Module limits 

Model context window limits registry.

There is intentionally no built-in per-model table. Real per-model values come from two sources, in priority order:

1. provider runtime metadata (e.g. Copilot reports real context/output),
2. user overrides persisted in model_limits.json.

Anything without a match falls back to a single global default (DEFAULT_MAX_CONTEXT_TOKENS / DEFAULT_MAX_OUTPUT_TOKENS). This keeps the registry from going stale as models churn — see token_budget.rs.

Structs

ModelLimit

Model limit configuration (user-overridable).

ModelLimitsRegistry

Registry for model limits with built-in defaults and user overrides.

Constants

DEFAULT_MAX_CONTEXT_TOKENS

Global default context window applied to any model without a provider metadata value or a user override. 1M reflects the current mainstream range across frontier models (Claude 3.5, GPT-4o, Gemini 1.5, etc.).

DEFAULT_MAX_OUTPUT_TOKENS

Global default maximum output tokens.

DEFAULT_MODEL_PATTERN

Sentinel pattern used for the single global fallback limit.

DEFAULT_SAFETY_MARGIN

Default safety margin for token counting errors (floor; scales with context window via ModelLimit::get_safety_margin).

Functions

create_budget_for_model

Create a token budget for a specific model.

default_model_limit

Build the single global default limit (1M context / 128K output).

get_default_config_path

Get the default configuration file path.

is_default_limit

Whether a user override is a no-op — identical to the global default, so it carries no information and need not be persisted (diff-only storage).

load_model_limits_from_unified_config

Load user model limits from the unified config.json model_limits value.