Context counting is a preflight occupancy seam. It counts the complete provider-neutral inference.Request before invocation so admission policy can compare input tokens with a model context limit. It is not a response-usage wrapper and it does not count generated output.

Purpose

type ContextCount struct {
	Model       model.ModelKey
	InputTokens content.TokenCount
	Quality     CountQuality
}

type ContextCounter interface {
	CountContext(context.Context, inference.Request) (ContextCount, error)
	CounterCapability() CounterCapability
}

InputTokens is the occupancy of the complete request, including system text, conversation, tools, structured-output schema, and dialect-specific encoding. Quality distinguishes provider-exact, local-exact, and heuristic estimates.

Harness consumes this preflight result for context limits and automatic compaction thresholds.

Flow

%%{init: {"theme":"base","themeVariables":{"background":"#111827","primaryColor":"#1f2937","primaryTextColor":"#f9fafb","primaryBorderColor":"#60a5fa","lineColor":"#94a3b8","secondaryColor":"#172033","tertiaryColor":"#0f172a","fontFamily":"Inter, ui-sans-serif, system-ui"}}}%%
flowchart LR
    Q["complete Request"] --> C["ContextCounter"]
    C --> M["ContextCount: model, input, quality"]
    M --> A["admission policy"]
    A --> I["invoke or reject"]
    U["response Usage"] -. separate postflight data .-> R["billing/telemetry"]

Source and proof

Run go test ./contextcount.