Runner.runAsync terminates prematurely after tool execution with Gemini 3 Flash (gemini-3-flash-preview)

[victorbash400]

Bug Report: Runner.runAsync terminates prematurely after tool execution with gemini-3-flash-preview

Environment

• Operating System: macOS Tahoe 26.2
• Library: @google/adk (Version 0.2.1)
• Node.js: >= 20.0.0
• google/genai version 1.32.0.

Description

When using gemini-3-flash-preview with the ADK Runner framework, the agent triggers a tool call (e.g., ls or read_file), and the Runner successfully executes the tool and injects the response. However, instead of performing a follow-up synthesis turn to process these results and respond to the user, the Gemini 3 Flash stream terminates immediately with finishReason: 'STOP'.

Comparison & Evidence

1. Gemini 3 Flash (Failure Case)

The model calls the tool, receives the output, and then the stream simply ends. No synthesis text is ever generated.

(Note: Screenshot shows the log ending with run_complete immediately after function_response without any following text content.)

2. Gemini 2.5 Flash (Success Case)

The model correctly continues the conversation after the tool result is injected, providing a summary of what it found.

(Note: Screenshot shows the model continuing with a text response after the tool completes.)

Observed Logs (Simplified)

// Gemini 3 Flash Log Trace
[stream] function_call ['ls']
[stream] tool_output received (403 chars)
[stream] event { hasContent: true, fc: [], fr: [], finishReason: 'STOP' } 
// ^ Model stops here instead of generating text describing the files.
[stream] run_complete { aborted: false, hasResponse: true, finishReason: 'STOP' }

Steps to Reproduce

1. Initialize LlmAgent with model: 'gemini-3-flash-preview'
2. Enable thinkingConfig
3. Give a prompt that requires a tool call (e.g., "list files in this directory")
4. Observe that the Runner.runAsync generator closes immediately after the function response is yielded, before any synthesis text is produced

Expected Behavior

The Runner should ensure that a follow-up turn is executed to allow the model to summarize or act upon the tool output, matching the behavior seen in gemini-2.5-flash.

[Varun-S10]

Hi @victorbash400, could you please provide more details regarding your setup? Specifically, are you using the Gemini API key or Vertex AI? Additionally, sharing your code would be very helpful for us to reproduce the issue.

[victorbash400]

I'm using the Gemini API key directly. Had a look at the adk code and identified issues with how ADK handles Gemini 3 models—specifically around tool calling loops and model version detection.

I've submitted PR #49 with fixes for:

• Preventing premature loop termination when Gemini 3 sends STOP after tool calls
• Adding proper gemini-3 model detection

Current status: Testing with a locally patched ADK version (code here). The patch improves stability, though occasional errors still occur—especially in the flash and lower reasoning pro models likely due to model-specific quirks or something i haven't caught yet