Skip to content

Mid-answer drops on GLM 5.3 are the caller hanging up, not Novita - #15

Draft
Marketen wants to merge 1 commit into
mainfrom
ncr/6-mid-answer-drops-on-glm-5-3-are-the-call
Draft

Marketen wants to merge 1 commit into
mainfrom
ncr/6-mid-answer-drops-on-glm-5-3-are-the-call

Conversation

@Marketen

Copy link
Copy Markdown
Contributor

Why

In these failures the caller had already closed the connection: the gateway keeps reading Novita's stream, its request context gets cancelled, the read fails, and that read error is logged as a provider fault with a 502. I reproduced it exactly - a deliberately aborted request through the same endpoint produced an identical line (mid-stream read error, provider fault, 502), while the same abort before the first token was correctly recorded as client_canceled/499. Novita was healthy throughout: the same caller finished a 169-second stream on the same model and provider two minutes later…

In these failures the caller had already closed the connection: the gateway keeps reading Novita's stream, its request context gets cancelled, the read fails, and that read error is logged as a provider fault with a 502. I reproduced it exactly - a deliberately aborted request through the same endpoint produced an identical line (mid-stream read error, provider fault, 502), while the same abort before the first token was correctly recorded as client_canceled/499. Novita was healthy throughout: the same caller finished a 169-second stream on the same model and provider two minutes later, and th…

Prepared in the Nexus Control Room; approved by Marc.

Co-Authored-By: Nexus Control Room agent <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant