Stealth Capture → AI (macOS)

English | 中文

Many online assessments, coding interviews, and proctored hiring tests track when you leave the exam tab — via the Page Visibility API and window focus/blur events. On pages like detect-tab-focus, OS-level screenshots do not fire any JS events. What actually gets flagged is: switching tabs, switching windows, or switching to another app (for example, opening the ChatGPT, Claude, or Gemini desktop app to paste a screenshot).

This tool is for studying those detection boundaries on local demo pages. It keeps the browser in the foreground during an assessment-style session, runs capture and AI inference in the background, and shows results in a non-activating overlay — so the exam page stays focused.

---

Detection boundary (what tab-focus-test can / cannot see)

Action	Page detects it?
Switch to another tab	✅ visibilitychange → hidden
Switch to another browser window	✅ window blur
Switch to ChatGPT / Claude / Gemini / Terminal / Raycast	✅ window blur
Silent macOS screenshot (screencapture -x, Hammerspoon snapshot)	❌
Background process calling Claude/OpenAI API	❌
On-screen overlay showing text (without activating a window)	❌
Built-in ⌘⇧4 region select (crosshair briefly steals focus)	⚠️ may trigger blur
Copy to clipboard (browser stays foreground)	❌

---

Recommended architecture

┌─────────────────────────────────────────────────────────┐
│  Chrome / Safari (always frontmost app, tab stays visible) │
│  ┌─────────────────────────────────────┐                │
│  │  assessment / interview question page  │                │
│  └─────────────────────────────────────┘                │
└─────────────────────────────────────────────────────────┘
         ▲                              │
         │ no focus switch                 │ global hotkey
         │                              ▼
┌────────────────┐    PNG     ┌──────────────────────┐
│ Hammerspoon    │ ────────► │ ask_ai.py            │
│ silent capture │           │ POST → Claude/GPT    │
└────────────────┘           └──────────┬───────────┘
                                        │ text
                                        ▼
                           ┌──────────────────────┐
                           │ Overlay              │
                           │ (non-activating)     │
                           │ shows AI answer      │
                           └──────────────────────┘

---

Approach comparison

Approach A: Hammerspoon + Python (recommended — included in this repo)

• Pros: Fully local and controllable; hotkey capture of window or region; overlay does not steal focus; swap any API provider.
• Cons: Requires Hammerspoon; first run needs Screen Recording + Accessibility permissions.
• Best for: Local proctoring research, interview-platform debugging, and iteration.

Approach B: Second monitor / phone for AI lookup

• Pros: Zero code; browser keeps focus on the primary display.
• Cons: Poor UX; clicking an app on the second monitor still triggers blur on the primary screen.
• Best for: One-off temporary use.

Approach C: Switch to ChatGPT / Claude / Gemini app to paste the image

• Verdict: Always triggers blur, directly conflicts with the page’s focus-detection goal — not usable during an online assessment or live interview.

---

Quick start (Approach A)

1. Install dependencies

brew install hammerspoon
pip3 install anthropic   # or openai

Or run the installer:

git clone https://github.com/autocue-interview-tools/stealth-capture.git
cd stealth-capture
./install.sh

2. Configure API key

export ANTHROPIC_API_KEY="sk-ant-..."
# optional: custom model
export STEALTH_AI_MODEL="claude-sonnet-4-20250514"

Recommended: add to ~/.zshrc or set via Hammerspoon os.setenv on startup.

Supported providers (via STEALTH_AI_PROVIDER): openrouter (default), gemini, openai, anthropic.

3. Install Hammerspoon config

If you did not use ./install.sh, copy manually:

mkdir -p ~/.hammerspoon/stealth-capture
cp stealth_capture.lua ~/.hammerspoon/stealth_capture.lua
cp ask_ai.py ~/.hammerspoon/stealth-capture/ask_ai.py

Add to ~/.hammerspoon/init.lua:

require("stealth_capture")

4. Permissions

System Settings → Privacy & Security:

• Screen Recording: enable Hammerspoon
• Accessibility: enable Hammerspoon

Restart Hammerspoon (menu bar icon → Reload Config).

5. Hotkeys

Hotkey	Action
⌥ S	Drag to select a screen region → send to AI → show answer in overlay
⌥ A	Full-screen capture (no drag) → send to AI → show answer in overlay
⌥ R	Reopen last AI answer
Esc (while overlay visible)	Close overlay without switching focus

6. Verify

1. Open detect-tab-focus in Chrome (simulates a typical online assessment focus tracker); confirm status is Tab Focused.
2. Press ⌥ S.
3. On the page: no overlay should appear, and the Event Log should not gain a BLUR entry.
4. An AI answer overlay should appear on screen.

---

Why the overlay does not trigger blur

Hammerspoon hs.canvas with:

• level = floating (above normal windows)
• collectionBehavior = canJoinAllSpaces + stationary
• No win:focus() call; canvas does not accept keyboard focus by default

The browser remains NSWorkspace.shared.frontmostApplication, so document.hasFocus() stays true.

---

Extensions

• Capture browser content only: use hs.window.filter for Chrome / Safari in Lua, or crop the toolbar after window capture.
• Streaming output: change ask_ai.py to SSE and refresh the overlay incrementally (use hs.timer to avoid blocking).

---

Safety & compliance

This tool is designed for local demo pages such as detect-tab-focus, to understand how online assessments, interviews, and proctoring platforms detect focus — and to build your own anti-cheating logic. Bypassing monitoring in real online assessments, live interviews, certification exams, or employer proctoring may violate rules or law. Use only in authorized scenarios.