Relax GitHub CI

Inspect the current PR

Identify the repository and PR from the request or current branch. Capture the current head and required checks before choosing an operation:

gh pr view <PR> --repo redai-studio/Relax --json number,url,headRefOid,headRefName,headRepository,headRepositoryOwner,baseRefName,state
gh pr checks <PR> --repo redai-studio/Relax --required

Read .github/ci/config.json for workflow/job targets. Match runs to the PR, head SHA, workflow and head repository; an old run can have a recent update time after a rerun. Inspect logs for failed required or specifically requested checks.

Use exact target keys: a workflow name such as ci, or workflow/job such as ci/pre-commit or integration/gpu-async. ci selects the pre-commit/CPU workflow; all selects every configured workflow. Job targets have no unqualified aliases.

For the selected workflow on the current head, prefer an active run, then a non-cancelled run; use a cancelled run only when none survive. Within each group, choose the highest run number. Concurrency can cancel a higher-numbered duplicate while a lower-numbered run continues. For fork runs with an empty pull_requests array, match both head branch and head repository. Deployment workflows are outside the configured CI targets.

Choose the operation

Read failed logs before deciding whether a retry can help. A deterministic code or configuration error needs a fix; evidence of an infrastructure failure or a flaky test can justify a retry. Prefer failed-job or specific-job reruns over restarting unrelated successful work.

Track retries of the same failure on the same PR head. If more than three reruns still fail, stop retrying, collect the head SHA, affected workflow/job, attempt URLs and representative logs, and report the issue to SigureMo for maintainer investigation; send a message only within the user's authorization. Reclassify the failure after a code change rather than treating the old retries as validation of the new head.

Prefer gh run or the REST API when the authenticated credential has Actions write access. See rerun operations for retrying failed jobs, workflows or individual jobs. To cancel a workflow, use gh run cancel <RUN_ID> --repo redai-studio/Relax or POST repos/redai-studio/Relax/actions/runs/<RUN_ID>/cancel.

Use the narrowest scope that addresses the request. A job rerun includes dependent jobs; workflow cancellation includes sibling matrix jobs. Refresh the PR head immediately before a write and reassess if it changed.

For an explicitly requested exemption, read bypass and CI maintenance. CI team membership and an authorized exemption request are both required.

PR comment interface

Use comment commands when direct Actions access is unavailable and the actor is eligible. Rerun/cancel comments are available to the PR author and users with repository write/maintain/admin permission. The command action silently ignores /rerun, /cancel, and /help comments whose GitHub user.type is Bot, without replying. Put a command on the first line of a new PR comment; later lines are ordinary comment text.

Targets with a job field support rerun only. Use a prepared body file with gh pr comment <PR> --repo redai-studio/Relax --body-file <FILE> and retain the returned URL. Editing a comment does not execute rerun/cancel/help again. The command action records source comment IDs to prevent duplicate execution.

Verify the result

Inspect the affected run after writing. Distinguish a request being accepted from a new attempt starting or finishing. Verify cancellation reaches a terminal state. If an API call times out or returns EOF, read back the run or comment before retrying.

For comment commands, check the reply and its workflow link, then verify the actual run. Neither API nor comment reruns create an absent workflow run or approve a fork workflow.

For automation changes, read the relevant action/workflow and maintenance reference. Use gh-stack for native stacked PR operations.