Verify ui

/verify-ui

Single-backend skill. Calls verify_ui_flow on the rolepod-uiproof MCP server and surfaces the structured result. No fallback (D-024) — if the tool is unavailable, this skill fails with a clear diagnostic.

When to use

• A diff changes user-visible behavior on a web target.
• A URL is reachable (dev server is running, or the target is deployed).
• You want to prove the UI works AND has no console errors / failed requests / regressed visuals — code-level tests can't do that.

When NOT to use

• Backend-only diffs (no UI change).
• Doc, config, or build-tool changes with no behavior surface.
• No dev server / target available — ask the user to spin one up first.
• iOS / Android targets — mobile is partially supported (basic input); console / network / set_env / evaluate are web-only.

Modes

• mode: 'assert' (default) — assertions describe what the feature should do; pass = feature works.
• mode: 'reproduce' — assertions describe what the bug looks like; pass = bug reproduces. When minimize: true (default) the tool then removes steps one-by-one to find the shortest still-reproducing sequence and writes replay-minimized.json next to replay.json.

Inputs

open — context setup

{ "platform": "web", "url": "https://...", "browser": "chromium" }

Optional: viewport, headless, user_agent, locale. UA / locale / timezone MUST be set here — they cannot change mid-session.

steps — UI actions in order

Each step is one of:

• { "kind": "click", "query": "Submit" }
• { "kind": "type", "query": "Email", "text": "[email protected]", "clear_first": true }
• { "kind": "key", "key": "Enter" }
• { "kind": "wait_for", "condition": { ... } }
• { "kind": "navigate", "url": "https://..." }
• { "kind": "hover", "query": "More" }
• { "kind": "drag", "from_query": "Card A", "to_query": "Column 2" }
• { "kind": "fill_form", "fields": [ { "query": "Name", "value": "Alice" }, { "query": "Subscribe", "value": true, "kind": "checkbox" } ] }
• { "kind": "upload", "query": "Avatar", "file_path": "/abs/path/to/file.png" }
• { "kind": "dialog", "action": "accept" } — place BEFORE the action that triggers the dialog
• { "kind": "set_env", "viewport": { "width": 375, "height": 812 } } — also accepts offline, geolocation, color_scheme, reduced_motion, extra_headers, network_throttle, cpu_throttle
• { "kind": "switch_page", "index": 1 } — multi-page (popups, target=_blank)
• { "kind": "evaluate", "script": "return document.title" } — gated by ROLEPOD_ALLOW_EVAL=1
• { "kind": "scroll", "direction": "down", "amount": 1500 } — scroll by a delta (amount defaults to 400px); use to fire scroll-reveal / IntersectionObserver content
• { "kind": "settle" } — scroll the full page to trigger all reveals + lazy loads, network-idle, then return to top (web-only). Reach for this on reveal/animation-heavy pages so text_visible assertions see the settled state instead of an opacity:0 placeholder.

expect — assertions

• { "kind": "text_visible", "text": "..." }
• { "kind": "text_absent", "text": "..." }
• { "kind": "url_matches", "pattern": "regex" }
• { "kind": "ref_in_state", "query": "Submit", "state": "enabled" }
• { "kind": "no_console_errors", "exclude_patterns": ["3rd-party.com"] }
• { "kind": "no_failed_requests", "exclude_patterns": ["/analytics"], "allow_4xx": false }
• { "kind": "request_made", "url_pattern": "/api/checkout", "method": "POST", "min_count": 1 }
• { "kind": "response_status", "url_pattern": "/api/me", "status": 200 }

capture — evidence

Default: ["screenshot"]. Available:

• screenshot — final.png
• console — console.json (filtered errors+warnings, ring buffer up to 1000)
• har — network.har (full HAR)
• video — videos/*.webm
• trace — trace.zip (Playwright trace; view with npx playwright show-trace)
• a11y_tree — a11y_tree.json (final snapshot)

