tasks/v2/04_autocomplete_search_multi_combobox_v2.yaml

- id: autocomplete_freeform-antd-v2-T01
  name: 'Drawer labels: add one suggested tag and one custom tag'
  canonical_type: autocomplete_freeform
  implementation_source: antd
  implementation_variant: Select.tags (autoClearSearchValue=false)
  implementation_component: 'AntD: Select (mode="tags", autoClearSearchValue=false) inside Drawer'
  task_template: select_many
  secondary_template: confirm_cancel
  browsergym_goal: Open the “Edit labels” drawer. In the “Incident labels” field, keep the existing tag `backend`, add the suggested tag `urgent`, add the custom tag `sev-zero`, and end with exactly those three tags. Then click “Apply labels”. The task will finish automatically when done.
  ui_copy: Open the “Edit labels” drawer. In the “Incident labels” field, keep the existing tag `backend`, add the suggested tag `urgent`, add the custom tag `sev-zero`, and end with exactly those three tags. Then click “Apply labels”. The task will finish automatically when done.
  setup_description: |-
    The page uses a drawer_flow layout with compact spacing and medium clutter. A release triage dashboard remains visible in the background with charts, counters, and a recent incidents list. Clicking “Edit labels” opens a right-side drawer. Inside the drawer are two visually similar tag inputs: “Incident labels” and “Archive labels”. Both are Ant Design Select controls in tags mode with search, and both render selected values as removable chips.

    Initial state:
    - Incident labels: backend
    - Archive labels: legacy

    For the Incident labels field, dropdown suggestions include urgent, customer-visible, infra, backend. The value `sev-zero` is not in the suggestions and must be committed as a custom tag. Search text is not cleared automatically after each selection, so residual input can remain visible until it is explicitly committed or removed. The drawer footer contains “Cancel” and “Apply labels”. Changes are committed only when “Apply labels” is clicked.
  scene_context:
    theme: light
    spacing: compact
    layout: drawer_flow
    placement: off_center
    scale: default
    instances: 2
    guidance: text
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: drawer
  design_intent:
    active_factors:
    - instances
    - overlay_model
    - confirmation_model
    factor_rationale: Two nearly identical editable tag fields inside a drawer create wrong-instance pressure, and the residual search buffer means the agent must distinguish between typed-but-uncommitted text and actual chips before applying the change.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 3
      density_choice_interference: 3
      depth_layering: 3
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 5
    justification: This is hard because the agent must edit the correct tags field, preserve an existing chip, add one option-backed value and one custom value, audit the final chip set, and then commit via the drawer-local action.
  success_trigger:
    human_readable:
    - The committed value of “Incident labels” is exactly {backend, urgent, sev-zero}.
    - “Archive labels” remains exactly {legacy}.
    - The drawer-local “Apply labels” control has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        target_instance: Incident labels
        values:
        - backend
        - urgent
        - sev-zero
        non_target_instances_must_remain:
          Archive labels:
          - legacy
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Apply labels
      require_correct_instance: true
      target_instance_label_or_id: Incident labels
      terminal_condition: true
  negative_cases:
  - Adding the right text in the input but leaving it uncommitted as residual search text.
  - Editing “Archive labels” instead of “Incident labels”.
  - Keeping any extra chip such as customer-visible or infra.
  - Stopping with the right chips visible but without clicking “Apply labels”.
  expected_interaction_path:
  - Open the drawer.
  - Focus the Incident labels field, not Archive labels.
  - Use the dropdown to add urgent, then commit the custom value sev-zero as a chip.
  - Verify the final Incident labels chip set is exactly backend, urgent, sev-zero.
  - Click Apply labels.
  notes: Track both the chip array and any live search buffer so uncommitted text does not count as a final value.
- id: autocomplete_freeform-antd-v2-T02
  name: 'Settings panel alias: suggestion selection on the second field only'
  canonical_type: autocomplete_freeform
  implementation_source: antd
  implementation_variant: AutoComplete with custom option labels
  implementation_component: 'AntD: AutoComplete with grouped custom labels in settings panel'
  task_template: search_and_select
  secondary_template: confirm_cancel
  browsergym_goal: In the routing settings panel, set only the “Escalation alias” field to `plat-eu — Platform Europe` by selecting it from the suggestions list. Do not change “Owner alias”. Then click “Save routing”. The task will finish automatically when done.
  ui_copy: In the routing settings panel, set only the “Escalation alias” field to `plat-eu — Platform Europe` by selecting it from the suggestions list. Do not change “Owner alias”. Then click “Save routing”. The task will finish automatically when done.
  setup_description: |-
    Layout is settings_panel with compact spacing, off-center placement, and high clutter. Above the target area are unrelated toggles, segmented controls, and two compact selects. The target area contains two Ant Design AutoComplete inputs with the same custom option renderer:
    - Owner alias
    - Escalation alias

    Both inputs open a suggestion popup with aliases rendered as `shortcode — team name`, including:
    - plat-us — Platform US
    - plat-eu — Platform Europe
    - plat-ops — Platform Operations
    - platform — Shared Platform

    Initial state:
    - Owner alias: platform — Shared Platform
    - Escalation alias: (empty)

    The control allows arbitrary typing, but the benchmark checker distinguishes free text from a committed suggestion selection. A panel-local button labeled “Save routing” commits changes.
  scene_context:
    theme: light
    spacing: compact
    layout: settings_panel
    placement: off_center
    scale: default
    instances: 2
    guidance: text
    clutter: high
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - instances
    - clutter
    - confirmation_model
    factor_rationale: The old family often collapsed to direct typing, so this task forces suggestion commitment on one of two similar fields inside a dense settings panel and makes the untouched sibling part of the success boundary.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 4
      density_choice_interference: 4
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 4
      disambiguation_load: 5
    justification: The alias itself is short, but the field is one of two similar freeform autocompletes in a cluttered panel and success requires a real suggestion commit plus preservation of the non-target field.
  success_trigger:
    human_readable:
    - The committed value of “Escalation alias” is exactly `plat-eu — Platform Europe` and its selection origin is a suggestion commit.
    - “Owner alias” remains `platform — Shared Platform`.
    - “Save routing” has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: Escalation alias
        value: plat-eu — Platform Europe
        selection_origin: suggestion
        non_target_instances_must_remain:
          Owner alias: platform — Shared Platform
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save routing
      require_correct_instance: true
      target_instance_label_or_id: Escalation alias
      terminal_condition: true
  negative_cases:
  - Typing the text into Escalation alias without committing a suggestion.
  - Changing Owner alias as well as Escalation alias.
  - Selecting plat-us — Platform US or plat-ops — Platform Operations.
  - Leaving the panel unsaved.
  expected_interaction_path:
  - Focus the Escalation alias field.
  - Type a prefix such as plat to reveal suggestions.
  - Choose plat-eu — Platform Europe from the popup.
  - Leave Owner alias unchanged.
  - Click Save routing.
  notes: Expose whether the final value came from suggestion selection or raw free text.
- id: autocomplete_freeform-antd-v2-T03
  name: Table row city binding with diacritic near-misses
  canonical_type: autocomplete_freeform
  implementation_source: antd
  implementation_variant: AutoComplete inside editable table row
  implementation_component: 'AntD: AutoComplete inside table cell with row-local save'
  task_template: table_operation
  secondary_template: search_and_select
  browsergym_goal: In the deployment table, find the row “Service B”. In that row only, set “Rollout city” to `San José (CR)` by choosing it from the suggestion list, then click the “Save” action for Service B. The task will finish automatically when done.
  ui_copy: In the deployment table, find the row “Service B”. In that row only, set “Rollout city” to `San José (CR)` by choosing it from the suggestion list, then click the “Save” action for Service B. The task will finish automatically when done.
  setup_description: |-
    The page uses a table_cell layout with compact spacing, small scale, and medium clutter. A deployment table shows three rows: Service A, Service B, and Service C. The “Rollout city” column is editable for Service B and Service C, each with an Ant Design AutoComplete input and a row-local “Save” button.

    The suggestion list contains several confusable entries:
    - San Jose
    - San José
    - San José (CR)
    - Santo Domingo
    - São Paulo

    Initial state:
    - Service B Rollout city: (empty)
    - Service C Rollout city: São Paulo

    The input accepts arbitrary text, but only a suggestion-origin selection counts for this task. Saving is scoped to the row where the edit was made.
  scene_context:
    theme: light
    spacing: compact
    layout: table_cell
    placement: bottom_left
    scale: small
    instances: 2
    guidance: text
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: table
  design_intent:
    active_factors:
    - layout
    - instances
    - scale
    factor_rationale: A dense table row with two editable autocomplete instances makes row binding and small-target acquisition primary, while the city list uses accent and suffix near-misses to keep direct typing from being the whole problem.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 5
      density_choice_interference: 4
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 4
      disambiguation_load: 5
    justification: 'This is a row-scoped high-pressure task: the target row is dense, the editable cells are small, the options are confusable, and the final row-local Save must be applied to the correct row.'
  success_trigger:
    human_readable:
    - Service B “Rollout city” equals `San José (CR)` with suggestion-origin selection.
    - Service C remains `São Paulo`.
    - The Service B row Save action has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: Service B
        field: Rollout city
        value: San José (CR)
        selection_origin: suggestion
        non_target_instances_must_remain:
          Service C__Rollout city: São Paulo
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: save-service-b
      require_correct_instance: true
      target_instance_label_or_id: Service B
      terminal_condition: true
  negative_cases:
  - Typing San José (CR) as raw text without suggestion commit.
  - Saving the Service C row instead of Service B.
  - Selecting San Jose or San José instead of San José (CR).
  - Changing Service C while editing Service B.
  expected_interaction_path:
  - Locate the Service B row.
  - Open the Rollout city suggestions for Service B and search with San.
  - Select San José (CR).
  - Click Save on the Service B row.
  notes: The checker should bind to a row identifier and record both the committed city and selection origin.
- id: autocomplete_freeform-antd-v2-T04
  name: Reference-card employee match in modal invite flow
  canonical_type: autocomplete_freeform
  implementation_source: antd
  implementation_variant: AutoComplete with custom option render in Modal
  implementation_component: 'AntD: AutoComplete with avatar options inside Modal'
  task_template: match_reference
  secondary_template: confirm_cancel
  browsergym_goal: Click “Add reviewer”. In the modal, set the “Reviewer” field to match the person shown in the reference card and choose that person from the suggestions list. Then click “Invite reviewer”. The task will finish automatically when done.
  ui_copy: Click “Add reviewer”. In the modal, set the “Reviewer” field to match the person shown in the reference card and choose that person from the suggestions list. Then click “Invite reviewer”. The task will finish automatically when done.
  setup_description: |-
    Layout is modal_flow with compact spacing and medium clutter. A review dashboard in the background shows a reference card labeled “Next reviewer”, including a small avatar, the display name “Olivia Chen”, and the team label “Platform”. Clicking “Add reviewer” opens a modal with one target AutoComplete labeled “Reviewer”.

    The modal’s suggestion list uses custom option rendering with avatar + name + team, including:
    - Olivia Chen — Platform
    - Olivia Cheng — Platform
    - Oliver Chen — Platform
    - Olivia Chen — Design

    The input accepts arbitrary text, but only a suggestion-origin selection counts. The modal footer contains “Cancel” and “Invite reviewer”.
  scene_context:
    theme: light
    spacing: compact
    layout: modal_flow
    placement: center
    scale: default
    instances: 1
    guidance: mixed
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: false
    confirmation_model: explicit_confirm
    precision_surface: mixed
    secondary_supporting_primitive: dialog
  design_intent:
    active_factors:
    - guidance
    - overlay_model
    - confirmation_model
    factor_rationale: The target is grounded from an on-page reference card rather than copied from the prompt, and the nested modal-plus-suggestions flow ensures the task measures both visual matching and suggestion commitment.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 3
      density_choice_interference: 4
      depth_layering: 4
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 3
    justification: The task is short but genuinely hard because several near-duplicate people are present, the correct answer is specified by a reference card, and the modal commit is explicit.
  success_trigger:
    human_readable:
    - The committed Reviewer value is `Olivia Chen — Platform` and it was selected from the suggestion list.
    - The modal “Invite reviewer” action has been confirmed.
    canonical_predicate:
      predicate_type: matches_reference
      target_state:
        reference_id: next-reviewer-card
        value: Olivia Chen — Platform
        selection_origin: suggestion
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Invite reviewer
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Selecting Olivia Cheng — Platform, Oliver Chen — Platform, or Olivia Chen — Design.
  - Typing “Olivia Chen” without choosing a suggestion.
  - Closing the modal or clicking Cancel instead of Invite reviewer.
  - Ignoring the team label shown in the reference card.
  expected_interaction_path:
  - Read the name and team on the reference card.
  - Open the modal.
  - Type Olivia to filter suggestions and choose Olivia Chen — Platform.
  - Click Invite reviewer.
  notes: Use the reference card text to derive the expected answer, but still require suggestion-origin selection in the checker.
- id: autocomplete_freeform-mui-v2-T01
  name: Collapsed chips in dark panel with explicit save
  canonical_type: autocomplete_freeform
  implementation_source: mui
  implementation_variant: Autocomplete (multiple, freeSolo, limitTags=1, filterSelectedOptions)
  implementation_component: 'MUI: Autocomplete (multiple, freeSolo, limitTags=1) in dashboard panel'
  task_template: select_many
  secondary_template: confirm_cancel
  browsergym_goal: In the dark “Ticket labels” panel, edit only the “Labels” field so the selected chips are exactly `backend`, `urgent`, and `p0`. The field starts with `backend`. Then click “Save labels”. The task will finish automatically when done.
  ui_copy: In the dark “Ticket labels” panel, edit only the “Labels” field so the selected chips are exactly `backend`, `urgent`, and `p0`. The field starts with `backend`. Then click “Save labels”. The task will finish automatically when done.
  setup_description: |-
    Layout is dashboard_panel with dark theme, compact spacing, off-center placement, and medium clutter. The target panel contains two MUI Autocomplete controls rendered as chip inputs:
    - Labels (target)
    - Watch tags (distractor)

    The Labels field is configured with multiple, freeSolo, filterSelectedOptions, and limitTags=1. When the input is not focused, only one visible chip and a “+N” counter are shown. Initial state:
    - Labels: backend
    - Watch tags: triage

