The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes external, potentially untrusted data from git commits, PRs, and source code which could contain malicious prompt injection attempts.
Ingestion points: SKILL.md specifies that the agent should take scope from $ARGUMENTS (commit, range, branch, PR#) and read the touched source code files.
Boundary markers: Absent. The instructions do not define delimiters or specific safety warnings to prevent the agent from following instructions embedded within the code it is reviewing.
Capability inventory: The skill utilizes shell commands (git status) and file system access to read and potentially write source files.
Sanitization: Absent. There are no defined procedures for the agent to sanitize or escape the content of the files being analyzed before processing.
[COMMAND_EXECUTION]: The skill directs the agent to execute git commands and potential project-specific compilers.
Evidence: The 'Target' section explicitly requires running git status --short to orient the review.
Context: While standard for a development tool, this involves interacting with the shell using inputs that may be derived from the environment or $ARGUMENTS.
[DYNAMIC_EXECUTION]: The skill instructs the agent to generate and compile code snippets at runtime to verify proposed remedies.
Evidence: Standard 11 instructs the agent to "Compile the construct you propose, in a scratch file outside the repo, using the project's own type-checker."
Risk: This involves script generation and execution of external build tools (like tsc) on agent-generated content, though the instruction specifies using a scratch file to minimize impact.

type-precision-review