Defaults

• close_on_finish: true
• minimize: true (only consulted when mode: 'reproduce')

Outputs

• run_id, passed, failed_at_step, failure_reason, final_url_or_screen
• evidence_paths: { screenshots, replay_bundle, console?, a11y_tree?, har?, trace?, video? }
• minimized (only on mode: 'reproduce' + passed: true + minimize: true)

Process

1. Build the verify_ui_flow input.
2. Call the tool.
3. Report the structured result. On failure include failed_at_step + failure_reason + relevant evidence paths (screenshot, console.json if console errors caused the failure).

Default suggestion

For ANY user-visible flow, default-include no_console_errors and no_failed_requests in expect. Real UI bugs surface as console errors or 5xx responses far more often than as wrong text.

Evidence routing

Run artifacts are saved under:

• Standalone: .rolepod-uiproof/artifacts/<prefix>_<ts>_<uuid>/
• With rolepod parent (detected via the marker file <git-root>/.rolepod/parent-active written by the parent's SessionStart hook): <git-root>/.rolepod/evidence/<ts>-rolepod-uiproof-<skill>/

Either way the run directory contains a manifest.json per Extension Protocol v1, so the parent's check-work skill can aggregate results into the verify phase report. Standalone users can read the manifest themselves — same shape.

If the tool is unavailable

The /verify-ui skill needs the rolepod-uiproof MCP server, which is not currently available. Confirm the plugin is installed and try again, or check that npx -y @rolepod/uiproof is reachable.

Do not attempt this work via Playwright MCP, Chrome DevTools MCP, or any other backend from inside this skill. Multi-backend routing is the job of the parent rolepod plugin's check-work / debug-issue skills.

Examples

Success — verify checkout flow with no errors

User: "Verify https://shop.example.com/checkout works — fill the form, submit, expect a success page and no errors."

{
  "mode": "assert",
  "open": { "platform": "web", "url": "https://shop.example.com/checkout" },
  "steps": [
    { "kind": "fill_form", "fields": [
      { "query": "Name", "value": "Alice" },
      { "query": "Email", "value": "[email protected]" },
      { "query": "Card", "value": "4242 4242 4242 4242" }
    ]},
    { "kind": "click", "query": "Pay" },
    { "kind": "wait_for", "condition": { "kind": "text_visible", "text": "Thank you" } }
  ],
  "expect": [
    { "kind": "text_visible", "text": "Thank you" },
    { "kind": "no_console_errors" },
    { "kind": "no_failed_requests", "exclude_patterns": ["/analytics"] },
    { "kind": "response_status", "url_pattern": "/api/checkout", "status": 200 }
  ],
  "capture": ["screenshot", "console", "har"]
}

Failure with evidence

When no_console_errors fails, the result surfaces:

{
  "passed": false,
  "failure_reason": "Expectations failed: expect[1] no_console_errors",
  "evidence_paths": {
    "screenshots": ["…/final.png"],
    "console": "…/console.json"
  }
}

Open console.json to inspect the errors.

Dialog handling

User: "When the user clicks Delete, a confirm dialog appears. Verify that accepting it deletes the row."

{
  "steps": [
    { "kind": "dialog", "action": "accept" },
    { "kind": "click", "query": "Delete" },
    { "kind": "wait_for", "condition": { "kind": "text_absent", "text": "Row A" } }
  ],
  "expect": [ { "kind": "text_absent", "text": "Row A" } ]
}

The dialog step arms a one-shot handler; the next trigger (the click) fires it. Un-armed dialogs are auto-dismissed.

Responsive + dark mode

User: "Verify mobile dark-mode layout."

{
  "steps": [
    { "kind": "set_env", "viewport": { "width": 375, "height": 812 }, "color_scheme": "dark" }
  ],
  "expect": [ { "kind": "text_visible", "text": "Menu" } ],
  "capture": ["screenshot"]
}