bug: `run --file` hardcodes `text/plain` MIME type for all files

[alexey-pelykh]

Problem

The run command's --file flag hardcodes text/plain MIME type for all non-directory files, making binary file attachments (images, audio, video) non-functional.

In packages/opencode/src/cli/cmd/run.ts:

const mime = (await Filesystem.isDir(resolvedPath))
  ? "application/x-directory"
  : "text/plain"

Expected behavior

MIME type should be detected from the file (e.g. via extension or magic bytes) so that opencode run --file image.png -m "describe this" correctly passes the image as image/png.

Reproduction

opencode run --format json --file photo.jpg -m "describe this image"

The file is sent as text/plain regardless of actual content type, so the model receives corrupted/unusable binary data.

Suggested fix

Use file extension or content-based MIME detection (e.g. mime-types package or Node's built-in detection) instead of the hardcoded "text/plain" fallback.

[andrea-tomassi]

I've been debugging this while working on a vision pipeline that uses opencode run --file screenshot.png with a vLLM-hosted multimodal model via the openai-compatible provider.

Root cause in run.ts: Line 331 hardcodes mime: "text/plain" for all --file arguments. This causes prompt.ts to route images through the text/Read path (line 1065) instead of the binary/base64 multimodal path (line 1184+).

I noticed PR #23291 adds MIME sniffing via magic bytes in message-v2.ts and read.ts, but the run.ts --file code path isn't covered there — so this issue would remain open even after that PR merges.

Additional downstream issue: Even after fixing the MIME in run.ts, there's a second bug in @ai-sdk/openai-compatible@2.0.41. The provider at line 158 constructs data:${mediaType};base64,${convertToBase64(part.data)}, but part.data arrives as a full data URL string (e.g. data:image/png;base64,...) from convertToModelMessages. Since convertToBase64() returns strings as-is, the result is double-encoded: data:image/png;base64,data:image/png;base64,.... This doesn't affect the dedicated @ai-sdk/openai or @ai-sdk/anthropic providers (they have their own message conversion), which likely explains why it's gone unnoticed — it only impacts openai-compatible users sending images via file attachment.

I'm planning to file the AI SDK issue on vercel/ai separately. Happy to submit a PR for the run.ts fix here if that's welcome.

[antl3x]

@thdxr @kitlangton any ETA ?