The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the Command Code CLI (cmd) and Git to perform and review coding tasks. The relay.mjs script strictly validates all CLI arguments (such as session IDs and model names) against whitelist regular expressions to prevent command injection.
[PRIVILEGE_ESCALATION]: The skill facilitates "full-trust" runs by using the implementation agent's --yolo mode, which bypasses internal sandboxing. This high-privilege access is clearly documented, and the skill provides a multi-step review process (diff checking, gate re-running, and manual commits) to ensure human oversight.
[DYNAMIC_EXECUTION]: The skill utilizes a Node.js relay script to manage implementer sessions and capture detailed event logs. This script uses only standard Node.js libraries and implements file-level integrity checks (SHA256 hashing) to detect unauthorized file modifications during restricted runs.
[INDIRECT_PROMPT_INJECTION]: As an implementation tool, the skill processes natural language "briefs" that could potentially contain subverting instructions for the implementer agent.
Ingestion points: Instructions are read from a "brief" text file or standard input.
Boundary markers: Briefs are structured using XML-style tags (<task>, <action_safety>) to provide clear intent delimiters for the implementer agent.
Capability inventory: The implementer agent has full write access to the repository and can execute shell commands.
Sanitization: The skill enforces a strict "review and land" policy where the orchestrator must manually verify all changes and re-run project test suites before committing code.

commandcode-delegate