[BUG] Muse Spark 1.2 Contributor Free on OpenCode Zen does not work with images

[FloPinguin]

If you try to send a message with an image to that model, you get:

Sorry, your request failed. Please try again.

Client Request Id: 30cd20ce-06bb-4e35-a7ff-de425a21177f

Reason: OpenCode Zen API request failed (400) model=muse-spark-1.2-contributor-free payloadBytes=155759: {"id":"chatcmpl_tzmu7e4mn9","object":"chat.completion","created":1787408810,"model":"muse-spark-1.2-contributor-free","choices":[{"index":0,"message":{"role":"assistant"},"finish_reason":null}]}: OpenCodeRequestError: OpenCode Zen API request failed (400) model=muse-spark-1.2-contributor-free payloadBytes=155759: {"id":"chatcmpl_tzmu7e4mn9","object":"chat.completion","created":1787408810,"model":"muse-spark-1.2-contributor-free","choices":[{"index":0,"message":{"role":"assistant"},"finish_reason":null}]} at buildOpenCodeRequestError (c:\Users\Flo.vscode\extensions\ltmoerdani.opencode-copilot-chat-0.6.0\out\errors.js:39:12) at streamOpenCodeResponse (c:\Users\Flo.vscode\extensions\ltmoerdani.opencode-copilot-chat-0.6.0\out\streaming.js:360:73) at process.processTicksAndRejections (node:internal/process/task_queues:104:5) at async streamChatCompletions (c:\Users\Flo.vscode\extensions\ltmoerdani.opencode-copilot-chat-0.6.0\out\streaming.js:70:5) at async OpenCodeProvider.provideLanguageModelChatResponse (c:\Users\Flo.vscode\extensions\ltmoerdani.opencode-copilot-chat-0.6.0\out\extension.js:2702:13)

[ltmoerdani]

Hey @FloPinguin, thanks for the report and for including the payload details, that saved me a repro round.

I dug into it on our side:

• models.dev lists muse-spark-1.2-contributor-free as image-capable (attachment: true, modalities.input includes image), so the extension advertises it as vision-capable and sends input_image with a plain string URL to https://opencode.ai/zen/v1/responses. That matches the Responses API contract, and Zen's own docs put this model on /v1/responses, so routing on our side checks out.
• The 400 you got is the odd part. The body is a chat.completion with an empty assistant message and finish_reason: null, not a normal error object. Best I can tell, the upstream endpoint returned an empty completion and the gateway turned that into a 400. I can't verify that from here, so take it as a guess, but nothing in the request format looks wrong.

My working theory is that the free contributor tier on Zen doesn't accept images yet even though the metadata says it does.

Workarounds in the meantime:

• set a vision proxy model in the extension settings (opencodego.visionProxyModelId), images get described by that model and sent to Muse as text
• or use muse-spark-1.2 (paid) or another vision model when you need image input

One thing that would help: does it also fail with a small PNG (1x1 or under 500KB)?

[FloPinguin]

The interesting thing is that it works with OpenCode (OpenChamber UI), with the same image and also with the OpenCode Zen provider.

I tried different images, always the same problem.

[Fahad090NP]

PEACE BE UPON YOU

Thanks for the detailed report — the error body actually tells the story.

Root cause: your stack trace is from v0.6.0, where muse-spark-1.2-contributor-free was sent through the chat-completions endpoint. The Muse gateway rejects chat-completions for these models (they require the Responses API), and the 400 body you got — a chat.completion object with an empty message and finish_reason: null — is the gateway failing that request shape. Commit 5586836 (released in 0.7.0) routes all muse-* models to the Responses endpoint and maps images to the correct input_image format, so this exact failure should be gone on current builds.

Also hardened in #189: models.dev lists image input for both Muse variants, but the extension's bundled offline fallback didn't — meaning with no network/metadata cache, the picker reported the model as non-vision and VS Code stripped attachments up front. Both variants are now in the fallback set.

Please update to the latest release (0.7.0+) and retry with an image. If it still fails there, paste the new error + diagnostics and we'll dig into the Zen Responses path specifically.

Fi Amanillah