Critical gaps across all 13 tools. Descriptions exist but are often vague or lack actionable guidance for LLM selection. No parameter descriptions at all (all schemas show empty properties). No documented output schemas. Error handling and recovery guidance absent. Tool names follow verb_noun convention (positive), but lack of schema detail and parameter guidance makes this server unsuitable for production agent use.

• Define input schemas for all 13 tools. At minimum, add type, description, and constraints for every parameter. E.g., chief_audit should have lane (string, optional), include_orphans (boolean, optional), limit (number, 1-20, optional).
• Document output schemas for each tool. Specify the structure of responses: field names, types, and whether fields are always present or conditional. Use JSON Schema format.
• Expand descriptions to 50-200 characters with WHEN/WHY/WHAT guidance. E.g., 'Dispatch a pending task to an external worker. Call this after plan_tasks() to start execution. Returns immediately; check status with get_worker_status().'
• Add parameter descriptions to all tools. For chief_config_help, document what 'provider' filters (e.g., 'Optional provider name: openai, anthropic, etc.'). For chief_audit, explain limit range and orphan inclusion behavior.
• Implement error handling patterns: categorize errors as retryable vs. fatal, include the invalid value in error messages, and provide recovery guidance. E.g., 'Task not found: task_id must match an entry in .chief/tasks.json. Call chief_doctor() to list available tasks.'