Long local-model runs finish but the chat reports an error

Troubleshooting
, , , ,
1

Adapted community discussion: Paraphrased and condensed from an anonymized public OpenClaw Discord support conversation from September 2026. Technical guidance was reviewed for OpenClaw 2026.9.3.

I am testing a slow local model through the browser interface. A long task can run for more than 15 minutes and complete its work, but the chat reports an error instead of returning the final reply. The model server does not report a context overflow. Is an OpenClaw timeout cutting off the response?

2

There are two relevant limits: agents.defaults.timeoutSeconds covers the whole agent run, while models.providers.<provider>.timeoutSeconds covers the provider request and stream. Check both effective values and follow the logs while reproducing the problem:

openclaw config get agents.defaults.timeoutSeconds
openclaw config get models.providers.PROVIDER_ID.timeoutSeconds
openclaw logs --follow

Completed work with a missing final reply points more toward an expired run or provider stream than context overflow. For jobs this slow, a background session can also be more resilient.

3

What commands should I use to raise those limits from the terminal?

4

Set the whole-run limit above the provider limit. These example values are in seconds:

openclaw config set agents.defaults.timeoutSeconds 3600
openclaw config set models.providers.omlx.timeoutSeconds 1800
openclaw config get agents.defaults.timeoutSeconds
openclaw config get models.providers.omlx.timeoutSeconds

Replace omlx with the actual key used under models.providers if yours differs. Current reference: local models.

5

I changed both values and am testing now. Do I need another step to make them survive a restart?

6

No. openclaw config set writes persistent configuration, so those values survive Gateway and computer restarts. You can rerun the two config get commands to verify what was saved.

7

Raising the timeouts worked. The local model still took a while, but the chat completed without the error.

8

That confirms the timeout was the immediate problem rather than context overflow. The remaining delay is model generation speed; for faster interactive work, reduce the model’s reasoning or output budget, or use a smaller local model.