Test pages for browser automation challenges. Each page demonstrates different web features that can be tricky to automate.
Simple login form with username and password fields
Targets for INPUT_TEXT {{key}} tokens (Tab, arrows, Space, Escape, Ctrl+A/C/V) with a keydown log for assertions
The keystroke-token targets hosted inside an iframe, so a field XPath resolves into the frame
Multi-line textarea target for the INPUT_TEXT Paste (clipboard + Ctrl+V) toggle; echoes value plus char/line counts for assertions
Canvas signature boxes (signature_pad library + raw pointer events) for the CLICK Draw click type. Shows strokes, points, isTrusted and the exported PNG (?pad=library|raw|both, ?variant=iframe)
Login backed by a real Set-Cookie session; page auto-redirects to the dashboard when the cookie is present. For testing session-persistence/browser-state behavior.
Three tabbed portals with identical flow but different fields — for testing parameterized components
Run the admin fingerprint collector from the demo site without logging in
Patient ID lookup form that rejects invalid IDs (for testing input_required recovery)
Missing expected element / precondition scenarios (no-result, conditional action, stalled post-state) that classify as UPSTREAM_ERROR / PREREQUISITE_NOT_MET — for the Fix-with-Builder-Agent button
Package tracking form with conditional popup for certain inputs
Drag-and-drop file uploader with metadata display
Controllable upload target for network-signal classification (?mode=sync|commit|async, ?status=, ?delay=, ?commitStatus=, ?verdict=). Issues a real browser POST so the network listener captures server-side upload failures (5xx/4xx/slow, upload-ok-but-commit-failed, async accept/reject).
Basic select dropdowns including one inside an iframe
Various select types: native, custom, searchable, multi-select, radio listbox, datalist
Password change forms (?variant=kinnser|availity|ecw|onehealthcare|reject-first|alphanumeric)
Standalone 2FA setup screens (?variant=dismissable|method-selection|authenticator|authenticator-qr-only|sms-verify|email-verify)
Stateful login → TFA setup → code entry → dashboard. After setup, subsequent logins ask for code only. Add ?reset=true to clear state (?variant=dismissable|authenticator|authenticator-qr-only)
Form with breakable selectors for XPath recovery propagation testing (?broken=true swaps selectors)
Healthcare prior-auth portal simulating framework migration drift. Full self-healing cycle: Stage 0 LLM static attempt → COMPUTER_USE fallback → pre-click capture → crystallize → verify → promote. (?drift=false|true|hard ?delay=<ms>)
Button with drifting id (changes every page load) for testing the full healing promotion pipeline: AI_FALLBACK → XPath capture → writeback → static graduation
Button + input with per-variant attributes for XPath cascade tier testing (?variant=default|data-testid|id-attr|aria|placeholder|role|text-content|class-only|broken|heavy-dom|multi-action|shifted-attrs|table|delayed-render)
Full CRM dashboard: login (admin/password), cookie banner, sidebar nav, 15-row data table, modal contact form with 3-tab wizard, toast notifications, 1000+ DOM elements (?variant=default|logged-out|modal-open|cookie-banner|slow-load|404|duplicate-contacts, ?reset=true)
Healthcare claims portal: login (provider/claims123), patient search with shuffled columns, 4-step claim wizard, accordion sidebar, CAPTCHA overlay, version drift, session expiry (?variant=default|expired-session|captcha|maintenance|rate-limit|slow-results|deleted-patient|password-expired|force-logout, ?version=1|2, ?reset=true)
Full flow: login → password expired → dashboard with claim form. Tests end-to-end password update recovery.
Dense insurance claims form for enrichment testing
Two identical Submit buttons; an XPath selector matches both, triggering the deterministic multiple-matching classification under INCORRECT_FORM_INPUTS
Dependency-free start page for desktop/RDP workflows, replacing example.com as the Start node URL. A fixed 'CloudCruise Start Page' heading lets a vision check confirm the browser landed.
Repro of CareTend's cascading Workflows hover menu (Workflows → Infusion/SPRx Workflow → ePrescription Manager). Tests whether teleport-style clicks keep the cascade open or whether the pointer must dwell/hover along the trajectory. ?mode=hover|hover-strict|click|sticky, ?openDelay=ms (submenu render delay), ?grace=ms. Live HUD shows open levels + last-opened leaf; stable data-testids on every menu node.
Repro of the crowded ePrescription Manager grid: dense filter row, crammed column funnel icons, and a vertically + horizontally scrollable ~105-row table. Working Type filter (funnel → 'Is equal to' → type 'New Rx' → Count updates) plus a Filter Search box; rows carry <No Matching Patient>/<Multiple Matches> sentinels. All synthetic data (no PHI). ?rows=N, ?type=New+Rx. Live scrollTop/scrollLeft + count readouts.
