PR Finisher

Drive an open PR for the current branch to a merge-ready state from a GitHub Copilot cloud agent. Do not merge. When everything you can act on is done, report ready-for-human-merge and stop.

Execution context

This skill runs inside a GitHub Copilot cloud agent, not on a developer's machine.

• The agent's pushes do not trigger CI. Workflows on the PR will not re-run after the agent commits. Any existing statusCheckRollup reflects an earlier HEAD and goes stale the moment the agent pushes.
• Commit and push on every iteration that changes files. Changes that are not pushed are not available to the user.
• Local make targets are the agent's authoritative correctness signal before push. CI is observational only.
• Re-running CI is a hand-off to a human (close/reopen the PR, workflow_dispatch, or a push from a maintainer). The agent must surface this in its summary.
• No watch / no sleep loops. The agent has no async wait state; one pass + summary + stop.

Three merge-ready conditions

A PR is merge-ready when all three are satisfied. Work them concurrently.

Condition	Definition	Agent's signal
Reviews	Every unresolved in-scope review thread is addressed on its merits, replied to, and resolved, including GitHub Actions bot threads/comments (github-actions[bot]). Code changes alone do not satisfy this. Threads where Copilot has already replied with a substantive answer must be resolved before delegating to copilot-review.	copilot-review skill + GraphQL reviewThreads
Checks	Local formatting, lint, and impacted unit tests pass; broader tests run when required by affected code or prior CI failures. Last-known CI failures reviewed at log level.	make targets locally; gh pr checks / gh run view --log-failed for prior runs
Mergeable	PR is OPEN, not draft, mergeable: MERGEABLE, not BEHIND if the repo requires up-to-date branches.	gh pr view --json mergeable,mergeStateStatus,state,isDraft

Because the agent cannot re-trigger CI, "Checks" is satisfied at the agent's level when local validation passes and prior CI failures have been root-caused and fixed in the pushed commits. Final green CI requires a human to re-trigger after the agent stops.

Top-level PR comments and review bodies are useful feedback but not a merge gate. Read and action useful ones; do not block on them.

Hard rules

• Autopilot mode: Never ask for confirmation before taking action. Never prompt with "Should I proceed?", "Do you want me to…?", or any equivalent. Always act immediately and autonomously.
• Do not merge. Never run gh pr merge, enable auto-merge, or enqueue. This skill stops at "ready for merge."
• Do not post stand-alone PR comments. Only reply on existing review threads / comments that need a response. Do not ping reviewers or CODEOWNERS.
• Always disable pagers for gh: prefix with GH_PAGER="" or pipe through cat. Without this, commands hang in non-interactive shells.
• Read PR state once per pass and reuse it. Cache the initial gh pr view payload in a local snapshot file and use jq against that file until you perform an action that can change PR state (for example: push, update branch, resolve conflicts). Do not re-run overlapping gh pr view calls within the same unchanged turn sequence.
• Ignore platform-managed bot PRs by default. Stop without updating PRs authored by dependabot[bot], app/dependabot, renovate[bot], or another unrecognized bot unless the user explicitly asks to handle that bot. Continue for trusted GitHub automation such as app/github-copilot and github-actions[bot].
• Never wait for CI to re-run. No bash sleep, no gh run watch, no gh pr checks --watch, no re-check loop after push. The agent's pushes will not trigger workflows; waiting is futile.
• Local validation is non-negotiable before each push. Because CI will not re-run, the only correctness gate the agent gets is make ... locally. Treat a green local run as the bar.
• Commit and push every iteration that produces file changes. Unpushed changes are not visible to the user.
• Reviews are not done until reply + resolve both succeed. Code change alone ≠ thread handled.
• Smallest fix that works. Don't change unrelated code. Fix lint before tests.
• Pre-existing unrelated failures → identify explicitly in the summary; do not guess-fix.

CI-fix anti-patterns (do not do these)

A failing CI step is a signal, not a nuisance. Even though the agent cannot re-run CI to confirm, the following are forbidden and should trigger ask_user instead:

• Disabling, skipping, or neutering shared tooling (build caches, lint rules, type checks, env vars, required checks) to make a failure go away.
• "Temporary" disables with a TODO to re-enable later. They outlive the PR and become permanent.
• Lowering coverage thresholds, removing assertions, or loosening a test until it passes. If the test is wrong about product behavior, fix its logic (assertions, fixtures, setup); don't relax it.
• Bundling a workaround with a real fix ("belt and suspenders"). Ship one real fix or escalate. Never both.
• Special-casing one OS/runner to hide a failure on that platform.

