Multi-turn sessions return empty response after first tool call

[vcorr]

Multi-turn conversations break immediately after the first tool execution in a session when using the TypeScript ADK. The agent returns an empty response for the second turn (no text content), and subsequent turns fail with API errors (400) because the conversation history contains a model message with empty parts.

I have verified that this issue does not exist in the Python ADK (v1.21.0), which handles the exact same agent logic and conversation flow correctly.

To Reproduce
Steps to reproduce the behavior:

1. Initialize an LlmAgent with at least one tool (e.g., FunctionTool) and InMemorySessionService.
2. Turn 1: Send a user message that triggers a tool call (e.g., "What is the shipping policy?" -> calls searchKnowledge).
   • Result: Agent correctly calls the tool and returns the answer. (✅ Works)
3. Turn 2: Send a follow-up user message in the same session (e.g., "What is 2+2?").
   • Result: The runner.runAsync loop completes, but the final model response contains no text parts. (❌ Fails - Empty Response)
4. Turn 3: Send another user message.
   • Result: The API request fails with error 400 because the history now includes the empty model response from Turn 2.

Expected behavior
The agent should continue the conversation normally after a tool call, maintaining context and providing a text response to subsequent user messages, just as the Python ADK does.

Screenshots
N/A (Console output confirms empty response)

Desktop (please complete the following information):

• OS: macOS 14.6

• TS version/environment: Node.js v18+, TypeScript 5.x

• ADK version: @google/adk v0.2.0 (npm)

• Model used: gemini-2.0-flash (Vertex AI)

[Varun-S10]

Hi @vcorr,
We have tried this, but we are unable to reproduce the issue on our side.
Could you please share more details or relevant code so that we can reproduce the issue?

[vcorr]

Here is the source code for the test agent:

import { FunctionTool, LlmAgent } from '@google/adk';
import { z } from 'zod';

const getCurrentTime = new FunctionTool({
  name: 'get_current_time',
  description: 'Returns the current date and time. Use when the user asks what time or date it is.',
  parameters: z.object({}),
  execute: () => {
    return { 
      current_time: new Date().toISOString(),
      status: 'success',
    };
  },
})

export const rootAgent = new LlmAgent({
  name: 'tool_agent',
  model: 'gemini-3-flash-preview',
  description: 'Tool agent',
  instruction: `You are a helpful agent that can use following tools
  - current time`,
  tools: [getCurrentTime]
});

This works fine with first invocation and returns the current time correctly but after that the agent is unable to answer anything and shows the error:
"Unable to submit request because it must include at least one parts field, which describes the prompt input. Learn more: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/gemini"

I also updated to the latest 0.2.1, but it behaves the same way.

[Varun-S10]

Hi @vcorr,
The code you provided is working well on our end using these versions: @google/adk (^0.2.1), @google/adk-devtools (^0.2.1), and typescript (^5.9.3). Could you please try it using the Gemini secret key? You can find the instructions in the documentation here. Please let us know if you encounter any further issues.

[vcorr]

I can confirm that it indeed works with GEMINI_API_KEY!
For some reason this reproduces only with GOOGLE_GENAI_USE_VERTEXAI=TRUE. If you set that to FALSE and supply the API KEY, then it works. Any change to get this fixed for Vertex auth too?

[Varun-S10]

Hi @vcorr, we are in the process of configuring Vertex AI to replicate this in adk-js. We will let you know once we have completed this step.