Deterministic surface for validating LOOP-based positional grid harvesting. Fixed 19-row viewport with dedicated scroll-up/down buttons (data-testid scroll-up-row / scroll-down-row) that advance the grid by exactly one row per click, plus fully-identical adjacent New Rx rows so positional capture can be tested against content-dedup. All synthetic data (no PHI). ?rows=N (capped 2000), ?type=New+Rx.
Deterministic surface for BOOL_CONDITION[LLM_VISION] model bake-offs. Ten ?variant= states reproducing the cases where cheap bool models actually disagree: a spinner fully visible vs partially hidden behind a dialog vs mid-fade, a static refresh icon that is not a spinner, height/weight populated with '0' vs real values vs empty, and a selected row vs a hover-tinted one. Ground truth for each check is on data-expected (invisible to vision) so a workflow can assert without leaking the answer into the screenshot. All synthetic data (no PHI). ?variant=clean|spinner-full|spinner-occluded|spinner-faded|refresh-icon|fields-zero|fields-values|row-selected|row-hover|popup.
Classic checkbox "I'm not a robot" captcha test (?sitekey=)
Invisible captcha with score verification (0.0-1.0)
Standard hCaptcha widget on hCaptcha's own hosts (?mode=checkbox|invisible|late, ?size=compact, ?sitekey=)
Turnstile widget in the embed shapes the solver detects, on Cloudflare test sitekeys (?embed=implicit|explicit|id|attrs|shadow|script-sibling|interstitial|decoy, ?state=auto-pass|click|invisible|block, ?sitekey=)
AWS WAF CAPTCHA puzzle via the JavaScript API (?mode=inline|button|modal|late, ?skipTitle=1, ?dynamicWidth=1, ?locale=, ?noLang=1, ?jsapi=, ?apiKey=)
Distorted-character transcription challenges across many layouts, carriers and controls, for detection + solving tests (?variant=classic|inline|aria|canvas|background|refresh|all)
One trigger per Chrome-owned surface: JS dialogs, OS file/print windows, permission bubbles, device pickers, HTTP auth, cert errors, native select/date/color popups, validation and password bubbles, hangs. Logs open/settle times to window.__chromeDialogLog so detection layers can be scored per dialog. ?danger=1 exposes a tab-crash trigger; ?lang=fr triggers the translate bubble.
DISMISSIBLE popup that auto-dismisses on a timer before the injected close-recovery click executes. Exercises the popup-gone no-op path: recovery should take no action and resume to the Continue button. ?dismiss_after=0 keeps it visible (recovery must actively close it); ?dismiss_after=<ms> tunes the timing.
Test opening popup windows and handling multiple windows
Simple popup window with close button
Success popup - target window in multi-window test
Trap popup - decoy window in multi-window test
15 scenarios for unexpected UI state: 1–4 DISMISSIBLE (cookie banner, newsletter, promo, survey), 5–12 DECISION_REQUIRED generic (duplicate record, destructive confirm, stale record, scheduling conflict, outstanding balance, coverage warning, duplicate submission, threshold exceeded), 13–15 DECISION_REQUIRED 1:1 prod reproductions (ECW code-already-exists Yes/No, Noridian Consent to Monitoring Accept/Reject, ECW Duplicate Patient Warning Proceed/Cancel)
Fully-fledged multi-PAGE EHR portal — real route navigation (login → dashboard → patients → chart → coding → review → done) with clickable sidebar nav, a patient results table with selectable rows, and chart tabs. ?seed=N carries across pages and deterministically scatters DISMISSIBLE / DECISION_REQUIRED popups at varied timings (on entry, after a delay, after the step's action). Knobs: ?density, ?only=decision|dismissible, ?popups=N. Login is never interrupted; every seed exercises ≥1 decision modal
1:1 ecwcloud.com — multi-step modal chain: duplicate-patient warning → code-already-exists; tests counter reset on verified dismissal between modals
1:1 essentials.availity.com — bottom-rectangle consent banner spanning full width with 4 buttons (Accept All / Reject All / Customize / ×); tests non-centered modal placement
1:1 eprg.wellmed.net — full-screen overlay with scrollable legal text and a single 'I Acknowledge' CTA disabled until scroll-to-end; tests scroll-required single-CTA case
1:1 app.azaleahealth.com — bottom-right corner toast (not overlay) with 3 buttons (Yes / Maybe Later / Don't ask again); tests non-overlay corner placement
1:1 thespot.fcso.com — single-CTA modal whose click causes navigation to a /login flow; tests identifyRecoveryNode smart-resume after state-loss
1:1 app.propertymeld.com — wrong-modal (Edit Meld Residents) appears unexpectedly on Save Note; only × close button, no labeled CTAs; tests close-only dismissal path
1:1 www.portal.jnjwithme.com — single-CTA acknowledge modal whose action terminates the workflow as failed (patient already enrolled, no recovery path)
1:1 ecw.cmc-pa.com — information-dense modal with embedded error list inside the body, OK / Cancel buttons; tests information-rich modal extraction
1:1 app.weinfuse.com — feature-prompt modal triggered after file upload with 3 buttons (Scan Now / Maybe Later / Don't Show Again) and illustration; tests feature-prompt 3-option dismissal
1:1 familycarecenter.insynchcs.com — wrong-context modal (Zip Search) appearing when opening a gender dropdown; modal has a text-input field plus 2 buttons (V2 gap: modal_action MVP supports buttons only, this exposes the input-inside-modal limit)
Multiple levels of nested iframes with different sandbox attributes
Form elements inside iframe with security context display
The Login + TFA Setup Flow inside a same-origin iframe, so the TFA code field lives in a child frame. Query params pass through to /tfa-login-flow.
