The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted external code and diffs, creating a surface for indirect prompt injection.
Ingestion points: The workflow accepts an initial_review_tree, fix diff content, and Actor output as primary inputs.
Boundary markers: The skill utilizes specific markdown headers and JSON schema constraints to separate untrusted data from the core instructions.
Capability inventory: The skill explicitly prohibits performing high-privilege operations such as git commit, git push, PR creation, or version releases.
Sanitization: While the skill canonicalizes finding rows into structured JSON, it does not describe specific sanitization or filtering logic for the raw code content it processes.
[DYNAMIC_EXECUTION]: The skill dynamically resolves and loads other project-specific skills to provide capabilities and context to the judge and fix agents.
Evidence: The Execution section of SKILL.md specifies resolving project skills and injecting their paths into prompts. The prompt templates in references/prompts-and-formats.md include placeholders like {matching SKILL.md paths} for loading skills at runtime based on the project environment.

gentle-ai-judgment-day