Module model_limits 

Each driver choice’s per-turn output cap, context window, input ceiling and input budget, resolved once at launch.

A transport: http capsule names a model and the driver that serves it. The driver’s bundled murmur.yaml declares, in its models: table, each model’s max_output_tokens, context_window and optionally max_input_tokens (read by murmur_artifact::parse_model_limits). For every driver choice, the primary and each inference.alternates entry:

Limit	"manifest"	else "driver"	else
output cap	inference.max_tokens	the table’s max_output_tokens	refused, E-RUN-052
context window	context.max_tokens	the table’s context_window	none, "none"
input ceiling	—	the table’s max_input_tokens	none, "none"

The manifest’s two keys are session-wide: each applies to every choice as written, never clamped to a declared ceiling. A model id matches a table entry by exact string equality. A choice whose model the table does not list draws W-RUN-009 unless the manifest sets both keys.

Every inference call has one input budget, min(input ceiling, context window − output cap) with the cap the call sends (ResolvedModelLimits::input_budget_for). Compaction’s threshold and the seed budget are fractions of it. With no window there is no budget: for that choice compaction does not run and no seed budget applies. A choice whose budget is 0 is refused, E-RUN-055.

inference.compaction.max_tokens is checked here too: above the primary driver’s ceiling for the compaction model it is refused, E-RUN-054; at or above the compaction model’s window it leaves compaction calls no input budget, E-RUN-055.

Staging and mur doctor both resolve through resolve_model_limits, so the two never disagree on a number or on the warning’s text. transport: process resolves nothing: the harness owns its model and its output.

Structs

ChoiceModelLimits

One driver choice’s resolution.

DriverManifest

A driver choice’s driver, as resolution reads it.

InputBudget

The most input tokens one inference call may carry, and which limit set it.

ResolvedModelLimits

One driver choice’s output cap, context window, input ceiling and input budget in force, and where each came from. The model_limits object of session_start and of each inference_choices element.

Enums

BudgetBasis

Which term of min(input ceiling, context window − output cap) an input budget is.

LimitSource

Where a limit in force came from, as session_start names it.

ModelLimitsError

Why a launch is refused: E-RUN-052, E-RUN-054 for ModelLimitsError::CompactionCapAboveCeiling, or E-RUN-055 for ModelLimitsError::NoInputBudget and ModelLimitsError::CompactionNoInputBudget.

Functions

limit_or_none

n, or none for an absent limit, as the refusals, mur doctor and MURMUR.md print it.

model_limits_warning_line

W-RUN-009 as staging and mur doctor print it, message being ChoiceModelLimits::warning.

resolve_model_limits

Resolves the output cap, context window, input ceiling and input budget of every driver choice inference declares, the primary first and the alternates in manifest order. Empty under transport: process.