Single-row table extraction inside an iframe. Use //tr[@data-row-id='row-2'] as the container XPath and short child XPaths like /td[@class='typeDesc'].
W3Schools-style result frame: createElement('iframe') + appendChild (about:blank) + document.write, recreated on each Run. Add ?delay=2000 to reproduce the slow-render injection race.
Frame served with script-src 'none' so the main-world inject is blocked; target is streamed in after ?delay ms. Realistic trigger for the isolated-world XPath fallback's early-bail.
Sandboxed iframe without allow-same-origin → opaque origin, so the parent can't read contentDocument (same as cross-origin). Forces the iframe relay path; regression guard for cross-origin find/click on a single host.
Embeds https://example.com — a real cross-origin frame that allows framing. contentDocument is null to the parent, so only the iframe relay reaches it. Regression guard for the relay path; target the 'Example Domain' heading.
Iframes mounted inside open and closed shadow roots — querySelectorAll('iframe') and a document-scoped MutationObserver don't cross shadow boundaries, so the frame may never be injected.
3+ levels of same-origin nesting with a click/type target in the deepest frame. Native (coordinate) actions must compose the frame offset across every level. ?level=N, ?delay=ms.
Same 3+ level same-origin nesting as Iframe Deep Nested Actions, but the click/type target in the deepest frame lives inside an open shadow root instead of light DOM. Reproduces PE-767: a DOM/aria-tree walk that recurses into nested iframes but never checks element.shadowRoot sees an empty host at the deepest level. ?level=N, ?delay=ms.
Full-viewport same-origin iframe with 20 scrollable sections. Tests full-page scroll+stitch during error screenshots. 'Simulate navigation gap' button blanks the iframe src for 400ms to reproduce the relay-disconnect window (iframeInject port disconnects on navigation, connectedIframes goes empty). Before fix: screenshot = single viewport. After fix: full 20-section scroll via direct DOM fallback.
Scaled (transform: scale) and scrolled-container iframes — native click coordinates must account for the shifted real position of the frame.
Frame src holds the HTTP response for ?delay ms (default 3s), so the frame sits at about:blank through a real network wait before any content exists.
loading='lazy' iframe far below the fold — contentWindow stays about:blank until scrolled into view, so injection must handle a frame that exists long before it loads.
The same iframe element is renavigated/reloaded to a new document after injection; the extension must re-establish frameId and listeners on a frame it has already seen.
Click/type target inside a srcdoc frame (document URL is about:srcdoc, not about:blank) to exercise the marker/injection logic on that path.
Same form at top level, same-origin iframe, and cross-origin iframe. Isolates whether pointer-action (click/input) pre-action screenshots get dropped on the iframe relay path vs the top-level path.
Origin chain parent → cross-origin → back to parent's origin; the postMessage relay must hop across a cross-origin boundary and back to reach the innermost target.
Injects many iframes in one synchronous burst to stress the MutationObserver's 100ms debounce — checks none are dropped. ?count=N.
Removes the iframe and appends a fresh one reusing the same id (SPA route churn) — tests that stale frameId references and listeners are not reused.
Simulates chart notes structure with nested PDF viewer
Clinical notes form with embedded PDF iframe
Full-viewport iframe with 20 tall sections. ?sandboxed=1 loads the iframe with sandbox='allow-same-origin' (scripts disabled → iframeInject can never run → relay always disconnected, but contentDocument accessible). Cleanly isolates the direct DOM fallback fix in getScrollInfo: WITHOUT fix = single viewport screenshot; WITH fix = full 20-section scroll.
Scroll to elements below the fold on a long page
Scrollable containers: log viewer, chat history, horizontal carousel
Terms and Conditions modal with scroll-to-bottom to enable OK button
Nested scrollable containers: sidebar + main content with inner scrollable areas
Content that loads on scroll (pagination trigger, lazy loading)
Scrollable iframes, sticky headers, and CSS scroll-snap
Scrollable dropdown option lists: custom ARIA, native select, hidden JCF-style
Collapsible sections that push content below the fold when expanded
Windowed list (only visible rows in DOM) and overflow:hidden clipped content
Scrollable containers inside shadow DOM and floating popover/tooltip
Smooth scroll, bidirectional grid, CSS transform positioning, keyboard tab focus
Wide strip requiring horizontal scroll to reveal off-screen columns (?layout=page|container). Live window.scrollX readout for verifying left/right scroll.
