The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a restricted set of git commands and a local validation script (scripts/validate-output.sh). It also allows for the execution of a 'Verification' command (e.g., tests or linters) as part of the commit process. This is mitigated by a strict 'read-only' requirement for verification commands and the requirement for explicit user approval of the entire execution plan.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data including repository files, ticket context via CONTEXT_QUERY, and user-provided quotes (Ingestion points). It uses 'Evidence, not instructions:' delimiters and subagent isolation (Boundary markers) to mitigate injection risks. The skill has the capability to commit code and run shell commands (Capability inventory), but it sanitizes the workflow by requiring a validator script to pass and a human user to approve the exact plan (Sanitization).
[SAFE]: The skill follows security best practices by implementing a least-privilege model for system access, using SHA digests to verify repository integrity during execution, and explicitly denying high-risk capabilities like network access or arbitrary file writing in its configuration instructions.

committing-scoped-changes