Superglue exposes a single tool 'search_documentation' with acceptable naming and schema structure. However, the tool lacks critical context for production use: no documented output schema, no error handling guidance, and no explanation of what 'documentation' refers to (API docs? Internal wiki? User manual?). The description is 257 characters (within baseline 194±158), but uses vague language ('relevant documentation excerpts') without explaining what data structure is returned or how the LLM should interpret results. The input parameter 'query' is well-typed (string) and described, meeting basic parameter standards, but the tool is fundamentally underspecified for reliable agent composition. The tool appears to be registered directly in llm-tools.ts, but without visible output schema declaration in the source, the tool cannot guide downstream planning or chaining.

• Document the output schema explicitly. Add a section to the tool description or a separate schema definition stating: 'Returns an object with: {results: [{title: string, excerpt: string, source_url: string, relevance_score: number}], total_results: number, has_more: boolean}' (adjust to actual schema). Include pagination parameters (limit, offset) if the tool supports them.
• Clarify what documentation system is being searched. Specify in the tool description: 'Searches the Superglue API documentation, including endpoint definitions, authentication patterns, and rate limit policies. For scraping external documentation, use the separate scrape_external_documentation tool.' This prevents ambiguity and tells LLMs when to use this tool vs. others.
• Add error recovery guidance. Extend the description: 'If no results are returned, try broadening the query or searching for related terms (e.g., search for "auth" instead of "OAuth 2.0 flow"). If documentation has not been indexed, contact the admin to trigger an index refresh.'
• Replace example values with formal constraints. Instead of 'e.g., authentication, batch processing, rate limits' in the parameter description, either (a) provide an enum of valid search categories, or (b) specify the search algorithm ('Full-text search across API documentation; results ranked by relevance'). This prevents LLMs from assuming those are the only valid searches.
• Add pagination and limit parameters to the tool if large result sets are possible. If the tool returns all matches, add a 'limit' parameter (default 10, max 100) to cap results and reduce token cost. Document: 'Returns up to {limit} results; if total_results > limit, use offset to fetch additional pages.'

Inferred effective spec: 2026-07-28+.

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