Target off-screen both down AND right — exercises to-element 2D auto-reveal (scrolls both axes). Live scrollX/scrollY readout.
Like the 2D page but the whole canvas is tiled edge-to-edge with visually-distinct cells (coordinate + token + hue), so every scroll step registers pixel change — no blank stretch that defeats the no-DOM pixel-shift detector. Target is the far bottom-right cell (down AND right). Live scrollX/scrollY readout.
Bordered container scrollable in both axes with a far-right target inside — exercises region find-element / full-container horizontal. Live container.scrollLeft readout.
Target near the top edge that can't be centered (page won't scroll above 0); loads scrolled to the bottom. Verifies scroll-to-element succeeds when the target is revealed but un-centerable at the start boundary.
Wide horizontal grid with a smaller vertical-only notes pane overlapping its center. A horizontal scroll must pick the grid, not the vertical decoy — verifies the builder's axis-filtered container selection. Live grid.scrollLeft + decoyPane.scrollTop readouts.
Numbered rows under a fixed header that appears only after scrolling off the exact top — full-page screenshot stitching must keep rows contiguous with no header duplication mid-page. layout-shift mode swaps in a color bar that's removed on scroll plus a nav header that goes position:fixed (?mode=appear|always|none|layout-shift, ?rows=N, ?headerHeight=N)
Live read of wheel/scroll event ratio, inter-event CV, burst structure, and wheel-to-click gap. Mirrors the admin /admin/fingerprint scoring for localhost iteration.
Measures where native CLICK and INPUT_TEXT actions actually land versus the element center. Y-ladder separates a constant offset from a y-proportional scale error; height-ladder finds the box height at which clicks start missing. Also flags Ctrl+A selections that escape the input.
Drives Chrome's real load state so the tab throbber can be correlated with telemetry. Use one knob at a time: ?mainDelay=8000 (server-held document), ?iframeHang=5000 (hanging iframe, add &iframeOrigin=opaque or &iframeDeferMs=3000), ?afterMs=6000 (slow app, no network), ?fakeSpinner=1 (negative control - cosmetic spinner, zero network)
Progressive loading with overlay, skeletons, lazy content, and delayed dropdowns (?delay=X&stealth=true hides indicators)
Reports the browser's public exit IP alongside the host's own, so a proxied browser is distinguishable from an unproxied one (?marker=X names the sample).
Button appears after configurable delay (?delay=X seconds)
Click target does not exist until ?appearAfter=X seconds after load (default 25), past a node's 20s wait, with a click counter that must read 1. Tests the post-failure XPath recheck.
Whole page stays in a loading state for 20 seconds before rendering content
Rich prior-auth list (200+ DOM nodes). Click '+ Create New Request' — form appears after 60s delay. Tests PSL classification and stale cursor fix.
Shows a 503 Service Unavailable page on first paint, then swaps to a normal dashboard after ?dwell seconds — for testing that a stale outage frame in S1 history is not mis-classified as SERVICE_UNAVAILABLE (?dwell=, ?status=)
UHC-style portal: click Create New Request during table load → Information Loading modal → silent fail. Tests PSL prior-frame classification. (?loadDelay=N seconds, default 6)
Form submit → server-side 302 redirect — instant full page navigation kills CS mid-execution
Noridian pattern: processing overlay → massive DOM rewrite (200+ nodes) triggers Plasmo CS teardown/remount
Configurable redirect loop with various redirect methods
Hover dropdowns, hover cards, and scroll-to-agree modal
Test downloading a dynamically generated PDF file
Request list whose View button opens the request PDF inline in a new tab
Result page with no file export and a site Print button
Main-frame navigations to PDF URLs (link, location.href, redirect, form submit) with inline/none/attachment dispositions
Inline PDF produced by a native form POST; a GET of the same URL returns a different "No data found" PDF
PDF opened via target=_blank, window.open, and popup windows with inline/none/attachment dispositions
PDFs as iframe src (sub_frame) with inline/none/attachment dispositions, plus JS-set src
PDFs via embed and object tags with inline/none/attachment dispositions
Client-side generated blob: and data: PDFs — navigation, new tab, iframe, embed, anchor download
Network-fetched PDF wrapped in a blob (configurable MIME) shown in a 'Progress Note Preview' modal iframe — mirrors Aarista/charta. Tests blob capture and new-tab auto-download vs. the dead download stub
Three distinct PDFs as iframes on one page — tests multi-file capture (upload all)
Single-use-token PDF (proves token-safe tshark capture) and a ~15MB PDF (large reassembly)
Configurable large file download (up to 500MB) for testing file handling
Tests 10 different download methods with a known binary file to diagnose corruption
Controllable download target for network-signal classification (?mode=sync|async, ?status=, ?delay=, ?drip=1&dripMs=, ?size=, ?redirect=). Triggers a real browser download so the network listener captures server-side download failures (404/500/503, slow/drip timeout, redirect-to-login, async export).
