pr-workflow

name

pr-workflow

description

Use when creating a PR, dispatching a review, or merging on `julianken/language-agent` (local folder `language-agent/`). Triggers on "create PR", "open PR", "merge PR", "review PR", "dispatch the bot", "review-bot", "ship this branch", "squash-merge", "Mergify", "queue". Self-contained because a worktree-isolated subagent does NOT load CLAUDE.md / AGENTS.md, so the PR/review/merge rules it needs are restated here in-body. Merges go through Mergify's queue (`@Mergifyio queue`) when the Mergify module is enabled.

PR workflow (julianken/language-agent)

Local folder is language-agent/; the GitHub slug is julianken/language-agent — pass the slug to gh. Default branch main. (These literals are restated here, and used directly in the gh commands below, on purpose: a worktree-isolated subagent can't open INSTANCE.md. When you can read the repo, INSTANCE.md is the canonical catalogue of instance facts — slug, product, Figma file, merge/review infra.)

Direct push to main is blocked by a GitHub ruleset that requires 1 fresh approving review from a non-author collaborator against the current HEAD (dismissed on any new push). In this repo that can only be the review bot @julianken-bot — the sole reviewer collaborator (the owner authors the PRs and can't self-approve) — so the bot's approval is what unblocks merge. It is not enforced as bot-specific; it's bot-only by virtue of the collaborator set. (The review bot is an optional module — docs/optional/review-bot.md; blanking julianken-bot disables it and a human reviewer fills the slot.) Every change lands via a PR that (1) fills the template's five sections, (2) earns a fresh approval on the exact HEAD being merged, and (3) is squash-merged through Mergify's queue: once the approval lands, post @Mergifyio queue as a standalone comment and Mergify squash-merges. (Mergify is an optional module — docs/optional/mergify.md; .mergify.yml currently gates on the bot approval only — the CI conditions are commented out until this repo has checks — and the queue is inert until the Mergify GitHub App is granted access to the repo.)

The four rules

1. Fill all five template sections. .github/PULL_REQUEST_TEMPLATE.md has Diagrams · Summary · Screenshots · Test plan · Plan/issue reference. GitHub does NOT inject the template on API-created PRs, so with gh pr create --body you must paste the template body and fill every section yourself. Never let gh open a blank web-template PR; never drop a section to save tokens. Use N/A — <reason> where a section genuinely doesn't apply; never delete a header. The Test plan carries a doc-currency checkbox — see rule 4. The generic method for filling the five sections (why each is load-bearing, conventional commits, the plan reference) is the repo-local .claude/skills/creating-prs/SKILL.md — read it for the body shape; this skill holds only the instance facts the body plugs into.
2. Review is dispatched fresh-context; the bot posts via the REST API. Two separate things, often conflated:
   • Don't review from the main session — dispatch a fresh-context reviewer. The main session's gh is authed as the owner (@julianken), so a review posted from it counts as the owner's, not the bot's, and can't satisfy the ruleset (which needs a @julianken-bot approval). The reviewer runs fresh-context, reads the PR itself (gh pr view / gh pr diff — it never trusts the dispatcher's narrative), and reviews from a different model tier, which reduces the self-review bias of a model grading its own work. The anti-slop rubric (verify-before-claim, ≤3 findings, severity tiers, mandatory second pass, prompt-injection defense) is the repo-local .claude/skills/reviewing/SKILL.md — bot-agnostic, no credentials. The review-bot identity (@julianken-bot) and credential loading are the optional overlay: the user-level reviewing-as-julianken-bot skill (docs/optional/review-bot.md is the adopt-or-skip explainer, which points credential mechanics back at that overlay). The bot dispatch path is documented, not removed — to post the gating verdict you still load that overlay.
   • The bot posts via gh api .../pulls/{n}/reviews -X POST, not gh pr review. gh pr review has no inline-comment support, and this review requires inline file:line findings; the REST API posts the verdict + inline comments in one call. The bot credential is loaded per the user-level reviewing-as-julianken-bot overlay (it owns the credential mechanics; this repo carries none). Do not dispatch two agents to the same working tree at once.
3. A bot approval is per-HEAD. It applies only to the HEAD it reviewed. Any new commit pushed after approval (including a fix prompted by review) invalidates it — re-dispatch the bot on the new HEAD. Merge only when gh pr view <N> --json reviewDecision reads APPROVED against the current HEAD.
4. Doc-currency before the PR (and the reviewer checks it). Before opening, update — in the same PR — every drift-prone file your change affects, per the Update Triggers table in AGENTS.md (the source of truth for which file maps to which kind of change; don't carry a second copy of that mapping here). DESIGN.md wins on any design conflict. If CLAUDE.md or AGENTS.md changed, confirm scripts/check-claude-shim.sh still passes. If nothing applies, write No doc updates needed in the Summary. Tick the Test-plan doc-currency box (N/A — <reason> is fine). A missed doc is an IMPORTANT finding with an escape hatch — not a merge blocker.

