The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates by ingesting and analyzing untrusted data from source code diffs and external specifications. This creates a surface for indirect prompt injection where malicious instructions embedded in the code being reviewed could attempt to influence the agent's logic or the fixes it suggests.
Ingestion points: Source code files are read via git diff and git status, and project specifications are ingested from README.md files or external documentation.
Boundary markers: The skill uses instructional headers and templates but lacks strict data/instruction isolation markers for the processed code.
Capability inventory: The skill can perform file-write operations (to apply fixes) and execute shell commands (git utilities).
Sanitization: There is no evidence of sanitization or filtering of the ingested code diffs before they are processed by the agent.
[COMMAND_EXECUTION]: The skill relies on local shell command execution to interface with the version control system. It uses git to retrieve changes, check branch status, and resolve commit metadata. These operations are limited to the local repository environment.
[DYNAMIC_EXECUTION]: The skill includes shell script templates in references/execution.md that are intended to be executed at runtime. These scripts compute a project fingerprint using git ls-files and git hash-object to implement a caching mechanism for audit reports.

code-review