Stage

Multi-model adversarial code review for agent outputs. Use when reviewing core logic changes, 5+ files, 200+ lines, or when user says "对抗审查/adversarial review/cross-model review". The reviewer agent inspects coder outputs through a cross-validation lens to catch single-model blind spots.

Change impact scan for Python projects. When modifying function signatures, interfaces, data models, config structs, or shared modules, automatically grep all call sites and dependent files to prevent incomplete changes. Outputs a structured markdown impact report.

Pitfall recording and evolution tracking. When AI makes a mistake or user corrects output, record a PIT entry. Same-type PITs reaching 2+ trigger rule extraction; 4+ trigger skill crystallization. Use when user says "踩坑/pitfall/PIT/记录错误/犯错日志/evolve mistake/踩坑记录".

Use when developing, modifying, refactoring, or reviewing the task-orchestrator multi-agent orchestration engine. Covers design principles, architecture decisions, and development guardrails derived from the Multi-Agent Harness design reference. Triggered by changes to src/openagents_orchestration/, adding new orchestration features, or questions about orchestrator design.

Run a static code review over a target directory. Scans Python/JS/TS files for common issues (long functions, TODOs, missing docstrings, unused imports, security smells) and produces a structured markdown report.

Clean and transform tabular data (CSV/JSON). Performs deduplication, null-row removal, string trimming, and column normalization. Useful for ETL tasks that don't need pandas — uses only Python standard library.

Generate API documentation (markdown) from a Python module by AST parsing. Extracts module docstrings, function signatures, class hierarchies, and type hints. Pure Python, no LLM dependency.

Run a suite of health checks (bash commands, file existence, content matching) and produce a structured pass/fail report. Pure Python, no LLM dependency.

Batch-create directory structures and files in one shot. Given a directory tree and file contents, creates everything atomically. Use this when a coder agent needs to scaffold a multi-file project skeleton without burning steps on individual write_file calls.

Generate pytest test skeletons for a Python module by AST parsing. Extracts function signatures and produces parametrised tests, edge-case stubs, and fixture hints. Pure Python, no LLM dependency.

Aggregate web content from a list of URLs into a structured research brief. Fetches each URL, strips HTML to text, deduplicates paragraphs, and assembles a markdown report. The caller provides sources — this skill does not search.