Review Loop

PR Review Loop

Autonomously drive a PR (and its sibling backport PRs) through this loop:

pause CI → self-review the diff → request Copilot review → fix every finding
        → re-review (loop until clean)
        → relaunch tests → on failure: fix + re-enter loop → on success: append " [ready]"

The goal is a PR that is green and has no outstanding Copilot findings, marked ready for human review — without ever merging it.

When to use

• The user wants a hands-off "get this PR reviewed and green" cycle.
• A change spans a main PR + one or more backport-*-to-stable/* PRs that must stay in sync.

Prerequisites

• gh authenticated (gh auth status) with permissions for everything the loop touches: pull-requests: write (request reviews, edit the title), issues: write (add/remove the skip_all label), and actions: write (cancel/rerun workflow runs). A classic token with the repo + workflow scopes covers all three.
• go ≥ 1.23 on PATH, for the CLI below.
• Working dir inside the target repo; the feature branch is pushed.

The CLI

The choreography — resolving the PR set, the skip_all lifecycle, reading Copilot's findings, tagging — lives in the ci-feedback-loop CLI rather than in this file, so its idempotency is tested rather than asserted. Define the alias once per shell:

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

Every ci_fb review subcommand takes the same optional PR references (123, #123, or a PR URL) and, given none, falls back to the current branch's PR. Each one expands the set with the open backports of every seed, so sibling PRs stay in lockstep without you tracking them by hand.

ci_fb review prs        # resolve and print the PR set this loop will drive

Contracts this skill obeys

Defined in AGENTS.md — read there, not here:

• "PR review rules" — never merge; stop at [ready]; triage every Copilot finding; propagate a valid finding to every sibling backport.
• "CI cost and skip labels" — what the labels mean and when to apply them.
• "Agent collaboration rules" — staging and commit-message rules for every fix this loop pushes.

skip_all is this loop's pause button and nothing else. ci_fb review pause adds it, ci_fb review resume removes it, and ci_fb review ready refuses to tag a PR that still carries it — a paused PR has no CI signal, so it cannot be ready by definition.

Procedure

1. Pause CI

ci_fb review pause

Adds skip_all, strips any [ready] tag (a paused PR is not ready), and cancels the PR's in-progress runs to free the runners.

2. Self-review the diff before asking for a machine review

Read and follow the code-quality-review skill (.github/skills/code-quality-review/SKILL.md) on the current branch's diff.

Do this before step 3, not after. Copilot reviews the lines that exist when it is asked; if a structural fix is still coming, its findings land on code that is about to disappear, and you spend a round-trip answering threads on dead lines.

• Apply the findings you agree with.
• On a sibling backport set, run it once on the originating PR and carry the same fixes across — the diffs are meant to stay identical.
• An empty report is a normal outcome. Do not manufacture a refactor to justify the step.

Skip this step only for a diff that cannot have structure: a lockfile bump, a renovate update, a one-line constant change.

3. Request a Copilot review

Some repos auto-request Copilot on every push (branch ruleset "Review new pushes") — there, a push is enough. To request explicitly, pass the bot login (a display name like "Copilot" will not resolve):

for n in $(ci_fb review prs); do
  gh pr edit "$n" --add-reviewer copilot-pull-request-reviewer
done

If that fails because the reviewer is not requestable in this repo, rely on auto-review on push or request it from the GitHub UI. Do not spin here.

4. Wait for the review to land

ci_fb review state

Prints one PR #<n> <STATE> line per PR. COMMENTED / CHANGES_REQUESTED / APPROVED mean the review landed — go to step 5. NONE or PENDING means it has not: stop and re-check later, do not proceed with nothing to triage. This is a single non-blocking check, never a busy-wait. If it stays NONE well past the usual few minutes, the request never landed — see step 3.

5. Read and triage every finding

ci_fb review findings

Prints Copilot's review summary and every inline finding, per PR, with the comment id you need to reply. Human review threads are deliberately excluded — they are for the humans; do not auto-reply to them.

For each Copilot finding:

• Fix it in code with the minimal correct change, or
• Reply on that specific thread explaining why it is intentionally not addressed: gh api /repos/<owner>/<repo>/pulls/<n>/comments/<comment_id>/replies -f body=...
• Never silently ignore a finding.
• Leave thread resolution to the human author.

6. Push fixes, then re-review

Commit under the staging and message rules in AGENTS.md. Then re-request the review (step 3) and return to step 4. Loop steps 3–6 until the newest Copilot review adds no new actionable findings.

7. Relaunch the tests

ci_fb review resume --rerun

Removes skip_all and re-runs each PR's latest completed run. The rerun is needed because removing a label is not a workflow trigger; pushing a fix re-triggers them too, in which case plain ci_fb review resume is enough.

The rerun is deliberately full, not --failed: the heavy jobs were skipped while paused, not failed, so there is nothing for --failed to select.

8. Watch the tests

Use the ci-feedback-loop procedure (.github/skills/ci-feedback-loop/SKILL.md) — ci_fb locate / summarize / logs / artifacts.

9. On failure → fix → re-enter the loop

• Map the failure to its source, read it, apply the minimal real fix.
• To re-run only the failed jobs of a completed run: gh run rerun --failed <run-id>.
• Never edit a workflow just to skip a check to go green.
• Then go back to step 1: pause is idempotent and strips the [ready] tag again, since the regression made the PR not-ready.

10. On success → append the [ready] tag

ci_fb review ready

Appends [ready] once, at the end of each title. It refuses while skip_all is still on the PR or while any check is failing or pending, so a green, un-paused PR is a precondition rather than a thing you remember to verify.

This is where the loop stops. Tagging is the signal; merging is the human's.

Anti-patterns

• Don't merge. This skill stops at the [ready] tag.
• Don't request the Copilot review before step 2 — findings on code you are about to restructure are wasted round-trips.
• Don't silently drop a Copilot finding — fix it or reply with a rationale.
• Don't skip/disable a failing check to force green — fix the root cause.
• Don't reach for --force on ci_fb review ready. It exists for a human who knows why a check is red; an agent using it is lying about the state.
• Don't block the session on long polls — re-check later.

References

• ci-feedback-loop — the CLI: CI status/logs/artifacts and the review subcommands.
• code-quality-review — the strict self-review run at step 2.
• AGENTS.md → "PR review rules", "CI cost and skip labels", "Agent collaboration rules".