CI Sweeper Triage Loop

Use this skill to identify repeated or impactful CI failures and classify bounded automated-fix candidates. Pair it with skills/loop-engineering/SKILL.md for global loop safety rules.

Inputs

• Failed checks list with messages and logs.
• Commit that introduced failures, when available.
• Retry count and history from prior runs.
• Current ci-sweeper-state.md, loop-run-log.md, and loop-budget.md.
• Current denylist, allowlist, and pause flags.

Output

Write only actionable findings:

## High Priority
- Check:
  Root cause class: lint | test | type | runtime | dependency | config
  Repro confidence: high | medium | low
  Minimal fix path:
  Attempt:

## Watch

## Escalate
- Flakes:
- Infra/noise:
- Repeated failures with no stable fix:

## Suggested Loop Action
- open worktree + minimal-fix
- ESCALATE_HUMAN

Append a short run summary to ci-sweeper-state.md and one JSON line or table row to loop-run-log.md, matching the repository's existing loop-log format.

Rules

• Cap attempts at 3 per item.
• Reproduce or identify the failure class before proposing a fix.
• Prefer the smallest diff that addresses the failing check.
• Do not auto-fix dependency-major upgrades, auth, storage migrations, attestation/snapshot code, release publishing, production infrastructure, or other denylisted paths.
• Route flakes, infrastructure failures, repeated low-confidence failures, and exhausted attempts to ESCALATE_HUMAN.
• For low-risk candidates, create an isolated worktree and dispatch a minimal-fix implementer followed by an independent verifier.

Verification Gate

The verifier must check:

• The original failing command or closest local equivalent.
• Relevant tests, lint, type checks, or build commands.
• Scope diff against the declared minimal fix path.
• The full Engram gate when agent, script, CI, connector, dependency, or policy surfaces changed: bash docs/harness/bin/sensors.sh.
• The denylist before any PR, merge, or external write.

Completion Gate

Mark the run complete only when the state files include the attempt count, confidence, evidence, next action, and clear escalation for denied or uncertain items.