Three tools with minimal parameter descriptions and no output schemas documented. scan_project and project_memory have empty input schemas ({}), violating schema requirements. explain_code has proper Zod schema with enum constraint on language parameter, but descriptions lack actionable detail. No error handling guidance, no recovery paths. Tool names follow verb_noun pattern (scan_, explain_) but descriptions are generic and under 200 chars. Missing parameter-level descriptions for explain_code's 'path' and 'language' fields in the actual tool registration, only visible in Zod definition, not in inputSchema passed to registerTool.

• Add explicit input schemas to scan_project and project_memory. Even if they take no parameters, define inputSchema as a proper JSON Schema object (e.g., {type: 'object', properties: {}, required: []}) rather than empty {}.
• Document output schemas for all three tools. Specify what fields scan_project returns (e.g., languages: string[], stack: string[], layers: object), what memoryBlock structure looks like, and what markdown structure explain_code produces.
• Expand tool descriptions to 100-150 chars with WHEN/WHY context. Example: 'Scan the current project to build a map of languages, stack, and architecture. Call this first to understand the codebase structure before explaining specific files.'
• Move Zod parameter descriptions into the inputSchema object passed to registerTool(). Use JSON Schema 'description' fields so LLMs see parameter guidance at registration time, not buried in implementation code.
• Add error recovery guidance. When explain_code fails due to missing API key, return: 'No API key configured. Run `vexi` in a terminal to set up, then retry.' This tells the LLM the exact next step.