openclaw - 💡(How to fix) Fix long model timeouts . [1 participants]

Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

LLM generated summary : When performing complex tasks, the system's model selection is unreliable (e.g., asking for a specific model fails) and performance degrades severely as context grows, even if context clean-up is performed. To me it would timeout and these is not way that , I or the LLM can change the timeouts it would just remove the setting on gateway restart .

Steps to reproduce

using a large local model qwen3.5:27b on a ( strix halo ) at over 20% ctx , or about 50% on qwen3.5:9b would timeout , then fall back to qwen3.5:27b , timeout again and run out of models to use.

Expected behavior

LLM generated , The system should reliably allow the user to select any installed model (e.g., ollama/qwen3.5:27b) via a clear, working CLI command. Performance should remain stable across the full range of models, regardless of context size, assuming the context size is manageable (e.g., under 50k tokens).

Actual behavior

LLM generated , Model switching commands fail or result in the system sticking to the default model despite explicit requests (e.g., asking for 27B results in 9B response). Performance degrades noticeably when context accumulates, indicating that the model parameters are the primary limiting factor, not the context window ceiling. To me , being able to set a larger timeout ?

OpenClaw version

4.23 , the latest .

Operating system

ubuntu server 20.04 ?

Install method

via your install .

Model

ollama local , various qwen3.5 and gemma4 models

Provider / routing chain

openclaw / gateway on VM > ollama direct on host .

Additional provider/model setup details

NAME ID SIZE MODIFIED
qwen3.5:9b 6488c96fa5fa 6.6 GB 4 days ago
qwen3.5:35b-a3b 3460ffeede54 23 GB 5 days ago
qwen3.5:27b 7653528ba5cb 17 GB 5 days ago
gemma4:26b 5571076f3d70 17 GB 5 days ago
gemma4:31b 6316f0629137 19 GB 10 days ago
gemma4:e4b c6eb396dbd59 9.6 GB 10 days ago

Logs, screenshots, and evidence

17:11:34
warn
agent/embedded
Profile ollama:default timed out. Trying next account...
17:11:34
warn
agent/embedded
embedded run failover decision
17:11:34
error
diagnostic
lane task error: lane=main durationMs=122445 error="FailoverError: LLM request timed out."
17:11:34
error
diagnostic
lane task error: lane=session:agent:main:main durationMs=122447 error="FailoverError: LLM request timed out."
17:11:34
warn
model-fallback/decision
model fallback decision
17:12:19
warn
model-fallback/decision
model fallback decision

Impact and severity

ask it to make a skill with qwen3.5:9b , this is what I want you to do , make this skill with these rules , making the skill but as the context window is filling up the model is getting slower to until timeout , swaps modes , I dead system without a /new or /clear.

Additional information

ollama overrides . [Service] Environment="HSA_OVERRIDE_GFX_VERSION=11.5.1" Environment="HSA_ENABLE_SDMA=0" Environment="OLLAMA_HOST=0.0.0.0" Environment="OLLAMA_NUM_PARALLEL=4 Environment=OLLAMA_REQUEST_TIMEOUT=10m # API request timeout Environment=OLLAMA_KEEP_ALIVE=-1 # Model keep-alive duration Environment=OLLAMA_MAX_QUEUE=512 # Maximum queued requests