    Suggestions include backend, urgent, p0, p1, frontend, qa. The panel-local action “Save labels” commits the edited state.
  scene_context:
    theme: dark
    spacing: compact
    layout: dashboard_panel
    placement: off_center
    scale: default
    instances: 2
    guidance: text
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - theme
    - instances
    - confirmation_model
    factor_rationale: Collapsed chip display makes state auditing harder, and the sibling chip field plus explicit save step prevent the task from collapsing to “type two strings and leave.”
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 3
      density_choice_interference: 3
      depth_layering: 2
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 5
    justification: A hidden +N chip counter, a dark compact panel, and a nearby distractor chip field make this an exact-set state-tracking task rather than a simple free-solo input.
  success_trigger:
    human_readable:
    - The committed Labels value is exactly {backend, urgent, p0}.
    - Watch tags remains exactly {triage}.
    - Save labels has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        target_instance: Labels
        values:
        - backend
        - urgent
        - p0
        non_target_instances_must_remain:
          Watch tags:
          - triage
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save labels
      require_correct_instance: true
      target_instance_label_or_id: Labels
      terminal_condition: true
  negative_cases:
  - Stopping with the right chips visible in the draft state but not clicking Save labels.
  - Adding p1 or frontend as an extra chip.
  - Editing Watch tags instead of Labels.
  - Leaving a typed but uncommitted input buffer after adding chips.
  expected_interaction_path:
  - Focus the Labels field so hidden chips expand if needed.
  - Add urgent and p0 while keeping backend.
  - Verify Watch tags still contains only triage.
  - Click Save labels.
  notes: Record the controlled chip array at commit time; visual chip collapse should not affect correctness.
- id: autocomplete_freeform-mui-v2-T02
  name: Backup alias only with grouped suggestion headings
  canonical_type: autocomplete_freeform
  implementation_source: mui
  implementation_variant: Autocomplete (freeSolo, groupBy, openOnFocus, autoHighlight)
  implementation_component: 'MUI: Autocomplete (freeSolo, grouped options) in settings panel'
  task_template: search_and_select
  secondary_template: confirm_cancel
  browsergym_goal: In the alias settings panel, set only the “Backup alias” field to `alpha-admin` by selecting it from the grouped suggestion list. Do not change “Primary alias”. Then click “Apply aliases”. The task will finish automatically when done.
  ui_copy: In the alias settings panel, set only the “Backup alias” field to `alpha-admin` by selecting it from the grouped suggestion list. Do not change “Primary alias”. Then click “Apply aliases”. The task will finish automatically when done.
  setup_description: |-
    The page uses a settings_panel layout with compact spacing, top-left placement, and high clutter. The target card contains two single-value MUI Autocomplete fields:
    - Primary alias
    - Backup alias

    Both use freeSolo with grouped options and openOnFocus. The options are grouped under headings “Admins”, “Analysts”, and “Automation”, and include:
    - alpha-admin
    - alpha-analyst
    - alpha-auto
    - beta-admin

    autoHighlight is enabled, so the first visible option becomes highlighted when the popup opens. Initial state:
    - Primary alias: beta-admin
    - Backup alias: (empty)

    The panel action “Apply aliases” commits both fields.
  scene_context:
    theme: light
    spacing: compact
    layout: settings_panel
    placement: top_left
    scale: default
    instances: 2
    guidance: text
    clutter: high
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - instances
    - clutter
    - guidance
    factor_rationale: Grouped headings and auto-highlight add real interaction-knowledge pressure because clicking the field is not enough; the agent must target the second alias field and commit the intended grouped option rather than the first highlighted near-match.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 4
      density_choice_interference: 4
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 4
      disambiguation_load: 5
    justification: Wrong-field errors and wrong-highlight errors are both plausible here, and the explicit Apply step means the agent has to verify the final state before committing.
  success_trigger:
    human_readable:
    - Backup alias equals `alpha-admin` and the value came from the grouped suggestion list.
    - Primary alias remains `beta-admin`.
    - Apply aliases has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: Backup alias
        value: alpha-admin
        selection_origin: suggestion
        non_target_instances_must_remain:
          Primary alias: beta-admin
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Apply aliases
      require_correct_instance: true
      target_instance_label_or_id: Backup alias
      terminal_condition: true
  negative_cases:
  - Accepting the first highlighted suggestion in the wrong group.
  - Typing alpha-admin as raw input without suggestion commit.
  - Changing Primary alias.
  - Leaving the panel unsaved.
  expected_interaction_path:
  - Focus Backup alias.
  - Use the grouped popup to find and select alpha-admin.
  - Confirm Primary alias is still beta-admin.
  - Click Apply aliases.
  notes: The checker should distinguish suggestion-origin values from freeSolo raw text.
- id: autocomplete_freeform-mui-v2-T03
  name: Async office picker in modal with explicit save
  canonical_type: autocomplete_freeform
  implementation_source: mui
  implementation_variant: Autocomplete (freeSolo, async search-as-you-type)
  implementation_component: 'MUI: Autocomplete with async options inside Modal'
  task_template: search_and_select
  secondary_template: confirm_cancel
  browsergym_goal: Click “Add office”. In the modal, use the “Office” autocomplete to select `San Sebastián Office` from the async suggestions list, then click “Save office”. The task will finish automatically when done.
  ui_copy: Click “Add office”. In the modal, use the “Office” autocomplete to select `San Sebastián Office` from the async suggestions list, then click “Save office”. The task will finish automatically when done.
  setup_description: |-
    Layout is modal_flow with compact spacing, centered placement, and medium clutter. A button labeled “Add office” opens a modal containing one MUI Autocomplete labeled “Office”. The control is configured for async search-as-you-type: a loading indicator appears while options are being fetched, and the visible option list updates after each query. Built-in filtering is disabled so the server-returned list is used as-is.

    The async suggestions for `San` include:
    - San Diego Office
    - San José Office
    - San Sebastián Office
    - Santiago Office

    The input permits arbitrary text, but only an async suggestion selection counts. The modal action “Save office” commits the selection.
  scene_context:
    theme: light
    spacing: compact
    layout: modal_flow
    placement: center
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: false
    confirmation_model: explicit_confirm
    precision_surface: mixed
    secondary_supporting_primitive: dialog
  design_intent:
    active_factors:
    - overlay_model
    - confirmation_model
    - clutter
    factor_rationale: The option list appears behind a modal boundary and is intentionally asynchronous, so the task pressures the “wait, re-ground, then commit” loop that plain static autocompletes often miss.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 3
      density_choice_interference: 4
      depth_layering: 4
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 2
    justification: The target text is readable, but the async update and explicit save boundary make this a real state-verification task instead of an instant fill.
  success_trigger:
    human_readable:
    - The committed Office value is `San Sebastián Office` and its selection origin is an async suggestion.
    - The modal Save office action has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        value: San Sebastián Office
        selection_origin: suggestion
        options_source: async
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save office
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Typing the target text and saving without selecting a suggestion.
  - Choosing San Diego Office, San José Office, or Santiago Office.
  - Closing the modal or canceling instead of saving.
  - Trying to select before the async suggestions have loaded and ending with no committed value.
  expected_interaction_path:
  - Open the modal.
  - Type San into the Office field and wait for the async list to populate.
  - Select San Sebastián Office from the returned options.
  - Click Save office.
  notes: Expose loading state and selection origin so the checker can enforce async option commitment.
- id: autocomplete_freeform-mui-v2-T04
  name: 'Allowed domains: replace one existing chip and add one custom domain'
  canonical_type: autocomplete_freeform
  implementation_source: mui
  implementation_variant: Autocomplete (multiple, freeSolo, filterSelectedOptions)
  implementation_component: 'MUI: Autocomplete (multiple, freeSolo) inside Drawer'
  task_template: select_many
  secondary_template: confirm_cancel
  browsergym_goal: Open “Allowed domains”. In the “Domains” field, remove `example.net`, keep `example.com`, add the suggested value `research.org`, add the custom value `labs.dev`, and end with exactly those three domains. Then click “Save domains”. The task will finish automatically when done.
  ui_copy: Open “Allowed domains”. In the “Domains” field, remove `example.net`, keep `example.com`, add the suggested value `research.org`, add the custom value `labs.dev`, and end with exactly those three domains. Then click “Save domains”. The task will finish automatically when done.
  setup_description: |-
    The page uses a drawer_flow layout with compact spacing, bottom-right placement, and medium clutter. Clicking “Allowed domains” opens a right-side drawer containing one MUI Autocomplete configured with multiple and freeSolo. Selected values are shown as chips inside the input.

    Initial chips:
    - example.com
    - example.net

    Suggested options include example.com, example.net, research.org, staging.local. The value labs.dev is not in the suggestions and must be committed as free text. filterSelectedOptions is enabled so already selected suggestions disappear from the list. The drawer-local action “Save domains” commits the chip set.
  scene_context:
    theme: light
    spacing: compact
    layout: drawer_flow
    placement: bottom_right
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: drawer
  design_intent:
    active_factors:
    - overlay_model
    - confirmation_model
    - placement
    factor_rationale: This task mixes deletion, suggestion-backed addition, and custom-value creation in one chip field, so the agent has to audit the exact final set before leaving the drawer.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 3
      density_choice_interference: 3
      depth_layering: 3
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 2
    justification: Removing one existing chip, preserving one, adding one suggestion-backed chip, and adding one custom chip creates more recovery/state-tracking burden than a plain add-only labels task.
  success_trigger:
    human_readable:
    - The committed Domains chip set is exactly {example.com, research.org, labs.dev}.
    - Save domains has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        values:
        - example.com
        - research.org
        - labs.dev
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save domains
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Leaving example.net selected.
  - Adding research.org but failing to commit labs.dev as a chip.
  - Adding an extra domain such as staging.local.
  - Leaving the correct chips visible but not clicking Save domains.
  expected_interaction_path:
  - Open the drawer.
  - Remove the example.net chip.
  - Add research.org from suggestions and labs.dev as a custom value.
  - Verify the final set is example.com, research.org, labs.dev.
  - Click Save domains.
  notes: Track final chip array and ensure free-text values are committed as chips rather than left in the text buffer.
- id: autocomplete_freeform-mantine-v2-T01
  name: Primary labels only with accept-on-blur disabled
  canonical_type: autocomplete_freeform
  implementation_source: mantine
  implementation_variant: TagsInput (acceptValueOnBlur=false, maxTags=3)
  implementation_component: 'Mantine: TagsInput (acceptValueOnBlur=false, maxTags=3) in settings panel'
  task_template: select_many
  secondary_template: confirm_cancel
  browsergym_goal: In the labels settings panel, edit only the “Primary labels” field so it ends with exactly `api`, `ui`, and `agent`. It starts with `api`. Do not change “Secondary labels”. Then click “Save labels”. The task will finish automatically when done.
  ui_copy: In the labels settings panel, edit only the “Primary labels” field so it ends with exactly `api`, `ui`, and `agent`. It starts with `api`. Do not change “Secondary labels”. Then click “Save labels”. The task will finish automatically when done.
  setup_description: |-
    Layout is settings_panel with compact spacing, off-center placement, and high clutter. Two Mantine TagsInput controls are stacked in the target section:
    - Primary labels (target) — initial chips: api
    - Secondary labels — initial chips: docs

    Both fields are configured with suggestions, acceptValueOnBlur={false}, and maxTags=3. Suggestions include api, ui, agent, docs, infra. Because acceptValueOnBlur is disabled, typing and then blurring does not create a chip; values are committed only by pressing Enter or clicking a suggestion. A panel-local “Save labels” button commits both fields.
  scene_context:
    theme: light
    spacing: compact
    layout: settings_panel
    placement: off_center
    scale: default
    instances: 2
    guidance: text
    clutter: high
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - instances
    - confirmation_model
    - clutter
    factor_rationale: The task combines a common Mantine-specific commit rule with two near-identical tags fields, so the agent has to manage chip commitment and wrong-instance risk simultaneously.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 4
      density_choice_interference: 3
      depth_layering: 2
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 5
    justification: This is difficult because blurring does not commit values, the target field has a max tag cap, and a sibling TagsInput with overlapping suggestions is present in the same cluttered panel.
  success_trigger:
    human_readable:
    - Primary labels equals exactly {api, ui, agent}.
    - Secondary labels remains exactly {docs}.
    - Save labels has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        target_instance: Primary labels
        values:
        - api
        - ui
        - agent
        non_target_instances_must_remain:
          Secondary labels:
          - docs
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save labels
      require_correct_instance: true
      target_instance_label_or_id: Primary labels
      terminal_condition: true
  negative_cases:
  - Leaving typed text uncommitted because the field was blurred instead of submitting a tag.
  - Editing Secondary labels.
  - Exceeding the max tag cap and ending with the wrong chip set.
  - Forgetting to click Save labels.
  expected_interaction_path:
  - Focus Primary labels.
  - Commit ui and agent as chips via Enter or suggestion clicks.
  - Leave Secondary labels untouched.
  - Click Save labels.
  notes: The checker should inspect the controlled value array, not the live search text.
- id: autocomplete_freeform-mantine-v2-T02
  name: Reviewer autocomplete with first-option trap
  canonical_type: autocomplete_freeform
  implementation_source: mantine
  implementation_variant: Autocomplete (selectFirstOptionOnChange)
  implementation_component: 'Mantine: Autocomplete (selectFirstOptionOnChange) in compact inline surface'
  task_template: search_and_select
  secondary_template: confirm_cancel
  browsergym_goal: In the compact review card, set the “Reviewer” field to `Ana María` by choosing it from the suggestions list, then click “Apply reviewer”. The task will finish automatically when done.
  ui_copy: In the compact review card, set the “Reviewer” field to `Ana María` by choosing it from the suggestions list, then click “Apply reviewer”. The task will finish automatically when done.
  setup_description: |-
    The page uses an inline_surface layout anchored near the top-left with compact spacing, small scale, and medium clutter. The target card contains one Mantine Autocomplete labeled “Reviewer”. selectFirstOptionOnChange is enabled, so when the user types, the first option is automatically selected for keyboard submission.

    Suggestions include:
    - Ana Marin
    - Ana María
    - Ana-Marie
    - Annika Ma

    Initial state is empty. The card footer contains “Apply reviewer”. The field still accepts arbitrary input because Mantine Autocomplete is not a searchable select; the checker therefore requires suggestion-origin selection rather than raw text entry.
  scene_context:
    theme: light
    spacing: compact
    layout: inline_surface
    placement: top_left
    scale: small
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - scale
    - placement
    - guidance
    factor_rationale: This task exploits a documented autocomplete behavior that can make the first option too “sticky”, so the agent must actively choose the right diacritic variant instead of assuming Enter is safe.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 4
      density_choice_interference: 4
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 3
      disambiguation_load: 2
    justification: The compact top-left placement raises acquisition cost, and the option list is intentionally full of near-miss names where auto-selecting the first candidate is often wrong.
  success_trigger:
    human_readable:
    - Reviewer equals `Ana María` with suggestion-origin selection.
    - Apply reviewer has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        value: Ana María
        selection_origin: suggestion
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Apply reviewer
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Submitting Ana Marin because it appears first.
  - Typing Ana María as raw text without selecting a suggestion.
  - Selecting Ana-Marie or Annika Ma.
  - Leaving the card unapplied.
  expected_interaction_path:
  - Focus the Reviewer field.
  - Type Ana and inspect the suggestion list carefully.
  - Click Ana María rather than accepting the first near-match.
  - Click Apply reviewer.
  notes: Record whether the final value came from suggestion selection or raw input.
- id: autocomplete_freeform-mantine-v2-T03
  name: Drawer replacement under maxTags saturation
  canonical_type: autocomplete_freeform
  implementation_source: mantine
  implementation_variant: TagsInput (maxTags=3, clearable)
  implementation_component: 'Mantine: TagsInput (maxTags=3, clearable) inside Drawer'
  task_template: select_many
  secondary_template: confirm_cancel
  browsergym_goal: Open the “Edit rollout tags” drawer. In the “Rollout tags” field, replace the existing tags so the final set is exactly `alpha`, `delta`, and `epsilon`. Then click “Apply rollout tags”. The task will finish automatically when done.
  ui_copy: Open the “Edit rollout tags” drawer. In the “Rollout tags” field, replace the existing tags so the final set is exactly `alpha`, `delta`, and `epsilon`. Then click “Apply rollout tags”. The task will finish automatically when done.
  setup_description: |-
    The page uses a drawer_flow layout with compact spacing, bottom-left placement, and medium clutter. Clicking “Edit rollout tags” opens a left drawer containing one Mantine TagsInput labeled “Rollout tags”.

