Skip to content

Daily AIC guardrail fails closed for 24h after a re-run whose agent job was cancelled ("Cannot verify the agent producer artifact for its job attempt") #67970

Closed

Description

@ThomasStarkInc

Version

gh-aw v0.89.15 (github/gh-aw-actions@045beb2), workflow with max-daily-ai-credits: 2000.

What happened

  1. Run A of a pull_request workflow: attempt 1 — agent job failed and uploaded the agent artifact; conclusion uploaded usage.
  2. Run A was re-run (attempt 2). The attempt-2 agent job was cancelled before it uploaded a new agent artifact; the attempt-2 conclusion job uploaded a new usage artifact.
  3. Every later run of the workflow in the repo then failed in activation → "Check daily workflow token guardrail":
[daily-workflow-aic] Selected guardrail artifact: {"runId":<A>,"artifactName":"usage"}
[daily-workflow-aic] AIC inspection failed: {"status":"transient_error","error":"Cannot verify the agent producer artifact for its job attempt"}
##[error]Daily workflow AI Credits are unknown: Cannot verify the agent producer artifact for its job attempt

Re-running the blocked runs reproduces it deterministically. It only stops once run A leaves the rolling 24h window (or run A is deleted — deleting it fixed the next run immediately).

Why (from setup/js/daily_aic_component_coverage.cjs at the pinned SHA)

  • Component selection prefers the newest non-skipped attempt, so agent resolves to the cancelled attempt-2 job.
  • sumCoveredComponents (attempt > 1) requires a producer artifact named agent created between that job's started_at and completed_at. The only agent artifact is from attempt 1, so it throws.
  • The error is reported as transient_error, but it is permanent for that run, and the guardrail fails closed for all runs of the workflow for 24h.

Expected

A cancelled re-run of the agent job should not block every later run for 24h. For example: fall back to the latest attempt that has a verifiable producer artifact, or count a cancelled attempt with no producer artifact as zero / bounded usage, or surface it as a non-transient, per-run accounting gap with a warning instead of failing closed.

Notes

v0.89.21 adds provesJobExecutionNotStarted (cancelled or failed job with no runner and no steps → zero). That may not cover an agent job that was cancelled after it was assigned a runner, which is likely the case here. We couldn't confirm, because we deleted run A to unblock the repository.

Activity

  1. locked and limited conversation to collaborators on Oct 12, 2026
  2. unlocked this conversation on Oct 12, 2026
  3. ThomasStarkInc commented on Oct 12, 2026

    @ThomasStarkInc
    Author

    Additional trigger, same family: this also happens on attempt 1 when the workflow's own concurrency group (cancel-in-progress: true) cancels a PR run mid-agent because of a quick follow-up push. Every later run then fails in activation for 24h:

    [daily-workflow-aic] AIC inspection failed: {"status":"transient_error","error":"Missing accounting for executed agent component in run <A> (attempt 1, job <J>, conclusion cancelled): agent/token_usage.jsonl is missing; ..."}
    ##[error]Daily workflow AI Credits are unknown: Missing accounting for executed agent component ...
    

    So two pushes in quick succession on a PR are enough to block the workflow repo-wide for 24 hours. Deleting the cancelled run unblocks it immediately again. A cancelled agent job with no accounting file could be treated as bounded or unknown usage for that run only (warning), rather than failing closed for every subsequent run.

  4. locked and limited conversation to collaborators on Oct 12, 2026
  5. unlocked this conversation on Oct 12, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions