TaskSwarm defines 10 tools with basic schemas and descriptions in Chinese. All tools have input schemas with type definitions and descriptions present. However, descriptions are minimal (10-50 chars), lack actionable context for LLM selection, and omit prerequisites/side effects. No output schemas documented. Parameter descriptions are terse. Error handling exists but lacks recovery guidance. Tool naming follows verb_noun pattern (plan_create, task_claim) but some names are ambiguous (plan_get vs board; state is generic). No tool annotations (readOnlyHint, destructiveHint). Composition is reasonable, tools chain logically (plan_create → task_ready → task_claim → task_update). Security: workspace parameter allows path traversal; no explicit permission gates documented.

• Expand all tool descriptions to 50-200 chars. Example: 'plan_get' → 'Retrieve the full task tree with wave-based decomposition status. Returns all tasks, dependencies, and current state. Call this to see what work is ready for dispatch.'
• Document output schemas for all tools. Example: plan_get returns {tasks: [{id, title, status, dependencies, notes}], waves: [{id, tasks}], summary: {total, completed, blocked}}.
• Add tool annotations: mark plan_reset, task_claim, task_update, task_add, state as destructive=true; mark plan_get, task_ready, task_notes, board as readOnly=true.
• Rename 'board' to 'get_progress_board' and 'state' to 'manage_session_state' for clarity and consistency with verb_noun naming.
• Validate workspace parameter: reject absolute paths and paths containing '..'. Whitelist safe base directories or require workspace to be a relative subdirectory of cwd.