model

Overview ¶

Package model provides model selection, cost tracking, and escalation chains.

This package helps choose the appropriate LLM model for different task types and tracks token usage for cost estimation. It is designed to be task-type agnostic - define your own task types and map them to model tiers.

Model Selection ¶

selector := model.NewSelector(
    model.WithThinkingModel(model.ModelOpus),
    model.WithDefaultModel(model.ModelSonnet),
    model.WithFastModel(model.ModelHaiku),
)
m := selector.SelectForTier(model.TierThinking)

Cost Tracking ¶

tracker := model.NewCostTracker()
tracker.Record(model.ModelSonnet, 1000, 500)  // input, output tokens
cost := tracker.EstimatedCost()

Model Escalation ¶

Escalation chains define how to retry with more capable models:

state := model.NewEscalationState(&model.DefaultEscalation, model.ModelSonnet)
for !state.Exhausted() {
    resp, err := tryRequest(state.CurrentModel)
    if err == nil {
        return resp, nil
    }
    state.RecordFailure(err)  // May escalate to next model
}

Package model provides model selection, escalation chains, and cost tracking.

The package implements a tiered model selection strategy:

• Thinking tier (opus): Complex reasoning, architecture, risk assessment
• Default tier (sonnet): Implementation, review, general tasks
• Fast tier (haiku): Search, simple transforms, high-volume

This package is designed to be task-type agnostic. Higher-level packages (like devflow) can define their own task types and map them to model tiers.