Problem with Agent Plugin

[osseonews]

We installed the agent plugin and set it up. We are using Digital Ocean Serverless Inference platform: https://docs.digitalocean.com/products/inference/how-to/si-endpoints/

It fully supports Responses API and Completions API. Actually, we tried both, using useCompletionApi: false/true in the set up.

No matter what we do, we get back the following response in the API: ERROR (User logs): Agent response streaming failed:

We confirmed all required information and the model used, and we actually tested it on Digital Ocean playground and it works there, but in Adminforth, we just get back and API error about streaming. What is this error? Any idea on how to fix?

What's strange is that we if ask the API endpoint: "What is the capital of France" from Admin forth, it answer back: "Paris", but then "Error:" with the error per last comment.

[yaroslav8765]

Hi! We'll research your problem and let you know as soon as possible when solution will be found

[osseonews]

If it helps, here are the logs. Seems to maybe be an issue with the tools? Because we do get answer back for simple questions, like "What is the capital....?"

[11:37:01.875] INFO (User logs): AdminForth Agent callable tools: get_resource, get_user_location, navigate_user, fetch_skill, fetch_tool_schema
[11:37:02.510] ERROR (User logs): Agent response streaming failed:

[ivictbor]

@osseonews , we'll check it, yes, compatible APIs are really not always compatible so we have to test it. It is API error, related to DO endpoints

Can you send also code snippet of adapter instance? So it shows params like - which URL is used, which model, etc

Also can you check DO endpoint directly

// test-do-chat.mjs
const apiKey = process.env.DO_INFERENCE_KEY || process.env.DIGITALOCEAN_TOKEN;

if (!apiKey) {
  throw new Error("Set DO_INFERENCE_KEY or DIGITALOCEAN_TOKEN");
}

const res = await fetch("https://inference.do-ai.run/v1/chat/completions", {
  method: "POST",
  headers: {
    "Authorization": `Bearer ${apiKey}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    model: process.env.DO_MODEL || "llama3.3-70b-instruct",
    messages: [
      { role: "system", content: "You are a helpful assistant." },
      { role: "user", content: "Say hello in one short sentence." },
    ],
    max_completion_tokens: 64,
    temperature: 0.2,
  }),
});

const json = await res.json();

if (!res.ok) {
  console.error(json);
  process.exit(1);
}

console.log(json.choices?.[0]?.message?.content);
console.log("usage:", json.usage);

[osseonews]

yeah, we tested the endpoint with CURL and it works fine. Copied below. It's not a Digital Ocean issue. Actually, when using adminforth with get back the response, but then it is always follwed by an error per issue above. The result is never saved to the database either. I think there is an issue with your tools calls. Because the error always follows the callable tools. We noticed also that if we ask if we have a specific table in the database, the agent responds correcly, but then if we ask to get specific information about a resource, it fails with the error above. So I think something with tools has a bug.

curl https://inference.do-ai.run/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $MODEL_ACCESS_KEY" \
-d '{
  "model": "openai-gpt-oss-120b",
  "messages": [
    {
      "role": "user",
      "content": "What is the capital of France?"
    }
  ],
  "max_tokens": 100
}'

[ivictbor]

Same for DO responses API

const apiKey = process.env.DO_INFERENCE_KEY || process.env.DIGITALOCEAN_TOKEN;

if (!apiKey) {
  throw new Error("Set DO_INFERENCE_KEY or DIGITALOCEAN_TOKEN");
}

const res = await fetch("https://inference.do-ai.run/v1/responses", {
  method: "POST",
  headers: {
    "Authorization": `Bearer ${apiKey}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    model: process.env.DO_MODEL || "openai-gpt-oss-20b",
    input: "Say hello in one short sentence.",
    max_output_tokens: 64,
    temperature: 0.2,
    stream: false,
  }),
});

const json = await res.json();

if (!res.ok) {
  console.error(json);
  process.exit(1);
}

console.log(json.output_text);

console.log(
  json.output?.flatMap(item => item.content || [])
    ?.filter(part => part.type === "output_text")
    ?.map(part => part.text)
    ?.join("\n")
);
console.log("usage:", json.usage);

If you have ability to allocate some temporarily token with e.g. 1$ limit on it and send on email to ypechorkin@devforth.io then we can check it all faster and try to perform some adoption

[osseonews]

OK, I'll send over later.