dsh-verify v0.8.3

dsh-verify v0.8.3

Independent browser acceptance testing for AI-agent deliverables. Agents self-test and pass; real browsers tell the truth.

What's new

• Glama MCP server live — dsh-verify is now published on Glama (quality score 83%, server coherence A, tool definition quality A)
• npm 0.8.3 — installable via npm i -g dsh-verify
• MCP server — dsh-verify-mcp lets any AI agent (Claude Code, Cursor, Copilot...) run real-browser acceptance checks on its own deliverables
• Multi-browser — chromium / firefox / webkit via --browser or spec.browser
• Visual regression — screenshot diffing with style-change assertions
• AI-drafted checklists — generate_and_verify drafts acceptance specs from plain-language requirements
• GitHub Actions — CI workflow for agent-delivered web artifacts

Quick start

npm i -g dsh-verify
dsh-verify --spec demo/fixed.json --out /tmp/report

dsh-verify v0.7.0

v0.7.0 — visual regression + AI-drafted checklists + MCP + GitHub Action

Agents self-test and pass. Real browsers tell the truth.

What's new

• Visual regression — capture_baseline / expect_screenshot: first run creates a pixel baseline, later runs diff real-Chromium screenshots (pngjs), fail over the threshold, and embed a red-highlight diff image in the report. --update-baselines refreshes expected changes.
• AI-drafted checklists — dsh-verify gen --url <url> [--prompt ...] [--run]: learns the page in a real browser, an LLM drafts the checklist, deterministic Chromium executes it. The AI only drafts — it never judges.
• MCP server — claude mcp add dsh-verify -- npx -y -p dsh-verify dsh-verify-mcp: any MCP-capable agent (Claude Code / Cursor / Copilot) can verify its own deliverables. Tools: verify_spec, verify_url, generate_and_verify, health.
• GitHub Action — uses: 263311487-ux/dsh-verify/.github/actions/dsh-verify: one-step acceptance in CI with report artifact upload. The repo dogfoods it (fixed build must PASS, buggy build must FAIL).
• DSH plugin — installable via dsh plugin add dsh-verify.

Verified

• 16 self-tests, all green, including: buggy build detection through MCP and through the Action, AI-drafted spec PASS on fixed / FAIL on buggy, visual baseline create → PASS → page change FAIL (98% diff) → update → PASS.
• Demo: fixed vs buggy — same page, one missing CSS rule, different verdicts.

npm i -g dsh-verify · npx dsh-verify --help · MIT

v0.3.0

dsh-verify — independent browser acceptance testing for agent deliverables. Agents self-test and pass; real browsers tell the truth.

What's in v0.3.0

• Multi-spec glob: dsh-verify --spec 'specs/*.json' runs many specs, per-spec HTML reports, aggregated verdict, exit 0 only if all pass
• expect_navigation: wait for a URL change after a click (multi-page flows)
• Docker image: pinned to mcr.microsoft.com/playwright:v1.62.1-jammy (Chromium matches Playwright version; ⚠️ built locally, runtime verification pending on a Docker host)
• Console & network assertions (v0.2): expect_console_errors, expect_network_errors (4xx / failed requests)
• CI-friendly output (v0.2): --json machine-readable verdict; failed steps printed to stdout
• Engine self-tests: 7 cases (node:test), including the .dark bug reproduction

The story

A 4-agent web team self-reviewed "no issues found" — but in a real browser the dark-mode toggle did nothing (.dark was toggled in JS, never styled in CSS). Same page, one missing CSS rule: buggy build FAILS 10/11, fixed build PASSES 11/11. Receipts in evidence/.

Install

git clone https://github.com/263311487-ux/dsh-verify
cd dsh-verify && npm install && npx playwright install chromium
npm run demo:buggy   # FAIL — the exact bug
npm run demo:fixed   # PASS

CI self-verifies on every push: engine tests + fixed must PASS + buggy must FAIL.