The Agent Skills Directory

[SAFE]: The skill design follows the principle of least privilege by restricting subagents to read-only operations and limiting output to the .claude/plan/ directory. It explicitly prevents automatic execution of generated commands, requiring manual user intervention. The processing of local codebase files is a standard function for this tool and is protected by a human-in-the-loop validation step. [Indirect Surface Evidence]: 1. Ingestion points: Phase 2 context retrieval from the workspace. 2. Boundary markers: No explicit delimiters for codebase content. 3. Capability inventory: Writing implementation plans and calling the Agent tool. 4. Sanitization: Not implemented, mitigated by mandatory user review.

ac-plan