osborn 0.8.27 → 0.8.29

package/dist/index.js

@@ -554,15 +554,15 @@ async function main() {
      * If interrupted: gather spoken text + JSONL context. Does NOT send to Claude yet —
      * that happens when the user's transcribed message arrives via chat().
      */
-    async function handleSpeechDone(handle, fullText) {
+    async function handleSpeechDone(handle, fullText, fullBlockText) {
         if (!handle.interrupted) {
             lastInterruption = null;
             return;
         }
-        // fullText is what was being spoken when interrupted (passed from tts_say handler).
-        // No word-level cutoff for say() — only generateReply pipeline has that — but Claude
-        // knows its own output from JSONL, so the full block is enough context.
-        console.log(`🔇 Speech interrupted. Was speaking (${fullText.length} chars): "${fullText}"`);
+        // fullText is the synchronized (word-accurate) transcript of what was actually spoken
+        // before the interruption. fullBlockText is the original full TTS segment text.
+        const fullBlockLen = fullBlockText?.length ?? fullText.length;
+        console.log(`🔇 Speech interrupted. Heard (${fullText.length} chars): "${fullText}" [full block was ${fullBlockLen} chars]`);
         // Read last 10 assistant messages from JSONL (Claude's full untruncated output).
         // SessionMessage.text is pre-joined from all text content blocks.
         let recentMessages = '';
@@ -1109,10 +1109,62 @@ async function main() {
                 if (handle && typeof handle.addDoneCallback === 'function') {
                     // SpeechHandle — track it and register interruption callback
                     currentSpeechHandle = handle;
+                    // Wall-clock timer: capture when audio actually starts playing (first frame)
+                    // Used as fallback if LiveKit's playbackPosition is 0 (race condition)
+                    let playbackStartedAt = null;
+                    const audioOutputRef = currentSession?._activity?.agentSession?.output?.audio;
+                    if (audioOutputRef && typeof audioOutputRef.on === 'function') {
+                        const onPlaybackStarted = () => {
+                            playbackStartedAt = Date.now();
+                            console.log(`🔊 [${sayId}] audio first frame out (playbackStarted)`);
+                            audioOutputRef.off('playbackStarted', onPlaybackStarted);
+                        };
+                        audioOutputRef.on('playbackStarted', onPlaybackStarted);
+                    }
                     handle.addDoneCallback((sh) => {
                         if (sh.interrupted) {
                             console.log(`🔇 [${sayId}] session.say INTERRUPTED`);
-                            handleSpeechDone(sh, data.text);
+                            const audioOutput = currentSession?._activity?.agentSession?.output?.audio;
+                            const sdkTranscript = audioOutput?.lastPlaybackEvent?.synchronizedTranscript;
+                            const sdkPlaybackSec = audioOutput?.lastPlaybackEvent?.playbackPosition ?? 0;
+                            let spokenText;
+                            let method;
+                            if (sdkTranscript) {
+                                // Best case: LiveKit gave us word-accurate transcript (requires alignedTranscript TTS)
+                                spokenText = sdkTranscript;
+                                method = 'sdk-transcript';
+                            }
+                            else if (sdkPlaybackSec > 0) {
+                                // Second: LiveKit gave us playback duration — estimate chars from it
+                                const CHARS_PER_SEC = 14;
+                                const charCount = Math.min(Math.round(sdkPlaybackSec * CHARS_PER_SEC), data.text.length);
+                                const slicePoint = data.text.lastIndexOf(' ', charCount) || charCount;
+                                spokenText = slicePoint > 0 ? data.text.slice(0, slicePoint) : data.text;
+                                method = 'sdk-position';
+                            }
+                            else if (playbackStartedAt !== null) {
+                                // Third: use our wall-clock timer from first audio frame
+                                const elapsedSec = (Date.now() - playbackStartedAt) / 1000;
+                                const CHARS_PER_SEC = 14;
+                                const charCount = Math.min(Math.round(elapsedSec * CHARS_PER_SEC), data.text.length);
+                                const slicePoint = data.text.lastIndexOf(' ', charCount) || charCount;
+                                spokenText = slicePoint > 0 ? data.text.slice(0, slicePoint) : data.text;
+                                method = 'wall-clock';
+                            }
+                            else {
+                                // Fallback: interrupt fired before first frame — pass full block
+                                spokenText = data.text;
+                                method = 'full-block-fallback';
+                            }
+                            console.log('🔇 Interruption estimate:', JSON.stringify({
+                                method,
+                                sdkPlaybackSec,
+                                isSynced: !!sdkTranscript,
+                                spokenChars: spokenText.length,
+                                fullChars: data.text.length,
+                                heard: spokenText.slice(0, 80) + (spokenText.length > 80 ? '...' : '')
+                            }));
+                            handleSpeechDone(sh, spokenText, data.text);
                         }
                         else {
                             console.log(`✅ [${sayId}] session.say DONE`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.8.27",
+  "version": "0.8.29",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {

package/.claude/settings.local.json

@@ -1,9 +0,0 @@
-{
-  "permissions": {
-    "allow": [
-      "Bash(ps:*)",
-      "Bash(osascript:*)",
-      "Bash(curl -s http://localhost:3000)"
-    ]
-  }
-}

