The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands to manage the local git environment, including resolving branch names (git rev-parse), calculating diffs (git diff), and committing applied fixes. It also invokes repository-specific scripts defined in files like package.json, Makefile, justfile, or Taskfile to perform verification checks.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the repository and a local findings file, creating a vulnerability surface where malicious code comments or data could influence agent behavior.
Ingestion points: Local source code files, repository configuration files, and the findings log at .agent/{branch}/findings.jsonl (SKILL.md, Step 1).
Boundary markers: No explicit instructions are provided to the agent to treat finding descriptions or repository content as untrusted data during the application phase.
Capability inventory: The skill has the ability to write to the file system (applying fixes), execute arbitrary repository scripts, and dispatch further sub-agents.
Sanitization: There is no evidence of sanitization or filtering applied to the findings data before it is used to drive code modifications.
[DYNAMIC_EXECUTION]: The skill executes a Node.js fingerprinting script using a path resolved at runtime via the $AGENT_SUPPORT variable. It also dynamically discovers and executes build or test targets specified in the repository's documentation or configuration files during its verification step (SKILL.md, Sub-step D).

review-branch