Repository navigation
OpenRouter/ Kimi K2 just stops working #1329
Description
It said "Let me run the specific tests to see what's failing" but did not proceed. So I said "do it" and it still did not do anything. The status line does not say "working", indicating execution has stopped.
Metadata
Metadata
Assignees
Labels
No labels
Activity
tnypxl commentedon Jul 25, 2025
I wouldn't think this is an opencode issue. I think this is on openrouter's side. When I was doing model comparisons on via their site, Kimi k2 would stop working there as well. You might research if this issue is experienced on other platforms that offer this model.
esafak commentedon Jul 25, 2025
I disagree. From the user's perspective it's opencode's problem; I am using it, and openrouter is a dependency. I don't have access to the openrouter API's response to opencode, and I don't care to. If openrouter times out or returns an error, opencode should deal with it.
What would be a good diagnostic is for opencode to display which provider openrouter selected.
pepperpepperpepper commentedon Jul 26, 2025
Can confirm that Kimi K2 through Groq works flawlessly. Never had this particular issue.
tnypxl commentedon Jul 26, 2025
I think openrouter is responding successfully but with incomplete content from the LLM. Unless you grab the stop reason from the LLM, opencode probably won't think there is an issue. I've seen this in Zed editor and in Neovim.
esafak commentedon Jul 26, 2025
It should be inspecting that and reporting the issue, not pretending nothing happened.
thdxr commentedon Jul 26, 2025
the problem is openrouter is hanging and not giving us anything - we need to add a timeout here
andreagrandi commentedon Jul 27, 2025
I confirm this problem. I experienced the same problem yesterday. At the beginning it seems to do something (it changed a few lines in a file), then completely stopped working.
The problem is that if I switch to a model which works (ie
sonnet-4from GitHub Copilot) it behaves in the same way. I had to completely close the app and start a new session directly with GitHub Copilot / Sonnet-4 to continue working.I had to stop trying because I was burning paid tokens without doing any work 🫤
andreagrandi commentedon Jul 27, 2025
the problem is also (not fixable by you, I know) Openrouter not giving us anything AND charging us anyway 🙄
qydxyx commentedon Jul 28, 2025
Pretty sure it's openrouter's issue. when using function call, most providers of kimi k2 will stop for no reason. It's suggested to use official moonshot api. Some have also mentioned that groq api works well, and they offer k2 model with greater speed .
esafak commentedon Jul 28, 2025
I reported it in their Discord. They said they are going to look into it, may stop supporting tool calling until it is fixed, and suggest using Groq or Moonshot as the provider.
DanielTowerz commentedon Jul 30, 2025
Moonshot as provider presents the same behavior for me.
mizchi commentedon Jul 30, 2025
I discovered by looking at the Roo code that the tools all worked when I set it to openai-compatible.
In my environment, the tools couldn't be used without this setting.
{ "$schema": "https://opencode.ai/config.json", "provider": { "moonshot": { "npm": "@ai-sdk/openai-compatible", "options": { "apiKey": "...", "baseURL": "https://openrouter.ai/api/v1" }, "models": { "moonshotai/kimi-k2": {} } } } }It works with tools.
esafak commentedon Jul 30, 2025
If so this is something that ought to be done in opencode, not in personal config files.
daleal commentedon Aug 2, 2025
I have some weird example where the model (groq's K2 through an openrouter preset) actually made a change correctly, and then started outputting the same change to the console (1 change at a time, stopping immediately after showing the change), as if it didn't have access to what it had just done. And then I told it to just move on and it carried on just outputting 1 diff to the console and stopping there
rekram1-node commentedon Dec 28, 2025
[automated] Closing due to 90+ days of inactivity. Feel free to reopen if you still need this!