5,"_489":660,"_491":3356,"_493":494,"_495":31,"_496":662,"_41":656,"_497":663,"_499":500,"_442":664,"_91":501},[3357],{"_489":666,"_156":-5,"_667":668,"_91":669},{"_449":3359,"_436":437,"_451":439},{"_509":672,"_480":656,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3361,"_436":437,"_451":439},{"_452":-5,"_468":675,"_41":-5,"_470":3362,"_91":472},[],{"_449":3364,"_436":437,"_451":439},{"_452":3365,"_41":680,"_470":3366,"_91":472},[],[],{"_449":3368,"_436":437,"_451":439},{"_480":684,"_482":685,"_91":484},{"_449":3370,"_436":437,"_451":439},{"_487":-5,"_489":688,"_491":3371,"_493":494,"_495":690,"_496":31,"_41":684,"_497":691,"_499":500,"_442":112,"_91":501},[3372],{"_489":693,"_91":570},{"_449":3374,"_436":437,"_451":439},{"_509":696,"_480":684,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3376,"_436":437,"_451":439},{"_452":-5,"_468":699,"_41":-5,"_470":3377,"_91":472},[],{"_449":3379,"_436":437,"_451":439},{"_452":3380,"_41":704,"_470":3381,"_91":472},[],[],{"_449":3383,"_436":437,"_451":439},{"_480":708,"_482":709,"_91":484},{"_449":3385,"_436":437,"_451":439},{"_487":712,"_489":713,"_491":3386,"_493":494,"_495":31,"_496":31,"_41":708,"_497":715,"_499":500,"_442":112,"_91":501},[3387],{"_489":717,"_156":718,"_667":719,"_91":669},{"_449":3389,"_436":437,"_451":439},{"_509":722,"_480":708,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3391,"_436":437,"_451":439},{"_452":-5,"_468":725,"_41":-5,"_470":3392,"_91":472},[],{"_449":3394,"_436":437,"_451":439},{"_452":3395,"_41":730,"_470":3396,"_91":472},[],[],{"_449":3398,"_436":437,"_451":439},{"_480":734,"_482":735,"_91":484},{"_449":3400,"_436":437,"_451":439},{"_487":738,"_489":739,"_491":3401,"_493":494,"_495":31,"_496":31,"_41":734,"_497":741,"_499":500,"_442":112,"_91":501},[3402],{"_489":743,"_167":504,"_156":505,"_91":506},{"_449":3404,"_436":437,"_451":439},{"_509":746,"_480":734,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3406,"_436":437,"_451":439},{"_480":749,"_482":750,"_91":484},{"_449":3408,"_436":437,"_451":439},{"_487":753,"_489":754,"_491":3409,"_493":494,"_495":31,"_496":31,"_41":749,"_497":756,"_499":500,"_442":112,"_91":501},[3410],{"_489":758,"_156":504,"_667":668,"_91":669},{"_449":3412,"_436":437,"_451":439},{"_509":761,"_480":749,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3414,"_436":437,"_451":439},{"_452":-5,"_468":764,"_41":-5,"_470":3415,"_91":472},[],{"_449":3417,"_436":437,"_451":439},{"_452":3418,"_41":769,"_470":3419,"_91":472},[],[],{"_449":3421,"_436":437,"_451":439},{"_480":773,"_482":774,"_91":484},{"_449":3423,"_436":437,"_451":439},{"_487":777,"_489":778,"_491":3424,"_493":494,"_495":31,"_496":31,"_41":773,"_497":780,"_499":500,"_442":112,"_91":501},[3425],{"_489":782,"_167":504,"_156":505,"_91":506},{"_449":3427,"_436":437,"_451":439},{"_509":785,"_480":773,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3429,"_436":437,"_451":439},{"_452":-5,"_468":788,"_41":-5,"_470":3430,"_91":472},[],{"_449":3432