CI Automation

CI Automation

Monitors CI for an existing PR, classifies failures, and spawns a debugger agent to fix regressions. PR creation and pushing is handled by the user — this skill only looks after CI.

Reference for the CI loop. The root SKILL.md covers the status vocabulary — this file has the debugger format and allowlist.

Actions by Status

"done" + phase: "ci_done"

All green. Run gh pr ready <N> to mark the draft PR ready for review, then CronDelete all CI cron jobs. Truly finished.

"schedule_cron" + phase: "ci_monitor"

CI still running. CronCreate with the on_complete fields (cron_interval, prompt, durable=true). Save the job ID, session exits.

"need_agent" + phase: "ci_debug"

1. Spawn debugger via Agent tool: run_in_background=true, mode=bypassPermissions
2. Capture agent_id from result
3. Wait for completion, then save the result JSON to the feed_file path from the task spec (use the Write tool), and run: python orchestrator.py <file> --ci-check --feed debugger --feed-file <feed_file>

Why --feed-file, not echo ... | python ...: the Bash permission matcher is whole-string prefix matching — echo '{...}' | python orchestrator.py ... matches no Bash(python *) allow rule and is blocked in Auto/restricted modes. A plain python ... --feed-file <path> command matches and is auto-approved.

Permission caveat for the debugger spawn: some harnesses ignore the Agent tool's mode parameter, so the debugger inherits the parent session's permission mode instead of bypassPermissions. And an explicit permissions.deny rule for git commit/git push blocks the debugger in every mode. The orchestrator now fails fast with an actionable message if it detects such a deny — if you see that error, allow git commit/git push (see allowlist below) and re-run.

Debugger Result Format

{
  "agent_id": "c3fa28...",
  "agent_name": "debugger",
  "fixes_applied": [
    {
      "check_name": "pull / linux-bionic-cuda12.1 / test_ops",
      "failure": "Description of the failure",
      "verdict": "caused_by_us",
      "rationale": "Why the refactoring caused this",
      "change": "What code change was made"
    }
  ],
  "unrelated": [
    {
      "check_name": "pull / win-vs2019 / test_ops",
      "failure": "Description",
      "verdict": "unrelated",
      "rationale": "Why this is not our fault"
    }
  ]
}

• agent_id: Required after spawn.
• agent_name: Must be "debugger".

Max Fix Rounds

5 rounds max. On limit: orchestrator leaves a PR comment and marks ready.

Required Bash Allowlist

{
  "permissions": {
    "allow": [
      "Bash(gh:*)",
      "Bash(git:*)",
      "Bash(lintrunner:*)",
      "Bash(python orchestrator.py:*)"
    ]
  }
}

git commit / git push must not be denied. The debugger pushes its fixes, and a permissions.deny entry for Bash(git commit *) / Bash(git push *) blocks the agent in every permission mode (deny wins over allow and bypassPermissions). If your setup has git guardrails (e.g. the git-guardrails skill) that deny these, either:

• remove them from permissions.deny (add to permissions.allow instead), or
• expect CI automation to stop at the "fixes ready" stage and push manually.

The orchestrator emits an actionable error at the start of the debug phase if it detects such a deny, so you'll know before spawning a blocked agent.

Related

• CI state machine: ci_ops.py
• CI operations: scripts/ci.py
• Debugger prompt: agent/prompts/debugger.md