package/.claude/skills/browser-apply/SKILL.md

@@ -1,114 +0,0 @@
-# Skill: Browser Apply — Step-by-Step Workday Application
-
-Automate Workday job applications interactively, one step at a time. Each step takes a screenshot, confirms what's on screen, fills the current page, and waits before proceeding.
-
-**This skill uses the Playwright MCP tools** (`mcp__playwright__browser_*`) for direct browser control — no scripts needed.
-
-## When to Use
-- Any Workday ATS application (`*.wd1.myworkdayjobs.com`)
-- Any multi-step JS-heavy job application form
-- When you want visible, confirmable progress at each step
-
-## Key Principle: Step-by-Step, Not One Big Script
-
-Do NOT write a monolithic automation script. Instead:
-1. Navigate to the URL
-2. Take a screenshot — confirm what's on screen
-3. Fill only the current step's fields
-4. Take another screenshot — confirm fields filled correctly
-5. Ask the user "Ready for next step?" before clicking Next
-6. Click Next, wait for page load, screenshot again
-7. Repeat for each step
-
-This approach catches rendering issues, unexpected fields, and errors before they cascade.
-
-## Step-by-Step Execution Pattern
-
-### Step 0 — Open the browser
-Use: `mcp__playwright__browser_navigate` with the applyManually URL
-
-Then: `mcp__playwright__browser_take_screenshot` — show the user what loaded
-
-### Step 1 — Create Account / Sign In
-Take a snapshot with `mcp__playwright__browser_snapshot` to see element refs.
-Fill fields using `mcp__playwright__browser_fill_form` or individual `mcp__playwright__browser_type` calls.
-Screenshot to confirm. Then ask user before clicking Create Account / Sign In.
-
-### Step 2 — Start Application
-If "Start Your Application" screen appears with Apply Manually button:
-Screenshot it. Click "Apply Manually" using `mcp__playwright__browser_click`.
-Screenshot after.
-
-### Step 3 — My Information
-Snapshot → fill each field → screenshot → ask user before clicking Next.
-
-Fields to fill:
-- First Name, Last Name, Phone
-- Address, City, State (dropdown), Zip
-- Work authorization: Yes
-- Sponsorship: No
-
-### Step 4 — My Experience
-Snapshot → click Add for each job entry → fill title/company/dates/description → save each → screenshot.
-Then add education entries.
-Ask user before clicking Next.
-
-### Step 5 — Application Questions
-Snapshot to see all questions. Fill each one. **Always confirm salary expectation with user before filling** — never guess. Screenshot. Ask before Next.
-
-### Step 6 — Voluntary Disclosures
-Select "I do not wish to answer" / "Prefer not to disclose" for all. Screenshot. Ask before Next.
-
-### Step 7 — Self Identify
-Fill name and date. Select disability option. Screenshot. Ask before Next.
-
-### Step 8 — Review
-Screenshot the full review page. Confirm with user before clicking Submit.
-
-### Step 9 — Confirm submission
-Screenshot the confirmation dialog. Save it.
-
-## Candidate Data (Osborn Ojure)
-
-- Email: osbornojure@gmail.com
-- Password: Workday2026!
-- First: Osborn, Last: Ojure
-- Phone: 3127185561
-- Address: 1234 N Michigan Ave, Chicago, IL 60601
-
-Jobs:
-1. Meta API Consultant at Prehype / Audos — April 2024 to Present
-2. Full Stack Developer, Freelance — January 2016 to Present
-
-Education:
-1. A.S. Information Systems
-2. B.S. Psychology
-
-## Workday data-automation-id Selector Reference
-
-| Field | Selector |
-|---|---|
-| Email | `input[type="email"]` |
-| Password | `input[type="password"]` |
-| First name | `[data-automation-id="legalNameSection_firstName"]` |
-| Last name | `[data-automation-id="legalNameSection_lastName"]` |
-| Phone | `[data-automation-id="phone-number"]` |
-| Address | `[data-automation-id="addressSection_addressLine1"]` |
-| City | `[data-automation-id="addressSection_city"]` |
-| Zip | `[data-automation-id="addressSection_postalCode"]` |
-| Job title | `[data-automation-id="jobTitle"]` |
-| Company | `[data-automation-id="company"]` |
-| Description | `[data-automation-id="description"]` |
-| Next button | `[data-automation-id="bottom-navigation-next-btn"]` |
-| Create Account | `[data-automation-id="click_filter"][aria-label="Create Account"]` |
-
-## Critical Rules
-- headless: false always (Workday renders blank in headless)
-- Confirm salary with user before every submission — never auto-fill
-- After "You already applied to this job" error — that confirms a previous submission worked
-- Use `{ force: true }` on Workday buttons — overlay click filters block normal clicks
-- Always wait for networkidle or waitForSelector after navigation before interacting
-
-## Playwright Install Location
-Run scripts from: `/Users/newupgrade/Desktop/Developer/osborn/frontend`
-(playwright is in `node_modules` there)

package/.claude/skills/markdown-to-pdf/SKILL.md

@@ -1,29 +0,0 @@
-# Skill: Markdown to PDF
-
-Export Markdown documents as formatted PDF files.
-
-## When to use
-When the user wants to create a PDF from a Markdown file, spec, or research findings.
-
-## How to execute
-
-Option 1 — Using md-to-pdf (best quality):
-```bash
-npx --yes md-to-pdf "<MARKDOWN_PATH>"
-```
-This creates a PDF alongside the source file with the same name.
-
-Option 2 — Using pandoc (if available):
-```bash
-pandoc "<MARKDOWN_PATH>" -o "<OUTPUT_PATH>.pdf" --pdf-engine=wkhtmltopdf
-```
-
-Option 3 — Using markdown-pdf:
-```bash
-npx --yes markdown-pdf "<MARKDOWN_PATH>" -o "<OUTPUT_PATH>.pdf"
-```
-
-## Output
-- Save the PDF to the session workspace (e.g., `library/{name}.pdf`)
-- Confirm the output path and file size to the user
-- If the source is spec.md, name the output `spec-export.pdf`

