The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from the codebase being reviewed (via git diffs and file reads) and passes this data to multiple specialized sub-agents. This creates a surface for indirect prompt injection attacks where malicious code could contain instructions designed to influence or override agent behavior. 1. Ingestion points: Git diff output and source file contents are read and interpolated into sub-agent prompts. 2. Boundary markers: The skill implements a standardized Agent Prompt Template that includes clear section headers and an explicit warning to sub-agents to ignore instructions found in the analyzed code. 3. Capability inventory: The skill and its sub-agents have access to Bash, Read, Write, and Task tools. 4. Sanitization: The raw diff data and file contents are not sanitized before being passed to the sub-agents.

deep-review