fix: control Codex goals through the ACP extension - #554
Conversation
Pausing or resuming a goal was delivered as a chat message (`/goal pause`, `/goal resume`), so it needed the session's ACP prompt slot — the same slot the running goal holds open across the agent's own continuations. The dispatch watcher deferred those turns with `guard-noop-active-session` and the banner disabled every button while it waited, so a paused goal's Resume button did nothing until the user pressed Stop. Goal actions now split by what they do. `pause` and `clear` move durable state only, so they travel out of band on `_lody/session/goal` and land while a prompt is running. `set` and `resume` start work, so they ride a Lody-owned turn carrying `_meta.lody.goalControl`: no command text in the conversation, and the resulting turns still belong to an assistant entry the user can see and cancel. An action that arrives while a turn is draining waits for that turn instead of being dropped, newest action wins, and the banner's pending state now expires. Goal buttons follow the runtime's advertised `goalActions` instead of an `agentType === 'codex'` check, so any agent implementing the extension gets them. Model: claude-opus-5 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Remove redundant goal fields, helpers, waiter values, and test details; update the Codex adapter pointer and document the ablation matrix and remaining P1 findings. Validation: targeted tests, full typecheck and lint, formatting, docs, and public boundary checks passed. Full check was stopped after reproducing the known component act failure. Model: gpt-5
Restore the pre-review formatting; no behavior changes. Model: gpt-5
Update acp-extension-codex from d7969e6 to merged PR #39 commit 33d897b. Validation: 161 targeted tests and docs check passed; pnpm format completed. Codex typecheck fails because the current Core pointer lacks LodyWorktreeProject and worktreeProject capabilities. Full pnpm check timed out. The Core integration remains unresolved; this commit only updates the requested pointer. Model: gpt-5
Preserve goal capabilities alongside independent plan mode and pin Core to a merge containing both goal transport and worktree project contracts. Retain merged Codex PR #39 and upstream prompt ownership fixes. Validation: Core build, workspace typecheck and lint, 205 targeted tests, docs and public-boundary checks passed. pnpm check was terminated at five minutes during tests; the full suite did not complete. Model: gpt-5
Pin Core to its 0.1.4 release and Codex to the matching dependency update. Record successful registry-only installation, Codex typechecks, and 52 targeted tests. Validation: workspace typecheck, formatting, and docs checks passed. Full pnpm check timed out at two minutes during lint; full-suite success is not claimed. Model: gpt-5
Update the gitlink to merged PR #40 (5f0aab0). Its tree is identical to the previously validated ce7b0e8 commit; reuse the registry-backed typechecks and 52 passing tests. Docs check and diff check passed; full checks were not rerun for this pointer-only change. Model: gpt-5
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 925094aec4
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| const response = await this.deps.controlSessionGoal({ | ||
| sessionId: request.params.sessionId as SessionId, | ||
| action: request.params.action, | ||
| ...(request.params.objective ? { objective: request.params.objective } : {}), | ||
| userId: request.params.userId, |
There was a problem hiding this comment.
When resume or set arrives after the previous turn has drained, this awaits controlSessionGoal before sending the RPC response. That call awaits startGoalTurn → continueSession → the goal's persistent agentClient.prompt, while both goal RPC clients time out after 10 seconds. A normally resumed goal can therefore start successfully but still be reported as failed, with the server responding only when the long-running goal eventually stops. Start the turn asynchronously after ownership is established and return turn_started immediately.
AGENTS.md reference: packages/loro-streams-rpc/AGENTS.md:L48-L55
Useful? React with 👍 / 👎.
| // No live agent, or an action that needs a turn: the turn boots the session | ||
| // when necessary and lets the agent client pick its transport at prompt time. |
There was a problem hiding this comment.
