Three tools with basic descriptions and partial schemas. All tools have descriptions (10-100 chars, below the 194-char baseline), but parameter descriptions are sparse or missing. Input schemas present but lack type constraints and validation guidance. No output schemas documented. Error handling absent. Naming is verb-forward but descriptions lack WHEN/WHY context. No tool annotations (readOnlyHint, destructiveHint). Composition is reasonable (speak, listen, conversation_turn are distinct), but conversation_turn combines two concerns (listen + speak) without clear separation.

• Document output schemas for each tool. E.g., listen() returns {transcribed_text: string, confidence: number, language: string, duration_seconds: number}. speak() returns {status: 'success'|'error', audio_duration_seconds: number}.
• Expand parameter descriptions with ranges and units. E.g., 'max_duration: Maximum recording duration in seconds (1 - 300, default 30). Longer durations consume more memory.' Replace vague 'silence_threshold' with 'Audio level threshold for silence detection (0.0 - 1.0, where 0.01 = very quiet, 0.1 = moderate noise). Lower values detect silence sooner.'
• Add error handling with recovery guidance. Wrap model initialization, audio device access, and transcription in try-catch blocks. Return structured errors: {error: string, code: 'DEVICE_NOT_FOUND'|'MODEL_LOAD_FAILED'|'TRANSCRIPTION_TIMEOUT', suggestion: string}.
• Split conversation_turn into separate listen_and_respond tool OR document why both operations must be atomic. If atomic, clarify in description: 'Executes listen → process → speak as a single transaction. Use separate listen() and speak() tools for finer control.'
• Add tool annotations to tool registration. Mark all three as readOnlyHint=true (no state changes). Mark speak() and listen() as idempotentHint=true (safe to retry). This enables LLM retry logic.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry