Module limits 

Per-model context/output limits for recognized providers, so the agent can size context against the real window instead of guessing.

The provider’s own API does not expose a context window: the OpenCode Go (zen) /models endpoint returns only id/object/created/owned_by, and rig leaves Model::context_length empty for every OpenAI-compatible listing (it is only populated for providers whose listing reports a window, e.g. Groq/Moonshot/OpenRouter). The authoritative catalog is models.dev, whose opencode-go entry (api = https://opencode.ai/zen/go/v1) carries limit.context / limit.output per model. [OPENCODE_GO] is a snapshot of that table.

The live measure of how full the context is comes from the provider’s own response usage (prompt_tokens → crate::message::Usage::input_tokens), not from here; these limits only say how much fits.

Structs

ModelLimit

Context window and output ceiling advertised for a model.

Constants

DEFAULT_CONTEXT_WINDOW

Context window assumed when neither the model catalog nor config supplies one: a conservative baseline so an unknown model still compacts in time.

DEFAULT_RESERVE_CAP

Cap on how many tokens are held back for the model’s reply. The advertised output ceiling can be huge (OpenCode’s deepseek-v4.1-flash is 384k), so it is capped to keep most of the window usable for input.

Functions

is_opencode_go

True for endpoints served by OpenCode’s Go (zen) API.

model_limit

Limits for model served from base_url, when we recognize the provider. None = unknown provider/model (the caller falls back to a default budget).