,"_436":437,"_451":439},{"_452":3433,"_41":793,"_470":3434,"_91":472},[],[],{"_449":3436,"_436":437,"_451":439},{"_480":797,"_482":798,"_91":484},{"_449":3438,"_436":437,"_451":439},{"_487":636,"_489":637,"_491":3439,"_493":494,"_495":31,"_496":31,"_41":797,"_497":802,"_499":500,"_442":112,"_91":501},[3440],{"_489":641,"_91":570},{"_449":3442,"_436":437,"_451":439},{"_509":644,"_480":797,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3444,"_436":437,"_451":439},{"_452":-5,"_468":808,"_41":-5,"_470":3445,"_91":472},[],{"_449":3447,"_436":437,"_451":439},{"_452":3448,"_41":813,"_470":3449,"_91":472},[],[],{"_449":3451,"_436":437,"_451":439},{"_480":817,"_482":818,"_91":484},{"_449":3453,"_436":437,"_451":439},{"_487":821,"_489":822,"_491":3454,"_493":494,"_495":824,"_496":662,"_41":817,"_497":825,"_499":500,"_442":664,"_91":501},[3455],{"_489":827,"_91":570},{"_449":3457,"_436":437,"_451":439},{"_509":830,"_480":817,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3459,"_436":437,"_451":439},{"_452":-5,"_468":833,"_41":-5,"_470":3460,"_91":472},[],{"_449":3462,"_436":437,"_451":439},{"_452":3463,"_41":838,"_470":3464,"_91":472},[],[],{"_449":3466,"_436":437,"_451":439},{"_480":842,"_482":843,"_91":484},{"_449":3468,"_436":437,"_451":439},{"_487":846,"_489":847,"_491":3469,"_493":494,"_495":31,"_496":31,"_41":842,"_497":849,"_499":500,"_442":112,"_91":501},[3470],{"_489":851,"_91":570},{"_449":3472,"_436":437,"_451":439},{"_509":854,"_480":842,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3474,"_436":437,"_451":439},{"_452":-5,"_468":857,"_41":-5,"_470":3475,"_91":472},[],{"_449":3477,"_436":437,"_451":439},{"_452":3478,"_41":862,"_470":3479,"_91":472},[],[],{"_449":3481,"_436":437,"_451":439},{"_480":866,"_482":867,"_91":484},{"_449":3483,"_436":437,"_451":439},{"_487":870,"_489":871,"_491":3484,"_493":494,"_495":31,"_496":31,"_41":866,"_497":873,"_499":500,"_442":112,"_91":501},[3485],{"_489":875,"_91":570},{"_449":3487,"_436":437,"_451":439},{"_509":878,"_480":866,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3489,"_436":437,"_451":439},{"_452":-5,"_468":881,"_41":-5,"_470":3490,"_91":472},[],{"_449":3492,"_436":437,"_451":439},{"_452":3493,"_41":886,"_470":3494,"_91":472},[],[],{"_449":3496,"_436":437,"_451":439},{"_480":890,"_482":891,"_91":484},{"_449":3498,"_436":437,"_451":439},{"_487":894,"_489":895,"_491":3499,"_493":494,"_495":31,"_496":31,"_41":890,"_497":897,"_499":500,"_442":112,"_91":501},[3500],{"_489":899,"_91":570},{"_449":3502,"_436":437,"_451":439},{"_509":902,"_480":890,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3504,"_436":437,"_451":439},{"_452":-5,"_468":905,"_41":-5,"_470":3505,"_91":472},[],{"_449":3507,"_436":437,"_451":439},{"_452":3508,"_41":910,"_470":3509,"_91":472},[],[],{"_449":3511,"_436":437,"_451":439},{"_480":914,"_482":915,"_91":484},{"_449":3513,"_436":437,"_451":439},{"_487":918,"_489":919,"_491":3514,"_493":494,"_495":530,"_496":31,"_41":914,"_497":921,"_499