When a daemon restart or memory eviction leaves a durable goal but no live Session, pause and clear enter this fallback and become prompt turns. After the ACP client is restored, buildGoalControlPrompt resolves these advertised actions to request and throws because it only accepts prompt metadata or the legacy slash bridge, so an active persisted goal can no longer be paused or cleared. Restore the session and then invoke controlGoal through _lody/session/goal instead of putting status-only actions into a prompt.
AGENTS.md reference: packages/shared/AGENTS.md:L78-L83
Useful? React with 👍 / 👎.
| if (this.getExecutionSnapshot(sessionId).hasActiveTurn) { | ||
| this.deps.logger.warn( | ||
| `[${sessionId}] Dropping queued goal ${pending.control.action}: session stayed busy` | ||
| ); | ||
| return; |
There was a problem hiding this comment.
If three successive turns acquire the session before the queued resume or set can start, this deletes and silently drops the action after the caller already received accepted: true with disposition: 'queued'. That contradicts the guarantee in specs/session-goal-control.md that a work-starting action waits and is not dropped, and the caller has no completion channel or durable retry after its pending UI state expires. Keep accepted work queued durably until it executes, or avoid acknowledging responsibility before that guarantee can be maintained.
AGENTS.md reference: AGENTS.md:L12-L17
Useful? React with 👍 / 👎.
2 participants
Related issue
Problem / pressure
An active Codex goal holds the session's only ACP v1 prompt open across the agent's own continuations — that is what
GoalPromptLifecycle(#465) is for, and it is what keeps a goal's turns attributable to one conversation entry.Lody's goal controls were chat messages. Pressing Pause or Resume wrote a pending user turn (
/goal pause,/goal resume) and asked the CLI to dispatch it, butresolveSessionDispatchActionreturnsnoop('active-session')whenever a turn is active, so the turn sat pending until the goal's own prompt closed. Meanwhile the banner'spendingGoalCommandonly cleared when the goal status actually changed, with no timeout, and it disabled every goal button.The user-visible result: a goal shows paused, Resume does nothing, and the only way out is pressing Stop. Pause had already accumulated two compensations for the same root cause — the Stop button sends its own
/goal pause, and the adapter pauses the goal itself when a prompt is cancelled — both of which exist because the bridge could not deliver a command during a prompt.Summary
Goal actions now split by what they do, not by which agent is running.
pause/clearmove durable state only. They go out of band on the_lody/session/goalextension request and take effect mid-prompt. No turn, no prompt slot, no queue.set/resumestart work, and ACP v1 gives a client exactly one way to own running work: its own prompt. They ride a Lody-owned turn carrying_meta.lody.goalControl, so no command text enters the conversation while the resulting turns still belong to an assistant entry the user can see and cancel.waitForTurnRelease) instead of being dropped. One queued action per session, newest wins, bounded to three turn waits.goalActionsinstead ofagentType === 'codex', and its pending state expires after 60s so a slow action cannot leave the controls dead.A goal turn creates an assistant entry with no user message and carries no run configuration, so resuming a goal cannot silently change model or mode.
New machine RPC
session/goal, subject to the same session access verification as a chat message.Depends on two submodule PRs (see Submodules below). The pointers here currently reference their branch heads and must be re-pinned to the merged commits before this merges, the same way #465 pinned its adapter commit.
