dsh-anti-detect-browser

Interactive browser tools for dsh (DeepSeek Harness) — dual-backend, architecture mirrored from the Hermes Agent browser tool. Gives dsh's agent real page interaction: navigate, click, fill, evaluate JS, screenshot — not just static fetch.

Tools

| Tool | What it does | |------|--------------| | browser_navigate | Open a URL; returns title, page text and a numbered list of interactive elements | | browser_snapshot | Re-list the current page (refs go stale after DOM changes) | | browser_click | Click an element by ref, returns a fresh snapshot | | browser_fill | Type text into an input/textarea ref (or select an option) | | browser_eval | Evaluate JS in the page context (Chromium backend only) | | browser_screenshot | Save a viewport PNG | | browser_close | Close the session; next call starts fresh |

Both backends return the same shape: {url, title, text, elements} where each element is {ref, tag, kind, text, href}.

Backends

Selection follows the same semantics as Hermes's is_camofox_mode:

• CAMOFOX_URL unset (default) → local full Chromium via Playwright. Executable discovery, in order:

  1. DSH_BROWSER_EXECUTABLE_PATH (also honors AGENT_BROWSER_EXECUTABLE_PATH)
  2. system Chrome/Chromium on PATH (google-chrome, chromium, chromium-browser, chrome)
  3. Playwright cache chromium-* builds (newest first, both chrome-linux64/ and chrome-linux/ layouts)

  chromium_headless_shell is never used — no extension hooks, broken fonts/codecs, trivially fingerprinted. If no full Chromium exists the tool throws with install guidance instead of falling back.

• CAMOFOX_URL set → Camoufox (anti-detect Firefox) via the camofox-browser REST API. Tab 404 (server GC) auto-recreates, same as Hermes. browser_eval returns an explicit unsupported error in this mode (the REST API has no JS eval).

Install

dsh plugin --profile web add dsh-anti-detect-browser -w
systemctl restart dsh-web   # however you run dsh

To enable the Camoufox backend, start a camofox-browser server (npm start in its checkout, or docker run -p 9377:9377 -e CAMOFOX_PORT=9377 jo-inc/camofox-browser) and inject the URL into dsh:

systemctl edit dsh-web      # [Service] → Environment=CAMOFOX_URL=http://localhost:9377
systemctl restart dsh-web

Environment variables

| Variable | Effect | |----------|--------| | CAMOFOX_URL | Switch to the Camoufox backend (server base URL) | | CAMOFOX_USER_ID / CAMOFOX_SESSION_KEY | Share a browser identity with an external integration | | DSH_BROWSER_EXECUTABLE_PATH | Force a specific Chromium binary (Chromium backend) | | DSH_BROWSER_SHOT_DIR | Screenshot output directory (default <home>/workspace/dsh-browser-shots) |

Development

The published tarball contains only the runtime (index.js, cordis.patch.yml + manifest). The smoke tests — Chromium backend against real pages, Camoufox backend against a mock REST server (including the tab-GC 404-recreate path) — live in the plugin's source checkout, not in this package. At minimum run node --check index.js after any edit, then restart dsh.

How it differs from dsh-browser

The npm package dsh-browser (by ben7am1n) covers similar ground with a CSS-selector-driven toolset (fetch HTML → guess a selector → click). This plugin ports the Hermes Agent approach instead:

• Ref-based snapshots — browser_navigate/browser_snapshot list every visible interactive element with a stable ref; the agent clicks by ref in one call instead of fetching HTML and guessing selectors across two.
• Camoufox anti-detect backend — set CAMOFOX_URL and the same tools run on an anti-fingerprinting Firefox (camofox-browser REST API) with server-GC-safe tab handling; nothing like it exists in dsh-browser.
• Deterministic binary discovery — env override → system Chrome → Playwright cache; full chromium-* builds only, chromium_headless_shell never used.

License

MIT