package/.claude/skills/pdf-to-markdown/SKILL.md

@@ -1,28 +0,0 @@
-# Skill: PDF to Markdown
-
-Convert PDF documents to readable Markdown text.
-
-## When to use
-When the user provides a PDF file path and wants to read, search, or work with its contents.
-
-## How to execute
-
-Option 1 — Using the built-in Read tool:
-The Read tool can directly read PDF files. Use `pages` parameter for large PDFs (max 20 pages per request).
-
-Option 2 — Full extraction via CLI (for better formatting or batch processing):
-```bash
-npx --yes pdf-parse-cli "<PDF_PATH>"
-```
-
-Option 3 — Using pdftotext (if available):
-```bash
-pdftotext -layout "<PDF_PATH>" -
-```
-
-## Output
-Save the converted content to the session workspace as `library/{filename}.md` with:
-- Document title and source path at the top
-- Preserved heading structure where detectable
-- Tables converted to Markdown tables where possible
-- Page numbers as section markers

package/.claude/skills/playwright-browser/SKILL.md

@@ -1,90 +0,0 @@
-# Skill: Playwright Browser Automation
-
-Automate web browser interactions — navigate pages, click buttons, fill forms, take screenshots, and extract content.
-
-## When to use
-- Navigate to a URL and interact with it
-- Click buttons or links by their text or role
-- Fill form fields and submit data
-- Take screenshots of web pages
-- Extract text or structured data from pages
-- Automate multi-step web workflows (e.g. join a room, test a UI flow)
-
-## How to execute
-
-Uses `@playwright/cli` via npx — no global install needed. Token-efficient: uses element references (e.g. `e15`) instead of pixel coordinates.
-
-### First time only — install browser binaries
-```bash
-npx playwright install chromium
-```
-
-### Step 1 — Open a URL
-```bash
-npx @playwright/cli open https://localhost:3000
-```
-
-### Step 2 — Get page structure and element references
-```bash
-npx @playwright/cli snapshot
-```
-Returns an accessibility tree with element IDs like e1, e2, e15. Use these in subsequent commands.
-
-### Step 3 — Interact with elements
-```bash
-npx @playwright/cli click e15
-npx @playwright/cli fill e3 "some text"
-npx @playwright/cli press e3 Enter
-npx @playwright/cli select e7 "optionValue"
-npx @playwright/cli check e9
-npx @playwright/cli hover e12
-```
-
-### Take a screenshot
-```bash
-npx @playwright/cli screenshot --path=/tmp/page.png
-```
-
-### Take a screenshot at a specific viewport size (mobile check)
-```bash
-npx @playwright/cli screenshot --viewport-size=375,812 --path=/tmp/page-mobile.png
-```
-Common mobile sizes: `375,812` (iPhone 14), `390,844` (iPhone 14 Pro), `412,915` (Pixel 7), `768,1024` (iPad).
-
-### Close the browser
-```bash
-npx @playwright/cli close
-```
-
-### Named sessions (persistent state across commands)
-```bash
-npx @playwright/cli -s=myflow open https://localhost:3000
-npx @playwright/cli -s=myflow snapshot
-npx @playwright/cli -s=myflow fill e3 "abc123"
-npx @playwright/cli -s=myflow click e5
-npx @playwright/cli -s=myflow close
-```
-
-## Complete example — join Osborn voice room
-```bash
-npx @playwright/cli open http://localhost:3000
-npx @playwright/cli snapshot
-npx @playwright/cli fill e3 "abc123"
-npx @playwright/cli click e4
-npx @playwright/cli screenshot --path=/tmp/osborn-joined.png
-npx @playwright/cli close
-```
-
-## Complete example — check mobile layout
-```bash
-npx @playwright/cli open http://localhost:3000
-npx @playwright/cli screenshot --viewport-size=375,812 --path=/tmp/mobile-375.png
-npx @playwright/cli close
-```
-
-## Notes
-- Runs headless by default. Add --headed to see the browser window.
-- Install browsers first if needed: npx playwright install chromium
-- Element IDs are session-scoped — run snapshot again after page changes
-- Use `--viewport-size=WIDTH,HEIGHT` to simulate mobile screen sizes (e.g. `375,812` for iPhone 14)
-- Use `--storage-state=/tmp/state.json` to save and restore session state (cookies, localStorage) across runs

