Refs #4158. First cut of multimodal: send attached images to vision-capable models. Video / keyframe decoding is intentionally out of scope.

The gap

Images dragged/pasted into chat are saved to .reasonix/attachments/ and referenced as @…png. Today the controller turns that into a text block — [image attachment available… use a vision MCP tool] — and provider.Message.Content is a plain string, so the model never receives image data. The issue's root-cause read (openai chatMessage.Content is *string) was correct.

What changed

• provider.Message.Images []string (data URLs). Agent history is []provider.Message, so one field carries images through history, session persistence, and both providers.
• openai: chatMessage.Content → any; a []chatContentPart array (text + image_url) for a vision user turn, the same string/null shape as before otherwise.
• anthropic: a {type:image, source:{type:base64,media_type,data}} block beside the text block (parsed from the data URL via the shared provider.ParseImageDataURL).
• Gating — a per-model vision config flag (forwarded through Config.Extra). Only vision models get image parts; text-only models never receive them (they'd 400) and their prompt-cache prefix is byte-for-byte unchanged. Set vision = true on the DeepSeek model entry once its multimodal API lands.
• Threading — the controller resolves image @refs to data URLs and passes them via agent.WithUserImages(ctx, …), mirroring WithParentSession, so agent stays free of an internal/control import.

Cache safety

Images live only in user turns; embedding one doesn't retroactively invalidate the prefix before it (append-only history). Text-only flows are untouched because of the vision gate.

Tests

provider (ParseImageDataURL), openai / anthropic (image parts/blocks when vision, plain content when not), control (image @ref → data URL on the turn). go vet + the touched packages all green locally.

Known tradeoffs / follow-ups

• Images store inline as data URLs in session history → .jsonl bloat. A path-and-resolve-at-send design would keep history compact (deferred).
• Video → keyframes (ffmpeg) is a separate, larger effort.
• Anthropic vision works today; DeepSeek's main models are text-only until their multimodal API ships, so the immediate beneficiaries are Claude and other OpenAI-compatible vision endpoints.