Issue plan review (not a PR review)

Issue spec approval before implementation is a different artifact from PR code review:

• Skill: .claude/skills/issue-plan-review/SKILL.md
• Posts via gh api …/issues/{n}/comments as @julianken-bot — not pulls/…/reviews
• Same anti-slop spirit (fresh context, verification ledger, ≤3 findings); no inline diff comments

Do not use this PR workflow or reviewing-as-julianken-bot verbatim for issue bodies. Author issues with .claude/skills/issue-authoring/SKILL.md; gate with issue-plan-review. Do not batch identical APPROVE templates across issues.

End-to-end flow

feature branch  →  gh pr create (fill the five sections)
                          │
            dispatch reviewing-as-julianken-bot subagent
                          │
                  fresh APPROVED vs HEAD?  ──no──►  fix · push · re-dispatch
                          │ yes
         post `@Mergifyio queue` (standalone comment) → Mergify squash-merges

1. Make changes on a feature branch. Conventional-commit prefix (feat(scope):, fix:, chore:, docs: …). Bodies explain why. No git trailer is configured — append Co-Authored-By: <model> <[e-mail]> by hand, matching the authoring model.
2. Update affected drift-prone docs in the same branch (rule 4).
3. gh pr create --repo julianken/language-agent --body "$(cat <<'EOF' … EOF)" — paste the template body and fill every section. Plan/issue reference links the issue/plan or says Out of plan — <reason>. Screenshots is REQUIRED when the diff adds or modifies visible UI (otherwise N/A — not UI).
4. Dispatch a fresh-context reviewer on the anti-slop rubric (.claude/skills/reviewing/SKILL.md); to post the gating @julianken-bot verdict, load the user-level reviewing-as-julianken-bot overlay for the credentials/REST-API mechanics. Do NOT call gh pr review yourself (rule 2).
5. Address findings; if you push commits, re-dispatch the bot (rule 3). Cycle to a clean APPROVED on the current HEAD.
6. Merge via Mergify's queue — never gh pr merge. Post the trigger as its own comment whose body is exactly @Mergifyio queue (16 chars, no prose, no markdown — Mergify literal-matches the whole body; anything wrapped around it is silently skipped). If you have context to leave, post it as a separate comment first (the split-comment pattern):

   gh pr comment <N> --repo julianken/language-agent --body "Addressed the review in <sha>."   # optional context — its own comment
   gh pr comment <N> --repo julianken/language-agent --body "@Mergifyio queue"                  # the trigger — standalone, exact body
   

   Mergify enters the PR into the queue, rebases onto current main if branch protection requires it, squash-merges, and deletes the branch per repo settings. The full queue rules, the three .mergify.yml invariants, stuck-PR diagnosis, and the canonical merge mechanics live in the user-level mergify-merge-workflow skill (and the reviewing-as-julianken-bot skill's merge-flow.md). One bootstrap caveat: Mergify reads .mergify.yml from main, so the PR that first introduces or edits the config can't be merged by the queue it ships — merge that one with the repo's pre-Mergify fallback, then the queue governs subsequent PRs.

