Add native OpenClaw image-generation support via ZenMux Gemini/Vertex route

[abnershang]

Summary

The official OpenClaw plugin currently registers ZenMux only as an LLM provider. OpenClaw therefore cannot select ZenMux for its native image.generate capability, even though the ZenMux catalog includes image-output models such as google/gemini-3.1-flash-image.

Current behavior

After installing @zenmux/openclaw-plugin@0.1.2 and completing OAuth:

• zenmux appears in the LLM provider registry.
• The plugin manifest does not declare contracts.imageGenerationProviders.
• The plugin does not call api.registerImageGenerationProvider(...).
• Configuring agents.defaults.imageGenerationModel.primary as zenmux/google/gemini-3.1-flash-image does not work without an additional adapter.

Routing the model through an OpenAI-compatible /images/generations endpoint is also incorrect. The working ZenMux route is the Gemini/Vertex-compatible endpoint:

POST https://zenmux.ai/api/vertex-ai/v1/publishers/google/models/gemini-3.1-flash-image:generateContent

with generationConfig.responseModalities: ["TEXT", "IMAGE"]. Image bytes are returned in candidate content parts as inlineData / inline_data.

Expected behavior

The official plugin should expose ZenMux as an OpenClaw image-generation provider and reuse the existing ZenMux OAuth credential/refresh flow.

Suggested model reference:

zenmux/google/gemini-3.1-flash-image

Suggested implementation

• Add imageGenerationProviders: ["zenmux"] to the plugin manifest contract.
• Register an image-generation provider with api.registerImageGenerationProvider.
• Resolve request-ready credentials through OpenClaw runtime model auth so OAuth refresh remains provider-owned.
• Map OpenClaw prompt, input images, aspect ratio, resolution, and output format to ZenMux Vertex generateContent.
• Parse base64 image parts from candidates[].content.parts[].inlineData and inline_data.
• Keep response-size and timeout limits bounded.
• Add unit tests for request mapping, OAuth credential resolution, error responses, and image parsing.

Verification performed

A small local companion provider using the route above generated a real 1024×1024 PNG successfully with:

• provider: zenmux
• model: google/gemini-3.1-flash-image
• route: vertex-generateContent

One earlier request returned a successful response without image data, so the implementation should surface that case clearly and allow OpenClaw's configured fallback chain to proceed.