Three tools with OpenAI-compatible chat/completion APIs. All have descriptions (194 - 250 chars, within baseline 34 - 392 range) and input schemas with typed parameters. No output schemas documented. Tool names use verb_noun pattern (chat.send, completions.create, models.list) but are dot-separated rather than underscore-separated, deviating from baseline conventions. Error handling descriptions mention in-band HTTP-200 errors but provide no recovery guidance for LLMs. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite all being read-only operations.

• Add parameter descriptions to all three tools. For 'model', specify format (e.g., 'Model identifier string, e.g. gpt-4-turbo'). For 'messages', describe structure (e.g., 'Array of {role, content} objects'). For 'stream', clarify behavior ('true returns SSE stream; false returns single JSON response').
• Document output schemas explicitly. Define the structure of chat.completion (fields: id, choices, usage, etc.), text_completion (fields: text, finish_reason, etc.), and model catalog (fields: data array with model objects).
• Add readOnlyHint: true to all three tools in the tool registration to signal they are safe to retry and have no side effects.
• Enhance error descriptions with recovery guidance. E.g., 'If finish_reason is "error", the model failed; retry with a shorter prompt or different model parameter.'
• Standardize tool names to underscore convention: chat_send, completions_create, models_list for consistency with production baselines.