package/.claude/skills/shadcn/SKILL.md

@@ -1,232 +0,0 @@
-# Skill: shadcn/ui Components
-
-Add and configure shadcn/ui components in a Next.js or React project.
-
-## When to use
-When the user wants to add UI components (buttons, dialogs, cards, forms, tables, etc.) using shadcn/ui — the copy-paste component library built on Radix UI and Tailwind CSS.
-
-## Setup (first time only)
-
-Initialize shadcn in the project root (where package.json lives):
-```bash
-npx shadcn@latest init
-```
-
-This creates `components.json` and sets up `src/components/ui/`. Answer the prompts to match your project's style preferences.
-
-## Add a component
-
-```bash
-npx shadcn@latest add <component-name>
-```
-
-## Add multiple components at once
-```bash
-npx shadcn@latest add button card dialog input form
-```
-
-## Commonly used components
-
-| Component | Install name | Description |
-|-----------|-------------|-------------|
-| Button | `button` | Clickable button with variants |
-| Card | `card` | Container with header/content/footer |
-| Input | `input` | Text input field |
-| Label | `label` | Form label |
-| Textarea | `textarea` | Multi-line text input |
-| Select | `select` | Dropdown select |
-| Checkbox | `checkbox` | Checkbox with label |
-| Switch | `switch` | Toggle switch |
-| Dialog | `dialog` | Modal dialog |
-| Sheet | `sheet` | Slide-in panel (drawer) |
-| Popover | `popover` | Floating content panel |
-| Tooltip | `tooltip` | Hover tooltip |
-| Dropdown Menu | `dropdown-menu` | Contextual dropdown |
-| Command | `command` | Command palette / search |
-| Badge | `badge` | Small status indicator |
-| Avatar | `avatar` | User avatar with fallback |
-| Separator | `separator` | Visual divider |
-| Table | `table` | Data table |
-| Tabs | `tabs` | Tabbed interface |
-| Accordion | `accordion` | Collapsible sections |
-| Sonner | `sonner` | Toast notifications (preferred over toast) |
-| Alert | `alert` | Inline alert message |
-| Alert Dialog | `alert-dialog` | Confirmation dialog |
-| Progress | `progress` | Progress bar |
-| Skeleton | `skeleton` | Loading placeholder |
-| Scroll Area | `scroll-area` | Custom scrollbar container |
-| Calendar | `calendar` | Date picker calendar |
-| Form | `form` | React Hook Form integration |
-| Slider | `slider` | Range slider |
-| Toggle | `toggle` | Toggle button |
-| Navigation Menu | `navigation-menu` | Site navigation |
-| Breadcrumb | `breadcrumb` | Page navigation breadcrumbs |
-| Collapsible | `collapsible` | Expandable/collapsible section |
-| Context Menu | `context-menu` | Right-click context menu |
-| Menubar | `menubar` | Application menu bar |
-| Resizable | `resizable` | Resizable panel groups |
-
-## Import pattern
-
-```tsx
-import { Button } from "@/components/ui/button"
-import { Card, CardContent, CardHeader, CardTitle, CardDescription, CardFooter } from "@/components/ui/card"
-import { Input } from "@/components/ui/input"
-import { Label } from "@/components/ui/label"
-import {
-  Dialog,
-  DialogContent,
-  DialogDescription,
-  DialogHeader,
-  DialogTitle,
-  DialogTrigger,
-} from "@/components/ui/dialog"
-```
-
-## Example — Button variants
-
-```tsx
-<Button>Default</Button>
-<Button variant="destructive">Delete</Button>
-<Button variant="outline">Cancel</Button>
-<Button variant="ghost">Ghost</Button>
-<Button variant="link">Link</Button>
-<Button size="sm">Small</Button>
-<Button size="lg">Large</Button>
-<Button disabled>Disabled</Button>
-```
-
-## Example — Card
-
-```tsx
-<Card>
-  <CardHeader>
-    <CardTitle>Card Title</CardTitle>
-    <CardDescription>Card description goes here</CardDescription>
-  </CardHeader>
-  <CardContent>
-    <p>Card content here</p>
-  </CardContent>
-  <CardFooter>
-    <Button>Action</Button>
-  </CardFooter>
-</Card>
-```
-
-## Example — Form with validation (React Hook Form + Zod)
-
-```bash
-npx shadcn@latest add form input
-npm install zod react-hook-form @hookform/resolvers
-```
-
-```tsx
-import { useForm } from "react-hook-form"
-import { zodResolver } from "@hookform/resolvers/zod"
-import * as z from "zod"
-import { Form, FormControl, FormField, FormItem, FormLabel, FormMessage } from "@/components/ui/form"
-import { Input } from "@/components/ui/input"
-import { Button } from "@/components/ui/button"
-
-const formSchema = z.object({
-  email: z.string().email("Invalid email address"),
-  password: z.string().min(8, "Password must be at least 8 characters"),
-})
-
-export function LoginForm() {
-  const form = useForm<z.infer<typeof formSchema>>({
-    resolver: zodResolver(formSchema),
-    defaultValues: { email: "", password: "" },
-  })
-
-  function onSubmit(values: z.infer<typeof formSchema>) {
-    console.log(values)
-  }
-
-  return (
-    <Form {...form}>
-      <form onSubmit={form.handleSubmit(onSubmit)} className="space-y-4">
-        <FormField
-          control={form.control}
-          name="email"
-          render={({ field }) => (
-            <FormItem>
-              <FormLabel>Email</FormLabel>
-              <FormControl>
-                <Input placeholder="you@example.com" {...field} />
-              </FormControl>
-              <FormMessage />
-            </FormItem>
-          )}
-        />
-        <FormField
-          control={form.control}
-          name="password"
-          render={({ field }) => (
-            <FormItem>
-              <FormLabel>Password</FormLabel>
-              <FormControl>
-                <Input type="password" {...field} />
-              </FormControl>
-              <FormMessage />
-            </FormItem>
-          )}
-        />
-        <Button type="submit" className="w-full">Sign in</Button>
-      </form>
-    </Form>
-  )
-}
-```
-
-## Example — Toast notifications (Sonner)
-
-```bash
-npx shadcn@latest add sonner
-```
-
-```tsx
-// In your root layout — add the Toaster once
-import { Toaster } from "@/components/ui/sonner"
-export default function RootLayout({ children }) {
-  return (
-    <html>
-      <body>
-        {children}
-        <Toaster />
-      </body>
-    </html>
-  )
-}
-
-// Then anywhere in your app
-import { toast } from "sonner"
-
-toast("Event created")
-toast.success("Profile saved")
-toast.error("Something went wrong")
-toast.promise(saveData(), {
-  loading: "Saving...",
-  success: "Saved!",
-  error: "Failed to save",
-})
-```
-
-## The cn() utility
-
-shadcn uses `clsx` + `tailwind-merge` via a `cn()` helper for conditional classes:
-
-```tsx
-import { cn } from "@/lib/utils"
-
-<div className={cn("base-class", isActive && "active-class", className)} />
-```
-
-## Notes
-- Components are **copied into your project** (not a node_modules dependency) — edit them freely
-- Requires **Tailwind CSS** configured in the project
-- Uses **Radix UI** primitives for accessibility
-- Components land in `src/components/ui/` by default
-- Run `npx shadcn@latest diff` to see if upstream components have changed
-- Check `components.json` for path and style config
-- Use Sonner (`sonner`) instead of the legacy `toast` component for notifications

