#Openclaw in loop
1 messages · Page 1 of 1 (latest)
That symptom can come from two very different places, and the fix depends on which one you’re actually seeing:
1) It’s “one message that keeps getting edited/extended” (Telegram streaming preview)
On Telegram, OpenClaw can send a temporary preview message and keep editing it while the model streams. If your model never hits EOS / never stops, it looks like an infinite loop.
Common causes with llama.cpp + Qwen:
- wrong chat template (Qwen needs the Qwen template; the wrong one can prevent proper EOS)
- llama.cpp running with something like ignore-EOS, or no practical max_tokens / n_predict limit
- decoding params that encourage repetition (too low temp + weak repeat penalty)
What to try (llama.cpp side):
- ensure you’re using the correct Qwen chat template
- ensure generation has a hard cap (max_tokens / n_predict) and you’re not ignoring EOS
- bump repetition controls (e.g., repeat penalty) so it can’t spiral
2) It’s sending the same finished answer again and again (delivery/session loop)
If Telegram shows multiple separate messages (or the same final answer re-sent repeatedly), that’s usually not the model—it’s OpenClaw replaying a stuck state.