cursor-buddy 0.0.10 → 0.0.11

package/README.md

@@ -1,7 +1,13 @@
 # cursor-buddy
 
 
-https://github.com/user-attachments/assets/3cdfe011-aee2-4c8e-b695-34f83a972593
+
+
+https://github.com/user-attachments/assets/def0876a-d63c-4e31-b633-9be3fb2b79b5
+
+
+
+
 
 
 AI Agent that lives in your cursor, built for web apps. Push-to-talk voice assistant that can see your screen and point at things.
@@ -15,6 +21,7 @@ Customize its prompt, pass custom tools, choose between browser or server-side s
 - **DOM snapshot context** — AI sees a token-efficient representation of your visible page structure
 - **Voice responses** — Browser or server TTS, with optional streaming playback
 - **Cursor pointing** — AI can point at UI elements it references
+- **Tool call bubbles** — Visual feedback for tool execution with customizable display
 - **Voice interruption** — Start talking again to cut off current response
 - **Framework agnostic** — Core client written in Typescript, adapter-based architecture
 - **Customizable** — CSS variables, custom components, headless mode
@@ -82,6 +89,89 @@ export default function RootLayout({ children }) {
 
 That's it! Hold **Ctrl+Alt** to speak, release to send.
 
+## How It Works
+
+```mermaid
+flowchart LR
+    subgraph Input
+        A[Hold hotkey] --> B[Mic + Speech Recognition]
+        A --> C[Screenshot + DOM Snapshot]
+    end
+
+    subgraph Transcription
+        B --> D{Browser transcript?}
+        D -->|Yes| E[Use browser transcript]
+        D -->|No| F[Server transcription]
+    end
+
+    subgraph Processing
+        E --> G[Send to AI with context]
+        F --> G
+        C --> G
+        G --> H[AI Response]
+        H -->|point tool called| I[Animate cursor to @ID]
+    end
+
+    subgraph Output
+        H --> J[Speak response via TTS]
+    end
+```
+
+1. User holds the hotkey
+2. Microphone captures audio and browser speech recognition starts when available
+3. At the same time, a screenshot and token-efficient DOM snapshot of the viewport are captured in the background. This runs in parallel with speech capture to minimize latency
+4. User releases hotkey
+5. The client prefers the browser transcript; if it is unavailable or empty in `auto` mode, the recorded audio is transcribed on the server
+6. The already-captured screenshot + DOM snapshot are sent to the AI model. Each element has an `@ID` (e.g., `@12`) that the AI can reference.
+7. AI responds with text and can optionally call the `point` tool to indicate an element on screen by its `@ID` from the DOM snapshot
+8. Response is spoken in the browser or on the server based on `speech.mode`,
+    and can either wait for the full response or stream sentence-by-sentence
+    based on `speech.allowStreaming`
+9. If the AI calls the point tool, the cursor animates to the target element's current position (it resolves the element from the snapshot registry and computes its center point)
+10. **If user presses hotkey again at any point, current response is interrupted**
+
+## DOM Snapshot
+
+The DOM snapshot is a token-efficient representation of the visible page structure that gives the AI context about what's on screen.
+
+When the user holds the hotkey, cursor-buddy traverses the visible DOM and builds a lightweight text representation. Each interactive or meaningful element is assigned a unique `@ID` that the AI can reference when pointing.
+
+- **Enables pointing** — The AI can say "click the submit button @42" and the cursor will animate to that exact element
+- **Token efficient** — Only visible, relevant elements are included (no hidden elements, scripts, or styles)
+- **Semantic context** — The AI understands the page structure, not just pixels from the screenshot
+
+For a simple login form, the snapshot might look like:
+
+```
+# viewport 1280x720
+@20 body "Sign In Email Password Sign In Forgot password?" [x=0 y=0 w=1280 h=720]
+  @19 main "Sign In Email Password Sign In Forgot password?" [x=440 y=200 w=400 h=320]
+    @18 form "Sign In Email Password Sign In Forgot password?" [x=440 y=200 w=400 h=320]
+      @1 h1 "Sign In" [x=580 y=220 w=120 h=32]
+      @4 div "Email" [x=460 y=270 w=360 h=56]
+        @2 label "Email" [x=460 y=270 w=40 h=20]
+        @3 input [type="email"] [placeholder="Enter your email"] [x=460 y=294 w=360 h=32]
+      @7 div "Password" [x=460 y=340 w=360 h=56]
+        @5 label "Password" [x=460 y=340 w=64 h=20]
+        @6 input [type="password"] [placeholder="Enter your password"] [x=460 y=364 w=360 h=32]
+      @8 button "Sign In" [type="submit"] [x=460 y=420 w=360 h=40]
+      @9 a "Forgot password?" [href="/forgot"] [x=540 y=476 w=120 h=20]
+```
+
+Each line contains: `@ID tag "text content" [attributes] [bounding box]`
+
+The AI sees this alongside the screenshot. When it wants to guide the user to enter their email, it can call `point(@3)` and the cursor will animate to that input field.
+
+### What Gets Captured
+
+| Included | Excluded |
+|----------|----------|
+| Visible elements in viewport | Hidden elements (`display: none`, `visibility: hidden`) |
+| Interactive elements (buttons, inputs, links) | Script and style tags |
+| Text content (truncated if long) | Elements outside viewport |
+| Element attributes (type, placeholder, href) | Inline styles and classes |
+| Semantic structure | Comment nodes |
+
 ## Server Configuration
 
 ```ts
@@ -136,12 +226,22 @@ createCursorBuddyHandler({
   speechBubble={(props) => <CustomBubble {...props} />}
   waveform={(props) => <CustomWaveform {...props} />}
 
+  // Tool display configuration
+  toolDisplay={{
+    "*": { minDisplayTime: 1500 },           // Default for all tools
+    web_search: { label: "Searching..." },   // Custom label
+    internal_tool: { mode: "hidden" },       // Hide from UI
+  }}
+  renderToolBubble={(props) => <CustomToolBubble {...props} />}
+
   // Callbacks
   onTranscript={(text) => {}}    // Called when speech is transcribed
   onResponse={(text) => {}}      // Called when AI responds
   onPoint={(target) => {}}       // Called when AI points at element
   onStateChange={(state) => {}}  // Called on state change
   onError={(error) => {}}        // Called on error
+  onToolCall={(event) => {}}     // Called when a tool is invoked
+  onToolResult={(event) => {}}   // Called when a tool completes
 />
 ```
 
