-
1.0.6: make the thinking switch real and collect reasoning tokens
iomgaa released this
2026-08-02 20:17:30 +08:00 | 198 commits to main since this releaseissue #5: enable_thinking=False was a silent no-op on minimax and openai
sources. The parameter shape stays at provider level; a model-level
capability table now says whether a model can honour the switch, and
models that cannot are refused at assembly instead of pretending.issue #6: reasoning_tokens is collected from the usage payload, so the
cost of thinking can be attributed separately from the cost of
answering. The telemetry port grows from 21 to 22 fields.Verified against the live gateway across 137 real calls.
Downloads