    Configuration:
    - maxTags=3
    - clearable
    - suggestions: alpha, beta, gamma, delta, epsilon

    Initial chips are alpha, beta, gamma, so the field is already saturated. The target set requires preserving alpha while removing beta and gamma, then adding delta and epsilon. The drawer footer contains “Cancel” and “Apply rollout tags”.
  scene_context:
    theme: light
    spacing: compact
    layout: drawer_flow
    placement: bottom_left
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: drawer
  design_intent:
    active_factors:
    - confirmation_model
    - placement
    - clutter
    factor_rationale: 'A saturated editable tags field turns this into a recovery task: the agent must free capacity before adding replacements, verify the exact final set, and then commit inside an off-center drawer.'
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 3
      density_choice_interference: 3
      depth_layering: 3
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 2
    justification: The maxTags cap creates a real edit-and-recover burden instead of a simple add-only action, and the save boundary makes the exact final chip set matter.
  success_trigger:
    human_readable:
    - Rollout tags equals exactly {alpha, delta, epsilon}.
    - Apply rollout tags has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        values:
        - alpha
        - delta
        - epsilon
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Apply rollout tags
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Leaving beta or gamma selected.
  - Clearing all tags and failing to restore alpha.
  - Adding an extra tag beyond the required three.
  - Stopping with the right chips in the draft state but not clicking Apply rollout tags.
  expected_interaction_path:
  - Open the drawer.
  - Remove beta and gamma to free capacity.
  - Add delta and epsilon while keeping alpha.
  - Click Apply rollout tags.
  notes: The clear button exists, but using it would force re-entering alpha as well; the checker only cares about the final exact set at commit time.
- id: autocomplete_freeform-mantine-v2-T04
  name: Modal office autocomplete with auto-select-on-blur hazard
  canonical_type: autocomplete_freeform
  implementation_source: mantine
  implementation_variant: Autocomplete (autoSelectOnBlur)
  implementation_component: 'Mantine: Autocomplete (autoSelectOnBlur) inside Modal'
  task_template: match_reference
  secondary_template: confirm_cancel
  browsergym_goal: Click “Assign office”. In the modal, set the “Office” field to match the office shown in the reference pill, select it from the suggestions list, and click “Save office”. The task will finish automatically when done.
  ui_copy: Click “Assign office”. In the modal, set the “Office” field to match the office shown in the reference pill, select it from the suggestions list, and click “Save office”. The task will finish automatically when done.
  setup_description: |-
    Layout is modal_flow with compact spacing, center placement, and medium clutter. A reference pill on the background card shows the target office code `ZRH-HQ`. Clicking “Assign office” opens a modal containing one Mantine Autocomplete labeled “Office”.

    The field uses autoSelectOnBlur and suggestions:
    - ZRH-HQ
    - ZRH-Lab
    - ZRH-Helpdesk
    - ZRM-HQ

    The control still accepts any string because it is an Autocomplete, not a Select. The benchmark checker therefore requires that the final value match the reference pill and come from a suggestion commit. The modal footer has “Cancel” and “Save office”.
  scene_context:
    theme: light
    spacing: compact
    layout: modal_flow
    placement: center
    scale: default
    instances: 1
    guidance: mixed
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: false
    confirmation_model: explicit_confirm
    precision_surface: mixed
    secondary_supporting_primitive: dialog
  design_intent:
    active_factors:
    - guidance
    - overlay_model
    - confirmation_model
    factor_rationale: The reference pill grounds the target visually, while autoSelectOnBlur means a careless blur can commit the wrong highlighted option; the task therefore pressures suggestion auditing before modal save.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 3
      density_choice_interference: 4
      depth_layering: 4
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 2
    justification: The office code is short, but the modal flow and blur-driven commit behavior make this harder than an ordinary free text field, especially with multiple ZRH-like distractors.
  success_trigger:
    human_readable:
    - Office equals the reference code `ZRH-HQ` and the final value came from a suggestion selection.
    - Save office has been clicked.
    canonical_predicate:
      predicate_type: matches_reference
      target_state:
        reference_id: office-reference-pill
        value: ZRH-HQ
        selection_origin: suggestion
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save office
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Typing ZRH-HQ as raw text without selecting the suggestion.
  - Auto-selecting ZRH-Lab or ZRH-Helpdesk by blur.
  - Choosing ZRM-HQ because the code looks similar.
  - Closing the modal without saving.
  expected_interaction_path:
  - Read the reference pill.
  - Open the modal and type ZRH.
  - Select ZRH-HQ from the suggestion list.
  - Click Save office.
  notes: The checker should derive the target from the reference pill and still require suggestion-origin commitment.
- id: autocomplete_restricted-antd-v2-T01
  name: Clear only the fallback region in a compact routing panel
  canonical_type: autocomplete_restricted
  implementation_source: antd
  implementation_variant: Select (allowClear, showSearch)
  implementation_component: 'AntD: Select (showSearch, allowClear) in settings panel'
  task_template: clear_reset
  secondary_template: confirm_cancel
  browsergym_goal: In the routing settings panel, clear the “Fallback region” field so it has no selection. Do not change “Primary region”. Then click “Save routing”. The task will finish automatically when done.
  ui_copy: In the routing settings panel, clear the “Fallback region” field so it has no selection. Do not change “Primary region”. Then click “Save routing”. The task will finish automatically when done.
  setup_description: |-
    Layout is settings_panel with compact spacing, off-center placement, and high clutter. The target area contains two Ant Design Select controls with allowClear enabled:
    - Primary region — selected value: us-east-1
    - Fallback region — selected value: eu-west-1

    Both controls share the same small clear affordance on the right side and identical option lists. Above them are unrelated checkboxes, a text input, and status chips. Hovering or focusing a field reveals its clear button. The panel action “Save routing” commits the current selections.
  scene_context:
    theme: light
    spacing: compact
    layout: settings_panel
    placement: off_center
    scale: default
    instances: 2
    guidance: text
    clutter: high
  component_context:
    overlay_model: none
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - instances
    - clutter
    - confirmation_model
    factor_rationale: The difficulty comes from locating and activating the correct embedded clear affordance among two identical selects in a cluttered compact panel, then explicitly saving the preserved/cleared state.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 5
      density_choice_interference: 2
      depth_layering: 1
      feedback_dynamics: 4
      semantic_observability: 3
      disambiguation_load: 5
    justification: This is a hard clear task because the actionable target is a small suffix control on one of two nearly identical selects, and the wrong field is easy to clear by mistake.
  success_trigger:
    human_readable:
    - Fallback region is committed as empty / null.
    - Primary region remains us-east-1.
    - Save routing has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: Fallback region
        selected_option: null
        non_target_instances_must_remain:
          Primary region: us-east-1
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save routing
      require_correct_instance: true
      target_instance_label_or_id: Fallback region
      terminal_condition: true
  negative_cases:
  - Clearing Primary region instead of Fallback region.
  - Selecting a different fallback option rather than clearing.
  - Opening the dropdown and closing it without changing the committed value.
  - Clearing the right field but not clicking Save routing.
  expected_interaction_path:
  - Locate Fallback region.
  - Reveal and click its clear affordance.
  - Leave Primary region unchanged.
  - Click Save routing.
  notes: Expose the controlled selected value after save; clearing the search text alone must not count.
- id: autocomplete_restricted-antd-v2-T02
  name: Drawer severity selection from a non-searchable long list
  canonical_type: autocomplete_restricted
  implementation_source: antd
  implementation_variant: Select (showSearch=false, virtual=false)
  implementation_component: 'AntD: Select (non-searchable) inside Drawer'
  task_template: scroll_find
  secondary_template: confirm_cancel
  browsergym_goal: Open the “Escalation policy” drawer. In the “Severity level” field, open the dropdown, scroll the list, and select `Level 72`. Then click “Apply policy”. The task will finish automatically when done.
  ui_copy: Open the “Escalation policy” drawer. In the “Severity level” field, open the dropdown, scroll the list, and select `Level 72`. Then click “Apply policy”. The task will finish automatically when done.
  setup_description: |-
    The page uses drawer_flow with compact spacing, bottom-right placement, and medium clutter. Clicking “Escalation policy” opens a right-side drawer. Inside is one Ant Design Select labeled “Severity level”. Search is disabled for this instance.

    The dropdown contains a fixed-height, internally scrollable list of 80 options from Level 01 to Level 80. Level 72 is not visible when the dropdown first opens. The drawer also contains a read-only summary card and two unrelated toggles. “Apply policy” commits the drawer state.
  scene_context:
    theme: light
    spacing: compact
    layout: drawer_flow
    placement: bottom_right
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: true
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: drawer
  design_intent:
    active_factors:
    - internal_scroll_region
    - overlay_model
    - confirmation_model
    factor_rationale: This preserves a high-value restricted-select failure mode from v1—within-list scrolling when search is unavailable—but raises pressure with a drawer boundary and explicit apply step.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 4
      density_choice_interference: 4
      depth_layering: 4
      feedback_dynamics: 3
      semantic_observability: 2
      disambiguation_load: 2
    justification: The task is intrinsically hard because the option list is long and non-searchable, the scroll region is inside the dropdown rather than the page, and the result still must be explicitly applied.
  success_trigger:
    human_readable:
    - Severity level equals Level 72.
    - Apply policy has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        selected_option: Level 72
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Apply policy
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Scrolling the page or drawer instead of the dropdown list.
  - Selecting a neighboring option such as Level 71 or Level 73.
  - Typing into the field even though search is disabled.
  - Selecting Level 72 but not clicking Apply policy.
  expected_interaction_path:
  - Open the drawer.
  - Open Severity level.
  - Scroll within the dropdown list until Level 72 is visible.
  - Select Level 72 and click Apply policy.
  notes: Disable virtualization for this task or expose a stable list container so dropdown scrolling is deterministic.
- id: autocomplete_restricted-antd-v2-T03
  name: Modal time zone requires real option commit
  canonical_type: autocomplete_restricted
  implementation_source: antd
  implementation_variant: Select (showSearch, optionFilterProp="label")
  implementation_component: 'AntD: Select (showSearch) inside Modal'
  task_template: search_and_select
  secondary_template: confirm_cancel
  browsergym_goal: Click “Edit profile”. In the modal, set the “Time zone” field to `UTC+05:30 — Kolkata` by selecting the matching option from the dropdown, then click “Save profile”. The task will finish automatically when done.
  ui_copy: Click “Edit profile”. In the modal, set the “Time zone” field to `UTC+05:30 — Kolkata` by selecting the matching option from the dropdown, then click “Save profile”. The task will finish automatically when done.
  setup_description: |-
    Layout is modal_flow with compact spacing, centered placement, and medium clutter. Clicking “Edit profile” opens a modal containing one Ant Design Select labeled “Time zone”. The Select uses showSearch and optionFilterProp="label", so typing filters by the displayed label rather than the internal value code.

    The dropdown contains 52 time zones spanning UTC‒12:00 through UTC+14:00, requiring scrolling to reach the target. Similar options near the target include:
    - UTC+05:00 — Karachi
    - UTC+05:00 — Tashkent
    - UTC+05:30 — Kolkata
    - UTC+05:30 — Colombo
    - UTC+05:45 — Kathmandu

    Initial value is UTC+00:00 — UTC. Typing only changes the popup filter; the committed value changes only after an option selection. The modal footer contains “Cancel” and “Save profile”.
  scene_context:
    theme: light
    spacing: compact
    layout: modal_flow
    placement: center
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: true
    confirmation_model: explicit_confirm
    precision_surface: discrete
    secondary_supporting_primitive: dialog
  design_intent:
    active_factors:
    - overlay_model
    - confirmation_model
    - clutter
    factor_rationale: The target list is confusable and the search box is only a filter, so the task measures whether the agent understands that typed search text is not the same thing as a committed restricted value.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 3
      density_choice_interference: 4
      depth_layering: 4
      feedback_dynamics: 4
      semantic_observability: 3
      disambiguation_load: 2
    justification: The option list is semantically dense, and the modal save boundary raises recovery cost if the agent stops after filtering instead of actually selecting.
  success_trigger:
    human_readable:
    - Time zone equals `UTC+05:30 — Kolkata` as a committed selected option.
    - Save profile has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        selected_option: UTC+05:30 — Kolkata
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save profile
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Leaving the field at UTC+00:00 — UTC after typing a search query.
  - Selecting UTC+05:00 — Karachi, UTC+05:45 — Kathmandu, or UTC+06:00 — Dhaka.
  - Closing or canceling the modal instead of saving.
  - Typing the label and assuming that is enough without selecting the option.
  expected_interaction_path:
  - Open the modal.
  - Open Time zone and type 05:30 or Kolkata.
  - Select UTC+05:30 — Kolkata from the filtered list.
  - Click Save profile.
  notes: The checker should read the committed selected value, not the transient search input.
- id: autocomplete_restricted-antd-v2-T04
  name: Server B region only in compact editable table
  canonical_type: autocomplete_restricted
  implementation_source: antd
  implementation_variant: Select in table cell
  implementation_component: 'AntD: Select inside editable table row with row-local save'
  task_template: table_operation
  secondary_template: search_and_select
  browsergym_goal: In the server table, set the “Region” field for `Server B` to `eu-central-1`, then click “Save” for Server B. Do not change the other rows. The task will finish automatically when done.
  ui_copy: In the server table, set the “Region” field for `Server B` to `eu-central-1`, then click “Save” for Server B. Do not change the other rows. The task will finish automatically when done.
  setup_description: |-
    Layout is table_cell with compact spacing, small scale, and medium clutter. A three-row server table lists Server A, Server B, and Server C. The Region column is editable for Server B and Server C only. Each editable cell contains an Ant Design Select with showSearch enabled and a row-local Save action.

