Stretch exposes 3 tools with partial schema coverage. health and list_tools have minimal descriptions (under 100 chars). dispatch_intent has a well-structured input schema with required fields and descriptions, but lacks output schema documentation. No tool has documented return types or error recovery guidance. Parameter descriptions exist for dispatch_intent but are generic ('Intent payload containing kind-specific data'). No tool declares idempotency, destructiveness, or read-only hints. Error handling is absent, no guidance on retryability or recovery paths.

• Add explicit JSON Schema output definitions for all three tools. Document what fields health returns (kernel_name, version, status), what list_tools returns (array of tool objects with name/description/schema), and what dispatch_intent returns (result object with status, data, or error details).
• Convert dispatch_intent's 'kind' parameter from free-form string to an enum: ['tool_call', 'policy_query', 'agent_delegate', 'resource_fetch', 'custom']. Add similar constraints to 'caller_id' (e.g., UUID format) and 'tenant' (optional, alphanumeric).
• Expand health and list_tools descriptions from ~50 chars to 100 - 200 chars. Include WHEN to call them (e.g., 'Call health to verify kernel availability before dispatching intents') and WHAT they return (e.g., 'Returns kernel name, version, and operational status').
• Add error handling guidance to dispatch_intent: document which errors are retryable (network timeouts, rate limits) vs. user-fixable (invalid kind, missing caller_id) vs. fatal (policy violation). Return structured error responses with actionable messages.
• Declare tool safety hints: mark health and list_tools with readOnlyHint=true; mark dispatch_intent with destructiveHint=true if it can modify state (e.g., agent_delegate kind). This helps agents reason about side effects.

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

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