Automerge

Auto-merge (review-gated)

The fresh-context review is the thing that flips auto-merge on — nothing reaches main on green CI alone for the dangerous categories. This is the Gate (run-the-gate) applied to a PR diff at merge time instead of a spec before code, plus the verified-sweep refute guard, wired to gh pr merge --auto.

GitHub does the actual merge: enabling auto-merge holds the PR until the required readiness check is green and the branch is up-to-date, then merges and deletes the branch. "Enabled" ≠ "merged now."

Auto-merge NEVER updates a BEHIND branch. allow_update_branch only permits the update — nothing calls it (verified 2026-07-03: #859 sat BEHIND indefinitely; the old claim here that BEHIND PRs "self-update" was false). Every squash-merge re-BEHINDs whatever is queued behind it, so a merge train stalls after the first car. After enabling auto-merge, check mergeStateStatus and update BEHIND branches yourself (Step 3a); the pr-babysitter skill re-does this on every pass for the whole board.

The one property this lives or dies on

Classification fails TOWARD review. A false positive costs one wasted Opus review. A false negative auto-merges an unreviewed money/rules/routing/schema change on green CI — the disaster this whole system exists to prevent. When a changed path is unfamiliar or you can't confidently say it's harmless, treat it as GATE. Never drift toward "just merge it."

Pipeline

/automerge <N>
  └─ preconditions: PR open, not draft, auto-merge not already enabled
  └─ classify  (gh pr diff <N> --name-only  vs the denylist below)
       ├─ EXEMPT → enable auto-merge → done
       └─ GATE   → fresh Opus diff review (zero history)
                    ├─ P1s > 0 → post P1s as PR comment, STOP (you fix; re-run = NEW round)
                    └─ clean   → fresh REFUTER (tries to find a missed P1)
                                  ├─ refuted/unverified → post reason, STOP
                                  └─ cleared            → enable auto-merge → done

Step 1 — preconditions

gh pr view <N> --json isDraft,state,autoMergeRequest,mergeStateStatus

• state != OPEN or isDraft == true → stop, report why.
• autoMergeRequest != null → already enabled, nothing to do.

Step 2 — classify the diff

gh pr diff <N> --name-only

GATE if any changed path matches the denylist (verified to exist 2026-06-06):

Plus the catch-all: any changed path you don't recognize as plainly safe → GATE. Docs (*.md), docs/**, golden/test-only changes, design-token sweeps under lib/core/theme/tokens/ + widget styling, fastlane/**, asset files → EXEMPT.

Step 3a — EXEMPT → enable auto-merge

Resolve the repo default method at runtime (do not pin):

m=$(gh repo view --json viewerDefaultMergeMethod -q .viewerDefaultMergeMethod)
case "$m" in SQUASH) f=--squash;; REBASE) f=--rebase;; *) f=--merge;; esac
gh pr merge <N> --auto $f

Then clear the BEHIND stall (auto-merge will never do this itself):

s=$(gh pr view <N> --json mergeStateStatus -q .mergeStateStatus)
[ "$s" = "BEHIND" ] && gh pr update-branch <N>

Report: classified EXEMPT (which check let it through), auto-merge enabled, branch updated if it was BEHIND, will merge on green readiness.

Step 3b — GATE → review → refute → enable

Review. Spawn a fresh reviewer — Agent, subagent_type: general-purpose, model: opus, zero history. Prompt = the full contents of diff-reviewer-prompt.md (this skill's dir) + the PR number. The agent runs gh pr diff <N> itself and verifies against live code. Do NOT paste this session's reasoning in. Beyond the code rubric, the reviewer also checks spec conformance (if the PR body has a Spec: link, the diff must not drift from the Gate-approved spec — un-gated scope creep or an unbuilt acceptance box is a [P1]) and RED evidence (a bug-fix must ship a regression test + pasted failing-before-fix output).

• Verdict has P1s → post them as a PR comment (gh pr comment <N> --body ...), STOP. Auto-merge stays OFF. You fix + push; re-running /automerge <N> is a NEW round with a NEW fresh agent (never SendMessage to continue the old one — continuation re-imports the context the freshness exists to escape).
• Verdict clean (0 P1s) → go to Refute. A clean review alone does NOT enable auto-merge.

Refute (the verified-sweep guard — a PASS is a done-claim, and fabricated/sycophantic done-claims are this repo's signature failure; delete_branch_on_merge makes the merge effectively irreversible). Spawn a SECOND fresh Agent (opus, zero history). Give it the diff + the reviewer's clean verdict; its job is to REFUTE — find one P1 the reviewer missed, default to refuted: true on any uncertainty. Model the prompt on .claude/workflows/verified-sweep.js's refute prompt.

• refuted: true (or refuter returns nothing) → post the reason as a PR comment, STOP. Treat unverified as refuted.
• refuted: false with concrete evidence it actively confirmed clean → enable auto-merge (Step 3a commands, including the BEHIND update).

The merge tail — hand it to pr-babysitter

Enabling auto-merge is not the finish line: branches go BEHIND after every squash-merge (and never self-update), CI goes red, reviewers leave P1s, merged worktrees linger. The pr-babysitter skill (this repo) owns that tail — per pass it updates BEHIND PRs, runs THIS pipeline on any open PR still missing auto-merge, and surfaces P1/red-CI/conflict PRs for the human. Under /loop it runs until the board is clear. Use it instead of hand-polling gh pr view.

Non-negotiables

• Classification fails toward GATE. Unknown path = GATE.
• One reviewer / one refuter = one fresh Agent call each, zero history. Round 2 must not know round 1 existed.
• A clean review never enables auto-merge on its own — the refuter must also clear.
• A green test suite is NOT the review (worked example #185: rules suite green while the client write violated the rule).
• Never bypass to direct merge. The skill only ever enables native auto-merge; GitHub + readiness do the merge.
• Repo-default merge method, resolved at runtime — don't hardcode --squash.