fix(coding-agent): answer ACP prompts only once the session admits the next turn

[parkerpettit]

Linear: ENG-5376

ACP could answer session/prompt after its first idle check even though injected work had restarted the session. An immediate follow-up was then rejected as Agent is already processing.

Without this change:

1. The current ACP turn finishes.
2. waitForHeadlessCompletion() observes an idle session.
3. Injected work, such as a subagent message, starts another turn.
4. ACP returns end_turn.
5. The client sends its next prompt and Prime Agent rejects it because the session is busy.

Re-check connection.waitForIdle() before returning the ACP response. This closes the admission race seen in multi-turn Verifiers runs.

#881 fixes a different gap: waitForIdle() did not include a continuation already scheduled by compact.run. One makes the idle check complete; this PR puts a final check at the response boundary.

Verification:

• ACP response waits for the final admission check
• an immediate second prompt is accepted

Note

Queue ACP follow-up prompts behind in-flight work and cancel queued turns before delivery

• ACP prompt handler no longer throws when a prompt is admitted before cancellation; it returns stopReason="cancelled" and passes streamingBehavior:"followUp", queueIfBusy:true to promptAndWait
• AgentSession.promptAndWait accepts an optional AbortSignal via options.signal; on abort it cancels queued session-owned turn actions matching the agentMessageId before delivery and rejects the completion
• New daemon server capability owned_prompt_cancellation (default-on) and cancelOwned flag on cancel_prompt_admission commands allow cancelling session-owned prompts that have not started; gated by protocol 7, schema revision 20
• Daemon schema bumped from 19 to 20 in daemon-protocol.ts; DaemonAgentConnection sends cancelOwned:true when the server advertises the capability
• Behavioral Change: cancel_prompt_admission with cancelOwned=true now aborts an owned admission's controller in daemon-mode.ts; clients below protocol 7 / schema 20 will not send the flag and owned cancellation is a no-op for them

^{Macroscope summarized 7ce5609.}

---

Note

Medium Risk
Changes daemon worker lifecycle, persisted launch environment, and ACP turn-boundary semantics that verifiers and embedders depend on; behavior is heavily tested but mis-timed idle or env handling could still break reconnect or scoring.

Overview
ACP session/prompt now awaits connection.waitForIdle() before returning end_turn, so clients are not told the turn finished while injected work (e.g. subagent messages) has already restarted the session and would reject an immediate follow-up with “Agent is already processing.”

ACP session lifecycle and scoring metadata: session/new reserves the single-session slot before the first await, subscribes before getInitialSnapshot(), and fails setup cleanly if the snapshot read errors. After headless completion, a session_info_update carries namespaced quiescence (outstandingSubagents, remainingAutonomousContinuations) plus autonomous state from a live roster snapshot; snapshot failures at emission time propagate instead of reporting a false zero.

Daemon / ACP residency: Normal ACP sessions (with a session file) use resident workers so disconnect/reconnect can reattach; --no-session ACP stays client-owned. Resident creates always forward launchEnv from the caller (model endpoint, tokens, proxy, etc.). The supervisor persists launchEnv on resident worker descriptors for recovery after restart and strips it when promoting to client-owned, with integration tests for env across worker recovery and live IPython namespace across ACP reconnect.

^{Reviewed by Cursor Bugbot for commit ac97743. Bugbot is set up for automated code reviews on this repo. Configure here.}

[snimu]

Missing CHANGELOG entry: this fixes a user-visible ACP race (immediate follow-up prompt rejected with "Agent is already processing"). Please add a bullet under ## [Unreleased] in packages/coding-agent/CHANGELOG.md.

[parkerpettit]

Reworked per review: ACP prompts now queue behind in-flight work (streamingBehavior: "followUp", queueIfBusy: true) instead of waiting for whole-session idleness; the final waitForIdle() is removed, so the stop reason comes from the status captured after the turn the response gates and the response is never held open by detached work. Regression tests inject real work after the first idle observation; CHANGELOG entry added.

[macroscopeapp]

Macroscope has since reviewed this pull request. An earlier review was skipped by a cost limit; a review has now completed, so that notice no longer applies.