LanguageModel

Stable identifier for telemetry / routing ("litertlm/gemma-3-1b", "openai/gpt-5").

Provider tag for the model — "openai", "anthropic", "litert-lm", etc. Mirrors v6's LanguageModelV3.provider (per historical parity gap #40). Lets routing layers (a fallback chain, a tier picker) make provider-aware decisions without parsing modelId.

Map of URL regex → media-tag list describing which URL shapes the provider accepts inline (so the agent doesn't have to download + base64-encode them itself). Mirrors v6's LanguageModelV3.supportedUrls. Empty default — most on-device providers don't accept URLs and fall back to base64.

One-shot completion. Requires explicit low-level opt-in at direct call sites.

Streaming completion. Cold until collected, then drives one upstream call per collection. Requires explicit low-level opt-in at direct call sites. Streaming has no default deadline; configure the provider transport or wrap the call in a higher-level CallConfig.timeout.

Streaming completion plus metadata available before stream collection. Implementations must keep the returned stream cold, and must not flowOn/emit from a foreign context — the collection context is the caller's choice. Providers can override this to expose request bodies and response headers while preserving the v6 doStream result shape.