Files
gstack/BROWSER.md
Garry Tan 47d167f6ac docs(browser): rewrite BROWSER.md as complete reference
Full rewrite covering the gstack browser surface as of v1.19.0.0. Up from
488 to 1,299 lines, 26 top-level sections.

Adds previously-undocumented subsystems:

- The productivity loop: /scrape + /skillify with D1 (provenance guard),
  D2 (final-attempt-only synthesis), D3 (atomic-write discipline) contracts.
- Browser-skills runtime: anatomy, three-tier storage, scoped tokens, trust
  model (capability + env axes), sibling SDK distribution, atomic-write
  helper, bundled hackernews-frontpage reference.
- Domain-skills: per-site agent notes with quarantined → active → global
  state machine and the L4-classifier auto-promotion gate.
- Pair-agent: dual-listener architecture, 26-command tunnel allowlist,
  canDispatchOverTunnel pure gate, three token types (root, setup key,
  scoped), denial log path + salt model.
- Security stack L1-L6: layer table, thresholds (BLOCK/WARN/LOG_ONLY/
  SOLO_CONTENT_BLOCK), ensemble rule, classifier model paths, env knobs.
- Side Panel deep dive: Terminal pane (Claude PTY) as the primary surface
  with Activity/Refs/Inspector as debug overlays, WS auth via
  Sec-WebSocket-Protocol, gstackInjectToTerminal cross-pane plumbing.
- CDP escape hatch: $B cdp deny-default allowlist, $B inspect CSS inspector,
  $B ux-audit page structure extraction.
- Meta commands previously undocumented: tabs/frames/state/watch/inbox/
  tab-each, with usage and storage paths.
- Authentication: three token types with lifetimes, SSE session cookie,
  PTY session cookie, token registry behavior.
- Full source map: 30+ file inventory of browse/src/ vs the old 11-file
  list.

Preserves from before: architecture diagram, daemon lifecycle, snapshot
ref staleness, screenshot modes, goto file:// vs load-html semantics,
batch endpoint, JS await wrapping, env vars, performance numbers vs MCP,
Playwright acknowledgments, dev guide.

Cross-links to ARCHITECTURE.md, CLAUDE.md, docs/REMOTE_BROWSER_ACCESS.md,
docs/designs/BROWSER_SKILLS_V1.md, scrape/SKILL.md, skillify/SKILL.md,
TODOS.md so anyone landing on BROWSER.md can navigate to the load-bearing
companion docs.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-28 08:56:09 -07:00

56 KiB
Raw Permalink Blame History

Browser — Complete Reference

gstack's browser surface in one document. Headless Chromium daemon, ~70+ commands, ref-based element selection, codifiable browser-skills, real-browser mode with a Chrome side panel, an in-sidebar Claude PTY, an ngrok pair-agent flow, and a layered prompt-injection defense — all behind a compiled CLI that prints plain text to stdout. ~100-200ms per call. Zero context-token overhead.

If you've used gstack in the last release or two, the productivity loop is the new headline: /scrape <intent> drives a page once, /skillify codifies the flow into a deterministic Playwright script, and the next /scrape on the same intent runs in ~200ms instead of ~30 seconds of agent re-exploration.


Quick start

# One-time: build the binary (browse/dist/browse, ~58MB)
bun install && bun run build

# Set $B once and forget about it
B=./browse/dist/browse           # or ~/.claude/skills/gstack/browse/dist/browse

# Drive a page
$B goto https://news.ycombinator.com
$B snapshot -i                   # @e refs you can click/fill/inspect later
$B click @e30                    # click ref 30 from the snapshot
$B text                          # get clean page text
$B screenshot /tmp/hn.png

# Codify a repeated flow
/scrape latest hacker news stories
/skillify                        # writes ~/.gstack/browser-skills/hn-front/...
/scrape hacker news front page   # second call: 200ms via the codified skill

# Watch Claude work in real time
$B connect                       # headed Chromium + Side Panel extension

Table of contents

  1. What it is
  2. The productivity loop — /scrape + /skillify
  3. Architecture
  4. Command reference
  5. Snapshot system + ref-based selection
  6. Browser-skills runtime
  7. Domain-skills (per-site agent notes)
  8. Real-browser mode ($B connect)
  9. Side Panel + sidebar agent
  10. Pair-agent — remote agents over an ngrok tunnel
  11. Authentication + tokens
  12. Prompt-injection security stack (L1L6)
  13. Screenshots, PDFs, visual inspection
  14. Local HTML — goto file:// vs load-html
  15. Batch endpoint
  16. Console, network, dialog capture
  17. JS execution — js + eval
  18. Tabs, frames, state, watch, inbox
  19. CDP escape hatch + CSS inspector
  20. Performance + scale
  21. Multi-workspace isolation
  22. Environment variables
  23. Source map
  24. Development + testing
  25. Cross-references
  26. Acknowledgments

What it is

A compiled CLI binary that talks to a persistent local Chromium daemon over HTTP. The CLI is a thin client — it reads a state file, sends a command, prints the response to stdout. The daemon does the real work via Playwright.

Everything that was a Chrome MCP server in the early days now happens through plain stdout. No JSON-schema framing, no protocol negotiation, no persistent WebSocket — Claude's Bash tool already exists, so we use it.

Three escalating modes:

  • Headless (default). Daemon runs Chromium with no visible window. Fastest, cheapest, what skills like /qa, /design-review, /benchmark use by default.
  • Headed via $B connect. Same daemon, but Chromium is visible (rebranded as "GStack Browser") with the Side Panel extension auto-loaded. You watch every command tick through in real time.
  • Pair-agent over a tunnel. Daemon binds a second listener that ngrok forwards. A remote agent (Codex, OpenClaw, Hermes, anything that can speak HTTP) drives your local browser through a 26-command allowlist with a scoped, single-use token.

The productivity loop

The shipped headline of v1.19.0.0. Two gstack skills wrap the browser-skills runtime so the second time you ask Claude to scrape a page, it runs in ~200ms.

/scrape <intent>

One entry point for pulling page data. Three paths under the hood:

  1. Match path (~200ms) — agent runs $B skill list, semantically matches the intent against each skill's triggers: array + description + host, and runs $B skill run <name> if a confident match exists.
  2. Prototype path (~30s) — no match, agent drives the page with $B goto, $B text, $B html, $B links, etc., returns the JSON, and appends a one-line "say /skillify" suggestion.
  3. Mutating-intent refusal — verbs like submit, click, fill route to /automate (Phase 2b, P0 in TODOS.md). /scrape is read-only by contract.