package/.claude/skills/youtube-transcript/SKILL.md

@@ -1,24 +0,0 @@
-# Skill: YouTube Transcript
-
-Fetch and save transcripts from YouTube videos.
-
-## When to use
-When the user asks to get a transcript, subtitles, captions, or summary from a YouTube video URL.
-
-## How to execute
-
-yt-dlp is installed on this system. Use this exact command:
-
-```bash
-yt-dlp --skip-download --write-auto-sub --sub-lang en --convert-subs srt -o "/tmp/yt-%(id)s" "<VIDEO_URL>"
-```
-
-This downloads auto-generated English subtitles as an SRT file to /tmp/yt-{video-id}.en.srt
-
-Then read the SRT file and strip the timing markers to get clean transcript text.
-
-## Output
-Save the cleaned transcript to the session workspace as `library/youtube-{video-id}-transcript.md` with:
-- Video title and URL at the top
-- Cleaned transcript text (strip SRT timing markers and duplicate lines)
-- Key timestamps preserved as section markers if meaningful breaks exist

package/caresource-apply.js

@@ -1,50 +0,0 @@
-const { chromium } = require('playwright');
-
-(async () => {
-  const browser = await chromium.launch({ headless: false });
-  const page = await browser.newPage();
-  
-  await page.goto('https://caresource.wd1.myworkdayjobs.com/CareSource/job/Remote/AI-Developer_R10487/apply', { waitUntil: 'networkidle' });
-  await page.waitForTimeout(2000);
-  
-  // Click Apply Manually
-  console.log('Clicking Apply Manually...');
-  await page.click('text=Apply Manually');
-  await page.waitForTimeout(3000);
-  
-  // Fill in email
-  console.log('Filling email field...');
-  const emailInput = await page.$('input[type="email"]');
-  if (emailInput) {
-    await emailInput.fill('osbornojure@gmail.com');
-    console.log('Email filled: osbornojure@gmail.com');
-  }
-  
-  // Fill in password
-  console.log('Filling password fields...');
-  const passwordInputs = await page.$$('input[type="password"]');
-  if (passwordInputs.length >= 2) {
-    await passwordInputs[0].fill('workday2026!');
-    await passwordInputs[1].fill('workday2026!');
-    console.log('Passwords filled');
-  }
-  
-  await page.waitForTimeout(1000);
-  
-  // Click Create Account
-  const createBtn = await page.$('text=Create Account');
-  if (createBtn) {
-    console.log('Clicking Create Account...');
-    await createBtn.click();
-    await page.waitForTimeout(3000);
-  }
-  
-  // Take screenshot
-  await page.screenshot({ path: '/tmp/caresource-step1.png' });
-  console.log('Screenshot saved to /tmp/caresource-step1.png');
-  
-  // Get current page text
-  const text = await page.evaluate(() => document.body.innerText);
-  console.log('\n--- PAGE CONTENT ---');
-  console.log(text);
-})();

package/caresource-apply.mjs

@@ -1,34 +0,0 @@
-import { chromium } from 'playwright';
-
-(async () => {
-  const browser = await chromium.launch({ headless: false });
-  const page = await browser.newPage();
-  
-  console.log('🌐 Navigating to CareSource...');
-  await page.goto('https://caresource.wd1.myworkdayjobs.com/CareSource/job/Remote/AI-Developer_R10487/apply', { waitUntil: 'networkidle' });
-  await page.waitForTimeout(2000);
-  
-  console.log('📋 Clicking Apply Manually...');
-  await page.click('text=Apply Manually');
-  await page.waitForTimeout(3000);
-  
-  console.log('📧 Filling email...');
-  const emailInput = await page.$('input[type="email"]');
-  if (emailInput) {
-    await emailInput.fill('osbornojure@gmail.com');
-  }
-  
-  console.log('🔐 Filling passwords...');
-  const passwordInputs = await page.$$('input[type="password"]');
-  if (passwordInputs.length >= 2) {
-    await passwordInputs[0].fill('workday2026!');
-    await passwordInputs[1].fill('workday2026!');
-  }
-  
-  await page.waitForTimeout(1000);
-  await page.screenshot({ path: '/tmp/caresource-step1.png' });
-  
-  console.log('✅ Account form filled. Screenshot saved.');
-  console.log('\n📝 Browser is now open. You can review before submitting.');
-  console.log('Keep this window open and ready to proceed.\n');
-})();

