Symptom: Qwen3.8-Flash-Next and GLM-5.3-Flash repeat plans in reasoning, rethinking again and again in each block without apparent awareness of prior turns and being slow to progress under Hermes Agent.
Cause: By default, Hermes Agent strips reasoning from assistant replays for custom providers, whatever the backend and the chat template might have set it to. This can cause problems for models that accept (or indeed expect) preserved reasoning. Other frameworks may have similar defaults, though I am not familiar with any specific examples.
Fix: Set model.reasoning_echo totrue in Hermes Agent config.
Just thought this might be non-obvious and useful for anyone else having loopy local models in agentic harnesses. That would be all.
1 Like
This is a great find! I was seeing looping on two MoE models myself (Gemma4-26b and qwen3.6-a35b) after getting a ways into context. It would basically hit some reasoning sentence in its thinking that kept getting repeated ad infinitum. Maybe this is the reason.
Does this increase token usage over time? I couldn’t find a straight answer on the config key in the hermes wiki as to how it feeds back into the next prompt.
Hello and welcome,
It will by definition increase prefill token consumption, since the framework is now sending previous turns’ reasoning chains along with assistant reply. Expect context to fill up much faster for long reasoning, too.
Though reasoning loops may end up consuming more decode tokens, depending on usage.