openclaw - 💡(How to fix) Fix 🔴 [CONFIRMED BUG] WebUI image attachments dropped — Built-in Qwen plugin filters qwen3.6-plus on Coding Plan baseUrl + custom provider key mismatch [1 comments, 1 participants]

Bug Description

WebUI (Control UI) image attachments are silently dropped with the warning parseMessageWithAttachments: X attachment(s) dropped — model does not support images, even when the configured model explicitly declares vision support (input: ["text", "image"]).

The same model works correctly via Telegram channel. This indicates the issue is specific to WebUI attachment handling, not model configuration.

Environment

Model Configuration

{
  "id": "qwen3.6-plus",
  "name": "qwen3.6-plus",
  "input": ["text", "image"],
  "contextWindow": 1000000,
  "maxTokens": 65536
}

This model is officially confirmed to support visual understanding by the provider (Alibaba Coding Plan).

Reproduction Steps

1. Configure a vision-capable model (alibailian/qwen3.6-plus) with input: ["text", "image"]
2. Restart Gateway to ensure model catalog is loaded
3. Open Control UI (WebUI) in browser
4. Upload 2 images (1 PNG, 1 JPG) via the chat interface
5. Observe: Images are dropped, no visual analysis occurs

Expected Behavior

WebUI should accept and forward image attachments to the vision-capable model for analysis, same as Telegram channel.

Actual Behavior

• WebUI: Images are dropped with warning parseMessageWithAttachments: 2 attachment(s) dropped — model does not support images
• Telegram: Same model, same config, same images — works correctly

Evidence / Logs

1. Gateway startup confirms model loaded

12:58:47 INFO  agent model: alibailian/qwen3.6-plus
12:58:47 INFO  ready (9 plugins: acpx, active-memory, browser, device-pair, memory-core, memory-wiki, phone-control, talk-voice, telegram; 21.2s)

2. WebUI image upload FAILS

13:01:01 WARN  parseMessageWithAttachments: 2 attachment(s) dropped — model does not support images

3. Media understanding subsystem also fails

12:21:16 WARN  media-understanding: image: failed (0/1) reason=Unknown model

4. Telegram channel WORKS (same model, same session)

• Telegram image sent at 12:21 — successfully analyzed
• Model correctly identified as vision-capable via Telegram

Code Analysis

The bug appears to be in resolveGatewayModelSupportsImages() (session-utils-CIR-C-as.js):

async function resolveGatewayModelSupportsImages(params) {
    if (!params.model) return true;
    try {
        const modelEntry = (await params.loadGatewayModelCatalog())
            .find((entry) => entry.id === params.model && 
                            (!params.provider || entry.provider === params.provider));
        if (modelEntry) {
            if (modelEntry.input?.includes("image")) return true;
            return false;
        }
        return false;  // Falls through here
    } catch {
        return false;
    }
}

Root cause hypothesis: The modelEntry lookup fails for WebUI sessions because:

1. params.provider format differs between WebUI and Telegram channels
2. loadGatewayModelCatalog() does not include custom provider models (only pi-coding-agent built-in models)
3. The entry.id / entry.provider values in the model catalog do not match the WebUI resolved model ref

Supporting evidence: media-understanding logs Unknown model for qwen3.6-plus, confirming the model catalog lookup returns undefined.

Workaround

Use Telegram channel for image uploads — works correctly with the same model configuration.

Additional Context

• Gateway restart does NOT fix the issue
• Config hot-reload does NOT fix the issue
• Telegram channel works correctly (proves model config is valid)
• Issue is 100% reproducible on WebUI

Session Status (at time of report)

Model: alibailian/qwen3.6-plus
Context: 109k/1.0m (11%)
Compactions: 0
Runtime: direct