":500,"_442":112,"_91":501},[3515],{"_489":923,"_91":570},{"_449":3517,"_436":437,"_451":439},{"_509":926,"_480":914,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3519,"_436":437,"_451":439},{"_452":-5,"_468":929,"_41":-5,"_470":3520,"_91":472},[],{"_449":3522,"_436":437,"_451":439},{"_452":3523,"_41":934,"_470":3524,"_91":472},[],[],{"_449":3526,"_436":437,"_451":439},{"_480":938,"_482":939,"_91":484},{"_449":3528,"_436":437,"_451":439},{"_487":942,"_489":943,"_491":3529,"_493":494,"_495":31,"_496":31,"_41":938,"_497":945,"_499":500,"_442":112,"_91":501},[3530],{"_489":947,"_167":948,"_156":949,"_91":506},{"_449":3532,"_436":437,"_451":439},{"_509":952,"_480":938,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3534,"_436":437,"_451":439},{"_452":-5,"_468":955,"_41":-5,"_470":3535,"_91":472},[],{"_449":3537,"_436":437,"_451":439},{"_452":3538,"_41":960,"_470":3539,"_91":472},[],[],{"_449":3541,"_436":437,"_451":439},{"_480":964,"_482":965,"_91":484},{"_449":3543,"_436":437,"_451":439},{"_487":968,"_489":969,"_491":3544,"_493":494,"_495":31,"_496":31,"_41":964,"_497":971,"_499":500,"_442":112,"_91":501},[3545],{"_489":973,"_156":974,"_667":948,"_91":669},{"_449":3547,"_436":437,"_451":439},{"_509":977,"_480":964,"_41":-5,"_167":511,"_512":-5,"_91":513},{"_449":3549,"_436":437,"_451":439},{"_452":-5,"_468":980,"_41":-5,"_470":3550,"_91":472},[],{"_449":3552,"_436":437,"_451":439},{"_452":3553,"_41":985,"_470":3554,"_91":472},[],[],{"_449":3556,"_436":437,"_451":439},{"_452":3557,"_41":990,"_91":991},[3558],{"_45":993,"_994":3559,"_91":45},[],{"_449":3561,"_436":437,"_451":439},{"_452":3562,"_454":-5,"_41":-5,"_455":-5,"_95":309,"_91":457},[3563],{"_45":993,"_91":1000},{"_436":437,"_86":3565},{"_102":-5,"_41":439,"_440":3566,"_442":1005},[],{"_442":3568,"_436":437},{"_1008":3569,"_91":1010},[],{"_449":3571,"_436":437,"_451":1013},{"_452":3572,"_454":-5,"_41":-5,"_455":-5,"_95":309,"_91":457},[3573,3574],{"_45":1018,"_91":1000},{"_45":1017,"_91":1000},{"_449":3576,"_436":437,"_451":1013},{"_452":3577,"_454":-5,"_41":-5,"_455":-5,"_95":1022,"_91":457},[3578],{"_45":1024,"_91":1000},{"_1027":3580},{"_1029":-5,"_1030":-5,"_1031":1032,"_1033":1034,"_493":494,"_1035":15,"_1036":-5,"_1037":-5,"_41":437,"_1038":1039,"_167":-5,"_156":1040,"_1041":185,"_499":1042,"_442":3581,"_76":3582,"_1045":1034},{"_91":446},[],[],[3585],{"_41":1050,"_1051":3586,"_1053":1054,"_1055":1054,"_452":3587,"_442":1057,"_454":-5,"_1058":31,"_1059":3588,"_1061":283,"_1062":-5},{"_95":309,"_167":-5,"_1059":3592},{"_1067":45,"_1068":3591},{"_1063":3589},{"_1065":3590},[],[1070],{},{"_91":457,"_95":92,"_452":3603},{"_91":1105,"_1106":3598},{"_91":1080,"_1081":1082,"_1083":1084,"_1085":3596,"_1087":-5},{"_91":1085,"_1088":494,"_127":128,"_1089":-5,"_308":1090,"_1091":3597,"_1093":1094,"_1095":1096,"_1097":1098,"_1099":1100,"_1101":-5},{"_1102":1103,"_1104":-5},[3599],{"_156":1109,"_1110":-5,"_1111":3600,"_1113":-5},[3601],{"_1115":662,"_1116":1117,"_452":3