package/dist/conversation-brain.d.ts

@@ -1,92 +0,0 @@
-/**
- * Conversation Brain - Gemini 2.5 Pro powered conversation manager
- *
- * This is the "smart brain" that:
- * 1. Keeps conversation alive with relevant questions
- * 2. Builds context until we understand what user wants
- * 3. Dispatches background research agents
- * 4. Receives progress updates and decides when to execute
- * 5. Handles direct commands immediately
- */
-export interface ConversationMessage {
-    role: 'user' | 'assistant' | 'system';
-    content: string;
-    timestamp: Date;
-}
-export interface ResearchTask {
-    id: string;
-    query: string;
-    status: 'pending' | 'running' | 'completed' | 'failed';
-    result?: string;
-    startedAt?: Date;
-    completedAt?: Date;
-}
-export interface BrainDecision {
-    action: 'speak' | 'research' | 'execute' | 'clarify' | 'direct_command';
-    speech?: string;
-    researchQueries?: string[];
-    executeTask?: string;
-    directCommand?: string;
-    reasoning?: string;
-}
-export interface BrainState {
-    conversationHistory: ConversationMessage[];
-    userGoal: string | null;
-    userGoalConfidence: number;
-    pendingResearch: ResearchTask[];
-    completedResearch: ResearchTask[];
-    readyToExecute: boolean;
-    executionPlan: string | null;
-}
-export declare class ConversationBrain {
-    private llm;
-    private state;
-    private workingDir;
-    private onSpeak;
-    private onStateChange;
-    constructor(config: {
-        workingDir: string;
-        onSpeak: (text: string) => Promise<void>;
-        onStateChange: (state: string) => Promise<void>;
-    });
-    /**
-     * Process user input and decide what to do
-     */
-    processUserInput(input: string): Promise<BrainDecision>;
-    /**
-     * Receive research results from background agents
-     */
-    receiveResearchResult(taskId: string, result: string, success: boolean): void;
-    /**
-     * Check if we should provide a status update
-     */
-    shouldProvideUpdate(): boolean;
-    /**
-     * Generate a status update based on completed research
-     */
-    generateStatusUpdate(): Promise<string | null>;
-    /**
-     * Get current state for debugging/display
-     */
-    getState(): BrainState;
-    /**
-     * Reset conversation state
-     */
-    reset(): void;
-    private buildAnalysisPrompt;
-    private parseDecision;
-    private updateState;
-    /**
-     * Create research tasks for background agents
-     */
-    getPendingResearchTasks(): ResearchTask[];
-    /**
-     * Mark a research task as running
-     */
-    markResearchRunning(taskId: string): void;
-}
-export declare function createConversationBrain(config: {
-    workingDir: string;
-    onSpeak: (text: string) => Promise<void>;
-    onStateChange: (state: string) => Promise<void>;
-}): ConversationBrain;

