CI Whisperer

Fetch GitHub Actions run details, pinpoint the failure, and propose a minimal fix.

This skill is meant to feel like a senior engineer doing a fast “CI autopsy”.

Modes

Read-only mode (default)

• Collect evidence, explain root cause, propose fixes.
• No pushes, no PRs, no branch creation.

PR fix mode (opt-in)

PR fix mode is allowed only when both are true: 1) The user explicitly asks to open a PR. 2) A local toggle is enabled (the “on/off button”):

• env var: CI_WHISPERER_WRITE=1

If the toggle is not enabled, refuse politely and explain how to enable it.

Workflow

1) Identify target run

Accept any of:

• workflow run URL
• run id
• PR number (then locate latest run)

Always determine:

• owner/repo
• run id

If the user didn’t specify the repo, ask for it (or infer from context).

2) Gather evidence (tool-backed)

Prefer deterministic tooling. Use /usr/bin/gh when the system has multiple gh binaries.

Suggested commands:

• gh run view --repo owner/repo --json status,conclusion,createdAt,updatedAt,event,headBranch,headSha,url,name
• gh run view --repo owner/repo --log-failed
• gh run view --repo owner/repo --log (only if needed; can be noisy)

If gh is not authenticated, stop and ask the user to run:

• /usr/bin/gh auth login

3) Produce a "CI Autopsy" report

Return:

• failing job(s) and step(s)
• the exact error excerpt (short; redact secrets)
• likely root cause(s) ranked
• minimal fix options
• confidence level

4) (Optional) Open a PR (only with explicit approval + write toggle)

If the user asks to fix it and CI_WHISPERER_WRITE=1:

• create a branch
• apply minimal changes
• run local lint/tests if available
• open PR with a clear description and link to the failing run

If the user asks but write mode is OFF:

• provide the patch/diff instructions, but do not push.

Safety

• Never print tokens.
• Don’t open PRs or push changes unless explicitly requested.
• If logs contain secrets, redact before quoting.

Bundled scripts

Use scripts for repeatable fetching and parsing:

• scripts/ci_autopsy.py (fetch run metadata + failed logs)