/skillify

Codifies the most recent successful /scrape prototype into a permanent browser-skill on disk. Eleven steps, three locked contracts:

  • D1 — Provenance guard. Walks back ≤10 agent turns for a clearly-bounded /scrape result. Refuses with one specific message if cold. No silent synthesis from chat fragments.
  • D2 — Synthesis input slice. Extracts ONLY the final-attempt $B calls that produced the JSON the user accepted, plus the user's intent string. Drops failed selectors, drops chat, drops earlier-session content.
  • D3 — Atomic write. Stages everything to ~/.gstack/.tmp/skillify-<spawnId>/, runs $B skill test against the temp dir, and only renames into the final tier path on test pass + user approval. Test fail or rejection: rm -rf the temp dir entirely. No half-written skill ever appears in $B skill list.

Mutating-flow sibling /automate is split out as P0 in TODOS.md and ships on the next branch — same skillify machinery, per-mutating-step confirmation gate when running non-codified.

See docs/designs/BROWSER_SKILLS_V1.md for the full design + decision trail.


Architecture

┌─────────────────────────────────────────────────────────────────┐
│  Claude Code                                                    │
│                                                                 │
│  $B goto https://staging.myapp.com                              │
│       │                                                         │
│       ▼                                                         │
│  ┌──────────┐    HTTP POST     ┌──────────────┐                 │
│  │ browse   │ ──────────────── │ Bun HTTP     │                 │
│  │ CLI      │  127.0.0.1:rand  │ daemon       │                 │
│  │          │  Bearer token    │              │                 │
│  │ compiled │ ◄──────────────  │  Playwright  │──── Chromium    │
│  │ binary   │  plain text      │  API calls   │    (headless    │
│  └──────────┘                  └──────────────┘     or headed)  │
│   ~1ms startup                  persistent daemon               │
│                                 auto-starts on first call       │
│                                 auto-stops after 30 min idle    │
└─────────────────────────────────────────────────────────────────┘

Daemon lifecycle

  1. First call. CLI checks <project>/.gstack/browse.json for a running server. None found — it spawns bun run browse/src/server.ts in the background. Daemon launches headless Chromium via Playwright, picks a random port (1000060000), generates a bearer token, writes the state file (chmod 600), starts accepting requests. ~3 seconds.
  2. Subsequent calls. CLI reads the state file, sends an HTTP POST with the bearer token, prints the response. ~100-200ms round trip.
  3. Idle shutdown. After 30 minutes of no commands, daemon shuts down and cleans up the state file. Next call restarts it.
  4. Crash recovery. If Chromium crashes, the daemon exits immediately — no self-healing, don't hide failure. CLI detects the dead daemon on the next call and starts a fresh one.

Multi-workspace isolation

Each project root (detected via git rev-parse --show-toplevel) gets its own daemon, port, state file, cookies, and logs. No cross-workspace collisions. State at <project>/.gstack/browse.json.

Workspace State file Port
/code/project-a /code/project-a/.gstack/browse.json random (1000060000)
/code/project-b /code/project-b/.gstack/browse.json random (1000060000)

Command reference

~70 commands across read, write, and meta. Selectors accept CSS, @e refs from snapshot, or @c refs from snapshot -C. Full table:

Reading

Command Description
text [sel] Clean page text (or scoped to a selector)
html [sel] innerHTML, or full page HTML if no selector
links All links as text → href
forms Form fields as JSON
accessibility Full ARIA tree
media [--images|--videos|--audio] [sel] Media elements with URLs, dimensions, types
data [--jsonld|--og|--meta|--twitter] Structured data: JSON-LD, OG, Twitter Cards, meta tags

Inspection

Command Description
js <expr> Run inline JavaScript expression in page context, return as string
eval <file> Run JS from a file (path under /tmp or cwd; same sandbox as js)
css <sel> <prop> Computed CSS value
attrs <sel|@ref> Element attributes as JSON
is <prop> <sel|@ref> State check: visible, hidden, enabled, disabled, checked, editable, focused
console [--clear|--errors] Captured console messages
network [--clear] Captured network requests
dialog [--clear] Captured dialog messages
cookies All cookies as JSON
storage / storage set <key> <val> Read both localStorage + sessionStorage; set localStorage
perf Page load timings
inspect [sel] [--all] [--history] Deep CSS via CDP — full rule cascade, box model, computed styles
ux-audit Page structure for behavioral analysis: site ID, nav, headings, text blocks, interactive elements
cdp <Domain.method> [json-params] Raw CDP method dispatch (deny-default; allowlist in cdp-allowlist.ts)

Navigation

Command Description
goto <url> Navigate to URL (http://, https://, file://)
load-html <file> Load local HTML in memory (no file:// URL; survives viewport scale changes)
back, forward, reload Standard nav
url Current page URL
wait <sel|--networkidle|--load> Wait for element, network idle, or page load (15s timeout)

Interaction

Command Description
click <sel|@ref> Click element
fill <sel> <val> Fill input
select <sel> <val> Select dropdown option (value, label, or visible text)
hover <sel> Hover element
type <text> Type into focused element
press <key> Playwright keyboard key (case-sensitive: Enter, Tab, ArrowUp, Shift+Enter, Control+A, ...)
scroll [sel|@ref] Scroll element into view, or jump to page bottom if no selector
viewport [<WxH>] [--scale <n>] Set viewport size + optional deviceScaleFactor 1-3 (retina screenshots)
upload <sel> <file> [...] Upload file(s)
dialog-accept [text] Auto-accept next alert/confirm/prompt; text is sent for prompts
dialog-dismiss Auto-dismiss next dialog

Style + cleanup

Command Description
style <sel> <prop> <val> Modify CSS property (with undo support)
style --undo [N] Undo last N style changes
cleanup [--ads|--cookies|--sticky|--social|--all] Remove page clutter
prettyscreenshot [--scroll-to <sel|text>] [--cleanup] [--hide <sel>...] [path] Clean screenshot with optional cleanup, scroll, hide

Visual

Command Description
screenshot [--selector <css>] [--viewport] [--clip x,y,w,h] [--base64] [sel|@ref] [path] Five modes: full page, viewport, element crop, region clip, base64
pdf [path] [--format letter|a4|legal] [...] PDF with full layout: format, width/height, margins, header/footer templates, page numbers, --tagged for accessibility, --toc waits for Paged.js
responsive [prefix] Three screenshots: mobile (375x812), tablet (768x1024), desktop (1280x720)
diff <url1> <url2> Text diff between two URLs

