UIC: luna stalls before first output on 10 of 308 page calls, and a stall is not retried

[bbertucc]

Iris Maintainer Agent here.

On iris.equalify.uic.edu, us.openai.gpt-5.6-luna often goes silent before its first output. Iris doesn't retry that, so the page is lost.

What the logs show

From 2026-09-01 to 09-24, luna made 308 page calls on the UIC deployment:

• 10 hit the 120 s first-output stall. No other model did.
• 3 returned an empty stream after 42, 82 and 83 seconds. This was [bug]: Bedrock Fail #480, and fix(providers): a stream that delivered nothing is sent again, not called a partial document #481 now retries it once.

A first-output stall throws StalledStreamError, which is not retried. The page fails, and the document is delivered without it (Partial documents).

Options

1. Move the page agent off luna on this deployment. This is a config change in the private deploy repo, and it is the user's decision to make: it changes cost and output. Say the word and I'll make it.
2. Retry a first-output stall once, as fix(providers): a stream that delivered nothing is sent again, not called a partial document #481 does for an empty stream. That adds up to 120 s to an affected page, and it may pay for the prompt twice: I have not checked whether Bedrock bills a call that never produced output. The risk is doubling the wait on a model that is down.

I'd do 1 first. Until someone decides, I'm leaving this open.

🤖 Generated with Claude Code