Browser Viz Verify

Browser Visualization Verify Skill

Purpose

NetClaw ships several skills that generate browser-based visualization outputs — threejs-network-viz, canvas-network-viz, drawio-diagram, uml-diagram, markmap-viz. Until now, confirming that a generated file actually rendered correctly (no blank canvas, no JavaScript errors, no broken layout) required a human to manually open it in a browser. This skill closes that gap: it opens the file itself in a controlled, headless Chrome session, takes a screenshot, reads the console for errors, and optionally runs a Lighthouse audit — then reports a clear verdict.

This skill never navigates to a remote URL and never touches the persistent authenticated browser profile used by browser-gui-inspect. It only ever opens local file:// paths. See specs/048-chrome-devtools-browser-inspection/ for the full spec, research, and design behind this skill.

MCP Server

This skill uses tools provided by the chrome-devtools-mcp server (official Chrome DevTools team package, chrome-devtools-mcp on npm), spawned locally via npx -y chrome-devtools-mcp@latest over stdio transport. See mcp-servers/chrome-devtools-mcp/README.md.

Environment Variables

None. chrome-devtools-mcp takes all configuration as CLI flags, not env vars. This skill always runs headless against a local file — it never needs --headless=false or the persistent profile's authenticated sessions.

Workflow: Verify a Generated Visualization (P1 — MVP)

Input:

• file_path (required) — absolute path to the generated visualization HTML file
• run_audit (optional, default false) — whether to also run a Lighthouse audit

Step 1: Confirm the file exists

Before opening a browser session, verify file_path exists on disk. If it does not, stop and return:

Copy code
verdict: "file_not_found"

Do not attempt to open a browser session for a path that doesn't exist.

Step 2: Open the file in a headless session

Copy code
Tool: new_page
Args: { "url": "file://<file_path>" }

Step 3: Wait for the page to settle

Async-rendered visualizations (three.js scenes, canvas drawings) may take a moment to finish drawing. Use a bounded wait rather than capturing immediately:

Copy code
Tool: wait_for
Args: { "text": <a string known to appear once rendering completes, e.g. a legend label>, "timeoutMs": 15000 }

If the wait times out, stop and return:

Copy code
verdict: "timed_out"

Step 4: Capture a screenshot

Copy code
Tool: take_screenshot

Save the result as the screenshot_path artifact, following the same output-directory convention as other NetClaw skill artifacts.

Step 5: Read the console

Copy code
Tool: list_console_messages

Classify the result:

• No errors/warnings present → verdict: "rendered_clean"
• One or more error/warning messages present → verdict: "rendered_with_errors", and include the message text in console_errors — don't just report a blank/inconclusive screenshot when the console explains exactly what broke.

Step 6: Optional Lighthouse audit

If run_audit is true:

Copy code
Tool: lighthouse_audit

Summarize the findings (performance, accessibility) into audit_summary. Skip this step entirely when run_audit is false — it adds meaningful runtime and isn't needed for a quick render check.

Step 7: Close the session

Copy code
Tool: close_page

Output

Copy code
{
  "verdict": "rendered_clean" | "rendered_with_errors" | "timed_out" | "file_not_found",
  "screenshot_path": "<path to captured screenshot>",
  "console_errors": ["<message>", ...],
  "audit_summary": { ... }   // present only if run_audit was true
}

Invoking from Other Skills

threejs-network-viz, canvas-network-viz, drawio-diagram, uml-diagram, and markmap-viz should each call browser-viz-verify automatically immediately after writing their generated HTML file, passing the freshly written file's path. This turns "did it render?" from a question the operator has to answer manually into a QA step NetClaw answers itself, within the ~30-second budget described in the spec's success criteria (SC-001). If browser-viz-verify returns rendered_with_errors or timed_out, surface that to the operator alongside the generated file rather than silently reporting success.

Error Handling

Constitution Notes

This skill is purely observational (Constitution I) — it never performs a write against any system, requires no ITSM gating (III), and needs no bespoke audit trail beyond DefenseClaw's generic tool-call logging (IV) since it is read-only and touches no external or authenticated system.