Built-in Tool Refactoring Guide

When to Use This Skill

Use this skill when:

• Refactoring existing built-in MCP tools in LibrAgent.
• Reviewing tools that cause AI context bloat or "cognitive overload".
• Consolidating redundant or overlapping tools (e.g., exportFile and exportZip).
• Fixing issues where internal callback tools are exposed to the AI agent.

Core Design Principles

1. Context Economy & Cognitive Load

AI agents operate within strict context limits. Every exposed tool consumes tokens and attention.

• Hide Internal Callbacks: Never expose internal system callbacks or UI triggers (e.g., resumeFromWait, getUserAnswer, circuitBreak) to the all_tools() export. The backend router (call_tool) can handle these without the AI needing to know they exist.
• Concise Descriptions: Keep tool descriptions brief. Avoid paragraphs of prerequisites or error-handling advice. The schema itself should document parameters.
• Progressive Disclosure: For discovery tools (like listTools), provide a compact summary (names only) when queried broadly. Show detailed schemas only when a specific, narrowed query is provided.

2. Consolidation & Auto-Switching

Reduce the sheer number of tools by combining highly similar operations into a single tool with smart, data-driven behavior.

• Merge Sibling Tools: Combine tools that achieve the exact same goal (e.g., merging exportFile and exportZip into a single export tool that accepts an array of paths).
• Data-Driven Logic: Let the backend determine the execution mode (e.g., if one file is passed -> export raw; if multiple files -> export as ZIP). The AI shouldn't have to make micro-decisions about file counts or formats.

3. Strict Synchronization & State Integrity

Data integrity must be guaranteed at the backend level.

• Synchronous Verification: Whenever a tool or configuration is created or updated, the system must verify its validity synchronously before committing it to the database.
• Backend as Source of Truth: Do not rely on the frontend to fire asynchronous validation events. If an entity is in the database, it must be verified and operational.
• No "Ghost" Entities: If a registered entity (like an external MCP server) lacks a cache due to a transient failure, it must still be visible in tool listings with an appropriate warning, rather than being silently filtered out.

4. Don't Repeat Yourself (DRY) in Integration

Built-in tools should leverage central architectures.

• Use Central Managers: Reuse central components like MCPServerManager or McpServerService::verify_config instead of writing raw tokio::process::Command or HTTP clients inside specific tool handlers.
• Single Source of Validation: Formulate validation logic in the core service layer and reuse it across UI endpoints, tool handlers, and agent routines.

5. Cross-Platform UX Considerations

• Process Stealth: On Windows, always apply the CREATE_NO_WINDOW flag (0x08000000) when spawning background or verification processes via std::os::windows::process::CommandExt to prevent distracting terminal flashes.

Refactoring Workflow

1. Audit tools.rs: Identify all tools exported via all_tools(). Remove any tool that is strictly an internal callback.
2. Review Descriptions: Condense verbose descriptions in MCPTool definitions.
3. Consolidate Sibling Tools: Check for tools with similar prefixes (e.g., read_foo, read_bar) and combine them if their core intent is identical.
4. Check Handlers: Ensure handlers use central service functions instead of duplicating connection/execution logic.
5. Verify State Updates: Ensure any DB mutations happen after successful synchronous verification.