The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill relies on external configuration data from CLAUDE.md to resolve file system paths for git operations.
Ingestion points: The data path and repo-alias are parsed from the local CLAUDE.md file in the project root.
Boundary markers: Absent; there are no instructions to the agent to treat the parsed paths as potentially untrusted or to validate them against a whitelist.
Capability inventory: The skill uses the Bash tool to perform repository status checks, file staging (git add), and commits.
Sanitization: Absent; the skill does not verify that the resolved paths are within the intended project scope or check for shell metacharacters that could lead to command injection.
[COMMAND_EXECUTION]: The skill executes Git commands via the Bash tool, using paths and messages derived from user input and project files.
Evidence: Sequential bash calls such as git -C <data-path> status and git -C <data-path> commit -m "<message>" are defined in the flow section.
[PROMPT_INJECTION]: The skill includes instructions to bypass human confirmation gates when a specific flag is provided, increasing the risk of unintended file system modifications.
Evidence: The --yes flag is documented to "skip human confirmation gate — auto-approve the commit proposal," allowing the agent to commit changes without explicit user approval.

commit-repo