Gobble has 18 tools with mixed quality. Naming is generally good (verb-based: list, search, read, get, embed, summarize, spawn, abort). However, critical gaps exist: (1) Many tool descriptions lack specificity about when to use each tool vs. alternatives (e.g., search_books vs. search_contents vs. search_episodes are semantically similar but their distinctions are unclear). (2) Parameter descriptions are minimal or generic (e.g., 'Book identifier to summarize' repeated across tools). (3) Return type documentation is missing or incomplete for most tools, only inferred from code inspection. (4) Error handling guidance is absent, no recovery suggestions in descriptions. (5) The orchestrator tools (opencode_spawn, opencode_prompt, opencode_abort, opencode_get, opencode_list) appear to invoke an external 'OpenCode' system, but their schemas and error semantics are not documented. (6) Parameter validation rules are not described (e.g., what are valid book_ids? what format are session_ids?). Conservative scoring reflects these gaps against production baselines.

• Document return types for every tool. In tool registration (if using fastmcp decorators), add 'Returns' sections describing the structure of successful responses and error responses.
• Disambiguate search tools by adding contextual guidance: 'Use search_books for title/author metadata queries. Use search_contents for semantic similarity over book text. Use search_episodes for keyword matching in transcripts.' Include this in each tool's description.
• Expand parameter descriptions with format hints: e.g., 'book_id: Alphanumeric identifier (see list_books to discover valid IDs)' and 'session_id: UUID format returned by opencode_spawn'.
• Add recovery guidance to error-prone tools. E.g., book_information: 'If book not found, try search_books(query=...) to find the correct book_id.' get_transcript: 'If filename format is unclear, call list_episodes(show_name=...) to see available transcripts.'
• Clarify the OpenCode execution model: 'opencode_spawn queues a task asynchronously. Results are polled via opencode_get() or retrieved on the next agent turn. This is non-blocking, the agent can continue other work while the task runs in the background.'
• Mark side-effecting tools explicitly: 'embed_book: Modifies state, adds book to vector store. This is safe to retry (idempotent). get_youtube_transcript: Modifies state, downloads and stores transcript. opencode_spawn: Modifies state, creates a new background session.'
• Add max_results and pagination to search tools: 'search_contents returns up to 10 results (configurable). Use the returned next_cursor to retrieve additional pages if available.' Update the schema to include limit/offset or cursor parameters.

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

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