A Claude Code skill that runs a deep end-to-end audit of a desktop app (Windows, macOS, Linux) or a web app/site. You get a tiered report and an annotated screenshot gallery covering two things:
- Functional: every feature is exercised through the real UI down to storage, network and integrations. Parallel area agents run in isolated sandboxes, and every serious finding is re-checked independently.
- Eyes: every surface is captured in every state, and same-kind surfaces are compared side by side. This catches the context menu that doesn't match the others and the overloaded card. It also includes a heuristic walkthrough, pixel measurements, and persona sessions that use the product like a person would.
git clone git@github.com:sectapunterx/deep-e2e-auditor.git ~/.claude/skills/deep-e2e-auditor
cd ~/.claude/skills/deep-e2e-auditor
pip install pillow # visual tools + report
cd scripts/web && npm run setup # web/Electron drivers: Playwright + Chromium + axe (optional)On Windows the path is %USERPROFILE%\.claude\skills\deep-e2e-auditor.
In Claude Code, inside the project you want audited:
/deep-e2e-auditor # both halves, whole product
/deep-e2e-auditor --visual # eyes only
/deep-e2e-auditor https://staging.example.com --quick
The skill reads the repo first, then asks one round of questions: scope, sandboxes, what must
never be touched, and real-data permission. It then proves a harness, maps features and surfaces,
and fans out agents. It writes everything to .deep-e2e/runs/<date>/ in the project, which is
git-excluded locally and never committed.
| Path | |
|---|---|
SKILL.md |
the workflow |
references/ |
safety rules, intake, harness recipes per platform/framework, functional method, eyes method, report and verification |
templates/ |
agent briefs (functional, eyes surfaces, persona, verification), finding schema |
scripts/win/ |
capture (all windows of a process, occlusion-proof), input (posted or real), UI Automation tree |
scripts/mac/, scripts/linux/ |
the same four jobs for macOS (JXA + screencapture) and X11/Xvfb (xdotool + ImageMagick + AT-SPI) |
scripts/web/ |
Playwright session runner and site crawler (screens × widths × themes, axe, console, overflow) |
scripts/visual/ |
contact sheets, pixel measurements (contrast, palette, density, diff), annotation |
scripts/report/ |
builds REPORT.md and gallery.html from the findings |
- Windows desktop drivers: tested on a Qt 6 app.
- Web drivers: tested with Playwright 1.63.
- macOS and Linux drivers: follow the platform APIs but are not tested on real hardware. The harness phase checks every command once before agents rely on it.