honest epoch
What I'm trying to do & What's happening instead:
I am trying to use the OpenClaw TUI and Telegram channel (@Friday_anasclawbot) with local Ollama models on a CPU-only VPS.
Instead of responding, the TUI hangs indefinitely (or resets the stream after 30s). When I bypass the TUI and test via CLI, the agent times out after exactly 120 seconds, completely ignoring my config override to wait 300 seconds.
However, my Ollama instance is working fine. A raw curl to Ollama's native /api/chat using qwen2.5:1.5b returns a successful response in ~20 seconds. It seems OpenClaw is either dropping the stream or hard-capping the wait time at 2 minutes before the model can finish its first inference.
My setup:
Hosting: Contabo VPS (Ubuntu Linux, 8-core CPU, no GPU) 400gb ssd, 24gb ram.
OpenClaw version: 2026.4.23 (a979721)
Node: v22.22.2
Providers: Local Ollama (Models loaded and tested: qwen2.5:1.5b, llama3.1:latest, gemma2:latest, gemma4:latest)
Error messages / logs:
Running the standard CLI smoke test:
openclaw infer model run --model ollama/qwen2.5:1.5b --prompt "Reply with exactly: smoke-ok" --json
Yields this timeout exactly 2 minutes later:
20:14:55 [agent/embedded] Profile ollama:default timed out. Trying next account...
20:14:55 [agent/embedded] [llm-idle-timeout] ollama/qwen2.5:1.5b produced no reply before the idle watchdog; retrying same model
20:16:59 [agent/embedded] Profile ollama:default timed out. Trying next account...
20:16:59 [agent/embedded] embedded run failover decision: runId=6bedbd6c... stage=assistant decision=surface_error reason=timeout from=openclaw/gateway-injected
(It then outputs a fallback error: Error: Gateway is only bound to loopback even though my config bind is clearly set to lan).