merge-ready

Merge ready

Drive the current work to the point where the only remaining step is the maintainer's own review and merge. The deliverable is a pushed branch with a clean pnpm ci:check, checkpoint commits along the way, and an open PR with a high-level description plus review instructions.

Never merge the PR. The maintainer always reviews last.

0. Figure out the starting point

This skill composes with feature work — it is not only a review pass:

• Invoked alongside a build request ("build X, make it merge-ready"): implement the feature/fix first, committing as you go, then continue below. The review phases cover all changes on the branch vs origin/main, not just the last edit.
• Invoked on existing work ("make this branch merge-ready"): start directly at step 1. The scope is git diff origin/main...HEAD plus anything uncommitted.

1. Sync with main

• git fetch origin main. If the branch is behind, merge origin/main in and resolve conflicts (favor main's version for code this branch didn't intentionally change).
• Checkpoint: commit the merge before starting review, so conflict resolution is auditable separately from review fixes.

2. Multi-axis subagent review

Spawn independent review subagents in parallel, one per axis, each given repo access and the complete branch scope:

• committed changes: git diff origin/main...HEAD
• staged changes: git diff --cached
• unstaged changes: git diff
• untracked files: git status --short, followed by reading every in-scope untracked file

Do not let an uncommitted or newly created file escape review merely because it is absent from origin/main...HEAD.

1. Unnecessary complexity — thin wrappers, needless indirection, single-use abstractions, defensive guards for impossible states, dead config. This codebase deliberately stays simple.
2. Security — authz on new endpoints (org/project scoping), SSRF, injection, secrets handling, anything user-input-shaped reaching D1/R2/external APIs.
3. Billing & metering — ways a user could trigger DataForSEO/provider spend without being metered, charged-but-failed paths, retry/loop amplification, endpoints with unexpectedly high per-call user cost. Credits are billed via Autumn; uncounted spend is a revenue leak.
4. Library & project idioms — TanStack (Router/Query/Start) used idiomatically; patterns match how the rest of the codebase already does it (shared application/provider error boundaries, db/schema conventions, existing component patterns). Flag novel patterns where an established one exists.
5. Vibe-coded cruft — leftover scaffolding, stale comments narrating the edit history, console.logs, TODO-without-owner, copy-pasted near-duplicates, files/exports nothing uses.

Each reviewer returns findings with file:line, severity (blocker / should-fix / nitpick), and a one-line rationale. Tell reviewers explicitly: this is an early-stage product — do not chase theoretical edge cases; mark anything debatable as nitpick.

3. Verify findings — never blindly accept

For each blocker and should-fix finding, spawn verification subagents (in parallel) that adversarially check the finding against the actual code and verdict APPLY / APPLY-MODIFIED / REJECT with reasoning. Drop rejected findings. Nitpicks don't need verification — they're reported, not necessarily fixed.

Preserve review learnings

After verification, route durable learnings without forcing every review to change policy:

• If an APPLY or APPLY-MODIFIED finding reveals a recurring or high-risk repository invariant that existing .greptile/ context and CI do not capture, use maintain-greptile-rules and apply its promotion bar.
• Keep one-off bugs as code fixes and regression tests. Put deterministic mechanical checks in CI or lint instead of Greptile.
• When a small tooling, documentation, or workflow frustration occurs, use papercuts to append it to .agents/PAPERCUTS.md; do not derail merge-ready work to fix it.

4. Fix, check, loop

• Apply verified blocker/should-fix fixes. Apply nitpicks only when trivial and clearly right; otherwise list them in the PR for the maintainer to judge.
• Checkpoint: commit fixes in logical groups (e.g. one commit per axis or per concern) so the fix history is reviewable on its own.
• Run pnpm ci:check (prettier, knip, tsc, oxlint). Fix failures and re-run until clean. If a fix was substantial (not formatting/lint), run a quick re-review of just that change.
• Loop until ci:check passes and no verified findings remain unaddressed.

5. Push and open/update the PR

• Push the branch. Open a PR against main if one doesn't exist; otherwise update the existing PR's description.
• PR description requirements:
  • High-level — what changed and why, written for a human skimming. No file paths, no per-file changelog.
  • How to review — a short ordered guide: what to look at first, what the risky/judgment-call areas are, what was deliberately left out of scope.
  • Review notes — unfixed nitpicks and any REJECT verdicts worth a second opinion, clearly labeled as such.
• Report back: PR link, one-paragraph summary, and anything that still needs the maintainer's judgment. Do not merge.