feat: GStack Browser — double-click AI browser with anti-bot stealth#695
Merged
feat: GStack Browser — double-click AI browser with anti-bot stealth#695
Conversation
…modification New browse/src/cdp-inspector.ts with full CDP inspection engine: - inspectElement() via CSS.getMatchedStylesForNode + DOM.getBoxModel - modifyStyle() via CSS.setStyleTexts with headless page.evaluate fallback - Persistent CDP session lifecycle (create, reuse, detach on nav, re-create) - Specificity sorting, overridden property detection, UA rule filtering - Modification history with undo support - formatInspectorResult() for CLI output Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…yscreenshot CLI Server endpoints: POST /inspector/pick, GET /inspector, POST /inspector/apply, POST /inspector/reset, GET /inspector/history, GET /inspector/events (SSE). CLI commands: inspect (CDP cascade), style (live CSS mod), cleanup (page clutter removal), prettyscreenshot (clean screenshot pipeline). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…, quick edit Extension changes for the visual CSS inspector: - inspector.js: element picker with hover highlight, CSS selector generation, basic mode fallback (getComputedStyle + CSSOM), page alteration handlers - inspector.css: picker overlay styles (blue highlight + tooltip) - background.js: inspector message routing (picker <-> server <-> sidepanel) - sidepanel: Inspector tab with box model viz (gstack palette), matched rules with specificity badges, computed styles, click-to-edit quick edit, Send to Agent/Code button, empty/loading/error states Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
browser-manager.ts changes:
- context.on('page') listener: automatically tracks tabs opened by the user
(Cmd+T, right-click open in new tab, window.open). Previously only
programmatic newTab() was tracked, so user tabs were invisible.
- page.on('close') handler in wirePageEvents: removes closed tabs from the
pages map and switches activeTabId to the last remaining tab.
- syncActiveTabByUrl: match Chrome extension's active tab URL to the correct
Playwright page for accurate tab identity.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Prevents parallel sidebar agents from interfering with each other's tab context. Three-layer fix: - sidebar-agent.ts: passes BROWSE_TAB=<tabId> env var to each claude process, per-tab processing set allows concurrent agents across tabs - cli.ts: reads process.env.BROWSE_TAB and includes tabId in command request body - server.ts: handleCommand() temporarily switches activeTabId when tabId is present, restores after command completes (safe: Bun event loop is single-threaded) Also: per-tab agent state (TabAgentState map), per-tab message queuing, per-tab chat buffers, verbose streaming narration, stop button endpoint. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Extension changes:
- sidepanel.js: per-tab chat history (tabChatHistories map), switchChatTab()
swaps entire chat view, browserTabActivated handler for instant tab sync,
stop button wired to /sidebar-agent/stop, pollTabs renders tab bar
- sidepanel.html: updated banner text ("Browser co-pilot"), stop button markup,
input placeholder "Ask about this page..."
