Single tool 'runChatCompletion' has a well-structured input schema with proper types and descriptions for messages and options. However, the tool lacks critical production-grade patterns: no output schema documentation, no error handling guidance, no idempotency markers, and no security/permission declarations. The description is adequate (67 chars) but generic. Parameters are typed but lack constraints (e.g., temperature range, model enum). No pagination, no recovery guidance for API failures. This is a basic wrapper around OpenAI with minimal agent-optimization.

• Document the output schema explicitly: 'Returns {content: string, totalTokens: number}. content is the assistant's response text; totalTokens is the total tokens consumed (for cost tracking).'
• Add parameter constraints: temperature (0 - 2, default 0.7), model (enum: gpt-4o, gpt-4-turbo, gpt-3.5-turbo, or custom via env), messages (non-empty array).
• Implement error classification: distinguish retryable errors (timeout, rate_limit) from user-fixable (invalid_model, empty_messages) from fatal (auth_failed). Return actionable guidance: 'Rate limited. Retry after 60s.' or 'Invalid model. Supported: gpt-4o, gpt-4-turbo.'
• Add tool annotations: mark as destructiveHint=true (incurs API costs) and idempotentHint=false (repeated calls with same input produce different outputs due to temperature/randomness).
• Inject LLM_API_KEY via secure server-side configuration (environment at startup, not exposed as parameter). Document that the tool requires OPENAI_API_KEY to be set before server initialization.