fix: propagate thoughtSignature during streaming

[RobinClowers]

Link to Issue or Description of Change

• Closes: fix: prevent premature loop termination for Gemini 3 tool calls #49

Gemini thinking models only provide thoughtSignature on the first functionCall part in a streaming turn. When the model issues multiple concurrent function calls, subsequent parts lack the signature. The API requires matching thoughtSignature values on all function response parts, causing 400 errors when the ADK sends responses back.

Problem:
Gemini thinking models only provide thoughtSignature on the first functionCall part in a streaming turn. When the model issues multiple concurrent function calls, subsequent parts lack the signature. The API requires matching thoughtSignature values on all function response parts, causing 400 errors when the ADK sends responses back.

Solution:
Save the first thoughtSignature seen on a function call part and inject it into any subsequent function call parts that are missing it.

Testing Plan

We are using this patch (applied at runtime) in ADK in production today.

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally.

 Test Files  104 passed | 11 skipped (115)
      Tests  1093 passed | 21 skipped (1114)

Manual End-to-End (E2E) Tests:

Setup: Create an agent with 2+ simple tools and a prompt that encourages parallel calls:

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

const getWeather = new FunctionTool({
  name: 'get_weather',
  description: 'Get current weather for a city',
  inputSchema: z.object({ city: z.string() }),
  execute: async ({ city }) => ({ city, temp: '72F', condition: 'sunny' }),
});

const getTime = new FunctionTool({
  name: 'get_time',
  description: 'Get current time in a city',
  inputSchema: z.object({ city: z.string() }),
  execute: async ({ city }) => ({ city, time: '2:30 PM' }),
});

const agent = new LlmAgent({
  name: 'test_agent',
  model: 'gemini-3-flash-preview', // must be a thinking model
  instruction: 'You help with travel planning. Use tools in parallel when possible.',
  tools: [getWeather, getTime],
});

const runner = new InMemoryRunner({ agent, appName: 'test' });
const session = await runner.sessionService.createSession({
  appName: 'test',
  userId: 'test',
});

for await (const event of runner.runAsync({
  userId: 'test',
  sessionId: session.id,
  newMessage: createUserContent(
    'What is the weather and time in Seattle, Tokyo, and London?'
  ),
})) {
  console.log(event.author, JSON.stringify(event.content?.parts?.map(p =>
    p.functionCall ? `call:${p.functionCall.name}(sig:${!!p.thoughtSignature})` :
    p.functionResponse ? `resp:${p.functionResponse.name}` :
    p.text ? `text:${p.text.slice(0, 60)}` : '?'
  )));
}

What to look for:

• Without the fix: The run fails with a 400 error on the second model call (after function responses are sent back). You'll see the function calls go out, the responses come back, then a crash.
• With the fix: The run completes. In the logged output, all function call parts should show sig:true. The model produces a final text response summarizing weather and time.

Key conditions to reproduce:

• Must use a thinking model (gemini-3-flash-preview, gemini-3-pro-preview)
• Must use streaming (which is the default for Runner)
• The prompt must trigger 2+ concurrent function calls in a single turn — asking about multiple cities with multiple tools is a reliable way to get this
• It's somewhat nondeterministic — the model might choose sequential calls instead. The multi-city prompt usually works but you may need to retry

Checklist

• I have read the CONTRIBUTING.md document.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end.
• Any dependent changes have been merged and published in downstream modules.

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[ScottMansfield]

Thank you for the patch! I will take a look. Meanwhile, the CLA is needed for any contributions, please look at filling it out. It might be a company one for you if this is on behalf of your company.

https://cla.developers.google.com/

[RobinClowers]

@ScottMansfield my org already had one, I amended my commit to use my work email.

[ScottMansfield]

all done, thank you!