[Skill] pr-review: Automated kernel PR review checks

[sunway513]

Skill

pr-review

Priority: P1 (High — catches common PR errors before CI)

Motivation

Kernel PRs frequently have subtle errors that pass CI but cause runtime failures or documentation drift: parameter names that don't match function signatures, dtype support claims that aren't implemented, missing boundary checks for edge-case shapes, and import paths that reference internal Triton functions instead of exported APIs. A skill automates these checks, catching errors that human reviewers often miss.

What This Skill Should Do

1. Signature validation: For every function call in the PR diff, verify parameter names match the actual def signature (catches common mistakes like fc1_scale vs w1_scale, residual vs res)
2. Dtype claim verification: If PR claims support for new dtypes (FP8, FP4, BF16), verify the kernel actually handles those types (check dispatch logic, not just docstrings)
3. Boundary condition checks: Verify the kernel handles edge cases — M=1 (single token), non-aligned shapes, empty inputs, shapes where N < BLOCK_N
4. Import path validation: Check that from aiter import X paths reference actual exports in __init__.py, not internal modules
5. Documentation consistency: If PR modifies function signatures, check if corresponding docs/ guides need updating

Acceptance Criteria

• Can parse PR diff and extract all function calls with their parameter names
• Detects parameter name mismatches against actual function signatures
• Verifies dtype dispatch logic matches claimed dtype support
• Checks for common boundary condition issues (M=1, non-aligned N/K)
• Validates import paths against init.py exports
• Flags documentation that needs updating when signatures change