Copilot PR Autopilot

Drive any GitHub pull request through repeated rounds of Copilot code review until the agent has done its job — every Copilot finding has a reply from the agent (fix-acknowledgement, decline-with-rationale, or explicit escalate-to-user hand-off). Remaining open threads, if any, are deliberate hand-offs to the human merge owner — they’re not loop failures. Repository-agnostic — works on any repo that has Copilot Code Review enabled, run from a machine with gh CLI installed and authenticated (see Prerequisites).

When to Use This Skill

• The user asks to “request Copilot review” or “run a Copilot review loop” on a PR.
• A PR is functionally complete and the user wants a final correctness pass via repeated automated review rounds.
• A previous Copilot review on the PR has left open threads that need triage, fixing, replying, and resolving.

When NOT to Use This Skill

• The PR is still under active design — wait until the structure is stable; otherwise findings churn round-over-round.
• The user wants human reviewer feedback, not Copilot’s.

Prerequisites

• gh CLI installed and authenticated against the target repository.
• PowerShell on PATH — Windows PowerShell 5.1+ (powershell.exe) or PowerShell 7+ (pwsh). Both are tested.
• Copilot Code Review is the primary use case (01-request-review.ps1 uses GraphQL requestReviewsByLogin to trigger Copilot). It is NOT a hard requirement — if 01-request-review.ps1 fails because Copilot isn’t enabled on the repo / account, the agent can still drive existing review threads (human, advanced-security, etc.) to completion by running steps 3–8 once as a single iteration; just skip the trigger + wait. There is no auto-detect for “Copilot unavailable” — the agent makes that decision after the trigger fails (the script can’t reliably tell “Copilot disabled” from “Copilot enabled but not yet triggered” from API state alone).

Permissions: who can run the full loop

The full multi-round autopilot (steps 1 → 9 → 1) needs Triage or Write permission on the target repo, because GitHub’s only public API for adding the Copilot bot as a reviewer (requestReviewsByLogin) is gated on that permission. Verified against the public REST + GraphQL surface in this PR’s commit history — there is no public-API path for bot reviewers without write permission.

In single-iteration mode the loop’s convergence boolean is Converged: true iff OpenThreadsAwaitingReply == 0 (the agent’s side is done). The maintainer-side re-trigger then drives any additional rounds.

Every script dot-sources scripts/_lib.ps1 (opens in a new tab) which runs Assert-GhReady on load: if gh is missing OR gh auth status fails, the script halts before any work with a single actionable error message naming the install command and gh auth login. The agent should surface that message to the user verbatim and stop the loop — do not retry or work around it.

Step-by-Step Workflow

The loop: steps 1 → 2 → 3 → 4 → 5 → 6 → 7 → 8 → 9, then back to step 1 if Converged: false. Repeat the 1→9 round until step 9 returns Converged: true; only then run step 10 once and call task_complete. At every 10th round, the parent runs the round-cap recap gate (opens in a new tab) before looping back — recap all prior rounds and stop if the loop has drifted out of the PR’s original scope.

Each round runs steps 1–9; step 10 is a one-time cleanup after convergence. The parent agent coordinates; every sub-agent step runs in a fresh context with a bounded budget. Cross-cutting protocol (time-boxing, extension, single-iteration fallback): orchestration.md (opens in a new tab).

1. Request review (parent) — see 01-request-review.md (opens in a new tab)
2. Wait for review (sub-agent, 20-min cap) — see 02-wait.md (opens in a new tab)
3. List + categorize open threads (sub-agent, 5 min) — see 03-list-threads.md (opens in a new tab)
4. Triage (sub-agent, 5 min per ≤5 threads) — see 04-triage.md (opens in a new tab)
5. Fix (sub-agents, parallel max 5, 5 min each) — see 05-fix.md (opens in a new tab)
6. Build + test per repo conventions (sub-agent, 10 min) — see 06-build-test.md (opens in a new tab)
7. Commit + push (parent) — see 07-commit-push.md (opens in a new tab)
8. Reply (always) + resolve (conditional) (sub-agent drafts, parent posts) — see 08-reply-resolve.md (opens in a new tab)
9. Convergence verify (sub-agent, 3 min) — see 09-convergence.md (opens in a new tab)
   • Converged: false → loop back to step 1 for another round (re-trigger, wait, list, triage, fix, push, reply, re-check). Each round addresses Copilot’s findings on the previous round’s HEAD; the loop terminates as soon as Copilot has nothing new to say AND every open thread has a reply from the agent.
   • Converged: true → exit the loop, run step 10 once, call task_complete with the proof.
   • Every 10th round (10, 20, 30…) → run the round-cap recap gate (opens in a new tab) before looping back. Recap ALL prior rounds against the PR’s original scope and pick a verdict: CONTINUE, REVERT-AND-SHIP (drop drifted commits, ship the in-scope ones), or HAND-OFF (escalate to the user). This is the circuit breaker that stops a runaway bot-review loop.
10. Cleanup outdated (parent, post-convergence, once) — see 10-cleanup.md (opens in a new tab)

Convergence is computed by scripts/02-check-review-status.ps1 (opens in a new tab) as a single Converged: true boolean. Do not call task_complete until it returns true; print the proof (HeadOid, LatestCopilotReview.commitOid, submittedAt) in the completion message.

