CI Feedback Loop

CI Feedback Loop

Pull failing GitHub Actions context (status → failed jobs → step logs → artifacts) into the local working session so the agent can fix issues without round-tripping through the GitHub UI.

A small Go CLI (in this skill folder) wraps the gh CLI for orchestration, parses logs/artifacts, and writes everything under debug/ci/<run_id>/ — the debug/** path is already covered by .gitignore.

When to Use

• A PR is open and one or more workflow runs failed
• User asks to "debug CI", "check the pipeline", "fix the failing tests in CI", "fetch the CI logs"
• Iterating on a feature branch where commits keep breaking specific jobs
• Need cluster-state artifacts (kubectl dumps, helm logs, pod descriptions) attached to a failed run

Prerequisites

• gh CLI authenticated (gh auth status)
• go ≥ 1.23 on PATH
• Working directory inside the target git repo on the feature branch

Setup

The CLI lives in .github/skills/ci-feedback-loop as its own Go module. Invoke it with the helper alias (run once per shell):

ci_fb() { (cd "$(git rev-parse --show-toplevel)/.github/skills/ci-feedback-loop" && go run . "$@"); }

Then use ci_fb <subcommand> from anywhere in the repo. All examples below assume this alias is defined.

Available subcommands: locate, summarize, logs, artifacts. Pass --help to any of them for flags.

Procedure

1. Locate the target run

ci_fb locate                       # list 10 most recent runs (non-blocking)
ci_fb locate --latest              # print only the latest RUN_ID (non-blocking)
ci_fb locate <RUN_ID>              # print this RUN_ID (non-blocking)
ci_fb locate --wait                # block until every in-progress run finishes (hard timeout)
ci_fb locate --latest --wait       # wait, then print the latest RUN_ID
ci_fb locate <RUN_ID> --wait       # wait specifically on this run

Default is non-blocking — listing returns immediately and flags any pending runs in the footer (⏳ N run(s) still in progress). The agent should re-invoke later rather than blocking the conversation.

When --wait is set, the command polls every 30 s without streaming logs, and aborts with a clear message after --wait-timeout seconds (default 1800 = 30 min). This prevents the calling agent from hanging.

The command emits a single RUN_ID=<id> line on stdout when resolved (with --latest or an explicit id). Capture it for later steps.

2. Summarize failures

ci_fb summarize <RUN_ID>

Prints workflow metadata + only the failed/cancelled jobs with their failing step numbers. Compact enough to keep in context — use it to decide which logs/artifacts to pull next.

3. Fetch focused logs

ci_fb logs <RUN_ID> [--job <substring>] [--tail 200] [--context 80]

Behavior:

• Downloads the run's full log archive once into debug/ci/<RUN_ID>/_run.zip
• Extracts to debug/ci/<RUN_ID>/logs/
• For each failed job (optionally filtered by --job), prints:
  • the last --tail lines of its log
  • the first error context (regex error|fail|fatal|panic, ±lines, capped at --context lines)

Read the printed excerpts directly. Open debug/ci/<RUN_ID>/logs/<job>.txt only if excerpts are insufficient.

4. Download artifacts (cluster state, etc.)

ci_fb artifacts <RUN_ID>                       # list only
ci_fb artifacts <RUN_ID> --name cluster-dump   # download matching

Stored under debug/ci/<RUN_ID>/artifacts/<name>/. Always list first, pick by name (e.g. cluster-dump, pod-logs-kind, helm-status), then download.

After download, inspect the most useful files:

• *.yaml / *.json — read directly
• *.log / *.txt — tail or grep for the error keywords from step 2
• *.tar.gz / *.zip — extract in place before reading

5. Diagnose and fix

With logs + artifacts in debug/ci/<RUN_ID>/:

1. Map the error message to the source (workflow step, action, script, app code)
2. Read that source file before editing
3. Apply the minimal fix
4. Verify locally if possible (lint, unit test, dry-run)

6. Push and re-loop

Commit the fix under the repository's staging and message rules — AGENTS.md, "Agent collaboration rules". This skill does not own the commit convention; do not reach for git add -u here.

Then re-enter the loop:

# Either: poll later non-blockingly
ci_fb locate
# Or: explicitly wait (safe — hard timeout)
ci_fb locate --latest --wait

Repeat steps 2–6 until the pipeline is green.

Cleanup

debug/** is gitignored. Delete anytime: rm -rf debug/ci.

Anti-patterns

• Don't dump full logs into the conversation. Use summarize + logs excerpts.
• Don't download every artifact blindly — list first, then pick by --name.
• Don't push-and-pray. Always re-watch the run after fixing.
• Don't edit the workflow YAML to "skip" a check just to make CI pass — fix the underlying issue.

References

• main.go — the full CLI (cobra subcommands)
• go.mod
• AGENTS.md → "Agent collaboration rules" — staging and commit-message rules.