    Options include:
    - us-east-1
    - us-west-2
    - eu-central-1
    - eu-west-3

    Initial state:
    - Server B: us-east-1
    - Server C: eu-west-3
  scene_context:
    theme: light
    spacing: compact
    layout: table_cell
    placement: off_center
    scale: small
    instances: 2
    guidance: text
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: table
  design_intent:
    active_factors:
    - layout
    - scale
    - instances
    factor_rationale: Dense row layout plus two editable same-type instances makes row binding and row-local save the primary pressure rather than the select choice itself.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 5
      density_choice_interference: 3
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 3
      disambiguation_load: 5
    justification: 'This is a classic wrong-row risk task: the options are easy, but the table layout is compact, two rows are editable, and the save control is row-scoped.'
  success_trigger:
    human_readable:
    - Server B Region equals eu-central-1.
    - Server C remains eu-west-3.
    - Server B Save has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: Server B
        selected_option: eu-central-1
        non_target_instances_must_remain:
          Server C: eu-west-3
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: save-server-b
      require_correct_instance: true
      target_instance_label_or_id: Server B
      terminal_condition: true
  negative_cases:
  - Changing Server C instead of Server B.
  - Selecting eu-west-3 instead of eu-central-1.
  - Changing the right row but not clicking the Save action for that row.
  - Editing both editable rows.
  expected_interaction_path:
  - Locate the Server B row.
  - Open its Region select and choose eu-central-1.
  - Leave Server C unchanged.
  - Click Save on the Server B row.
- id: autocomplete_restricted-mui-v2-T01
  name: Clear billing channel only with the embedded clear button
  canonical_type: autocomplete_restricted
  implementation_source: mui
  implementation_variant: Autocomplete (clearable, freeSolo=false)
  implementation_component: 'MUI: Autocomplete (restricted) in compact settings panel'
  task_template: clear_reset
  secondary_template: confirm_cancel
  browsergym_goal: In the channels panel, clear only the “Billing channel” field so it has no value. Do not change “Shipping channel”. Then click “Save channels”. The task will finish automatically when done.
  ui_copy: In the channels panel, clear only the “Billing channel” field so it has no value. Do not change “Shipping channel”. Then click “Save channels”. The task will finish automatically when done.
  setup_description: |-
    The page uses a settings_panel layout with compact spacing, top-right placement, and high clutter. Two MUI Autocomplete controls appear side by side:
    - Shipping channel — value: Email
    - Billing channel — value: SMS

    Both are restricted (freeSolo=false) and clearable, so a clear indicator appears inside the input when the field has a value. The panel is dense: there are several switches, helper texts, and an unrelated select directly above the fields. “Save channels” commits the panel state.
  scene_context:
    theme: light
    spacing: compact
    layout: settings_panel
    placement: top_right
    scale: default
    instances: 2
    guidance: text
    clutter: high
  component_context:
    overlay_model: none
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - instances
    - placement
    - confirmation_model
    factor_rationale: A small embedded clear affordance inside one of two neighboring autocompletes is easy to mis-ground, especially in a crowded top-right panel; the explicit save step makes preservation of the other field part of success.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 5
      density_choice_interference: 2
      depth_layering: 1
      feedback_dynamics: 4
      semantic_observability: 3
      disambiguation_load: 5
    justification: This is hard because the action target is the clear button, not the whole field, and it sits in a crowded compact panel with an almost identical neighboring input.
  success_trigger:
    human_readable:
    - Billing channel is null / empty.
    - Shipping channel remains Email.
    - Save channels has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: Billing channel
        selected_option: null
        non_target_instances_must_remain:
          Shipping channel: Email
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save channels
      require_correct_instance: true
      target_instance_label_or_id: Billing channel
      terminal_condition: true
  negative_cases:
  - Clearing Shipping channel instead of Billing channel.
  - Replacing Billing channel with a different value rather than clearing it.
  - Removing the visible text without clearing the selected value.
  - Forgetting to click Save channels.
  expected_interaction_path:
  - Reveal the clear indicator for Billing channel.
  - Clear Billing channel only.
  - Verify Shipping channel is still Email.
  - Click Save channels.
  notes: Use the committed selected value for checking; transient input text should not count.
- id: autocomplete_restricted-mui-v2-T02
  name: Manager selection with clearOnBlur false and near-duplicate names
  canonical_type: autocomplete_restricted
  implementation_source: mui
  implementation_variant: Autocomplete (freeSolo=false, clearOnBlur=false)
  implementation_component: 'MUI: Autocomplete (restricted) in settings panel'
  task_template: search_and_select
  secondary_template: confirm_cancel
  browsergym_goal: In the assignment panel, set the “Manager” field to `Dana Wu` by selecting it from the dropdown, then click “Apply assignment”. The task will finish automatically when done.
  ui_copy: In the assignment panel, set the “Manager” field to `Dana Wu` by selecting it from the dropdown, then click “Apply assignment”. The task will finish automatically when done.
  setup_description: |-
    Layout is settings_panel with compact spacing, bottom-right placement, and medium clutter. The target section contains one restricted MUI Autocomplete labeled “Manager”. clearOnBlur is set to false, so filtering text can remain visible when focus moves away if no option has been selected.

    The option list includes:
    - Dana Wu
    - Daniel Wu
    - Dina Wu
    - Dana Xu

    Initial value is empty. The panel action “Apply assignment” commits the value.
  scene_context:
    theme: light
    spacing: compact
    layout: settings_panel
    placement: bottom_right
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - placement
    - confirmation_model
    - clutter
    factor_rationale: Because clearOnBlur is false, the field can look “worked on” without actually holding a committed option, so the task explicitly tests whether the agent distinguishes filtered text from a selected value.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 3
      density_choice_interference: 4
      depth_layering: 2
      feedback_dynamics: 4
      semantic_observability: 3
      disambiguation_load: 2
    justification: Several near-duplicate names are present, and the component is configured so filtered text can persist after blur, which makes false-positive state perception more likely.
  success_trigger:
    human_readable:
    - Manager equals Dana Wu as a committed selected option.
    - Apply assignment has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        selected_option: Dana Wu
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Apply assignment
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Leaving filter text in the input without actually selecting Dana Wu.
  - Selecting Daniel Wu, Dina Wu, or Dana Xu.
  - Applying the panel while the selected option is still null.
  - Using keyboard blur or outside click as if it committed the value.
  expected_interaction_path:
  - Focus Manager and type Dana.
  - Select Dana Wu from the popup.
  - Verify the field shows a committed selection, not just search text.
  - Click Apply assignment.
  notes: Record selected option separately from inputValue so free text residue cannot satisfy the checker.
- id: autocomplete_restricted-mui-v2-T03
  name: Async warehouse picker in modal
  canonical_type: autocomplete_restricted
  implementation_source: mui
  implementation_variant: Autocomplete (freeSolo=false, async load-on-open)
  implementation_component: 'MUI: Autocomplete (restricted, async load-on-open) inside Modal'
  task_template: open_and_select
  secondary_template: confirm_cancel
  browsergym_goal: Click “Edit warehouse”. In the modal, open the “Warehouse” field, wait for the options to load, choose `Frankfurt-03`, and click “Save warehouse”. The task will finish automatically when done.
  ui_copy: Click “Edit warehouse”. In the modal, open the “Warehouse” field, wait for the options to load, choose `Frankfurt-03`, and click “Save warehouse”. The task will finish automatically when done.
  setup_description: |-
    Layout is modal_flow with compact spacing, centered placement, and medium clutter. Clicking “Edit warehouse” opens a modal containing one restricted MUI Autocomplete labeled “Warehouse”. The field uses async load-on-open rather than static options: a loading spinner is shown after the popup opens, and the list becomes selectable only after the results arrive.

    Loaded options include Frankfurt-01, Frankfurt-02, Frankfurt-03, and Paris-01. Initial value is Frankfurt-01. The modal action “Save warehouse” commits the selection.
  scene_context:
    theme: light
    spacing: compact
    layout: modal_flow
    placement: center
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: false
    confirmation_model: explicit_confirm
    precision_surface: discrete
    secondary_supporting_primitive: dialog
  design_intent:
    active_factors:
    - overlay_model
    - confirmation_model
    - clutter
    factor_rationale: This tests whether the agent can manage an async open-boundary flow inside a modal instead of treating the first empty popup state as final.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 2
      density_choice_interference: 3
      depth_layering: 4
      feedback_dynamics: 4
      semantic_observability: 3
      disambiguation_load: 2
    justification: The options themselves are simple, but the interaction is multi-step and asynchronous, so premature clicks or premature save actions are easy failure modes.
  success_trigger:
    human_readable:
    - Warehouse equals Frankfurt-03.
    - Save warehouse has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        selected_option: Frankfurt-03
        options_source: async
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save warehouse
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Saving Frankfurt-01 because the options were never changed.
  - Clicking before the async options have loaded and ending with no committed change.
  - Choosing Frankfurt-02 or Paris-01.
  - Closing or canceling the modal.
  expected_interaction_path:
  - Open the modal.
  - Open Warehouse and wait for the spinner to resolve.
  - Choose Frankfurt-03.
  - Click Save warehouse.
  notes: Expose loading state and selected option in the wrapper; raw input text should not matter because freeSolo is false.
- id: autocomplete_restricted-mui-v2-T04
  name: Node 3 availability zone in editable dashboard table
  canonical_type: autocomplete_restricted
  implementation_source: mui
  implementation_variant: Autocomplete in table row
  implementation_component: 'MUI: Autocomplete (restricted) inside row editor with save'
  task_template: table_operation
  secondary_template: open_and_select
  browsergym_goal: In the node table, set the “Availability zone” for `Node 3` to `us-east-1c`, then click “Save” for Node 3. Do not change any other row. The task will finish automatically when done.
  ui_copy: In the node table, set the “Availability zone” for `Node 3` to `us-east-1c`, then click “Save” for Node 3. Do not change any other row. The task will finish automatically when done.
  setup_description: |-
    The page uses a dashboard_panel layout with compact spacing, off-center placement, and medium clutter. A node health panel above the table shows charts and counters. The table itself contains three rows: Node 1, Node 2, and Node 3. Only Node 2 and Node 3 have editable Availability zone cells, each implemented with a restricted MUI Autocomplete and a row-local Save action.

    Valid zones are us-east-1a, us-east-1b, us-east-1c, and us-east-1d. Initial values:
    - Node 2: us-east-1b
    - Node 3: us-east-1a
  scene_context:
    theme: light
    spacing: compact
    layout: dashboard_panel
    placement: off_center
    scale: default
    instances: 2
    guidance: text
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: table
  design_intent:
    active_factors:
    - instances
    - layout
    - confirmation_model
    factor_rationale: The target cell is embedded in a cluttered dashboard and there are two same-type editable row controls, so row binding and row-local commit are the real sources of pressure.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 4
      density_choice_interference: 2
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 3
      disambiguation_load: 5
    justification: The option set is easy, but wrong-row editing is a real risk and the save action is local rather than global, making this a robust disambiguation task.
  success_trigger:
    human_readable:
    - Node 3 Availability zone equals us-east-1c.
    - Node 2 remains us-east-1b.
    - Node 3 Save has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: Node 3
        selected_option: us-east-1c
        non_target_instances_must_remain:
          Node 2: us-east-1b
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: save-node-3
      require_correct_instance: true
      target_instance_label_or_id: Node 3
      terminal_condition: true
  negative_cases:
  - Editing Node 2 instead of Node 3.
  - Selecting the wrong availability zone.
  - Changing the correct row but failing to click its Save action.
  - Changing both editable rows.
  expected_interaction_path:
  - Locate Node 3.
  - Open its Availability zone editor and choose us-east-1c.
  - Leave Node 2 untouched.
  - Click Save for Node 3.
- id: autocomplete_restricted-mantine-v2-T01
  name: Clear backup locale only in a dense locale panel
  canonical_type: autocomplete_restricted
  implementation_source: mantine
  implementation_variant: Select (clearable)
  implementation_component: 'Mantine: Select (clearable, searchable) in settings panel'
  task_template: clear_reset
  secondary_template: confirm_cancel
  browsergym_goal: In the locale settings panel, clear only the “Backup locale” field so it has no selection. Do not change “Primary locale”. Then click “Save locale settings”. The task will finish automatically when done.
  ui_copy: In the locale settings panel, clear only the “Backup locale” field so it has no selection. Do not change “Primary locale”. Then click “Save locale settings”. The task will finish automatically when done.
  setup_description: |-
    Layout is settings_panel with compact spacing, top-left placement, and high clutter. Two Mantine Select controls are shown:
    - Primary locale — en-US
    - Backup locale — de-DE

    Both are searchable and clearable, so a clear button appears in the right section when the field has a value. Several helper descriptions, switches, and a radio group are packed around the two fields. “Save locale settings” commits the panel state.
  scene_context:
    theme: light
    spacing: compact
    layout: settings_panel
    placement: top_left
    scale: default
    instances: 2
    guidance: text
    clutter: high
  component_context:
    overlay_model: none
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - instances
    - clutter
    - confirmation_model
    factor_rationale: The embedded clear affordance is small and visually similar across both fields, so this task is about correctly targeting one value-preserving clear action inside a dense settings panel.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 5
      density_choice_interference: 2
      depth_layering: 1
      feedback_dynamics: 4
      semantic_observability: 3
      disambiguation_load: 5
    justification: The target action is tiny and duplicated, and the panel is cluttered enough that a wrong-field clear is plausible even for a careful agent.
  success_trigger:
    human_readable:
    - Backup locale is null / empty.
    - Primary locale remains en-US.
    - Save locale settings has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: Backup locale
        selected_option: null
        non_target_instances_must_remain:
          Primary locale: en-US
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save locale settings
      require_correct_instance: true
      target_instance_label_or_id: Backup locale
      terminal_condition: true
  negative_cases:
  - Clearing Primary locale.
  - Replacing Backup locale with another value instead of clearing it.
  - Removing visible text without clearing the controlled value.
  - Not clicking Save locale settings.
  expected_interaction_path:
  - Reveal the clear affordance for Backup locale.
  - Clear Backup locale only.
  - Verify Primary locale is still en-US.
  - Click Save locale settings.
  notes: Use the committed selected option for checking; transient search text must not count.
- id: autocomplete_restricted-mantine-v2-T02
  name: Non-searchable project code in nested scroll
  canonical_type: autocomplete_restricted
  implementation_source: mantine
  implementation_variant: Select (searchable=false)
  implementation_component: 'Mantine: Select (non-searchable) inside nested scroll panel'
  task_template: scroll_find
  secondary_template: confirm_cancel
  browsergym_goal: In the nested “Timesheet link” panel, open the “Project code” field, scroll the dropdown list, select `PRJ-087`, and click “Save project link”. The task will finish automatically when done.
  ui_copy: In the nested “Timesheet link” panel, open the “Project code” field, scroll the dropdown list, select `PRJ-087`, and click “Save project link”. The task will finish automatically when done.
  setup_description: |-
    Layout is nested_scroll with compact spacing, center placement, and medium clutter. The page itself has an outer scroll container. Inside it, a scrollable side panel titled “Timesheet link” contains one Mantine Select labeled “Project code”. Search is disabled for this Select.

