This server has three explicitly registered tools with schemas and descriptions. Tool naming follows verb_noun convention (generate_image, backend_status, browser_visibility), which is correct. However, parameter descriptions vary in quality and completeness. The generate_image tool has comprehensive schema documentation with proper constraints (min/max, enums), but lacks clear guidance on behavior (e.g., what happens on fallback from api to chatgpt-web). Error handling returns basic error text but lacks recovery guidance. Output schema is documented via structuredContent in code, but this is implicit rather than explicitly declared in the tool definition. The server does not implement tool annotations (readOnlyHint, destructiveHint) despite having tools with clear READ-ONLY (backend_status) and WRITE (generate_image, browser_visibility) semantics.

• Add tool annotations to all three tools: set readOnlyHint=true on backend_status; set destructiveHint=true on generate_image and browser_visibility. This enables agents to reason about side effects and ask for confirmation when appropriate.
• Enhance error messages with recovery guidance. Instead of returning raw error text, categorize errors (retryable, user-fixable, fatal) and suggest next steps: 'API authentication failed: Check GPT_IMAGE_API_KEY environment variable.' or 'ChatGPT session expired: Call browser_visibility action=show to re-authenticate.'
• Document the response schema formally in each tool's description or via a JSON Schema output declaration. Specify what fields the agent can expect: {status: 'success'|'partial'|'failed', backend: string, images: array, metadata: object, ...}.
• Constrain the 'size' parameter to an enum. Replace free-form string with enum of OpenAI-supported sizes (e.g., '256x256', '512x512', '1024x1024', 'auto') so LLMs cannot invent invalid values.
• Document fallback behavior explicitly in generate_image description: 'If the selected backend fails, automatically falls back to the next available backend (api → chatgpt-web). Set backend to a specific name to prevent fallback and get an error instead.'
• Add idempotency note: 'Image generation is non-deterministic. Identical prompts may produce different images on retry. For reproducibility, capture the image URLs and do not re-generate unless the user explicitly requests new variations.'
• Clarify browser_visibility behavior when browser is not running: 'If action=show and start_browser=false and no browser is open, returns an error. Set start_browser=true to launch the browser and retry.'