The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted code from git diffs and pull requests to perform reviews. This content could contain malicious instructions designed to manipulate the agent's behavior during the analysis or the subsequent fix application phase. \n
Ingestion points: Code diffs from git diff, pull request data from gh pr diff or glab mr diff, and untracked files via ls-files. \n
Boundary markers: None. The skill does not define specific delimiters to isolate untrusted code from instructions. \n
Capability inventory: Execution of shell commands (git, gh, glab), file system modification (in the --fix loop), and network-enabled search tools. \n
Sanitization: None. The skill reads and processes external content directly.\n- [COMMAND_EXECUTION]: The skill constructs and executes shell commands using user-provided $ARGUMENTS. While it includes steps to verify git references, unvalidated input in the arguments string could lead to command injection if not handled securely by the underlying platform.\n- [DATA_EXFILTRATION]: The skill transmits session-specific metadata, including sessionId, sessionRepository, and agentVendor, to the external 'Baz' service. These identifiers are used for session correlation within the vendor's infrastructure.\n- [DYNAMIC_EXECUTION]: The skill dynamically generates code changes and executes local build or test tools (e.g., project tests and linters) when the --fix flag is used.\n- [EXTERNAL_DOWNLOADS]: The skill fetches pull request and repository data from well-known services such as GitHub and GitLab using official CLI tools.

review