Printable certification page with 15 iframes. Use with trigger_print in a tab-churn loop to measure per-iteration degradation (?patient, ?cert_date, ?office, ?provider, ?delay)
Multi-step form with random popups and conditional fields
OASIS-style clinical assessment across 3 tabs
Display page showing tracking status (reached via /shipping)
Login with 2FA, then paginated claims table with lookup and EOB downloads
Tabbed medical portal (Claims, Eligibility, Auth) for testing enrichment classification
Video→workflow testbed: conditional reveal (radio→wound section), custom non-native dropdown, 0..N add-wound loop, format-sensitive date, validation gating, and destructive buttons. Variants: ?wounds=yes|no, ?date=us|iso, ?pain=short|long, ?strict=1
Stable id/class/position on every load. GREEN path — graduated xpath resolves cleanly.
Rotating button id; role/aria stable. YELLOW path — cascade recovers and fingerprint matches.
?variant=a|b swaps Submit and Cancel at the same position. Tests fingerprint-mismatch immediate revert.
?show=false removes the button entirely. Tests cascade-exhausted AI fallback and hard failure.
Stable input id/name/label for input_text graduation. GREEN path for the input_text action type.
Rotating input id/name, stable label/aria/testid. YELLOW path for input_text cascade recovery.
Five Submit buttons reshuffled on every load. Tests the observing/candidate → disqualified path.
?broken=true rotates a graduated button's id without removing the element. Tests post-graduation silent xpath failure → revert.
?broken=true renames the table id so STATIC extract_datamodel misses. With enable_llm_dom_fallback=true on the node and ENABLE_LLM_DOM_FALLBACK=true on the backend, the run rescues via LLM_DOM.
Patient table inside iframe-within-iframe. STATIC only traverses one iframe level; LLM_DOM with recursive serialization extracts correctly.
Claims data rendered on <canvas> (simulating PDF viewer). Zero DOM text nodes — only LLM_VISION can read the rendered pixels.
Medication table inside shadow root within shadow root. XPath cannot cross shadow boundaries; requires LLM_DOM with shadow serialization or LLM_VISION.
Schedule table renders after JS delay (?delay=5000). With ?randomize=true the container ID changes each load. STATIC fires against empty DOM; LLM_DOM captures the settled state.
Benefits data inside a cross-origin iframe. contentDocument is null; LLM_DOM cannot serialize it. Only LLM_VISION (screenshot compositing) works.
200 claims in a virtual-scrolling list — only ~20 rows exist in the DOM at once. Tests partial extraction and scroll+extract patterns.
Puts a known string on the OS clipboard (Copy button, or select the textarea + Ctrl/Cmd+C) so a STATIC EXTRACT_DATAMODEL node sourced from {{clipboard}} reads it back exactly, for native/no-DOM apps (RDP/DCV) where content isn't in the page DOM. Also has a copy-per-row harvest loop (Copy next row + ?delay= to simulate copy-sync lag) for testing extract inside a LOOP. ?variant=simple|multiline|unicode|xml, ?text=custom. Synthetic data (no PHI).
MRN field lives on an unopened Insurance tab; Demographics (name, DOB) is open by default. ?tab=insurance opens it directly. Tests zero-match-from-wrong-tab, not a broken selector.
Record reads correctly, then fully unmounts/remounts under a new node identity after ?churnDelay= (default 1200ms), simulating a background sync. Tests the 'said it didn't but did' misclassified-failure case. ?churn=false disables.
Line items sit inside a closed accordion; opening it loads items in two waves (3 immediately, the rest 800ms later). ?open=true loads pre-expanded and settled. Tests zero/partial-match from lazy render timing.
Fully working line-items table with a Re-extract button simulating a recovery retry; by default each retry appends a duplicate copy into session context. ?dedupe=true replaces instead of appending. Tests array-duplication risk from an un-cleared partial write.
Results table is correct and extractable on load, then the node is removed for an empty state after ?detachDelay= (default 1000ms), simulating a background refetch. Tests a post-extract health-check misreporting an already-successful read as failed. ?detach=false disables.
Fires the same URL (/api/method-echo) with GET/POST/PUT/PATCH/DELETE, each returning a method-tagged payload. Verifies EXTRACT_NETWORK's HTTP-method filter selects the right response when one URL is hit by many methods. ?auto=1 fires all on load; ?methods=GET,POST limits which.
