The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to execute a local git check-ignore command within the peer_profiles tool. This is a security feature designed to warn users if their configuration files containing API keys are accidentally being tracked by version control.
[EXTERNAL_DOWNLOADS]: The peer_ask tool facilitates network requests to external OpenAI-compatible API endpoints using urllib.request. This behavior is the core purpose of the skill, enabling cross-model deliberation. The implementation includes an "Egress Consent" mechanism (Requirement R5) to ensure the operator approves all outgoing transmissions.
[DATA_EXPOSURE]: The skill manages sensitive API keys through environment variables or local configuration files located in the user's home directory (~/.peer-review/profiles.json). The code includes explicit redaction logic to prevent keys from leaking into error logs and ensures the peer_profiles tool does not return secret values to the agent.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (artifacts and specs) by embedding them into a "packet" for analysis by another model.
Ingestion points: The peer_ask tool reads a payload file from the .peer-review/ directory containing the artifact to be reviewed.
Boundary markers: The skill uses structured Markdown headers and specific system prompts defined in round-prompts.md to guide the challenger model.
Capability inventory: The server has the capability to perform network requests and local git status checks.
Sanitization: The skill intentionally avoids sanitizing the artifact content to maintain "transmission fidelity" (Requirement R15), relying instead on protocol-level separation and model guardrails to manage injection risks.

cross-model-peer-review