fix: retry with resume message when model returns empty response

[gurjot-05]

Bug

Some models (notably Gemini 2.5 Pro/Flash) intermittently return empty content (parts: [], candidatesTokenCount: 0, finishReason: STOP) after processing tool results. This is especially common under concurrent load and with streaming + thinking enabled.

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:

• Gemini 2.5 Pro via vertex_ai — after tool execution in orchestrator agents
• Intermittent: ~15-29% failure rate under parallel load (7+ concurrent sessions)
• Affects both streaming (SSE + thoughts) and non-streaming modes

Example session showing the bug:

Event: agent calls nth_research → completes successfully
Event: agent calls review_research → completes successfully (status: "approved")
Event: agent should call draft_response next, but returns parts: [] ← EMPTY
Agent loop ends. User sees nothing.

Related: #3525

Root Cause

There are three distinct failure modes, all leading to the same silent halt:

Mode 1: Non-streaming empty response

In BaseLlmFlow.run_async():

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

An event with parts: [] passes is_final_response() — no function calls, no function responses, not partial — so the loop breaks silently.

Mode 2: Streaming + thinking, no final response yielded

With streaming + thinking enabled, the LiteLLM adapter yields thought chunks as partial=True. When finish_reason=stop arrives with no text content and no accumulated text/reasoning:

• lite_llm.py finalization requires (text or reasoning_parts) which is False
• No final aggregated response is ever yielded from the streaming generator
• last_event is either None or partial=True from a thought chunk
• The loop breaks on not last_event or last_event.partial — before any retry logic

Mode 3: Model returns content=None

When the model returns content=None, _postprocess_async filters it out (returns without yielding if no content and no error). last_event stays None, and the loop breaks.

Fix

Two-layer fix addressing the root cause and adding defense-in-depth:

Layer 1: lite_llm.py — Ensure streaming always yields a final response

Added a fallback after the streaming loop: if the model produced no meaningful output at all (no text, no reasoning, no tool calls), yield an explicit empty non-partial LlmResponse so downstream retry logic can detect and handle it.

This converts Mode 2 into Mode 1, making it catchable by the retry logic.

Layer 2: base_llm_flow.py — Unified retry for all empty response modes

Restructured run_async() to check for empty responses before normal termination conditions. The retry now handles all three cases:

1. not last_event — no events yielded at all (Mode 3)
2. last_event.partial with no meaningful content — streaming thought-only chunks, no final response (Mode 2, defense-in-depth)
3. last_event.is_final_response() with no meaningful content — non-streaming empty response (Mode 1)

When an empty response is detected:

• A resume message is injected: "Your previous response was empty. Please resume execution from where you left off."
• This gives the model context about why it's being re-prompted, improving recovery rate vs. silent retry
• _MAX_EMPTY_RESPONSE_RETRIES = 2 prevents infinite loops

False positive prevention

• Partial events with real content (normal streaming): _has_meaningful_content() returns True — no retry
• Agent transfer events: Author check (last_event.author == agent.name) prevents retrying events from other agents
• Thought-only parts: Treated as non-meaningful (they have no user-visible content)

Tests

Existing tests: 384 passed (0 failures)

PR tests (test_empty_response_retry.py): 12 passed

Comprehensive scenario tests (test_empty_response_all_scenarios.py): 12 passed

#	Scenario	Expected	Result
1	Non-streaming empty parts:[] then recovery	1 retry, good response	PASS
2	Thought-only final response	1 retry	PASS
3a	No events at all (last_event=None)	Retries up to max	PASS
3b	Partial + empty content	1 retry, then recovery	PASS
3b'	Partial + thought-only content	1 retry, then recovery	PASS
4	Partial WITH real text content	No retry (normal break)	PASS
5	Empty exhausts max retries	Exactly 2 retries, then stop	PASS
6	Empty, Empty, Good (multi-retry recovery)	2 retries, then good	PASS
7	LiteLLM fallback response is non-partial	is_final_response()=True	PASS
8	Whitespace-only text	1 retry	PASS
9	Function call is meaningful	Not retried	PASS
10	Partial empty then partial with content	1 retry, then normal break	PASS

Test plan

• pytest tests/unittests/flows/llm_flows/ — 384 passed, 0 failed
• All 12 original PR tests pass
• All 12 comprehensive scenario tests pass
• Agent transfer tests not affected (author check prevents false positives)
• Verified with production logs from real orchestrator agent running Gemini 2.5 Pro — both non-streaming and streaming+thinking failure modes are covered

[adk-bot]

Response from ADK Triaging Agent

Hello @gurjot-05, thank you for your contribution!

It looks like the Contributor License Agreement (CLA) check has failed. Before we can merge this PR, you will need to sign the CLA. You can do so by following the instructions at https://cla.developers.google.com/.

Signing the CLA is a one-time process and is required for all contributions. Thanks!

[rohityan]

Hi @gurjot-05 , Thank you for your contribution! We appreciate you taking the time to submit this pull request. Your PR has been received by the team and is currently under review. We will provide feedback as soon as we have an update to share.

[rohityan]

Hi @GWeale , can you please review this.

[gurjot-05]

Thank you @rohityan I look forward to your feedback. Let me know if there’s anything else I can provide in the meantime.

[pras2309]

@gurjot-05, please rebase the branch.