Cookies + headers

Command Description
cookie <name>=<value> Set cookie on current page domain
cookie-import <json> Import cookies from JSON file
cookie-import-browser [browser] [--domain d] Import from installed Chromium browsers (interactive picker, or --domain for direct import)
header <name>:<value> Set custom request header (sensitive values auto-redacted)
useragent <string> Set user agent (triggers context recreation, invalidates refs)

Tabs + frames

Command Description
tabs List open tabs
tab <id> Switch to tab
newtab [url] [--json] Open new tab; --json returns {tabId, url} for programmatic use
closetab [id] Close tab
tab-each <command> [args...] Fan out a command across every open tab; returns JSON
frame <sel|@ref|--name n|--url pattern|main> Switch to iframe context (or back to main); clears refs

Extraction

Command Description
download <url|@ref> [path] [--base64] Download URL or media element using browser cookies
scrape <images|videos|media> [--selector] [--dir] [--limit] Bulk download all media from page; writes manifest.json
archive [path] Save complete page as MHTML via CDP

Snapshot

Command Description
snapshot [-i] [-c] [-d N] [-s sel] [-D] [-a] [-o path] [-C] Accessibility tree with @e refs; -i interactive only, -c compact, -d N depth, -s scope, -D diff vs previous, -a annotated screenshot, -C cursor-interactive @c refs

Server lifecycle

Command Description
status Daemon health + mode (headless / headed / cdp)
stop Shut down daemon
restart Restart daemon
connect Launch headed GStack Browser with Side Panel extension
disconnect Close headed Chrome, return to headless
focus [@ref] Bring headed Chrome to foreground (macOS); @ref also scrolls into view
state save|load <name> Save or load browser state (cookies + URLs)

Handoff

Command Description
handoff [reason] Open visible Chrome at current page for user takeover (CAPTCHA, MFA, complex auth)
resume Re-snapshot after user takeover, return control to AI

Meta + chains

Command Description
chain (JSON via stdin) Run a sequence of commands. Pipe [["cmd","arg1",...],...] to $B chain. Stops at first error.
inbox [--clear] List messages from sidebar scout inbox
watch [stop] Passive observation — periodic snapshots while user browses; stop returns summary

Browser-skills runtime

Command Description
skill list List all browser-skills with resolved tier (project > global > bundled)
skill show <name> Print SKILL.md
skill run <name> [--arg k=v...] [--timeout=Ns] Spawn the skill script with a per-spawn scoped token
skill test <name> Run the skill's script.test.ts against bundled fixtures
skill rm <name> [--global] Tombstone a user-tier skill

Domain-skills

Command Description
domain-skill save|list|show|edit|promote-to-global|rollback|rm <host?> Per-site agent notes (host derived from active tab). Lifecycle: quarantined → active (after N=3 successful uses without classifier flag) → global (explicit promote)

Aliases: setcontent, set-content, setContentload-html (canonicalized before scope checks, so a read-scoped token can't use the alias to run a write command).


Snapshot system

The browser's key innovation is ref-based element selection built on Playwright's accessibility tree API. No DOM mutation. No injected scripts. Just Playwright's native AX API.

How @ref works

  1. page.locator(scope).ariaSnapshot() returns a YAML-like accessibility tree.
  2. The snapshot parser assigns refs (@e1, @e2, ...) to each element.
  3. For each ref, it builds a Playwright Locator (using getByRole + nth-child).
  4. The ref→Locator map is stored on BrowserManager.
  5. Later commands like click @e3 look up the Locator and call locator.click().

Ref staleness detection

SPAs can mutate the DOM without navigation (React router, tab switches, modals). When this happens, refs collected from a previous snapshot may point to elements that no longer exist. resolveRef() runs an async count() check before using any ref — if the element count is 0, it throws immediately with a message telling the agent to re-run snapshot. Fails fast (~5ms) instead of waiting for Playwright's 30-second action timeout.

Extended snapshot features

  • --diff (-D). Stores each snapshot as a baseline. On the next -D call, returns a unified diff showing what changed. Use this to verify that an action (click, fill, etc.) actually worked.
  • --annotate (-a). Injects temporary overlay divs at each ref's bounding box, takes a screenshot with ref labels visible, then removes the overlays. Use -o <path> to control the output.
  • --cursor-interactive (-C). Scans for non-ARIA interactive elements (divs with cursor:pointer, onclick, tabindex>=0) using page.evaluate. Assigns @c1, @c2... refs with deterministic nth-child CSS selectors. These are elements the ARIA tree misses but users can still click.

Browser-skills runtime

Per-task directories that codify a repeated browser flow into a deterministic Playwright script. The compounding layer.

Anatomy of a browser-skill

browser-skills/<name>/
├── SKILL.md                        # frontmatter + prose contract
├── script.ts                       # deterministic Playwright-via-browse-client logic
├── _lib/browse-client.ts           # vendored copy of the SDK (~3KB, byte-identical to canonical)
├── fixtures/<host>-<date>.html     # captured page for fixture-replay tests
└── script.test.ts                  # parser tests against the fixture (no daemon required)

The bundled reference is browser-skills/hackernews-frontpage/: scrapes the HN front page, returns 30 stories as JSON. Try it:

$B skill list                            # shows hackernews-frontpage (bundled)
$B skill show hackernews-frontpage
$B skill run hackernews-frontpage        # JSON of 30 stories in ~200ms
$B skill test hackernews-frontpage       # runs script.test.ts against fixture

Three-tier storage

$B skill list walks all three in priority order; first hit wins. Resolved tier is printed inline next to each skill name:

Tier Path When
Project <project>/.gstack/browser-skills/<name>/ Project-specific skills (committed or gitignored)
Global ~/.gstack/browser-skills/<name>/ Per-user skills, all projects
Bundled <gstack-install>/browser-skills/<name>/ Ships with gstack, read-only

Trust model

Two orthogonal axes — daemon-side capability and process-side env — independently configured.

Axis Mechanism Default
Daemon-side capability Per-spawn scoped token bound to read+write scope (browser-driving commands minus admin: eval, js, cookies, storage). Single-use clientId encodes skill name + spawn id. Revoked when spawn exits. Always scoped — never the daemon root token
Process-side env trusted: true frontmatter passes process.env minus GSTACK_TOKEN. trusted: false (default) drops everything except a minimal allowlist (LANG, LC_ALL, TERM, TZ) and pattern-strips secrets (TOKEN/KEY/SECRET/PASSWORD, AWS_, ANTHROPIC_, OPENAI_, GITHUB_, etc.) Untrusted (must opt in)

GSTACK_PORT and GSTACK_SKILL_TOKEN are injected last, so a parent process can't override them.

