Single tool with clear naming (search_web) and reasonable description (194 chars). Input schema present with typed parameters and descriptions. However, output schema is undocumented (returns string, not structured), error handling is generic ('Error searching...' string), and no guidance on recovery or retry logic. Missing tool annotations (readOnlyHint despite READ_ONLY risk classification). Descriptions are adequate but lack dependency hints or format constraints. No pagination support despite potential for large result sets.

• Define output schema as structured object: {content: string, citations: [{url: string, title?: string}], search_query: string, model_used: string}. Document in tool docstring.
• Replace generic error handling with categorized responses: 'Retryable: API rate limit exceeded. Wait 60s and retry.' or 'User-fixable: Query too long (>500 chars). Shorten and retry.'
• Add readOnlyHint: true to tool schema to signal to clients this is a safe, cacheable read operation.
• Convert recency parameter to enum: {type: 'string', enum: ['day', 'week', 'month', 'year'], default: 'month'} instead of prose description.
• Add max_results parameter (default 20, range 1-50) and document in description: 'Limits citations returned to prevent context window exhaustion.'

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

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