The Agent Skills Directory

[DATA_EXPOSURE]: The skill implements a proactive security measure by explicitly instructing the agent not to commit sensitive files such as .env or credentials. This aligns with industry best practices for secret management.
[COMMAND_EXECUTION]: The skill utilizes standard, read-only and repository-management commands (git status, git diff, git commit) via the Bash tool. These operations are restricted to the intended purpose of managing a local git repository.
[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection as it ingests untrusted data from repository diffs and status reports to generate commit messages.
Ingestion points: Reads repository state via git status and file contents via git diff in SKILL.md (Phase 1).
Boundary markers: None explicitly defined for the diff content.
Capability inventory: Uses Bash and Read tools for repository interaction.
Sanitization: Relies on the agent's logic to summarize diffs rather than direct interpolation into shell commands.
[SAFE]: The skill-report.json file contains metadata and a prior security audit that correctly identifies several static analysis findings (such as markdown code fences being mistaken for executable backticks) as false positives. The referenced GitHub repository in the metadata is used for provenance and does not involve remote code execution.

git-commit