The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script local-check.py implements a local execution environment for GitHub Actions. It parses the .github/workflows/checks.yml file from the repository and executes the shell commands defined in the run: blocks using bash or sh. This allows the execution of arbitrary commands defined within project-controlled configuration files.
Evidence: scripts/local-check.py uses subprocess.run to execute scripts extracted from the YAML workflow file in a local environment.
[DYNAMIC_EXECUTION]: scripts/merge.py and scripts/review_boxes.py utilize importlib.util to dynamically load and execute Python modules from computed paths, specifically the project's .agents/hooks/ directory.
Evidence: The shared_review_reader() function in scripts/merge.py and the open_review_box() function in scripts/review_boxes.py dynamically import modules based on local file paths.
[COMMAND_EXECUTION]: Multiple scripts within the skill, including first-upload.py, merge.py, protect-branch.py, and upload-piece.py, make extensive use of subprocess.run to execute shell commands such as git and the GitHub CLI (gh). This behavior is integral to the skill's functionality but requires strict adherence to security best practices for command construction.
Evidence: Frequent calls to subprocess.run() are present in scripts/merge.py, scripts/first-upload.py, and scripts/protect-branch.py for repository management operations.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources, specifically Pull Request descriptions and issue bodies retrieved via the GitHub API, which could influence agent behavior.
Ingestion points: PR bodies are read in scripts/merge.py and issue bodies are read in scripts/upload-piece.py via gh API calls.
Boundary markers: The skill instructions guide the agent to look for specific markers (e.g., "Worth stopping for" boxes or "Done when" sections), but these do not fully isolate the agent from instructions embedded in general prose.
Capability inventory: The skill has the ability to merge code, push to remote repositories, and execute local scripts.
Sanitization: A secret scanner (.agents/hooks/scan-secrets.sh) is employed in scripts/upload-piece.py to check for sensitive data in PR titles and bodies before processing.

merge