What problem does it solve? Python code in AI/MLOps and agent codebases has sharp edges that formatters and linters miss: async bugs that stall event loops, agent loops without iteration or budget bounds, unsafe model deserialization, and silent reproducibility failures. This Skill runs a systematic review that catches these correctness, security, and reproducibility issues before human review. ## Core Features & Use Cases - Static analysis integration: Runs pyright (or mypy) and ruff check on changed files, but only when the project already configures them — it never installs tools into the project. - Domain-specific checklist: Walks a structured checklist covering core correctness and typing, async correctness, agent-building invariants (iteration bounds, tool-argument validation, prompt-injection surface), ML reproducibility and serialization safety (torch.load, pickle, seeds), and packaging hygiene. - Structured findings: Reports each issue as file:line with a category (correctness/security/reproducibility/performance), severity (must-fix/should-fix/consider), and a suggested fix. - Use Case: Before opening a PR on an LLM agent codebase, say "review this agent code" to catch missing await calls, unbounded tool-use loops, eval/exec on model output, and torch.load without weights_only=True. ## Quick Start Ask the assistant to review the changed Python files in this project for typing, async, agent-loop, and reproducibility issues.
