The server declares 3 tools with consistent naming patterns (verb_noun: search_inside, search_items, get_item) and reasonable descriptions (97-121 chars each, within baseline 34-392 range). All tools have input schemas with type declarations. However, critical gaps exist: (1) Parameter descriptions in schemas are minimal, 'Words to search for inside the text' and 'Search query for the catalogue' provide context but lack constraints on length, format, or valid value ranges; (2) Output schemas are not visible in the source code provided, 'runSearchInside', 'runSearchItems', 'runGetItem' functions exist but their return types/structures are not shown, preventing verification that outputs are documented for LLM reasoning; (3) No evidence of error handling guidance, callers do not know which errors are retryable, user-fixable, or fatal; (4) Pagination is supported (limit, page params) but no documentation of total count, next_cursor, or max result limits in descriptions. Tool annotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint) are correctly declared at server level, but no per-tool error classification or recovery guidance is visible.

• Document and declare output schemas for all three tools. Include sample JSON responses showing field names, types, and nesting. At minimum, document: (1) which fields are returned, (2) which IDs (archive_id, item_id, reference) are included for chaining downstream calls, (3) whether results are paginated and how pagination is indicated (total_count, next_cursor, has_more). This is critical for LLM planning and per the rubric.
• Add explicit constraints to numeric parameters. For 'limit', state: 'Maximum number of results per archive (1 - 100, default 10).' For 'page', state: 'Page number, 1-indexed; omit if using cursor-based pagination.' For 'query', state: 'Search string, 1 - 255 characters; supports AND/OR operators' or similar.
• Document error cases and recovery paths for each tool. E.g., 'If a specific archive is unavailable, results from other archives are still returned (partial success). If all archives are unavailable, returns an error with message: "All archives unreachable. Retry after 30 seconds." (retryable error).' This guides LLM error handling per the recovery-guide pattern.
• Clarify pagination semantics in tool descriptions. State: 'Pagination is 1-indexed. Response includes total_count (total results across all archives) and results (per-page items). If page exceeds available results, an empty list is returned.' Without this, LLMs may misinterpret pagination boundaries.
• Add a 'capabilities' or 'limitations' section to the server instructions (buildInstructions function already exists). Document which archives support each tool: e.g., 'search_inside is supported by Internet Archive and data.bnf.fr, but not Library of Congress (which has no full-text API). search_items is supported by all three archives.' This prevents LLM confusion when a tool returns fewer results than expected.

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

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