Two tools with basic schemas and descriptions, but significant gaps in parameter documentation, output schema clarity, and error handling guidance. create_thought has a context parameter (object type) with no field constraints or examples. validate_thought lacks output schema documentation. Neither tool describes what happens on error or how to recover. Descriptions are adequate (100-150 chars) but lack WHEN/WHY guidance. No tool annotations (readOnlyHint, destructiveHint). Parameter descriptions are present but generic, 'Metadata about the environment' does not explain what fields context should contain or what values are valid.

• Document the full output schema for create_thought as a JSON object with fields: {envelope: {id, timestamp}, cognition: {intent, layer, confidence, context}, payload}. Include type and description for each field.
• Replace context parameter description with: 'Metadata dict with required keys: domain (string: production|staging|test), agent_id (string), and optional keys: session_id, trace_id. Example: {"domain": "production", "agent_id": "agent-001"}'.
• Add error handling guidance: 'On validation error, returns error message. Retryable if JSON is malformed (fix syntax). Non-retryable if schema violation (check RFC-001 compliance).'
• Add tool annotations via FastMCP: @mcp.tool(readOnlyHint=True, idempotentHint=True) to both tools to signal they are safe to retry and do not modify state.
• Refactor validate_thought to return structured JSON: {"valid": bool, "errors": [string], "message": string} instead of plain text. Enables LLM to branch on valid/invalid without string parsing.