My conversation metrics doesn't show info about th...
# questionsf
faint-greece-70506
03/03/2026, 1:43 PMMy conversation metrics doesn't show info about the current context window.
Also: where do I even define the max context tokens OpenHands is allowed to use?
s
salmon-lock-64866
03/03/2026, 8:43 PMThe max_tokens values should come from litellm. If your using a model not in its tables then I think you can create a config.toml for openhands to override it locally. There is a config.template.toml in OpenHands git repo that shows values you can modify. Sometimes litellm will look at env bars to override values but I don't think this one is supported that way.
f
faint-greece-70506
03/04/2026, 8:33 AM@salmon-lock-64866 Thank you for pointing me in the right direction. Do you know where I should put that file? I grepped for the example file inside the openhands container and couldn't find it. In fact, the only *.toml files inside the container (except for .venv) is /app/pyproject.toml which doesn't seem to be related.
The example config.template.toml doesn't explain where it ought to exist. So I don't know where to put it
faint-greece-70506
03/04/2026, 8:44 AMPutting it on /app/config.toml don't seem to work
faint-greece-70506
03/04/2026, 9:05 AMYeah I don't think config.toml applies to V1...
I gotta say, OpenHands itself is awesome. Better than a lot of other agentic tools I've used. But the documentation is extremely lacking
faint-greece-70506
03/04/2026, 9:17 AMThe reason this is important to me is because I have no idea of knowing what the actual context size is at any given moment. And as such, I am not able to make judgements about what value to set the condenser to. Which wouldn't even be a problem if the condenser was based on tokens and not events. Super frustrating. All other tools let's you define max tokens
s
salmon-lock-64866
03/04/2026, 4:35 PMReally? Putting it at /app/config.toml worked for me in recent past but I don't currently use to confirm.
f
faint-greece-70506
03/04/2026, 9:11 PMI find it confusing that OpenHands uses event-based condensing. Do you know the reasoning behind this? What's the upside? To me it just makes it unpredictable, since one event can consist of more tokens than another