- sidepanel.css: tab bar styles, stop button styles, loading state fixes
- background.js: chrome.tabs.onActivated sends browserTabActivated to sidepanel
with tab URL for instant tab switch detection
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
sidebar-agent.test.ts (new tests):
- BROWSE_TAB env var passed to claude process
- CLI reads BROWSE_TAB and sends tabId in body
- handleCommand accepts tabId, saves/restores activeTabId
- Tab pinning only activates when tabId provided
- Per-tab agent state, queue, concurrency
- processingTabs set for parallel agents
sidebar-ux.test.ts (new tests):
- context.on('page') tracks user-created tabs
- page.on('close') removes tabs from pages map
- Tab isolation uses BROWSE_TAB not system prompt hack
- Per-tab chat context in sidepanel
- Tab bar rendering, stop button, banner text
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…inspector # Conflicts: # browse/src/server.ts # browse/src/sidebar-agent.ts
…tion Merged main's security improvements (XML escaping, prompt injection defense, allowed commands whitelist, --model opus, Write tool, stderr capture) with our branch's per-tab isolation (BROWSE_TAB env var, processingTabs set, no --resume). Updated test expectations for expanded system prompt. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Pre-existing bug found by Codex: ALLOWED_TYPES in background.js was missing all inspector message types (startInspector, stopInspector, elementPicked, pickerCancelled, applyStyle, toggleClass, injectCSS, resetAll, inspectResult). Messages were silently rejected, making the inspector broken on ALL pages. Also: separate executeScript and insertCSS into individual try blocks in injectInspector(), store inspectorMode for routing, and add content.js fallback when script injection fails (CSP, chrome:// pages). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When inspector.js can't be injected (CSP, chrome:// pages), content.js provides a basic picker using getComputedStyle + CSSOM: - startBasicPicker/stopBasicPicker message handlers - captureBasicData() with ~30 key CSS properties, box model, matched rules - Hover highlight with outline save/restore (never leaves artifacts) - Click uses e.target directly (no re-querying by selector) - Sends inspectResult with mode:'basic' for sidebar rendering - Escape key cancels picker and restores outlines Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two action buttons in the inspector toolbar: - Cleanup (🧹): POSTs cleanup --all to server, shows spinner, chat notification on success, resets inspector state (element may be removed) - Screenshot (📸): POSTs screenshot to server, shows spinner, chat notification with saved file path Shared infrastructure: - .inspector-action-btn CSS with loading spinner via ::after pseudo-element - chat-notification type in addChatEntry() for system messages - package.json version bump to 0.13.9.0 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
16 new tests in sidebar-ux.test.ts: - Inspector message allowlist includes all inspector types - content.js basic picker (startBasicPicker, captureBasicData, CSSOM, outline save/restore, inspectResult with mode basic, Escape cleanup) - background.js CSP fallback (separate try blocks, inspectorMode, fallback) - Cleanup button (POST /command, inspector reset after success) - Screenshot button (POST /command, notification rendering) - Chat notification type and CSS styles Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Quick actions toolbar (🧹 Cleanup, 📸 Screenshot) now appears above the chat input, always visible. Both inspector and chat buttons share runCleanup() and runScreenshot() helper functions. Clicking either set shows loading state on both simultaneously. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Tests that chat toolbar exists (chat-cleanup-btn, chat-screenshot-btn, quick-actions container), CSS styles (.quick-action-btn, .quick-action-btn.loading), shared runCleanup/runScreenshot helper functions, and cleanup inspector reset. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…emoval Massively expanded CLEANUP_SELECTORS with patterns from uBlock Origin and Readability.js research: - ads: 30+ selectors (Google, Amazon, Outbrain, Taboola, Criteo, etc.) - cookies: OneTrust, Cookiebot, TrustArc, Quantcast + generic patterns - overlays (NEW): paywalls, newsletter popups, interstitials, push prompts, app download banners, survey modals - social: follow prompts, share tools - Cleanup now defaults to --all when no args (sidebar button fix) - Uses !important on all display:none (overrides inline styles) - Unlocks body/html scroll (overflow:hidden from modal lockout) - Removes blur/filter effects (paywall content blur) - Removes max-height truncation (article teaser truncation) - Collapses empty ad placeholder whitespace (empty divs after ad removal) - Skips gstack-ctrl indicator in sticky removal Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- setActionButtonsEnabled() toggles .disabled class on all cleanup/screenshot buttons (both chat toolbar and inspector toolbar) - Called with false in updateConnection when server URL is null - Called with true when connection established - runCleanup/runScreenshot silently return when disconnected instead of showing 'Not connected' error notifications - CSS .disabled style: pointer-events:none, opacity:0.3, cursor:not-allowed Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
