The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a custom Node.js relay script (scripts/relay.mjs) to execute the claude CLI and git commands. This script follows security best practices by using child_process.spawn with array arguments to avoid shell injection and validates parameters like model names and session IDs.\n- [PROMPT_INJECTION]: The skill creates an indirect prompt injection surface by passing a task brief to a sub-agent. Ingestion points: Instructions are read from stdin or a file via scripts/relay.mjs. Boundary markers: The skill utilizes XML-style blocks and clear templates to structure instructions for the sub-agent. Capability inventory: The delegated Claude session can perform file system operations and execute shell commands. Sanitization: The primary orchestrator is explicitly instructed to review the sub-agent's diffs and re-run all project gates independently.\n- [SAFE]: The skill implements significant security controls, such as a read-only mode that strips write/shell capabilities and a requirement for Claude's shell sandbox to be enabled on supported platforms (macOS, Linux, WSL2).\n- [SAFE]: Detailed documentation is provided to guide the orchestrator in maintaining a secure Review and Land workflow, ensuring that code changes are verified and committed by the primary agent rather than the implementer sub-agent.

claude-delegate