602,"_1119":57},[1120,185,1121,185,1122,185,1123,185,1124,185,1125,185,1126,1127,1128,1129,1130,1131,1132,1133,1134,1135,1136,1137,1138,1139,1140,1141,1142,1143,1144,1145,1146,1147,1148,1149,1150,1151,1152,1153,1154,1155,1156,1157,1158,1159,1160,1161,1162,185,1124,185,1163,185,1164,185,1165,1166,1167,185,1168,185,1169,1170,1171,1172,185,1173,185,1174,185,1175,1176,1177,1178,1179,185,1180,185,1181,185,1182,185,1183,185,1184,185,1124,185,1185,185,1186,185,1187,185,1188,1189,185,1190,185,1191,185,1192,185,1193,185,1194,1195,1196,1197,185,1198,185,1199,185,1200,185,1201,185,1202,1203,1204,185,1205,185,1206,1207,1208,1209,185,1210,185,1211,1212,1213,185,1214,185,1215,185,1216,1217,1218,1219,1220,185,1221,185,1222,185,1223,1224,1225,1226,1227,1228,1229,1230,1231,1232,1233,1234,1235,1236,1237,185,1238,185,1239,185,1124,185,1240,185,1241,185,1242,185,1243,185,1244,185,1245,1246,1247,1248,185,1249,185,1250,1251,1252,185,1253,1254,1255,1256,1257,185,1258,185,1259,185,1260,185,1261,185,1262,185,1263,185,1264,185,1265,1266,1267,1268,1269,1270,1271,1272,1273,1274,1275,1276,1277,185,1278,185,1279,185,1280,185,1281,1282,1283,185,1284,185,1285,185,1286,1287,1288,185,1289,185,1290,185,1291,1292,1293,185,1294,185,1295,185,1296,1297,1298,1299,1300,1301,185,1302,185,1303,185,1304,185,1305,1306,1307,1308,1309,185,1310,185,1311,185,1312,185,1313,1314,1315,1316,1317,1318,185,1319,185,1320,185,1321,185,1322,1323,1324,1325,1326,185,1327,1328,1329,1330,1331,1332,1333,1334,1335,1336,1337,1330,1331,1338,1339,1330,1331,1340,1341,1330,1342,1343,1344,1345,1346,1330,1342,1343,1347,1345,1348,1330,1342,1349,1350,1351,185,1352,1353,1354,1355,1356,1357,1358,1359,1360,1361,1362,1363,185,1364,1365,1366,1367,1368,1369,1370,1371,185,1372,1373,1374,1375,1376,1377,1378,1379,1380,185,1381,185,1382,185,1383,185,1384,185,1385,1386,1387,1388,1389,1390,1391,185,1392,185,1393,185,1394,1395,1396,1397,1398,185,1399,185,1400,185,1401,1402,1403,1404,1405,1406,1407,1408,1409,1410,1411,1412,185,1413,185,1414,185,1415,1416,1417,1418,1419,185,1420,185,1421,185,1422,1423,1424,1425,185,1426,185,1427,185,1428,1429,1430,1431,1432,1433,1434,1435,1436,1437,1380,185,1438,185,1439,185,1440,185,1441,185,1442,1443,1444,1445,1446,1447,1448,185,1449,1450,1451,1452,1453,1454,185,1455,1456,1457,185,1458,185,1459,185,1460,185,1461,185,1462,185,1463,185,1464,185,1465,185,1466,1467,1468,1469,1470,1471,1472,1473,1474,1475,1476,185,1477,185,1478,185,1479,185,1480,185,1481,185,1482,185,1483,185,1484,185,1485,185,1486,185,1487,185,1380,1488,1489,1490,1491,1492,1493,1494,1495,1496,1497,1498,1499,1500,1501,1502,1503,1498,1504,1505,1506,1507,1508,1509,1510,1511,1512,1513,1514,1498,1504,1505,1515,1516,1517,1518,1500,1519,1520,1498,1499,1500,1521,1522,1520,1498,1499,1523,1524,1489,1490,1525,1489,1490,1526,1380,185,1527,185,1528,185,1529,185,1530,1531,1532,1533,1534,1535,1536,1537,1538,185,1539,185,1540,185,1541,1542,185,1543,1544,185,1545,185,1546,185,1547,185,1548,1549,1550,1551,185,1552,185,1553,185,1554,1555