code-audit-fanoutlisted

# code-audit-fanout ## When to use USE for any code review request where correctness matters more than speed. Especially: - ML training loops, loss functions, custom layers - Signal-processing pipelines (FFT, filters, modulation) - 5G PHY simulators (channel models, MIMO precoding, decoders) - Any function touching complex-valued tensors, dB scales, or hardware-determinism primitives - Diffs > 50 lines before merge SKIP for trivial edits (typo fix, rename, formatting). ## The 5 specialist agents Each spawned in parallel in ONE assistant message. Each has a narrow scope, a required source list, and a citation contract. ### Specialist 1 — static-analyzer **Scope:** Read-only static analysis. **Tools:** Read, Grep, Glob, Bash **Checks:** - `ruff check --select ALL` output - `mypy --strict` output (if config present, else `mypy`) - `pyright` if available - Function-level cyclomatic complexity > 10 - Bare `except:`, `except Exception:`, swallowed errors - Mutable default args - Print statements left in code (vs logging) **Citation requirement:** every flagged rule must reference the exact ruff / mypy / pyright rule ID and link to the rule docs. ### Specialist 2 — library-api-auditor **Scope:** Verify every imported API call against official docs. **Tools:** Read, Grep, WebFetch, WebSearch **Required sources (in order of precedence):** - PyTorch — https://docs.pytorch.org/docs/stable/ - NumPy — https://numpy.org/doc/stable/reference/ - SciPy — https://docs.scipy.org/doc/sci