This PR fix the problem surfaced by Robert above: ...
# proj-agent-canvas
m
This PR fix the problem surfaced by Robert above: https://github.com/OpenHands/agent-canvas/pull/1148 The problem was just with OpenHands provider. In fact, we were persisting the invalid shape
model: "litellm_proxy/*"
+
base_url: null
I could not reproduce the bug... But it will be cool if someone also try to reproduce it
c
Out of curiosity… Sorry, I don’t see well in the video, what backend were you using?
I think we should not mess with ‘translating’
openhands/
anywhere else except at transport time, that is, in llm.py in the SDK
but I think maybe I’ve seen some code on the cloud is translating too, and it seems error prone to me
m
local
c
I saw in the PR description that apparently we store the LLM as “translated”
Not sure why…
We only need to send to litellm the “actual” provider, it seems to me
(at some point I was even suggesting to register
openhands/
as provider with litellm, and then we’d never have to translate anywhere in the codebase, but I don’t think people want quite that)
m
TBH: i don’t know why we are doing that. I just saw that the translation had a corner case that was generating the bug and i fixed accordingly. I suppose there is a good reason for that
c
It might also be a mistaken assumption that has taken a life of its own 😅
🤔
So after one round trip, the UI can legitimately receive
model: "litellm_proxy/..."
for a profile the user selected as OpenHands.
Ugh so user sends
openhands/
to agent-server and gets back
litellm_proxy/
, I won’t lie, it feels like bugs waiting to happen, and translations all over the codebases to nudge them back