chatGLM

中文说明 / Chinese README: README-zh.md

Cross‑platform SwiftUI client for Zhipu GLM models.
Provides a ChatGPT‑style experience with chat, image generation, video generation, and image understanding, targeting iOS, iPadOS, macOS, and visionOS.

On macOS, chatGLM can read the file currently being edited in Xcode / VSCode and other editors via Accessibility and Automation (Apple Events).
Details: MacEditorIntegration.md.

---

Features

• Multi‑platform SwiftUI app

  • Single codebase targeting iOS, iPadOS, macOS, and visionOS.
  • Adaptive layouts for iPhone, iPad, and Mac, with animated background and chat bubbles.

• Chat, images, and video

  • Chat mode with streaming responses.
  • Image generation mode (text‑to‑image).
  • Video generation mode (text‑to‑video).

• Image understanding

  • Attach a local image.
  • The app extracts basic info (size, format, etc.) and injects it as extra context for the chat model.

• macOS editor integration

  • Discover common editors (Xcode, VSCode, JetBrains IDEs, terminal tools, etc.).
  • With user‑granted Accessibility + Automation permissions, read the current file (or at least the current editor buffer) from the target app and attach it as context for the model.
  • Current version is read‑only: it doesn’t auto‑modify your code files; it focuses on analysis and suggestions.
  • See MacEditorIntegration.md and chatGLM/MacEditorIntegration/.

---

Architecture Overview

High‑level structure:

• chatGLMApp.swift

  • App entry point.
  • Loads the shared chat UI.

• ContentView.swift

  • Main SwiftUI view.
  • Provides the ChatGPT‑style chat interface, background gradient, message list, mode picker, and the “rainbow glow” input bar.
  • On macOS, integrates the editor picker button that lets you choose a running editor app and attach its current file as context.

• ChatModels.swift

  • Domain models:
    • ChatMessage – user/assistant messages, optional image/file attachment metadata.
    • ChatMode – .chat, .image, .video.

• ChatViewModel.swift

  • Core business logic:
    • Maintains message history, input state, and current mode.
    • Chat mode:
      • Calls the async chat completion API.
      • Polls the async result.
      • Streams the final answer into the UI with a “typewriter” effect.
    • Image mode:
      • Calls the image generation API.
      • Displays returned image URLs.
    • Video mode:
      • Calls the video generation async API.
      • Shows a card with the resulting video URL.
    • File / image context:
      • For attached images: injects a short description into the prompt.
      • For attached files (from file picker or editor integration): adds a truncated snippet plus path to the prompt so the model can reason about the code.

• PlatformImage.swift, MultilineInputView.swift, etc.

  • Small platform‑adaptation helpers and UI components.

macOS‑specific integration

• Folder: chatGLM/MacEditorIntegration/
  • EditorIntegration.swift
    • Discovers known editors using NSWorkspace and NSRunningApplication.
    • Uses Accessibility (AXUIElement) to try to obtain the current document path via:
      • kAXFocusedWindowAttribute
      • kAXMainWindowAttribute
      • kAXFocusedUIElementAttribute + climbing parents to find kAXDocumentAttribute.
    • For Xcode:
      • First attempts AppleScript to get the current document’s POSIX path.
      • If both Accessibility and AppleScript fail to yield a path, it falls back to:
        • Traversing the Accessibility window tree to find the code editor text area.
        • Reading the full editor buffer text.
        • Writing it to a temporary file inside chatGLM’s sandbox.
        • Returning that temp file URL as “current file” context.
    • Exposes this via EditorIntegrationManager and EditorAppPickerView.
  • SettingsWindowController.swift
    • Hosts SettingsView inside an NSWindow on macOS, separate from the main chat window.

For a detailed explanation of permissions and behavior, see MacEditorIntegration.md.

---

LLM Networking Layer

• Folder: chatGLM/LLM/

• ZhipuModels.swift

  • Request/response types for Zhipu APIs:
    • Async chat: ZhipuAsyncChatRequest, ZhipuAsyncTaskResponse.
    • Image generation: ZhipuImageGenerationRequest, ZhipuImageGenerationResponse.
    • Video generation: ZhipuVideoGenerationRequest.
    • Async result polling: ZhipuAsyncResultResponse.
  • ZhipuModel enum with common models:
    • glm-4.5-flash – chat.
    • glm-4.1v-thinking-flash – vision / image understanding.
    • cogview-3-flash – image generation.
    • cogvideox-flash – video generation.

• ZhipuAPIClient.swift

  • A thin wrapper around URLSession with convenient methods:
    • createAsyncChatCompletion → /api/paas/v4/async/chat/completions
    • createImageGeneration → /api/paas/v4/images/generations
    • createVideoGeneration → /api/paas/v4/videos/generations
    • fetchAsyncResult → /api/paas/v4/async-result/{id}

No third‑party networking library is required.

---

Configuration & Running

1. Open the project

   • Open chatGLM.xcodeproj in Xcode.

2. Configure your Zhipu API key

   • In your Run scheme, add an environment variable:

     • ZHIPU_API_KEY=<your Zhipu API key>

3. Run on a target platform

   • Select the desired destination:
     • iOS Simulator / physical device.
     • My Mac (macOS app).
     • visionOS simulator (if installed).
   • Build and run.

4. Grant macOS permissions (optional, but recommended)

   • For editor integration on macOS:
     • In “Privacy & Security → Accessibility”, allow chatGLM to control your Mac.
     • On first AppleScript use (e.g., Xcode integration), approve Automation permission when macOS prompts.
   • After that, you can pick a running editor in the chat input bar and chatGLM will attach the current file (or editor buffer) as context for the model.