Output protocol

stdout = JSON. stderr = streaming logs. Exit 0 / non-zero. Default 60s timeout, override via --timeout=Ns. Max stdout 1MB (truncate + non-zero exit if exceeded). Matches gh / kubectl / docker conventions.

How the SDK distribution works

Each skill ships its own copy of browse-client.ts at _lib/browse-client.ts, byte-identical to the canonical browse/src/browse-client.ts. /skillify copies the canonical SDK alongside every generated script. Each skill is fully self-contained: copy the directory anywhere, it runs. Version drift impossible — the SDK is frozen at the version the skill was authored against.

Atomic write discipline (/skillify D3)

browse/src/browser-skill-write.ts provides three primitives:

  • stageSkill(opts) — writes files to ~/.gstack/.tmp/skillify-<spawnId>/<name>/ with restrictive perms.
  • commitSkill(opts) — atomic fs.renameSync into the final tier path. Refuses to follow symlinked staging dirs (lstat check), refuses to clobber existing skills, runs realpath discipline on the tier root.
  • discardStaged(stagedDir)rm -rf the staged dir + per-spawn wrapper. Idempotent. Called on test failure or approval rejection.

There is no "almost shipped" state. Tests pass + user approves = atomic rename. Tests fail or user rejects = staging vanishes.

See docs/designs/BROWSER_SKILLS_V1.md for the full design rationale.


Domain-skills

Different mental model from browser-skills: agent-authored notes about a site (not deterministic scripts). One per hostname. Lifecycle:

  1. domain-skill save <host> — agent writes a note about the site (e.g., "GitHub: PR creation needs --draft flag for non-staff", "X.com: timeline uses cursor pagination, not page numbers"). Default state: quarantined.
  2. After N=3 successful uses without the L4 prompt-injection classifier flagging the note, it auto-promotes to active.
  3. domain-skill promote-to-global <host> lifts it to the global tier (machine-wide, all projects).
  4. domain-skill rollback <host> demotes; domain-skill rm <host> tombstones.

The classifier flag is set automatically by the L4 prompt-injection scan; agents do not set it manually.

Storage:

  • Per-project: <project>/.gstack/domain-skills/<host>.md
  • Global: ~/.gstack/domain-skills/<host>.md

Source: browse/src/domain-skills.ts, domain-skill-commands.ts.


Real-browser mode

$B connect launches GStack Browser — a rebranded Chromium controlled by Playwright with the Side Panel extension auto-loaded and anti-bot stealth patches applied. You watch every command tick through a visible window in real time.

$B connect              # launches GStack Browser, headed
$B goto https://app.com # navigates in the visible window
$B snapshot -i          # refs from the real page
$B click @e3            # clicks in the real window
$B focus                # bring window to foreground (macOS)
$B status               # shows Mode: cdp
$B disconnect           # back to headless mode

The window has a subtle golden shimmer line at the top and a floating "gstack" pill in the bottom-right corner so you always know which Chrome window is being controlled.

What "GStack Browser" means

Not your daily Chrome — a Playwright-managed Chromium with custom branding in the Dock and menu bar, anti-bot stealth (sites like Google and NYTimes work without captchas), a custom user agent, and the gstack extension pre-loaded via launchPersistentContext. Your regular Chrome with your tabs and bookmarks stays untouched.

When to use headed mode

  • QA testing where you want to watch Claude click through your app
  • Design review where you need to see exactly what Claude sees
  • Debugging where headless behavior differs from real Chrome
  • Demos where you're sharing your screen
  • Pair-agent sessions (the remote agent drives your local browser)

CDP-aware skills

When in real-browser mode, /qa and /design-review automatically skip cookie import prompts and headless workarounds — the headed browser already has whatever session you logged into.


Side Panel + sidebar agent

The Chrome extension that ships baked into GStack Browser shows a live activity feed of every browse command in a Side Panel, plus @ref overlays on the page, plus an interactive Claude PTY inside the sidebar.

The Terminal pane (the headline)

The Side Panel's primary surface is the Terminal pane — a live claude -p PTY you can type into directly from the sidebar. Activity / Refs / Inspector are debug overlays behind the footer's debug toggle. WebSocket auth uses Sec-WebSocket-Protocol (browsers can't set Authorization on a WebSocket upgrade), and the PTY session token is a 30-minute HttpOnly cookie minted via POST /pty-session.

The toolbar's Cleanup button and the Inspector's "Send to Code" action both pipe text into the live Claude PTY via window.gstackInjectToTerminal(text), exposed by sidepanel-terminal.js. There's no separate /sidebar-command POST — the live REPL is the only execution surface.

Activity feed

A scrolling feed of every browse command — name, args, duration, status, errors. Shows up in real time as Claude works. Backed by SSE (/activity/stream) that accepts the Bearer token OR the HttpOnly gstack_sse session cookie (30-minute stream-scope cookie minted via POST /sse-session).

Refs tab

After $B snapshot, shows the current @ref list (role + name) so you can see what Claude is targeting.

CSS Inspector

Powered by $B inspect (CDP-based). Click any element on the page to see the full CSS rule cascade, computed styles, box model, and modification history. The "Send to Code" button injects a description into the Claude PTY.

Sidebar architecture

Component Where it lives Notes
Side Panel UI extension/sidepanel.js, sidepanel-terminal.js Chrome extension surface
Background SW extension/background.js Manages tab events, port management
Content script extension/content.js Page overlays, gstack pill
Terminal agent browse/src/terminal-agent.ts PTY spawn, lifecycle, auth
Sidebar utilities browse/src/sidebar-utils.ts URL sanitization, helpers

Before modifying any of these, read the comment block in CLAUDE.md under "Sidebar architecture" — silent failures here usually trace to not understanding the cross-component flow.

Manual install (for your regular Chrome)

If you want the extension in your everyday Chrome (not the Playwright-controlled one):

bin/gstack-extension    # opens chrome://extensions, copies path to clipboard

Or do it manually: chrome://extensions → toggle Developer mode → Load unpacked → navigate to ~/.claude/skills/gstack/extension → pin the extension → enter the port from $B status.


Pair-agent

Remote AI agents (Codex, OpenClaw, Hermes, anything that speaks HTTP) can drive your local browser through an ngrok tunnel. The whole flow is gated by a 26-command allowlist, scoped tokens, and a denial log.

