Repository navigation
Conversation
|
Happy to rebase or reshape the section if it doesn't fit the page's conventions — the gateway config in the diff was verified end-to-end against a live OpenHands setup, and I can attach a short screen recording of the flow if that helps review |
enyst
left a comment
There was a problem hiding this comment.
I'm an AI agent based on Opus 5.5, helping Engel Nyst (@enyst) with project work.
Thanks for the disclosure. The generic part of this section is already on the OpenAI page (Using an OpenAI Proxy). It isn't Moonshot-specific, since any gateway works the same way. The remaining part names one gateway, which is the same question as #850, and maintainers have declined such vendor examples before (#504, #561, #619, and OpenHands/OpenHands#16857).
For the outdated kimi-k2-0711-preview recommendation, #634 already refreshes this page with Moonshot's own models. I'd suggest closing this in favor of that.
Summary of changes
Adds a short "Using an OpenAI-Compatible Gateway" section to the Moonshot provider page: the three
Advancedsettings (custom model with theopenai/prefix, base URL, API key) plus one verified gateway option for Kimi K3.Why
The Moonshot page currently only covers the official platform, and its recommended model (
kimi-k2-0711-preview) predates the current Kimi releases. Users who access Moonshot models through OpenAI-compatible gateways find no guidance on this page, while the same pattern is already documented elsewhere (for example, the LiteLLM proxy page).Disclosure
I run Wallaby, the gateway named as the verified example. Happy to swap in a different example or leave the section generic if you'd prefer.
Testing
The
openai/kimi-k3+ base URL configuration is what we run in Agent Canvas, verified end to end on 1.23.0 and 1.24.0 (chat, scheduled automations, webhook integrations).