Three tools with complete schemas and descriptions, but significant gaps in parameter documentation and error handling. Tool names are action-oriented (transcribe, speak, health) and descriptions are present (avg 150 chars). However, many parameters lack descriptions or constraints. The 'speak' tool has good parameter coverage, but 'transcribe' and 'health' are minimal. No output schemas documented. Error handling is absent, no recovery guidance, no categorization of retryable vs fatal errors. No tool annotations (readOnlyHint/destructiveHint). Circuit breaker logic is implemented server-side but not exposed to the LLM.

• Document output schemas for all three tools. E.g., transcribe returns {text: string, language: string, confidence: number}; speak returns {file_path: string, duration_ms: number}; health returns {whisper_healthy: bool, kokoro_healthy: bool, details: object}.
• Add enum constraints for 'voice' parameter in speak: {enum: ['af_sky', 'af_bella', 'am_adam', 'zf_xiaobei', 'zf_xiaoni', 'zm_yunxi', 'jf_alpha', 'jf_gongitsune']}. Remove example list from description.
• Add min/max bounds to numeric parameters: speed 0.5 - 2.0, leading_pad_ms 0 - 5000, trailing_pad_ms 0 - 5000. Document in parameter descriptions.
• Implement error recovery guidance. E.g., if transcribe fails with timeout, suggest 'Whisper service may be down. Call health() to check endpoint status, or increase TRANSCRIBE_TIMEOUT.'
• Add tool annotations: transcribe with readOnlyHint=true; speak with destructiveHint=true (creates files). health with readOnlyHint=true.