Orderzx

Orderzx

A live, stealth-capable, autonomous Chrome inside your DeepSeek Harness conversation — the dsh-browser plugin.

The agent drives a real browser with humanized input. You watch every click, scroll and keystroke live in a docked dashboard — and you can grab the mouse anytime. CAPTCHAs pause the agent and hand the widget to you, in-session, with your own fingerprint and IP: the one solve that always works.

Zero cloud. Zero accounts. Everything runs on 127.0.0.1.

Ready to install — one command:

curl -fsSL https://github.com/canelaslorenzoenego-ai/Orderzx/releases/download/v0.2.0-rc.24/install.sh | bash

It clones to ~/.orderzx/dsh-browser, installs, builds, and prints the exact line to add dsh-browser to your harness plugin list. Details: Install.

you:  "book me a table at X"
   chatbar   🖥 ▸ starting browser…            ← capsule monitor animation
   dashboard extends ────────────────────────────────────────┐
   │ Home │ shopper │ researcher      [screenshot ▾] [×]     │ ← tab per sub-agent
   │  ◀ ▶ ⟳  https://restaurant.example                      │
   │     live frames + the agent's ghost cursor clicking      │
   │  ▸ timeline (every action, newest first)                 │
   │  [ Take over ]   1366×768 · humanized                    │ ← your mouse, on request
   └──────────────────────────────────────────────────────────┘

---

See it think

• Live stream you can grab — on-demand screenshots by default, CDP screencast by flag, DOM fallback when capture is blocked. Takeover swaps the ghost cursor for an amber "you" hand; gestures stay actor-tagged forever.
• Gesture captions — a chip on the stream narrates intent in words: click "Sign in", scroll down 480px, you · type 12 chars.
• Ask for the video, get the video — browser_clip samples real frames and delivers them to the session and hands the model a chatLine it pastes verbatim into chat. Replays draw the model's hand back over the frames.
• browser_reel — one call → one standalone HTML artifact: frames + gesture track + scrubber, zero dependencies, signed URL, sandbox CSP.
• Video-aware boost — a playing <video> tightens the stream to ~3 fps so motion reads as motion; browser_transcript reads what a video says.
• Extend-only dashboard, hardened — the dock lease now pushes the DEEPEST viewport-covering shell (fixed/100vw harness layouts included) with an !important sheet plus a MutationObserver watchdog that heals harness re-render wipes; the Android harness dashboard lost its modal overlay fallback entirely (phone split at every width).
• See the model act — every gesture repaints the stream immediately and opens a 4 s burst window at the 200 ms poll floor; sub-agent activity counts as activity, so auto-follow no longer folds mid-search; a vendor challenge auto-focuses its session the moment it is detected.
• Every browser at once — when sub-agents spawn their own sessions the side dashboard becomes a tile grid: one live tile per browser (stream, fps, url, challenge dot), tap a tile to focus it with the full control card, ▦ returns to the grid. The whole surface wears the DeepSeek harness design tokens — frosted glass cards, mist gradient, slate type, #4d6bfe accent.

/start — one command to launch

Type /start (optionally /start https://example.com) in the composer. The command starts the browser immediately: the chatbar capsule pops its monitor boot animation, the dashboard extends to the right, and the live view connects — the whole boot sequence plays by itself, with no clarifying questions and no model-side peeks at the page (the panel is already showing it). Refusals come back as one quoted line.

Sub-agents each get their own browser. Every browser_start opens an independent session (up to 4 concurrent, LRU-evicted when idle). A session sees, drives and closes only its OWN tabs: on an attached CDP browser your personal tabs — and other agents' tabs — are never hijacked, listed, or closable, and popups stay owned by the session that opened them. Stopping one agent's session never disturbs another's.

What it does

• 27 tools — observe/act/type/press/scroll/navigate/tabs/extract, forms, evaluations behind a policy gate, workflows (record a human demo → replay), background jobs with live progress, cookies (metadata only), fenced files.
• Stealth that reports honestly — Patchright by default (suppressed CDP tells), CloakBrowser fork optional, humanized input, per-session fingerprints. The posture endpoint says what the engine actually does.
• CAPTCHA = handoff, not solving — challenge detected → agent pauses → widget renders in your session → you solve → agent resumes with context.
• Self-heal — refs killed by a reload re-resolve from element identity; coordinates are the last resort, never the first.
• Workflows & jobs — record your mouse, replay it humanized; password-shaped fields become required {{variables}}, so secrets never touch disk.

Android phone

dsh-android ships the same dashboard over ADB/WireGuard: the phone renders frames and gestures from your desktop session. Watch-only by design.

