What problem does it solve? Letting an agent loose with write permissions on an unfamiliar or critical codebase risks destructive mistakes before anyone reviews its reasoning. This Skill enforces a two-mode contract: the agent explores in read-only Plan Mode, drafts an auditable plan.md with findings and a proposed architecture, and only transitions to writing after a human signs the plan. ## Core Features & Use Cases - Read-only exploration: Configures the agent with permission_mode="plan" and a read-only toolset (Read, Glob, Grep) so it can map structure, conventions, and dependencies without touching disk. - Hard write barrier via hooks: A PreToolUse hook denies write-capable tools (Write, Edit, NotebookEdit, and Bash commands with redirections) while Plan Mode is active, rather than merely discouraging them. - Signed approval artifact: Produces a frozen plan.md that a human signs; any change to the plan forces a return to Plan Mode and re-approval. - Use Case: You inherit an unfamiliar Python microservices repo and need to add JWT authentication to the payments service. The agent explores read-only, proposes the middleware architecture in plan.md, and writes code only after you sign off. ## Quick Start Ask the agent to explore an unfamiliar repository in Plan Mode, write its findings and proposed architecture into plan.md, and wait for your signed approval before making any code changes.
