plat 589

← crew / functions

PLAT-589: Completed Crew ask call remains running when its trigger binding is unavailable

Field	Value
State	deployed
Priority	P2
Product	crew
Area	functions
Summary	Excellence Crew aq finished its Code ask call, but the caller kept reporting running: saved call remained queued, polling reported a missing internal trigger, and the supervisor ignored lookup errors.

What happened

Read-only investigation on Excellence, 2026-10-06. Both the Code workspace linkedin-scraper-de7a2ec2 and Crew aq-ae5e5fac are registered to the same user, b2bb7148f87b3338355f92e8c59c9c43. The call correctly used the Crew's main chat under the same-owner routing rule.

• Call: fn-42ca174e-8d28-48d0-a16c-9ec558ddb729, implicit ask.
• Trigger: b2fc312e-481f-4659-b819-b6155894ac66.
• Run: b5dc296e-ea71-56ed-bbd8-655eaf6fe2ed.
• Session: work:project:ae5e5fac-4534-4405-bf06-ba306f39ac60.
• schedule.log and Crew/aq-ae5e5fac/schedule-runs.json agree: started 2026-10-06T08:51:55.982887829Z, succeeded 2026-10-06T08:52:31.650257863Z, duration 35,616 ms, with a nonempty final answer saved. Browser timeline records also show the completion painted.
• The persisted function call still has status: queued, no result, and its original updated_at. The user's screenshot shows the caller describing the call as running, with a background agent still waiting.
• At 2026-10-06T10:56:48+02:00, ask_function_update failed with trigger not found on the attached workflow.
• At inspection, the Crew's runtime workflow.json has no triggers. This is not enough to establish which writer removed the binding or exactly when; dispatch had successfully resolved it earlier.

The deployed builder revision was 581ccc6606c2caa3f54c4e25964136d5b45df98d. The same relevant code remains on main at the time of this report: getInternalProductTriggerRun requires a currently enabled trigger before it reads the run; superviseCrewFunctionCall ignores lookup errors; and get_function_call suppresses them as well. Thus an unavailable binding can prevent a completed run from settling the function call, while the caller continues to receive its old status without the lookup failure.

Global Monitor is mounted for Crew/Code and has no per-user disable flag in the inspected code. It filters out the currently selected session and hides when no activity remains. No monitor-paint diagnostic was found, so this investigation does not establish why its button was absent on the user's screen. A completed Crew run should disappear from it; the screenshot's background-agent footer can independently remain while the call watcher waits.

Fix

The writer was identified: at 08:51:56Z, ReadWorkflowManifest treated Crew/aq-ae5e5fac as an AgentWorks workflow and pruned its triggers and identity. Its protection recognized only Chats/Work/projects/..., not the new shared Crew root. The manifest changelog records that removal at the exact start of the call. The same destructive read was reproduced locally.

• Recognize Crew runtime paths through workspaceref.AnyCrewProject, keeping product-owned manifests read-only under both old and new path formats.
• After three consecutive run-lookup failures (normal polling: about 20 seconds from the first failed lookup), fail the call with the actual error and release its notification watcher. Successful reads reset the retry count.
• get_function_call reports lookup errors instead of returning a stale status. Existing binding/caller authorization and revocation checks remain in place.
• A combined regression drives a same-owner private Code ask through its tool, durable trigger/run record, Crew runtime read, function supervisor and caller notification registry. It verifies the answer and saved completion, and that both success and binding revocation settle the background watcher. The model answer is supplied by the test; this is not a live model end-to-end check.
• Extended the existing manifest-preservation regression to shared Crew roots. Relevant existing function, trigger, authorization and browser tests pass. Restoring the old recognizer reproduces the lost trigger and missing answer.

Existing Excellence call recovered

The owner registry, source and target IDs, successful saved run and destructive manifest changelog were checked before restoring this call's single missing binding. No task was reissued and no other binding was changed. Backup: /srv/agents/state/repairs/plat-589/aq-workflow-before-20261006T093348Z.json. The existing supervisor picked up the saved answer: call status became completed, with a result, at 2026-10-06T09:33:58.045295892Z.

Left

• Monitor absence was not independently reproduced. The stale call/watcher issue is fixed and the existing call is recovered; if the monitor is absent during future live background work, capture the header response and selected session then.

Deployment evidence — 2026-10-06

Excellence release agents-99dc2842-20261006115535 records builder revision 99dc284251, which contains this fix. Public and agent health checks passed; slot self-test completed with 156 passed, zero failed, 16 skipped.