cef Agent Skills & AI Rules

# CEF Tools Contributor and Agent Guidelines

This file applies to `//cef/tools` and its descendants.

## Generator inventory and coverage

- Every direct-child, non-test `make_*.py` generator must have a matching
  `make_*_test.py` module.
- `make_generator_inventory_test.py` enforces the direct-child inventory. When
  adding or removing a generator, update its expected count intentionally and
  keep nested generators, helpers, and existing tests out of the inventory.
- Exercise every top-level generator function through at least one successful
  test. Cover meaningful explicit error branches, writer behavior, and the
  generator's CLI style; do not treat filename coverage as function coverage.
- Use standard-library `unittest`, `tempfile`, and `unittest.mock`. Run from the
  CEF root:

  ```sh
  python3 -m unittest discover -s tools -p '*_test.py'
  python3 -m unittest discover -s tools -p 'make_example_test.py' -v
  ```

- Run the complete suite twice for generator-wide work. These modules share
  parser state, so the second run detects ordering and cleanup leaks.

## Hermetic fixtures and goldens

- Put deterministic inputs and reviewed goldens under
  `tools/testdata/make_generators/`; use `generator_test_util.py` for shared
  fixture, manifest, archive, and CLI helpers.
- Mock or fix time, year, Git metadata, versions, platforms, environment
  variables, external commands, and formatter results when they affect output.
- Goldens must originate from a reviewed baseline, never from ignored
  working-tree output, and tests must not contain an auto-update path.
- Use inline expectations for small output and reviewed full-file goldens for
  larger generated C/C++ content. Add targeted assertions that explain the
  behavior protected by a golden.
- Prove that customized CppToC/CToCpp bodies survive regeneration, source
  fixtures are not modified, and relevant paths work when they contain spaces.

## Writer contracts

- Test every writer on its first generation and an identical second
  generation.
- For `write_file_if_changed`, assert changed/unchanged returns and preserved
  mtime. For unconditional writers, assert stable bytes across rewrites.
- Multi-output writers must report a change when any output changes. Test
  repairs where only one output was corrupted.
- Pre-publication failures must leave outputs missing or unchanged. Characterize
  partial I/O-failure output before changing it; atomic publication is a
  separate behavior change.

## CLI contracts

- Cover invalid status, stdout, stderr, and partial artifacts, plus successful
  stdout, single-file, multi-output, argparse, and distribution CLI styles.
- If a generator function returns metadata with its content, such as
  `(content, customized)`, unpack the result at the CLI boundary and write only
  the string content. Do not change the underlying generator return contract
  merely to fix a CLI.
- Assert progress messages exactly when callers may depend on them, and return
  explicit nonzero statuses for rejected invocations.

## Parser and module state

- Restore parser notification context and other module globals with
  `try/finally`; add test cleanup so failing regressions cannot leak state.
- Generator modules intended for reuse must be importable without parsing
  arguments, exiting, mutating files, calling Git, or launching subprocesses.
  Put orchestration in `main(argv=None)` and use a conventional guarded entry
  point when refactoring for importability.

## Output equivalence

- Establish an exact pre-refactor baseline and keep testability refactors
  separate from behavior fixes.
- Testability refactors must preserve generated bytes, paths, ordering,
  comments, whitespace, guards, encoding, newlines, modes, and symlink targets.
- Compare fixes to the latest approved baseline while retaining the original
  refactor baseline as an immutable reference.
- Return-value, status, diagnostic, and CLI changes must not alter generated
  files unless that is their documented purpose.
- For translator changes, run a repository-scale comparison covering C API,
  CppToC, CToCpp, wrapper, dylib, and GYPI outputs.
- For distribution changes, compare the complete normalized tree. Compare
  archive member paths, types, uncompressed bytes, and modes; normalize archive
  timestamps only. Do not use raw archive-container hashes as equivalence.
- Use the same Python interpreter as the recorded baseline. Python-version
  differences can change comment wrapping and produce misleading byte deltas.

## Bug-fix workflow

- Reproduce a bug with a focused regression before changing production code.
- Keep coverage/refactor checkpoints separate from behavior fixes so output
  differences cannot be attributed to the wrong change.
- Keep unrelated bugs independently reviewable. Closely related instances of
  the same root cause may be grouped when explicitly requested.

## Formatting and validation

- From the CEF root, format only explicit changed Python paths:

  ```sh
  python3 tools/fix_style.py tools/make_example.py tools/make_example_test.py
  ```

- Inspect formatter output immediately. The formatter may rewrap unrelated
  legacy code; revert that churn with targeted patches.
- Compile every changed generator, helper, and test.
- Run `git diff --check` and scan untracked files for trailing whitespace.
- Re-run focused tests after formatting, then run the complete suite twice.