17 new tests: - cleanup defaults to --all on empty args - CLEANUP_SELECTORS overlays category (paywall, newsletter, interstitial) - Major ad networks in selectors (doubleclick, taboola, criteo, etc.) - Major consent frameworks (OneTrust, Cookiebot, TrustArc, Quantcast) - !important override for inline styles - Scroll unlock (body overflow:hidden) - Blur removal (paywall content blur) - Article truncation removal (max-height) - Empty placeholder collapse - gstack-ctrl indicator skip in sticky cleanup - setActionButtonsEnabled function - Buttons disabled when disconnected - No error spam from cleanup/screenshot when disconnected - CSS disabled styles for action buttons Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Instead of brittle CSS selectors, the cleanup button now sends a prompt to the sidebar agent (which IS an LLM). The agent: 1. Runs deterministic $B cleanup --all as a quick first pass 2. Takes a snapshot to see what's left 3. Analyzes the page semantically to identify remaining clutter 4. Removes elements intelligently, preserving site branding This means cleanup works correctly on any site without site-specific selectors. The LLM understands that "Your Daily Puzzles" is clutter, "ADVERTISEMENT" is junk, but the SF Chronicle masthead should stay. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Deterministic cleanup improvements (used as first pass before LLM analysis): - New 'clutter' category: audio players, podcast widgets, sidebar puzzles/games, recirculation widgets (taboola, outbrain, nativo), cross-promotion banners - Text-content detection: removes "ADVERTISEMENT", "Article continues below", "Sponsored", "Paid content" labels and their parent wrappers - Sticky fix: preserves the topmost full-width element near viewport top (site nav bar) instead of hiding all sticky/fixed elements. Sorts by vertical position, preserves the first one that spans >80% viewport width. Tests: clutter category, ad label removal, nav bar preservation logic. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…y nav 22 new tests covering: - Cleanup button uses /sidebar-command (agent) not /command (deterministic) - Cleanup prompt includes deterministic first pass + agent snapshot analysis - Cleanup prompt lists specific clutter categories for agent guidance - Cleanup prompt preserves site identity (masthead, headline, body, byline) - Cleanup prompt instructs scroll unlock and $B eval removal - Loading state management (async agent, setTimeout) - Deterministic clutter: audio/podcast, games/puzzles, recirculation - Ad label text patterns (ADVERTISEMENT, Sponsored, Article continues) - Ad label parent wrapper hiding for small containers - Sticky nav preservation (sort by position, first full-width near top) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…, rebrand - Add GSTACK_CHROMIUM_PATH env var for custom Chromium binary - Add BROWSE_EXTENSIONS_DIR env var for extension path override - Move auth token to /health endpoint (fixes read-only .app bundles) - Anti-bot stealth: disable navigator.webdriver, fake plugins, languages - Custom user agent: Chrome/<version> GStackBrowser (auto-detects version) - Rebrand Chromium plist to "GStack Browser" at launch time - Update security test to match new token-via-health approach
- scripts/app/gstack-browser: dual-mode launcher (dev + .app bundle) - scripts/build-app.sh: compiles binary, bundles Chromium + extension, creates DMG - Rebrands Chromium plist during build for "GStack Browser" in menu bar - 389MB .app, 189MB compressed DMG, launches in ~5s
…ision 5-phase roadmap from .app wrapper through Chromium fork, 9 capability visions, competitive landscape, architecture diagrams, design system.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Show attempt count in loading screen ("Connecting... attempt 3")
- After 5 failed attempts, show debug details (port, connected, token)
so stuck users can see exactly what's failing
- Add 4 tests: getPort includes token, tryConnect uses token,
dead state exists with MAX_RECONNECT_ATTEMPTS, reconnectAttempts visible
Root cause: extension service worker starts before Bun.serve() is listening. First checkHealth() fails, next attempt is 10 seconds later. User stares at "Connecting..." for 10 seconds. Fix: retry every 1s for up to 15 attempts on startup, then switch to 10s polling once connected (or after 15s gives up). Sidebar should connect within 1-2 seconds of server becoming available. 3 new tests verify the fast-retry → slow-poll transition.
