openclaw - 💡(How to fix) Fix Ollama: gateway times out at 60s despite model responding in <10s via direct curl [1 comments, 2 participants]

Summary

OpenClaw's gateway times out on all Ollama requests after exactly 60 seconds, while the same model responds correctly in <10s via direct curl. The timeout fires even with a pre-warmed model already resident in VRAM. The Ollama stack is healthy — the break is in the gateway's HTTP streaming path to Ollama.

Diagnostic evidence

Both models in the fallback chain (qwen2.5:3b, qwen2.5:7b) time out at the same threshold. The gateway logs show the timeout fires in the lane task layer:

[agent/embedded] embedded run failover decision: reason=timeout provider=ollama/qwen2.5:3b
[diagnostic] lane task error: lane=main durationMs=60371 error="FailoverError: LLM request timed out."
[diagnostic] lane task error: lane=session:agent:quinn:main durationMs=60373
[model-fallback/decision] candidate_failed reason=unknown next=ollama/qwen2.5:7b
[diagnostic] lane task error: lane=main durationMs=60733 error="FailoverError: LLM request timed out."

Key observations

1. Not a cold-start problem. Pre-warmed model (already resident, 2279MB VRAM), expires_at confirming it was loaded — same 60s timeout.
2. Not a model performance problem. Direct curl /api/chat stream:true returns first token in 4.2 seconds.
3. Hardcoded 60s timeout. Both fallback models hit exactly 60s. No config key found in configuration-reference.md or schema to adjust this. agents.defaults.timeoutSeconds: 1200 and openclaw agent --timeout do not affect it.
4. Gateway-first, embedded-second, both fail. The diagnostic output shows Gateway agent failed; falling back to embedded — the embedded path uses the same timeout, so fallback does not help.
5. Exit code unreliable. openclaw agent returned exit 0 on the first failure run and exit 1 on the second — cannot be used for programmatic pass/fail detection.

Environment

• OS: Ubuntu 24.04.4 LTS (kernel 6.17.0-20-generic)
• OpenClaw: v2026.4.2
• Hardware: Lenovo V15 G4 ABP — AMD Ryzen 7 7000, 40GB RAM, AMD Radeon iGPU (Vulkan)
• Ollama: localhost:11434, OLLAMA_VULKAN=1, OLLAMA_FLASH_ATTENTION=1
• Model: qwen2.5:3b (primary), qwen2.5:7b (fallback)
• Provider config: baseUrl: "http://127.0.0.1:11434", api: "ollama", apiKey: "ollama-local"

Expected behavior

1. A configurable models.providers.ollama.requestTimeoutMs (or equivalent) key so local model latency can be accommodated.
2. Alternatively: the gateway streaming path should begin the timeout clock from first-token receipt, not from connection open — since local models may have high prompt-eval latency before any tokens stream.

Workaround

Direct curl http://localhost:11434/api/chat works correctly. openclaw agent against Ollama-backed agents is non-functional on this hardware until resolved.