CI Feedback Loop

After pushing commits to a PR branch, monitor check status, diagnose failures, attempt fixes, and escalate if unable to resolve.

When to Use

• After pushing commits to a PR branch in the auto-advance loop
• When a user asks to monitor and fix CI failures on a PR
• When CI checks fail and the agent should attempt automated remediation

Prerequisites

• gh CLI installed and authenticated with repo scope
• A PR exists with check runs (GitHub Actions, external CI, etc.)

Configuration

Overview

The skill runs a monitor-diagnose-fix loop:

Push to PR branch
      │
      ▼
 Poll checks ◄──────────────┐
      │                      │
  ┌───┴───┐                  │
  │Result?│                  │
  └───┬───┘                  │
      │                      │
 pass → Done                 │
 pending → Wait (backoff) ───┘
 fail → Retrieve logs
            │
            ▼
      Analyze failure
            │
       ┌────┴────┐
       │Fixable? │
       └────┬────┘
            │
       yes → Attempt fix → Push → Re-poll
       no  → Retries left?
                yes → Re-trigger (flaky)
                no  → Escalate (PR comment)

Operations

1. Poll Checks

When: After any push to the PR branch (initial or fix attempt). Implementation: Load operations/poll-checks.md Quick summary: Polls gh pr checks --json with exponential backoff until all checks resolve or timeout.

2. Retrieve Logs

When: One or more checks have failed. Implementation: Load operations/retrieve-logs.md Quick summary: Fetches failure logs via gh run view --log-failed with API fallback.

3. Analyze Failure

When: Failure logs have been retrieved. Implementation: Load operations/analyze-failure.md Quick summary: Classifies failure type (code error, flaky/infra, configuration) and extracts actionable context.

4. Attempt Fix

When: Failure is classified as fixable and retries remain. Implementation: Load operations/attempt-fix.md Quick summary: Agent applies fix, pushes to PR branch, decrements retry counter, re-enters poll loop.

5. Escalate

When: Retry limit reached, timeout exceeded, or failure classified as unfixable. Implementation: Load operations/escalate.md Quick summary: Posts structured PR comment with failure summary, attempted fixes, and next steps for human.

End-to-End Flow

When invoked, run operations sequentially in a loop:

Input

• pr_number (required): PR number to monitor
• repo (optional): Repository in owner/repo format. Defaults to current repo.

Loop

1. poll-checks(pr_number) → status
2. if status == "pass" → return success
3. if status == "timeout" → escalate("Checks timed out")
4. if status == "fail":
   a. retrieve-logs(failed_checks) → logs
   b. analyze-failure(logs) → diagnosis
   c. if diagnosis.fixable AND retries > 0:
      - attempt-fix(diagnosis) → push
      - retries -= 1
      - goto 1
   d. if NOT diagnosis.fixable OR retries == 0:
      - escalate(diagnosis, attempts)
      - return failure

Output

• Success: All checks passed (possibly after fixes). Agent can continue auto-advance.
• Failure: Escalation comment posted on PR. Agent should stop auto-advance and wait for human.