[Gemini 3 Flash Preview] Inconsistent thought_signature generation in parallel function calls causes 400 errors and potential silent data loss – Gemini API

Gemini 3 Flash Preview (gemini-3-flash-preview) exhibits a critical bug where it inconsistently generates thought_signature fields for parallel function calls. This causes:

1. 400 INVALID_ARGUMENT errors when returning function results
2. Position-based, non-deterministic signature generation
3. Makes Gemini 3 Flash unsuitable for production use with multi-tool scenarios

• Model: gemini-3-flash-preview
• API: Vertex AI (Node.js SDK)
• Region: global
• SDK: @google/genai v1.38.0
• Date Tested: January 23, 2026
• Baseline: Same code works flawlessly with gemini-2.5-flash

Gemini 3 Flash Preview inconsistently generates thought_signature fields in multi-tool responses:

• Only the first 1-2 function calls receive signatures
• Remaining function calls have NO signature field (hasThoughtSignatureField: false)
• The cutoff point is non-deterministic (varies between requests)
• Subsequent API requests fail with: 400: INVALID_ARGUMENT - "Unable to submit request because function call in the X. content block is missing a thought_signature"

We implemented comprehensive debug logging that captures the raw API response directly from Google’s SDK, proving this is an API-level bug, not a client implementation issue.

Prompt: User asks for comprehensive report requiring multiple data sources

Google’s Raw Response (3 parallel tool calls):

{
  "toolCalls": [
    {
      "toolName": "get_resource_list",
      "hasThoughtSignatureField": true,
      "thoughtSignatureValue": "CiEBjz1rX0ilumvPg6k0sMMUtISoQCLJv2bPJ7Vv9r+Xw0o=",
      "thoughtSignatureLength": 48
    },
    {
      "toolName": "get_detailed_results", 
      "hasThoughtSignatureField": true,
      "thoughtSignatureValue": "CiEBjz1rX0ilumvPg6k0sMMUtISoQCLJv2bPJ7Vv9r+Xw0o=",
      "thoughtSignatureLength": 48
    },
    {
      "toolName": "analyze_data",
      "hasThoughtSignatureField": false,
      "thoughtSignatureValue": undefined
    }
  ]
}

Result: Position 1-2 have signatures | Position 3 missing → 400 error when returning position 3 result

Error Message:

Unable to submit request because function call `analyze_data` 
in the 6. content block is missing a `thought_signature`.

Prompt: User explicitly requests multiple operations in sequence

Google’s Raw Response (3 parallel tool calls):

{
  "toolCalls": [
    {
      "toolName": "analyze_data",
      "hasThoughtSignatureField": true,
      "thoughtSignatureValue": "CiEBjz1rX0ilumvPg6k0sMMUtISoQCLJv2bPJ7Vv9r+Xw0o=",
      "thoughtSignatureLength": 48
    },
    {
      "toolName": "get_detailed_results",
      "hasThoughtSignatureField": false,
      "thoughtSignatureValue": undefined
    },
    {
      "toolName": "get_resource_list",
      "hasThoughtSignatureField": false,
      "thoughtSignatureValue": undefined
    }
  ]
}

Result: Position 1 has signature | Position 2-3 missing → 400 error when returning position 2 result

Error Message:

Unable to submit request because function call `get_detailed_results` 
in the 4. content block is missing a `thought_signature`.

Critical observation: The tool get_resource_list:

• HAD signature in Test Case 1 (position 1)
• NO signature in Test Case 2 (position 3)

This proves the bug is position-based, not related to specific tools.

Conclusion: Gemini 3 Flash generates signatures for approximately the first 1-2 tool calls, then stops. The cutoff is non-deterministic.

• Scope: Any application using Gemini 3 Flash with 3+ parallel function calls
• Frequency: Occurs consistently when 3+ tools are called in parallel
• Reliability: Non-deterministic cutoff makes it impossible to predict which calls will fail
• Production Impact: Cannot reliably use Gemini 3 Flash for multi-tool scenarios

1. Configure Vertex AI with gemini-3-flash-preview model (global region)
2. Define 3 or more tool/function declarations
3. Send a prompt that triggers 3+ parallel function calls
4. Capture raw API response from generateContentStream
5. Observe: Only first 1-2 tool calls will have thought_signature field
6. Return all tool results (with signatures where provided)
7. Observe: 400 INVALID_ARGUMENT error for tool calls missing signatures

To rule out implementation issues, we:

Properly capture thought signatures from streaming responses
Correctly preserve signatures through tool execution lifecycle
Accurately return signatures byte-for-byte to the API
Added extensive debug logging at every step of the flow
Confirmed with raw API response logs that Google simply doesn’t send signatures for all tool calls

The same implementation works flawlessly with Gemini 2.5 Flash, proving this is a Gemini 3 API issue.

Evidence from our production logs showing the raw Google response:

[llm] 2026-01-23 16:19:40 [debug]: [RAW GOOGLE RESPONSE] Tool call received
{"toolName":"analyze_data","hasThoughtSignatureField":true,
 "thoughtSignatureValue":"CiEBjz1rX0ilumvPg6k0sMMUtISoQCLJv2bPJ7Vv9r+Xw0o="}

[llm] 2026-01-23 16:19:40 [debug]: [RAW GOOGLE RESPONSE] Tool call received
{"toolName":"get_detailed_results","hasThoughtSignatureField":false}

[llm] 2026-01-23 16:19:41 [debug]: [RAW GOOGLE RESPONSE] Tool call received
{"toolName":"get_resource_list","hasThoughtSignatureField":false}

These logs capture the response directly from the SDK, before any application processing.

Use gemini-2.5-flash instead.

This model:

• Doesn’t use thought signatures
• Handles parallel function calls reliably
• Produces consistent, predictable results

1. Why are thought signatures inconsistently generated for parallel function calls?
2. Is this a known limitation of the preview API?
3. What is the intended maximum number of parallel function calls with thought signatures?
4. Is there a timeline for fixing this in Gemini 3 Flash?
5. Should Gemini 3 Flash be considered production-ready for function calling scenarios?

Please fix Gemini 3 Flash Preview to consistently generate thought_signature for ALL function calls in a parallel response, or clearly document any limitations on the number of parallel tool calls supported.

We have extensive debug logs, raw API responses, and reproduction code available. Happy to provide additional evidence or collaborate on debugging.

Environment: Enterprise production system with comprehensive monitoring and logging infrastructure.

This bug makes Gemini 3 Flash unsuitable for production use with multi-tool function calling. We’re staying on Gemini 2.5 Flash until this is resolved.