fix: retry when model returns empty response after tool execution

[akashbangad]

Bug

Some models (notably Claude, and some Gemini preview models) return an empty content array (parts: []) after processing tool results. ADK's is_final_response() treats this as a valid completed turn because it only checks for the absence of function calls — not the presence of actual content. The agent loop stops and the user sees nothing.

Observed with:

• Claude (Opus/Sonnet/Haiku) via AnthropicLlm — after run_shell, computer_use tool results
• Gemini preview models — after tool execution with streaming enabled

Example session history showing the bug:

Event 19: agent calls run_shell({"command": "cloudflared --version"})
Event 20: tool responds: {"output": "cloudflared version 2026.3.0", "exit_code": 0}
Event 21: agent responds with parts: [] ← EMPTY, agent loop ends, user sees nothing

Root Cause

In BaseLlmFlow.run_async() (line 757):

if not last_event or last_event.is_final_response() or last_event.partial:
    break

And is_final_response() in event.py:

return (
    not self.get_function_calls()
    and not self.get_function_responses()
    and not self.partial
    and not self.has_trailing_code_execution_result()
)

An event with parts: [] passes all these checks — no function calls, no function responses, not partial — so is_final_response() returns True and the loop breaks.

Fix

Added a retry mechanism in BaseLlmFlow.run_async():

1. _has_meaningful_content(event) — helper that checks if an event actually contains content worth showing (non-empty text, function calls, inline data, etc.)
2. When is_final_response() is True but the event has no meaningful content, the loop continues instead of breaking, re-prompting the model
3. A maximum retry count (_MAX_EMPTY_RESPONSE_RETRIES = 2) prevents infinite loops if the model keeps returning empty responses

Tests

Added 10 new tests in test_empty_response_retry.py:

_has_meaningful_content tests (7):

• test_no_content — None content → not meaningful
• test_empty_parts — parts: [] → not meaningful
• test_only_empty_text_part — text="" → not meaningful
• test_only_whitespace_text_part — text=" \n " → not meaningful
• test_non_empty_text — actual text → meaningful
• test_function_call — function call → meaningful
• test_function_response — function response → meaningful

Integration tests (3):

• test_empty_response_retried_then_succeeds — empty response triggers retry, second call succeeds
• test_empty_response_stops_after_max_retries — stops after max retries to prevent infinite loop
• test_non_empty_response_not_retried — normal responses are not retried

All 10 tests pass. All 356 pre-existing flows/llm_flows/ tests pass.

pytest tests/unittests/flows/llm_flows/test_empty_response_retry.py -v

Closes #3754
Related: #3467, #4090, #3034

🤖 Generated with Claude Code