@strong-ice-84145 Is it true that when we use ACP with OHE, LiteLLM Proxy is not used for ACP agents? The agent called via ACP is making its own model routing choices and we don't pass LiteLLM proxy configuration through to those agent?

It's not true. There are two setups for Claude Code / Codex via ACP: subscription and API key. If you only setup API keys, then it will use API keys. So ACP can use LiteLLM Proxy in OHE

In OHE if we use api key setup, we need to explicitly set OPENAI_BASE_URL to the enterprise litellm proxy, no? It doesn't automatically inherit those settings or allow us to specify that we want to use the litellm installed by OHE by default, right?

we need to explicitly set OPENAI_BASE_URL to the enterprise litellm proxy,

Yes absolutely. But isn't that parameter already explicited set for non-ACP openhands agent?

It looked to me like the code goes out of its way to unset variables related to litellm proxy when using ACP so as not have unexpected behavior. We have a customer interested in ACP that still wants to route everything through litellm proxy--I think you saw that from @acceptable-analyst-7859 -- so probably ideal scenario for them would be if enterprise agent settings had a checkbox for that which would have the effect of setting appropriate flags / env vars so that the various acp agents would use the proxy so users don't have to figure out what their OHE proxy address is, etc. Not a must do, but would make things easier for them.

Whats a bit tricky is that there are 2 authentication methods for ACP: subscription (much cheaper) and api key (litellm)

Ah yes. And of course the reality is we don't control the model calls for ACP -- that is important for Chris to relate to the customer.

no we don't. But if you set the base url then those calls are sent to your litellm proxy. And you can control the aliases in your litellm proxy!