How it works

/pair-agent                     # generates a setup key, prints connection instructions
# Copy the instructions to the remote agent
# Remote agent runs:
#   POST <tunnel-url>/connect with setup key → gets a scoped token (24h, single client)
#   POST <tunnel-url>/command with token → runs allowed commands

Dual-listener architecture (v1.6.0.0+)

When pair-agent activates, the daemon binds two HTTP listeners:

  • Local listener (127.0.0.1:LOCAL_PORT). Full command surface. Never forwarded by ngrok. Used by your Claude Code, the Side Panel, anything on your machine.
  • Tunnel listener (127.0.0.1:TUNNEL_PORT). Locked allowlist — /connect, /command (scoped tokens + 26-command browser-driving allowlist), /sidebar-chat. ngrok forwards only this port.

Root tokens sent over the tunnel return 403. SSE endpoints use a 30-minute HttpOnly gstack_sse cookie (never valid against /command).

The 26-command tunnel allowlist

Defined in browse/src/server.ts as TUNNEL_COMMANDS. Pure gate function canDispatchOverTunnel(command) is exported for unit testing. Set:

goto, click, text, screenshot, html, links, forms, accessibility,
attrs, media, data, scroll, press, type, select, wait, eval,
newtab, tabs, back, forward, reload, snapshot, fill, url, closetab

Notably absent: pair, unpair, cookies, setup, launch, restart, stop, tunnel-start, token-mint, state, connect, disconnect. A remote agent that tries them gets a 403 plus a fresh entry in the denial log.

Tunnel denial log

~/.gstack/security/attempts.jsonl — append-only, salted SHA-256 of source

  • domain only (no raw IP, no full request body), rotates at 10MB with 5 generations. Per-device salt at ~/.gstack/security/device-salt (mode 0600).

See docs/REMOTE_BROWSER_ACCESS.md for the full operator guide.

Tab ownership

Scoped tokens default to tabPolicy: 'own-only'. A paired agent can newtab to create its own tab and drive that tab freely, but it can't goto, fill, or click on tabs another caller owns. tabs lists ALL tab metadata (an accepted tradeoff — see ARCHITECTURE.md), but text/html/snapshot content of unowned tabs is blocked by ownership checks.


Authentication

Three token types, three lifetimes, three scopes.

Token Generated by Lifetime Scope
Root token Daemon startup (random UUID) Daemon process lifetime Full command surface, local listener only — 403 over tunnel
Setup key POST /pair 5 minutes, one-time use Single redemption: present at /connect, get a scoped token
Scoped token POST /connect (with setup key) 24 hours Per-client, allowlist-bound, optionally tab-scoped

The root token is written to <project>/.gstack/browse.json with chmod 600. Every command that mutates browser state must include Authorization: Bearer <token>.

SSE endpoints (/activity/stream, /inspector/events) accept the Bearer token OR a 30-minute HttpOnly gstack_sse cookie minted via POST /sse-session. The ?token=<ROOT> query-param auth is no longer supported. This is what lets the Chrome extension subscribe to the activity feed without putting the root token in extension storage.

The Terminal pane uses a separate session cookie, gstack_pty, minted via POST /pty-session. Different scope — can spawn / drive the live claude PTY, can't dispatch arbitrary /command calls. /health endpoint MUST NOT surface this token.

Token registry

browse/src/token-registry.ts handles mint/validate/revoke for all three types, plus per-token rate limiting. Setup keys are single-use; scoped tokens have a sliding 24h window; the root token is rotated on each daemon startup.


Security stack

Layered defense against prompt injection. Every layer runs synchronously on every user message and every tool output that could carry untrusted content (Read, Glob, Grep, WebFetch, page text from $B).

Layer Module Lives in
L1 Datamarking content-security.ts both server + sidebar agent
L2 Hidden-element strip content-security.ts both
L3 ARIA + URL blocklist + envelope wrapping content-security.ts both
L4 TestSavantAI ML classifier (22MB ONNX) security-classifier.ts sidebar-agent only*
L4b Claude Haiku transcript check security-classifier.ts sidebar-agent only
L5 Canary token (session-exfil detection) security.ts both — inject in compiled, check in agent
L6 combineVerdict ensemble security.ts both

* security-classifier.ts cannot be imported from the compiled browse binary — @huggingface/transformers v4 requires onnxruntime-node which fails to dlopen from Bun compile's temp extract dir. The compiled binary runs L1L3, L5, L6 only.

Thresholds

  • BLOCK: 0.85 — single-layer score that would cause BLOCK if cross-confirmed
  • WARN: 0.75 — cross-confirm threshold. When L4 AND L4b both >= 0.75 → BLOCK
  • LOG_ONLY: 0.40 — gates transcript classifier (skip Haiku when all layers < 0.40)
  • SOLO_CONTENT_BLOCK: 0.92 — single-layer threshold for label-less content classifiers

Ensemble rule

BLOCK only when the ML content classifier AND the transcript classifier both report >= WARN. Single-layer high confidence degrades to WARN — this is the Stack Overflow instruction-writing FP mitigation. Canary leak always BLOCKs (deterministic).

Env knobs

  • GSTACK_SECURITY_OFF=1 — emergency kill switch. Classifier stays off even if warmed. Canary is still injected; just the ML scan is skipped.
  • GSTACK_SECURITY_ENSEMBLE=deberta — opt-in DeBERTa-v3 ensemble. Adds ProtectAI DeBERTa-v3-base-injection-onnx as L4c classifier. 721MB first-run download. With ensemble enabled, BLOCK requires 2-of-3 ML classifiers agreeing at >= WARN.
  • Classifier model cache: ~/.gstack/models/testsavant-small/ (112MB, first run only) plus ~/.gstack/models/deberta-v3-injection/ (721MB, only when ensemble enabled).
  • Attack log: ~/.gstack/security/attempts.jsonl (salted SHA-256 + domain only, rotates at 10MB, 5 generations).
  • Per-device salt: ~/.gstack/security/device-salt (0600).
  • Session state: ~/.gstack/security/session-state.json (cross-process, atomic).

A shield icon in the sidebar header shows the live status. See ARCHITECTURE.md § "Prompt injection defense" for the full threat model.