@@ -170,6 +270,65 @@ createCursorBuddyHandler({
 - `speech.allowStreaming: true` — Speak completed sentence segments as the chat
   stream arrives.
 
+## Tool Display
+
+When the AI uses tools (like web search), bubbles appear near the cursor showing the tool's status. Configure how tools are displayed:
+
+```tsx
+<CursorBuddy
+  endpoint="/api/cursor-buddy"
+  toolDisplay={{
+    // Default settings for all tools
+    "*": {
+      minDisplayTime: 1500,  // Minimum time to show bubble (ms)
+    },
+
+    // Per-tool configuration
+    web_search: {
+      label: "Searching the web...",  // Static label
+      // Or dynamic label based on status:
+      // label: (args, status) => status === "completed" ? "Found results" : "Searching..."
+    },
+
+    // Hide internal tools from UI
+    internal_logging: {
+      mode: "hidden",
+    },
+
+    // Custom render for specific tool
+    data_fetch: {
+      render: (props) => (
+        <div className="my-custom-bubble">
+          {props.status === "pending" ? "Loading..." : "Done!"}
+        </div>
+      ),
+    },
+  }}
+/>
+```
+
+### Tool Call States
+
+| Status | Description |
+|--------|-------------|
+| `pending` | Tool called, waiting for result |
+| `awaiting_approval` | Needs user consent (for tools with `needsApproval`) |
+| `approved` | User approved, executing |
+| `denied` | User denied the tool call |
+| `completed` | Finished successfully |
+| `failed` | Execution failed |
+
+### Approval Keyboard Shortcuts
+
+When a tool requires approval, use these keyboard shortcuts:
+
+| Key | Action |
+|-----|--------|
+| **Y** or **Enter** | Approve the tool call |
+| **N** or **Escape** | Deny the tool call |
+
+Shortcuts are automatically enabled when a tool is awaiting approval and disabled otherwise. They are ignored when focus is in an input field or textarea.
+
 ## Customization
 
