Composer: Queue never reaches the agent, and Redirect stops it and fails the step #29

New issue

Closed

opened 2026-09-14 23:47:23 -04:00 by cmoriarty · 1 comment

cmoriarty commented

2026-09-14 23:47:23 -04:00

Owner

The composer's two verbs do not reach the agent.

Queue (⏎) logs human.queue_prompt and nothing delivers it.
Redirect (⌘⏎) stops the agent and fails the step, which fails the run.

What happened (production, soundcheck run `run_01M2HF4SNRBVJ6TA8ZX11T2Z4H`, `openspec.apply[1]`)

time	seq	event
23:29:21	2821	agent runs `npm test 2>&1 \| tail -30`
23:32:30	2824	`human.queue_prompt` "how are the unit tests?" (never reaches opencode: no user message follows)
23:32:52	2825	`human.redirect` "status of the unit test suite?"
23:32:52	2826	`oc.session.error` MessageAbortedError, then `session.idle`
23:32:56	2833	`osf.step.failed` predicate_failed "group 1: 0/2 done"; groups 2–4 skipped, run failed

The redirect's prompt was never sent.

Cause

AgentExecutor.queue_prompt() and AgentExecutor.redirect() (src/osf/engine/executors.py) exist, but nothing calls them:

Routes: step_prompt, step_queue and step_redirect only append the human event.
Scheduler: it maps human.redirect to CancelIntent.INTERRUPT in _CANCEL_INTENTS, so Redirect is exactly Stop. The executor's wait loop returns ABORTED, _finish checks the completion predicate, and the step fails.
human.queue_prompt: it is read by nothing but boot recovery's "undelivered prompts" count, which already expects an executor to deliver it and write admit_prompt.
A second defect in the uncalled methods: both re-append the human event, so wiring them up as-is would loop.

Also seen, not a Braid bug

The agent looked stuck because soundcheck's vitest run takes 5–12 s per test in the production container (events.test.ts alone took 102 s). | tail -30 hid all progress until exit. Worth a separate look at the VM's CPU and PGlite start-up cost.

Done when

human.queue_prompt for a running agent step is delivered to its session (prompt_async, an admit_prompt operation with the same message id) without appending another human event.
human.redirect aborts the turn, then sends the text, and the step keeps running: its wait loop never ends the step on the redirect's own abort, and nothing marks the step interrupted.
Tests cover both: the executor on the event, and the scheduler not treating redirect as a cancel.
The failed soundcheck run is resumed once the fix is deployed.

The composer's two verbs do not reach the agent. - **Queue** (⏎) logs `human.queue_prompt` and nothing delivers it. - **Redirect** (⌘⏎) stops the agent and fails the step, which fails the run. ## What happened (production, soundcheck run `run_01M2HF4SNRBVJ6TA8ZX11T2Z4H`, `openspec.apply[1]`) | time | seq | event | |---|---|---| | 23:29:21 | 2821 | agent runs `npm test 2>&1 \| tail -30` | | 23:32:30 | 2824 | `human.queue_prompt` "how are the unit tests?" (never reaches opencode: no user message follows) | | 23:32:52 | 2825 | `human.redirect` "status of the unit test suite?" | | 23:32:52 | 2826 | `oc.session.error` MessageAbortedError, then `session.idle` | | 23:32:56 | 2833 | `osf.step.failed` predicate_failed "group 1: 0/2 done"; groups 2–4 skipped, run failed | The redirect's prompt was never sent. ## Cause `AgentExecutor.queue_prompt()` and `AgentExecutor.redirect()` (`src/osf/engine/executors.py`) exist, but nothing calls them: - **Routes:** `step_prompt`, `step_queue` and `step_redirect` only append the human event. - **Scheduler:** it maps `human.redirect` to `CancelIntent.INTERRUPT` in `_CANCEL_INTENTS`, so Redirect is exactly Stop. The executor's wait loop returns ABORTED, `_finish` checks the completion predicate, and the step fails. - **`human.queue_prompt`:** it is read by nothing but boot recovery's "undelivered prompts" count, which already expects an executor to deliver it and write `admit_prompt`. - **A second defect in the uncalled methods:** both re-append the human event, so wiring them up as-is would loop. ## Also seen, not a Braid bug The agent looked stuck because soundcheck's `vitest run` takes 5–12 s *per test* in the production container (events.test.ts alone took 102 s). `| tail -30` hid all progress until exit. Worth a separate look at the VM's CPU and PGlite start-up cost. ## Done when - `human.queue_prompt` for a running agent step is delivered to its session (`prompt_async`, an `admit_prompt` operation with the same message id) without appending another human event. - `human.redirect` aborts the turn, then sends the text, and the step keeps running: its wait loop never ends the step on the redirect's own abort, and nothing marks the step interrupted. - Tests cover both: the executor on the event, and the scheduler not treating redirect as a cancel. - The failed soundcheck run is resumed once the fix is deployed.

cmoriarty referenced this issue from a commit

2026-09-14 23:57:42 -04:00

fix(composer): Queue and Redirect reach the agent; Redirect no longer fails the step (#29)

cmoriarty referenced this issue from a commit

2026-09-15 01:02:51 -04:00

fix(composer): a queued message's chip clears when the message is delivered, not when the step stops (#29)

cmoriarty commented

2026-09-26 15:56:39 -04:00

Author

Owner

Shipped in acbc533 and 2ca5a7c, change archived as 2026-09-26-composer-verbs-reach-the-agent (2334877), live in production (4f19d6f).

Queue delivers the message to the agent's session after the current turn. The queued chip clears when the message is delivered, not when the step stops.
Redirect stops the turn in progress and delivers the new message, and the step keeps running. It no longer fails the step.

Verified in production on run run_01M3F8PTF3EWBCB721BP869E9N (soundcheck#32, minimalist), during openspec.propose.proposal:

Queue: human.queue_prompt (seq 54114) was followed by opencode recording it as oc.message.user (54117), and the agent carried on.
Redirect: human.redirect (54187) aborted the turn (MessageAbortedError, 54188), was delivered as oc.message.user (54196), and the step stayed running on attempt 1.

Shipped in acbc533 and 2ca5a7c, change archived as `2026-09-26-composer-verbs-reach-the-agent` (2334877), live in production (`4f19d6f`). - **Queue** delivers the message to the agent's session after the current turn. The queued chip clears when the message is delivered, not when the step stops. - **Redirect** stops the turn in progress and delivers the new message, and the step keeps running. It no longer fails the step. **Verified in production** on run `run_01M3F8PTF3EWBCB721BP869E9N` (soundcheck#32, `minimalist`), during `openspec.propose.proposal`: - Queue: `human.queue_prompt` (seq 54114) was followed by opencode recording it as `oc.message.user` (54117), and the agent carried on. - Redirect: `human.redirect` (54187) aborted the turn (`MessageAbortedError`, 54188), was delivered as `oc.message.user` (54196), and the step stayed `running` on attempt 1.

cmoriarty closed this issue

2026-09-26 15:56:40 -04:00

cmoriarty referenced this issue

2026-09-29 09:44:35 -04:00

A node added to a built-in pipeline stalls runs already in flight #102

cmoriarty referenced this issue from a commit

2026-09-29 10:10:11 -04:00

fix(engine): a run keeps the built-in pipeline it was admitted with, whatever is deployed while it runs (#102)