fix(coding-agent): reject headless idle waiters when a post-compaction continuation cannot start

[snimu]

Follow-up to #1494 and #800; ports the failure semantics from #881 (closed in favor of #1494's settlement placement).

Linear: ENG-5400 — https://linear.app/primeintellect/issue/ENG-5400/headless-idle-must-surface-post-compaction-continuation-start-failures

Problem

#1494's post-compaction continuation settlement is resolve-only: _runScheduledPostCompactionContinue swallows a continuation that fails to start (.catch(() => undefined)), and _settlePostCompactionContinue resolves the waiters. A headless caller (ACP session/prompt, print mode) waiting on waitForHeadlessIdle then observes a clean finish for a continuation that never ran — the turn silently loses its tail.

#881 solved this by rejecting idle waiters, but built the wait into waitForIdle globally, which would have exposed interactive and extension callers to a new rejection path. #1494's placement (headless-scoped waitForHeadlessIdle wrapper, waitForIdle untouched) is the right architecture, so this PR ports only the failure semantics onto it.

Change

• The settlement is now one-shot (settled flag): a failed settlement is cleared in the same step and never re-exposed to later waiters.
• _settlePostCompactionContinue(error?) rejects waiters on a non-retryable start failure; an error settles immediately, while a resolve still defers to a scheduled/timer-pending retry that owns the settlement next.
• The "already processing" retry loop is unchanged; "continue from" errors (turn already completed — nothing left to continue) still settle as a clean finish. That benign race was previously invisible under the blanket swallow; feat(acp): harden resident session lifecycle #1494's own "headless idle includes a successful post-compaction continuation" test caught the distinction when the swallow came off.
• Interactive waitForIdle is untouched — a placement-contract test pins that it keeps resolving through a failed continuation.

isSessionActive integration and all #1494 settle sites (cancel, dispose, abort, update-restart, ownership change) are unchanged; they settle as resolves.

Consumer

#1494's ACP error boundary already handles a rejected waitForHeadlessCompletion: the turn gets one correlated error response-boundary and never an invented terminal-quiescence update. This PR completes that path — before it, the rejection branch was unreachable from the continuation side.

Tests

Two, both pinning behavior this PR introduces (adapted from #881 — credit @parkerpettit):

• rejects headless idle waiters when a continuation cannot start
• does not expose a failed continuation to later headless idle waiters

Validation

• npm run check, build clean
• 595 tests across the affected suites (compaction-continuation, compaction, acp-mode, print-mode, recursion, queue, daemon connections): all passing

+56/−6 in three files.

---

Note

^{Cursor Bugbot is generating a summary for commit 20d4f80. Configure here.}

Note

Reject headless idle waiters when post-compaction continuation fails to start in AgentSession

• Introduces PostCompactionContinuationSettlement with a settled flag so a failed settlement is one-shot and not re-exposed to later waiters.
• _settlePostCompactionContinue now accepts an optional Error and rejects the settlement promise on failure instead of always resolving.
• _runScheduledPostCompactionContinue calls _settlePostCompactionContinue with the error when agent.continue() fails for reasons other than already processing.
• Interactive idle behavior is unchanged; only ACP and print-mode headless waiters are affected.
• Risk: callers of AgentSession.waitForHeadlessIdle that did not handle rejection may now see unhandled promise rejections when a scheduled continuation cannot start.

^{Macroscope summarized eb0a42c.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 20d4f80. Configure here.}