,1556,1557,1558,185,1559,185,1560,185,1561,185,1562,185,1563,185,1564,1565,1566,1567,1568,1569,1570,1571,1572,1573,185,1574,185,1575,1576,1577,1578,1579,185,1580,185,1581,185,1582,185,1583,1584,1585,185,1586,185,1587,1588,1589,1590,1591,185,1592,185,1593,1594,1595,185,1596,185,1597,185,1598,1599,1600,185,1601,185,1602,185,1603,185,1604,185,1605,1606,1607,1608,185,1609,185,1610,185,1611,1612,1613,1614,1615,1616,1617,1618,1619,1620,1621,1622,1623,1624,1625,185,1626,185,1627,185,1380,1628,1629,1630,1631,1632,1629,1633,1631,1634,1380,185,1635,1636,1637,185,1638,185,1639,185,1640,1641,1642,185,1643,185,1644,185,1645,185,1646,185,1647,185,1648,1649,1650,1651,1652,185,1653,1654,1655,1656,1657,1658,1659,185,1660,185,1661,1662,1663,1664,1665,1666,1667,1668,1669,1670,185,1671,185,1672,185,1673,185,1674,185,1675,1676,1677,1678,1679,185,1680,185,1681,185,1682,185,1322,1323,1683,1684,1685,1686,1687,1688,185,1689,1690,1691,1692,1693,1694,185,1695,1696,1697,185,1698,1699,1700,1380,185,1701,185,1702,185,1703,185,1704,1705,1706,1707,185,1708,185,1709,185,1710,185,1711,185,1712,185,1713,185,1714,1715,1716,1717,1718,185,1719,185,1720,185,1721,185,1722,1723,1724,1725,1726,1727,185,1728,185,1729,185,1730,185,1731,1732,1733,185,1734,185,1735,185,1736,185,1737,1738,185,1739,185,1740,185,1741,185,1742,185,1743,185,1744,1745,1746,1747,185,1748,185,1749,185,1750,185,1751,185,1752,185,1753,1754,1755,1756,185,1757,185,1758,185,1759,185,1760,1761,185,1762,1763,1764,1765,185,1766,1767,1768,185,1769,185,1770,1771,1772,1773,1774,185,1775,1776,1777,1778,185,1779,185,1780,185,1781,185,1782,1783,1784,1785,185,1786,185,1787,185,1788,185,1789,185,1790,185,1791,185,1792,185,1793,185,1794,185,1795,185,1796,1797,1798,1799,185,1800,185,1380,1801,1802,1803,1380,185,1804,1805,1806,1807,1808,1809,185,1810,1811,1812,1813,185,1814,185,1380,1815,1816,1817,1818,1380,185,1819,185,1820,185,1821,1822,1823,1824,185,1825,185,1826,185,1380,1827,1828,1829,1830,1831,1832,1380,185,1833,1834,1835,1836,185,1837,185,1838,185,1839,185,1840,185,1841,185,1842,185,1843,185,1844,185,1845,185,1846,185,1847,185,1848,185,1849,185,1850,185,1322,1851,1852,1853,1854,1855,1856,1857,1858,1859,1860,1861,1862,1863,1864,1380,185,1865,185,1866,185,1867,1868,1869,185,1870,185,1871,1872,1873,1874,1875,1876,1877,185,1878,185,1879,185,1880,185,1881,185,1380,1882,1883,1884,1885,1886,1887,1888,1889,1890,1891,1892,1893,1894,1895,1896,1897,1898,1899,1900,1901,1902,1903,1904,1895,1905,1906,1907,1908,1909,1910,1911,1908,1909,1912,1380,185,1913,1914,1915,1916,1917,1918,1919,185,1920,1921,1922,1923,1924,185,1925,185,1926,185,1927,185,1928,185,1929,185,1930,185,1931,1932,1933,1934,1935,1936,1937,185,1938,185,1939,185,1940,1941,1942,1943,1944,185,1945,1946,1947,1948,1949,185,1950,1951,1952,1953,185,1954,185,1955,185,1956,185,1957,185,1958,1959,1960,1961,1962,1963,1964,1965,1966,1967,1968,1969,1970,1971,1972,1973,1974,1975,1976,1977,1978,185,1979,1980,1981,1982,1983,1984,1985,1986,1987,1988,1989,1990,1991,185,1992,1993