Gotchas

The bundled scripts enforce the hard correctness invariants (trigger landing via copilot_work_started event id, Converged requiring HEAD-match + zero-awaiting + at-HEAD review, single-iteration fallback semantics, PR-state guard). Trust them — don’t re-derive. The notes below cover decisions the scripts can’t make for you:

• Reply to every open thread; resolve only when the loop owns the disposition. For fix and decline threads, reply + resolve. For escalate-to-user threads, reply with the analysis but leave the thread OPEN (08-reply-and-resolve.ps1 -NoResolve) so the human merge owner can act on it. See 08-reply-resolve.md (opens in a new tab).
• Copilot threads are loop-owned; human / advanced-security / other-bot threads default to escalate-to-user. Auto-resolving a human review thread can hide unaddressed concerns. See 04-triage.md (opens in a new tab) for the rubric.
• One focused commit per round, not one per PR. Bundling rounds destroys the audit trail of which finding drove which change and breaks git bisect. See 07-commit-push.md (opens in a new tab).
• Build/test/lint with the repo’s own commands (per its CONTRIBUTING / AGENTS / README / package.json / Makefile) before pushing a fix. Discovery procedure: 06-build-test.md (opens in a new tab).
• Push back with written rationale when a Copilot finding would over-engineer the design for a hypothetical edge case. Auto-accepting every suggestion erodes the design — see the decline path in 04-triage.md (opens in a new tab).
• Scripting traps (gh api graphql -F type-coercion, git stash push -m positional parsing, the three GraphQL traps for the reviewer mutation) are documented in references/api-quirks.md (opens in a new tab). Read before modifying any script.

Troubleshooting

References

• references/orchestration.md (opens in a new tab) — cross-cutting loop control: time-boxing & extension protocol, sub-agent delegation map, single-iteration fallback, and loop-wide notes.
• Per-step contracts (one NN-*.md per step): references/01-request-review.md (opens in a new tab) (parent), references/02-wait.md (opens in a new tab), references/03-list-threads.md (opens in a new tab), references/04-triage.md (opens in a new tab) (includes the fix-vs-decline rubric), references/05-fix.md (opens in a new tab), references/06-build-test.md (opens in a new tab), references/07-commit-push.md (opens in a new tab) (parent), references/08-reply-resolve.md (opens in a new tab), references/09-convergence.md (opens in a new tab) (includes the round-cap recap gate), references/10-cleanup.md (opens in a new tab) (parent).
• references/api-quirks.md (opens in a new tab) — verified GitHub API behavior, dead-ends, and the GraphQL traps for the reviewer mutation.
• Templates (one per reply type): templates/reply-fix.md (opens in a new tab) — accepted-fix pattern; templates/reply-decline.md (opens in a new tab) — declined-with-rationale pattern; templates/reply-drift.md (opens in a new tab) — PR-description / comment / test-plan drift acknowledgement; templates/reply-partial.md (opens in a new tab) — partial fix with deferred follow-up. Cross-cutting reply guidance and anti-patterns live in references/08-reply-resolve.md (opens in a new tab).
• scripts/_lib.ps1 (opens in a new tab) — shared helpers (Invoke-Gh, Invoke-GhGraphQL, Resolve-RepoCoords); dot-sourced by every script.
• scripts/01-request-review.ps1 (opens in a new tab) — trigger Copilot review and verify pickup via the copilot_work_started event.
• scripts/02-check-review-status.ps1 (opens in a new tab) — single-shot snapshot of the PR’s Copilot review state; emits Converged: true only when all three conditions hold.
• scripts/03-list-open-threads.ps1 (opens in a new tab) — every unresolved PR review thread from all reviewers (Copilot, humans, github-advanced-security, etc.).
• scripts/08-reply-and-resolve.ps1 (opens in a new tab) — post a reply and resolve in one call.
• scripts/10-cleanup-outdated.ps1 (opens in a new tab) — safety net for outdated Copilot threads.