Skip to content

[Bug]: Vision-capable model (qwen3-vl-plus) not used for native image input — agent falls back to external image tools #634

Description

@guolvlin-cn

Problem

When a vision-capable model (e.g. qwen3-vl-plus) is selected via /model, the agent does not pass image input directly to the model as multimodal content. Instead it falls back to external image-understanding tools (the understand-image skill, which calls the MiniMax understand_image MCP tool).

If that external tool is unavailable (MCP server not connected), the agent cannot read the image at all — despite running on a native vision model.

Repro

  1. Register a vision model (astra admin model ...), e.g. qwen3-vl-plus
  2. Select it via /model
  3. Provide a local image path and ask the agent to read/describe the image
  4. Observe: the agent does NOT pass the image bytes to the model; it attempts understand_image (external MCP) instead

Expected

When the active model is vision-capable, the agent should pass image input natively as multimodal content to the model, without requiring an external image tool.

Actual

Image input is not routed to the vision model. The agent depends on external image tools; when they are unavailable, image reading fails entirely.

Environment

  • astra 0.1.0
  • model: qwen3-vl-plus (high)
  • input: local image file path

Related

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    kind/bugBug or incorrect behavior

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions