The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from code changes and repository files that could contain embedded instructions intended to manipulate the agent's behavior.
Ingestion points: The skill reads working tree content, staged diffs, commit ranges, branches, and pull requests as defined in the 'Establish the review contract' section of SKILL.md.
Boundary markers: The skill lacks explicit instructions to use boundary delimiters or specific warnings to ignore instructions embedded within the reviewed code.
Capability inventory: The skill instructs the agent to use 'reproduction commands already supported by the repository' and other build/test tools in the 'Validate proportionately' section of SKILL.md.
Sanitization: The instructions do not require the agent to sanitize or escape the content of the reviewed code before analysis or execution.
[COMMAND_EXECUTION]: The skill directs the agent to execute shell commands for verification, such as linters, type checks, builds, and repository-defined reproduction scripts. Executing reproduction commands from potentially malicious pull requests introduces a security risk if the environment is not sufficiently isolated.
[NO_CODE]: The skill consists entirely of instructional markdown files (SKILL.md and referenced files) and does not package any scripts or binary executables.

superior-code-review