Strong foundation with clear naming, comprehensive descriptions, and well-structured schemas. All 6 tools follow verb_noun pattern (at_search, at_get_text, at_case_search, at_get_case_text, at_list_collections, at_coverage). Descriptions are detailed (100-250 chars) and explain WHAT, WHEN, and dependencies. Input schemas use proper JSON Schema with types and enums. Tool annotations (readOnlyHint, idempotentHint) present. Error handling includes structured codes with recovery guidance. Main gaps: output schemas not explicitly documented in source; some parameter descriptions could be more prescriptive about format/constraints; no batch operations despite search-then-fetch pattern.

• Document output schemas explicitly in tool registration. Add a 'returns' field or inline schema showing SearchResult fields (hits[], total_count, has_next_page) and CaseText fields (text, format, source_url, citation). This enables LLMs to plan multi-step queries.
• Add length and character constraints to free-text parameters. E.g. 'suchworte: Free text search (1-500 chars, alphanumeric + spaces + common punctuation)' and 'titel: Title search (1-200 chars)'. Include an example: 'e.g., "Datenschutz" or "DSGVO"'.
• Return pagination metadata in search results. Include total_count (total hits available), has_next_page (boolean), and next_page_number (if applicable). This lets agents know when to stop iterating.
• Add cross-parameter validation hints. Document that at_get_text's content_url must come from at_search results and must be a ris.bka.gv.at URL. Similarly for at_get_case_text. This prevents silent mismatches.
• Ensure all error paths return structured ToolError with a [code] prefix. Audit exception handlers to confirm no raw HTTPStatusError, RisError, or generic Exception escapes to the LLM. Add a test that verifies all error cases produce valid codes.