fix(coding-agent): resume suspended input pump on programmatic admission

[twaldin]

Problem

Any abort — requestAbort() from Esc/TUI interrupt, a daemon abort, a goal abort, or abortForUpdateRestart() — sets _sessionInputPumpSuspended = true. The user-typed prompt path (_prompt) clears that flag when the session is not streaming, but the programmatic admission path (queueAgentMessagePrompt → _queuePreparedPrompt → _admitSessionInput) never does. After one abort, every programmatically queued item — agent messages from children, heartbeat prompts, injected wake prompts — sits in the queue forever while the session is idle. Only a human-typed prompt or an explicit resume_queue revives delivery.

Reproduced clean-room on a pristine 0.7.0 install with a stub provider, driven over the daemon socket:

scenario	result
long tool call / long streaming turn with items queued	delivered at boundary — fine
idle session, items queued	delivered immediately — fine
abort during streaming with items queued	starved forever (queuedCount persisted >90s idle)
then one user-typed prompt	flushes the entire starved backlog in ~2s
starved queue + heartbeat cron due	heartbeat silently skipped every fire

The starved queue also blackholes heartbeats: shouldDeferHeartbeatCronJob defers while unfinishedActionCount > 0, and a skipped fire is lost until the next full interval because nextRunAt is recomputed from the skip time (adjacent to #820).

Related: #820 (heartbeat fires dropped when coincident with busy), #821 (send retry duplication), #823 (queued agent mail delivery). None of them names the suspended-pump starvation itself.

Fix

• _admitSessionInput: when a programmatic turn admission arrives while the session is not streaming and the pump is suspended, clear the suspension, notify checkpoint waiters, and schedule the pump — the same resume semantics _prompt() already gives user-typed input.
• recordDispatchResult: a skipped heartbeat fire reschedules to the next scheduler tick (1 minute) instead of the next full interval, so a deferred fire retries instead of being lost.

Testing

• New regression test resume-suspended-input-pump.test.ts: queued agent mail after an idle abort delivers; a backlog queued before an abort drains once a later programmatic admission arrives. Both fail without the fix.
• Updated cron-jobs.test.ts skip-reschedule expectation to the new retry-next-tick semantics.
• vitest run on cron-jobs, agent-session-queue, agent-session-prompt, agent-session-action-races: no new failures versus the unpatched baseline.

---

This PR was prepared with agent assistance.

Note

Fix AgentSession.admitSessionAction to resume suspended input pump on programmatic admission

• After an abort suspends the input pump, queued agent messages, heartbeats, and wake prompts would starve. This fix resumes the pump on two paths: coalesced follow-up re-fires and direct programmatic admissions of turn actions.
• Skipped heartbeat cron jobs now reschedule to retry at now + 1 minute instead of waiting for the full schedule interval, preventing long stalls after a skip.
• Regression tests in resume-suspended-input-pump.test.ts cover the three resume scenarios.
• Behavioral Change: skipped heartbeat jobs retry sooner than before; non-heartbeat and once-kind jobs retain prior scheduling behavior.

^{Macroscope summarized 9f77b4d.}

[twaldin]

Filed #1000 with the full analysis: this also explains hour-scale starvation of subagent agent-messages at idle sessions (not only post-abort). The branch now covers the coalesced early-return path too, with a failing-first regression test. -- posted by an AI agent on behalf of twaldin

[sethkarten]

Thank you for the report and proposed work. This root cause is now covered by maintainer-owned stacked PR #1162, authored independently from upstream/main.

We did not inspect or reuse this PR's diff, branch, commits, implementation code, or tests; its public description/comments were used only as a bug report. To keep one review surface, this PR is superseded by #1162 and is being closed.

The complete review stack is #1158–#1165. It is being left unmerged for human review after CI and review-bot findings are cleared.