Fires a no-cors fetch to a Cloudflare endpoint that publishes DNS HTTPS records (Chrome reliably picks HTTP/3). Pair with EXTRACT_NETWORK on the same URL to verify whether our tshark-on-cc_tun0 capture stack sees HTTP/3 traffic. The trace response body contains an http= line revealing which protocol Chrome actually used. ?target=<url>, ?iter=N, ?auto=0.
Fires a burst of concurrent requests at an external origin (?origin=, default postman-echo.com — must NOT be *.cloudcruise.com/.app, which cc_tshark_event_mapper.py hard-skips) plus one named marker request, to reproduce the concurrent-HTTP/2-connection shape behind a real EXTRACT_NETWORK decode gap (session 9d72c5f4: 22 TLS handshakes logged, only 4 decoded, none from the busy app domain). ?concurrency=, ?minDelay=/?maxDelay= (client-side stagger), ?bytes=. Also fires a same-origin control burst (?control=0 to skip) that should NEVER be captured, proving the hard-skip filter rather than the decode race. Must be served over TLS/HTTP2 (not the plain-HTTP dev server).
One distinct, fixed-position 'Download Report' button with stable pixels. Click it via LLM_VISION: first run grounds + caches the crop, a repeat run re-locates it with no VLM (DIRECT). ?theme=dark flips colors but keeps structure — Sobel matching should still hit.
Two visually-identical blank input fields. A DIRECT crop cross-matches both, so the password field binds to its 'Password' label anchor + offset and verifies landing before clicking. ?filled=true prefills email to prove the label anchor stays stable.
A row of N identical 'Select' buttons (?count=N, default 5). None is visually unique, so the matcher's top-2 margin is ~0 and it must abstain and defer to the VLM rather than guess (precision over recall).
Same fixed slot, different element by query: ?state=a blue 'Submit' (default), ?state=b green 'Cancel', ?show=false nothing. Cache on a, re-run on b — the cached crop should fail structural match and abstain, not mis-click.
Four-step sequential flow (Start → Add Details → Review → Submit), each step revealing the next distinct LLM_VISION target. Chains four vision clicks into four independent cache entries; a repeat run should produce four consecutive L1 hits.
Eval substrate for the dynamic-intent guard. ?case=<id> renders one realistic widget with a single target (data-testid="target"). 15 dynamic targets that must NOT cache (today's date, latest order, default card, primary contact, recommended plan, cheapest/oldest, next-available slot, selected/unread/new/your-task/pending, top result, highest bid) + 6 fixed controls whose labels merely contain a relative word and MUST stay cacheable (Today button, Sort-Newest toggle, Default-View tab, Primary tab, New button, Featured filter).
Five-step click-through for the prompt-only dynamic classifier + cache persistence: 2 static targets (Open records, Submit) that build then hit, 2 dynamic targets that reshuffle every load (the Globex Inc. account row and the highlighted tile) that must log skip_dynamic and re-ground, plus 1 tricky-static control ('Sort by Newest') that must stay cacheable. A red banner flags a dynamic target that got cached and mis-fired.
Fixture for desktop readiness (act as soon as the screen is ready, not after a fixed delay). Every transition keeps the old screen under a veil + spinner for ?delay seconds (+ ?jitter); names, dates and rows are random per load. ?flow=wizard: plain transitions, the same Next button on consecutive screens, back-to-back OK dialogs with different titles, a Save button that turns grey to blue, a no-op click and an invisible copy. ?flow=grid: type then click an untouched Close, a Search that closes its panel at once while the grid refills later (two speeds, stale first row), a sort arrow flipping inside a per-run table. ?flow=popup: a popup in two random layouts (the button moves), an optional Acknowledge step before Review / Reprint (branch), three items in a row (loop). ?clock=1 adds a live clock. ?variants=3 draws the content from one of 3 fixed datasets per load, so masks learned across runs converge.
Dense CareTend replica of the New Patient / Prescriber workflow's main path for desktop readiness: Workflows flyout, ePrescription Manager (Service Not Available, blue spinner, Filter Search, column filter popups with a blank option and an 'Is equal to' operator), View Message and View XML (Loading Message from Hub), Select Prescriber with a New prescriber form (Prof. Desig. Add New, New Value and Duplicate Descriptions popups, Companies), Select Patient, the patient record and its Demographics / Insurance / Medical Info / Diagnosis tabs. Every screen loads behind a veil like the real one; all data is fake and random per load. ?speed= scales loading, ?jitter= adds randomness, ?start= opens one screen directly.
Pixel-identical button whose onClick is rewired to a destructive action on run 2 (?variant=build|attack). Cache is purely visual — predicted FALSE HIT (confident wrong click).
Safe button cached at a slot; on run 2 a pixel-identical destructive decoy takes that slot while the real button moves out of search_radius=240 (?variant=build|attack). Predicted FALSE HIT.