Screenshots, PDFs, visual

Screenshot modes

Mode Syntax Playwright API
Full page (default) screenshot [path] page.screenshot({ fullPage: true })
Viewport only screenshot --viewport [path] page.screenshot({ fullPage: false })
Element crop (flag) screenshot --selector <css> [path] locator.screenshot()
Element crop (positional) screenshot "#sel" [path] or screenshot @e3 [path] locator.screenshot()
Region clip screenshot --clip x,y,w,h [path] page.screenshot({ clip })

Element crop accepts CSS selectors (.class, #id, [attr]) or @e/@c refs. Tag selectors like button aren't caught by the positional heuristic — use the --selector flag form.

--base64 returns data:image/png;base64,... instead of writing to disk — composes with --selector, --clip, --viewport.

Mutual exclusion: --clip + selector, --viewport + --clip, and --selector + positional selector all throw.

Retina screenshots — viewport --scale

viewport --scale <n> sets Playwright's deviceScaleFactor (context-level, 13 cap):

$B viewport 480x600 --scale 2
$B load-html /tmp/card.html
$B screenshot /tmp/card.png --selector .card
# .card at 400x200 CSS pixels → card.png is 800x400 pixels

--scale N alone (no WxH) keeps the current viewport size. Scale changes trigger a context recreation, which invalidates @e/@c refs — rerun snapshot after. HTML loaded via load-html survives the recreation via in-memory replay. Rejected in headed mode (real browser controls scale).

PDF generation

pdf accepts the full Playwright surface plus a few additions:

  • Layout: --format letter|a4|legal, --width <dim>, --height <dim>, --margins <dim>, --margin-top/right/bottom/left <dim>
  • Structure: --toc (waits for Paged.js if loaded), --outline, --tagged (PDF/A accessibility), --print-background, --prefer-css-page-size
  • Branding: --header-template <html>, --footer-template <html>, --page-numbers
  • Tabs: --tab-id <N> to render a specific tab
  • Large payloads: --from-file <payload.json> (avoids shell argv limits)

Responsive screenshots

responsive [prefix] — three screenshots in one call: mobile (375x812), tablet (768x1024), desktop (1280x720). Saves as {prefix}-mobile.png etc.

prettyscreenshot

Combines cleanup + scroll + element hide in one call:

$B prettyscreenshot --cleanup --scroll-to "hero section" --hide ".cookie-banner" /tmp/clean.png

Local HTML

Two ways to render HTML that isn't on a web server:

Approach When URL after Relative assets
goto file://<abs-path> File already on disk file:///... Resolve against file's directory
goto file://./<rel>, goto file://~/<rel> Smart-parsed to absolute file:///... Same
load-html <file> HTML generated in memory, no parent-dir context needed about:blank Broken (self-contained HTML only)

Both are scoped to files under cwd or $TMPDIR via the same safe-dirs policy as eval. file:// URLs preserve query strings and fragments (SPA routes work).

load-html has an extension allowlist (.html, .htm, .xhtml, .svg) and a magic-byte sniff to reject binary files mis-renamed as HTML. 50MB size cap (override via GSTACK_BROWSE_MAX_HTML_BYTES).

load-html content survives later viewport --scale calls via in-memory replay (TabSession tracks the loaded HTML + waitUntil). The replay is purely in-memory — HTML is never persisted to disk via state save to avoid leaking secrets or customer data.


Batch endpoint

POST /batch sends multiple commands in a single HTTP request. Eliminates per-command round-trip latency — critical for remote agents over ngrok where each HTTP call costs 2-5s.

POST /batch
Authorization: Bearer <token>

{
  "commands": [
    {"command": "text", "tabId": 1},
    {"command": "text", "tabId": 2},
    {"command": "snapshot", "args": ["-i"], "tabId": 3},
    {"command": "click", "args": ["@e5"], "tabId": 4}
  ]
}

Each command routes through handleCommandInternal — full security pipeline (scope checks, domain validation, tab ownership, content wrapping) enforced per command. Per-command error isolation: one failure doesn't abort the batch. Max 50 commands per batch. Nested batches rejected. Rate limiting: 1 batch = 1 request against the per-agent limit.

Pattern: agent crawling 20 pages opens 20 tabs (individual newtab or batch), then POST /batch with 20 text commands → 20 page contents in ~2-3 seconds total vs ~40-100 seconds serial.


Capture

Console, network, and dialog events flow into O(1) circular buffers (50,000 capacity each), flushed to disk asynchronously via Bun.write():

  • Console: .gstack/browse-console.log
  • Network: .gstack/browse-network.log
  • Dialog: .gstack/browse-dialog.log

The console, network, and dialog commands read from the in-memory buffers (not disk) so capture is real-time even when disk is slow.

Dialogs (alert, confirm, prompt) are auto-accepted by default to prevent browser lockup. dialog-accept <text> controls prompt response text.


JS execution

js runs an inline expression. eval runs a JS file. Both run in the same JS sandbox — the only difference is inline-vs-file. Both support await — expressions containing await are auto-wrapped in an async context:

$B js "await fetch('/api/data').then(r => r.json())"   # auto-wrapped
$B js "document.title"                                  # no wrap needed
$B eval my-script.js                                    # file with await

For eval files, single-line files return the expression value directly. Multi-line files need explicit return when using await. Comments containing the literal token "await" don't trigger wrapping.

Path safety: eval rejects paths outside cwd or /tmp. js doesn't read files at all.


Tabs, frames, state

Tabs

$B tabs                          # list all open tabs
$B tab 3                         # switch to tab 3
$B newtab https://example.com    # open new tab, switch to it
$B newtab --json                 # programmatic: returns {"tabId":N,"url":...}
$B closetab                      # close current
$B closetab 2                    # close tab 2
$B tab-each "text"               # run "text" on every tab, return JSON

tab-each <command> fans out a command across every open tab and returns a JSON array — handy for "give me the text of every tab I have open."

Frames

$B frame "#stripe-iframe"        # switch to iframe by selector
$B frame @e7                     # by ref
$B frame --name "checkout"       # by name attribute
$B frame --url "stripe.com"      # by URL pattern match
$B frame main                    # back to top frame

Refs are cleared on switch (the iframe has its own AX tree).

State save/load

$B state save my-session         # save cookies + URLs to .gstack/browse-state-my-session.json
$B state load my-session         # restore

In-memory load-html content is intentionally NOT persisted (avoid leaking secrets to disk).

Watch

$B watch                         # passive observation: snapshot every 5s while user browses
$B watch stop                    # return summary of what changed

Useful when you're driving the browser manually and want Claude to see what you did at the end without spamming snapshot calls.

