fix: prevent premature loop termination for Gemini 3 tool calls

[victorbash400]

Gemini 3 models send an explicit STOP finish reason after tool calls, causing runAsyncImpl to terminate before processing the tool response.

Changes:

• Track function responses in runAsyncImpl loop
• Continue loop if pending function response exists
• Include gemini-3 in isGemini2Model() for code executor compat

[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.

[dnlrbz]

it contradicts the function name isGemini2Model, it would be better to create a separate function for checking gemini-3

[victorbash400]

fixed

[kalenkevich]

It seems to be wrong as it will not break when there is final response but there is no any function response. It can lead to infinity loop as we are using unsafe while (true) practice.

I think we might need to find out another solution for that problem.

[victorbash400]

Thanks for the review ,based on my reading of the code, it looks like the loop does break when there is a final response and no function response. In that case, isFinalResponse(lastEvent) is true and hasFunctionResponse remains false, so the condition evaluates to true && true and exits the loop.

Basically this:

while (true) {
  let hasFunctionResponse = false;  // resets each iteration
  
  for await (const event of this.runOneStepAsync(context)) {
    if (getFunctionResponses(event).length > 0) {
      hasFunctionResponse = true;
    }
  }
  
  if (!lastEvent || (isFinalResponse(lastEvent) && !hasFunctionResponse)) {
    break;  // line 1512
  }
}

The hasFunctionResponse flag resets to false at the start of each iteration. As we process events from runOneStepAsync, if we encounter any function responses, we set it to true. After processing all events, we check whether we received a final response without any function responses. If so, we break. The only case where we continue looping is when hasFunctionResponse is true, meaning we just processed a tool response and need to send it back to the model for another turn.

Note the while(true) pattern was already present in the original code:

while (true) {
  // ...
  if (!lastEvent || isFinalResponse(lastEvent)) {
    break;
  }
}

I only modified the break condition, not the loop structure .

[DiegoLerma]

Confirming this issue - same behavior observed

I'm experiencing the exact same problem with gemini-3-flash-preview. The Runner.runAsync terminates immediately after receiving the tool response with finishReason: 'STOP', preventing the model from generating a synthesis response.

Environment:

• OS: Ubuntu 24.04.1 LTS (Kernel 6.14.0-37-generic, x86_64)
• Node.js: v20.18.3
• Package Manager: pnpm 10.17.0
• TypeScript: 5.8.3

Dependencies:
"@google/adk": "0.2.1",
"@google/genai": "1.34.0"

Use Case:
Multi-agent A2A flow where a MainAgent delegates to a SearchAgent (wrapped with AgentTool). The SearchAgent executes GOOGLE_SEARCH, returns results successfully, but the MainAgent never generates a follow-up response.

SSE Event Flow Observed:

phase: thinking → agent_start → search_start → phase: searching → tool_call
→ agent_complete → phase: analyzing → tool_result → done (NO content event)

The tool_result contains complete data from the sub-agent, but the stream terminates without any content event from the parent agent.

Workaround:

Switching to gemini-2.5-flash resolves the issue - the model correctly continues generating responses after tool execution.

Looking forward to this fix being merged. Thanks @victorbash400 for the PR!