Four tools with complete JSON schemas and detailed descriptions in Chinese. Naming follows verb_noun pattern (voice_dispatch_task, voice_list_sessions, voice_query_session, voice_cancel_task). Descriptions are comprehensive (200-500+ chars) and explain WHEN to use each tool. However, descriptions are in Chinese, making them inaccessible to English-speaking LLMs and violating the principle that descriptions must be prompt-engineered for the model. Parameter descriptions are present but verbose. No explicit error handling guidance or recovery paths documented. Output schemas are not formally documented, only inferred from description text. Missing tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels (WRITE vs READ_ONLY).

• Translate all tool and parameter descriptions to English.
• Add tool annotations: mark voice_dispatch_task and voice_cancel_task with destructiveHint=true; mark voice_list_sessions and voice_query_session with readOnlyHint=true.
• Document output schemas formally. For voice_dispatch_task, specify: {task_id: string, status: 'queued'|'running'|'completed'|'failed', message: string}. For voice_query_session, specify the exact fields returned.
• Add error handling guidance. E.g., 'If cwd is not found, the tool will fall back to the default vococo repository. If session_id is invalid, the tool returns an error, try voice_list_sessions to find valid IDs.'
• Simplify parameter descriptions. Remove implementation details (worktree, git mechanics). Focus on: what value to pass, what it controls, and valid options (enums where applicable).