Both tools lack input schemas entirely and have minimal descriptions. The 'codex' tool description is 67 chars but provides no actionable guidance on parameters, expected inputs, or outputs. 'codex-reply' description is 65 chars and equally vague. No parameter documentation visible. Tool definitions appear to be inferred from a verification script rather than explicit MCP server registration. No evidence of structured output schemas, error handling guidance, or composition patterns.

• Create explicit MCP server implementation with proper tool registration. Define each tool with name, description (50-200 chars), and complete inputSchema including parameter types and descriptions.
• Add comprehensive parameter documentation. For 'codex' tool, document all expected parameters (e.g., command, args, context) with types, constraints, and examples.
• Document output schemas for both tools. Specify what fields are returned, their types, and what data the LLM should expect for downstream tool chaining.
• Implement error handling patterns. Return structured errors with error codes, human-readable messages, and recovery guidance (e.g., 'Try calling search_codex() first').
• Add tool annotations (destructiveHint, readOnlyHint, idempotentHint) to clarify which operations modify state and which are safe to retry.