OpenTraderWorld provides 7 tools with explicit schemas and descriptions visible in core/otw-core/src/agent/tools.rs. Naming follows verb-noun patterns (otw_catalog, otw_read, otw_write, memory_load, memory_save, memory_delete, load_skill), which is solid. Descriptions are present and moderately detailed (avg ~150 chars), explaining what each tool does and when to use it. However, several critical issues reduce the score: (1) The otw_catalog, otw_read, otw_write tools accept vague 'query' and 'body' parameters that are untyped JSON objects with minimal constraint documentation, LLMs cannot predict valid structures. (2) The memory and skill tools lack output schema documentation; what fields do they return? (3) No error handling guidance: what happens when a memory slug doesn't exist? When an API endpoint 404s? (4) No pagination support documented for list-like endpoints. (5) Tool descriptions lack dependency hints (e.g., 'call otw_catalog first to discover endpoints'). The server does NOT include sampled tool annotations (readOnlyHint, destructiveHint, idempotentHint) visible in the schema; Risk tags are only in source comments, not in the tool definitions themselves. Overall, the tools are functional but lack the polish and LLM-optimized guidance expected of production-grade agent tools.

• Add JSON Schema constraints for 'query' and 'body' parameters in otw_read and otw_write. For otw_read, document expected query key-value pairs (e.g., 'start_date' (ISO 8601), 'limit' (1-1000)). For otw_write, include examples of valid body structures per HTTP method.
• Document output schemas for all tools. Example for otw_read: 'Returns {status: string, data: object, pagination?: {limit, offset, total}}'. For memory_load: 'Returns {slug: string, markdown: string, created_at: string, updated_at: string}'.
• Change 'method' parameter in otw_write to an enum: type='string', enum=['POST', 'PUT', 'DELETE'].
• Add error handling guidance to each tool description. Example: 'If the endpoint returns 404, the path may be invalid, call otw_catalog to discover available endpoints.' For memory_load: 'If the slug is not found, returns error code MEMORY_NOT_FOUND. Available slugs can be discovered via otw_catalog.'
• Implement a memory_delete_with_confirmation tool that returns a confirmation token, then a finalize_deletion tool that uses the token. Or document that memory_delete requires a confirmation parameter (e.g., 'confirm_slug_deletion' = true) to prevent accidental deletes.
• Document pagination for otw_read. Add optional 'limit' (1-1000, default 50) and 'offset' (default 0) parameters. Describe response structure: 'Returns {data: [...], pagination: {limit, offset, total_count, has_more}}'.
• Replace example values in load_skill description with a constraint. Instead of 'e.g. Portfolio Rebalance, Signal Scout', add 'Call otw_catalog to discover available skill names' or provide an explicit enum of valid skill names.