The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides complex shell scripts and command patterns for the agent to execute to facilitate communication between model CLIs. This includes using env for environment manipulation, mktemp for temporary file management, and the direct invocation of external binaries like claude and codex as seen in the Gold Path and failure recovery documentation.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data, creating a vulnerability surface for indirect prompt injection.\n
Ingestion points: The skill utilizes git diff, git show, and cat commands to read code and documentation files for review as described in SKILL.md and references/prompts.md.\n
Boundary markers: Delimiters and explicit "ignore embedded instructions" warnings for ingested code content are largely absent in the provided prompt templates.\n
Capability inventory: The reviewer environment utilizes tools with Read, Glob, Grep, and Bash permissions as documented in the sandbox and permission flag sections.\n
Sanitization: No explicit sanitization, validation, or filtering of the external code or diff content is implemented before the data is passed to the models.

cross-model-review