The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs Git operations using a Python helper script (scripts/git-squash.py) that wraps the git binary.
The script uses subprocess.run with arguments passed as a list, which is a secure implementation that avoids shell injection vulnerabilities.
The operations are limited to standard Git maintenance tasks such as git reset --soft, git commit, git log, and git status.
The script includes safety checks to ensure the working tree is clean and that the repository state has not changed between the planning and application phases.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external data (Git commit messages and diffs) which serves as a surface for indirect prompt injection.
Ingestion points: The preflight function in scripts/git-squash.py reads commit subjects and author information from the Git log.
Boundary markers: The SKILL.md instructions explicitly scope the agent's attention to the "plan's commits and the net diff" to generate the final commit message.
Capability inventory: The agent has the capability to write a new commit message to a file and execute a git commit via the helper script.
Sanitization: The skill relies on the agent's own semantic evaluation of the diff versus the commit messages to ensure the final output is accurate, acting as a natural check against malicious instructions embedded in commit history.
[EXTERNAL_DOWNLOADS]: The skill instructions reference the use of uv, a well-known and widely used Python package manager, to execute the local helper script. This is a standard development practice and does not involve fetching untrusted code from remote sources during execution.

git-squash