[Bug]: MiniMax-M3 image capability on Anthropic route bypasses MiniMax-VL-01 and silently loses vision

Bug type

Regression (previous MiniMax VL routing fix no longer covers the newly image-capable M3 catalog entry)

Beta release blocker

No

Summary

OpenClaw 2026.7.2-beta.5 (ee929db) marks minimax/MiniMax-M3 as accepting input: ["text", "image"] while the provider uses MiniMax's Anthropic-compatible endpoint. OpenClaw therefore treats the active model as natively vision-capable and the image tool returns the image to M3 as Loaded 1 image for direct visual inspection instead of invoking the plugin-owned MiniMax-VL-01 image-understanding provider.

MiniMax's compatible Anthropic API documentation says image/document input is not supported on that route: https://platform.minimax.io/docs/api-reference/text-anthropic-api

This silently loses image understanding rather than producing a provider error. It is the M3 version of the routing regression previously tracked in #64244.

Steps to reproduce

1. Install OpenClaw 2026.7.2-beta.5 (ee929db) and enable the bundled minimax plugin.
2. Configure MiniMax with api: "anthropic-messages" and baseUrl: "https://api.minimax.io/anthropic".
3. Use minimax/MiniMax-M3 as the active chat model with the bundled/catalog capability input: ["text", "image"].
4. Send an image and ask the agent to inspect it, or call the image tool on a local image.

Expected behavior

OpenClaw should route image understanding through the plugin-owned minimax/MiniMax-VL-01 provider and return a textual image analysis, as documented for MiniMax and as intended by the fix cited when #64244 was closed.

Actual behavior

The transcript records:

assistant stopReason=toolUse: image(...)
toolResult: Loaded 1 image for direct visual inspection.

The following M3 response says the tool returned no description and repeatedly plans to retry/analyze, but never obtains visual content. In the observed Telegram turn it emitted 22,755 characters of planning text and stopped at Let me finalize. without an actual answer.

There was no 4xx response, so normal model fallback did not activate.

Root cause evidence

• The beta.5 MiniMax catalog and provider docs declare M3 with input: ["text", "image"].
• The image tool prioritizes the modelHasVision native path before configured image-model fallback selection.
• The native path returns the image block to the active M3 Anthropic-compatible request rather than calling MiniMax-VL-01.
• MiniMax documents that the compatible Anthropic API does not accept image/document input.

Verified workaround

Override M3 as text-only and explicitly select the plugin-owned vision provider:

{
  "agents": {
    "defaults": {
      "imageModel": {
        "primary": "minimax/MiniMax-VL-01"
      }
    }
  },
  "models": {
    "providers": {
      "minimax": {
        "models": [
          {
            "id": "MiniMax-M3",
            "input": ["text"]
          }
        ]
      }
    }
  }
}

After restarting, a no-delivery test with M3 produced this image-tool result:

Text: "VISION TEST 731"
Blue Shape: Square, left side
Orange Shape: Circle, right side

The final assistant answer was one concise sentence with the correct text and shapes.

Suggested fix

Make image capability route-aware for MiniMax M3, or special-case MiniMax chat models on the Anthropic-compatible route so image-tool resolution prefers the plugin-owned MiniMax-VL-01 media provider instead of the native active-model path. Add a regression test equivalent to the M2.7 coverage referenced in #64244, but with M3's new catalog entry.

Environment

• OpenClaw: 2026.7.2-beta.5 (ee929db)
• Install: npm global
• OS: Linux
• Provider: minimax
• Chat model: MiniMax-M3
• Chat API: anthropic-messages
• Image provider: bundled MiniMax plugin / MiniMax-VL-01