Critical failures across all dimensions. Two tools (run_process, run_shell) have completely empty input schemas with no properties defined. Four tools lack parameter descriptions entirely. Descriptions are present but generic (10-60 chars, well below the 194-char baseline). No output schemas documented. No error handling guidance. The framework appears to be a C++ agent toolkit with tool definitions, but the MCP server implementation is not visible, only header files showing tool structs. Cannot verify actual MCP registration, serialization, or protocol compliance from provided code.

• Add complete input schemas to run_process and run_shell. Define required parameters (command, args, timeout_ms, etc.) with types and descriptions. Currently they accept no documented inputs.
• Expand descriptions for run_process and run_shell to 50-200 chars. Explain: What command does it execute? Does it support shell syntax or only direct process invocation? What are the timeout limits?
• Document output schemas for all tools. E.g., run_python_snippet should return {stdout: string, stderr: string, exit_code: int, timed_out: bool}. search_knowledge already returns JSON array in code but not documented in tool definition.
• Add error handling guidance to descriptions. E.g., 'If timeout_ms is exceeded, returns timed_out=true. Retry with longer timeout or split code into smaller snippets.' Guides LLM recovery.
• Clarify parameter constraints in descriptions. E.g., 'timeout_ms must be 1-300000 (5 minutes max)' for run_python_snippet. 'limit must be 1-100' for search_knowledge. Prevents invalid LLM inputs.