v0.4.1 — VLM UI + Persistence + HTTP

[magicnight]

Summary

Third and final v0.4.1 PR — lights up the user-facing surfaces for vision-language models. Closes the v0.4.1 rollout begun in #33 (Foundation) and #34 (Engine).

Plan: `docs/superpowers/plans/2026-05-10-v0.4.1-vlm.md`.

What lands

Chat input — image picker + thumbnail strip

• Paperclip button in `ChatInputView` opens SwiftUI `.fileImporter` (image UTTypes only: jpeg / png / webp / gif / heic / bmp).
• Horizontal thumbnail strip above the text field; click × on any thumbnail to drop it.
• Disabled (with explanatory tooltip) when the loaded model isn't a VLM.
• Image-only messages on a VLM are now legitimate sends.

Chat bubbles — inline thumbnails

• `ChatMessageView` renders a 96pt `LazyVGrid` of attached images above the text bubble.
• Click a thumbnail → open in Preview via `NSWorkspace`.

Conversation persistence

• `StoredMessage.images: [ImageAttachment]` round-trips through `ConversationStore` (decoder defaults to empty when key absent — pre-v0.4.1 chats unchanged).
• `save(_:)` internalises external image URLs into `//images/<image-uuid>.<ext>` so chats survive the user moving the picked file.
• `delete(id:)` tears down the per-conversation directory.

OpenAI multimodal HTTP

• `/v1/chat/completions` accepts OpenAI's `content` array shape:
  ```json
  {"role":"user","content":[
  {"type":"text","text":"What's this?"},
  {"type":"image_url","image_url":{"url":"data:image/png;base64,…"}}
  ]}
  ```
• Plain-string `content` continues to work — decoder tries string first, falls through to `[Part]`.
• base64 data URLs → tmpfile-backed `ImageAttachment`. Caps: 10 MB / image, 4 images / message; `http(s)://` and `file://` URLs are not fetched.
• Ollama `/api/chat` / `/api/generate` stay text-only — Ollama uses a separate top-level `images: [base64]` field, follow-up.

Test plan

• `swift test --package-path MacMLXCore` — 115/115 green (incl. 4 new `ConversationStoreImagesTests`)
• Local `xcodebuild macMLX` — green
• Existing `chatCompletionsNonStreaming` test confirms string-form fallback path on the multimodal decoder
• Manual smoke (post-merge, on hardware):
  • `macmlx pull mlx-community/SmolVLM-Instruct-4bit`
  • Load it in the GUI, attach a JPEG, send "what is this?", get a coherent response.
  • `curl localhost:8000/v1/chat/completions -d '{ "model": "SmolVLM-Instruct-4bit", "messages": [{"role":"user","content":[{"type":"text","text":"describe"},{"type":"image_url","image_url":{"url":"data:image/png;base64,…"}}]}]}'` returns a sensible reply.

Stack

Builds on `main` after:

• v0.4.1 — VLM Foundation (data model + detection) #33 — Foundation (data model + library detection)
• v0.4.1 — VLM Engine (MLXSwiftEngine branch via MLXVLM) #34 — Engine (MLXSwiftEngine VLM branch via `MLXVLM.VLMModelFactory`)

After this PR merges, v0.4.1 is feature-complete and a tag is reasonable.

🤖 Generated with Claude Code