Taifeng is a pre-alpha agent microkernel (v2026.9.17.23) with 10 tools spanning financial data, user input collection, skill orchestration, and script execution. Critical gaps: 7 of 10 tools lack visible input schemas in the provided source; descriptions are present but often terse (Chinese, 10-30 chars); no documented output schemas; no error handling guidance; no parameter type constraints or validation rules visible. The codebase shows sophisticated audit/journal infrastructure (audit_tool.py) but tool definitions themselves are underspecified for LLM consumption. Tools like `run_script` (IRREVERSIBLE risk) and `kill_skill` (DESTRUCTIVE) lack confirmation patterns or permission gates. Composition is reasonable (spawn_skill + await_skills + join_skill form a coherent async pattern), but parameter documentation is minimal.

• Add explicit JSON Schema input definitions for all 10 tools. Include type, required fields, and constraints (e.g., symbol enum for lookup_stock_price, amount min/max for convert_currency).
• Expand descriptions to 50-200 chars in English. For each tool, state: (1) What it does, (2) When to call it vs. similar tools, (3) What it returns. Example: 'spawn_skill: Asynchronously start a child skill by name. Returns skill_id for use with await_skills/join_skill. Use when parallelizing independent tasks.'
• Document output schemas for all tools. Specify field names, types, and which fields are required for chaining (e.g., search_skills must return skill_id, confidence; call_skill must return result, status).
• Add confirmation/dry-run patterns to destructive tools (kill_skill, run_script). Implement a 'confirm' parameter or separate 'preview_script' tool to let agents verify before executing irreversible operations.
• Add error classification and recovery guidance to tool descriptions. E.g., 'run_script: Returns {status, output, error}. On timeout: retryable. On permission denied: fatal. On script error: return stderr for LLM to debug.'