Approve $50 → $5000: meaning flips but Sobel structure stays >accept_score=0.85 (?variant=build|attack, ?digits=long stresses the boundary). Predicted FALSE HIT (long → safe abstain).
Featureless field anchored on a distant heading; run 2 inserts an identical field between them so heading+offset drifts onto the twin and loose verify passes (?variant=build|attack). FALSE HIT on the anchored path (Obs 15).
Forces an L1 miss with a transient decoy present so the unconditional build() overwrites the good entry with bad pixels/coords (?variant=build|miss-poison|attack). Predicted durable POISON → FALSE HIT.
DIRECT locate uses scales=[1.0] only; run 2 enlarges the real target out of band and plants a same-1.0x decoy in-window (?variant=build|attack, ?zoom=, ?decoy=off). FALSE HIT (decoy on) / SAFE ABSTAIN (decoy off).
No per-field labels, so the featureless box is forced to anchor on a distant 'Billing' heading that can't co-move; run 2 inserts an identical box between them so heading+offset drifts onto the twin and loose verify passes (?variant=build|attack). Predicted FALSE HIT — the genuine anchored-drift trigger #4 missed.
Structure-preserving relabel on a fixed-geometry, same-color button: Confirm→Cancel or Pay $50→£50 (?variant=build|attack, ?case=confirm-cancel|currency). Sobel stays >accept_score=0.85 so the cache HITS the meaning-flipped button — the case #3 only abstained on by luck. Predicted FALSE HIT; motivates an OCR/CLIP semantic veto.
Target invoice row sits below the fold → a grounding/CLICK node must scroll the container until it's visible, then act. ?scroll=inner|page, ?count=N (vary target depth between build/replay), ?target=N, ?present=false (remove target → end-of-scroll FAIL). OCR-friendly unique row ids for the cold OCR-gated pass.
Icon-only cold-path weak spot: a distinctive star badge below the fold among repeated low-distinctiveness filler glyphs. No text → cold OCR pre-filter can't help; warm pass relocates the cached glyph. ?scroll=inner|page, ?count=N, ?target=N, ?present=false.
Target lives in a scrollable side panel (not the page), off-center so it doesn't sit under the agent's fixed blind-scroll cursor — tests whether scroll-to-target works for paneled UIs. ?side=left|right, ?count=N, ?target=N, ?present=false.
People names (non-sequential, fixed array) in a small centered overflow-y-auto box — removes the ordinal cue invoice numbers gave the grounder, to test whether it can scroll to a target it can't locate by number. ?count=N, ?target=N, ?present=false.
Names in a WIDE fixed-height scrollable container that spans past both observed scroll aim points (x~495, x~960) so the wheel reliably hits it — isolates the numbers-vs-names scroll-inference question without the small-box physical-miss confound. ?count=N, ?target=N, ?present=false.
NextGen-style modal: form fields + OK/Cancel/Clear/Search and a SMALL scrollable results grid (own scrollbar) in the lower portion, target location row below its fold. Reproduces the customer pattern where the wheel must land inside a small off-center sub-region. ?offset=left|center|right, ?count=N, ?target=N, ?present=false.
Tight desktop-style data grid where every row shares a long common prefix (Capitol Pain Institute-Corporate / -Dixie / -Elizabethtown) and the click target is the NAME cell — probes whether the bbox-from-click template crop bleeds adjacent rows and whether a prefix-dominated crop is distinctive enough to relocate the right row. ?dataset=locations|names, ?target=N, ?count=N, ?gap=N (row padding px).
A deterministic grid of uniquely-labelled tiles (Amber Falcon, Cobalt Otter, ...) for exercising the L1 visual-cache PRELOAD at scale — a workflow clicks N distinct tiles so the run relies on N cache entries. Clicking only adds a 'done' highlight, so the layout never shifts and cached crops keep matching across cold/warm runs. ?count=N (default 30), ?cols=N (default 6).
A CareTend-style 'Select Generic Equivalent' grid whose column headers each carry a tiny, closely-spaced filter-funnel icon (plus a small dropdown ▼ and stacked ▲/▼ number spinners) — the hostile targets for high-precision visual grounding. Every control writes a verifiable data-testid='last-action' readout (and funnels open a labeled data-testid='filter-popup') so a workflow can confirm WHICH element was clicked. ?density=compact|normal, ?cols=5|7|9 (funnel spacing), ?dupes=on|off (repeated row names), ?scale=0.75|1|1.25 (resolution/DPI).
The expected #submit-btn control is absent by default so a workflow selector matches nothing (produces XPATH_INCORRECT). ?present=true renders it (control/success case). Used by the unified-eval e2e harness.
Clear 404 UI, returns HTTP 404 by default; ?soft=true for a soft 404 (HTTP 200 + 404 content). Produces PAGE_NOT_FOUND.
