Image Attachments Flow

The web composer now supports image uploads for both "send to brain" and direct arm prompts.

Upload path

• The browser posts multipart form data to POST /api/uploads/images.
• The API stores the file under ~/.octopai/uploads/, records metadata in uploaded_media, and returns a signed contentUrl.
• Signed image delivery is served from GET /uploads/:id/content?token=... so harnesses and prompts can reference the image without API auth headers.

Prompt delivery

• Web-to-arm prompts send attachments to POST /api/arms/:id/prompt.
• The API checks the arm's provider/model modalities through /api/opencode/providers.
• opencode-api and opencode-tui receive native file parts when the selected model supports image input.
• Text-only harnesses, or models without image input, fall back to appending an ATTACHED IMAGES section with signed URLs to the prompt text.
• Distributed arms carry the same attachment payload through NATS so remote agents use the same rules as local harnesses.

Brain and task context

• Web-to-brain messages store attachment metadata in the mail headers (X-Coleo-Attachments).
• The brain parses those headers, keeps attachments in task context.attachments, and includes them in get_full_briefing output.
• This keeps screenshots available even when the initial human request becomes a queued task instead of an immediate direct prompt.