What problem does it solve? AI agents can invoke tools that mutate state, send messages, or change configuration without oversight, causing irreversible damage or unauthorized communication. This Skill enforces a governance layer that classifies every tool call, mandates logging of destructive actions, and blocks unapproved external communication. ## Core Features & Use Cases - Tool Classification Policies: Categorizes tool calls into internal, research, external communication, and configuration types, each with a defined policy (free use, draft-first, or explicit approval). - Mandatory Logging & Failsafe: Requires every state-modifying tool call to be logged in a session log, and forces agents to report tool failures instead of silently retrying or inventing data. - Draft-First Communication: Distinguishes operational self-delivery to the executive from third-party communication, requiring draft approval before any external message is sent. - Isolated Config Validation: Provides a reference pattern using PATH shims and fake binaries to validate configuration mutations (e.g., hermes config set) without touching the main runtime. - Use Case: When an agent needs to push code, send an email, or install a package, this Skill forces it to classify the action, generate a draft for approval, log the operation, and refuse blacklisted commands like rm -rf. ## Quick Start Ask the agent to check how many skills are installed and the last modification date of FEATURES.md, and verify it uses real tools with logged, verifiable output instead of guessing.