Results table is empty by default so a workflow expecting rows sees 'No results' (produces EMPTY_STATE). ?rows=N populates the table.
503 outage page, returns HTTP 503 by default; ?status=200 serves the same content with 200. Produces SERVICE_UNAVAILABLE.
Fifteen-module legacy EHR on WebForms-style postbacks, AngularJS 1.8, Knockout 3.5, Ext-style grids, a real HTML 4 frameset, and jQuery + Bootstrap 3. The launcher lists every module plus 20 recovery scenarios as deterministic ?fault= links.
Five-step WebForms wizard over 37 fields with per-step server validation; partially entered steps persist in ViewState rather than the DOM.
Provider day view of 120 near-identical Available slots, each its own __doPostBack target. Dense ambiguous-target grid for grounding tests.
Expand/collapse result rows with nested analyte tables, H/L/critical flags, select-all, and a critical-value acknowledgement gate.
Debounced medication typeahead rendered outside the input, live sig preview, and a static-backdrop interaction alert requiring a checkbox plus a typed override reason.
Knockout 3.5 observable service lines with computed line and encounter totals. foreach-generated rows carry no stable ids.
Document queue with an inline PDF preview iframe, a real file input, and a scanner simulation driving a multi-step progress modal.
Real HTML 4.01 <frameset>: folders, list, and preview are three sibling documents navigating each other via parent.frames. Not iframes.
ASP.NET WebForms simulation: __VIEWSTATE, __EVENTVALIDATION, ctl00$cphMain$ control names, full-page postbacks. Sign in as clinician / meridian2026.
240-record GridView with __doPostBack sorting and paging; search criteria live in ViewState rather than the URL.
AngularJS 1.8 chart with ng-repeat lists, late ng-if mounting, ng-disabled gating, and a scope-bound modal. ?fault=ghost discards input typed before the model binds.
Ext-style buffered grid in an iframe: only visible rows exist in the DOM, ids regenerate per render pass, and combobox options render in a body-level floating layer.
jQuery 1.9 + Bootstrap 3 note editor with a contenteditable iframe body and a sign-off modal gated on an attestation checkbox.
CMS-1500 style form post with server-side validation, post/redirect/get, and an acknowledgement page. Main target for the mid-form and silent-no-op faults.
Renders success-flag once its own change handler sees the required text in full; ?suppress=flag withholds the flag on a match. Read by e2e/cases/actions/input-text/basic.
Renders success-flag once its own change handler sees the required option committed; ?suppress=flag withholds the flag on a match. Read by e2e/cases/actions/input-select/basic.
Target sits four viewport heights down; success-flag is gated on an IntersectionObserver, not on a scroll event. ?suppress=flag withholds the flag. Read by e2e/cases/actions/scroll/basic.
Reads the bytes it received and renders success-flag only if name and contents match public/e2e/upload-fixture.txt; ?suppress=flag withholds the flag. Read by e2e/cases/actions/file-upload/basic.
Three pages of two fixed rows; Next page wraps back to the first. No success-flag — the verdict is the accumulated dataset. Read by e2e/cases/actions/loop/pagination.
Uncontrolled field and dropdown starting on values the case never uses, so live state differs from the initial markup. No success-flag — reading raw state is the feature under test. Read by e2e/cases/actions/extract-datamodel/raw-state.
Mints a session cookie only for the exact credential the vault provisions; anything else comes back as ?error=invalid with no session. Read by e2e/cases/flows/authenticated-extraction.
Cookie-gated claims table, three server-rendered pages of two fixed rows that wrap, plus a statement download. Redirects to the login when the session is missing, so unauthenticated extraction finds nothing. Read by e2e/cases/flows/authenticated-extraction.
Cookie-gated detail view the case opens in a second tab with TAB_MANAGEMENT. Counts its own acknowledge clicks, so a node that runs twice or runs against the opener's DOM shows up in the extracted data. ?claim=CLM-4104 picks the claim; ?ack_step=2 makes each click count twice, to show the assertion failing. Read by e2e/cases/flows/authenticated-extraction.
Search-as-you-type API where every request carries a sig query param computed as sha256(query string) — a pure function of the request, no external secret. Fixture for computed-request-signature detection (ADR-0021 retrospective finding 5).
Same as above, but sig = sha256(SALT + query string), where SALT is a constant hardcoded in a served .js file rather than derivable from the request alone. Fixture for the salted computed-signature case.
Record the specialty a patient is referred to and save.
Choose the payer a provider is enrolling with and save.
Assign the attending provider for an admission, grouped by clinic.
Pick the state and county where a service was rendered.
Search the formulary and pick a product to order.
Record the reason for today's visit.
Update a patient's state of residence.
Choose the kind of appointment to book.
Confirm coverage type and plan network on file.
Set contact methods, preferred language and reminder consent.
Set the reminder channel, preferred language and reminder consent.
Part of the CloudCruise browser automation platform