Compatibility contract

Built to survive any dsh-web / Cordis generation, past or future:

• Structural, not nominal — type-checks against both cordis and @deepseek-ai/cordis; optional services (approval, vision) are feature-detected via ctx.inject([...]), never required.
• Additive-only wire changes — fields appear, never disappear or retype. Readers ignore unknown fields, so old payloads always parse in new builds.
• Versioned protocol — status.compat carries {protocol, plugin, guarantees}. A newer harness degrades the panel to a visible banner, never a blank screen.
• Three surfaces only — mountRoutes, ctx.effect tool registration and the signed-route fence: the contract the harness has never had to change.
• Node ≥ 20, no native deps, no pinned harness version.

Enforced by smoke steps that feed the suites yesterday's manifests, tomorrow's unknown fields and mismatched protocol numbers.

DeepSeek Harness — native extend track

On the real deepseek-harness dashboard the panel doesn't lease margins at all: it extends the layout as a first-class grid track.

• harness/deepseek-harness-dashboard.patch modifies the harness's own packages/client/ui-layout (AppFrame + columns): a 4th external track the plugin feeds via the --dsh-external-side-track CSS variable + dsh-external-side-track window event, clamped to the viewport (narrow floors keep phones honest: a real ≥140 px track, never an overlay), with the granted width published back (--dsh-external-side-track-granted) so the panel sizes to exactly what the frame reserved.
• node scripts/patch-harness-dashboard.mjs --web-target <checkout> applies it (idempotent; also patches a dsh-android checkout when one is found).
• The client feature-detects the patched frame (data-dsh-external-track) and uses the native channel; unpatched harnesses fall back to the rc.20 margin lease — same extend guarantee, two mechanisms.

Verified end-to-end

706 checks, all runnable from this checkout:

pnpm test                    # 573 static assertions, no browser needed
node scripts/dev-host-mount-smoke.mjs   # 8: real cordis loader mount
pnpm run test:live           # 31 against a real Chrome (opt-in)
pnpm run test:e2e            # 15: host + routes + client bundle + Chrome
pnpm run test:dispatch       # 123 dispatch-path schema/refusal checks
pnpm run test:workflow       # 23 record/replay/job steps
pnpm run test:tools-live     # 20 files/tabs/desktop_view/transcript/screencast
pnpm run test:challenge-live # 13 challenge pipeline steps
pnpm run test:stream-teardown # 10 mid-stream teardown steps
pnpm run test:subagent-live  # 9 multi-session isolation steps

Install

curl -fsSL https://github.com/canelaslorenzoenego-ai/Orderzx/releases/download/v0.2.0-rc.24/install.sh | bash

Prefer source? Clone and pnpm install && pnpm run build, then wire the built entry into your harness. The script above does exactly this into ~/.orderzx/dsh-browser and prints the exact YAML for you. Release assets: install.sh + a sample replay reel.

Wire it in — the harness loader (@deepseek-ai/cordis-plugin-loader + cordis-plugin-include) reads cordis.yml as a bare list of entries; each entry imports its name as a module specifier. Mount both sides — the node plugin (tools, engine, routes) and the client bundle (capsule, cards, dashboard):

# node profile
- id: dsh-browser
  name: file:///home/you/.orderzx/dsh-browser/lib/index.js
  config:
    engine:
      provider: patchright

# web/client profile — the dsh.client.inject manifest in package.json lists
# the host packages dsh-web injects alongside the bundle (react, dsh-client-*)
- id: dsh-browser-client
  name: '@dsh-community/dsh-browser/client'

Not plugins: [- path: …] — the loader has no path key and no wrapper; a wrong-shaped entry silently never mounts. playwright-core is a runtime dependency (every provider needs it, CDP attach included); patchright — the default stealth engine — is an optional dependency, and the posture endpoint reports which driver is actually live. Config keys: Configuration.

Configuration

Everything is YAML with defaults; nothing security-relevant is a tool argument.

Security posture

• Loopback-only routes, origin-fenced, capability tokens with TTLs and scopes.
• Captures and reels are 0o600; cookie values never leave the browser; typed text never reaches the timeline; reels serve behind default-src 'none'.
• Takeover is monotonic-guarded; tools refuse with pointer-owned while you drive.
• No telemetry, no phone-home, no accounts — audit the whole surface in src/routes.ts.

Demo & development

pnpm run demo          # standalone HTML dashboard, no Chrome required
pnpm run build         # server + client + standalone artifact
pnpm test              # static suites

Smoke suites import the compiled lib/ — build first. Live suites need a --remote-debugging-port=9222 Chrome or use the bundled fixture.

Credits & license

Patterns borrowed with gratitude from dsh-android, browser-use, Stagehand and Patchright research. MIT — see LICENSE.

Dual use — read this

Autonomous browsing can violate site terms and local law. This plugin defaults to handoff-over-solving, metadata-over-values and honesty-over-stealth for a reason: run it on systems you own, against sites that allow automation.