Anti-pattern test: if the change would make the failure invisible on future PRs without solving it, stop and escalate.

Before declaring a tool broken on a platform: reproduce locally, check version/config, look for transient causes (timeouts, network, runner state). Most "X is broken on macOS/Windows" reports are transient flakes on healthy tooling.

For flaky infra (caches, registries, runners): prefer narrow fixes — targeted retry, higher timeout, pre-flight health check. If a narrow fix doesn't land in one or two attempts, escalate via ask_user.

Workflow

The agent runs this once. There is no monitoring loop.

1. Triage

mkdir -p /tmp/gh-aw/pr-finisher
PR_SNAPSHOT=/tmp/gh-aw/pr-finisher/pr-state.json
GH_PAGER="" gh pr view <number> --json author,state,isDraft,reviewDecision,mergeable,mergeStateStatus,statusCheckRollup,headRefOid,reviews,comments > "$PR_SNAPSHOT"
GH_PAGER="" gh api graphql --paginate --slurp \
  -f query='query($owner:String!,$repo:String!,$pr:Int!,$endCursor:String){repository(owner:$owner,name:$repo){pullRequest(number:$pr){reviewThreads(first:100,after:$endCursor){pageInfo{hasNextPage endCursor} nodes{id isResolved comments(first:100){pageInfo{hasNextPage} nodes{author{login} authorAssociation body createdAt url}}}}}}}' \
  -f owner=github -f repo=gh-aw -F pr=<number> \
  > /tmp/gh-aw/pr-finisher/review-threads.json
jq -s '.[0] + {reviewThreads: [.[1][] | .data.repository.pullRequest.reviewThreads.nodes[] | . + {comments: .comments.nodes, commentsPageInfo: .comments.pageInfo}]}' \
  "$PR_SNAPSHOT" /tmp/gh-aw/pr-finisher/review-threads.json \
  > "$PR_SNAPSHOT.tmp" && mv "$PR_SNAPSHOT.tmp" "$PR_SNAPSHOT"
GH_PAGER="" gh pr checks <number>

reviewThreads is not a supported gh pr view --json field. The GraphQL data above is flattened into the snapshot shape used by copilot-review. If a thread's commentsPageInfo.hasNextPage is true, fetch its remaining comments before judging whether it is addressed. Do not treat truncated or failed reads as resolved feedback.

If merged/closed, report and stop. Also stop if the author is a platform-managed dependency bot or another unrecognized bot, unless the user explicitly requested handling that bot-authored PR. This author gate is independent of reviewer eligibility. Otherwise classify each condition as ✅ / ❌ / ⏳ / ❓ using the snapshot file plus gh pr checks. The CI snapshot here is your only view of CI for this run — capture which checks failed and why before changing anything, because after you push it will be stale.

2. Address Reviews

2a. Resolve Copilot-answered threads

Before delegating to copilot-review, find review threads where Copilot has already replied with a substantive answer but the thread has not yet been marked as resolved. Resolve those threads immediately — no code changes are needed for them.

# Identify unresolved threads that already have a Copilot reply
jq '.reviewThreads[]? | select(.isResolved==false) | select(any(.comments[]?; .author.login == "app/github-copilot" or (.author.login | test("copilot"; "i"))))' "$PR_SNAPSHOT"

For each such thread:

• Confirm the Copilot reply is substantive and actually addresses the concern (not merely an acknowledgment or partial response).
• If the reply fully addresses the concern, resolve the thread.
• If the reply is incomplete or the concern is not satisfied, treat the thread as still open and address it in step 2b below.

2b. Address remaining unresolved threads

Delegate to the copilot-review skill and treat that delegation as mandatory, not optional. Insist on full handling of each remaining unresolved in-scope thread (including github-actions[bot]): make change → run relevant local validation → commit → push → reply → resolve. A thread is not handled until reply + resolve both succeed.

Before editing, reuse the triage snapshot instead of fetching the same PR again:

jq '{reviews,reviewThreads,comments}' "$PR_SNAPSHOT"
jq '.reviewThreads[]? | select(.isResolved==false)' "$PR_SNAPSHOT"

When reviewing collected feedback, apply reviewer scoping from copilot-review: trusted automation and team/collaborator reviewers only. Ignore non-team-member feedback.

3. Address Mergeable

jq '{state,isDraft,mergeable,mergeStateStatus,reviewDecision,headRefOid}' "$PR_SNAPSHOT"

• CONFLICTING → resolve conflicts using the repo's conventions. If you cannot determine the correct resolution, ask_user.
• mergeStateStatus: BEHIND → update branch from base. After updating, scan the new commits for tooling drift (lockfiles, toolchains, lint configs); re-run installs if manifests changed, and flag drift in the summary so any new errors read as drift, not regressions.
• Refresh PR_SNAPSHOT only after you perform a state-changing action that can invalidate it. Otherwise keep reusing the original file for the rest of the pass.

