`providerMetadata` lost in client-side UIMessage pipeline — breaks Gemini 3 thoughtSignature on multi-turn tool calls

[houmark]

Describe the bug

PR #401 added providerMetadata to ToolCall and ToolCallStartEvent types, and correctly wired it through the server-side ToolCallManager (used by chat()) and the Gemini adapter (processStreamChunks → formatMessages).

However, the client-side pipeline (StreamProcessor → UIMessage → ModelMessage) does not carry providerMetadata at all. This means:

1. First request works: chat() on the server uses ToolCallManager, which preserves providerMetadata on ToolCall objects. The Gemini adapter reads toolCall.providerMetadata?.thoughtSignature in formatMessages() — all good.

2. Second request fails: The client sends the conversation history back to the server. The StreamProcessor processed the first response's TOOL_CALL_START events, but dropped providerMetadata when creating ToolCallPart UIMessage parts. When converting back to ModelMessage.toolCalls for the second request, providerMetadata is missing. The Gemini adapter's formatMessages() finds no thoughtSignature, and Gemini returns HTTP 400:

Function call is missing a thought_signature in functionCall parts.
This is required for tools to work correctly, and missing thought_signature
may lead to degraded model performance.

Root cause — 5 locations in @tanstack/ai core

#	File	Function/Type	Issue
1	types.ts	ToolCallPart	Missing providerMetadata field
2	stream/types.ts	InternalToolCallState	Missing providerMetadata field
3	stream/message-updaters.ts	updateToolCallPart()	Doesn't accept or store providerMetadata
4	stream/processor.ts	handleToolCallStartEvent()	Doesn't pass chunk.providerMetadata to updateToolCallPart() or InternalToolCallState
5	messages.ts	buildAssistantMessages()	Doesn't include providerMetadata on toolCalls when converting ToolCallPart → ModelMessage.toolCalls

Additionally, modelMessageToUIMessage() and getCompletedToolCalls() also drop providerMetadata.

Why this wasn't caught

The test in PR #401 (preserves thoughtSignature in functionCall parts when sending history back to Gemini) only tests the server-side chat() flow where ToolCallManager correctly handles providerMetadata. The client-side StreamProcessor → UIMessage → ModelMessage path is not tested.

The server-side chat() loop works fine for single-session server-only scenarios. The bug manifests in client-server architectures where:

• The client uses ChatClient / StreamProcessor to process SSE events into UIMessages
• On the next user message, UIMessages are converted to ModelMessages via uiMessageToModelMessages() and sent to the server
• The server passes these messages to chat() → adapter → Gemini API

Steps to reproduce

1. Use @tanstack/ai-gemini with gemini-3-flash-preview and thinking enabled
2. Send a message that triggers a tool call
3. First response succeeds (tool call + response)
4. Send a follow-up message
5. HTTP 400: "Function call is missing a thought_signature in functionCall parts"

Environment

• @tanstack/ai@0.9.1
• @tanstack/ai-gemini@0.8.4
• @tanstack/ai-client@0.7.4
• Model: gemini-3-flash-preview with thinkingConfig: { includeThoughts: true }
• Architecture: Client (ChatClient/StreamProcessor) ↔ Server (chat() + Gemini adapter)

Related

• Fixes the remaining gap from Gemini 3 Flash: Function call is missing a thought_signature in functionCall parts #216
• PR fix(ai-gemini): preserve thoughtSignature and fix tool call IDs for Gemini 3+ thinking models #401 fixed the server-side adapter but not the client-side pipeline