Three tools with complete input schemas and descriptions. search_memory has sophisticated parameter set (11 params) with good descriptions; get_document and gateway_health are simpler. All tools have verb-noun naming and non-empty descriptions (194-200 chars, within baseline 34-392 range). However, output schemas are not documented in the source code, LLMs cannot see what fields to expect from responses. No error handling guidance visible. Parameters lack enum constraints where applicable (e.g., 'tier' accepts 'episodic|semantic|procedural' but not declared as enum; 'group_by' accepts 'chunk|document' but not constrained). No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite all three being read-only operations.

• Document output schemas for all three tools. Specify what fields search_memory returns (e.g., results array with chunk_id, score, text, document_id, workspace), what get_document returns (content, doc_id, char_count), and what gateway_health returns (status, vector_layer_ok, lexical_layer_ok, timestamp).
• Convert 'tier' and 'group_by' parameters to JSON Schema enums instead of free-form strings. This prevents hallucinated values and makes the tool self-documenting.
• Add readOnlyHint: true to all three tool schemas to signal to agents these are safe, idempotent, and can be retried without side effects.
• Add error handling guidance to each tool description. E.g., search_memory: 'If no results found, try broadening the query or lowering min_vector_similarity. If workspace not found, call with workspace=null to search all workspaces.'
• Add min/max constraints to numeric parameters in search_memory schema: top_k (1-25), max_token_budget (100-100000), max_context_chars (500-10000), vector_weight (0.1-10.0), lexical_weight (0.1-10.0), min_vector_similarity (0.0-1.0).