,1994,1995,1996,1997,1998,1999,2000,185,2001,2002,2003,2004,2005,2006,2007,2008,2009,185,2010,2011,2012,2013,2014,2015,185,2016,2017,2018,2019,185,2020,2021,2022,2023,2024,185,2025,2026,2027,2028,2029,2030,185,2031,185,2032,185,2033,2034,2035,185,2036,185,2037,185,2038,2039,185,2040,2041,2042,2043,2044,2045,2046,2047,2048,2049,2046,2050,2051,2052,2053,2054,2046,1437,1380,185,2055,185,2056,185,2057,185,2058,2059,2060,2061,2062,2063,2064,2065,185,2066,2067,2068,2069,2070,185,2071,185,2072,185,2073,185,2074,2075,2076,2077,2078,2079,2080,2081,185,2082,185,2083,185,1322,1323,1327,1328,1329,2084,1332,1333,1335,1336,1337,2084,1338,1339,2084,1340,1341,2085,1343,1344,1345,1346,2085,1343,1347,1345,1380,185,2086,185,2087,185,2088,2089,2090,2091,185,2092,185,2093,185,2094,185,2095,185,2038,2096,2097,2098,2099,2100,2101,2102,2103,2104,2105,185,2106,2107,2108,2109,2110,2111,1437,1380,185,2112,185,2113,2114,2115,2116,2117,2118,2119,2120,185,2121,185,2122,185,2123,185,2124,185,2125,185,2126,185,2127,185,2128,185,2129,185,2130,185,2131,2132,2133,2134,2135,185,2136,185,2038,2137,2138,2139,2140,2141,2142,2143,2144,2145,2146,2147,2148,1437,185,2149,2150,2151,2152,2153,2154,2155,2156,2157,2158,2159,2160,2161,2162,2163,2164,1437,185,2165,2166,2167,2168,2169,2170,2171,2172,2173,2174,2175,2176,2177,2178,2179,185,2180,185,2181,2182,2183,1437,185,2184,2150,2185,2186,2187,1437,1380,185,2188,185,2189,185,2190,2191,185,2192,185,2038,2193,2194,185,2195,2196,2197,2198,185,2199,2200,185,2201,2202,185,2203,2204,2205,2206,2207,2208,1380,185,2209,185,2210,2211,2212,2213,2214,2215,2216,2217,2218,2219,2220,2221,2222,185,2223,185,2224,2211,2225,2226,2227,2228,2229,2230,2231,2232,2233,2234,2235,185,2236,185,2237,2211,2238,2239,2240,185,2241,185,2242,185,2243,185,2244,185,2038,2245,2246,185,2247,2248,185,2249,2250,2251,185,2252,2253,2254,2255,2256,2257,2258,2259,2260,2261,2262,2263,2105,185,2264,2265,2266,2267,2268,2110,185,2269,2270,1437,1380,185,2271,185,2272,2211,2273,2274,2275,2276,2277,2278,2279,2280,2281,2282,185,2283,185,2284,185,2285,185,2286,185,2287,185,2288,2289,2290,185,2291,185,2292,185,1380,2293,2294,2295,2296,2297,2298,2299,185,2300,2301,2302,2303,2304,185,2305,2306,2307,185,2308,2309,2310,2311,2312,2313,2314,185,2315,2316,2317,2318,2319,2320,1380,185,2321,185,1124,185,2322,185,2323,185,2324,2325,2326,2327,2328,2329,2330,2331,2332,2333,185,2334,185,2335,185,2336,2337,185,2338,185,2339,185,2340,185,2341,2342,2343,2344,2345,2346,2347,2348,185,2349,185,2350,185,2351,2352,2353,2354,2355,185,2356,185,2357,185,2358,185,2359,185,2360,2361,2362,2363,2364,185,2365,185,2366,185,1124,185,2367,185,2368,2369,2370,2371,2372,2373,2374,2375,2376,2377,2378,2379,2380,2381,2382,2383,2384,2385,2386,2387,2388,2389,2390,2391,2392,2393,2394,2395,2396,2397,2398,2399,2400,2401,2402,2403,2404,2405,2406,2407,2408,2409,2410,2411,2412,2413,2414,2415,2416,2417,2418,2419,2420,2421,2422,2423,2424,2425,2426,2427,2428,2429,185,2430,185,1124,185,2431,185,2432,185,2433,185,2434,2435,2436,