Support for the /v1/responses endpoint.

[Vodolay]

To use streaming with the Agent I had to revert to the /v1/completions endpoint. It be nice to have a version of the Agent which works with the separated streams from the /v1/responses endpoint without any hacks.

[ianmacartney]

Can you say more about what the issue was? My understanding is that using the openai provider to generate streaming text works - I'm not sure where you're providing the endpoint manually. If it's supported with the AI SDK then it should work just fine with the Agent component

[Vodolay]

If I use

export const agent = new Agent(components.agent, {
name: "apriel",
languageModel: vllm("Apriel...."),
instructions: You are a ...,
});

the agent connects to the /v1/response end point and I simply get this error in the pnpm convex dev terminal:
[CONVEX A(agent:chat)] Uncaught Error: error decoding response body
Obviously the response is then not rendered by the UI.

On the other hand if I use (Notice vllm.chat)

export const agent = new Agent(components.agent, {
name: "apriel",
languageModel: vllm.chat("Apriel...."),
instructions: You are a ...,
});

the model connects to the /v1/chat/completions endpoint and everything is fine.

The issue could be with the way /v1/response separates the streams or could be that this new endpoint is stateful or both.

[ianmacartney]

does this work with AI SDK directly? generateText({ model: vllm("Apriel..." ?

…

On Tue, Jan 6, 2026 at 7:54 AM Ivan T. Ivanov ***@***.***> wrote: *Vodolay* left a comment (get-convex/agent#205) <#205 (comment)> If I use export const agent = new Agent(components.agent, { name: "apriel", languageModel: vllm("Apriel...."), instructions: You are a ..., }); the agent connects to the /v1/response end point and I simply get this error in the pnpm convex dev terminal: [CONVEX A(agent:chat)] Uncaught Error: error decoding response body Obviously the response is then not rendered by the UI. On the other hand if I use (Notice vllm.chat) export const agent = new Agent(components.agent, { name: "apriel", languageModel: vllm.chat("Apriel...."), instructions: You are a ..., }); the model connects to the /v1/chat/completions endpoint and everything is fine. The issue could be with the way /v1/response separates the streams or could be that this new endpoint is stateful or both. — Reply to this email directly, view it on GitHub <#205 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AACZQW2WT3FOEKH7NHOPGFL4FPLB7AVCNFSM6AAAAACQRDKOISVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTOMJVGI2DANRUHA> . You are receiving this because you commented.Message ID: ***@***.***>