[1.239.4] — 2026-09-29
Patch from the v1.239.3 live regression.
Fixed
- Korean and Traditional Chinese evaluations reported a missing score summary. Those reports run to 10–11 thousand characters, and the summary comes last, so it fell past the 8192 output-token limit. Evaluations (and the auto-pipeline) now allow 16384 tokens, and when a provider still cuts an answer off, the first warning says so instead of listing missing blocks.
- The live check read a short German answer as not German (3 German stop words, no English ones). A list-shaped answer can hold few stop words; with no English ones at all, three are now enough.
- Tests 3778 → 3782.
Notes
- The deep-research and mode routes keep their 8192-token limit: their answers are not long enough to reach it in the live runs.