Shipped foundations plus where ai is heading in 2026. Planned items are aspirational, not yet implemented.
- Provider / Protocol / Vendor three-layer split
- Frozen-dataclass Context / StreamOptions / Model type model
- Streaming event union with .result() future
- Tool calls, web search, thinking levels, cancellation
- GitHub device-code auth → Copilot JWT
- Three wire protocols (openai-completions / anthropic-messages / openai-responses)
- Model catalog sync with 24h cache + offline fallback
- Usage / cost accounting with premium-request multipliers
- Anthropic-compatible /v1/messages proxy
- Additional first-party vendors behind the same Provider surface
- Provider-agnostic model-capability discovery
- Cross-provider routing / fallback policies
- Unified per-run cost budgets surfaced to callers
- Automatic prompt-cache block placement
- Structured-output / JSON-schema responses across protocols