openclaw - 💡(How to fix) Fix [Bug]: google-vertex/gemini-3.1-pro-preview returns empty assistant reply in 2026.4.5 [1 participants]

Bug type

Behavior bug (incorrect output/state without crash)

Summary

google-vertex/gemini-3.1-pro-preview returns an empty assistant reply in OpenClaw 2026.4.5, even though the model works in Vertex Playground and google-vertex/gemini-3-flash-preview works in the same OpenClaw environment.

Model

google-vertex/gemini-3.1-pro-preview

Provider / routing chain

openclaw -> google-vertex -> Vertex AI

Additional provider/model setup details

• Built-in google-vertex provider
• Vertex authentication is valid and requests do reach Vertex AI
• Tested with Vertex region us-central1
• Same OpenClaw install can successfully use google-vertex/gemini-3-flash-preview
• Vertex Playground can successfully use Gemini 3.1 Pro in the same general environment
• No private credentials, project IDs, tokens, or local secret values are included here

Steps to reproduce

1. Configure OpenClaw to use the built-in google-vertex provider
2. Set the model to google-vertex/gemini-3.1-pro-preview
3. Ensure Vertex authentication is valid
4. Send a normal chat message such as hello???
5. Observe that the request reaches Vertex AI, but the assistant reply is empty

Expected behavior

The assistant should return normal text output, similar to:

• Vertex Playground using Gemini 3.1 Pro
• google-vertex/gemini-3-flash-preview in the same OpenClaw install

Actual behavior

The session completes without a visible provider error, but the assistant reply is empty.

Observed session result:

• provider: google-vertex
• model: gemini-3.1-pro-preview
• stopReason: stop
• assistant text: ""
• output tokens: 0

So this does not appear to be a crash or fallback. It looks like a successful completion with blank output.

OpenClaw version

2026.4.5

Operating system

Linux 6.1.0-38-amd64 x86_64 GNU/Linux

Install method

npm global install

Node.js version

v24.14.1

Logs, screenshots, and evidence

Relevant observed behavior from the session trace:

• request goes to provider google-vertex
• model is gemini-3.1-pro-preview
• assistant response is saved as empty text
• output tokens are 0
• stop reason is stop
• no fallback and no obvious logged provider error

Cross-check evidence:

• google-vertex/gemini-3-flash-preview works in the same OpenClaw install
• Gemini 3.1 Pro works in Vertex Playground
• this suggests the request path and general Vertex setup are functional, but the OpenClaw runtime handling for this model is not returning usable text

Impact and severity

• Affected users/systems/channels: users running OpenClaw with the built-in google-vertex provider and google-vertex/gemini-3.1-pro-preview
• Severity: blocks practical use of Gemini 3.1 Pro on Vertex in OpenClaw conversations
• Frequency: observed consistently in this environment during testing
• Consequence: model appears unusable in real OpenClaw chats, and users may misdiagnose it as a GCP auth/quota issue even when Vertex Playground works

Additional information

This looks more like an OpenClaw provider/runtime handling issue than a Vertex AI configuration issue, because:

• Vertex Playground works with Gemini 3.1 Pro
• OpenClaw can successfully use google-vertex/gemini-3-flash-preview
• OpenClaw reaches Vertex with Gemini 3.1 Pro, but stores the response as an empty assistant message

This may be related to other google-vertex runtime issues already reported, but the symptom here is slightly different from visible crash reports:

• instead of throwing an error,
• the run finishes with stopReason=stop and empty text

Regression status: NOT_ENOUGH_INFO