retrigger-flaky-ci
Triage failed CI on a pull request or the active branch, decide whether the failures come from its changes, and rerun only failures that are unrelated. Use when asked to retrigger, rerun, or kick CI on a PR with failing checks.
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 2.2k
- Forks
- 349
01 前言 Front Matter
- name
- retrigger-flaky-ci
- description
- Triage failed CI on a pull request or the active branch, decide whether the failures come from its changes, and rerun only failures that are unrelated. Use when asked to retrigger, rerun, or kick CI on a PR with failing checks.
03 SKILL.md 正文
Retrigger flaky CI
Use this skill when a user asks to retrigger, rerun, or restart CI for a pull request, or for the PR backing the current branch. The request authorizes a rerun of failed jobs on that PR only. It does not authorize pushing, editing the PR, changing labels, or rerunning other PRs.
Never rerun first and investigate later. Read the failure, compare it with the PR's changes, then either warn the user or rerun.
1. Resolve the PR and its changes
Resolve the PR by number, or from the current branch when no number is given:
gh pr view <number> --json number,title,headRefName,headRefOid,files,statusCheckRollup,url
gh pr view --json number,title,headRefName,headRefOid,files,statusCheckRollup,url
Record the PR head SHA and the changed file list. When the request refers to the active branch, also compare local work with the PR:
git rev-parse HEAD
git status -sb
git fetch origin main --quiet
git diff --stat origin/main...HEAD
If local HEAD differs from the PR head, or there are uncommitted changes, the
CI run did not test them. Say so, and judge the failure against the PR head
that CI actually ran.
Pull request runs of Build and Test check out refs/pull/<number>/merge, so
they include the current main. A failure can come from main rather than from
the PR.
2. Read the actual failures
List the failing checks from statusCheckRollup. CI Gate only aggregates the
required jobs. Look at the jobs it reports as failed instead of CI Gate itself.
If checks are still queued or running, report that and stop. A partial run is
not grounds for a rerun.
Read the failed job logs:
gh run view <run-id> --log-failed
gh run view <run-id> --job <job-id> --log-failed
Find the specific failure in each failing job: [FAIL] and Failed <TestName>
lines with their Error Message and stack trace, error CS or error MSB
build errors, Test host process crashed, validator output, or the step that
reported ##[error]. Record the test or step, the error, and the top stack
frames that belong to the repository.
3. Decide whether the PR caused the failure
Treat a failure as related if any of the following is true:
- It is a build, analyzer, or compile error in a changed file, or in code that calls an API the PR changed.
- The failing test file is in the PR diff, or the test directly exercises a production owner, method, string resource, or contract the PR changed. Map the repository frames in the stack trace to the diff.
- It is a deterministic assertion about behavior, copy, localization, architecture ledgers, docs, skills, proof pools, or CI classification that the PR touched.
- It fails the same way on more than one attempt of the same head SHA.
- It reproduces locally on the PR head with a focused run, for example
dotnet test <project> --filter "FullyQualifiedName~<TestName>".
Treat a failure as unrelated only with positive evidence, such as:
-
The failing code path does not touch anything in the diff, and the diff adds no startup, background, or Gateway traffic that could reach it.
-
The same test or job fails on
mainor on unrelated PRs:gh run list --workflow "Build and Test" --status failure --limit 25 --json databaseId,headBranch,event,createdAt,displayTitle gh run view <other-run-id> --log-failed -
Retained CI evidence or an existing issue shows the same failure on the recorded base revision, and comparison confirms the PR does not introduce it.
-
It has the shape of a timing or infrastructure failure: a UI Automation element briefly missing (
NullReferenceExceptioninside a wait lambda), aTimeoutExceptionfrom a fixture wait, a wall-clock limit on a loaded runner, a crashed test host, a network or restore error, a runner shutdown, or a missing hosted tool.
A timeout counts as flaky only when the waited-for behavior is outside the diff. A timeout in a path the PR changed is a related failure until it is reproduced or ruled out locally.
If the evidence is mixed and a focused local run is possible, run it. If it remains ambiguous, do not rerun. Report the uncertainty and ask the user.
If the failure is unrelated but deterministic on main (it fails on every
recent main run), a rerun will not help. Report it as a main breakage
and do not rerun.
4a. Related: warn and do not rerun
Do not rerun. Tell the user that the failures come from the PR's changes. For each failure, give the job, the test or step, the error, and the changed file or owner that connects to it. Do not edit code unless the user asks for a fix.
4b. Unrelated: summarize, then rerun failed jobs
Show a short table before rerunning:
| Job | Failing test or step | Error | Why it is unrelated |
|---|
Then rerun only the failed jobs of each failed run. Rerunning a job also
reruns the jobs that depend on it, including CI Gate:
gh run rerun <run-id> --failed
gh run view <run-id> --json status,attempt,url
Report the run URL and the new status. Do not wait for the rerun to finish
unless the user asks. If gh run rerun is refused, for example because of
missing permission or a fork run that needs approval, report the error as is.
Do not rerun the same failure on the same head more than once. A second
identical failure on the same SHA reclassifies it as related, or as a real
main breakage. Stop rerunning and report the evidence and blocker. When
authorized to fix it, investigate the cause, implement the repair, and validate
it. Do not create a checked-in failure backlog.
Report
- PR number and title, head SHA, and whether local work matches it.
- Each failing job with its test or step and error.
- The related or unrelated decision for each failure, with its evidence (diff overlap, matching baseline failures, existing issue evidence, local reproduction).
- The action taken: no rerun with a warning, or the rerun command and the new run status.
04 同仓库更多 Skills
network-handoff
Use before generating, changing, or reviewing networking, IPC, or cross-process handoffs, including requests, retries, deadlines, reconnect, process launch, configuration apply, cancellation, and cleanup. Establish confirmation, identity, authority, reconciliation, and producer compatibility before implementation.
technical-documentation
Write, review, or restructure Windows node documentation and agent instructions using current source, task-focused explanations, and explicit fact preservation.
pr-simplify
Simplify your own Windows-node change to its smallest complete fix while preserving distinct regressions, platform contracts, and required proof. Use after implementation before PR preparation, after large repairs, or when asked to shrink a change or remove slop.
openclaw-testing
Choose proportional OpenClaw Windows node tests, builds, packaging checks, and runtime proof, then diagnose failures without weakening required gates.
openclaw-proof-validation
Plan and collect OpenClaw Windows validation/proof: tests, rubber-duck review, UI evidence, MCP output, and gateway runtime proof.
openclaw-debugging
Debug OpenClaw Windows Companion, setup, Gateway connection, node capabilities, packaged permissions, updates, installers, and system.run by selecting the failing boundary and proof path before changing code.