dsh-verify

中文 | English

Witness — The browser is the judge. The quality gate for agent-built web apps. Agents say done; the browser proves it. (Witness is the product name; dsh-verify is the package name — same thing.)

If Witness catches something for you, ⭐ star the repo — it's how this project stays alive.

You asked an AI to build a web app. It said "done." Does it actually work?

dsh-verify opens a real browser and checks — so you never have to take the agent's word for it.

The quality gate for agent-built web apps. Works with any agent — DeepSeek Harness (dsh), Claude Code, Cursor, Copilot, Codex — and with any CI. You write what a human would check in a browser; a real browser executes it and returns a PASS/FAIL verdict with receipts (screenshots + diff images).

No LLM judges the outcome. The browser is the judge.

Same task. Same AI. Two builds. One missing CSS rule — the agent's self-review passed, a real browser caught it.

Why this exists

We ran a 4-agent web team (spec writer → frontend dev → QA → reviewer). Their own review said:

✅ "All requirements met. No issues found."

In a real browser, the dark-mode toggle did nothing — the .dark class was toggled, but the CSS rule was never written. Every agent self-test passed because there was nothing in the page for the agents to run. No one opened a real browser.

That's the gap: agents verify against what they believe they built, not against what a user actually experiences. Unit tests and static checks can't catch a missing CSS rule.

Same page. Same JS. One missing CSS rule. Two different verdicts.

Why not just ...?

The agent graded its own homework. dsh-verify re-grades it in a real browser.

Use it three ways

From any AI agent (MCP)

claude mcp add dsh-verify -- npx -y -p dsh-verify dsh-verify-mcp

Then tell your agent, in plain words:

Verify http://localhost:3000 — click #dark-toggle, then check body background-color changed. Screenshot it.

Tools exposed: verify_spec (run a spec JSON), verify_url (inline checks, no files), generate_and_verify (the AI drafts the checklist, real Chromium executes it), health.

In CI (GitHub Action)

- uses: 263311487-ux/dsh-verify/.github/actions/dsh-verify@main
  with:
    spec: demo/fixed.json       # spec file or glob
    # url: https://staging.example.com   # optional override
    # out: dsh-verify-out               # report output dir (default)

The repo dogfoods it: the dogfood workflow asserts the fixed build passes and the buggy build fails on every push.

On the command line

npm install -g dsh-verify          # or: npx dsh-verify
npx playwright install chromium    # one-time browser download
npx dsh-verify --spec 'specs/*.json'
# [PASS] specs/home.json (5/5)
# [FAIL] specs/cart.json (4/5)
#   ❌ expect_text #total: got "0" want "99"

What's in the box

• Deterministic judge — a real headless Chromium (or Firefox / WebKit) executes human-style checks: click, fill, text, classes, computed styles, URLs, console errors, network errors, pixels.
• Receipts, not vibes — every run emits a self-contained HTML report with screenshots and red-highlighted diff images; --json for machines; exit 0/1 for CI.
• Visual regression — screenshot baselines, pixel-diff with thresholds (expect_screenshot), refresh with --update-baselines.
• AI-drafted checklists — dsh-verify gen --url ... --prompt "..." learns the page in a real browser, has an LLM draft the checklist, then executes it deterministically. The AI drafts; it never judges.
• Multi-browser — chromium | firefox | webkit per spec or --browser.
• Zero framework lock-in — a JSON spec is all there is. No config language, no SDK, no vendor.

Example spec

{
  "title": "my app",
  "serve": "dist",
  "browser": "chromium",
  "steps": [
    { "action": "goto", "path": "/index.html" },
    { "action": "click", "selector": "#count-btn", "count": 3 },
    { "action": "expect_text", "selector": "#count-btn", "text": "Clicked: 3" },
    { "action": "capture_style", "selector": "#page", "prop": "backgroundColor", "var": "bg_before" },
    { "action": "click", "selector": "#color-btn" },
    { "action": "expect_class", "selector": "#page", "class": "dark", "present": true },
    { "action": "expect_style_changed", "selector": "#page", "prop": "backgroundColor", "var": "bg_before" },
    { "action": "screenshot", "name": "final-state" }
  ]
}

Top-level fields: title, serve (static dir) or base (target URL), browser, steps. Run many at once with a glob; exit is 0 only if all pass.

The report

A self-contained HTML report — every step with a pass/fail badge, selector, and detail, plus screenshots:

Agent Arena — bring your agent

Real-browser benchmark for agent-built web apps: same 3 tasks, same human checks, open entry. Run your model on the board in ~10 minutes:

git clone https://github.com/263311487-ux/dsh-verify && cd dsh-verify
npm install && npx playwright install chromium
export LLM_API_KEY=sk-...          # any OpenAI-compatible model
node arena/run.mjs --agent "gpt-5/single" --task all --repeat 1 --submitter yourname

Your setup appears on the live leaderboard next to DeepSeek v4-flash / v4-pro: agent-arena. Full rules in docs/ARENA.md.

Prove it (run it yourself)

git clone https://github.com/263311487-ux/dsh-verify && cd dsh-verify
npm install && npx playwright install chromium
npm run demo:fixed    # → PASS (11/11)
npm run demo:buggy    # → FAIL (exit 1) — the missing .dark rule, caught
npm test              # engine self-tests

The repo's own CI runs exactly that — engine self-tests, then asserts fixed passes and buggy fails — so the tool verifies itself on every push.

Agent Arena — can agents ship working web apps?

Same task, same prompt, same human checks — different agents, graded by dsh-verify in a real browser. Latest run (2026-08-19): 44/48 runs passed across 2 models × 2 strategies × 3 tasks, 4 runs per cell. Two counterintuitive findings: the pricier v4-pro single-shot scored below the cheaper v4-flash single-shot (10/12 vs 11/12), and a real-browser self-check loop lifted v4-pro to 12/12 — while v4-flash's self-check crashed once when its own verification report came back as corrupt JSON. Every failure is reproducible and invisible to an LLM judge.

See docs/ARENA.md — methodology, the tasks, and how to run your own agent.

Badge your agent-built app

Built something with an AI agent? Prove it in a real browser and show the world:

[![agent deliverable: browser-verified](https://img.shields.io/badge/agent_deliverable-browser_verified-brightgreen?logo=playwright&logoColor=white)](https://github.com/263311487-ux/dsh-verify)

Add a spec, wire the GitHub Action, and the badge is earned, not claimed. See docs/verified-badge.md.

Read the full README on GitHub →