An AI agent said "all requirements met." The dark-mode toggle did nothing — one CSS rule was never written.
demo/buggy — what the agents said
"No issues found"
Every agent self-test passed, because nothing was on the page for the agents to run. No one opened a real browser.
demo/fixed — what a real browser says
FAIL → PASS
Same page, same JS. One missing CSS rule. The agent's self-review passed; a real browser caught it.
Same task. Same AI. Two builds. Only a real browser tells the difference.
Agents verify against what they believe they built — not against what a user actually experiences. A browser check can cover rendered behavior that existing checks may miss. dsh-verify re-grades the homework in a real browser.
Use it three ways
Works with any agent — DeepSeek Harness, Claude Code, Cursor, Copilot, Codex — and any CI.
CLI
⌘Terminal
Write what a human would check in a browser; run npx dsh-verify --spec spec.json. Verdict + screenshots in seconds.
MCP
⇄Agents
Connect the MCP server to Claude Code, Cursor or Copilot — agents can self-test against a real browser before saying done.
CI
▸GitHub Action
One step in any workflow. Installs dsh-verify and Chromium in isolation; your project stays untouched.
Receipts, not opinions
Each run produces an HTML report with per-step results and adjacent screenshot and diff images. Keep the images beside the report when sharing it.
HTML report with adjacent screenshot files — per-step results and pass/fail badges.
Star the repo
If Witness catches something for you — a star is how this project stays alive.
