The Agent Skills Directory

[COMMAND_EXECUTION]: Executes local utilities such as git, openssl, and jq, as well as model provider CLIs including claude, codex, grok, and cursor-agent to perform grounding and cross-model peer review.
[COMMAND_EXECUTION]: Utilizes a custom Python task runner (peer-job-runner.py) to manage detached sub-agent processes. This runner implements specific security controls to prevent unauthorized access in shared environments, including restrictive file permissions (0600) and owner-identity verification (os.fstat vs os.geteuid) before reading job results.
[EXTERNAL_DOWNLOADS]: The external-evidence-researcher sub-agent uses web tools to fetch information about technologies, maturity, and pitfalls from public internet sources.
[DATA_EXFILTRATION]: The "Cross-Model POV Panel" protocol transmits project summaries and verified facts to external model providers (OpenAI, Anthropic, xAI) for independent peer review. This behavior is explicitly documented and governed by a mandatory privacy notice and user consent mechanism described in references/cross-model-panel.md.
[PROMPT_INJECTION]: The skill processes untrusted external data from the web and project files (Indirect Prompt Injection surface). Risks are mitigated by the use of explicit boundary markers (<pov-payload>) and defensive instructions in the peer persona (references/agents/pov-peer.md) to treat input as data rather than instructions.

pov