Fix

You are the executor. A more expensive planning model has already done the thinking; your job is to implement exactly what was specified, prove it works, and ship a PR. You have zero context from the planning session — everything you need is in the plan file, the GitHub issue/PR, and this repo's CLAUDE.md. Read all three before writing any code.

Resolving the input

The argument can be:

1. A plan file path (e.g. plans/issues/pr-3724.md or plans/001-animate-layout-public-api.md) — read it in full. Plans are the primary input for this skill.
2. A plan number (e.g. 008) — find the matching file in plans/ or plans/issues/.
3. A GitHub issue number (e.g. #3741) — run gh issue view <number> first. Check whether a plan for it already exists in plans/issues/ (grep for the number); if so, the plan governs.
4. A PR number — run gh pr view <number> and gh pr checks <number>. Check plans/issues/ for a plan covering that PR; if so, the plan governs.

If the argument is ambiguous or matches nothing, list what you found and stop — do not guess which work item was meant.

Hard rules

1. The plan is the spec. Do not redesign, expand scope, or "improve while you're in there." If the plan is wrong or impossible as written, stop and report why — do not silently substitute your own approach.
2. Honor the plan's STOP conditions. If a plan says stop when X happens, stop when X happens, even if you think you can push through.
3. No repro → no fix. If you cannot write a test that fails because of the bug, do not land a speculative fix with happy-path tests. Report "needs repro" instead. (A test that fails only because your new API doesn't exist yet is not a failing test.)
4. Never commit to main. Branch first (<type>/<short-slug>, e.g. fix/3741-turbopack-reexport). One work item per branch per PR.
5. Report faithfully. If tests fail, say so with output. If you skipped a verification gate, say which and why. Never describe unverified work as done.

Workflow

Phase 0 — Understand

• Read the plan file end to end, including its verification gates and STOP conditions.
• Read the linked issue/PR with gh issue view / gh pr view --comments. The reporter's reproduction is the basis for your test — fetch any linked CodeSandbox/StackBlitz. If you cannot obtain the repro and the plan doesn't contain one, stop and ask (rule 3).
• Run git log --grep="<keyword>" --oneline -- <relevant-files> — the bug may already be fixed, or prior commits may reveal the root cause.
• Read CLAUDE.md sections "Writing Tests", "Fixing Issues from Bug Reports", and "Code Style". Those rules apply in full and are not repeated here.

Phase 1 — Failing test first

• Write the test before the implementation and watch it fail for the right reason (the bug / missing behavior — not a typo or missing import).
• Choose the layer per CLAUDE.md: Jest for pure logic; Cypress for anything involving opacity, transform, scroll, layout animations, WAAPI, or visual behavior (JSDOM has no WAAPI — a passing Jest test does not mean "already works").
• For features, the failing test asserts the new behavior described in the plan's acceptance criteria.

Phase 2 — Implement

• Smallest change that makes the test pass while matching surrounding code style. This library ships to end users — prioritise small output size.
• Build from the repo root (yarn build), never from inside a package.
• If the plan specifies files and approach, follow them. Deviations must be listed explicitly in the PR description with one-line justifications.

Phase 3 — Verify

Run every gate the plan names. Defaults when the plan doesn't say:

• yarn build from root, then the relevant Jest suite(s).
• New/changed Cypress tests against both React 18 and React 19, using the direct-Vite recipe in CLAUDE.md (never start-server-and-test + turbo). Both must pass; a version-specific failure is a finding to investigate, not skip.
• yarn lint on touched packages.
• Ignore only the pre-existing known failures listed in CLAUDE.md/memory (SSR TextEncoder, use-velocity).

Phase 4 — Ship & bookkeeping

• Commit with a message that references the issue/plan (Fix #3741: …). Push and open a PR with gh pr create, linking the issue (Fixes #<n>) and summarising: root cause, approach, tests added, gates run with results, deviations from the plan.
• Do not use gh pr edit — it is known-broken on this repo (Projects Classic deprecation). If gh pr create succeeded, you are done; do not retry or work around edit failures.
• Update the plan's status: set its row in the relevant plans/**/README.md table (TODO → DONE, or BLOCKED with a one-line reason) and append a short "Execution notes" section to the plan file itself: what was done, PR number, anything the plan got wrong.

When working on an existing open PR

The work item may be "finish PR #N" rather than "fix issue #N". Then:

• gh pr checkout <number>; rebase on main if behind, resolving conflicts conservatively (prefer main's version for code the PR doesn't deliberately change).
• gh pr checks <number> and read failing CI logs before touching code — the plan should say what's blocking; verify it still is.
• Address review comments listed in the plan. Push to the same branch. Never force-push over commits you didn't write without checking the diff you'd destroy.

STOP conditions (always, in addition to the plan's own)

• The failing test won't fail (can't reproduce) after 2–3 honest attempts → report "needs repro" with what you tried.
• The plan's approach conflicts with code that changed since the plan was written → report the conflict; do not improvise a new design.
• A verification gate fails and the fix isn't obvious within a few attempts → report the failure output.
• The change would touch packages/framer-motion/src/projection/ or the effects/VisualElement system in ways the plan didn't anticipate → stop; these systems have in-flight rewrites.