Single tool with critical definition gaps. Tool name 'image_generation' lacks action verb prefix (should be 'generate_image'). Description is 600+ chars in Chinese with embedded parameter docs, verbose and not LLM-optimized. Input schema present but minimal: only 'image_prompt' (string) with Chinese description. No output schema documented, tool returns dict with 'path' or 'error' keys, but structure is undocumented. Error handling returns raw dicts with 'error' and 'code' fields; no recovery guidance. No parameter constraints (min/max length, format validation). Tool performs external API call with 30s timeout but no timeout error handling documented. No idempotency guarantees stated. Security: API endpoint is environment-injected (good), but no rate limiting, no audit logging, no permission checks.

• Rename tool to 'generate_image' (verb_noun convention). Update description to ~100 chars in English: 'Generate an image from an English text prompt and save to local file. Returns file path on success.'
• Document output schema in tool docstring: 'Returns: {"path": str} on success, {"error": str, "code": int, "retryable": bool} on failure.'
• Add parameter constraints: image_prompt min 10 chars, max 1000 chars. Add description: 'English text prompt (10-1000 chars). Be specific and vivid to convey article mood/theme.'
• Enhance error handling: catch timeout separately, return {"error": "API timeout after 30s", "code": 504, "retryable": true}. For missing event_id, return {"error": "API did not return event_id", "code": 502, "retryable": true}.
• Add audit logging: log caller identity, prompt (first 50 chars), result status, and latency to stderr. Include rate-limit checks to prevent agent loops.