fix(provider/openai-compatible): trust user-declared modalities, dont inject fake ERROR text

[Alezander9]

Summary

packages/opencode/src/provider/transform.ts — unsupportedParts replaces image/file parts with an inline ERROR: Cannot read <name> (this model does not support <modality> input) text part when model.capabilities.input.<modality> is false. This is the right behavior for native providers where models.dev knows the truth. For @ai-sdk/openai-compatible (the user-configured proxy provider) it's a silent footgun: there is no models.dev entry to consult, so image/audio/video/pdf all default to false unless the user explicitly declares modalities in opencode.json, and bcode silently strips screenshots from vision-capable upstream models.

Repro

Browser Use cloud's V4 LLM gateway is an openai-compatible proxy in front of Anthropic. Worker opencode.json declared:

"models": { "claude-opus-4.7": { "name": "claude-opus-4.7 via V4 gateway" } }

Vision smoketest: agent screenshots example.com via browser_execute, writes its verdict to a file. Result every run: "I CANNOT SEE THE IMAGE". Gateway-side request audit confirmed zero image_url / file / data:image/ signals in the wire body even when the synthetic "Attached media from tool result:" user message was hoisted with file parts. The file part was being replaced by unsupportedParts before streamText serialized the body — the model saw the fabricated error text as if it came from the user.

Capability derivation reference (provider.ts:1298-1304):

image: model.modalities?.input?.includes("image") ?? existingModel?.capabilities.input.image ?? false,

When neither the user-declared model nor a models.dev fallback declares modalities, image defaults to false. unsupportedParts then replaces the file part with:

{
  type: "text",
  text: `ERROR: Cannot read "${filename}" (this model does not support image input). Inform the user.`,
}

Fix

Early-return for @ai-sdk/openai-compatible — forward image/file parts as-is. If upstream truly can't handle them, the provider call returns a real error (e.g. Anthropic 400 "unsupported media type") which is far more debuggable than fabricated capability text the model reads back to the user.

if (model.api.npm === "@ai-sdk/openai-compatible") return msgs

Native providers (@ai-sdk/anthropic, @ai-sdk/openai, @ai-sdk/google, @ai-sdk/amazon-bedrock) keep the existing check — models.dev IS authoritative for them, and the filter prevents a real 4xx upstream while giving the user a clear local error.

Why not require users to declare modalities?

That's the workaround on our side (declaring modalities: { input: ['text', 'image'] } in opencode.json), which we're also shipping. But:

1. The current behavior is a silent footgun — there's no warning, the model just receives weird text.
2. The error text masquerades as user input, which trains the model to "explain why it can't see the image" instead of surfacing the bug.
3. For user-configured proxies bcode genuinely doesn't know what's downstream; defaulting to "strip everything" is the wrong direction. "Forward and let upstream answer" is honest.
4. Diff is 1 line + comment.

Risk

Other openai-compatible users who point bcode at a non-vision endpoint will now receive an upstream 4xx instead of a local stripped-with-fake-error. Strictly an error-message-quality regression for them; the model never received the image in either case.

---

Summary by cubic

Forward media parts for @ai-sdk/openai-compatible instead of replacing them with fake "ERROR: Cannot read…" text. This fixes silent stripping of images/files and lets upstream return real, debuggable errors when a modality isn’t supported.

• Bug Fixes
  • Bypass unsupportedParts for @ai-sdk/openai-compatible; media is forwarded as-is.
  • Keep capability filtering for native providers (@ai-sdk/anthropic, @ai-sdk/openai, @ai-sdk/google, @ai-sdk/amazon-bedrock) where models.dev is authoritative.

^{Written for commit cbd37a8. Summary will update on new commits. Review in cubic}

[cubic-dev-ai]

No issues found across 1 file

_{Re-trigger cubic}