Replace useless "Connecting..." with real-time debug info: - "Looking for browse server... (attempt N)" - Shows port, server responding status, token status - Shows chrome.runtime errors if extension messaging fails - Tells user to run /open-gstack-browser if server not found
…ground Root cause: sidepanel asked background "are you connected?" but background's health check hadn't succeeded yet (1-10s gap). Sidepanel waited forever. Fix: when background says not connected, sidepanel hits /health directly with fetch(). Gets the token from the response. Bypasses background entirely for initial connection. Shows step-by-step debug info: "Checking server directly... port: 34567 / Trying GET /health..."
Two issues making the sidebar look broken when it's actually working: 1. "Timed out after 300s" error displayed after agent_done — this is a cleanup timer, not a real error. Now suppressed when no active session. 2. "(session ended)" text appended on every idle poll — removed entirely. The thinking spinner is cleaned up silently instead.
Ensures the child claude process connects to the existing headed browse server (port 34567) instead of spawning a new headless one. Without this, sidebar chat commands run in an invisible browser.
…wser When set, the browse CLI refuses to start a new server and exits with a clear error: "Server not available, run /open-gstack-browser to restart." The sidebar agent sets this so users never get an invisible headless browser when the headed one is closed.
5 tests: CLI checks env var before starting server, shows actionable error, sidebar-agent sets the flag + BROWSE_PORT, guard runs before lock acquisition to prevent stale lock files.
background.js checkHealth() never refreshed authToken from /health responses, so when the browse server restarted with a new token, all sidebar-command requests got 401 Unauthorized forever. Also: error placeholder text was #3f3f46 on #0C0C0C (nearly invisible). Now shows in red to match the error border. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Every empty catch {} in sidepanel.js, sidebar-agent.ts now logs with
[gstack sidebar] or [sidebar-agent] prefix. Chat poll 401s, stop agent,
tab poll, clear chat, SSE parse, refs fetch, stream JSON parse, queue
read/parse, process kill — all now visible in console.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Server-side silent catch blocks (22 instances) now log with [browse] prefix: chat persistence, session save/load, agent kill, tab pin/restore, welcome page, buffer flush, worktree cleanup, lock files, SSE streams. Also adds pickSidebarModel() — routes sidebar messages to sonnet for navigation/interaction (click, goto, fill, screenshot) and opus for analysis/comprehension (summarize, describe, find bugs). Sonnet is ~4x faster for action commands with zero quality difference. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- stopAgent slice 800→1000 to accommodate added error logging lines - Replace hardcoded opus assertion with model router assertions Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
… entries Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Previously the welcome page arrow hid immediately when the extension's content script loaded — but extension loaded ≠ sidebar open. Now the signal flow is: sidepanel connects → tells background.js → relays to content script → dispatches gstack-extension-ready → arrow hides. Adds welcome-page.test.ts: 14 tests verifying arrow, branding, feature cards, dark theme, and auto-hide behavior via real HTTP server. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
8 new tests verify the sidebarOpened → background → content → welcome signal chain. Updates stale "(session ended)" test that checked for text removed in a prior commit. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When the user sends a message and the server assigns it to a new tab (because Chrome's active tab changed), switchChatTab() was blowing away the optimistic user bubble and thinking dots with a welcome screen. Now preserves the current DOM if we're mid-send with a thinking indicator. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
SIDEBAR_MESSAGE_FLOW.md documents the full init timeline, message flow (user types → claude responds), auth token chain, arrow hint signal chain, model routing, tab concurrency, and known failure modes. CLAUDE.md now tells you to read it before touching sidebar files. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two fixes for the "browser died while chatting" problem: 1. /sidebar-command now calls resetIdleTimer(). Previously only CLI commands reset it, so the server would shut down after 30 min even while the user was actively chatting in the sidebar. 2. shutdown() now pkills the sidebar-agent daemon. Previously the agent survived server shutdown, kept polling a dead server, and spawned confused claude processes that auto-started headless browsers. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The 30-minute idle timeout only applies to headless mode now. In headed mode the user is looking at the Chrome window, so auto-shutdown is wrong. The browser stays alive until explicit disconnect or window close. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