package/dist/conversation-brain.js

@@ -1,360 +0,0 @@
-/**
- * Conversation Brain - Gemini 2.5 Pro powered conversation manager
- *
- * This is the "smart brain" that:
- * 1. Keeps conversation alive with relevant questions
- * 2. Builds context until we understand what user wants
- * 3. Dispatches background research agents
- * 4. Receives progress updates and decides when to execute
- * 5. Handles direct commands immediately
- */
-import { llm } from '@livekit/agents';
-import * as google from '@livekit/agents-plugin-google';
-// ============================================================
-// Conversation Brain Class
-// ============================================================
-export class ConversationBrain {
-    llm;
-    state;
-    workingDir;
-    onSpeak;
-    onStateChange;
-    constructor(config) {
-        // Use Gemini 2.0 Flash for brain decisions - it's faster and less likely to conflict
-        // with the Gemini Realtime voice session which uses a different model
-        this.llm = new google.LLM({
-            model: 'gemini-2.0-flash',
-            // Set lower temperature for more consistent decisions
-            temperature: 0.3,
-        });
-        // Add error handler to prevent unhandled rejection crashes
-        this.llm.on('error', (err) => {
-            const errorMsg = err.error?.message || String(err);
-            // Only log non-abort errors
-            if (!errorMsg.includes('aborted') && !errorMsg.includes('AbortError')) {
-                console.error('🧠 [Brain LLM] Error:', errorMsg);
-            }
-        });
-        this.workingDir = config.workingDir;
-        this.onSpeak = config.onSpeak;
-        this.onStateChange = config.onStateChange;
-        this.state = {
-            conversationHistory: [],
-            userGoal: null,
-            userGoalConfidence: 0,
-            pendingResearch: [],
-            completedResearch: [],
-            readyToExecute: false,
-            executionPlan: null,
-        };
-    }
-    /**
-     * Process user input and decide what to do
-     */
-    async processUserInput(input) {
-        // Add to conversation history
-        this.state.conversationHistory.push({
-            role: 'user',
-            content: input,
-            timestamp: new Date(),
-        });
-        // Build the analysis prompt
-        const prompt = this.buildAnalysisPrompt(input);
-        let stream = null;
-        let streamError = null;
-        try {
-            // Call Gemini 2.5 Pro for decision
-            const chatCtx = new llm.ChatContext();
-            chatCtx.addMessage({ role: 'user', content: prompt });
-            let response = '';
-            stream = this.llm.chat({ chatCtx });
-            // Create a promise that wraps the stream iteration with proper error handling
-            const collectResponse = async () => {
-                let result = '';
-                try {
-                    for await (const chunk of stream) {
-                        if (chunk.delta?.content) {
-                            result += chunk.delta.content;
-                        }
-                    }
-                }
-                catch (e) {
-                    streamError = e;
-                    throw e;
-                }
-                return result;
-            };
-            // Race the stream against a timeout to prevent hanging
-            const timeoutPromise = new Promise((_, reject) => {
-                setTimeout(() => reject(new Error('Brain timeout after 30s')), 30000);
-            });
-            response = await Promise.race([collectResponse(), timeoutPromise]);
-            // Parse the decision
-            const decision = this.parseDecision(response);
-            // Update state based on decision
-            this.updateState(decision);
-            // Add assistant response to history
-            if (decision.speech) {
-                this.state.conversationHistory.push({
-                    role: 'assistant',
-                    content: decision.speech,
-                    timestamp: new Date(),
-                });
-            }
-            return decision;
-        }
-        catch (err) {
-            const errorMsg = err.message || String(err);
-            // Only log if not an abort (which is expected when user interrupts)
-            if (!errorMsg.includes('aborted')) {
-                console.error('🧠 [Brain] Stream error:', errorMsg);
-            }
-            // Close stream if it exists to prevent further errors
-            if (stream) {
-                try {
-                    stream.close();
-                }
-                catch { }
-            }
-            return {
-                action: 'speak',
-                speech: "I'm having trouble processing that. Could you try again?",
-            };
-        }
-    }
-    /**
-     * Receive research results from background agents
-     */
-    receiveResearchResult(taskId, result, success) {
-        const task = this.state.pendingResearch.find(t => t.id === taskId);
-        if (task) {
-            task.status = success ? 'completed' : 'failed';
-            task.result = result;
-            task.completedAt = new Date();
-            // Move to completed
-            this.state.pendingResearch = this.state.pendingResearch.filter(t => t.id !== taskId);
-            this.state.completedResearch.push(task);
-            console.log(`🧠 Research completed: ${taskId.substring(0, 8)}... (${this.state.completedResearch.length} done)`);
-        }
-    }
-    /**
-     * Check if we should provide a status update
-     */
-    shouldProvideUpdate() {
-        // Provide update if research just completed
-        return this.state.completedResearch.length > 0 &&
-            this.state.pendingResearch.length === 0 &&
-            !this.state.readyToExecute;
-    }
-    /**
-     * Generate a status update based on completed research
-     */
-    async generateStatusUpdate() {
-        if (this.state.completedResearch.length === 0)
-            return null;
-        const researchSummary = this.state.completedResearch
-            .map(r => `- ${r.query}: ${r.result?.substring(0, 200) || 'No result'}`)
-            .join('\n');
-        const prompt = `Based on completed research, provide a brief conversational status update.
-
-RESEARCH RESULTS:
-${researchSummary}
-
-USER GOAL (so far): ${this.state.userGoal || 'Still understanding...'}
-
-Generate a 1-2 sentence update that:
-1. Summarizes what you learned
-2. Either asks a clarifying question OR proposes next steps
-3. Sounds natural and conversational
-
-Just the update text, no JSON.`;
-        let stream = null;
-        try {
-            const chatCtx = new llm.ChatContext();
-            chatCtx.addMessage({ role: 'user', content: prompt });
-            let response = '';
-            stream = this.llm.chat({ chatCtx });
-            for await (const chunk of stream) {
-                if (chunk.delta?.content) {
-                    response += chunk.delta.content;
-                }
-            }
-            return response.trim();
-        }
-        catch (err) {
-            console.error('🧠 [Brain] Status update error:', err.message || err);
-            if (stream) {
-                try {
-                    stream.close();
-                }
-                catch { }
-            }
-            return null;
-        }
-    }
-    /**
-     * Get current state for debugging/display
-     */
-    getState() {
-        return { ...this.state };
-    }
-    /**
-     * Reset conversation state
-     */
-    reset() {
-        this.state = {
-            conversationHistory: [],
-            userGoal: null,
-            userGoalConfidence: 0,
-            pendingResearch: [],
-            completedResearch: [],
-            readyToExecute: false,
-            executionPlan: null,
-        };
-    }
-    // ============================================================
-    // Private Methods
-    // ============================================================
-    buildAnalysisPrompt(userInput) {
-        const recentHistory = this.state.conversationHistory.slice(-6)
-            .map(m => `${m.role.toUpperCase()}: ${m.content}`)
-            .join('\n');
-        const researchContext = this.state.completedResearch.length > 0
-            ? `\nCOMPLETED RESEARCH:\n${this.state.completedResearch.map(r => `- ${r.query}: ${r.result?.substring(0, 300)}`).join('\n')}`
-            : '';
-        const pendingContext = this.state.pendingResearch.length > 0
-            ? `\nPENDING RESEARCH: ${this.state.pendingResearch.map(r => r.query).join(', ')}`
-            : '';
-        return `You are the brain of a voice AI coding assistant. Analyze this input and decide what to do.
-
-WORKING DIRECTORY: ${this.workingDir}
-
-CAPABILITIES:
-- Full internet access (web search, fetch URLs, API calls)
-- Read/write files in the working directory
-- Run shell commands (npm, git, etc.)
-- Search and analyze codebases
-
-CONVERSATION HISTORY:
-${recentHistory}
-
-CURRENT USER INPUT: "${userInput}"
-
-CURRENT UNDERSTANDING:
-- User Goal: ${this.state.userGoal || 'Unknown - still gathering context'}
-- Confidence: ${Math.round(this.state.userGoalConfidence * 100)}%
-- Ready to Execute: ${this.state.readyToExecute}
-${researchContext}
-${pendingContext}
-
-DECIDE WHAT TO DO:
-
-1. DIRECT_COMMAND - If user gives a simple, clear command:
-   - "read file X" → direct_command
-   - "run npm test" → direct_command
-   - "show me the package.json" → direct_command
-   - "search the web for X" → direct_command
-   - "look up X online" → direct_command
-
-2. CLARIFY - If request is ambiguous, ask a specific question to understand better
-
-3. RESEARCH - If we need more info, start background research (2-3 queries max)
-   - Search codebase, read docs, explore files
-   - Search the web for information
-   - Keep conversation going while research runs
-
-4. EXECUTE - ONLY if we have HIGH confidence (>80%) about what user wants
-   - Must have clear plan
-   - Only ONE execution task at a time
-
-5. SPEAK - Just respond conversationally (greetings, status, etc.)
-
-IMPORTANT: Keep speech SHORT and conversational. No markdown formatting.
-
-Respond in JSON:
-{
-  "action": "speak" | "research" | "execute" | "clarify" | "direct_command",
-  "speech": "What to say to user RIGHT NOW (keep it short, conversational, NO markdown)",
-  "researchQueries": ["query1", "query2"],  // Only if action=research
-  "executeTask": "detailed task description",  // Only if action=execute
-  "directCommand": "simple command to run",  // Only if action=direct_command
-  "updatedGoal": "What we think user wants now",
-  "goalConfidence": 0.0-1.0,
-  "reasoning": "Brief explanation"
-}`;
-    }
-    parseDecision(response) {
-        try {
-            // Extract JSON from response
-            const jsonMatch = response.match(/\{[\s\S]*\}/);
-            if (!jsonMatch) {
-                return {
-                    action: 'speak',
-                    speech: response.substring(0, 200),
-                };
-            }
-            const parsed = JSON.parse(jsonMatch[0]);
-            // Update goal tracking
-            if (parsed.updatedGoal) {
-                this.state.userGoal = parsed.updatedGoal;
-            }
-            if (typeof parsed.goalConfidence === 'number') {
-                this.state.userGoalConfidence = parsed.goalConfidence;
-            }
-            return {
-                action: parsed.action || 'speak',
-                speech: parsed.speech,
-                researchQueries: parsed.researchQueries,
-                executeTask: parsed.executeTask,
-                directCommand: parsed.directCommand,
-                reasoning: parsed.reasoning,
-            };
-        }
-        catch (err) {
-            console.error('Failed to parse brain decision:', err);
-            return {
-                action: 'speak',
-                speech: "Let me think about that...",
-            };
-        }
-    }
-    updateState(decision) {
-        // Track research tasks
-        if (decision.action === 'research' && decision.researchQueries) {
-            for (const query of decision.researchQueries) {
-                this.state.pendingResearch.push({
-                    id: `research-${Date.now()}-${Math.random().toString(36).substring(2, 6)}`,
-                    query,
-                    status: 'pending',
-                });
-            }
-        }
-        // Mark ready to execute
-        if (decision.action === 'execute') {
-            this.state.readyToExecute = true;
-            this.state.executionPlan = decision.executeTask || null;
-        }
-    }
-    /**
-     * Create research tasks for background agents
-     */
-    getPendingResearchTasks() {
-        return this.state.pendingResearch.filter(t => t.status === 'pending');
-    }
-    /**
-     * Mark a research task as running
-     */
-    markResearchRunning(taskId) {
-        const task = this.state.pendingResearch.find(t => t.id === taskId);
-        if (task) {
-            task.status = 'running';
-            task.startedAt = new Date();
-        }
-    }
-}
-// ============================================================
-// Factory function
-// ============================================================
-export function createConversationBrain(config) {
-    return new ConversationBrain(config);
-}

