Two tools with partial schema visibility. compat.chat has a detailed input schema with 11 parameters (base_url, api_key, model, prompt, timeout_ms, connect_timeout_ms, sampling, response_format, tools, streaming, retry, guardrails, cache, budget) but lacks output schema documentation. compat.embedding has minimal schema (4 params) and no output schema. Both descriptions are present but generic (194-char baseline suggests these are adequate length). Critical gaps: no output schemas documented, no error handling guidance, no parameter constraints (enums, ranges), no idempotent/destructive hints. Tool names lack action verbs (should be 'execute_chat_completion' or 'generate_embedding'). Parameters like 'sampling', 'response_format', 'tools' are documented but lack type details and constraints in descriptions.

• Rename tools to action-verb form: 'execute_chat_completion' and 'generate_embedding'. This makes intent explicit and helps LLMs select the right tool.
• Document output schemas for both tools. For compat.chat, specify: {choices: [{message: {role, content}, finish_reason}, ...], usage: {prompt_tokens, completion_tokens, total_tokens}}. For compat.embedding, specify: {data: [{embedding: number[], index}], model, usage}.
• Add parameter constraints in descriptions. E.g., 'sampling: object with temperature (0.0 - 2.0), top_p (0.0 - 1.0), top_k (1 - 100)' and 'response_format: enum of text|json|json_schema'.
• Add error handling guidance. E.g., 'If base_url is unreachable, verify the endpoint is running and accessible. Retryable: connection timeout, rate limit. Non-retryable: invalid API key, malformed request.'
• Declare tool annotations: compat.chat should include destructiveHint=true (it sends requests to external LLM endpoints with side effects). compat.embedding should include readOnlyHint=true (read-only operation).

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry