Operator read surface: system health, issue state, and last-progress on in-flight runs #2010

Open
opened 2026-11-14 16:46:57 +00:00 by sam · 0 comments
Owner

What's missing

The chat operator surface has no read path for three things an operator is routinely asked to judge. Evidence is from tasks 471 and 473 in this instance (observation window 2026-11-14T16:41Z–16:46Z).

1. No system/health read

There is no tool that reads daemon or service state, and settings_get exposes no status kind. The dashboard has /settings/status, but the operator surface cannot read it, so questions like "is the daemon up / is it accepting work" cannot be answered from the operator surface at all.

2. run_events carries no last-progress signal

run_events returns stage transitions only. For task 471 attempt 4 the trail ends at:

2026-11-14T16:41:49.126241Z  stage_start  baseline-fix

with nothing after it. At the same moment task_logs for the same task shows continuous agent model and tool calls:

2026-11-14T16:46:33.585511592Z  INFO  agent model call  stage=baseline-fix
2026-11-14T16:46:37.201616079Z  INFO  agent model call  stage=baseline-fix
2026-11-14T16:46:37.225720549Z  INFO  agent tool call   tool=shell

Read on its own, the event trail is indistinguishable from a run that hung 5 minutes ago. In this session that ambiguity produced a false report of a multi-hour stall, retracted only after a separate task_logs call. Each stage transition (or the run record) should carry the timestamp of last observed activity for the in-flight stage, so a single read answers "is this moving".

3. No single-issue read for state

issue_list is the only way to see an issue, and fetching one issue by number returns 404 on a private repo. Confirming that a task's target issue is still open requires paging the whole open-issue list, and absence from the list is the only evidence of closure. A single-issue read returning state (open/closed, labels, title) is needed.

Related but distinct: #2009 covers workflows re-checking issue state before running; this issue is about the operator surface being able to read it.

Done when

  • A health/status read is available to the operator surface for the daemon and its services.
  • run_events (or the run summary) reports last observed activity for the in-flight stage.
  • One issue can be read by number, including its open/closed state.

Provenance

Reported from chat session 01a1d476-cbbc-764b-b566-bff6e4161d58, which had to reconstruct liveness from task_logs after misreading the run_events trail for task 471.

## What's missing The chat operator surface has no read path for three things an operator is routinely asked to judge. Evidence is from tasks 471 and 473 in this instance (observation window 2026-11-14T16:41Z–16:46Z). ### 1. No system/health read There is no tool that reads daemon or service state, and `settings_get` exposes no status kind. The dashboard has `/settings/status`, but the operator surface cannot read it, so questions like "is the daemon up / is it accepting work" cannot be answered from the operator surface at all. ### 2. `run_events` carries no last-progress signal `run_events` returns stage transitions only. For task 471 attempt 4 the trail ends at: ``` 2026-11-14T16:41:49.126241Z stage_start baseline-fix ``` with nothing after it. At the same moment `task_logs` for the same task shows continuous agent model and tool calls: ``` 2026-11-14T16:46:33.585511592Z INFO agent model call stage=baseline-fix 2026-11-14T16:46:37.201616079Z INFO agent model call stage=baseline-fix 2026-11-14T16:46:37.225720549Z INFO agent tool call tool=shell ``` Read on its own, the event trail is indistinguishable from a run that hung 5 minutes ago. In this session that ambiguity produced a false report of a multi-hour stall, retracted only after a separate `task_logs` call. Each stage transition (or the run record) should carry the timestamp of last observed activity for the in-flight stage, so a single read answers "is this moving". ### 3. No single-issue read for state `issue_list` is the only way to see an issue, and fetching one issue by number returns 404 on a private repo. Confirming that a task's target issue is still open requires paging the whole open-issue list, and absence from the list is the only evidence of closure. A single-issue read returning state (open/closed, labels, title) is needed. Related but distinct: #2009 covers workflows re-checking issue state before running; this issue is about the operator surface being able to read it. ## Done when - A health/status read is available to the operator surface for the daemon and its services. - `run_events` (or the run summary) reports last observed activity for the in-flight stage. - One issue can be read by number, including its open/closed state. ## Provenance Reported from chat session 01a1d476-cbbc-764b-b566-bff6e4161d58, which had to reconstruct liveness from `task_logs` after misreading the `run_events` trail for task 471.
sam added the mediumfeature labels 2026-11-14 16:46:57 +00:00
sam added the deferred label 2026-11-15 17:26:08 +00:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: sam/archie#2010