| name | ci-triage |
| description | Decision tree for a red CI run. Use BEFORE retrying — blind re-runs burn the shared GitHub rate limit and hide real bugs. |
CI Failure Triage
Use this skill whenever a PR's CI goes red.
1. Get the signal, not the whole log
Never dump a full failed-run log into context (10-50KB+). Use the
wrapper's key-line extraction (github.ci.failed_log_signal_patterns,
capped by failed_log_max_lines). Pull the raw log only when a human
explicitly asks.
2. Classify the failure
| Symptom | Class | Action |
|---|
psycopg2 ... Connection refused / port 5432 / "Apply migrations" | runner infra (postgres OOM/contention) | Do NOT loop-retry. Escalate / wait off-peak / evidence-based admin override. |
| kube-apiserver readiness 500 / kube-proxy CrashLoop / node not-ready | runner infra (kind cluster) | Same. |
test_openapi_route_coverage_matrix... AssertionError: N == M | matrix drift | git merge origin/<default> into branch (brings updated matrix). Proven fix. |
Type error / tsc / Would reformat / ruff | real code | Fix precisely. Don't retry. |
| compose/kind smoke ✅ but lint-and-unit ❌ on an FE-only PR | code clean → infra | Treat the red as infra; pursue the infra path. |
3. Don't double-trigger
If a monitor/PM script or another agent is already auto-rerunning, do
NOT also push an empty commit or gh run rerun. Coordinate so exactly
one owner pulls the trigger. Empty-commit retrigger re-runs ALL checks
→ waste + runner contention.
4. Persistent failure = escalate, not loop
≥3 failures of the same check across reruns is NOT a transient flake.
Pull the real log, name the root cause, escalate (SRE / off-peak
retry / evidence-based admin override) instead of an infinite loop.
5. i18n / contract sync gotcha
When you i18n a user-facing string (raw → t('...')), grep
contract/snapshot tests for the raw string before merging — they may
assert on it and turn main red after merge.
