The server defines 4 well-structured search tools with consistent parameter sets and clear semantic distinctions. All tools have descriptions (50-100 chars), proper JSON input schemas with typed parameters, and descriptive parameter labels. However, descriptions lack depth on use-case differentiation, output schemas are not documented, and error handling is not specified. Tool naming follows verb_noun pattern correctly (search_*), but descriptions do not explain WHEN to use one search tool vs. another or what recovery looks like on failure.

• Document return type schema for all tools. Specify: results (array of {message: string, score: float, timestamp: string, workspace?: string, source?: string}), total_count (int), has_more (bool), next_offset (int). Include example JSON.
• Add 'Use this tool to...' statements that differentiate CLI vs IDE vs workspace vs global searches. Example for search_cli_history: 'Use this when the user mentions a CLI session specifically, or when searching for terminal-based workflow context. Falls back to search_global_history if uncertain.'
• Add error handling guidance to each tool description. Example: 'Returns empty results if index is not yet built (first call may take 30s); subsequent calls are fast. If query is empty, returns an error asking for at least 1 character.'
• Clarify 'threshold' semantics: 'Minimum cosine similarity (0=unrelated, 1=identical). Default 0.2 finds loosely related messages; raise to 0.5+ for strict relevance. Lower to 0.1 to cast a wider net and find tangential context.'
• Clarify 'context_size' behavior: 'Number of messages to return before and after each matched message. Symmetric (3 means 3 before, 3 after, for 7 total per match). If fewer messages exist, returns what is available.'
• Add parameter constraints inline. For max_results, add '(1-100)' to the description. For context_size, add '(0-10, default 3)'.
• Document pagination: 'offset: Skip this many results. Use this with max_results to paginate. Combine with total_count in response to detect end of results.'