 ### CSS Variables
@@ -192,6 +351,14 @@ Cursor buddy styles are customizable via CSS variables. Override them in your st
 
   /* Waveform */
   --cursor-buddy-waveform-color: #ef4444;
+
+  /* Tool bubbles */
+  --cursor-buddy-tool-bg: #ffffff;
+  --cursor-buddy-tool-text: #1f2937;
+  --cursor-buddy-tool-pending: #3b82f6;
+  --cursor-buddy-tool-approval: #f59e0b;
+  --cursor-buddy-tool-success: #22c55e;
+  --cursor-buddy-tool-error: #ef4444;
 }
 ```
 
@@ -245,6 +412,11 @@ function MyCustomUI() {
     isPointing,
     error,
 
+    // Tool state
+    toolCalls,       // All tool calls in current turn
+    activeToolCalls, // Visible, non-expired tool calls
+    pendingApproval, // Tool awaiting user approval, or null
+
     // Actions
     startListening,
     stopListening,
@@ -252,6 +424,11 @@ function MyCustomUI() {
     pointAt,         // Manually point at coordinates
     dismissPointing,
     reset,
+
+    // Tool actions
+    approveToolCall, // Approve a pending tool call
+    denyToolCall,    // Deny a pending tool call
+    dismissToolCall, // Dismiss a tool call bubble
   } = useCursorBuddy()
 
   return (
@@ -264,6 +441,19 @@ function MyCustomUI() {
       >
         Hold to speak
       </button>
+
+      {/* Render active tool calls */}
+      {activeToolCalls.map((tool) => (
+        <div key={tool.id}>
+          {tool.label}
+          {tool.status === "awaiting_approval" && (
+            <>
+              <button onClick={() => approveToolCall(tool.id)}>Yes</button>
+              <button onClick={() => denyToolCall(tool.id)}>No</button>
+            </>
+          )}
+        </div>
+      ))}
     </div>
   )
 }
@@ -362,21 +552,11 @@ client.stopListening()
 | `CursorRenderProps` | Props passed to custom cursor |
 | `SpeechBubbleRenderProps` | Props passed to custom speech bubble |
 | `WaveformRenderProps` | Props passed to custom waveform |
-
-## How It Works
-
-1. User holds the hotkey
-2. Microphone captures audio, waveform shows audio level, and browser speech recognition starts when available
-3. At the same time, a screenshot and token-efficient DOM snapshot of the viewport are captured in the background. This runs in parallel with speech capture to minimize latency
-4. User releases hotkey
-5. The client prefers the browser transcript; if it is unavailable or empty in `auto` mode, the recorded audio is transcribed on the server
-6. The already-captured screenshot + DOM snapshot are sent to the AI model. Each element has an `@ID` (e.g., `@12`) that the AI can reference.
-7. AI responds with text and can optionally call the `point` tool to indicate an element on screen by its `@ID` from the DOM snapshot
-8. Response is spoken in the browser or on the server based on `speech.mode`,
-    and can either wait for the full response or stream sentence-by-sentence
-    based on `speech.allowStreaming`
-9. If the AI calls the point tool, the cursor animates to the target element's current position (it resolves the element from the snapshot registry and computes its center point)
-10. **If user presses hotkey again at any point, current response is interrupted**
+| `ToolBubbleRenderProps` | Props passed to custom tool bubble |
+| `ToolCallState` | State of a tool call |
+| `ToolCallStatus` | `"pending" \| "awaiting_approval" \| "approved" \| "denied" \| "completed" \| "failed"` |
+| `ToolDisplayConfig` | Configuration for tool display |
+| `ToolDisplayOptions` | Options for a single tool |
 
 ## Security Best Practices
 
@@ -407,14 +587,8 @@ export async function POST(request: Request) {
 
   return handler(request)
 }
-
-export const GET = POST
 ```
 
-## TODOs
-
-- [ ] Medium: Proper test structure without relying on `as any` for audio and voice capture
-
 ## License
 
 MIT