Three tools with partial schema definitions and moderate descriptions. orchestrate has a well-structured input schema with type constraints (maxIterations 1-50, timeout 1000-600000ms) and a 95-char description. research_catalog_review and registry_import have enum-constrained parameters and reasonable descriptions (80-90 chars), but output schemas are not visible in source code, only inferred from mock responses. No tool descriptions explain when to use them vs alternatives, prerequisites, or error recovery paths. Parameter descriptions are present but generic (e.g., 'Optional context data' lacks detail on expected structure). No tool declares permissions, audit requirements, or idempotency guarantees. Error handling is absent from tool definitions, only mock responses show error cases. Overall: definitions are functional but lack LLM-optimization depth.

• Document output schemas explicitly in tool definitions. For orchestrate, specify: {task, status, output?, subtasks[], iterations, durationMs, error?}. For research_catalog_review, specify per-action response structure (list returns {pending[], count}, approve returns {approved}, etc.).
• Expand parameter descriptions with constraints and examples. Change 'Optional context data (record of unknown values)' to 'Optional context object with string keys and any values (e.g., {"domain": "ml", "budget": 50000}). Used to guide task decomposition.'
• Add error recovery guidance to tool descriptions. E.g., orchestrate: 'If status=timeout, increase timeout parameter or simplify task. If status=error, check agent state via list_experts tool.'
• Split research_catalog_review into four single-purpose tools: list_catalog_references(topic?), approve_reference(identifier), dismiss_reference(identifier), flush_catalog(). Each tool does one thing; LLM composes them.
• Declare permissions and idempotency. Add to registry_import description: 'Requires write:registry permission. Idempotent: repeated calls with same provider+modelId return same result without duplicate entries.'