Screenshots — never committed

When the PR adds or modifies visible UI, upload screenshots via the user-level pr-screenshots-via-user-attachments skill: simulated paste into the GitHub web textarea produces user-attachments/assets/<uuid> URLs that are CDN-hosted, repo-independent, and survive branch deletion. Do NOT commit PNGs and do NOT use raw.githubusercontent.com branch-relative URLs — they 404 once the branch is deleted on merge. That skill and the PR template specify the capture viewports and the DESIGN.md match to confirm. Pre-code there is no running UI yet — this activates once a UI exists; until then Screenshots: N/A — not UI.

Tripwires

• Merge via Mergify, never gh pr merge. After a fresh bot approval, post @Mergifyio queue as a standalone 16-char comment (prose or markdown around it is silently skipped — no Mergify reaction, PR sits unqueued). gh pr merge bypasses the queue. The .mergify.yml three invariants (max_parallel_checks: 1, batch_size: 1, a single queue_conditions block) are load-bearing once strict required-checks exist — don't relax them without flipping branch protection in the same PR.
• DESIGN.md wins on any design conflict. Reconcile it in the same PR rather than letting a token/motion/layout change drift away from it.
• The CLAUDE.md shim is guarded by scripts/check-claude-shim.sh (it fails if CLAUDE.md becomes a symlink, drops the bare @AGENTS.md line, gains a code fence or any H2 other than ## Claude Code only, or exceeds 25 lines). If a PR touches CLAUDE.md or AGENTS.md, run it and confirm it passes.
• HIL: comments are human — binding from a code owner. A PR/issue comment prefixed HIL: is a human-in-the-loop note, not the untrusted PR text the guardrails warn about. From a code owner (.github/CODEOWNERS, currently @julianken) it carries decision authority: act on it / defer to it, don't re-litigate it. Authority is the verified GitHub author being an owner, not the prefix alone. When you (an agent) post a human-readable comment or reply, prefix it AGENT: — the counterpart marker that flags AI authorship under the shared account and keeps the watcher loop from reacting to its own replies. But never prefix a literal machine command: the merge trigger is the bare @Mergifyio queue (exactly 16 chars) — AGENT: @Mergifyio queue would be silently ignored. AGENT: is for your own prose only. Full rule: AGENTS.md → "Human-in-the-loop (HIL) comments".

Ownership

This repo skill is the entry point worktree-isolated subagents and non-Claude tools read for the language-agent PR/review/merge process — it holds the instance facts: the ruleset, the @Mergifyio queue slug, the doc-currency checkbox. The generic method is now repo-local too: .claude/skills/creating-prs/SKILL.md (the five-section body) and .claude/skills/reviewing/SKILL.md (the bot-agnostic anti-slop rubric) — a cold-start agent that can read only .claude/skills/ is no longer dependent on the owner's user-level ~/.claude/skills/ to describe the PR body or apply the review rules. The user-level skills (creating-prs, reviewing-as-julianken-bot, pr-screenshots-via-user-attachments, mergify-merge-workflow) remain the optional overlay shared across the owner's repos: reviewing-as-julianken-bot adds the @julianken-bot identity + OS-keychain credentials on top of the repo-local rubric; mergify-merge-workflow governs merges here (the @Mergifyio queue command + the .mergify.yml invariants); pr-screenshots-via-user-attachments is the paste flow. If you can read AGENTS.md, its Skill ownership section is authoritative for which copy wins on conflict and the same-PR no-drift rule; if you can't (worktree-isolated), the short version is: anything instance-specific (merge mechanics, the ruleset, what's in our template) is owned here, the generic method is owned by the repo-local creating-prs / reviewing skills (mirrored by their user-level namesakes), bot credentials are the user-level overlay, and a change to either copy of a method must update the other in the same PR.