Single tool 'duo_review' has a well-crafted description (280+ chars) explaining what it does, when to use it, and what it returns. Input schema is complete with 7 parameters, all typed and described. Output schema (ReviewStructuredResult) is explicitly documented with typed fields. Error handling includes actionable recovery guidance (e.g., 'try search_users()'). However, tool naming lacks a clear action verb prefix (should be 'run_duo_review' or 'review_code'), and some parameter descriptions could be more prescriptive about constraints (e.g., timeoutMs range, model enum values). No tool annotations (readOnlyHint, destructiveHint, idempotentHint) present.

• Rename tool to 'run_duo_review' to follow verb_noun convention and signal action intent to LLMs.
• Add enum constraint to 'model' parameter: e.g., {"enum": ["gpt-4", "claude-3", "default"]} with description listing valid options.
• Add min/max bounds to 'timeoutMs': e.g., 'Override the run timeout (milliseconds, 1000 - 300000) for this call.'
• Declare tool annotations in MCP registration: readOnlyHint=true, idempotentHint=true to guide agent planning and retry logic.
• Document 'raw' field behavior: 'When Duo output cannot be parsed as JSON, the raw response is returned here with degraded=true. Agents should check degraded flag before relying on structured issues array.'