Gemini model integration issue with Agentica

[Funncy]

I created a simple project using Agentica and it worked as expected. However, after switching the model to Gemini 2.0 Flash, the following issue occurred when running a basic command.

code :

import { Agentica, IAgenticaVendor } from "@agentica/core";
import OpenAI from 'openai';
import typia from 'typia';
import { BashController } from './tools/bash';
import { FileReadController } from './tools/file_read';



async function main() {

  const agent = new Agentica({
    model: "gemini",
    vendor: {
      model: "gemini-2.0-flash",
      api: new OpenAI({
        apiKey: "",
        baseURL: "https://generativelanguage.googleapis.com/v1beta/openai/",
      }),
    } satisfies IAgenticaVendor,
    controllers: [   {
      protocol: 'class',
      name: 'bash',
      application: typia.llm.application<BashController, 'gemini'>(),
      execute: new BashController(),
    },
    {
      protocol: 'class',
      name: 'file_read',
      application: typia.llm.application<FileReadController, 'gemini'>(),
      execute: new FileReadController(),
    }
  ],
  });

  const result = await agent.conversate('안녕');

  console.log(result);
}

main();

result :

/Users/teddy/Development/flutter/projects/heroines/heroines/agent/node_modules/typia/src/internal/_assertGuard.ts:10
    else throw new TypeGuardError(props);
               ^
TypeGuardError: Error on json.assertParse(): invalid type on $input.id, expect to be string
    at Object._assertGuard (/Users/teddy/Development/flutter/projects/heroines/heroines/agent/node_modules/typia/src/internal/_assertGuard.ts:10:16)
    at _ao0 (/Users/teddy/Development/flutter/projects/heroines/heroines/agent/node_modules/@agentica/core/src/chatgpt/ChatGptCompletionMessageUtil.ts:18:5)
    at /Users/teddy/Development/flutter/projects/heroines/heroines/agent/node_modules/@agentica/core/lib/chatgpt/ChatGptCompletionMessageUtil.js:408:39
    at __assert (/Users/teddy/Development/flutter/projects/heroines/heroines/agent/node_modules/@agentica/core/lib/chatgpt/ChatGptCompletionMessageUtil.js:413:35)
    at /Users/teddy/Development/flutter/projects/heroines/heroines/agent/node_modules/@agentica/core/lib/chatgpt/ChatGptCompletionMessageUtil.js:416:44
    at Object.ChatGptCompletionMessageUtil.transformCompletionChunk (/Users/teddy/Development/flutter/projects/heroines/heroines/agent/node_modules/@agentica/core/lib/chatgpt/ChatGptCompletionMessageUtil.js:416:91)
    at /Users/teddy/Development/flutter/projects/heroines/heroines/agent/node_modules/@agentica/core/src/Agentica.ts:258:42
    at /Users/teddy/Development/flutter/projects/heroines/heroines/agent/node_modules/@agentica/core/src/internal/StreamUtil.ts:57:30
    at Generator.next (<anonymous>)
    at fulfilled (/Users/teddy/Development/flutter/projects/heroines/heroines/agent/node_modules/@agentica/core/lib/internal/StreamUtil.js:5:58) {
  method: 'json.assertParse',
  path: '$input.id',
  expected: 'string',
  value: undefined
}

[sunrabbit123]

Google Gemini does not emit the "id" property.

The ChatCompletionChunk should contain the "id" property.

I encountered this issue and received the following response:

{"choices":[{"delta":{"content":"I","role":"assistant"},"index":0}],"created":1742371939,"model":"gemini-2.0-flash","object":"chat.completion.chunk","usage":{"completion_tokens":0,"prompt_tokens":113,"total_tokens":113}}

https://github.com/openai/openai-node/blob/master/src/resources/chat/completions/completions.ts#L336-L387

[kakasoo]

@samchon

This seems to be an error caused by Google Model itself not responding to the specifications that they have defined. So, like setting tsconfig.json in Nestia, why don't you create an option that allows you to pass through even if there's an error, but only logging?

{
    ...
    "plugins": [
      { "transform": "typescript-transform-paths" },
      { "transform": "typia/lib/transform" },
      { 
        "transform": "@nestia/core/lib/transform",
        "validate": "validate",
        "stringify": "validate.log", // here
      },
    ]
    ...
}

[sunrabbit123]

Even if you remove that validation check through monkey-patching, issues will still arise in the tool_call section.

Based on my experiment, Gemini 2.0 Flash does not provide the index in the tool_calls property. This deviates from the OpenAI SDK's expected interface.

https://x.com/SunrabbitO84776/status/1902287632308015332

[samchon]

https://github.com/wrtnlabs/agentica/blob/168de1e9481a7982ccaba3ed98b4f8c7308307c8/packages/core/src/chatgpt/ChatGptCompletionMessageUtil.ts#L18

Gemini is not keeping the spec...