Context Limits

Harness resolves a hard input-token limit from model metadata and an explicit output reservation plus safety margin. The Inference context-counting guide explains the exact, provider-backed, and conservative counters that supply request token counts. The pure limit helper is:

func ResolveContextLimits(
	model model.ModelKey,
	limits model.ContextLimits,
	reservedOutput, safetyMargin content.TokenCount,
) (ResolvedContextLimits, error)

type ResolvedContextLimits struct {
	ReservedOutput content.TokenCount
	RawInputLimit  content.TokenCount
	InputLimit     content.TokenCount
}

The algorithm validates the model key and limits first. It clamps reservedOutput to MaxOutputTokens when that metadata is present, then uses WindowTokens - reserved when a window is known. If MaxInputTokens is known, it takes the smaller nonzero value. Finally it subtracts safetyMargin. Unknown or nonpositive results are rejected rather than guessed.

Resolve a limit

resolved, err := loop.ResolveContextLimits(
	model.Key(),
	contextLimits, // model.ContextLimits obtained from the selected model metadata.
	policy.ReservedOutput,
	policy.SafetyMargin,
)
if err != nil {
	var unknown *loop.ContextLimitUnknownError
	if errors.As(err, &unknown) {
		// Do not send a request whose denominator is unknown.
	}
	return err
}
fmt.Printf("input=%d output=%d margin=%d\n",
	resolved.InputLimit, resolved.ReservedOutput,
	resolved.RawInputLimit-resolved.InputLimit)

InputLimit is the admission denominator. RawInputLimit is the limit before the safety margin, and ReservedOutput records the effective reservation. The loop’s context policy supplies those values; it does not silently invent a timeout, output reservation, or margin.

Occupancy

OccupancyBasisPoints(used, limit) returns floor(used * 10_000 / limit) as event.BasisPoints, clamps at the full-scale value when used >= limit, and returns *loop.OccupancyError for a zero denominator. The multiplication is checked with 128-bit arithmetic so a large token count cannot overflow the display calculation.

Typed errors

Use errors.As; do not parse model names or error strings. A context-limit failure is an admission refusal, not permission to truncate an arbitrary conversation.

Source and proof

• Context limit resolution and occupancy arithmetic
• Context limit and occupancy tests
• Admission-error measurement proof