Visual explanation
Why the two transports are not symmetric:
sequenceDiagram participant UI participant CLI participant Agent Note over Agent: goal active — one prompt open across N native turns UI->>CLI: session/goal {pause} CLI->>Agent: _lody/session/goal (out of band) Agent-->>UI: goal snapshot: paused Note over Agent: current native turn drains, prompt returns UI->>CLI: session/goal {resume} alt a turn is still draining CLI-->>UI: accepted, disposition=queued Note over CLI: waitForTurnRelease, newest action wins end CLI->>Agent: prompt(_meta.lody.goalControl={resume}) Note over CLI,Agent: Lody owns this prompt, so the goal's<br/>turns attach to an assistant entryTransport selection (
apps/cli/src/agent/goal-control.ts):Before / after
/goal …as a user turnguard-noop-active-session/goal resumeappears in the transcriptagentType === 'codex'goalActionsTest plan
Ran:
pnpm typecheck(whole workspace),pnpm lint,pnpm lint:i18n,pnpm check:public-boundary,pnpm check:platform-boundaries,pnpm check:code-collab-imports,pnpm format— all pass.pnpm run docs check— no errors; the new Spec and note aredraft/implementedwith translation pending.@lody/shared1051 tests,@lody/loro-streams-rpc112,lody(CLI) 2639 — pass.New tests:
turnStart; the metadata parser accepts every advertised action and rejects a blank objective, an unknown action, and a future version.goal-control.test.ts): transport selection prefers the request for status-only actions, keeps work-starting actions on a prompt even when the agent lists them as control actions, falls back to the slash bridge for pre-split runtimes, and refuses unadvertised actions.session-execution-service.test.ts): out-of-band pause opens no turn; an unadvertised action is refused; a goal turn carries no run configuration; a resume queued behind a draining turn runs when the turn is released (driven by clearing the current turn — no timers, no sleeps); a newer action supersedes an older queued one.Not verified / known failures:
apps/cli/tests/gh-shim-script.test.tsfails in this sandbox (the spawned shim exits 1). Untouched by this change.@lody/componentstest suite fails wholesale here withReact.act is not a function— the runner resolves a production React build. 154 files, unrelated; the goal-specific component tests were run directly and pass.handles review slash commands through Codex app servertimes out. Reproduced on the pinned pre-change sources, so it is pre-existing.acp-extension-coreis unverified; inside this workspace the pnpm override resolves the local source.Submodules
Both must merge before this PR's pointers are re-pinned:
LodyAI/acp-extension-core→ branchfeat/goal-control-transports(commit9e7503c):controlActions/promptActionson the goal capability,LodyGoalPromptControl, version 0.1.2. Publish 0.1.2 before the adapter's pinned dependency resolves outside this workspace.LodyAI/acp-extension-codex→ branchfeat/goal-prompt-control(commit1a35bc0): accepts_meta.lody.goalControl, advertises both transports, and corrects stale claims indocs/goal-extension.md.I could not open those two PRs: the token available to me gets
Resource not accessible by integration (createPullRequest)on both repositories. The branches are pushed.Documentation
specs/session-goal-control.md(draft) — intent and guarantees..agents/notes/implemented/architecture/2026-09-09-goal-control-plane.md— the decision, the rejected alternatives, and the limits.packages/shared/AGENTS.md— the binding invariant.Both documents are English-only for now (
Translation: pending).🤖 Generated with Claude Code
Context handoff
Instructions for reviewing agents
apps/cli/src/agent/goal-control.ts(transport selection) andSessionExecutionService.controlSessionGoal+queueGoalTurn— the queue is the only new concurrency in this change and it shares the session's turn-ownership rules. AlsoCodexAcpServer.prompt's new metadata branch, which must reach exactly the same code as the slash command.resumemay not use the out-of-band request even when an agent advertises it (the adapter can self-start a turn there, which Lody cannot attribute); that a queued goal action is bounded to three turn waits and then dropped with a warning; thatACP_CAPABILITY_CACHE_VERSIONbumps to 8 rather than treating a missinggoalActionsas "assume Codex supports it".prompt(); the UI should never offer that, but the path is untested.Authoring context
/goal …prompt bridge with a real ACP control operation, defining the contract inacp-extension-coreand consuming it from the Codex adapter./goalstays available for other ACP clients. Not in scope: a first-class UI for setting a goal, and any other agent implementing the extension.session/goalmachine RPC (same access verification as chat); the capability cache version bump, which forces a re-probe; the goal turn deliberately carrying no run configuration.acp-extension-core@0.1.2; the pre-existing sandbox failures listed in the test plan were not chased.