The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes local command-line tools including git, gh, delta, and mergiraf for repository interaction and diff inspection. It also executes a local Python script (shared/scripts/severity.py) to compute finding severities based on parameters determined by the agent's analysis.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests and processes untrusted content from the codebase and PR metadata.
Ingestion points: Untrusted data enters the context through git diff outputs, GitHub PR descriptions, and file content reads.
Boundary markers: The instructions do not define explicit delimiters or instructions to ignore potential commands embedded in the analyzed code.
Capability inventory: The skill has the ability to write files to the .cheese/age/ directory and can trigger the /cure skill to apply automated fixes.
Sanitization: No sanitization of the input code or PR data is mentioned.
[DATA_EXFILTRATION]: No unauthorized data exfiltration patterns were observed. All network-related tool usage (via gh and git) is directed at legitimate repository management tasks.
[SAFE]: No hardcoded credentials, malicious persistence mechanisms, or multi-layer obfuscation were detected. The skill maintains a human-in-the-loop selection gate for fix application outside of autonomous mode and follows structured reporting best practices.

age