| name | report-incidental-flakes |
| description | Activate when a CI pipeline flaked and you need to re-run it. Best-effort -- before re-running, make sure the flake you just hit is tracked in the MIND backlog, so incidental flakes are not lost until the next scheduled sweep. Safe for coding/babysitting agents without Linear access; it degrades gracefully. |
Report an incidental flake
You hit a flake in a CI pipeline and are about to re-run it. Take that moment to make sure the flake is tracked, so the evidence survives your re-run. Best-effort: never let this block your real task.
You saw one flake, not a full sweep, so you have no authority to close flake tickets -- create or update only.
Steps
1. Verify Linear access -- degrade gracefully without it
latchkey services info linear should report "valid". If you cannot reach Linear (common for a coding agent):
- Still highlight the flake in your response and mark the test
@pytest.mark.flaky -- the repo's CLAUDE.md already mandates both, with or without this skill.
- Note that Linear filing was skipped; the scheduled
detect-flakes sweep will pick the flake up from CI later.
- Stop -- the rest of this skill needs Linear.
2. Confirm the flake is a pre-existing test, not your own code
Only flakes in code that already lives on main belong in the shared backlog. Check with git alone (no Linear):
- The failing test exists on
origin/main -- you did not add or rename it on this branch.
- The failure is not attributable to changes this branch made to that test or to the code it exercises (
git diff origin/main -- <relevant paths>).
If either check fails, the failure is most likely your bug, not a repo flake: do not file it -- debug it as part of your own work -- and stop here.
3. Hand the flake to manage-flakes as a partial set
Capture the evidence now, before the re-run replaces it: the failing test's id, the failure output, the branch CI ran on, and when. Assemble that as one flake record in the input shape the manage-flakes skill documents, then apply that skill, telling it this is a partial set -- one incidental flake, not a full window -- so it may create or update tickets but never close. It dedups against existing tickets and does the filing; read it for the rest.
