Bug Description

On n8n 2.21.7, using gemini-3.5-flash via the Vertex AI Chat Model node (@n8n/n8n-nodes-langchain.lmChatGoogleVertex) inside an AI Agent fails with 400 Bad Request from Vertex on the second turn — i.e. when n8n posts the tool result back to the model after a successful function call.

The underlying Vertex error is the familiar Function call is missing a thought_signature. This appears related to #23798 / #24889 and the fixes shipped in #24473 (single tool, n8n 2.6.0) and #24982 (parallel tool calls, n8n 2.7.0), but my scenario is sequential multi-turn with gemini-3.5-flash, which I don't believe either PR covered.

First request (model → tool call) works correctly. Second request (tool result → model continuation) fails with 400.

Swapping the model to gemini-2.5-flash in the same workflow works end-to-end, isolating the issue to the gemini-3.x thinking models + Vertex provider in n8n.

When n8n builds the follow-up contents array, the previous assistant message containing the functionCall part appears to be reconstructed without the thoughtSignature field that Vertex issued. Vertex requires every functionCall part in the conversation history to carry its original signature, not just those produced in parallel within one response.

To Reproduce

1. Create a workflow with an AI Agent (@n8n/n8n-nodes-langchain.agent, typeVersion 1.8).
2. Connect a Google Vertex Chat Model (lmChatGoogleVertex, typeVersion 1) with:
   • modelName: gemini-3.5-flash
   • options.temperature: 0.3
   • options.thinkingBudget: 0
3. Connect at least one Tool Workflow (toolWorkflow, typeVersion 2.1) to the AI Agent.
4. (Optional) Connect a MongoDB Chat Memory and a Structured Output Parser.
5. Send a user message that causes the model to invoke the tool.
6. Observe: the tool runs successfully and returns data.
7. Observe: the next request (tool result → model) fails with HTTP 400.

Switching the chat model to gemini-2.5-flash (same workflow, nothing else changed) makes the agent complete the turn without errors.

Expected behavior

After a tool returns, n8n should be able to send the tool result back to Vertex and receive the model's continuation, completing the agent turn. The thoughtSignature from the previous assistant functionCall part should be preserved when constructing the follow-up request so Vertex's validation passes — the same behavior already achieved for parallel tool calls in #24982, but applied to the sequential multi-turn case as well.

Debug Info

Debug info

core

• n8nVersion: 2.21.7
• platform: docker (self-hosted)
• nodeJsVersion: 24.15.0
• nodeEnv: production
• database: postgres
• executionMode: scaling (single-main)
• concurrency: -1
• license: enterprise (production)
• consumerId: fe26e0b6-c656-4849-a6dd-f2f22571b35c

storage

• success: all
• error: all
• progress: false
• manual: true
• binaryMode: database

pruning

• enabled: false

client

• userAgent: mozilla/5.0 (x11; linux x86_64) applewebkit/537.36 (khtml, like gecko) chrome/148.0.0.0 safari/537.36
• isTouchDevice: false

security

• secureCookie: false

cluster

• instanceCount: 14
• versions: 2.21.7
• instances:
  • instanceKey: 3991f098-bd73-4f18-82c9-89f2ad985046, hostId: worker-0f9952148b9a, instanceType: worker, instanceRole: unset, version: 2.21.7
  • instanceKey: 56dc2cd9-8216-419d-8b74-7b600b65b45d, hostId: worker-36b27e949b6a, instanceType: worker, instanceRole: unset, version: 2.21.7
  • instanceKey: 93d623d2-3329-43fc-a374-c3a987959f85, hostId: worker-ab4c9ac828da, instanceType: worker, instanceRole: unset, version: 2.21.7
  • instanceKey: 22d1e1e3-4373-4771-9aa3-5ee818624ea6, hostId: worker-a8388081d21a, instanceType: worker, instanceRole: unset, version: 2.21.7
  • instanceKey: dd092197-c7e1-4f25-a95b-b36ca24f038e, hostId: webhook-67f3db8bdcd3, instanceType: webhook, instanceRole: unset, version: 2.21.7
  • instanceKey: a5d6caec-30b1-435a-9096-caea54b980a2, hostId: worker-f0332492b7dd, instanceType: worker, instanceRole: unset, version: 2.21.7
  • instanceKey: 6d05b475-bf17-4301-bef7-2136f31c70ea, hostId: worker-9dbf86644342, instanceType: worker, instanceRole: unset, version: 2.21.7
  • instanceKey: c4f40aee-745a-4d23-9b6f-b0215039314c, hostId: worker-35a225b2ef1e, instanceType: worker, instanceRole: unset, version: 2.21.7
  • instanceKey: 7553f267-2ca3-42d7-9dda-fc97c0eb35fa, hostId: worker-5049867dba0d, instanceType: worker, instanceRole: unset, version: 2.21.7
  • instanceKey: 5edcfa2a-8942-4944-90fc-dc304d3809cf, hostId: worker-cc64bc4484ed, instanceType: worker, instanceRole: unset, version: 2.21.7
  • instanceKey: 0136166d-8ff3-4416-b219-3ad46c9207ef, hostId: worker-539d1d4f56ba, instanceType: worker, instanceRole: unset, version: 2.21.7
  • instanceKey: c0e07bb5-5f77-4b2c-b2ed-e83902c5e7f6, hostId: worker-f36cfaf7810b, instanceType: worker, instanceRole: unset, version: 2.21.7
  • instanceKey: 23337ccf-7a1c-4274-8ead-e19d655f34fa, hostId: main-a4c571afd829, instanceType: main, instanceRole: leader, version: 2.21.7
  • instanceKey: 6869027a-f393-401c-bbbe-8a9cb083ccff, hostId: worker-19f4f79b9321, instanceType: worker, instanceRole: unset, version: 2.21.7
• checks:
  • check: hostid-clash, status: succeeded, warnings: -
  • check: lifecycle, status: succeeded, warnings: -
  • check: split-brain, status: succeeded, warnings: -
  • check: version-mismatch, status: succeeded, warnings: -

Generated at: 2026-05-28T14:23:07.337Z

Operating System

Linux (Docker container: docker.n8n.io/n8nio/n8n:2.21.7 - Ubuntu 22.04)

n8n Version

2.21.7

Node.js Version

24.15.0

Database

PostgreSQL

Execution mode

queue

Hosting

self hosted