The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the local project environment to perform its audit functions.
Ingestion points: Files matching patterns such as *eval*.py, *.json, *.csv, and *.yaml are scanned in SKILL.md (Step 1) and contents are potentially read by the reviewer.
Boundary markers: The workflow relies on natural language instructions to separate the executor's intent from the data being reviewed, but does not implement cryptographic or robust delimiter-based isolation.
Capability inventory: The skill is granted Bash(*), Read, Write, Edit, Grep, and Glob tools, along with access to specific MCP review tools (mcp__codex__codex).
Sanitization: There is no evidence of sanitization or filtering of the content within the identified files before they are passed to the reviewer backend.
[COMMAND_EXECUTION]: The skill utilizes Bash(*) and Grep to perform discovery of experiment artifacts within the user's workspace, which is standard behavior for a project auditing tool.

experiment-audit