OpenCode mirror: sourced from ~/.config/opencode/skills/browser-crashtest-lab and mirrored for OpenCode CLI usage.

Browser Crashtest Lab

Overview

Run deterministic multi-pass browser audits with parallel worker processes and produce a final PASS/WARN/FAIL readiness report. Use this skill when broad surface testing is required, not just a single user flow.

Prerequisites

Run from any project directory:

command -v node >/dev/null 2>&1
command -v npx >/dev/null 2>&1

If missing, ask the user to install Node.js first.

Install runtime packages (project-local preferred):

pip3 install nodriver aiohttp

Lighthouse is pulled on demand with npx lighthouse.

For the second visual judge (NVIDIA Qwen), set:

export NVIDIA_API_KEY="..."

Ensure Codex CLI is available for the first visual judge:

command -v codex >/dev/null 2>&1

Skill Path

export OPENCODE_HOME="${OPENCODE_HOME:-$HOME/.config/opencode}"
export CRASHLAB="$OPENCODE_HOME/skills/browser-crashtest-lab/scripts/browser_crashtest_audit.mjs"

Quick Start

node "$CRASHLAB" \
  --url http://localhost:3000 \
  --workers 6 \
  --action-workers 6 \
  --link-workers 8 \
  --link-open-mode both \
  --design-zeugnis \
  --design-jury-mode dual \
  --design-fusion-mode union \
  --max-pages 120 \
  --max-actions-per-page 30 \
  --max-links-per-page 50 \
  --out-dir output/browser-crashtest

Then inspect:

• output/browser-crashtest/report.md
• output/browser-crashtest/report.json
• output/browser-crashtest/design-zeugnis.md
• output/browser-crashtest/design-zeugnis.json

Workflow

1. Run a wide crawl with high worker count.
2. Read report.md, especially failing pages and repeated patterns.
3. Re-run headed for targeted debugging:

node "$CRASHLAB" \
  --url http://localhost:3000/settings \
  --workers 1 \
  --link-open-mode browser \
  --headed \
  --max-pages 15 \
  --out-dir output/browser-crashtest-headed

1. Re-run after fixes and compare status and score deltas.

1. Default Codex judge mode is chat: the skill writes a strict JSON template + instructions for this Codex chat model.

• use --design-codex-judge-mode chat (default)
• fill output/browser-crashtest/design-judge-codex.json from this chat review
• rerun the same command (or use auto to try exec first, then chat/file fallback)

1. If you explicitly want CLI-driven codex runs, use:

• --design-codex-judge-mode exec

1. If Codex CLI execution is blocked/unavailable and you intentionally disable strict LLM requirement, use:

• --no-design-llm-required --design-codex-judge-mode file
• output/browser-crashtest/design-judge-codex.template.json
• output/browser-crashtest/design-judge-codex.instructions.md

Then rerun the same command.

Key Tuning Flags

• --action-workers <n>: Parallel micro-workers for button/click checks on each page.
• --link-workers <n>: Parallel micro-workers for link checks on each page.
• --link-open-mode browser|http|both:
• browser: open links in real tabs (default).
• http: fast status probes via browser context request API.
• both: run both and merge results for stricter audits.
• --design-zeugnis: Enable strict design jury and professor-style design report.
• --design-llm-required: Fail when any required judge is missing (default on).
• --design-jury-mode dual: Use Codex app visual judge + NVIDIA Qwen 3.5.
• --design-codex-judge-mode auto|exec|chat|file: Codex judge source (default chat).
• --design-codex-model <name>: Codex model override for local CLI judge runs (default gpt-5.3-codex).
• --design-codex-timeout-ms <n>: Timeout for Codex judge execution.
• --design-codex-attempts <n>: Codex judge attempts (default 3).
• --design-fusion-mode union: Fail if one judge fails.
• --design-qwen-endpoint <url>: Optional explicit endpoint override if base URL routing differs.
• --design-qwen-model <name>: Qwen model name (default qwen3.5-397b-a17b).
• --design-qwen-attempts <n>: Qwen judge attempts (default 3).
• --design-viewports 390x844,768x1024,1440x900: Multi-viewport evidence capture.
• --design-min-score-fail 920: Hard fail threshold on 1000 scale.
• --design-max-evidence-images <n>: Global cap for captured judge evidence.
• --design-codex-max-evidence-images <n>: Codex-specific evidence cap (default 6).
• --design-qwen-max-evidence-images <n>: Qwen-specific evidence cap (default 18).
• --design-codex-judge-file <path>: Optional explicit path for codex judge JSON.

What The Audit Enforces

• Crawl + coverage: Discover internal pages by BFS and test them in parallel processes.
• Link resilience: Open discovered links in a real browser by default and flag 4xx/5xx or request failures.
• Interaction resilience: Attempt safe button/control clicks with per-page parallel micro-workers.
• Runtime stability: Capture console errors/warnings, pageerror, and failed network requests.
• Accessibility baseline: Run axe-core (if installed).
• Best-practice baseline: Run Lighthouse categories (performance, accessibility, best-practices, seo) on sampled pages.
• Visual/UX heuristics: Compute a design score and classify:
• big-player-quality
• solid
• needs-polish
• high-risk
• Design Zeugnis: run dual-model jury and emit strict, evidence-based findings (P0..P3) plus release verdict.

Dual-Model Jury Mode

• Judge A (no API, default): this Codex chat model reviews evidence and writes design-judge-codex.json.
• Judge A optional CLI path: local codex exec with attached evidence images.
• Judge A fallback artifact path: design-judge-codex.template.json + design-judge-codex.instructions.md.
• Judge B (API): NVIDIA Qwen 3.5 (qwen3.5-397b-a17b) runs multimodal judgement.
• Fusion: union gate => if one judge says FAIL, final design verdict is FAIL.

Professorisches Design-Zeugnis

• Tone is intentionally hard and factual (--design-tone professor).
• Every finding must provide concrete evidence, impact, and a directly executable fix.
• Non-parseable/invalid judge output is retried and then treated as hard failure in strict mode.

Hard Gate Criteria

• Any P0 finding in deterministic checks or either model judge.
• Missing required judge output with --design-llm-required.
• Overall design score below --design-min-score-fail.
• Critical WCAG/Core Web Vitals issues.
• Codex CLI account quota/usage-limit errors in strict mode.

Judge Disagreement

• Disagreement is tracked in design-zeugnis.json -> fusion.disagreement.
• PASS requires both model judges to pass and no hard-fail condition.
• If judges disagree without P0, final verdict is at least WARN.

Optional DevTools MCP Escalation

When browser output is not enough (complex runtime/perf anomalies), start Chrome DevTools MCP:

"$OPENCODE_HOME/skills/browser-crashtest-lab/scripts/chrome_devtools_mcp_probe.sh" \
  output/browser-crashtest/devtools-mcp.log

Then connect from an MCP-capable client and inspect console, network, performance, and screenshots interactively.

Guardrails

• Avoid destructive actions by default; the script skips likely destructive controls (delete, pay, purchase, etc.).
• For authenticated apps, ensure the app is logged in before running wide audits.
• Keep artifacts under output/browser-crashtest* and avoid new top-level artifact folders.
• If a page is behind bot protection/CAPTCHA, document it and test critical flows manually.

References

Open only what is needed:

• Research-backed stack and links: references/stack-2026-02-26.md
• Worldclass standards and dual-judge policy: references/design-worldclass-standards-2026-02-27.md