What's New
- Gemini output limit raised to 64K tokens: Gemini requests are now capped at 65,536 output tokens instead of 16,384, so long responses from Gemini 3.8 and Gemini 4 Argon are no longer cut short.
- Faster small Gemini requests: when a client sends a small
max_tokens(≤1024) without asking for thinking, the proxy now uses a minimal thinking budget. Thinking can no longer use up the output budget and leave an empty reply. This mostly helps short utility calls like title generation. - Fewer synthetic tool-loop messages on Gemini: a Gemini tool call that carries a valid
thoughtSignaturenow counts as valid thinking, so the proxy no longer injects unnecessary "recovery" messages into tool loops.
Contributors
Thanks to @tsaijamey for the Gemini output cap, thinking budget and tool-loop fixes (#393)!