browse

if command -v npx &>/dev/null; then
  _UK_CFG="${SKILL_DIR:-$(cd "$(dirname "$0")" && pwd)}/../update-kit.config.json"
  [ -f "$_UK_CFG" ] || _UK_CFG="${SKILL_DIR:-$(cd "$(dirname "$0")" && pwd)}/../../update-kit.config.json"
  _UK_JSON=$(npx update-kit check --json --config "$_UK_CFG" 2>/dev/null || echo '{}')
  _UK_KIND=$(echo "$_UK_JSON" | python3 -c "import sys,json; print(json.load(sys.stdin).get('kind',''))" 2>/dev/null || echo "")
  if [ "$_UK_KIND" = "upgrade_available" ]; then
    _UK_VER=$(echo "$_UK_JSON" | python3 -c "import sys,json; print(json.load(sys.stdin).get('latest',''))" 2>/dev/null)
    echo "Browser Control Skill update available: $_UK_VER — run: cd ~/.claude/skills/browser-control && npx update-kit apply"
  fi
fi

/browse — Browser Control Router

Usage

/browse <instruction>          → auto-detect foreground vs background
/browse here <instruction>     → foreground mode (AppleScript, user's current page)
/browse bg <instruction>       → background mode (CDP Proxy, no user disruption)

Mode Resolution

Parse the first argument. If it is here or bg, use that mode explicitly. Otherwise, auto-detect.

Explicit Modes

/browse here — Foreground Mode

Use macOS AppleScript to operate on the user's current visible page
Target: front window → active tab
User can see every action in real time
Requires: macOS + Chrome + "Allow JavaScript from Apple Events"
Best for: "Help me on this page..." / "current page" / user is watching and supervising

/browse bg — Background Mode

Use CDP Proxy to operate in background tabs
Agent creates its own tabs, user's tabs are untouched
User continues browsing normally, zero disruption
Supports sub-agent parallel dispatch (multiple tabs simultaneously)
Requires: Node.js 22+ + Chrome remote debugging enabled
Best for: "Help me look up..." / "research in parallel..." / bulk research / parallel tasks

Auto-Detection (`/browse` without mode)

When no explicit mode is given, infer from context:

IS the platform macOS?
├─ NO → bg (CDP is the only option)
└─ YES
    ├─ Does the instruction reference the user's current page?
    │   ("这个页面", "当前页面", "this page", "current tab",
    │    "我正在看的", "the page I'm on", "帮我在这里...")
    │   └─ YES → here (user wants agent on THEIR visible page)
    │
    ├─ Does the instruction involve multiple targets or parallel work?
    │   ("同时", "并行", "这几个", "分别查", "simultaneously",
    │    "in parallel", "these N sites/pages")
    │   └─ YES → bg (AppleScript can't parallelize)
    │
    ├─ Does the instruction imply background / don't-disturb?
    │   ("后台", "别打扰我", "background", "while I work")
    │   └─ YES → bg
    │
    ├─ Does the instruction require advanced write operations?
    │   (keyboard input, file upload, isTrusted events, hover menus)
    │   └─ YES → bg (needs CDP capabilities)
    │
    └─ DEFAULT → here (on macOS, prefer the simpler zero-dep path)

On macOS, the default is here because:

Zero setup cost (AppleScript is built-in)
Most natural UX for single-page operations
User gets immediate visual feedback
If AppleScript hits a limitation (value doesn't stick, isTrusted check, file upload), auto-escalate to CDP within the session — no need to restart with /browse bg

Escalation Within Foreground Mode

When operating in here mode, if a basic AppleScript operation fails:

el.value = "..." doesn't stick (React/Vue) → try React-compatible event dispatch (still AppleScript)
Still fails → auto-escalate: start CDP Proxy, switch to CDP helper for that operation
Report to user: "AppleScript basic mode cannot complete this operation, switched to CDP enhanced mode."

The escalation is per-operation, not per-session. Simple reads can stay on AppleScript even after one operation escalated to CDP.

Execution

After mode resolution, load the full browser-control skill and execute:

Run platform detection and preflight for the resolved mode
Follow the browser-control SKILL.md instructions using the resolved mode's commands
Apply all safety rules (three-layer security) regardless of mode
In bg mode: if task has multiple independent targets, consider sub-agent parallel dispatch

Foreground Mode Quick Reference

# Read current page
osascript -e 'tell application "Google Chrome" to execute active tab of front window javascript "document.body.innerText.substring(0, 15000)"'

# Click element
osascript -e 'tell application "Google Chrome" to execute active tab of front window javascript "document.querySelector(\"SELECTOR\").click()"'

# Fill field
osascript -e 'tell application "Google Chrome" to execute active tab of front window javascript "var el=document.querySelector(\"SELECTOR\"); el.value=\"VALUE\"; el.dispatchEvent(new Event(\"input\",{bubbles:true})); \"done\";"'

# Navigate (new tab, never in user's current tab)
osascript -e 'tell application "Google Chrome" to tell front window to make new tab with properties {URL:"URL"}'

Background Mode Quick Reference

# Start proxy (if not running)
bash "${SKILL_DIR}/../browser-control/scripts/check-deps.sh"

# Create background tab
curl -s "http://localhost:3456/new?url=URL"

# Execute JS in background tab
curl -s -X POST "http://localhost:3456/eval?target=ID" -d 'JS_CODE'

# Click in background tab
curl -s -X POST "http://localhost:3456/click?target=ID" -d 'SELECTOR'

# Close background tab when done
curl -s "http://localhost:3456/close?target=ID"

CDP Helper (Advanced Write Operations, Both Modes)

python3 "${SKILL_DIR}/../browser-control/scripts/cdp-helper.py" type "text"
python3 "${SKILL_DIR}/../browser-control/scripts/cdp-helper.py" key Enter
python3 "${SKILL_DIR}/../browser-control/scripts/cdp-helper.py" upload "input[type=file]" /path/to/file
python3 "${SKILL_DIR}/../browser-control/scripts/cdp-helper.py" click <x> <y>

Full Reference

For complete documentation (DOM-to-Markdown, element indexer, virtual scroll, console/network interception, screenshots, sub-agent dispatch, site experience memory, safety system), see:

skills/browser-control/SKILL.md