`gemini-2.5-flash`: `finish_reason=MALFORMED_FUNCTION_CALL` when using streaming + tool definitions + dynamic thinking (no explicit `thinking_config`)

[bhakta0007]

Environment details

• Programming language: Python
• OS: macOS 26.2
• Language runtime version: Python 3.12.8
• Package version: google-genai==1.57.0

Steps to reproduce

1. Setup a new venv with python3, pip install google-genai==1.57.0
2. Run the attached script with: python3 gemini_malformed_reproducer.py --path1 --runs 20 (try with more runs if you cannot repro with 20)

Repro logs:

<snip>
======================================================================
[Run 7/20]  PATH 2: Native Google GenAI SDK — streaming + tools + thinking + MALFORMED retry
======================================================================
  model          : gemini-2.5-flash
  thinking_budget: 8192
  query          : 'give me a line chart of vulnerabilities over the last 4 weeks'

Both GOOGLE_API_KEY and GEMINI_API_KEY are set. Using GOOGLE_API_KEY.
Attempt 1 — dynamic thinking (no explicit budget, matches Torana):
--------------------------------------------------
  [chunk   1]  finish_reason = 'MALFORMED_FUNCTION_CALL'
--------------------------------------------------
  chunks received : 1
  finish_reason   : 'MALFORMED_FUNCTION_CALL'
  tool calls      : 0

⚠️  MALFORMED_FUNCTION_CALL detected on attempt 1
   → Retrying with thinking_budget=0 (CachingGemini workaround)

Attempt 2 — retry with thinking_budget=0:
--------------------------------------------------
  [chunk   1]  finish_reason = 'STOP'
  [chunk   1]  tool_call = 'ui_line_chart'
--------------------------------------------------
  chunks received : 1
  finish_reason   : 'STOP'
  tool calls      : 1

✅  RETRY SUCCEEDED: 'ui_line_chart' called with 5 top-level args
   labels   : ['Week 1', 'Week 2', 'Week 3', 'Week 4']
   datasets : 4 series
     - 'Critical': [120, 110, 90, 80]
     - 'High': [250, 220, 180, 150]
     - 'Medium': [400, 380, 350, 320]
     - 'Low': [600, 550, 500, 450]
   title    : 'Vulnerabilities Over Last 4 Weeks'

  [fail-fast] Bug reproduced — stopping. Use --no-fail-fast to continue.

======================================================================
STATISTICS  (0 runs)
======================================================================

  PATH 2 — Native SDK (with retry):
    Attempt 1 MALFORMED      :   1 / 7  (14%)
      → retry succeeded      :   1 / 1  (100%)
    Success on attempt 1     :   6 / 7  (86%)
    Overall success          :   7 / 7  (100%)

Note: i have also raised a bug with litellm to handle this scenario better: BerriAI/litellm#21744

gemini_malformed_reproducer.py

[Venkaiahbabuneelam]

Hi @bhakta0007 , Thanks for reaching out us!

To help us understand the issue, could you please share a code snippet of what you've tried so far? If you have a link to the relevant documentation, that would also be very helpful. Thanks!

[bhakta0007]

@Venkaiahbabuneelam

were you able to repro with the attached gemini_malformed_reproducer.py (attached to this ticket)?

[Venkaiahbabuneelam]

Hi @bhakta0007
Thank you for your feedback. We appreciate you taking the time to share your thoughts with us,

Yes, we tried to reproduce using gemini_malformed_reproducer.py script below are the

Relative Logs:
"╔══════════════════════════════════════════════════════════════════════╗
║ Gemini 2.5-flash MALFORMED_FUNCTION_CALL Bug Reproducer ║
╚══════════════════════════════════════════════════════════════════════╝
Run at : 2026-03-16T12:09:22
Model : gemini-2.5-flash
Thinking budget: 8192 tokens (forced — required to trigger bug)
Tools : 7 widget tools (line_chart, bar_chart, ...)
Query : 'give me a line chart of vulnerabilities over the last 4 weeks'
Runs : 1
Fail-fast : yes (stop on first bug hit)

NOTE: Bug is non-deterministic. Use --runs 10 to collect statistics.

======================================================================
PATH 1: LiteLLM (gemini/gemini-2.5-flash) — streaming + tools + thinking, NO retry

model : gemini-2.5-flash
thinking_budget: 8192
query : 'give me a line chart of vulnerabilities over the last 4 weeks'

Streaming (showing notable chunks only):

[chunk 2] finish_reason = 'stop'

RESULTS:
chunks received : 2
finish_reason : 'stop'
tool calls : 0
text response : ''

⚠️ BUG CONFIRMED (hidden by LiteLLM): finish_reason='stop', empty response, no tool.
LiteLLM normalised MALFORMED_FUNCTION_CALL → 'stop' and dropped the call.
The chart was never rendered — same user-visible outcome as explicit MALFORMED.

[fail-fast] Bug reproduced — stopping. Use --no-fail-fast to continue.

======================================================================
SUMMARY
PATH 1 (LiteLLM, GEMINI_NATIVE_BACKEND=false):
- No retry → MALFORMED means the tool call is lost
- Chart is never rendered; agent turn appears to succeed silently

PATH 2 (Native SDK, GEMINI_NATIVE_BACKEND=true, default):
- Detects MALFORMED_FUNCTION_CALL from finish_reason
- Retries the SAME turn with thinking_budget=0
- Retry succeeds → ui_line_chart called → chart rendered

Use --runs 10 to see bug reproduction rates across multiple turns.

BUG REPORT:
LiteLLM: https://github.com/BerriAI/litellm/issues
Include this script + output + package versions:

litellm     : 1.82.2
google-genai: 1.66.0"

---

I can see that bug you got fix for the same, Please let us know if you need any further assistance on this.

[bhakta0007]

@Venkaiahbabuneelam what is the reason for the gemini issue. Litellm bug i raised is a different issue where it was not capturing the correct failure reason

[Venkaiahbabuneelam]

Hey @bhakta0007,
Based on the logs and our reproduction, the reason for the issue is a formatting conflict between the "Thinking" block and the "Tool Call" block when streaming is enabled.

Thanks

[bhakta0007]

@Venkaiahbabuneelam - thank you for the insight. Is this something that will be fixed soon? Is there a workaround that I can use?

[bhakta0007]

@Venkaiahbabuneelam - do you have a timeline for the fix? is there a workaround?

[github-actions]

This issue has been marked as stale because it has been open for 7 days with no activity. It will be closed in 2 days if no further activity occurs.