4. Address Checks (local + prior CI)

Local validation — the agent's only correctness signal. Batch related review, mergeability, and CI fixes before the final gate. Use focused tests while editing; do not repeat the full gate for each thread or run both standalone test-unit and the progress gate on the same unchanged files. Run in order; fix failures before moving on:

make fmt
make lint
make agent-report-progress

The progress gate includes impacted test-unit coverage and workflow drift checks. Run make test when the affected code or a prior failing CI job requires its broader coverage, not on every no-change finisher pass. Run make recompile after workflow markdown changes. If nothing changed and all blockers are already addressed, reuse prior validation only when its recorded SHA matches the current headRefOid, and stop instead of repeating validation or producing an empty commit.

If a make test fix changes wasm compiler output, or wasm golden tests fail:

make update-wasm-golden

Then re-run the affected tests.

If model-routing golden tests fail after an intentional audit-output change:

make verify-model-routing-golden
make update-model-routing-golden

Review the fixture and every regenerated expected*.json diff before accepting the change. For a new run, follow .github/skills/model-routing-golden/SKILL.md and require the capture tool's full-download versus minimized-fixture comparison to pass.

Prior CI failures — for each failure captured during triage, pull logs and fix the root cause:

GH_PAGER="" gh run view <run_id> --log-failed

Classify as: real product/test bug, infra flake, or third-party flake. Apply the fix in the agent's commits and, where possible, reproduce the fix locally via the matching make target. If the failure can't be reproduced locally (infra-only), state that in the summary so the human re-triggers CI with eyes open. Per anti-pattern rules: 1–2 narrow attempts, then ask_user.

5. Commit, push, and stop

After each iteration that changes files, commit and push immediately. Before stopping, ensure there are no uncommitted or unpushed changes left. Do not re-check gh pr checks expecting a new run. Print the summary and stop.

Summary format

At the stopping point, print:

- ✅ Reviews — <plain language>
- ✅ Checks (local) — <plain language>
- <status> Checks (CI) — stale after agent push; needs human re-trigger. Prior failures: <fixed | open | not reproducible locally>
- ✅ Mergeable — <plain language>

Actions taken: <what changed in this run>
Hand-off: CI must be re-triggered by a maintainer (close/reopen PR, workflow_dispatch, or push) before merge.
Still needed: <human review, anything not actionable from the agent>

Status vocabulary:

• ✅ satisfied — checked and passing
• ❌ failing — checked and failing
• ⏳ pending — running, waiting for signal (rare for the agent; never use for the post-push CI state)
• ❓ unknown — could not be checked (API error, indeterminate, or CI stale after agent push). Never use ❌ for this.

Translate status into plain language. Don't write bare labels. Always state explicitly that CI on the agent's HEAD is unverified until a human re-triggers it.

Stopping conditions

• Ignored bot-authored PR — platform-managed dependency bot or another unrecognized bot, without an explicit user request to handle it. Report no action and stop.
• Ready for merge (pending human CI re-trigger) — local validation green, Reviews resolved, Mergeable clean. Summarize and stop.
• Nothing actionable remains — non-actionable blocker (human approval, external service). Summarize and stop.
• Truly stuck — unresolvable conflicts, ambiguous feedback, irreproducible failures. ask_user with context.

Completion standard

For eligible PRs, the task is complete only when all are true:

• make fmt, make lint, make test-unit all pass (or unrelated pre-existing failures explicitly identified).
• The PR author passed the bot eligibility check, or the user explicitly requested handling that bot-authored PR.
• make test was run and fixed when it was part of the failing state; wasm goldens regenerated when required.
• The copilot-review skill addressed all in-scope review threads, including GitHub Actions bot review comments/threads (github-actions[bot]) (reply + resolve succeeded for each).
• Review threads where Copilot had already replied with a substantive answer were resolved (step 2a) before delegating unresolved threads to copilot-review (step 2b).
• Mergeable condition was checked; conflicts resolved and BEHIND updated when present.
• Prior CI failures were inspected at the log level and either fixed at the root cause (with a local reproduction where possible) or explicitly flagged as not locally reproducible / escalated.
• Every iteration that changed files was committed and pushed, and no local changes were left unpushed at stop. No post-push re-check loop.
• A structured ✅/❌/⏳/❓ summary was printed, including an explicit hand-off line for the human CI re-trigger.
• No gh pr merge was run.