OneLamp defines 4 tools with basic schemas and descriptions, but critical gaps prevent production use. All tools have descriptions (10-150 chars, within range) and input schemas with typed parameters. However, output schemas are completely undocumented, LLMs cannot plan downstream calls or extract results. Parameter descriptions are minimal (e.g., 'The tool/agent surface calling this tool' for source_surface). No error handling guidance, no recovery paths, no validation rules. Tools lack idempotency markers and destructive operation warnings. The server is HTTP-based (good), but tool definitions appear inferred from shell scripts rather than explicit MCP registration, capping confidence.

• Document output schemas for all tools. Specify what fields get_context returns (e.g., ranked_results: [{source, content, relevance_score}]), what save_context returns (e.g., {success: bool, saved_id: string}), and what resume_session returns (e.g., {session_id, context_summary, timestamp}).
• Enhance parameter descriptions with actionable constraints. For 'query', add: 'Short task description (10-200 chars). Examples: "set up TypeScript", "fix auth bug". Avoid secrets.' For 'source_surface', explain: 'Identifies the calling agent for context filtering. Must be one of: claude-code, codex, cursor.'
• Add error handling guidance. Document: 'If query is empty, returns {results: []}. If user has no saved context, returns {results: []}. If API is unavailable, returns {error: "Service temporarily unavailable. Retry in 30s."}.'
• Add tool annotations (destructiveHint, idempotentHint) to save_context and save_session. Mark them as destructive=true so agents know these calls have irreversible side effects. Mark get_context and resume_session as readOnlyHint=true.
• Publish explicit MCP tool registration in the server implementation (not just shell hooks). Show the actual tool definitions with complete schemas, descriptions, and error contracts in the codebase.