One-click cookie import from the sidebar. Navigates the headed browser to /cookie-picker where you can select which domains to import from your real Chrome profile. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
README.md: updated Real browser mode and sidebar agent sections with model routing, cookie import button, no idle timeout in headed mode. Updated skill table entries for /browse and /open-gstack-browser. docs/skills.md: updated /open-gstack-browser deep dive with model routing and cookie import details. GSTACK_BROWSER_V0.md: added 6 new SHIPPED items to implementation status table (model routing, debug logging, idle timeout, cookie button, arrow hint, architecture doc). TODOS.md: marked "Sidebar agent Write tool + error visibility" as SHIPPED. Added new P2 TODO for direct API calls to eliminate claude -p startup tax. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Fifth example shows the parent agent workflow: navigate, extract CSS, write to file. The other four are all sidebar-only. This one shows co-presence — the Claude Code session that launched the browser can also control it directly. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Claude reads its own ~/.claude/projects/.../tool-results/ files as internal plumbing. These showed up as long unreadable paths in the sidebar. Now: describeToolCall returns empty for tool-result reads, and the sidebar skips rendering tool_use entries with no description. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
While the agent is working, tool calls stream live so you can watch progress. When the agent finishes, all tool calls collapse into a "See reasoning (N steps)" disclosure. Click to expand and see what the agent did. The final text answer stays visible. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Covers: tool-result file filtering, empty tool_use skip, reasoning disclosure collapse, idle timeout headed mode bypass, sidebar-command idle reset, shutdown sidebar-agent kill, cookie button, and model routing analysis-before-action priority. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Cookies now sits next to Cleanup and Screenshot as a primary action button (🍪 Cookies) instead of buried in the footer. Same behavior, more discoverable. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
"Select the domains of cookies you want to import to GStack Browser. You'll be able to browse those sites with the same login as your other browser." Also fixes stale test that expected hardcoded '--model', 'opus'. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
3x2 grid layout (was 2x2). New cards: "Import your cookies" (click 🍪 Cookies to import login sessions from Chrome/Arc/Brave) and "Or use your main agent" (your Claude Code terminal also controls this browser). Responsive: 3 cols > 2 cols > 1 col. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The arrow was centered vertically which put it behind the feature cards. Now positioned at top: 80px where there's open space and it's more visible. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
GStack Browser is a macOS .app you can double-click to launch an AI-controlled browser. Chromium opens with the sidebar extension baked in, Claude Code ready to go. No terminal needed.
What's new in this PR (on top of the sidebar-css-inspector branch):
scripts/build-app.shcreates a 389MB .app bundle (189MB DMG). Bundles browse binary + Chromium + extension. Launches in ~5s.navigator.webdriver = false, fake plugins, proper languages, CDP cleanup. Google and NYTimes work without captchas.GStackBrowser, no more "Chrome for Testing"docs/designs/GSTACK_BROWSER_V0.mdcovers the full 5-phase roadmap from wrapper through Chromium forkFrom the parent branch (sidebar-css-inspector):
Test Coverage
All tests pass. Auth security test updated to match new token-via-health approach.
Pre-Landing Review
Auth token moved from file to /health endpoint. Safe: /health is localhost-only (127.0.0.1). No external exposure.
Plan Completion
Plan went through 4 reviews (CEO, Design, Eng, Codex). All CLEARED. Codex caught 3 critical gaps (server bundling, auth file location, project binding) that survived 3 prior reviews. All resolved.
Test plan
scripts/build-app.sh)🤖 Generated with Claude Code