dsh-omni-browser

Control a real browser — and take the whole page with you. 16 omni_* tools that drive Chrome/Edge over CDP and capture everything a page is as one self-contained evidence folder.

Most browser plugins in the DSH marketplace give you a remote control. dsh-omni-browser gives you a remote control plus a forensic camera: DOM, accessibility tree, HAR with response bodies, console log, cookies/storage, full-page screenshots, PDF and a screencast — all written into one timestamped folder with a manifest.

No npm dependencies, no bundled Chromium, no cloud service. It starts the Chrome you already have and talks raw CDP.

---

Why another browser plugin?

	Control-only plugins	dsh-omni-browser
Click / type / navigate	✅	✅
Read DOM & text	✅	✅
Accessibility tree with stable refs	sometimes	✅ ax:<backend-node-id>
HAR 1.2 with response bodies	❌	✅
Console + uncaught exceptions	sometimes	✅
Cookies / localStorage / IndexedDB / CacheStorage / service workers	❌	✅
Full-page screenshot + element screenshot + PDF	partly	✅
Screencast frames of what actually happened	❌	✅
One-call evidence bundle + manifest	❌	✅ omni_capture
Bounded multi-page harvest	❌	✅ omni_harvest

Use it for: agent research on real sites, bug reports that must be reproducible, compliance/archival snapshots, scraping where you need the network truth (not just the rendered text), and "why did the page show that?" debugging.

---

Install

# from the DSH marketplace registry (once listed)
dsh plugin --profile desktop add github:wanwujiehuyi-web/dsh-omni-browser

# or from a local checkout, in the DSH GUI
# plugin_manager → install_bundle → /absolute/path/to/dsh-omni-browser

Requirements: DSH with the tools service, Node 22+, and a local Chrome / Edge / Chromium.

---

Quick start

omni_open        { "url": "https://example.com" }
omni_page        { "mode": "snapshot" }
omni_capture     { "name": "example" }

omni_capture returns something like:

{
  "ok": true,
  "dir": "D:\\work\\omni-capture\\example-20261105-151230",
  "manifestPath": ".../manifest.json",
  "fileCount": 14,
  "totalBytesText": "1.8 MB",
  "files": [{ "path": "page.png", "bytes": 184212 }, "..."]
}

and the folder contains:

manifest.json          schema + per-artifact status + timings
page.json              title, url, meta, outline, links, forms, counts
page.png               viewport screenshot
page-full.png          full-page screenshot (beyond the fold)
page.pdf               print-to-PDF (vector text)
dom.html               rendered DOM (outerHTML)
text.txt               readable innerText
ax.json                accessibility tree (roles/names/values/backend ids)
network.har.json       HAR 1.2: headers, timings, post data, bodies
network.json           compact request list
bodies/0001-*.json     response bodies (failures by default, all with bodies:"all")
console.json           console messages + uncaught exceptions
storage.json           cookies + localStorage/sessionStorage/IndexedDB/CacheStorage
cookies.json           cookies only
perf.json              runtime metrics
resources.json         resource tree with sizes

Every part degrades gracefully: if PDF or a response body is unavailable, the rest of the bundle is still written and the manifest records the error.

---

The 16 tools

Tool	What it does
omni_status	Is Chrome running, which tabs, version, profile. action: launch | close | restart
omni_open	Open a URL (reuse tab or newTab), wait for load/domcontentloaded/networkidle
omni_tabs	list / new / close / activate
omni_page	Read the page: snapshot, text, html, ax, links, forms, meta, outline, frames, resources
omni_act	click, dblclick, hover, focus, type, press, scroll, select, check, uncheck, upload, drag, wait, back, forward, reload — real input events
omni_eval	Evaluate JS in the page, get JSON back
omni_screenshot	viewport / fullpage / element / clip → real file
omni_pdf	Print current page to PDF
omni_network	list, body (one response body), har, stats, save, clear
omni_console	list / clear / save console + exceptions
omni_storage	all, cookies, local, session, indexeddb, cache, sw, save
omni_capture	one-call evidence bundle
omni_harvest	bounded crawl, one bundle per page
omni_record	screencast the tab for N seconds → JPEG frames + manifest
omni_download	enable (set download dir), wait, list
omni_perf	metrics, trace, coverage, resources

Element targets accept { "selector": "#id" }, { "text": "Sign in" }, { "ref": "ax:1234" } (from omni_page mode=ax) or { "x": 100, "y": 200 }.

---

Configuration

All optional — set via the plugin config in cordis.patch.yml, or environment variables:

Env	Meaning	Default
OMNI_CHROME_PATH	Browser executable	auto-detect Chrome/Edge/Chromium
OMNI_PORT	CDP port (0 = pick a free port)	0
OMNI_HEADLESS	1 / 0	1
OMNI_USER_DATA_DIR	Browser profile directory	~/.dsh/omni-browser/profile-<DSH profile>
OMNI_OUTPUT_DIR	Where captures go	<cwd>/omni-capture/<timestamp>/
OMNI_HOME	Plugin data home	~/.dsh/omni-browser

The plugin uses its own browser profile, so it never touches your personal Chrome session. Cookies you log into inside the managed browser persist between runs (same profile directory).

---

Privacy & safety

• Captures may contain secrets (session cookies, tokens in headers, page content). Treat omni-capture/ as sensitive data. Use includeValues: false with omni_storage to keep keys but redact values.
• The browser profile is isolated by default, but anything you type into it is real — use a dedicated account for risky automation.
• This tool is for sites you are allowed to access. It is not a scraping-as-a-service bypass and does not solve CAPTCHAs.

---

How it works

index.js            host entry: config, shared session, 16 tool registrations
lib/cdp.js          Chrome lifecycle + raw CDP transport (WebSocket, no deps)
lib/snapshot.js     page model: text/html/AX/links/forms/meta/outline/resources
lib/actions.js      trusted input actions
lib/collect.js      Network/Console collectors, HAR 1.2
lib/storage.js      cookies + Web Storage + IndexedDB + CacheStorage
lib/capture.js      evidence-bundle orchestration + manifest
lib/shots.js        screenshots + screencast frames
lib/pdf.js          print-to-PDF
lib/perf.js         metrics / trace / coverage / resource tree
lib/downloads.js    download behavior + wait/list
lib/harvest.js      bounded multi-page crawl

Tests run against a local fixture site (test/server.mjs) with a real headless Chrome — no internet required:

node test/smoke.mjs        # CDP transport health check
node test/tools.test.mjs   # all 16 tools end-to-end

---

License

MIT © 2026 wanwujiehuyi