[bug] image generate ignores --model and falls back to qwen-image-2.0

[65316]

Environment

• CLI: bl 1.10.1
• Skill: 1.10.1
• Node: v24.15.0
• OS: Microsoft Windows NT 10.0.26200.0 (Windows, x64)
• Region/base URL: domestic (https://dashscope.aliyuncs.com)
• Authentication: healthy API-key config (credential redacted)

Reproduce

Dry-run:

bl image generate --prompt "test human plush doll on cyan background" --model wan2.6-t2i --size "1024*1024" --watermark false --prompt-extend false --dry-run --output json

The dry-run correctly reports:

{
  "request": {
    "model": "wan2.6-t2i",
    "parameters": {
      "size": "1024*1024",
      "n": 1,
      "prompt_extend": false,
      "watermark": false
    }
  },
  "mode": "async"
}

Real request:

bl image generate --prompt "test human plush doll on cyan background" --model wan2.6-t2i --size "1024*1024" --watermark false --prompt-extend false --output json

The same behavior also occurred with:

bl image edit --image .\reference.jpg --prompt "[redacted user prompt about a human plush-doll sprite]" --model wan2.7-image --size "1024*1024" --watermark false --prompt-extend false --output json

Expected

The real request should use the explicitly selected wan2.6-t2i / wan2.7-image model, matching the dry-run payload, and honor watermark=false.

Actual

The command exits successfully but reports [Model: qwen-image-2.0] [Mode: sync]. The downloaded image visibly contains a Qwen-Image watermark even though --watermark false was passed. The --model selection appears to be ignored or silently replaced with the default model during execution.

Full output

urls:
  - [temporary OSS URL redacted]
saved:
  - [local output path redacted]
total: 1
[Model: qwen-image-2.0] [Mode: sync]
Exit code: 0

For image edit --model wan2.7-image, the result similarly ended with:

[Model: qwen-image-2.0] [Mode: sync] [Images: 1]
Exit code: 0

Already tried

• Confirmed CLI and installed skill versions are aligned at 1.10.1.
• Confirmed bl auth status is healthy.
• Tested both bl image generate and bl image edit.
• Tested two non-default model IDs: wan2.6-t2i and wan2.7-image.
• Confirmed --dry-run preserves the requested model while real execution reports the default model.
• Repeated real requests; behavior reproduced consistently.

Notes

• Frequency: always in this session (multiple requests)
• Invoked via: Codex agent on Windows PowerShell
• The mismatch is costly because successful exit status can consume image quota on an unintended model.

[65316]

Additional reproduction after successful domestic console login and free-tier discovery:nn- wan2.2-t2i-plus is selected correctly ([Model: wan2.2-t2i-plus] [Mode: async]) but returns HTTP 400 InvalidParameter: url error, request ID 59b5acc4-8241-9509-9e1d-1f2471187852.n- wanx-v1returns the same error, request ID1f4e2878-3ac5-9461-9254-fe0f9eba6205.n- z-image-turbo returns the same error, request ID fcaeb955-6862-9ae4-839c-f278daf4a510.n- After each error on Windows, the process also prints: Assertion failed: !(handle->flags & UV_HANDLE_CLOSING), file src\win\async.c, line 76.nnbl usage free --all` confirms all three models have remaining free quota. This suggests the non-Qwen async image request path is constructing an incompatible payload (the service complains about a URL even for text-to-image), while Qwen sync generation succeeds.