The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external CLI commands (codex exec and claude -p) to obtain judgments from alternative models. This is gated by environment detection to ensure model decorrelation.
[EXTERNAL_DOWNLOADS]: The skill requires the presence of codex and claude binaries on the system PATH. It performs version checks against lib/model-anchors.md to ensure the environment meets verified minimum requirements.
[PROMPT_INJECTION]: The skill incorporates potential vulnerability surfaces for indirect prompt injection as it processes output from external models. It mitigates this via a mandatory evidence chain:
Ingestion points: Data enters the session from external CLI output (codex or claude) in SKILL.md.
Boundary markers: The instructions explicitly enforce a "read-only sandbox guard" for Codex and disable tools/persistence for Claude calls.
Capability inventory: Capabilities are limited to stdout and specific CLI execution; operations like git commit and git push are explicitly forbidden.
Sanitization: All findings must pass through a manual four-option user gate (approve / edit / reject / skip) defined in lib/gate-contract.md before any action is taken.

advisor