RUN_CANCELLED missing from @ag-ui/core EventSchemas while adapter unions declare it; cancel is indistinguishable from provider failure

[ozzaii]

Context

We're evaluating AG-UI as the production wire for a music-production agent app (Mastra backend, assistant-ui frontend), with a bounded proof: our server emits AG-UI events over SSE, validated against @ag-ui/core's own EventSchemas, consumed by @ag-ui/client's HttpAgent (only fetch stubbed). Versions: @ag-ui/core@0.0.57, @ag-ui/client@0.0.57, @assistant-ui/react-ag-ui@0.0.50.

Four findings from that proof, receipts below — the first is the one that blocks us.

1. RUN_CANCELLED is rejected by EventSchemas, but adapter unions declare it

@ag-ui/core@0.0.57's EventSchemas rejects a RUN_CANCELLED event with invalid_union_discriminator, while @assistant-ui/react-ag-ui@0.0.50's own AgUiEvent union declares one — the two halves of the ecosystem disagree about whether the event exists.

Consequence: a server that cancels a run has no standard event to say so. The only conforming option is a coded RUN_ERROR.

2. runAgent resolves on RUN_ERROR, so cancellation is indistinguishable from provider failure

With cancel degraded to RUN_ERROR, HttpAgent.runAgent does not reject — it resolves. A caller cannot tell "the user stopped this" from "the model/provider blew up" unless it subscribes to the event stream and reads a non-standard code field. A first-class RUN_CANCELLED (or a typed reason on RUN_FINISHED) would fix both this and #1.

3. HttpAgentConfig has no credentials option

requestInit() emits only {method, headers, body}. Cookie-session auth (our case: same-site session cookie) requires replacing fetch wholesale. A credentials?: RequestCredentials passthrough would remove the workaround.

4. THINKING_TEXT_MESSAGE_* carries no messageId

Reasoning content that must stay id-addressable has to ride the id-keyed REASONING_MESSAGE_* channel instead, which materializes reasoning as a separate role: "reasoning" message rather than a part of the assistant turn (client-side consequence filed against assistant-ui separately).

Happy to provide the validation harness details or test cases for any of these.