185,2437,185,2438,185,2439,2440,2441,2442,2443,2444,2445,2446,2447,2448,2449,2450,2451,2452,2453,2454,2455,2456,185,2457,185,2439,2440,2458,2459,2460,2461,2462,2463,2464,185,2465,185,2439,2440,2466,2467,2468,2469,2470,2471,2472,2473,185,2474,185,2439,2440,2475,2476,2477,185,1124,185,2478,185,2479,185,2480,185,2481,185,2482,2483,2484,2483,185,2485,2483,2486,2483,185,2487,2483,2488,2483,2489,185,2490,185,2491,2483,2492,2493,2494,2495,2496,2497,2483,185,2498,185,1124,185,2499,185,2500,185,2501,185,2502,185,1428,1429,2503,2504,2505,2506,2507,2508,2509,2510,2267,2110,1437,1380,185,2511,185,2512,2513,1380,185,2514,185,2515,185,2516,185,2517,2518,2519,2520,185,2521,185,1124,185,2522,185,2523,185,2524,185,2525,2526,185,2527,185,2528,185,2529,185,2530,2531,2532,2533,185,2534,185,2535,185,2536,185,2537,185,2538,2539],[3604,3605,3606],{"_1067":45,"_45":2548},{"_1067":2545,"_156":1109,"_1115":2546,"_1116":2547},{"_1067":45,"_45":2544},[3609,3610,3611,3612],{"_2558":2559,"_247":182,"_2560":288,"_258":-5,"_1088":241,"_2561":-5,"_2562":2563,"_2564":2565,"_2566":-5,"_2567":-5,"_2568":2569,"_2570":15,"_2571":57,"_2572":52,"_2573":15},{"_91":457,"_95":309,"_452":3613},{"_295":296,"_297":185,"_2583":302,"_2584":128,"_305":-5,"_91":2585},{"_91":2578,"_297":2582,"_156":-5,"_2580":-5,"_2581":-5},{"_91":2578,"_297":2579,"_156":-5,"_2580":-5,"_2581":-5},[3614],{"_1067":45,"_45":2588},"task_e_69cbe79ea82c8323b2141fa4ca7fc806",1774970782.657163,"Github Mention: cancelling chats and subagents","has_generated_title","has_unread_turn","denormalized_metadata",{"_3646":31,"_3647":3648,"_3649":-5,"_3650":-5,"_3640":3651,"_3652":-5,"_3653":-5,"_3654":-5,"_3655":-5},"task_status_display",{"_3634":3635,"_3636":182,"_129":126,"_3637":-5,"_3638":1080},"creator_workspace","external_pull_requests",[3627],{"_41":3628,"_3629":82,"_3630":3631,"_3632":128},"github-1154710805-112","assistant_turn_id","pull_request",{"_287":288,"_278":289,"_290":291,"_292":57,"_293":15,"_294":15,"_295":296,"_297":298,"_299":251,"_300":126,"_301":302,"_303":304,"_305":306,"_307":-5,"_308":-5,"_309":3633,"_311":312,"_313":314,"_315":316,"_317":318},"codex_updated_sha",{"_272":273,"_167":273,"_319":-5,"_276":277,"_41":274},"latest_turn_status_display",{"_3639":82,"_111":112,"_3640":3641,"_93":3642,"_3643":-5,"_3644":-5,"_3645":1080,"_113":-5},"environment_label","fix_task_id","initial_intent","turn_id","diff_stats",{"_1093":1094,"_1095":1096,"_1097":1098},[],"image_asset_pointer_citations","issue_counts","intent","num_follow_ups","repo_name_map",{},"last_user_turn_created_at","last_turn_status",{"_1093":-5,"_1095":-5,"_1097":-5},"last_pushed_turn_id","start_branch_name","pushed_branch_name","cached_pull_request_data","dd",{"_3667":3668,"_3669":3670},"disablePrefetch","shouldPrefetchAccount","shouldPrefetchUser","shouldPrefetchSystemHints","promoteCss","disableSSR","stripModulepreloadImports","statsigGateEvaluationsPromise",["P",3666],"traceId","13917082262974375300","traceTime",1779206446351]