The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes data from the repository being reviewed, including code diffs and local instruction files (e.g., AGENTS.md, CLAUDE.md). This content is interpolated into prompts for subagents during the review fan-out stage.\n
Ingestion points: scripts/build_manifest.py gathers diffs and scripts/build_knowledge.py discovers local documentation and instructions. PR metadata is retrieved via the GitHub CLI.\n
Boundary markers: Reviewer templates in assets/PROMPT.md use structured sections and markdown delimiters to isolate context from instructions.\n
Capability inventory: The execution environment allows system command execution (git, gh, compozy) and file system writes for report artifacts.\n
Sanitization: The _redact function in scripts/repository_intelligence.py identifies and removes potential secrets from tool logs and environment variables before they are utilized or stored.\n- [COMMAND_EXECUTION]: The skill uses subprocess.run to interact with the version control system and external analysis tools as part of its standard workflow.\n
Evidence: Multiple scripts, including build_manifest.py and merge_findings.py, execute git and gh for repository analysis and PR updates.\n
Dynamic Execution: scripts/run_jobs.py manages the execution of reviewer tasks, while scripts/repository_intelligence.py runs analysis binaries such as graft and graphify.\n
Mitigation: A path validation check in scripts/repository_intelligence.py ensures that analysis binaries are executed from the expected local directory, preventing the use of tools from unauthorized paths.

wtk-deep-review