fix(streaming): mark the research span as ERROR when the stream fails

[miurla]

Problem

When the chat stream fails, the failure never reaches the tracing surface. The root research observation ends with level: DEFAULT and statusMessage: null even though the turn produced no answer at all.

This matters because tracing is the only surface that covers guest and ephemeral executions, and it is what gets scanned for error-level observations. A recent provider outage produced a large number of consecutive turns where every observation on the research path was recorded at DEFAULT with zero tokens and no output. The failure was only visible because title-generation runs through generateText, which throws and therefore does get recorded as an error. Had the chat path been the only caller, the outage would have produced no error-level observation whatsoever.

Root cause

The provider error arrives as an error chunk inside the response stream rather than as a thrown exception, so the surrounding try/catch never runs and the request completes with HTTP 200. The onError callback passed to toUIMessageStreamResponse does fire, and it logs the error and serializes it for the client, but nothing marks the span. endTracing() then calls rootSpan.end() on an unmodified span.

Fix

Capture the error in both entry points (create-chat-stream-response.ts and create-ephemeral-chat-stream-response.ts), from the stream onError callback and from the surrounding catch, and mark the root span with level: 'ERROR' plus a status message before it is ended.

The status message comes from a new describeStreamError helper that uses only the code and error fields of the existing public error payload. That payload is already sent to every client, so the trace cannot pick up user content, a provider response body, or credentials. The result is capped at 300 characters, which matters because some public payloads are passed through verbatim.

This is additive only. Response status codes, the serialized public error payload, and the existing logging are unchanged.

Client aborts are not affected: aborts travel as a distinct abort chunk and set isAborted on onFinish, so they never reach onError and will not be marked as errors.

Verification

• New unit tests cover a provider billing failure, an unclassified Error, a non-Error value, and the length cap, asserting the exact resulting message and that raw input text never appears in it.
• bun run test 295 passed, bun typecheck, bun lint, bun format:check and bun run build all pass.
• The billing case is asserted against the exact error shape observed in production, which resolves to provider_billing: The AI service is currently unavailable.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
morphic	Ready	Preview	Aug 3, 2026 10:53pm