Files
Hy-Memory/memhub/apps/cli/src
kristinhe b6a1e2432f fix(memhub): stop the model probe from starving itself
The generation probe asked for 256 output tokens. A thinking model spends
that budget before it writes anything — one measured 470 reasoning tokens to
return a two-field JSON — so the reply came back empty with
finish_reason=length.

The empty body was then read as "this gateway cannot do structured output",
which is a different fact, and that code fired the schema fallback with the
same starved budget. A working model was recorded as unavailable, and the
dialog told the person to compare model names.

Give the probe a budget that covers thinking plus the answer, report
truncation as truncation, and stop sending the disable-thinking preference:
it is a generation preference, and a model that refuses it (GLM 1210) failed
a probe the model itself would have passed.
2026-10-24 02:05:35 +08:00
..
2026-10-16 16:26:27 +08:00
2026-10-16 16:26:27 +08:00
2026-10-16 16:26:27 +08:00
2026-10-16 16:26:27 +08:00
2026-10-16 16:26:27 +08:00
2026-10-16 16:26:27 +08:00
2026-10-16 16:26:27 +08:00
2026-10-16 16:26:27 +08:00
2026-10-16 16:26:27 +08:00
2026-10-16 16:26:27 +08:00
2026-10-16 16:26:27 +08:00
2026-10-16 16:26:27 +08:00
2026-10-16 16:26:27 +08:00
2026-10-16 16:26:27 +08:00
2026-10-16 16:26:27 +08:00