Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

Under a working Vertex setup with GOOGLE_CLOUD_LOCATION=global, OpenClaw returns an empty assistant reply for google-vertex/gemini-3.1-pro-preview, while google-vertex/gemini-3-flash-preview and direct Vertex calls to gemini-3.1-pro-preview return normal text.

Steps to reproduce

1. Run OpenClaw 2026.4.5 with valid Google Vertex credentials and GOOGLE_CLOUD_LOCATION=global.
2. Set the model to google-vertex/gemini-3.1-pro-preview.
3. Keep the same config otherwise; in my case, thinking default was set to medium.
4. Send a simple prompt such as Reply with exactly: OK.
5. Observe that OpenClaw returns an empty assistant reply.
6. Change only the model to google-vertex/gemini-3-flash-preview under the same config and observe normal text output.

Expected behavior

google-vertex/gemini-3.1-pro-preview should return visible text in OpenClaw, because direct Vertex publisher API calls to the same model under the same credentials/project/location return normal text output.

Actual behavior

OpenClaw returns an empty assistant reply for google-vertex/gemini-3.1-pro-preview under a working Vertex configuration. Under the same configuration, google-vertex/gemini-3-flash-preview returns normal text.

OpenClaw version

v2026.4.5

Operating system

Debian 12 / Linux 6.1.0-38-amd64

Install method

npm global

Model

google-vertex/gemini-3.1-pro-preview

Provider / routing chain

openclaw -> bundled Google provider (google-vertex) -> Vertex AI publisher API

Additional provider/model setup details

• GOOGLE_CLOUD_LOCATION=global
• Vertex auth/project/env are valid
• The same environment works for:
  • google-vertex/gemini-3-flash-preview inside OpenClaw
  • direct Vertex publisher API calls to both gemini-3-flash-preview and gemini-3.1-pro-preview
• Thinking default in openclaw.json was set to medium

Logs, screenshots, and evidence

- Verified in OpenClaw:
  - `google-vertex/gemini-3-flash-preview` works again when region is set to `global`
  - `google-vertex/gemini-3.1-pro-preview` still returns empty replies under the same config
- Direct Vertex API tests for `gemini-3.1-pro-preview` returned visible text, which rules out a general Vertex outage or broken credentials.
- Code inspection of the OpenClaw dist bundle found the Google transport request path for Gemini 3 / 3.1 Pro uses a disabled-thinking path that can inject:
  - `thinkingConfig: { thinkingLevel: "LOW" }`
- Live repro against Vertex showed:
  - `gemini-3.1-pro-preview` with small output budget and `thinkingLevel: "LOW"` can return:
    - no visible `parts[].text`
    - `finishReason: "MAX_TOKENS"`
    - nonzero thought-token usage
  - `gemini-3.1-pro-preview` with `thinkingBudget: 0` returned visible text
- This supports that the empty reply is caused by the outgoing thinking config/path, not by broken Vertex auth.

Impact and severity

• Affected: users routing Gemini 3.1 Pro through google-vertex in OpenClaw
• Severity: High (the model returns no visible output, so it is not usable through this path)
• Frequency: Repeated during testing under the same working Vertex configuration
• Consequence: google-vertex/gemini-3.1-pro-preview cannot be used reliably in OpenClaw even though Vertex itself is working

Additional information

I traced the relevant bundled dist code and the effective request/response path appears to involve:

• dist/extensions/google/index.js
• dist/api-C29T3o68.js
• dist/anthropic-vertex-stream-BySayhWO.js
• dist/proxy-stream-wrappers-C_9sn9XU.js

Observed code-inspection direction for triage:

• google-vertex is routed through the Google transport path, not a separate Vertex-specific parser.
• In dist/anthropic-vertex-stream-BySayhWO.js, the request builder has a Gemini 3 / 3.1 Pro path that can populate thinkingConfig.
• In direct testing outside OpenClaw, a thinkingConfig path using thinkingLevel: "LOW" for gemini-3.1-pro-preview could reproduce a response with no visible text parts, while thinkingBudget: 0 returned visible text.
• Because my OpenClaw config had thinking default set to medium, this may indicate that OpenClaw is either:
  • sending an unexpected thinkingConfig for this model, or
  • not correctly respecting the configured thinking level, or
  • allowing hidden thinking to consume the visible output budget before any final text is emitted.