    The dropdown contains 100 project codes from PRJ-001 to PRJ-100 and is internally scrollable. PRJ-087 is initially offscreen. The panel footer has a “Save project link” action.
  scene_context:
    theme: light
    spacing: compact
    layout: nested_scroll
    placement: center
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: true
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - internal_scroll_region
    - layout
    - confirmation_model
    factor_rationale: The page and the dropdown both scroll, so the task directly tests whether the agent can operate the right scroll region before committing a non-searchable restricted value.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 4
      density_choice_interference: 4
      depth_layering: 3
      feedback_dynamics: 3
      semantic_observability: 2
      disambiguation_load: 2
    justification: The long code list is uniform and non-searchable, and the nested scroll structure makes it easy to manipulate the wrong viewport.
  success_trigger:
    human_readable:
    - Project code equals PRJ-087.
    - Save project link has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        selected_option: PRJ-087
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save project link
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Scrolling the panel or page instead of the dropdown.
  - Selecting PRJ-086 or PRJ-088.
  - Leaving the field empty after scrolling.
  - Not clicking Save project link.
  expected_interaction_path:
  - Open Project code.
  - Scroll within the dropdown until PRJ-087 is visible.
  - Select PRJ-087 and click Save project link.
- id: autocomplete_restricted-mantine-v2-T03
  name: On-call team only in a three-select routing card
  canonical_type: autocomplete_restricted
  implementation_source: mantine
  implementation_variant: Select (searchable, withCheckIcon=false)
  implementation_component: 'Mantine: Select (searchable, withCheckIcon=false) in dashboard panel'
  task_template: search_and_select
  secondary_template: confirm_cancel
  browsergym_goal: In the routing card, set only the “On-call team” field to `Platform Core` and then click “Apply routing”. Do not change “Primary team” or “Escalation team”. The task will finish automatically when done.
  ui_copy: In the routing card, set only the “On-call team” field to `Platform Core` and then click “Apply routing”. Do not change “Primary team” or “Escalation team”. The task will finish automatically when done.
  setup_description: |-
    The page uses a dashboard_panel layout with compact spacing, off-center placement, and high clutter. A routing card contains three Mantine Select controls with the same searchable option list:
    - Primary team — Payments
    - On-call team — (empty)
    - Escalation team — Infra

    The option list includes Platform Core, Platform Ops, Product Infra, Payments, and Infra. withCheckIcon={false} is enabled, so the active-option checkmark is removed from the dropdown. The card-local action “Apply routing” commits all three values.
  scene_context:
    theme: light
    spacing: compact
    layout: dashboard_panel
    placement: off_center
    scale: default
    instances: 3
    guidance: text
    clutter: high
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - instances
    - clutter
    - guidance
    factor_rationale: The active-option check icon is removed, so the task emphasizes post-selection verification and correct field binding across three same-type searchable selects in a cluttered dashboard card.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 4
      density_choice_interference: 4
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 3
      disambiguation_load: 5
    justification: Without an option check icon, it is easier to lose track of which select actually changed, and three same-type instances in one card make the task strongly disambiguation-heavy.
  success_trigger:
    human_readable:
    - On-call team equals Platform Core.
    - Primary team remains Payments.
    - Escalation team remains Infra.
    - Apply routing has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: On-call team
        selected_option: Platform Core
        non_target_instances_must_remain:
          Primary team: Payments
          Escalation team: Infra
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Apply routing
      require_correct_instance: true
      target_instance_label_or_id: On-call team
      terminal_condition: true
  negative_cases:
  - Setting Primary team or Escalation team instead of On-call team.
  - Selecting Platform Ops or Product Infra instead of Platform Core.
  - Applying routing with the target field still empty.
  - Changing more than one field.
  expected_interaction_path:
  - Focus On-call team.
  - Search for Platform and choose Platform Core.
  - Leave the other two selects unchanged.
  - Click Apply routing.
- id: autocomplete_restricted-mantine-v2-T04
  name: Backup approver in modal with three similar selects
  canonical_type: autocomplete_restricted
  implementation_source: mantine
  implementation_variant: Select (searchable, allowDeselect=false)
  implementation_component: 'Mantine: Select (searchable, allowDeselect=false) inside Modal'
  task_template: open_and_select
  secondary_template: confirm_cancel
  browsergym_goal: Click “Edit approvers”. In the modal, set only “Backup approver” to `Nina Shah`, then click “Save approvers”. Do not change the other approver fields. The task will finish automatically when done.
  ui_copy: Click “Edit approvers”. In the modal, set only “Backup approver” to `Nina Shah`, then click “Save approvers”. Do not change the other approver fields. The task will finish automatically when done.
  setup_description: |-
    Layout is modal_flow with compact spacing, centered placement, and medium clutter. Clicking “Edit approvers” opens a modal containing three searchable Mantine Select controls:
    - Primary approver — Alex Kim
    - Backup approver — (empty)
    - Escalation approver — Morgan Lee

    The shared option list is Alex Kim, Nina Shah, Morgan Lee, Priya Singh, and Dana Wu. allowDeselect={false} is enabled so selecting an already selected option will not clear it. The modal footer contains “Cancel” and “Save approvers”.
  scene_context:
    theme: light
    spacing: compact
    layout: modal_flow
    placement: center
    scale: default
    instances: 3
    guidance: text
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: false
    confirmation_model: explicit_confirm
    precision_surface: discrete
    secondary_supporting_primitive: dialog
  design_intent:
    active_factors:
    - instances
    - overlay_model
    - confirmation_model
    factor_rationale: This keeps the core difficulty on correct-instance selection across three same-type controls, but the modal boundary and explicit save step add the layered state verification that old low-pressure cases lacked.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 3
      density_choice_interference: 3
      depth_layering: 4
      feedback_dynamics: 3
      semantic_observability: 3
      disambiguation_load: 5
    justification: Three same-type selects in one modal create strong wrong-field risk, and the explicit modal save means the task cannot end immediately after the dropdown closes.
  success_trigger:
    human_readable:
    - Backup approver equals Nina Shah.
    - Primary approver remains Alex Kim.
    - Escalation approver remains Morgan Lee.
    - Save approvers has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: Backup approver
        selected_option: Nina Shah
        non_target_instances_must_remain:
          Primary approver: Alex Kim
          Escalation approver: Morgan Lee
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save approvers
      require_correct_instance: true
      target_instance_label_or_id: Backup approver
      terminal_condition: true
  negative_cases:
  - Changing Primary approver or Escalation approver.
  - Selecting Dana Wu or Priya Singh instead of Nina Shah.
  - Leaving Backup approver empty and still saving.
  - Closing the modal without clicking Save approvers.
  expected_interaction_path:
  - Open the modal.
  - Focus Backup approver and search for Nina.
  - Select Nina Shah.
  - Click Save approvers.
- id: select_with_search-antd-v2-T01
  name: Billing market with label-based search and explicit save
  canonical_type: select_with_search
  implementation_source: antd
  implementation_variant: Select (showSearch, optionFilterProp="label")
  implementation_component: 'AntD: Select (showSearch, optionFilterProp="label") in settings panel'
  task_template: search_and_select
  secondary_template: confirm_cancel
  browsergym_goal: In the market settings panel, set only the “Billing market” field to `Congo (DRC)` and then click “Save markets”. Do not change “Shipping market”. The task will finish automatically when done.
  ui_copy: In the market settings panel, set only the “Billing market” field to `Congo (DRC)` and then click “Save markets”. Do not change “Shipping market”. The task will finish automatically when done.
  setup_description: |-
    Layout is settings_panel with compact spacing, off-center placement, and high clutter. The card contains two Ant Design Select controls with showSearch:
    - Shipping market — Germany
    - Billing market — (empty)

    The option list is built from `{label, value}` objects, and optionFilterProp is set to `label` so filtering follows the visible labels. Confusable options include:
    - Congo (Republic)
    - Congo (DRC)
    - Congo (Brazzaville)
    - Cameroon

    “Save markets” commits the panel state.
  scene_context:
    theme: light
    spacing: compact
    layout: settings_panel
    placement: off_center
    scale: default
    instances: 2
    guidance: text
    clutter: high
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - instances
    - clutter
    - confirmation_model
    factor_rationale: This preserves a strong old search-and-choose pattern but raises pressure by making the correct field one of two similar selects in a crowded settings panel and by requiring an explicit save.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 4
      density_choice_interference: 4
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 3
      disambiguation_load: 5
    justification: The candidate labels are highly confusable and the task becomes truly hard once the agent also has to bind the action to the right one of two searchable selects and preserve the sibling field.
  success_trigger:
    human_readable:
    - Billing market equals Congo (DRC).
    - Shipping market remains Germany.
    - Save markets has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: Billing market
        selected_option: Congo (DRC)
        non_target_instances_must_remain:
          Shipping market: Germany
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save markets
      require_correct_instance: true
      target_instance_label_or_id: Billing market
      terminal_condition: true
  negative_cases:
  - Selecting Congo (Republic) or Congo (Brazzaville).
  - Changing Shipping market instead of Billing market.
  - Filtering the options with Congo but not selecting a value.
  - Forgetting to click Save markets.
  expected_interaction_path:
  - Focus Billing market.
  - Type Congo to narrow the visible labels.
  - Select Congo (DRC).
  - Click Save markets.
- id: select_with_search-antd-v2-T02
  name: Pipeline B region in compact row editor
  canonical_type: select_with_search
  implementation_source: antd
  implementation_variant: Select (showSearch) in table cell
  implementation_component: 'AntD: Select (showSearch) inside editable table row'
  task_template: table_operation
  secondary_template: search_and_select
  browsergym_goal: In the compact pipelines table, set the “Region” field for `Pipeline B` to `South Asia`, then click “Save” for Pipeline B. Do not change the other rows. The task will finish automatically when done.
  ui_copy: In the compact pipelines table, set the “Region” field for `Pipeline B` to `South Asia`, then click “Save” for Pipeline B. Do not change the other rows. The task will finish automatically when done.
  setup_description: |-
    Layout is table_cell with compact spacing, small scale, and medium clutter. Three rows are visible: Pipeline A, Pipeline B, and Pipeline C. Pipeline B and Pipeline C each have an editable Ant Design Select with showSearch in the Region column plus a row-local Save action.

    The region list includes South Asia, Southeast Asia, East Asia, Middle East, and Africa. Initial values:
    - Pipeline B: Europe
    - Pipeline C: East Asia
  scene_context:
    theme: light
    spacing: compact
    layout: table_cell
    placement: bottom_right
    scale: small
    instances: 2
    guidance: text
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: table
  design_intent:
    active_factors:
    - layout
    - instances
    - scale
    factor_rationale: The term-level confusability between South Asia and Southeast Asia is modest on its own, but compact row editing plus row-local save is what turns this into a strong disambiguation test.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 5
      density_choice_interference: 3
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 3
      disambiguation_load: 5
    justification: Small row controls, two editable rows, and a near-neighbor option name create a robust hard case even though the underlying option list is short.
  success_trigger:
    human_readable:
    - Pipeline B Region equals South Asia.
    - Pipeline C remains East Asia.
    - Save for Pipeline B has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: Pipeline B
        selected_option: South Asia
        non_target_instances_must_remain:
          Pipeline C: East Asia
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: save-pipeline-b
      require_correct_instance: true
      target_instance_label_or_id: Pipeline B
      terminal_condition: true
  negative_cases:
  - Selecting Southeast Asia instead of South Asia.
  - Editing Pipeline C instead of Pipeline B.
  - Selecting the correct option but clicking the wrong row Save or not saving at all.
  - Changing both rows.
  expected_interaction_path:
  - Locate Pipeline B.
  - Open its Region select and search for South.
  - Choose South Asia.
  - Click Save for Pipeline B.
- id: select_with_search-antd-v2-T03
  name: Drawer language picker with visible-only long list and apply
  canonical_type: select_with_search
  implementation_source: antd
  implementation_variant: Select (showSearch, listHeight custom, getPopupContainer parent)
  implementation_component: 'AntD: Select (showSearch) inside Drawer with parent popup container'
  task_template: scroll_find
  secondary_template: confirm_cancel
  browsergym_goal: Open the “Report settings” drawer. In the “Report language” field, open the dropdown, scroll the visible list until you can select `Zulu`, and then click “Apply settings”. The task will finish automatically when done.
  ui_copy: Open the “Report settings” drawer. In the “Report language” field, open the dropdown, scroll the visible list until you can select `Zulu`, and then click “Apply settings”. The task will finish automatically when done.
  setup_description: |-
    The page uses drawer_flow with compact spacing, top-right placement, and medium clutter. Clicking “Report settings” opens a right-side drawer. Inside is one Ant Design Select labeled “Report language” with showSearch enabled and a long alphabetized option list of roughly 120 languages.

    The dropdown is rendered inside the drawer’s parent container rather than the page body, and the visible list is fixed-height with internal scrolling. Initial selected value is English. Zulu is near the end of the list and not initially visible. “Apply settings” commits the drawer state.
  scene_context:
    theme: light
    spacing: compact
    layout: drawer_flow
    placement: top_right
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: true
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: drawer
  design_intent:
    active_factors:
    - internal_scroll_region
    - overlay_model
    - confirmation_model
    factor_rationale: This upgrades an old useful long-list pattern by moving it into a drawer and making the popup local to that drawer, so the agent must manage the correct overlay and commit boundary rather than a bare isolated-card scroll.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 3
      density_choice_interference: 4
      depth_layering: 4
      feedback_dynamics: 3
      semantic_observability: 2
      disambiguation_load: 2
    justification: The list is long and the target is offscreen, but the main hardness comes from operating the drawer-contained dropdown correctly and not confusing it with page-level scrolling.
  success_trigger:
    human_readable:
    - Report language equals Zulu.
    - Apply settings has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        selected_option: Zulu
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Apply settings
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Scrolling the drawer or page instead of the dropdown list.
  - Selecting Yoruba or another end-of-list language.
  - Typing Zulu without selecting it.
  - Leaving the drawer unapplied.
  expected_interaction_path:
  - Open the drawer.
  - Open Report language.
  - Scroll within the dropdown list until Zulu appears.
  - Select Zulu and click Apply settings.
  notes: Bind the popup to the drawer parent so overlay scoping is deterministic.
- id: select_with_search-antd-v2-T04
  name: Reference icon to visibility level with explicit apply
  canonical_type: select_with_search
  implementation_source: antd
  implementation_variant: Select (showSearch) with custom option render
  implementation_component: 'AntD: Select (showSearch) with icon+label options in dashboard panel'
  task_template: match_reference
  secondary_template: confirm_cancel
  browsergym_goal: In the sharing panel, set the “Visibility” field so it matches the icon shown in the reference tile, then click “Apply sharing”. The task will finish automatically when done.
  ui_copy: In the sharing panel, set the “Visibility” field so it matches the icon shown in the reference tile, then click “Apply sharing”. The task will finish automatically when done.
  setup_description: |-
    Layout is dashboard_panel with compact spacing, bottom-left placement, and high clutter. A small reference tile in the target card shows only an icon: 🔒. Below it is an Ant Design Select labeled “Visibility” with showSearch enabled and custom-rendered options:
    - 🌐 Public
    - 🏢 Internal
    - 🔒 Confidential
    - 🧪 Restricted Preview

    Initial value is Public. The panel includes unrelated stats and action chips. “Apply sharing” commits the new visibility level.
  scene_context:
    theme: light
    spacing: compact
    layout: dashboard_panel
    placement: bottom_left
    scale: default
    instances: 1
    guidance: visual
    clutter: high
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - guidance
    - clutter
    - confirmation_model
    factor_rationale: The target is given visually rather than textually, so the task measures whether the agent can read the reference tile, map it to the right searchable option, and still execute the explicit apply step in a cluttered panel.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 3
      density_choice_interference: 3
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 4
      disambiguation_load: 2
    justification: The option set is short, but the icon-only reference prevents a simple text copy path and the dense dashboard context makes quick target acquisition less reliable.
  success_trigger:
    human_readable:
    - Visibility equals Confidential, the option corresponding to the 🔒 reference icon.
    - Apply sharing has been clicked.
    canonical_predicate:
      predicate_type: matches_reference
      target_state:
        reference_id: visibility-reference-icon
        selected_option: Confidential
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Apply sharing
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Selecting Restricted Preview instead of Confidential.
  - Ignoring the icon and leaving Public or Internal selected.
  - Searching the right label but not selecting it.
  - Failing to click Apply sharing.
  expected_interaction_path:
  - Read the reference icon.
  - Open Visibility and choose the matching option.
  - Click Apply sharing.
- id: select_with_search-mui-v2-T01
  name: Grouped destination picker in modal with save
  canonical_type: select_with_search
  implementation_source: mui
  implementation_variant: Autocomplete (groupBy, freeSolo=false)
  implementation_component: 'MUI: Autocomplete (restricted grouped search) inside Modal'
  task_template: search_and_select
  secondary_template: confirm_cancel
  browsergym_goal: Click “Edit destination”. In the modal, set the “Destination” field to `San Sebastián (ES)` and click “Save destination”. The task will finish automatically when done.
  ui_copy: Click “Edit destination”. In the modal, set the “Destination” field to `San Sebastián (ES)` and click “Save destination”. The task will finish automatically when done.
  setup_description: |-
    Layout is modal_flow with compact spacing, center placement, and medium clutter. Clicking “Edit destination” opens a modal containing one restricted MUI Autocomplete labeled “Destination”. The option list is grouped by country. Relevant options include:
    - San Sebastián (ES)
    - San Sebastián (MX)
    - San Jose (US)
    - San José (CR)

    The field is empty initially. “Save destination” commits the modal value.
  scene_context:
    theme: light
    spacing: compact
    layout: modal_flow
    placement: center
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: false
    confirmation_model: explicit_confirm
    precision_surface: discrete
    secondary_supporting_primitive: dialog
  design_intent:
    active_factors:
    - overlay_model
    - confirmation_model
    - guidance
    factor_rationale: Grouped headings and location near-misses make the search results semantically dense, and the modal save boundary forces verification of the selected restricted option rather than the search query.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 3
      density_choice_interference: 4
      depth_layering: 4
      feedback_dynamics: 4
      semantic_observability: 3
      disambiguation_load: 2
    justification: Several San* destinations share almost the same name, so the task is about exact label discrimination plus proper modal commit.
  success_trigger:
    human_readable:
    - Destination equals San Sebastián (ES).
    - Save destination has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        selected_option: San Sebastián (ES)
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save destination
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Selecting San Sebastián (MX), San Jose (US), or San José (CR).
  - Typing the label without selecting it.
  - Closing or canceling the modal.
  - Saving the field while it is still empty.
  expected_interaction_path:
  - Open the modal.
  - Search for San Seb and inspect the grouped results carefully.
  - Select San Sebastián (ES).
  - Click Save destination.
- id: select_with_search-mui-v2-T02
  name: Backup clinic only in a two-field async form
  canonical_type: select_with_search
  implementation_source: mui
  implementation_variant: Autocomplete (search-as-you-type, freeSolo=false)
  implementation_component: 'MUI: Autocomplete (async restricted search) in inline surface'
  task_template: search_and_select
  secondary_template: confirm_cancel
  browsergym_goal: In the clinic form, set only the “Backup clinic” field to `Saint Mary — West` by selecting it from the results list, then click “Apply clinics”. Do not change “Primary clinic”. The task will finish automatically when done.
  ui_copy: In the clinic form, set only the “Backup clinic” field to `Saint Mary — West` by selecting it from the results list, then click “Apply clinics”. Do not change “Primary clinic”. The task will finish automatically when done.
  setup_description: |-
    Layout is inline_surface with compact spacing, top-right placement, and medium clutter. Two restricted MUI Autocomplete controls appear in a narrow form:
    - Primary clinic — Saint Mary — East
    - Backup clinic — (empty)

    Both use async search-as-you-type. Searching `Saint` returns:
    - Saint Mary — East
    - Saint Mary — West
    - Saint Martha — North

    The form action “Apply clinics” commits both fields.
  scene_context:
    theme: light
    spacing: compact
    layout: inline_surface
    placement: top_right
    scale: default
    instances: 2
    guidance: text
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - instances
    - placement
    - confirmation_model
    factor_rationale: This is a compact two-field selective search task where the interesting failure is choosing the wrong sibling field or the wrong near-match from an async result list.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 4
      density_choice_interference: 4
      depth_layering: 3
      feedback_dynamics: 4
      semantic_observability: 3
      disambiguation_load: 5
    justification: The option names are similar, the input is compact and off-center, and a sibling field with overlapping semantics raises the disambiguation load substantially.
  success_trigger:
    human_readable:
    - Backup clinic equals Saint Mary — West.
    - Primary clinic remains Saint Mary — East.
    - Apply clinics has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: Backup clinic
        selected_option: Saint Mary — West
        non_target_instances_must_remain:
          Primary clinic: Saint Mary — East
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Apply clinics
      require_correct_instance: true
      target_instance_label_or_id: Backup clinic
      terminal_condition: true
  negative_cases:
  - Changing Primary clinic instead of Backup clinic.
  - Selecting Saint Mary — East or Saint Martha — North.
  - Leaving the field as search text instead of a selected option.
  - Failing to click Apply clinics.
  expected_interaction_path:
  - Focus Backup clinic.
  - Type Saint and wait for results.
  - Select Saint Mary — West.
  - Click Apply clinics.
- id: select_with_search-mui-v2-T03
  name: Tertiary KPI only in a three-select dashboard strip
  canonical_type: select_with_search
  implementation_source: mui
  implementation_variant: Autocomplete (freeSolo=false) in dashboard strip
  implementation_component: 'MUI: Autocomplete (restricted search) in dashboard panel'
  task_template: search_and_select
  secondary_template: confirm_cancel
  browsergym_goal: In the KPI strip, set only the “Tertiary KPI” field to `Net Promoter Score`, then click “Apply KPIs”. Do not change the other KPI fields. The task will finish automatically when done.
  ui_copy: In the KPI strip, set only the “Tertiary KPI” field to `Net Promoter Score`, then click “Apply KPIs”. Do not change the other KPI fields. The task will finish automatically when done.
  setup_description: |-
    Layout is dashboard_panel with compact spacing, center placement, and medium clutter. Three searchable MUI Autocomplete fields appear in a horizontal KPI strip:
    - Primary KPI — Revenue
    - Secondary KPI — Customer Satisfaction
    - Tertiary KPI — (empty)

    The shared metric list includes Net Promoter Score, Customer Satisfaction, Revenue, Churn Rate, and Response Time. A button labeled “Apply KPIs” commits the strip.
  scene_context:
    theme: light
    spacing: compact
    layout: dashboard_panel
    placement: center
    scale: default
    instances: 3
    guidance: text
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - instances
    - clutter
    - confirmation_model
    factor_rationale: The task preserves a strong old multi-instance search pattern but adds a mandatory commit step, making it a cleaner separation between “found the label” and “actually saved the right one.”
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 3
      density_choice_interference: 3
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 3
      disambiguation_load: 5
    justification: The metric names are readable, but three same-type fields in a dashboard strip still create substantial wrong-instance risk.
  success_trigger:
    human_readable:
    - Tertiary KPI equals Net Promoter Score.
    - Primary KPI remains Revenue.
    - Secondary KPI remains Customer Satisfaction.
    - Apply KPIs has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: Tertiary KPI
        selected_option: Net Promoter Score
        non_target_instances_must_remain:
          Primary KPI: Revenue
          Secondary KPI: Customer Satisfaction
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Apply KPIs
      require_correct_instance: true
      target_instance_label_or_id: Tertiary KPI
      terminal_condition: true
  negative_cases:
  - Changing Primary KPI or Secondary KPI instead of Tertiary KPI.
  - Choosing Customer Satisfaction or another metric.
  - Typing promoter and not selecting a restricted option.
  - Forgetting to click Apply KPIs.
  expected_interaction_path:
  - Focus Tertiary KPI.
  - Search for promoter or NPS.
  - Select Net Promoter Score.
  - Click Apply KPIs.
- id: select_with_search-mui-v2-T04
  name: Framework from logo reference in compact architecture card
  canonical_type: select_with_search
  implementation_source: mui
  implementation_variant: Autocomplete (renderOption with icon, freeSolo=false)
  implementation_component: 'MUI: Autocomplete with custom option render in settings card'
  task_template: match_reference
  secondary_template: confirm_cancel
  browsergym_goal: In the architecture card, set the “Framework” field so it matches the logo shown in the reference tile, then click “Save architecture”. The task will finish automatically when done.
  ui_copy: In the architecture card, set the “Framework” field so it matches the logo shown in the reference tile, then click “Save architecture”. The task will finish automatically when done.
  setup_description: |-
    Layout is inline_surface with compact spacing, bottom-left placement, and medium clutter. A small reference tile shows the Next.js wordmark only. Below it is a restricted MUI Autocomplete labeled “Framework” whose options are custom rendered with icon + label:
    - Next.js
    - Nuxt
    - NestJS
    - Remix

    Initial value is Remix. The card action “Save architecture” commits the new framework choice.
  scene_context:
    theme: light
    spacing: compact
    layout: inline_surface
    placement: bottom_left
    scale: default
    instances: 1
    guidance: visual
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - guidance
    - placement
    - confirmation_model
    factor_rationale: The reference is visual and the option labels are deliberately confusable, so the task checks whether the agent can ground the logo correctly and still finish the apply boundary.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 3
      density_choice_interference: 4
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 4
      disambiguation_load: 2
    justification: The label set contains close lexical distractors, and the compact off-center card makes it harder to rely on a quick text skim.
  success_trigger:
    human_readable:
    - Framework equals Next.js, matching the reference logo.
    - Save architecture has been clicked.
    canonical_predicate:
      predicate_type: matches_reference
      target_state:
        reference_id: framework-logo-ref
        selected_option: Next.js
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save architecture
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Choosing Nuxt or NestJS because the names look similar.
  - Leaving Remix selected.
  - Typing Next without selecting the option.
  - Not clicking Save architecture.
  expected_interaction_path:
  - Inspect the reference logo.
  - Open Framework and choose the matching option.
  - Click Save architecture.
- id: select_with_search-mantine-v2-T01
  name: Timezone search without check icon in dark panel
  canonical_type: select_with_search
  implementation_source: mantine
  implementation_variant: Select (searchable, withCheckIcon=false)
  implementation_component: 'Mantine: Select (searchable, withCheckIcon=false) in settings panel'
  task_template: search_and_select
  secondary_template: confirm_cancel
  browsergym_goal: In the timezone panel, set the “Time zone” field to `UTC+05:30 — Kolkata`, then click “Save timezone”. The task will finish automatically when done.
  ui_copy: In the timezone panel, set the “Time zone” field to `UTC+05:30 — Kolkata`, then click “Save timezone”. The task will finish automatically when done.
  setup_description: |-
    Layout is settings_panel with dark theme, compact spacing, off-center placement, and medium clutter. One searchable Mantine Select labeled “Time zone” is shown in the target card. withCheckIcon={false} is enabled, so the active-option checkmark is removed from the dropdown.

    The dropdown contains 52 time zones spanning UTC‒12:00 through UTC+14:00, requiring scrolling to reach the target. Similar options near the target include UTC+05:00 — Karachi, UTC+05:00 — Tashkent, UTC+05:30 — Kolkata, UTC+05:30 — Colombo, UTC+05:45 — Kathmandu. Initial value is UTC+00:00 — UTC. “Save timezone” commits the card state.
  scene_context:
    theme: dark
    spacing: compact
    layout: settings_panel
    placement: off_center
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: true
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - theme
    - confirmation_model
    - guidance
    factor_rationale: Removing the dropdown check icon lowers observability of the currently active option, so the agent has to verify through the field value itself before saving in a dark compact panel.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 3
      density_choice_interference: 4
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 3
      disambiguation_load: 2
    justification: The option list is confusable, and removing the check icon makes it easier to misread the dropdown state while searching in a dark panel.
  success_trigger:
    human_readable:
    - Time zone equals UTC+05:30 — Kolkata.
    - Save timezone has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        selected_option: UTC+05:30 — Kolkata
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save timezone
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Selecting UTC+05:00 — Karachi or UTC+05:45 — Kathmandu.
  - Typing a search query without choosing the restricted option.
  - Leaving the original value selected.
  - Failing to click Save timezone.
  expected_interaction_path:
  - Open Time zone.
  - Search for 05:30 or Kolkata.
  - Choose UTC+05:30 — Kolkata.
  - Click Save timezone.
- id: select_with_search-mantine-v2-T02
  name: Material row binding in inventory table with save
  canonical_type: select_with_search
  implementation_source: mantine
  implementation_variant: Select (searchable) in table cell
  implementation_component: 'Mantine: Select (searchable) inside inventory table row'
  task_template: table_operation
  secondary_template: scroll_find
  browsergym_goal: In the inventory table, change the “Material” field for row 1 to `Walnut`, then click “Save” for row 1. Do not change row 2. The task will finish automatically when done.
  ui_copy: In the inventory table, change the “Material” field for row 1 to `Walnut`, then click “Save” for row 1. Do not change row 2. The task will finish automatically when done.
  setup_description: |-
    Layout is table_cell with compact spacing, small scale, and medium clutter. The inventory table has two editable rows. Each row has a searchable Mantine Select in the Material column and a row-local Save action.

    The material list is long and alphabetized; Walnut is initially offscreen. Initial values:
    - Row 1: Aluminum
    - Row 2: Steel

    The dropdown list is internally scrollable. Only row 1 should change.
  scene_context:
    theme: light
    spacing: compact
    layout: table_cell
    placement: off_center
    scale: small
    instances: 2
    guidance: text
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: true
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: table
  design_intent:
    active_factors:
    - layout
    - internal_scroll_region
    - instances
    factor_rationale: This combines long-list dropdown scrolling with row binding and a row-local save, which is materially harder than the old isolated-card version of the same material search.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 4
      density_choice_interference: 4
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 2
      disambiguation_load: 5
    justification: The long list and small table layout make acquisition and row disambiguation both meaningful, especially because only one row should change and the save action is local.
  success_trigger:
    human_readable:
    - Row 1 Material equals Walnut.
    - Row 2 remains Steel.
    - Row 1 Save has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: row-1
        selected_option: Walnut
        non_target_instances_must_remain:
          row-2: Steel
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: save-row-1
      require_correct_instance: true
      target_instance_label_or_id: row-1
      terminal_condition: true
  negative_cases:
  - Changing row 2 instead of row 1.
  - Selecting a nearby offscreen material such as Wenge or Willow.
  - Scrolling the page instead of the dropdown list.
  - Selecting Walnut but not clicking the row 1 Save action.
  expected_interaction_path:
  - Open Material for row 1.
  - Scroll the dropdown list until Walnut is visible.
  - Select Walnut.
  - Click Save for row 1.
- id: select_with_search-mantine-v2-T03
  name: Top-position dropdown in compact corner card
  canonical_type: select_with_search
  implementation_source: mantine
  implementation_variant: Select (searchable, comboboxProps position=top)
  implementation_component: 'Mantine: Select (searchable, dropdown forced above input)'
  task_template: search_and_select
  secondary_template: confirm_cancel
  browsergym_goal: In the compact timezone card, set “Time zone” to `UTC+05:45 — Kathmandu`, then click “Apply timezone”. The task will finish automatically when done.
  ui_copy: In the compact timezone card, set “Time zone” to `UTC+05:45 — Kathmandu`, then click “Apply timezone”. The task will finish automatically when done.
  setup_description: |-
    Layout is inline_surface with compact spacing, bottom-right placement, small scale, and medium clutter. One searchable Mantine Select labeled “Time zone” sits in a small card. Its dropdown is forced to appear above the input using comboboxProps position=top with flip disabled.

    The dropdown contains 52 time zones spanning UTC‒12:00 through UTC+14:00, requiring scrolling to reach the target. Similar options near the target include UTC+05:00 — Karachi, UTC+05:00 — Tashkent, UTC+05:30 — Kolkata, UTC+05:30 — Colombo, UTC+05:45 — Kathmandu. Initial value is UTC+00:00 — UTC. The card action “Apply timezone” commits the choice.
  scene_context:
    theme: light
    spacing: compact
    layout: inline_surface
    placement: bottom_right
    scale: small
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: true
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: popover
  design_intent:
    active_factors:
    - placement
    - scale
    - overlay_model
    factor_rationale: Forcing the dropdown above a small corner-anchored control adds acquisition pressure without making the task artificial, because the search semantics remain standard and the option set is still human-solvable.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 4
      density_choice_interference: 4
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 3
      disambiguation_load: 2
    justification: Small scale, corner placement, and an above-the-input dropdown increase target acquisition cost, while the time zone labels themselves are quite confusable.
  success_trigger:
    human_readable:
    - Time zone equals UTC+05:45 — Kathmandu.
    - Apply timezone has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        selected_option: UTC+05:45 — Kathmandu
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Apply timezone
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Selecting UTC+05:30 — Kolkata or UTC+06:00 — Dhaka.
  - Typing the target text without selecting it.
  - Leaving the default UTC value unchanged.
  - Failing to click Apply timezone.
  expected_interaction_path:
  - Open the select and handle the dropdown appearing above the field.
  - Search for 05:45 or Kathmandu.
  - Choose UTC+05:45 — Kathmandu.
  - Click Apply timezone.
- id: select_with_search-mantine-v2-T04
  name: Three KPI selectors in cluttered dashboard card
  canonical_type: select_with_search
  implementation_source: mantine
  implementation_variant: Select (searchable, withCheckIcon=false)
  implementation_component: 'Mantine: Select (searchable) in dashboard panel with three instances'
  task_template: open_and_select
  secondary_template: confirm_cancel
  browsergym_goal: In the dashboard KPI card, set only the “Tertiary KPI” field to `Net Promoter Score`, then click “Save KPI card”. Do not change the other two KPI fields. The task will finish automatically when done.
  ui_copy: In the dashboard KPI card, set only the “Tertiary KPI” field to `Net Promoter Score`, then click “Save KPI card”. Do not change the other two KPI fields. The task will finish automatically when done.
  setup_description: |-
    Layout is dashboard_panel with compact spacing, center placement, and high clutter. Three searchable Mantine Select controls appear in a KPI card:
    - Primary KPI — Revenue
    - Secondary KPI — Customer Satisfaction
    - Tertiary KPI — (empty)

    Shared options include Revenue, Customer Satisfaction, Net Promoter Score, Retention, and Churn. withCheckIcon={false} is enabled. The card footer has “Save KPI card”.
  scene_context:
    theme: light
    spacing: compact
    layout: dashboard_panel
    placement: center
    scale: default
    instances: 3
    guidance: text
    clutter: high
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: discrete
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - instances
    - clutter
    - confirmation_model
    factor_rationale: This leans into one of the most reliable v2 levers—same-type instance confusion inside cluttered dashboard context—while still keeping the primary component the searchable select itself.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 3
      density_choice_interference: 3
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 3
      disambiguation_load: 5
    justification: Three similar searchable selects in a cluttered KPI card create a wrong-instance failure family that remains human-solvable but strongly diagnostic.
  success_trigger:
    human_readable:
    - Tertiary KPI equals Net Promoter Score.
    - Primary KPI remains Revenue.
    - Secondary KPI remains Customer Satisfaction.
    - Save KPI card has been clicked.
    canonical_predicate:
      predicate_type: equals
      target_state:
        target_instance: Tertiary KPI
        selected_option: Net Promoter Score
        non_target_instances_must_remain:
          Primary KPI: Revenue
          Secondary KPI: Customer Satisfaction
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save KPI card
      require_correct_instance: true
      target_instance_label_or_id: Tertiary KPI
      terminal_condition: true
  negative_cases:
  - Changing Primary KPI or Secondary KPI.
  - Selecting Customer Satisfaction instead of Net Promoter Score.
  - Leaving the Tertiary KPI empty.
  - Not clicking Save KPI card.
  expected_interaction_path:
  - Focus Tertiary KPI.
  - Search for promoter or NPS.
  - Choose Net Promoter Score.
  - Click Save KPI card.
- id: combobox_editable_multi-antd-v2-T01
  name: Incident tags with persistent menu and near-miss options
  canonical_type: combobox_editable_multi
  implementation_source: antd
  implementation_variant: Select (mode="tags", autoClearSearchValue=false)
  implementation_component: 'AntD: Select (mode="tags", autoClearSearchValue=false) in settings panel'
  task_template: select_many
  secondary_template: confirm_cancel
  browsergym_goal: In the incident settings panel, make the “Incident tags” field end with exactly `canary`, `prod-hotfix`, and `rollback`, then click “Apply incident tags”. The task will finish automatically when done.
  ui_copy: In the incident settings panel, make the “Incident tags” field end with exactly `canary`, `prod-hotfix`, and `rollback`, then click “Apply incident tags”. The task will finish automatically when done.
  setup_description: |-
    Layout is settings_panel with compact spacing, off-center placement, and high clutter. The target field is an Ant Design Select in tags mode labeled “Incident tags”. Search text is not cleared automatically after each selection. Initial chips: canary.

    Suggestions include:
    - prod
    - production
    - prod-hotfix
    - rollback
    - preprod

    A visually similar second tags field labeled “Archive tags” sits directly below with initial chip legacy and must remain unchanged. “Apply incident tags” commits the panel state.
  scene_context:
    theme: light
    spacing: compact
    layout: settings_panel
    placement: off_center
    scale: default
    instances: 2
    guidance: text
    clutter: high
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - instances
    - clutter
    - confirmation_model
    factor_rationale: 'This focuses on the strongest editable-multi failure family from v1: selecting exact chips from a persistent search state while avoiding near-miss options and preserving a sibling field.'
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 3
      density_choice_interference: 4
      depth_layering: 2
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 5
    justification: Residual search text, highly similar production-related labels, and a second similar tags field make the exact-set audit meaningfully hard.
  success_trigger:
    human_readable:
    - Incident tags equals exactly {canary, prod-hotfix, rollback}.
    - Archive tags remains exactly {legacy}.
    - Apply incident tags has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        target_instance: Incident tags
        values:
        - canary
        - prod-hotfix
        - rollback
        non_target_instances_must_remain:
          Archive tags:
          - legacy
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Apply incident tags
      require_correct_instance: true
      target_instance_label_or_id: Incident tags
      terminal_condition: true
  negative_cases:
  - Selecting prod or production instead of prod-hotfix.
  - Leaving a typed value uncommitted in the search buffer.
  - Changing Archive tags.
  - Not clicking Apply incident tags.
  expected_interaction_path:
  - Focus Incident tags.
  - Add prod-hotfix and rollback while keeping canary.
  - Verify Archive tags still equals legacy.
  - Click Apply incident tags.
- id: combobox_editable_multi-antd-v2-T02
  name: Modal labels require chip commits before OK
  canonical_type: combobox_editable_multi
  implementation_source: antd
  implementation_variant: Select (mode="tags") inside Modal
  implementation_component: 'AntD: Select (mode="tags") inside Modal'
  task_template: confirm_cancel
  secondary_template: select_many
  browsergym_goal: Click “Edit labels”. In the modal, make the “Labels” field contain exactly `backend`, `urgent`, and `p0`, then click “OK”. The task will finish automatically when done.
  ui_copy: Click “Edit labels”. In the modal, make the “Labels” field contain exactly `backend`, `urgent`, and `p0`, then click “OK”. The task will finish automatically when done.
  setup_description: |-
    Layout is modal_flow with compact spacing, center placement, and medium clutter. Clicking “Edit labels” opens a modal containing one Ant Design Select in tags mode labeled “Labels”.

    Initial chips: backend
    Suggested options: backend, urgent, p0, p1, docs

    The component also accepts custom text, but only committed chips count. Typed text that has not been turned into a chip does not count. The modal footer contains “Cancel” and “OK”, and changes are committed only when “OK” is clicked.
  scene_context:
    theme: light
    spacing: compact
    layout: modal_flow
    placement: center
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: false
    confirmation_model: explicit_confirm
    precision_surface: mixed
    secondary_supporting_primitive: dialog
  design_intent:
    active_factors:
    - overlay_model
    - confirmation_model
    - clutter
    factor_rationale: This keeps the task component-centric but forces the agent to distinguish draft text from committed chips and then cross the explicit modal commit boundary.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 3
      density_choice_interference: 3
      depth_layering: 4
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 2
    justification: The exact set is short, but modal commit and uncommitted chip text create a strong recovery/verification burden.
  success_trigger:
    human_readable:
    - Labels equals exactly {backend, urgent, p0}.
    - OK has been clicked to commit the modal state.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        values:
        - backend
        - urgent
        - p0
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: OK
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Having the correct text in the input buffer but not as chips.
  - Leaving an extra chip such as p1 or docs.
  - Clicking Cancel or closing the modal.
  - Changing backend to something else instead of preserving it.
  expected_interaction_path:
  - Open the modal.
  - Add urgent and p0 as committed chips while keeping backend.
  - Verify no extra chips remain.
  - Click OK.
  notes: Track the committed chip array, not just visible text in the search input.
- id: combobox_editable_multi-antd-v2-T03
  name: Max-count saturated tags in a drawer
  canonical_type: combobox_editable_multi
  implementation_source: antd
  implementation_variant: Select (mode="tags", maxCount=4, maxTagCount="responsive")
  implementation_component: 'AntD: Select (mode="tags", maxCount=4) inside Drawer'
  task_template: select_many
  secondary_template: confirm_cancel
  browsergym_goal: Open the “Edit release labels” drawer. In the “Release labels” field, end with exactly `alpha`, `gamma`, `omega`, and `theta`, then click “Save labels”. The task will finish automatically when done.
  ui_copy: Open the “Edit release labels” drawer. In the “Release labels” field, end with exactly `alpha`, `gamma`, `omega`, and `theta`, then click “Save labels”. The task will finish automatically when done.
  setup_description: |-
    Layout is drawer_flow with compact spacing, top-right placement, and medium clutter. Clicking “Edit release labels” opens a drawer with one Ant Design Select in tags mode labeled “Release labels”.

    maxCount=4 is enabled, and the field starts already full with:
    - alpha
    - beta
    - gamma
    - delta

    Suggestions include alpha, beta, gamma, delta, omega, theta, lambda. maxTagCount is responsive, so not all selected chips may be visible when the field is unfocused. “Save labels” commits the drawer state.
  scene_context:
    theme: light
    spacing: compact
    layout: drawer_flow
    placement: top_right
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: drawer
  design_intent:
    active_factors:
    - confirmation_model
    - placement
    - clutter
    factor_rationale: A saturated field forces remove-then-add recovery, and responsive chip collapsing makes exact-set verification less trivial before the drawer-local save.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 3
      density_choice_interference: 3
      depth_layering: 3
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 2
    justification: The exact set is easy to state but not easy to execute because the field is full at the start and not all chips remain visible when unfocused.
  success_trigger:
    human_readable:
    - Release labels equals exactly {alpha, gamma, omega, theta}.
    - Save labels has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        values:
        - alpha
        - gamma
        - omega
        - theta
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save labels
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Leaving beta or delta selected.
  - Removing alpha or gamma and not restoring the intended set.
  - Trying to add a fifth tag without first removing one.
  - Not clicking Save labels.
  expected_interaction_path:
  - Open the drawer.
  - Remove beta and delta to free capacity.
  - Add omega and theta while keeping alpha and gamma.
  - Click Save labels.
- id: combobox_editable_multi-antd-v2-T04
  name: Rule B regions only in dense policy table
  canonical_type: combobox_editable_multi
  implementation_source: antd
  implementation_variant: Select (mode="tags") in table row
  implementation_component: 'AntD: Select (mode="tags") inside editable policy table row'
  task_template: table_operation
  secondary_template: select_many
  browsergym_goal: In the policy table, set the “Regions” field for `Rule B` to exactly `US`, `EU`, and `APAC`, then click “Save” for Rule B. Do not change Rule A. The task will finish automatically when done.
  ui_copy: In the policy table, set the “Regions” field for `Rule B` to exactly `US`, `EU`, and `APAC`, then click “Save” for Rule B. Do not change Rule A. The task will finish automatically when done.
  setup_description: |-
    Layout is table_cell with compact spacing, small scale, and medium clutter. Two policy rows are editable:
    - Rule A — Regions: US
    - Rule B — Regions: US

    Each row has an Ant Design tags-mode field in the Regions column and a row-local Save action. Suggestions include US, EU, APAC, LATAM, and MEA. The table is dense and both rows look nearly identical.
  scene_context:
    theme: light
    spacing: compact
    layout: table_cell
    placement: off_center
    scale: small
    instances: 2
    guidance: text
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: table
  design_intent:
    active_factors:
    - layout
    - instances
    - scale
    factor_rationale: This isolates row binding and exact-set checking inside a dense table rather than a forgiving card, which better separates grounding failures from option-choice failures.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 5
      density_choice_interference: 3
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 3
      disambiguation_load: 5
    justification: The selected set itself is easy, but the dense row context and row-local save make wrong-row edits very plausible.
  success_trigger:
    human_readable:
    - Rule B Regions equals exactly {US, EU, APAC}.
    - Rule A remains exactly {US}.
    - Save for Rule B has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        target_instance: Rule B
        values:
        - US
        - EU
        - APAC
        non_target_instances_must_remain:
          Rule A:
          - US
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: save-rule-b
      require_correct_instance: true
      target_instance_label_or_id: Rule B
      terminal_condition: true
  negative_cases:
  - Adding EU and APAC to Rule A instead of Rule B.
  - Leaving out one of the required regions.
  - Adding any extra region such as LATAM.
  - Changing Rule B correctly but not clicking its Save action.
  expected_interaction_path:
  - Locate Rule B.
  - Add EU and APAC to the Regions field while keeping US.
  - Leave Rule A unchanged.
  - Click Save for Rule B.
- id: combobox_editable_multi-mui-v2-T01
  name: Categories from visual reference with persistent popup
  canonical_type: combobox_editable_multi
  implementation_source: mui
  implementation_variant: Autocomplete (multiple, disableCloseOnSelect, filterSelectedOptions)
  implementation_component: 'MUI: Autocomplete (multiple) in dark dashboard panel'
  task_template: match_reference
  secondary_template: confirm_cancel
  browsergym_goal: In the dark routing panel, make the “Categories” field match the chips shown in “Target categories”, then click “Save routing”. The task will finish automatically when done.
  ui_copy: In the dark routing panel, make the “Categories” field match the chips shown in “Target categories”, then click “Save routing”. The task will finish automatically when done.
  setup_description: |-
    Layout is dashboard_panel with dark theme, compact spacing, and high clutter. The target card contains one MUI Autocomplete configured with multiple, disableCloseOnSelect, and filterSelectedOptions. A visual reference row labeled “Target categories” shows the required chips: News, Sports, Finance, Travel.

