| # CI Repair Loop |
|
|
| ## Objective |
|
|
| Turn a failing CI signal into a small, verified fix or a precise escalation note. |
|
|
| ## Trigger |
|
|
| - Schedule: poll failing checks on active branches. |
| - Event: CI check fails on a PR or protected branch. |
| - Manual bootstrap/debug command: "repair the failing CI check for this PR." |
|
|
| ## Intake |
|
|
| - Failed check name, logs, artifacts, commit SHA, changed files, and recent successful run. |
| - Local test commands from project docs. |
| - Known flaky-test list or CI status page. |
|
|
| ## Agents |
|
|
| - Investigator: extracts the failing command, error, and likely affected files. |
| - Implementer: makes the smallest patch consistent with the failure. |
| - Verifier: reruns the local or CI-equivalent command. |
| - Reporter: summarizes evidence and residual risk. |
|
|
| ## Workspace And Permissions |
|
|
| - Use a clean branch or worktree based on the failing commit. |
| - Allow read access to CI logs and artifacts. |
| - Allow local test, lint, typecheck, and build commands. |
| - Disallow unrelated refactors, dependency upgrades, broad formatting, or changes to protected CI config unless explicitly requested. |
|
|
| ## Durable State |
|
|
| - Failing check URL, command reproduced, exact error, attempted fixes, passing evidence, and unresolved blockers. |
|
|
| ## Loop Steps |
|
|
| 1. Fetch the failed check and logs. |
| 1. Identify the deterministic failing command or closest local equivalent. |
| 1. Delegate log extraction, patching, verification, and reporting to separate roles when useful. |
| 1. Reproduce locally when possible. |
| 1. Patch the smallest cause. |
| 1. Rerun the failing command first, then any adjacent cheap checks. |
| 1. If green, summarize and push. |
| 1. If not reproducible or blocked, write an escalation with logs and hypotheses. |
|
|
| ## Verification Gates |
|
|
| - The original failing command passes locally or in rerun CI. |
| - The patch is limited to the failure cause. |
| - New tests are added when the fix changes behavior. |
| - Output includes exact commands and relevant log excerpts. |
|
|
| ## Budget And Exit |
|
|
| - Max retries: 3 patch attempts. |
| - Max runtime: 45-90 minutes. |
| - Stop when the target check passes, the failure is proven flaky, the environment is unavailable, or the fix exceeds scope. |
|
|
| ## Escalation |
|
|
| Escalate for flaky infrastructure, missing credentials, third-party outages, nondeterministic failures without reproduction, or changes requiring owner approval. |
|
|
| ## Loop Instruction |
|
|
| ```text |
| Repair the failing CI check for <PR or branch>. |
| Start by identifying the exact failing command and reproducing it locally if possible. |
| Make the smallest scoped patch, rerun the failing command, and report evidence. |
| Do not change unrelated files or CI configuration unless the logs prove it is necessary. |
| ``` |
|
|
| Example automation: trigger on failed required checks, or poll active PRs for red CI every 30-60 minutes. |
|
|
| ## Failure Modes |
|
|
| - Treating a CI artifact as stale without checking commit SHA. |
| - Running only broad checks and missing the failing command. |
| - Fixing by weakening tests or deleting assertions. |
| - Chasing flaky failures without a retry budget. |
|
|
| ## References |
|
|
| - [Stop Babysitting Your Coding Agent. Give It Backpressure.](https://generativeprogrammer.com/p/stop-babysitting-your-coding-agent) - Turns tests, linters, builds, and traces into feedback loops for coding agents. |
| - [GitHub Agentic Workflows](https://github.github.com/gh-aw/) - Runs coding agents from GitHub events or schedules with repository-level guardrails. |
|
|
