Skip to content

A job stopped at its timeout concludes cancelled, so neither a required check nor a Run reads it as timed_out #388

Closed

Description

@don-petry

Under #182. Trial P-42 failed on one clause, and its rule says a fail opens an issue against the design before any code depends on it.

What P-42 saw (second machine-user window, 2026-10-28, canary pull request 65)

p42-checks.yml held one job, p42 / timed-out, to timeout-minutes: 1 with a two-minute sleep. It ran on push and on pull_request. The Maintainer cancelled the pull_request run by hand; nobody cancelled the push run (40620020164).

Read (2026-10-29, by the Maintainer's login) What GitHub says
The push run's jobs p42 / timed-out started 02:17:12Z, completed 02:18:42Z: conclusion cancelled, its sleep step cancelled
That job's check run (122591519154) conclusion cancelled, the same at 02:21Z and 03:53Z
Its annotations failure: "The job has exceeded the maximum execution time of 1m0s", then "The operation was canceled."
The workflow run conclusion cancelled, although its other jobs succeeded

So a job GitHub stops at its timeout concludes cancelled, and so does its run. No check and no run concluded timed_out. The fixture is trials/P-42.json (verdict fail; every other clause of P-42 held).

Where the design reads timed_out that GitHub does not send

  • A required check (§4.2's table of check conclusions): "A timed_out check is a failure, and goes back to the author." A Lens job that runs past its timeout reads instead as cancelled, which the Kernel defers as check_unreadable and shows as needing the Maintainer.
  • A Run (§3.10, the record of a Run): a Run "its conclusion failure or timed_out, is read as item, an attempt", and "a Run cancelled with no outcome is no attempt: it is read as superseded". The design notes that "a job killed at its timeout concludes timed_out" waits on P-18. It does not, so a Run that runs past its job timeout without writing an outcome is read as superseded. The attempts bound never counts it, and only the Run ceiling stops it.

The ruling (orchestrator, 2026-10-29): (a), on a condition

  • (a) Each job the Factory writes (the Run's, and each Lens it owns) stops its own work under a step timeout-minutes shorter than the job's. The job then concludes failure, which the Kernel already reads as an attempt, or as a failing check. tools/check_workflows.py holds each such workflow to a step limit below its job's.
  • The condition. (a) assumes a step's own timeout-minutes concludes the job failure, not cancelled, and no trial has asked that yet. It is built only if the trial below reads failure.
  • The fallback, (b), if that trial reads cancelled too: the records port reads a cancelled job whose started and completed times reach its timeout-minutes as timed_out. Those are facts, not text.
  • Not a candidate: reading the annotation's text. No text from outside reaches a decision.

Done when

  • A small trial on the canary: one workflow whose step timeout-minutes is shorter than its job's, the step running past it. Read the step's, the job's, the check run's and the run's conclusions, and keep the fixture.
  • If it reads failure: (a) is built, with the check that holds every Factory workflow to it.
  • If it reads cancelled: (b) is built in the records port, from the job's times against its timeout-minutes.
  • The design's §4.2 table and §3.10 say what a timeout concludes, and P-18's row says which part P-42 answered.

P-18, partly answered

P-18 asks whether a final step still runs, and uploads partial usage, when a job times out or is cancelled, and whether a job whose step exits non-zero concludes failure. P-42 answers neither question, but it settles the premise §3.10 leaves to P-18: a job killed at its timeout concludes cancelled, not timed_out. Its log showed the job's own "Complete job" step succeed after the sleep was cancelled; whether a final step of the workflow's runs is not read here.

What this blocks

The Kernel's reading of a Run that times out (the attempts bound), and a Lens job's timeout as a failing check. P-18's first question, whether a final step still runs and uploads partial usage when a job times out, is unanswered by this window.

Activity

  1. don-petry commented on Oct 29, 2026

    @don-petry
    ContributorAuthor

    The trial box: a step's own timeout concludes failure (2026-10-29)

    Run on canary pull request 73: three jobs, each with a step timeout-minutes: 1 running sleep 120, under a job timeout-minutes: 5.

    Job Step Job Check run A later if: always() step
    no continue-on-error failure failure failure ran
    continue-on-error: true on the step success success success ran
    continue-on-error: true on the job failure failure failure ran
    • Each step's annotation reads "The action '…' has timed out after 1 minutes."
    • The run concluded failure. Job-level continue-on-error changes only how the run counts that job, which this run cannot show apart, since its first job failed anyway.

    So rule (a) holds. A job that stops its own work under a step limit shorter than the job's concludes failure, which the Kernel already reads as an attempt, or as a failing check. continue-on-error on that step must not be allowed, since it turns the timeout into success.

    The fixture is trials/P-42.step-timeout.json, on the trials lane's P-42 pull request (record 0091). The canary pull request is closed and its branch deleted.

    🤖 Addressed by Claude Code

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions