Short Answer
When Gemini 3 and later models (gemini-4-argon, gemini-3.8-flash, gemini-3.1-pro-preview, and so on) run multi-turn function calling, there are two kinds of 400 errors related to the thought signature (thoughtSignature), and they need different fixes:
If you call through the OpenAI-compatible
/v1/chat/completions endpoint, none of this applies to you: the gateway manages signatures, and standard tool_calls / tool messages are all you need.
Telling the Two Errors Apart
Missing signature
parts array. This is a hard rule enforced by the model provider for the Gemini 3 family: every functionCall part in the current turn must carry its signature. A missing one fails with 400 on every route, and retrying does not help.
Invalid or corrupted signature
Why This Happens
When tools are enabled, Gemini 3 models attach an encryptedthoughtSignature to the same part as each functionCall, recording the reasoning state at that step. On the next request the model needs it to continue the reasoning chain, so the provider requires it to be sent back unchanged.
Missing: the client rebuilt the part
The most common cause. A custom framework or middleware keeps only the
name and args of the functionCall when assembling history and drops the sibling thoughtSignature; or streaming chunks are reassembled without merging the signature in.Missing: history was compressed or rewritten
The client trimmed or summarized the context, or merged several parts into one, and the signature went with it.
Invalid: route changed mid-conversation
The first turns were served by route A and later turns landed on route B, which does not recognize A’s signatures. The switch can come from you changing group or key, or from automatic failover.
Invalid: model changed mid-conversation
Starting on gemini-3.8-flash and switching to gemini-4-argon within the same conversation. Signatures are not portable across models.
For parallel function calls the model attaches a signature only to the first
functionCall part; the others legitimately have none, so do not fabricate one. For sequential calls (several steps within one turn) every step carries a signature and every one must be returned.Fix 1: Missing Signature, Change How the Client Replays History
1
Put the whole model reply back verbatim
With the official
google-genai SDK, simply contents.append(response.candidates[0].content); do not extract only the function_call field. In hand-written REST, the model turn in the second request’s contents must look like this:thoughtSignature and functionCall are two fields of the same part, not two parts.2
Check whether a framework or middleware strips the field
With LangChain, a custom agent framework, or an OpenAI-to-Gemini translation layer, print the actual request body of the second turn and confirm
thoughtSignature is still there. The position N in the error is parts[N]; look there directly.3
Do not rely on some models letting it through
Models and routes differ in how strictly they enforce missing signatures: some let it pass, others return 400. If the same code has always worked on gemini-3.8-flash and breaks on gemini-4-argon, the signature was most likely never sent and was merely tolerated before. Returning signatures as the provider requires is stable on every model.
Fix 2: Invalid Signature, Resend or Start Over
1
Resend once as is
When the APIYI gateway receives a signature 400 it automatically retries on another route, so a plain resend often succeeds. No changes needed.
2
Still 400: strip signatures from history turns and resend
Remove the
thoughtSignature field from the parts of every role: "model" turn in the history, then resend. You lose the reasoning context of those turns and answer quality may dip slightly, but the conversation can continue. Only strip history turns, never the current one.3
Or start a new conversation
The simplest and most reliable option. In a new conversation every signature comes from the same route from the first turn on.
4
One model, one key per conversation
When you need to switch model, key, or group, start a new conversation first. Do not switch inside a conversation that already has tool-call history.
No Signature Handling on the OpenAI-Compatible Endpoint
When you call Gemini through/v1/chat/completions, tool_calls messages have no signature field; the gateway saves and restores it during translation. In our tests on gemini-4-argon, two-turn conversations passed for all three shapes: a single call, two parallel calls, and text plus a call. Just return tool_calls and tool messages the standard OpenAI way.
Clients such as Codex that default to
/v1/responses need to switch to the chat protocol, since Gemini models do not support that endpoint yet. See Gemini 4 Argon Text Generation.FAQ
Only argon throws this; flash has always been fine. Is argon broken?
Only argon throws this; flash has always been fine. Is argon broken?
No. Returning signatures is the provider’s requirement for the whole Gemini 3 family; some models’ routes are simply lenient about missing ones. Moving to a strictly validating model exposes the problem, but the root cause is still a client that never sent the signature.
Can I cache a signature and reuse it across conversations?
Can I cache a signature and reuse it across conversations?
No. A signature is valid only on the route and within the conversation context that produced it. Copying it into another conversation yields Invalid.
Where is the signature in a streaming response?
Where is the signature in a streaming response?
In the same part as the
functionCall, inside the chunk that carries it. When reassembling the stream, merge this field together with functionCall into the final model reply before putting it back into history.Do plain-text turns need signatures too?
Do plain-text turns need signatures too?
Not strictly. The provider hard-validates signatures only on
functionCall parts; a missing signature on a text turn does not 400, but keeping it helps reasoning continuity, so return the whole reply verbatim anyway.Related Docs
Gemini Function Calling Guide
Full two-turn example, including how to return signatures
Gemini 4 Argon Text Generation
Multi-turn tool-call notes and client setup
How Do I Fix Claude's Invalid Thinking Signature Error?
The same class of problem on the Claude side
Multi-Turn Conversation Guide
How each provider expects history to be returned