ci-triage

Shared logic for classifying a failed CI workflow run and recording the outcome. Used by humans (or Claude in an interactive session) when triaging a red main workflow or a red check on an open PR — the latter via the global pr-lifecycle skill, whose CI-triage section delegates the taxonomy here.

This skill owns the taxonomy, the de-dup rules for the main-red issue, and the issue template. Repo-specific reproduction steps and failure patterns live in the consumer repo's CLAUDE.md / <repo>-dev-process (read by pr-lifecycle), and e2e classification is delegated to web-testing.

Taxonomy

Every failure lands in exactly one bucket. Be explicit — "unclear" is not a bucket, needs-human is.

Do not invent a fifth bucket. If the signal you see fits nothing above, it's needs-human.

Suspect commit identification

The GitHub workflow_run payload includes head_sha. That's the commit that triggered this run — on a push: main event, it is the merge commit.

1. Fetch the commit via mcp__github__get_commit.
2. If the commit message matches Merge pull request #N, the suspect PR is #N; its author is the suspect author.
3. Otherwise (direct push to main, squash-merge), use the commit author directly.

Never blame more than one commit per failure. If the previous main commit's CI was also red, link that issue rather than opening a new one.

The rolling main-red issue

One open main-red issue at a time. If main is broken for three days, that's one issue accumulating comments — not twelve.

Before filing:

1. mcp__github__list_issues with labels: main-red, state: open.
2. If one exists, append a comment:

Run #<run-id> also failed. Workflow <name>, bucket <bucket>, suspect <sha-short> (#<pr-or-none>, @<author>). <one-line-cause>.

3. Only open a new issue if none is open. Title: main is red: <workflow> (<bucket>). Labels: main-red, plus infra or needs-human if applicable.

When main goes green again (the next successful run on the same workflow), close the issue with a comment naming the green run id. This close step is manual.

Issue body template

**Workflow**: <workflow-name>
**Run**: <run-url>
**First failed step**: <step-name>
**Bucket**: <regression|flake|infra|needs-human>
**Suspect**: <sha-short> — <commit-subject> (PR #<n>, @<author>)

### Failure excerpt

<last 30 lines of the failing step, or the ripgrep-extracted error block>

### Reproduction

<one of:>
- Deterministic: `<exact command from the workflow yaml>`
- Flake: passed on <prev-sha-short>; rerun button: <run-url>/attempts/2
- Infra: <service name> — <symptom>
- Needs human: <why the logs are ambiguous>

### Next action

<one line — "revert #N and reland", "rerun", "fix <specific thing>", "human eyes">

Keep the excerpt tight. Dumping the full log helps nobody.

Reproducing locally

Someone invoking this skill via the global pr-lifecycle skill should reproduce before filing, using the repo's check gate (named in its CLAUDE.md / <repo>-dev-process). For a main failure caught from outside a PR, check out main at the suspect SHA and run the same commands locally before filing.

For e2e failures specifically, delegate classification to web-testing's triage mode — it handles regression-vs-flake for browser-driven tests (the ambiguous case).

Log access

WebFetch cannot read authenticated GitHub Actions logs (403). The GitHub MCP gives you:

• mcp__github__pull_request_read with method: get_check_runs — step names, status, timings (no log body).
• The workflow run's jobs_url via mcp__github__get_commit + check_suite traversal — same metadata.

Log bodies are not reliably accessible from the GitHub MCP. When the log body is unavailable, classify from step names + exit codes + the workflow yaml, and bias toward needs-human rather than guessing.

Flake-detection heuristic

A failure is flake only if both hold:

1. The same workflow on the previous main commit passed (check via mcp__github__list_commits + check runs on the prior sha).
2. The failing step's logs do not contain a symbol name, file path, or assertion message that appears in the suspect commit's diff.

One of those alone is not enough. A deterministic regression can pass on the prior commit; a real flake can mention a touched file by coincidence.

Constraints

• Never open a PR from this skill. Triage is read-only on the codebase.
• Never @-mention for infra or flake buckets — alert fatigue kills the signal.
• Never close a main-red issue without a green run id to cite.
• Scope: any GitHub-Actions–driven repo. The taxonomy and rolling main-red pattern are repo-agnostic; the consumer repo's CLAUDE.md can override triggers and label conventions.

Relationship to other surfaces