1 # Siti AI — visionOS app
2
3 A native SwiftUI chat app for visionOS, built on [Ed Swift](https://github.com/ondeinference/ed-swift), the agent SDK over the same on-device Onde inference engine that powers the Tauri app at the repo root (through `ed-agent`). This is a separate native client, not a Tauri build target; Tauri does not support visionOS.
4
5 Project layout and build-target conventions (window sizing, `xros`/`xrsimulator` platform settings, App Group entitlements, solid-image-stack app icon) follow the `SplitFireAIVision` target in the sibling `splitfire` repo (`../splitfire/frontend/xcode/splitfire/SplitFireAI.xcodeproj`). The chat engine wiring (streaming bridge, model lifecycle, conversation persistence) uses Ed's `stream`, `load(gguf:)` and `restoreHistory`.
6
7 ## Setup
8
9 Install [XcodeGen](https://github.com/yonaskolb/XcodeGen) if you don't have it:
10
11 ```
12 brew install xcodegen
13 ```
14
15 Generate the Xcode project and open it:
16
17 ```
18 cd xcode
19 xcodegen generate
20 open SitiVision.xcodeproj
21 ```
22
23 Set your development team under Signing & Capabilities (defaults to `2TQF86ZACD`, the same team as the Tauri app's iOS build — see `../src-tauri/tauri.conf.json`), pick a Vision Pro simulator or device, hit Run.
24
25 ### Developing against a local Ed Swift checkout
26
27 `project.yml` pulls `Ed` from the published [ed-swift](https://github.com/ondeinference/ed-swift) releases (`from: 1.1.0`), the same way `src-tauri/Cargo.toml` pins `ed-agent` from crates.io. To try an unreleased change, point the package at a sibling `onde-ed-swift` checkout instead:
28
29 ```yaml
30 packages:
31 Ed:
32 path: ../../onde-ed-swift
33 ```
34
35 Then build the visionOS slice in that checkout with `make visionos`. It needs the sibling `onde-ed` Rust workspace (`ED_REPO` defaults to `../onde-ed`) and a nightly Rust toolchain with `-Z build-std`, since visionOS is a tier-3 target.
36
37 ## What it does
38
39 On first launch, it downloads and loads the platform-default model (Qwen 2.5 3B on visionOS, same tier as macOS) and streams replies token-by-token. Conversations persist locally between launches. The model cache lives in the shared `group.com.ondeinference.apps` App Group, so a model already downloaded by the Tauri macOS/iOS build of Siti is reused here instead of being fetched twice.
40
41 Current MVP feature set:
42
43 - token-by-token streaming with a stop button
44 - local conversation persistence between launches
45 - model settings sheet for system prompt + sampling preset
46 - toolbar actions to reload the model or clear the conversation
47 - prompt suggestions when the conversation is empty
48
49 Not yet done: a real visionOS-native layout (ornaments, volumes, spatial windows) — this MVP is a single glass `WindowGroup`, functionally equivalent to the iOS/macOS chat view. Immersive-space exploration is a good next step once the 2D chat flow is verified end-to-end.
50
51 ## Files
52
53 | File | What's in it |
54 |------|---------------|
55 | `SitiVisionApp.swift` | Entry point. Calls `setupInferenceEnvironment()` at launch, sizes the default window. |
56 | `ContentView.swift` | The chat UI: header, empty state, message list, input bar, reload/clear actions, settings sheet. |
57 | `ChatViewModel.swift` | Bridges UniFFI streaming into Swift concurrency, manages model lifecycle, restores history, persists conversation. |
58 | `SitiDefaults.swift` | Siti's system prompt (mirrors `../src-tauri/src/chat/mod.rs`'s `CHAT_SYSTEM_PROMPT`), sampling presets, prompt suggestions. |
59 | `InferenceEnvironment.swift` | Points `HF_HOME`/`TMPDIR` at the shared App Group container so downloads land somewhere visionOS allows and are shared with the Tauri build. |
60 | `SitiVision.entitlements` | Generated by `xcodegen generate` from `project.yml`'s `entitlements.properties` — do not hand-edit, it's overwritten on every generate. |
61 | `project.yml` | XcodeGen spec. `SitiVision.xcodeproj` is generated from this, not checked in. |