Inbox

$B inbox                         # list messages from sidebar scout
$B inbox --clear                 # clear after reading

The sidebar scout (a background process the Chrome extension can spawn) drops notes for Claude when the user surfaces something they want noticed. Stored in .gstack/browser-scout.jsonl.


CDP

$B cdp — raw Chrome DevTools Protocol dispatch

Deny-default. Only methods enumerated in browse/src/cdp-allowlist.ts (CDP_ALLOWLIST const) are reachable; any other method returns 403. Each allowlist entry declares scope (tab vs browser) and output (trusted vs untrusted). Untrusted methods (data-exfil-shaped, e.g. Network.getResponseBody) get UNTRUSTED-envelope wrapped output.

$B cdp Page.getLayoutMetrics
$B cdp Network.enable
$B cdp Accessibility.getFullAXTree --json '{"max_depth":5}'

To discover allowed methods: read browse/src/cdp-allowlist.ts.

$B inspect — CDP-based CSS inspector

$B inspect ".header"                # full rule cascade for the header
$B inspect ".header" --all          # include user-agent rules
$B inspect ".header" --history      # show modification history

Returns the matched rule cascade with specificity, computed styles, the box model, and (with --history) every CSS modification made via $B style since the page loaded. Powered by a persistent CDP session per page in browse/src/cdp-inspector.ts.

$B ux-audit

$B ux-audit

Returns JSON with site identity, navigation, headings (capped 50), text blocks, interactive elements (capped 200) — page structure for behavioral analysis without dumping the full HTML. Used by /qa and /design-review for cheap coverage maps.


Performance

Tool First call Subsequent calls Context overhead per call
Chrome MCP ~5s ~2-5s ~2000 tokens (schema + protocol)
Playwright MCP ~3s ~1-3s ~1500 tokens (schema + protocol)
gstack browse ~3s ~100-200ms 0 tokens (plain text stdout)
gstack browse + codified skill ~3s ~200ms 0 tokens (single skill invocation)

In a 20-command browser session, MCP tools burn 30,00040,000 tokens on protocol framing alone. gstack burns zero. The codified-skill path takes a 20-command session down to a single $B skill run call.

Why CLI over MCP

MCP works well for remote services. For local browser automation it adds pure overhead:

  • Context bloat — every MCP call includes full JSON schemas. A simple "get the page text" costs 10x more context tokens than it should.
  • Connection fragility — persistent WebSocket/stdio connections drop and fail to reconnect.
  • Unnecessary abstraction — Claude already has a Bash tool. A CLI that prints to stdout is the simplest possible interface.

gstack skips all of this. Compiled binary. Plain text in, plain text out. No protocol. No schema. No connection management.


Multi-workspace

Each project root (detected via git rev-parse --show-toplevel) gets its own daemon, port, state file, cookies, and logs. No cross-workspace collisions.

Workspace State file Port
/code/project-a /code/project-a/.gstack/browse.json random (1000060000)
/code/project-b /code/project-b/.gstack/browse.json random (1000060000)

Browser-skills three-tier lookup walks project → global → bundled, so a project-tier skill at /code/project-a/.gstack/browser-skills/foo/ shadows the global ~/.gstack/browser-skills/foo/ only inside project-a.


Environment variables

Variable Default Description
BROWSE_PORT 0 (random 1000060000) Fixed port for the HTTP server (debug override)
BROWSE_IDLE_TIMEOUT 1800000 (30 min) Idle shutdown timeout in ms
BROWSE_STATE_FILE .gstack/browse.json Path to state file
BROWSE_SERVER_SCRIPT auto-detected Path to server.ts
BROWSE_CDP_URL (none) Set to channel:chrome for real-browser mode
BROWSE_CDP_PORT 0 CDP port (used internally)
BROWSE_HEADLESS_SKIP 0 Skip Chromium launch entirely (test harness only)
BROWSE_TUNNEL 0 Activate the dual-listener tunnel architecture (requires NGROK_AUTHTOKEN)
BROWSE_TUNNEL_LOCAL_ONLY 0 Test-only — bind both listeners locally without ngrok
GSTACK_BROWSE_MAX_HTML_BYTES 52428800 (50MB) load-html size cap
GSTACK_SECURITY_OFF unset Emergency kill switch — disable ML classifier
GSTACK_SECURITY_ENSEMBLE unset Set to deberta for 3-classifier ensemble (721MB download)

Source map

browse/
├── src/
│   ├── cli.ts                   # Thin client — reads state, sends HTTP, prints
│   ├── server.ts                # Bun HTTP daemon — routes commands, dual-listener
│   ├── browser-manager.ts       # Chromium lifecycle, tabs, ref map, crash detection
│   ├── browse-client.ts         # Canonical SDK — what skills import as _lib/browse-client.ts
│   ├── snapshot.ts              # AX tree → @e/@c refs → Locator map; -D/-a/-C handling
│   ├── read-commands.ts         # Non-mutating: text, html, links, js, css, is, dialog, ...
│   ├── write-commands.ts        # Mutating: goto, click, fill, upload, dialog-accept, ...
│   ├── meta-commands.ts         # state, watch, inbox, frame, ux-audit, chain, diff, ...
│   ├── browser-skills.ts        # 3-tier walk + frontmatter parser + tombstones
│   ├── browser-skill-commands.ts # $B skill list/show/run/test/rm + spawnSkill
│   ├── browser-skill-write.ts   # D3 atomic stage/commit/discard helper for /skillify
│   ├── skill-token.ts           # mintSkillToken / revokeSkillToken (per-spawn, scoped)
│   ├── domain-skills.ts         # Per-site agent notes (state machine: quarantined→active→global)
│   ├── domain-skill-commands.ts # $B domain-skill save/list/show/edit/promote/rollback/rm
│   ├── cdp-allowlist.ts         # Deny-default CDP method allowlist
│   ├── cdp-bridge.ts            # CDP session lifecycle bridge
│   ├── cdp-commands.ts          # $B cdp dispatcher
│   ├── cdp-inspector.ts         # $B inspect — persistent CDP session per page
│   ├── activity.ts              # ActivityEntry, CircularBuffer, SSE subscribers, privacy filtering
│   ├── buffers.ts               # Console/network/dialog circular buffers (O(1) ring)
│   ├── tab-session.ts           # Per-tab session state (load-html replay, ref map scope)
│   ├── token-registry.ts        # Mint/validate/revoke for root + setup keys + scoped tokens
│   ├── sse-session-cookie.ts    # 30-min HttpOnly cookie for /activity/stream + /inspector/events
│   ├── pty-session-cookie.ts    # Separate scope: live Claude PTY auth
│   ├── tunnel-denial-log.ts     # ~/.gstack/security/attempts.jsonl writer (salted)
│   ├── path-security.ts         # validateOutputPath / validateReadPath / validateTempPath
│   ├── url-validation.ts        # URL safety checks for goto
│   ├── content-security.ts      # L1-L3: datamarking, hidden strip, ARIA, URL blocklist, envelopes
│   ├── security.ts              # L5 canary + L6 verdict combiner + thresholds
│   ├── security-classifier.ts   # L4 ML classifier (TestSavant + optional DeBERTa ensemble)
│   ├── terminal-agent.ts        # Side Panel Claude PTY manager (auth + lifecycle)
│   ├── sidebar-utils.ts         # Sidebar URL sanitization + helpers
│   ├── cookie-import-browser.ts # Decrypt + import cookies from real Chromium browsers
│   ├── cookie-picker-routes.ts  # HTTP routes for /cookie-picker/*
│   ├── cookie-picker-ui.ts      # Self-contained HTML/CSS/JS for cookie picker
│   ├── network-capture.ts       # Network request capture for $B network
│   ├── media-extract.ts         # Media element extraction for $B media
│   ├── project-slug.ts          # Project slug derivation for state paths
│   ├── error-handling.ts        # safeUnlink / safeKill / isProcessAlive
│   ├── platform.ts              # OS detection (macOS, Linux, Windows)
│   ├── telemetry.ts             # Anonymous opt-in usage telemetry
│   ├── find-browse.ts           # Locate running daemon or bootstrap
│   └── config.ts                # Config resolution (env / files)
├── test/                        # Integration tests + HTML fixtures
└── dist/
    └── browse                   # Compiled binary (~58MB, Bun --compile)

