feat: send images to serverless agents from the playground and CLI

[devin-ai-integration]

Summary

This adds the client side of image turns. The backend support is in diggerhq/blue feat/turn-image-attachments, which adds attachments to POST /sessions/:id/turns. That PR needs to be deployed before the images reach the agent.

attachments?: { type: "image"; mediaType: string; data: string /* base64 */; name?: string }[]

Web playground (managed-agents/Detail.tsx, images.ts):

• The composer can take images from a picker button, a paste, or a drag and drop. Each image shows as a thumbnail you can remove, and sent images also appear in the user's message bubble.
• imageRefusal rejects anything that isn't PNG, JPEG, GIF or WebP, anything over 5 MB, and more than 4 images, with a message shown inline.
• A message can be sent with only an image. Queue, steer and interrupt still work for text; a message with images is queued.
• Images travel as AI SDK FileUIPart data URLs. ManagedAgentChatTransport turns them into attachments for runManagedAgent and continueManagedAgentSession. When a session is already active, admitManagedAgentInput sends them directly.

CLI:

opencomputer run <agent> [prompt] --image ./shot.png [--image ...]
opencomputer session send <session-id> [prompt] --image ./shot.png

• readTurnImages checks that each path is a regular file of at most 5 MB, finds the type from the file's magic bytes (the file extension is ignored), and allows at most 4 images.
• When there are images, the prompt is optional.
• createTurn(sessionId, input, key, attachments = []) only adds attachments to the request body when there are some, so text-only requests are sent exactly as before.

Verified locally: cli npm test (135 passed) and tsc; web vitest (299 passed), npm run build, and eslint on the changed files. The only remaining lint error in Detail.tsx (react-refresh/only-export-components on PROJECT_DETAIL_TABS) is already on main.

Link to Devin session: https://app.devin.ai/sessions/af9f2258a9af4cad96662ae2c820b4f1
Open in Devin Desktop: https://app.devin.ai/desktop/session/af9f2258a9af4cad96662ae2c820b4f1?variant=devin
Requested by: @motatoes

Polylane reviews this pull request when you ask:

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[devin-ai-integration]

Tested locally against a real Blue worker (https://github.com/diggerhq/blue/pull/133) and a real OpenCode/model runtime, with no API mocking.

Passed: picker preview and removal; image + text, where the base64 sent matched the original file byte for byte and the model correctly identified the orange triangle and blue circle in the fixture; image-only send via paste; drag and drop; inline notices for SVG, files over 5 MB and a 5th image; image + steer gets queued; a text-only request has no attachments key. The backend returned the expected 400 codes for an empty turn, invalid bytes, 5 images, an oversized image, and data passed with a data: prefix.

Real model recognition	Image-only response

Follow-ups from the run:

• A .txt renamed to .png only failed after Send, with a backend 400. Fixed in 121e8ee: the client now checks the file's bytes and shows the inline notice before sending.
• Image thumbnails don't show in history after a reload. The backend deliberately deletes image bytes once a turn finishes, so history only has the image name and hash.
• Turns admitted while a run is active only appear in the chat after a reload. A text-only control showed the same thing, so this was already the case and isn't caused by images.

Not covered: workerd forwarding (covered by an integration test instead), sending an image successfully from the CLI (a local route mismatch between /api/managed-agents and /v1 blocked it; the CLI's rejection of a non-image file was checked and passed), and Slack.