Imported from beartype/bearshape (AGENTS.md). Install upstream with npx skills add beartype/bearshape. Copyright stays with the author.

bearshape agents

bearshape checks runtime array shapes, dtypes and tree constraints through beartype. Scientific Python developers and library authors use its annotations at API boundaries alongside native backend types. Keep the public identity lowercase bearshape.

This is the shared Codex/Claude instruction file. CLAUDE.md points here. Human setup and commands live in CONTRIBUTING.md; tool usage lives in tools/README.md. Read docs/maintainers/production-readiness.md and the current plan before changing a contract. Historical plans record their original evidence; current owner decisions take precedence.

Working rules

• State assumptions before changing behavior. Resolve routine implementation choices within the approved scope; ask when a missing decision changes the public contract or acceptance criteria.
• Maintenance is now surgical by default. The rename is complete. No speculative features, duplicate helpers, unused abstractions or broad cleanup unrelated to the requested work.
• Preserve public annotation syntax and useful errors. Use explicit names, simple control flow and the existing 2-space Python indentation.
• Keep runtime behavior and static consumer behavior aligned. Passing syntax checks, skipped tests and inferred Any do not prove support.
• Do not spawn other agents unless the user requests delegation.

Current release contract

• GitHub ownership transferred on 2026-09-10: use beartype/bearshape and https://beartype.github.io/bearshape/. Preserve original authorship and historical plan evidence. PyPI ownership and trusted-publisher configuration are separate and remain unverified; future publishing must trust owner beartype, repository bearshape, workflow pypi.yml, environment pypi.

• bearshape is independently versioned; the candidate is 0.1.0rc0.

• The current dependency and test target is beartype 0.23.0rc0, with metadata >=0.23.0rc0,<0.24. Full native composite-union rollback is still incorrect. The owner chose full composition through supported upstream integration and permits a later candidate. Keep rc0 validation until a concrete replacement is implemented and tested; update metadata, lock, preflight and CI together.

• Do not replace beartype, monkeypatch upstream internals, infer union boundaries from bytecode, hide the defect with an xfail, or restrict composition to make CI pass. tools/probe_union.py reproduces the open release blocker.

• pyright, mypy, ty and pyrefly are supported through the consumer harness on Python 3.10–3.14. New checkers, including zuban, need an explicit plan.

• Native CuPy static typing is explicitly limited, as accepted by the owner. Its protocol fallback does not prove native ndarray method inference. Actual GPU runtime evidence is recorded separately in the handoff report.

• Release readiness remains withheld until the union integration and release prerequisites are satisfied. Merge approval does not authorize publication, deployment, access changes, ownership transfer or messages to maintainers.

Product boundaries

• Root imports must not load NumPy, JAX, Torch, CuPy or optree. Backend behavior belongs in its explicit module; do not widen the root API for shorter examples.
• Preserve named, anonymous, fixed, arithmetic and broadcastable dimensions, Scalar, Value(...), backend aliases, Like[...], scalar-like types and Tree[...]. Dimensions are runtime constraints, not static shape proofs.
• NumPy owns the broadest dtype/scalar/Like surface. Other backends use their own native arrays and selected converters. Like validation does not replace the caller's argument or retry failed conversions through a different backend.
• Bare @check manages memo scope. Combine it with @beartype, or use @check(conf=...), to perform type checking.
• Keep runtime tree structure syntax distinct from checker-supported containers. optree uses the default registry; there is no custom namespace API.
• Value expressions are trusted developer contracts, not a security sandbox.

Memo ownership, frame discovery, failed-alternative rollback, async contexts, decorator metadata, dtype normalization and static aliases need focused positive and negative regressions. Preserve lifetime/weak-reference evidence when changing state ownership. Read the feature-to-test map in the handoff report.

Major-work workflow

For features, contract changes, packaging, CI, release work and substantial refactors: create a feature branch and matching worktree, commit an ExecPlan following PLANS.md, and open a PR before implementation. Keep independent changes in focused PRs and document any dependency on another PR. Use the existing task goal when one exists; do not start duplicate goals.

Keep the plan current and validate each milestone. Present a concrete result before requesting merge approval. Reuse explicit approval already given in the conversation for that scope; do not ask again solely because work crossed a milestone or worktree. After approval and passing checks, merge, verify main and remove only clean merged branches/worktrees after preserving evidence.

For stacked PRs, merging a head into its actual base can mark the PR merged. Before approval, use a separate integration branch to validate combinations. After approval, preserve every reviewed head in the merged history, including conflict resolutions; verify reachability before cleanup. Never force main or delete another contributor's changes. Small, explicitly requested low-risk edits may skip the major-work workflow.

Validation

Use the locked uv toolchain. Start with uv sync --locked; install hooks with uv run --locked prek install. Use these routes from the repository root:

• Runtime changes: uv run --locked pytest tests/ --ignore=tests/test_typecheck.py -n auto.
• Typing changes: uv run --locked pytest tests/test_typecheck.py -q; this runs all four engines, source checks and real positive/negative/inference consumers with interpreter-matched settings. Direct checker commands are debugging aids.
• Normal hooks: uv run --locked prek run -a.
• Workflows: uv run --locked prek run actionlint -a --stage manual.
• Select relevant tox, docs/notebook, minimal and installed-archive checks from CONTRIBUTING.md and tools/README.md; do not invent environment factors.

Run a failing regression first when practical. Broaden validation when inputs or unresolved risks justify it, not by repeatedly rerunning unchanged expensive suites. Do not lower coverage, suppress diagnostics or remove negative fixtures to obtain green checks. Stage new files before running hooks; review formatting changes before committing. Git subprocesses targeting a different repository must clear the names reported by git rev-parse --local-env-vars; use the existing git_environment helper for release tooling and its fixtures. Real hook execution is part of validation when changing lifecycle stages.

CI, nightly and candidate validation share .github/workflows/validate.yml. Required lanes include CPU backends, four checkers, compatibility floors, docs, notebook and normal installed-wheel consumers. Minimal environments require an exact sync or a fresh venv; uv run --only-group can retain earlier packages. CuPy needs real CUDA hardware. CPU skips are not GPU evidence, and historical GPU results do not validate a changed package automatically.

Docs and handoff

Keep README concise and examples test-backed. Build and inspect changed rendered structures; execute changed runnable examples and notebook cells. Document runtime-only syntax and backend/checker limits explicitly. Update CHANGELOG for features and major user-visible changes.

Record source commits, versions, artifact hashes, commands and meaningful results in the plan or handoff report. Distinguish implementation completion, merge state and release readiness. Keep public repository/PyPI/Pages identities unchanged until the corresponding ownership operation has actually happened.