Four tools with basic schemas and descriptions, but significant gaps in parameter documentation and output structure. Tool names follow verb_noun convention (sjtu_*), but descriptions lack actionable context for LLM selection. Parameters have types but many lack descriptions (e.g., 'model' param in sjtu_text has description, but temperature/max_tokens lack guidance on ranges). No output schema documented. Error handling is minimal, SjtuApiError exists but provides no recovery guidance. Composition is reasonable (each tool has one responsibility), but tools lack idempotency guarantees and don't document prerequisites or when to use one over another.

• Document output schemas for all tools. E.g., sjtu_models should return {models: [{id: string, name: string, type: string}], total: int}. sjtu_text/vision/cheap_task should return {result: string, model_used: string, tokens_used: int} to enable chaining and cost tracking.
• Add range constraints to numeric parameters. Describe temperature as 'Sampling temperature (0.0 - 2.0, default 0.2)' and max_tokens as 'Maximum tokens in response (1 - 4096, optional)'.
• Expand tool descriptions to include selection criteria. E.g., sjtu_text: 'Use for text-only tasks. For image analysis, use sjtu_vision. For auto-routed tasks (summarize, classify, etc.), use sjtu_cheap_task.'
• Implement error categorization. Wrap SjtuApiError with context: 'API error (retryable): timeout after 30s. Retry with exponential backoff.' or 'API error (user-fixable): invalid model name. Call sjtu_models() to list available models.'
• Rename sjtu_cheap_task to sjtu_route_task or sjtu_execute_task to clarify that it intelligently routes work based on task type, not that it's a generic catch-all.