project-code-review

Review a diff against dynamic-neural-field-composer's own standards - backwards compatibility, Clean Code, tools/ helper reuse, C++20 idiom, hot-path performance, and test coverage. Use before opening

SKILL.md

GitHub

Project code review

Review git diff origin/main...HEAD (or the diff you were given) against the checks below. This deliberately overlaps .coderabbit.yaml so problems surface locally before CodeRabbit posts them on the PR.

1. Backwards compatibility (highest severity)

Not a finding to note - a stop condition. If the diff breaks compatibility, say so first and stop; do not continue reviewing style.

Look for: changed public signatures in include/, altered default parameter values, changed behaviour of an existing element, and any change to the .dnf serialization format that older files would not survive.

2. Clean Code

Per CONTRIBUTING.md:

  • Names that read like prose. n, tmp, data2 are findings.
  • Small, single-responsibility functions. A function that needs a comment to explain its sections wants to be several functions.
  • No raw owning pointers. shared_ptr / unique_ptr only.
  • No comments that explain what the code does. Only why, and only when non-obvious.
  • Clarity over cleverness - if a reader has to stop and decode a line, it is a finding.

3. Helper reuse

Every new free function: does something equivalent already live in tools/?

Search include/tools/math.h, utils.h, profiling.h, logger.h, fft_convolution.h, simd_dispatch.h, and include/exceptions/exception.h before accepting it.

Also flag helpers defined at a call site that should move into tools/ - a numeric or utility function sitting in an element .cpp is misplaced.

4. Modern C++20

Flag missed opportunities where they genuinely improve the code:

  • std::span instead of const std::vector<double>& for read-only array views
  • std::string_view for read-only string parameters
  • [[nodiscard]] on accessors and anything whose result must not be dropped
  • ranges, structured bindings, std::format over manual concatenation
  • enum class over bare enum

Do not flag idiom for its own sake. If the change is cosmetic and the file is consistent as-is, leave it.

5. Performance

This is a simulation library; the step loop runs thousands of times a second.

  • Copies of heavy objects - field matrices, kernel vectors, std::vector<double> returned by value from an accessor
  • shared_ptr copies (atomic refcount) inside step() or any per-iteration path
  • Allocation inside the step loop that could be hoisted or reused
  • Repeated map/string lookups in a hot path

6. Tests

  • New or changed behaviour in elements/, simulation/ or tools/ with no test
  • A new test file not added to tests/CMakeLists.txt - it is silently never compiled and the suite goes green having never run it. Check this explicitly every time.
  • Floating-point assertions without an explicit tolerance
  • A test that would pass against the unfixed code, i.e. does not actually pin the bug

Output

Order findings most severe first. For each: file:line, one sentence on what is wrong, and the concrete failure it causes or the standard it violates.

Say "clean" when it is clean. Do not manufacture findings to look thorough - a short honest review beats a padded one.