Repository navigation
Description
Summary
When a task chat checks back on its own work with defer (for example "CI is running, I'll check in 15 minutes and report"), the result of that wake-up never reaches the Concierge. The Concierge runs no turn, so the user never hears the result unless they ask.
The wake-up turn has no input row in the task chat, so it is classified by=system. A by=system turn that ends with no question and no approval is marked passive, and a passive task change never flushes on its own. This came in with #2011.
What happened (v1.1.198)
On one instance, two task chats were watching CI on PRs they had opened. Both had been sent work by their Concierge chat (by=concierge), and both said they would check CI later and report back.
- Task chat A (feat: basic vault for guardian-entered secrets held by the signer #2239): its own deferred wake-ups at 15:07 and 15:18 UTC posted "one check failed, rerun" and then "CI is green, ready to merge". The task row ended
donewith the new summary. - Task chat B (feat(web): focus a Concierge chat thread in place by filtering the chat to it #2244): its wake-ups at 14:55 and 15:13 UTC posted "CI 13/13 green, all review bots approve".
- The Concierge side: neither Concierge chat ran a turn between the task chat's earlier
by=conciergereport and the user asking "what's the status?" 15–40 minutes later. The user's turn had no<task_events>in its prompt. - For contrast: earlier deferred wake-ups of the same task chat that ended with a question card did reach the Concierge. Their change was
turn_end:completed question by=system, which is urgent.
From the 11 deferred wake-ups into task chats since 2026-11-12, every one that ended without a question (4 of 4) produced no Concierge turn.
Evidence
Tables:
routines/routine_runs: the wake-ups areresume_sessionroutines created by the task chat'sdefer. Theirchat_idis the task chat andfired_byisschedule.chat_messages/agent_turn_messages: the wake-up turns have only ananswerrow. There is noprompt/user row in the task chat.concierge_tasks: after the last wake-up,last_outcome = succeededandended_atare set, butoutcome_unsaid = 0. The result is not treated as owed.rome_agent_messages(roletrace): the Concierge chats have no turn in that window. They also show no failed turn (concierge_message_turns.turn_failed = 0), and nothing inaction_executionsfailed. This is not a lock or interruption issue.
Source (tag v1.1.198):
packages/core/src/concierge/events.ts:672:by = started ? "concierge" : quietRun || userRows.length === 0 ? "system" : "guardian". A turn with no input rows, which is what adeferwake-up is, becomessystem.events.ts:694-696:passive = nothingNew || (by === "system" && event.status === "completed" && !carriesApproval && !question).events.ts:704:setLastOutcome(..., { reported: !passive }). The result is not marked unsaid, so the restart catch-up won't tell it either.events.ts:754-758:urgentlooks only atstarted, and the change carriespassive: true.packages/core/src/concierge/task-events.ts:300:if (entry.passive) return Number.POSITIVE_INFINITY;. A passive row never flushes by itself. It goes out only with some other row's flush, or in a digest.packages/core/src/concierge/service.ts:1256-1261: the digest lists it only as<ran_on_schedule task title/>, without the conclusion.packages/core/agents/concierge.yaml:115: even when delivered,by=systemis described as "a scheduled wake-up worked in that chat directly. Stay silent unless …".packages/core/src/concierge/service.ts:1509-1536: the Concierge's own scheduled task runs post their brief asCONCIERGE_ACCOUNT, so they do have an input row. In practice,userRows.length === 0only matches wake-ups the task chat set itself, such asdefer→resume_session.packages/core/src/concierge/conclusion.ts:117: the task chat is told "Checks you defer while you work on the task are yours to keep". Following that instruction is exactly what makes its result invisible.
Repro
- In a Concierge chat, start a task. In the task chat, ask it to watch something and report back later. For example: "push, then check CI in a few minutes and report".
- The task chat answers ("I'll check in N minutes") and calls
defer. - When the wake-up fires, the task chat posts its result with a conclusion and no question card.
- Expected: the Concierge gets
turn_end:completed by=conciergeand forwards the conclusion.
Actual: no Concierge turn. The change sits as a passive in-memory mark until some other flush or a digest.
Proposed fix (one mechanism, ~10 lines)
- A turn with no input of its own carries on the task chat's previous input. In
events.ts, whenuserRows.length === 0, read the chat's latest user row with the existing webchat repo and apply the same test:- said by
CONCIERGE_ACCOUNT, or an answer to the task's card →by=concierge; - typed by the user →
by=guardian.
quietRun(awhen_it_mattersscheduled task run) stayssystem.
- said by
- Set
urgentfromby === "concierge"instead ofstarted, so the inherited report flushes within seconds like any other Concierge-started result. - Prompt: in
concierge.yaml:115, drop "or a scheduled wake-up". Wake-ups now inherit their task'sby, and the existingby=conciergerule already forwards the conclusion.
No new table, column, env var or setting. A prompt-only change cannot fix this, because the change never reaches the Concierge.
Activity