Known issues

Fixed issues move to the Changelog.

No cached tokens on turn 2 with 30 long conversations

Since 2026-10-29. Status: cache size limit; no fix planned.

30 conversations of about 2,400 tokens evict each other from the 6.35 GiB prefix cache between turns. Turn 2 is slower than turn 1. Details: Prefix caching.

Workaround: run long conversations in groups of 8 to 10.

First start after an engine update takes 11 minutes

Since 2026-10-29. Status: expected after each engine update.

The first start of a new engine version compiles kernels: 656 s measured, against 226 to 285 s for later starts. cold_start_s (150) and Retry-After underestimate the first start.

Workaround: stream without X-LLM-Wait (the request waits up to 15 minutes), or poll GET /v1/models until state is loaded.

Cloudflare rejects the Python urllib User-Agent

Since 2026-10-29. Status: Cloudflare rule; no change planned.

A request with the Python urllib default User-Agent (Python-urllib/3.x) fails with 403 and the plain-text body error code: 1010. The OpenAI SDKs, curl and Node fetch are not affected.

Workaround: set a User-Agent header, e.g. User-Agent: my-app/1.0.

Non-streamed requests end after 100 seconds

Since 2026-10-29. Status: Cloudflare limit; no change planned.

A request with no response headers after about 100 s fails with 524 (plain text). Affected: long non-streamed answers, thinking on hard questions, X-LLM-Wait above 90 on a loading model.

Workaround: stream. A stream starts within 10 s and stays open until the answer ends.