The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to process untrusted data from the repository's codebase, PR descriptions, and issue trackers.
Ingestion points: Git diffs, PR bodies (via GitHub CLI), source code files, and external documentation (SKILL.md, references/fan-out.md).
Boundary markers: Implements a multi-step workflow (Plan -> Review -> Reconcile -> Verify) that separates data collection from verdict generation to prevent malicious instructions in code from hijacking the review process.
Capability inventory: Reads repository files, executes a local helper bundle, and writes artifacts to the '.cheese/age/' directory (SKILL.md, references/packet.md).
Sanitization: Includes a verification pass where a specialized sub-agent validates findings against the source code to filter out hallucinations or injection-driven claims (references/fan-out.md).
[COMMAND_EXECUTION]: Orchestrates its lifecycle using a bundled Python archive located at 'skills/age/scripts/age.pyz'. This script is used for critical tasks such as routing review assignments, locking the production tree, and generating the final report (SKILL.md, references/commands.md).
[SAFE]: Features a dedicated 'review-lock' mechanism that records a digest of the production tree before analysis begins. The gated writer refuses to commit the final report if any production files were modified, ensuring the skill remains review-only and cannot maliciously alter the codebase (SKILL.md, references/packet.md).

age