typebench

Multimodal multi-axis code review for the typebench repo before merge — your own code, another agent's, or a teammate's PR. Runs two independent reviewers (a primary model + an independent second model) at high reasoning effort and synthesizes their findings. Adds benchmark-aware checks for measurement fidelity, record honesty, and failure completeness on top of the standard correctness/readability/architecture/security/performance review. Use whenever you are about to merge, or when asked "is this ready?", "review this", "check this change", "look this over". Reviewing AI-generated code is a stronger trigger, not a weaker one — false confidence is the dominant failure mode.

Explore the typebench codebase, surface architectural friction, and propose module-deepening refactors as actionable plan documents under docs/plans/. Use when asked to review architecture, "make this cleaner", "reduce coupling", "this module is doing too much", "the modules are just thrown there", "the tests are a junkyard", or to evaluate whether a module pulls its weight. Tuned to typebench's package layering (contracts ← engine ← {adapters, corpus} ← suite ← cli), its pydantic-free measured path, the Adapter Protocol grain, and the no-planning-jargon docstring norm. Produces a durable refactor plan, not edits — execution is a separate, approved step.

Python implementation and review skill for the typebench repo — a neutral, reproducible Python type-checker performance benchmark. Use every time you write, modify, refactor, or review Python code in this codebase — contracts, type safety, testability, security, measurement fidelity, and module boundaries all matter here. Apply on every Python edit, not only when the user explicitly asks for "clean code" or "best practices".

Git commit, branch, and PR conventions for the typebench repo — Conventional Commits with required scopes, atomic commits, meaningful bodies, a commit-as-save-point working pattern, and a hard no-AI-attribution rule. Use every time you create a git commit, write a commit message, stage changes, open a PR, or resolve a merge conflict in this repo. Apply these rules on every commit — a sloppy git history compounds faster than sloppy code.

Test-driven development skill for the typebench repo. Use whenever you are about to write or change product behavior — a new pipeline stage, a bug fix, a refactor of anything observable. Write the failing test first; use the Prove-It pattern for bugs. Apply on every behavioral change, not only when the user asks for tests. Covers typebench's conventions — `@pytest.mark.skipif` gating, the Typer `CliRunner`, `monkeypatch.setattr` for boundaries, the `fake_checker`/`StubAdapter` taxonomy harness, JSON round-trips with `extra="forbid"` — and the `ruff` + `pyrefly` + `pytest` floor.