Intelligence routing

Hava uses the lowest-cost, fastest intelligence level that can reliably complete a request, and can escalate after tool observations.

Levels

Fast handles simple conversation, direct recall, and lightweight answers. Standard handles ordinary tool use and moderate analysis. Deep is reserved for difficult code, planning, research, risk, or multi-step reasoning.

Cost principles

Routing is separate from tool selection. Hava avoids unnecessary retrieval, excessive memory context, repeated tool calls, and oversized tool results. Quality is the floor; latency and service cost are optimization targets.

Tracing

Development traces record intelligence level, selected model, reason, tool-call counts, runtime escalation, and total latency. They do not contain private prompt or response content.

Evals

Known regressions are kept as permanent evaluation cases so routing and cost changes can be checked against behavior quality.