Single tool 'verify_turn' has a well-structured input schema with proper JSON Schema types and descriptions. The tool description is comprehensive (380+ chars) and explains the deterministic fact-checking behavior clearly. However, the schema lacks explicit output documentation, and parameter descriptions could be more prescriptive about format constraints. The tool follows a clear single-responsibility pattern (verify claims), but lacks error categorization guidance for the LLM.

• Add explicit output schema documenting the structure of verdicts: e.g. {"verdicts": [{"claim": string, "verdict": "supported|contradicted|refused", "evidence": string, "citations": [string]}]}
• Enhance 'repo' parameter description: 'Absolute path to repo root (must exist and contain .git). Example: /home/user/myproject. Relative paths will be rejected.'
• Document error cases: 'If repo is not found, returns error code REPO_NOT_FOUND. If .truth index is missing, returns UNINDEXED with guidance to run `truth index <repo>` first.'
• Add constraint on 'message' parameter: specify max length (e.g. 10000 chars) to prevent context explosion when LLM passes entire conversation history.
• Clarify 'claims' array semantics: 'Each claim should be a single, testable assertion (10-200 chars). Claims longer than 500 chars will be truncated. Vague claims (e.g. "I made changes") will be refused as unverifiable.'