The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to perform shell operations to verify and execute external CLI tools (e.g., Gemini and Codex) for cross-model reviews.
Evidence: Commands such as which gemini, gemini --version, and piped execution strings are detailed in the "Cross-model escalation" section of SKILL.md.
Guardrails: The skill mandates path validation, binary testing, and explicit user authorization for every invocation to prevent unauthorized or accidental command execution.
[PROMPT_INJECTION]: The skill processes potentially untrusted data (code artifacts and contracts), which presents a risk of indirect prompt injection targeting either the reviewer model or the underlying execution shell.
Ingestion points: The ARTIFACT and CONTRACT inputs described in the Step 3: DOUBT section of SKILL.md.
Boundary markers: The instructions explicitly recommend passing content via stdin or temporary files to ensure that shell metacharacters in the artifact remain inert.
Capability inventory: Shell command execution via the gemini and codex command-line interfaces.
Sanitization: The skill advocates for the use of read-only sandboxes to prevent the external CLI from modifying the local workspace if it encounters malicious instructions within an artifact.

doubt-driven-development