Deterministic Browser Harness
Each page starts in a known state, exposes a visible task contract, and reveals a persistent completion code only after true success. The suite is tuned for debugging where an agent gets stuck, not for simulating a full product flow.
Field targeting, typing, clicking, simple submit flows, and explicit success checks.
Focused pages that isolate a single visible control or interaction pattern.
form-single-input-001
One text field, one correct value, one success path.
Enter `AURORA-17` in the Access code field, then submit the form.
form-multi-field-002
Three required fields with inline validation and one optional decoy field.
Fill First name with `Nora`, Last name with `Stone`, Email with `nora.stone@gettandem.app`, then submit.
dropdown-basic-003
A native select with decoy controls nearby and one valid option.
Choose `Orion Ops` from the Team assignment dropdown, then submit.
radio-basic-004
A simple radio group with similar-looking labels.
Select the deployment window labeled `06:30 PM`, then submit.
checkbox-basic-005
Two required checkboxes plus a select-all trap and a disabled decoy.
Check only `I have reviewed the release notes` and `I confirm the staging snapshot is current`, then submit.
button-target-006
A primary action surrounded by similar-sized decoy buttons.
Press the button labeled `Finalize Snapshot`.
enter-submit-007
A keyboard-only submission path with no clickable submit button.
Type `delta search ready` in the search field and submit with Enter.
textarea-basic-008
A multiline input task that checks textarea targeting and formatting.
Enter the exact three-line note shown in the reference block into the textarea, then submit.
slider-drag-019
A single slider that must be moved to an exact numeric value before submit.
Drag the slider until the current value is `72`, then submit.
autocomplete-combobox-020
A filtered suggestion list where typed text alone does not complete the task.
Type `or`, select `Orion Chen` from the suggestions, then submit.
file-upload-022
A native file input that accepts any local file and requires visible filename confirmation.
Upload any local file, confirm the filename appears, then submit.
Scroll management, wizards, modals, accordions, tabs, and dense row actions.
Layout-heavy tasks where the agent has to navigate sections, overlays, or repeated UI.
scroll-long-form-009
A deliberately tall page that separates the final submit area from the first fields.
Fill Project name with `Northwind`, Owner with `Ava`, Rollout zone with `Central`, then scroll to the bottom and submit.
wizard-steps-010
A three-step flow where only the last step reveals success.
Step 1: enter `BETA-9`. Step 2: choose `Manual Review`. Step 3: check the final confirmation and submit.
modal-confirm-011
A centered overlay that requires the agent to ignore the background page.
Open the modal, type `CONFIRM` inside it, then press `Confirm action`.
accordion-reveal-012
A hidden input flow behind one of several similar accordion sections.
Expand `Credential Rotation`, enter `rotate-88`, then submit the revealed form.
tabs-target-013
A tabbed settings shell with one actionable form inside the correct tab.
Switch to the `Notifications` tab, choose `Hourly`, then save the preference.
table-row-action-014
A dense table with repeated row actions and only one valid target row.
Find the row for `Helio Labs` and press its `Mark reviewed` button.
date-picker-021
A popover calendar that starts one month before the required date.
Open the date picker, move to April 2026, select `April 22, 2026`, then save.
filtered-table-action-023
A table that must be filtered before the correct repeated row action can succeed.
Search for `Mira`, then click `Approve access` for the `Mira Vault` row.
drag-drop-reorder-024
A reorderable list where the final item order must be verified before submit.
Drag `Compliance review` so it sits directly above `Release notes`, then submit.
Error handling, prerequisite gating, loading patience, and state re-checking.
Tasks that intentionally expose waiting, validation, and delayed state transitions.
validation-recovery-015
A recovery-specific task that requires noticing and correcting a validation failure.
Press submit once with the field empty to surface the validation error. Then enter `TANDEM-VALID` and submit again.
delayed-enablement-016
A prerequisite gate where the final button becomes enabled after a short delay.
Check `I understand this is a simulated dry run`, wait for the continue button to enable, then click continue.
async-spinner-017
A two-phase async interaction with a visible loading state and post-load action.
Press `Start analysis`, wait for loading to finish, then click `Publish result`.
Repeated modules, nearby decoys, full-label reading, and verifying the right target.
Pages that force the agent to disambiguate repeated controls and local context.
repeated-ui-018
Three nearly identical cards, each with the same action button.
In the card labeled `North Annex`, press `Arm sensor`.