May focused on making streaming and workflow execution more predictable. The work here is mostly infrastructure, but the user-facing result is straightforward: fewer edge cases, cleaner integrations, and safer caching behavior.
Unified streaming architecture
Streaming is now handled through a unified architecture across supported provider paths. This gives ModelRiver a more consistent way to process streamed responses, surface errors, and support OpenAI-compatible integrations.
For users, that means streaming responses should behave more consistently across providers instead of exposing provider-specific quirks to the application layer.
Workflow model aliases
We added support for model aliases in native workflows. Aliases make workflows easier to reason about when the underlying provider/model may change over time.
Instead of treating every model swap like an application integration change, teams can keep using workflow-level configuration and let ModelRiver handle the routing detail.
Better streaming error handling
OpenAI-compatible streaming errors now map more cleanly into ModelRiver's error handling path. That makes debugging failed streams easier and gives API clients more predictable failure behavior.
Response caching isolation fix
We fixed response caching isolation so cached responses stay scoped correctly. This protects users relying on cache behavior for speed and cost control while keeping workflow responses separated as intended.