package/dist/fast-llm.d.ts

@@ -1,15 +0,0 @@
-interface AcknowledgmentResult {
-    acknowledgment: string;
-    isCodingTask: boolean;
-    intent: 'code' | 'chat' | 'question' | 'permission';
-}
-/**
- * Fast LLM for immediate acknowledgment and intent classification
- * Target: < 200ms response time
- */
-export declare function getAcknowledgment(userInput: string): Promise<AcknowledgmentResult>;
-/**
- * Summarize Claude's response for voice output
- */
-export declare function summarizeForVoice(claudeResponse: string): Promise<string>;
-export {};

package/dist/fast-llm.js

@@ -1,81 +0,0 @@
-import Groq from 'groq-sdk';
-const groq = new Groq({
-    apiKey: process.env.GROQ_API_KEY,
-});
-/**
- * Fast LLM for immediate acknowledgment and intent classification
- * Target: < 200ms response time
- */
-export async function getAcknowledgment(userInput) {
-    const response = await groq.chat.completions.create({
-        model: 'llama-3.3-70b-versatile', // Fast model
-        messages: [
-            {
-                role: 'system',
-                content: `You are Osborn, a voice-enabled coding assistant. Classify user requests:
-
-CODING TASKS (isCodingTask: true) - Anything involving:
-- Files: read, write, create, edit, list, find, search
-- Directories: what directory, current folder, list files, project structure
-- Code: fix bugs, refactor, explain code, run tests
-- Terminal: run commands, install packages, git operations
-- Project: what's in this project, show me files, analyze codebase
-
-NOT CODING (isCodingTask: false):
-- General chat: hi, how are you, tell me a joke
-- Off-topic: weather, news, recipes
-- Permission responses: yes, no, approve, deny
-
-Respond in JSON:
-{
-  "acknowledgment": "brief friendly response",
-  "isCodingTask": true/false,
-  "intent": "code|chat|question|permission"
-}
-
-Examples:
-- "What directory are we in?" → {"acknowledgment": "Let me check", "isCodingTask": true, "intent": "code"}
-- "What files are in this project?" → {"acknowledgment": "I'll list them", "isCodingTask": true, "intent": "code"}
-- "Read package.json" → {"acknowledgment": "On it", "isCodingTask": true, "intent": "code"}
-- "Fix the bug" → {"acknowledgment": "Let me look", "isCodingTask": true, "intent": "code"}
-- "Hello" → {"acknowledgment": "Hi there!", "isCodingTask": false, "intent": "chat"}
-- "Yes, do it" → {"acknowledgment": "Got it", "isCodingTask": false, "intent": "permission"}`,
-            },
-            {
-                role: 'user',
-                content: userInput,
-            },
-        ],
-        temperature: 0.3,
-        max_tokens: 150,
-        response_format: { type: 'json_object' },
-    });
-    const content = response.choices[0]?.message?.content || '{}';
-    return JSON.parse(content);
-}
-/**
- * Summarize Claude's response for voice output
- */
-export async function summarizeForVoice(claudeResponse) {
-    if (claudeResponse.length < 200) {
-        return claudeResponse;
-    }
-    const response = await groq.chat.completions.create({
-        model: 'llama-3.3-70b-versatile',
-        messages: [
-            {
-                role: 'system',
-                content: `Summarize this coding assistant response for voice output.
-Keep it under 2 sentences. Focus on what was done and the result.
-Don't include code - just describe what happened.`,
-            },
-            {
-                role: 'user',
-                content: claudeResponse,
-            },
-        ],
-        temperature: 0.3,
-        max_tokens: 100,
-    });
-    return response.choices[0]?.message?.content || claudeResponse;
-}