Tool calling broken with OpenRouter

[alexanderniebuhr]

Whenever I tried to use ai-openrouter with an OpenAI model tool calling is broken, especially the loop functionallity. I see the input_complete message in the frontend stream, but the tool execute function is never called and returned to the llm. Switching to ai-openai everything works as docs say. I really like what you are building, but the docs are up to date and it seems like even official adaptes don't work. Also what is a way to debug is in when not using react or the devtools, I was adding logs to the node_modules folder 🤔

[merodiro]

I had the same issue with OpenRouter. The only way I was able to get it to work is by using the OpenAI adapter and changing the baseURL. I hope this changes soon, now that OpenRouter has become an official partner

const adapter = createOpenaiChat(
  "openrouter/auto",
  process.env.OPENROUTER_API_KEY!,
  {
    baseURL: "https://openrouter.ai/api/v1",
  }
);

[alexanderniebuhr]

Sounds bad, but that caused us to churn for now. Maybe we will look at this again in the future, but for now we moved to more stable options from the ecosystem.

[AlemTuzlak]

There was a bug that was introduced around this time into tool calling and we didn't have proper e2e testing and it introduced a regression. It should be fully resolved at this point!

[ishita-0301]

I've hit similar issues before. When tool execution silently breaks, debugging can be pretty painful, especially outside the React/devtools setup. I usually start by logging the raw messages exchanged with the model and verifying that the provider is actually returning valid tool calls instead of just emitting stream events.

On a related note, I've been looking at https://github.com/FailproofAI/failproofai which focuses on catching runtime agent failures like loops and repeated errors. Different problem, but having extra visibility into agent execution has been really helpful when tracking down issues like this.