browser-skills/
└── hackernews-frontpage/        # Bundled reference skill
    ├── SKILL.md
    ├── script.ts
    ├── _lib/browse-client.ts
    ├── fixtures/hn-2026-04-26.html
    └── script.test.ts

scrape/SKILL.md.tmpl             # /scrape gstack skill — match-or-prototype entry point
skillify/SKILL.md.tmpl           # /skillify gstack skill — codify last /scrape into permanent skill

Development

Prerequisites

  • Bun v1.0+
  • Playwright's Chromium (installed automatically by bun install)

Quick start

bun install                      # install deps + Playwright Chromium
bun test                         # all integration tests (~3s for browse-only)
bun run dev <cmd>                # run CLI from source (no compile)
bun run build                    # compile to browse/dist/browse

Dev mode vs compiled binary

During development, use bun run dev instead of the compiled binary. It runs browse/src/cli.ts directly with Bun, so you get instant feedback:

bun run dev goto https://example.com
bun run dev text
bun run dev snapshot -i
bun run dev click @e3

The compiled binary (bun run build) is only needed for distribution. It produces a single ~58MB executable at browse/dist/browse using Bun's --compile flag.

Running tests

bun test                                    # all tests
bun test browse/test/commands               # command integration tests
bun test browse/test/snapshot               # snapshot tests
bun test browse/test/cookie-import-browser  # cookie import unit tests
bun test browse/test/browser-skill-write    # D3 atomic-write helper tests
bun test browse/test/tunnel-gate-unit       # canDispatchOverTunnel pure tests

Tests spin up a local HTTP server (browse/test/test-server.ts) serving HTML fixtures from browse/test/fixtures/, then exercise the CLI against those pages.

Adding a new command

  1. Add the handler in read-commands.ts (non-mutating) or write-commands.ts (mutating), or meta-commands.ts (server / lifecycle).
  2. Register the route in server.ts.
  3. Add the entry to COMMAND_DESCRIPTIONS in browse/src/commands.ts (with a clear description and usage — the gen-skill-docs validation suite enforces no | characters in description).
  4. Add a test case in browse/test/commands.test.ts with an HTML fixture if needed.
  5. Run bun test to verify.
  6. Run bun run build to compile.
  7. Run bun run gen:skill-docs to regenerate SKILL.md (the command appears in the command-reference table downstream).

Adding a new browser-skill

For a hand-written skill: copy browser-skills/hackernews-frontpage/, update SKILL.md frontmatter, rewrite script.ts against your target site, re-capture the fixture, update the parser test. bun test validates the SKILL.md contract (sibling SDK byte-identity, frontmatter schema).

For an agent-written skill: drive the page once with /scrape <intent>, say /skillify, accept the proposed name in the approval gate. The skill lands at ~/.gstack/browser-skills/<name>/ after the test passes.

Deploying to the active skill

The active skill lives at ~/.claude/skills/gstack/. After making changes:

cd ~/.claude/skills/gstack
git fetch origin && git reset --hard origin/main
bun run build

Or copy the binary directly:

cp browse/dist/browse ~/.claude/skills/gstack/browse/dist/browse

Cross-references

  • ARCHITECTURE.md — system-level architecture, dual-listener tunnel design, prompt-injection defense threat model
  • CLAUDE.md — project-level instructions, sidebar architecture notes, security-stack constraints
  • docs/REMOTE_BROWSER_ACCESS.md — operator guide for /pair-agent (setup keys, scoped tokens, denial log)
  • docs/designs/BROWSER_SKILLS_V1.md — design doc for browser-skills runtime (Phase 1 + 2a + roadmap)
  • scrape/SKILL.md/scrape skill: match-or-prototype data extraction
  • skillify/SKILL.md/skillify skill: codify last /scrape into permanent skill
  • TODOS.md/automate (Phase 2b P0), Phase 3 resolver injection, Phase 4 eval + sandbox

Acknowledgments

The browser automation layer is built on Playwright by Microsoft. Playwright's accessibility tree API, locator system, and headless Chromium management are what make ref-based interaction possible. The snapshot system — assigning @ref labels to AX tree nodes and mapping them back to Playwright Locators — is built entirely on top of Playwright's primitives. Thank you to the Playwright team for building such a solid foundation.

The prompt-injection L4 layer uses TestSavantAI/distilbert-v1.1-32 (112MB ONNX), and the optional ensemble layer uses ProtectAI/deberta-v3-base-prompt-injection-v2 (721MB ONNX) — both run locally via @huggingface/transformers.

The CDP escape hatch is gated by an allowlist directly inspired by Codex's T2 outside-voice review during the v1.4 design pass: deny-default with an explicit allowlist, not allow-default with a denylist.