Single tool with strong naming (verb_noun: transcribe_audio), comprehensive parameter descriptions, and detailed tool description (340 chars). Input schema is fully defined with types, enums, and defaults. Tool annotations present (readOnlyHint, idempotentHint, destructiveHint, openWorldHint). Output schema not explicitly documented in visible code, function returns `str` with no structured schema definition. Error handling is present but returns JSON strings rather than structured error objects. Parameter descriptions are detailed and actionable (e.g., 'ISO-639-1 hint such as ar or en'). Mutually exclusive parameters (file_path, audio_url, audio_base64) are documented. No critical security issues, API key injected via environment, not parameters.

• Document output schema explicitly: define a TypedDict or Pydantic model for success (transcript text, model used, duration) vs error (error code, message, recovery hint). Return structured objects, not JSON strings.
• Add a `max_chars` parameter (default 25000) to let agents control truncation behavior. Document the default limit in the tool description.
• Enhance error responses with structured error codes (e.g., 'auth_failed', 'rate_limited', 'unsupported_format') alongside messages. This enables LLMs to classify errors and decide retry vs user-fixable vs fatal.
• Add a dependency hint in the language parameter description: 'Leave unset for auto-detection. Set only if auto-detection produces wrong results (e.g., Arabic detected as Farsi).'
• Consider adding a `dry_run` parameter for validation: 'If true, validate the file and model without calling OpenAI. Useful for checking format before transcription.'

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

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