The Agent Skills Directory

[SAFE]: The skill operates in a strictly read-only mode, with explicit instructions preventing the agent from modifying files, creating commits, or interacting with remote repository features like comments or branches.
[COMMAND_EXECUTION]: The skill allows for the execution of local project commands (such as tests or linters) to verify code defects. This is performed safely through instructions that explicitly prohibit the installation of new dependencies or the use of package runners that fetch external tools.
[PROMPT_INJECTION]: The skill is exposed to indirect prompt injection due to its requirement to process untrusted code diffs and repository files.
Ingestion points: Code changes, pull request diffs, and existing repository files provided by the user for review.
Boundary markers: The instructions mandate resolving exact review targets and using merge bases for comparison, which helps scope the input, though it does not specify unique textual delimiters for the ingested code.
Capability inventory: The agent can invoke pre-existing local diagnostic tools and project-specific commands.
Sanitization: The skill requires the agent to validate all findings against the surrounding implementation and repository history, providing a logical verification layer before any finding is reported.

code-reviewer