[BUG] CodeExecution emits a stray usage-only event that ends the CodeAct loop (already reported upstream)

[SuperCorleone]

Describe the Bug

After executing code, runPostProcessor clears the consumed model response's content but leaves its usageMetadata set. The flow then emits an extra content-less, usage-only event after the codeExecutionResult. BaseLlmFlow checks only the last event of the step, treats the usage-only event as final=true, and stops the loop — so the model is never called again with the code-execution result, breaking the CodeAct loop (model emits code → execute → emit result → should call model again for the final answer).

Root Cause

// CodeExecution.runPostProcessor, L299
llmResponseBuilder.content(Optional.empty());   // clears content but NOT usageMetadata

The builder (still carrying usageMetadata) is later built/emitted (around L160-166), producing the stray usage-only event.

Expected vs Observed

Expected:  executableCode → codeExecutionResult (final=false) → [next LLM call] → final answer (final=true)
Observed:  executableCode → codeExecutionResult (final=false) → usage-only event (final=true)  ← loop stops

Suggested Fix

Also clear usageMetadata when clearing content (or do not emit an event when content is empty), so the post-code response does not present as a final event.

Note

This matches a known upstream report (custom/non-built-in BaseCodeExecutor, streaming/SSE, provider that returns usageMetadata). Search github.com/google/adk-java before any action; contribute a Java reproduction to the existing issue rather than opening a duplicate.

Corresponding Test (generated)

The generated tests for this method exercise processResponse (the post-processor) but crash on a mocked LlmResponse.toBuilder() returning null, so they do not themselves assert the usageMetadata defect — which is why our triage initially dismissed them. Included verbatim for completeness; the defect is confirmed from source + the upstream issue, not from this test:

@Test
public void testProcessResponse_WithCodeExecutorAndEvents_ReturnsMappedResultWithEvents() {
    // Arrange
    Content content = Content.builder().build();
    LlmResponse mockResponse = mock(LlmResponse.class);
    when(mockResponse.partial()).thenReturn(Optional.of(false));
    when(mockResponse.content()).thenReturn(Optional.of(content));
    when(invocationContext.agent()).thenReturn(agent);
    when(agent.codeExecutor()).thenReturn(codeExecutor);
    when(agent.name()).thenReturn("test-agent");

    // Mock execution result to return events
    Event mockEvent = mock(Event.class);
    Flowable<Event> eventFlowable = Flowable.just(mockEvent);

    // Mock the code execution result
    when(codeExecutor.executeCode(any(), any())).thenReturn(executionResult);
    when(executionResult.stdout()).thenReturn("stdout");
    when(executionResult.stderr()).thenReturn("");
    when(executionResult.outputFiles()).thenReturn(ImmutableList.of());

    // Mock session to avoid null pointer
    ConcurrentHashMap<String, Object> sessionState = new ConcurrentHashMap<>();
    when(session.state()).thenReturn(sessionState);

    // Act
    Single<ResponseProcessor.ResponseProcessingResult> result =
        processor.processResponse(invocationContext, mockResponse);

    // Assert
    ResponseProcessor.ResponseProcessingResult processedResult = result.blockingGet();
    assertNotNull(processedResult);
    assertEquals(mockResponse, processedResult.updatedResponse());
    assertTrue(processedResult.events().iterator().hasNext());
    assertFalse(processedResult.transferToAgent().isPresent());
}

Environment

• Project: adk / google-adk (com.google.adk) · upstream: github.com/google/adk-java
• Version: 0.5.1-SNAPSHOT (local; upstream issue references 1.3.0) · Commit: <fill SUT commit>
• Java 17 · OS macOS

This input was generated by the test case generator TestFusion developed in our STAR lab.

[SuperCorleone]

Repeated