Diagnose Comet PR CI Failures with comet-github-ci-triage

Where this fits

You open a Comet PR and see a red GitHub Actions job. The agent summarizes the failure from the PR page, but the head has moved, the Codecov patch report is incomplete, and the local checks are green. That is exactly when comet-github-ci-triage earns its place: it makes the agent verify the true head SHA, pull failed job logs, and classify the failure before anyone writes a fix.

Why agents benefit

• Resolves CI checks against the current head SHA and mergeable status, so expired runs can't be mistaken for valid failures.
• Pulls job-level logs with gh run view --log-failed for evidence-based root causes.
• Classifies red jobs into production code, fixtures, runtime drift, toolchain, credentials, or coverage threshold issues.
• Reproduces locally only when a fix is authorized, and separates Windows sandbox noise from deterministic failures.
• Reports verified vs unverified states, preventing premature 'green' claims.

Practical scenarios

Stale red job after a force-push

A PR's latest commit is clean, but GitHub still shows a failed job from an older SHA. The skill flags the expired run as 'unverified' — no product change needed.

Codecov patch threshold fails without clear data

Codecov shows a coverage failure, but the page lacks file-level detail. The skill queries APIs for affected files and patch coverage metadata, then reports whether the threshold miss is real for the current head.

Local repro passes but CI stays red

A Windows job fails on path length or a missing placeholder executable, while local sandbox is green. The skill distinguishes runner/config noise from real production-code failure, so the agent doesn't 'fix' code that was never broken.

Add it to your agent workflow

Install the skill with:

npx skills add rpamis/comet --skill comet-github-ci-triage

Then ask the agent:

Use comet-github-ci-triage to diagnose PR #123. Do not modify code; report current head, failed jobs, root cause, and minimal action.

The skill first reads the maintainer contract, then fetches accurate remote state and logs. Only if the user asks for a fix does it reproduce locally and hand the confirmed scope to a fix skill.

Compare before adopting

Before adding a CI triage skill, compare evidence handling: Does it verify current head SHA and raw logs, or trust stale PR summaries? Check maintenance freshness — Comet is active, with 2.9k stars. Generic 'fix CI' prompts often miss Comet-specific Codecov behavior or Windows runner noise. Finally, consider workflow fit: the skill should read local repo state, call gh, and respect a 'diagnose first' boundary.

Why it is worth tracking

comet-github-ci-triage is worth watching for its strict diagnostic contract: verify remote state, cite logs, classify failure, and only then fix. That discipline reduces wasted actions on noisy CI. Track it alongside related skills on OpenAgentSkill; the stable index is https://www.openagentskill.com/skills/rpamis-comet-github-ci-triage. Re-evaluate if Comet's Actions or coverage workflows change, since the skill's value depends on matching those exact failure modes.