    Initial chips in the interactive field are News and Science. Suggestions include News, Sports, Finance, Travel, Science, Weather, Health. The popup stays open after each selection. The card action “Save routing” commits the final set.
  scene_context:
    theme: dark
    spacing: compact
    layout: dashboard_panel
    placement: off_center
    scale: default
    instances: 1
    guidance: visual
    clutter: high
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - guidance
    - clutter
    - confirmation_model
    factor_rationale: The persistent popup makes repeated selection efficient but also raises near-miss risk, and the visual reference means the agent must continually compare the live chip set to an external target before saving.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 3
      density_choice_interference: 3
      depth_layering: 2
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 2
    justification: This is hard because the exact-set comparison is externalized to a visual target row, and one existing distractor chip has to be removed before the final save.
  success_trigger:
    human_readable:
    - Categories equals exactly {News, Sports, Finance, Travel}.
    - Save routing has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        values:
        - News
        - Sports
        - Finance
        - Travel
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save routing
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Leaving Science selected.
  - Adding Weather or Health.
  - Matching part of the reference but not all four chips.
  - Stopping without clicking Save routing.
  expected_interaction_path:
  - Compare the live chip set to the Target categories row.
  - Remove Science.
  - Add Sports, Finance, and Travel while keeping News.
  - Click Save routing.
- id: combobox_editable_multi-mui-v2-T02
  name: Collapsed chips exact set under limitTags
  canonical_type: combobox_editable_multi
  implementation_source: mui
  implementation_variant: Autocomplete (multiple, limitTags=2)
  implementation_component: 'MUI: Autocomplete (multiple, limitTags=2) in settings panel'
  task_template: select_many
  secondary_template: confirm_cancel
  browsergym_goal: In the “Categories” field, end with exactly `News`, `Sports`, `Finance`, `Travel`, `Tech`, and `Health`, then click “Apply categories”. The task will finish automatically when done.
  ui_copy: In the “Categories” field, end with exactly `News`, `Sports`, `Finance`, `Travel`, `Tech`, and `Health`, then click “Apply categories”. The task will finish automatically when done.
  setup_description: |-
    Layout is settings_panel with compact spacing, center placement, and medium clutter. The target field is a single MUI Autocomplete configured with multiple and limitTags=2, so when it is not focused only two chips and a “+N” indicator are visible.

    Initial chips: none.
    Options: News, Sports, Finance, Travel, Tech, Health, Weather, Science.
    The panel action “Apply categories” commits the final selected set.
  scene_context:
    theme: light
    spacing: compact
    layout: settings_panel
    placement: center
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - confirmation_model
    - clutter
    - guidance
    factor_rationale: The selected set is intentionally large enough that the unfocused control will collapse chips, so the agent must re-focus or otherwise verify the full state before committing.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 2
      density_choice_interference: 3
      depth_layering: 2
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 2
    justification: Matching six values is exacting on its own, and limitTags makes post-selection verification harder than the visible chip count initially suggests.
  success_trigger:
    human_readable:
    - Categories equals exactly {News, Sports, Finance, Travel, Tech, Health}.
    - Apply categories has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        values:
        - News
        - Sports
        - Finance
        - Travel
        - Tech
        - Health
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Apply categories
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Selecting only some of the six categories.
  - Adding Weather or Science as extra chips.
  - Assuming the +N counter means the field is complete without checking the actual values.
  - Not clicking Apply categories.
  expected_interaction_path:
  - Open Categories.
  - Select the six required values.
  - Refocus or inspect the field to verify the full set despite chip collapse.
  - Click Apply categories.
- id: combobox_editable_multi-mui-v2-T03
  name: Allowed domains modal with one removal and two additions
  canonical_type: combobox_editable_multi
  implementation_source: mui
  implementation_variant: Autocomplete (multiple, freeSolo, filterSelectedOptions)
  implementation_component: 'MUI: Autocomplete (multiple, freeSolo) inside Modal'
  task_template: confirm_cancel
  secondary_template: select_many
  browsergym_goal: Click “Edit allowed domains”. In the modal, make the “Allowed domains” field contain exactly `example.com`, `research.org`, and `labs.dev`, then click “Save”. The task will finish automatically when done.
  ui_copy: Click “Edit allowed domains”. In the modal, make the “Allowed domains” field contain exactly `example.com`, `research.org`, and `labs.dev`, then click “Save”. The task will finish automatically when done.
  setup_description: |-
    Layout is modal_flow with compact spacing, center placement, and medium clutter. Clicking “Edit allowed domains” opens a modal containing one MUI Autocomplete with multiple and freeSolo labeled “Allowed domains”.

    Initial chips: example.com, example.net
    Suggestions include example.com, example.net, research.org, staging.local.
    labs.dev is not in suggestions and must be committed as custom text. The modal footer contains “Cancel” and “Save”.
  scene_context:
    theme: light
    spacing: compact
    layout: modal_flow
    placement: center
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: false
    confirmation_model: explicit_confirm
    precision_surface: mixed
    secondary_supporting_primitive: dialog
  design_intent:
    active_factors:
    - overlay_model
    - confirmation_model
    - clutter
    factor_rationale: 'This is an exact editable-multi state-repair task: one old chip must be removed, one suggestion-backed chip added, and one custom chip added before the modal can be safely committed.'
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 2
      density_choice_interference: 3
      depth_layering: 4
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 2
    justification: The chip set is short, but modal commit and mixed source-of-value behavior make it a real state auditing task rather than a simple add/remove operation.
  success_trigger:
    human_readable:
    - Allowed domains equals exactly {example.com, research.org, labs.dev}.
    - Save has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        values:
        - example.com
        - research.org
        - labs.dev
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Leaving example.net selected.
  - Adding staging.local or any extra domain.
  - Leaving labs.dev as uncommitted text.
  - Closing the modal without clicking Save.
  expected_interaction_path:
  - Open the modal.
  - Remove example.net.
  - Add research.org from suggestions and labs.dev as a custom chip.
  - Click Save.
- id: combobox_editable_multi-mui-v2-T04
  name: Regions field only in advanced filters drawer
  canonical_type: combobox_editable_multi
  implementation_source: mui
  implementation_variant: Autocomplete (multiple, disableCloseOnSelect)
  implementation_component: 'MUI: Autocomplete (multiple) inside Drawer with sibling field'
  task_template: open_and_select
  secondary_template: confirm_cancel
  browsergym_goal: Open “Advanced filters”. In the drawer, set only the “Regions” field to exactly `North America`, `Europe`, and `Asia`, then click “Save filters”. Do not change “Countries”. The task will finish automatically when done.
  ui_copy: Open “Advanced filters”. In the drawer, set only the “Regions” field to exactly `North America`, `Europe`, and `Asia`, then click “Save filters”. Do not change “Countries”. The task will finish automatically when done.
  setup_description: |-
    The page uses drawer_flow with compact spacing, top-right placement, and medium clutter. Clicking “Advanced filters” opens a drawer containing two MUI multiple Autocomplete fields:
    - Regions — initial chips: none
    - Countries — initial chips: Canada

    The Regions options are North America, South America, Europe, Asia, Africa, and Oceania. The drawer also contains a date range field and a checkbox. disableCloseOnSelect is enabled so the popup remains open while multiple values are chosen. The drawer-local action “Save filters” commits the state.
  scene_context:
    theme: light
    spacing: compact
    layout: drawer_flow
    placement: top_right
    scale: default
    instances: 2
    guidance: text
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: drawer
  design_intent:
    active_factors:
    - instances
    - overlay_model
    - confirmation_model
    factor_rationale: This is a stronger version of the old sibling-field filter task because the drawer save boundary and persistent popup force the agent to audit the target field and the untouched distractor field together.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 3
      target_acquisition: 3
      density_choice_interference: 2
      depth_layering: 3
      feedback_dynamics: 3
      semantic_observability: 3
      disambiguation_load: 5
    justification: The option set is straightforward, but the right-field-only requirement inside a drawer makes the task strongly about instance binding and commit discipline.
  success_trigger:
    human_readable:
    - Regions equals exactly {North America, Europe, Asia}.
    - Countries remains exactly {Canada}.
    - Save filters has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        target_instance: Regions
        values:
        - North America
        - Europe
        - Asia
        non_target_instances_must_remain:
          Countries:
          - Canada
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save filters
      require_correct_instance: true
      target_instance_label_or_id: Regions
      terminal_condition: true
  negative_cases:
  - Changing Countries instead of Regions.
  - Missing one of the required regions.
  - Adding an extra region such as Africa.
  - Leaving the drawer unsaved.
  expected_interaction_path:
  - Open the drawer.
  - Use the Regions field to select North America, Europe, and Asia.
  - Leave Countries unchanged.
  - Click Save filters.
- id: combobox_editable_multi-mantine-v2-T01
  name: Primary labels only with exact three-chip target
  canonical_type: combobox_editable_multi
  implementation_source: mantine
  implementation_variant: TagsInput (acceptValueOnBlur=false)
  implementation_component: 'Mantine: TagsInput (acceptValueOnBlur=false) in settings panel'
  task_template: select_many
  secondary_template: confirm_cancel
  browsergym_goal: In the labels panel, set only the “Primary labels” field to exactly `api`, `ui`, and `agent`, then click “Save labels”. Do not change “Secondary labels”. The task will finish automatically when done.
  ui_copy: In the labels panel, set only the “Primary labels” field to exactly `api`, `ui`, and `agent`, then click “Save labels”. Do not change “Secondary labels”. The task will finish automatically when done.
  setup_description: |-
    Layout is settings_panel with compact spacing, off-center placement, and high clutter. Two Mantine TagsInput fields are shown:
    - Primary labels — initial chips: api
    - Secondary labels — initial chips: docs

    Both fields use suggestions, acceptValueOnBlur={false}, and no duplicate chips by default. Suggestions include api, ui, agent, docs, infra. The panel action “Save labels” commits the values.
  scene_context:
    theme: light
    spacing: compact
    layout: settings_panel
    placement: off_center
    scale: default
    instances: 2
    guidance: text
    clutter: high
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - instances
    - confirmation_model
    - clutter
    factor_rationale: This shifts the old Mantine tags pattern upward by making sibling preservation explicit and by forcing Enter-or-suggestion chip commits instead of blur-based shortcuts.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 4
      density_choice_interference: 3
      depth_layering: 2
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 5
    justification: Two similar editable chip fields, no blur-commit shortcut, and explicit save create a strong wrong-instance-plus-state-auditing task.
  success_trigger:
    human_readable:
    - Primary labels equals exactly {api, ui, agent}.
    - Secondary labels remains exactly {docs}.
    - Save labels has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        target_instance: Primary labels
        values:
        - api
        - ui
        - agent
        non_target_instances_must_remain:
          Secondary labels:
          - docs
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save labels
      require_correct_instance: true
      target_instance_label_or_id: Primary labels
      terminal_condition: true
  negative_cases:
  - Editing Secondary labels.
  - Leaving a typed value uncommitted.
  - Adding infra or docs as extra chips.
  - Failing to click Save labels.
  expected_interaction_path:
  - Focus Primary labels.
  - Commit ui and agent as chips.
  - Verify Secondary labels remains docs.
  - Click Save labels.
- id: combobox_editable_multi-mantine-v2-T02
  name: Drawer replacement under maxTags saturation
  canonical_type: combobox_editable_multi
  implementation_source: mantine
  implementation_variant: TagsInput (maxTags=3, clearable)
  implementation_component: 'Mantine: TagsInput (maxTags=3) inside Drawer'
  task_template: select_many
  secondary_template: confirm_cancel
  browsergym_goal: Open the “Edit billing tags” drawer. In the “Billing tags” field, replace the existing tags so the final set is exactly `billing`, `overdue`, and `escalation`, then click “Apply billing tags”. The task will finish automatically when done.
  ui_copy: Open the “Edit billing tags” drawer. In the “Billing tags” field, replace the existing tags so the final set is exactly `billing`, `overdue`, and `escalation`, then click “Apply billing tags”. The task will finish automatically when done.
  setup_description: |-
    Layout is drawer_flow with compact spacing, bottom-left placement, and medium clutter. Clicking “Edit billing tags” opens a drawer containing one Mantine TagsInput labeled “Billing tags”.

    Initial chips: invoice, payment, refund
    maxTags=3 and clearable are enabled.
    Suggestions include billing, overdue, escalation, refund, payment.
    The drawer footer contains “Cancel” and “Apply billing tags”.
  scene_context:
    theme: light
    spacing: compact
    layout: drawer_flow
    placement: bottom_left
    scale: default
    instances: 1
    guidance: text
    clutter: medium
  component_context:
    overlay_model: nested_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: drawer
  design_intent:
    active_factors:
    - confirmation_model
    - placement
    - clutter
    factor_rationale: A fully saturated field converts the task into a remove-and-rebuild operation, which better exposes recovery and final-set auditing failures than add-only chip tasks.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 3
      density_choice_interference: 3
      depth_layering: 3
      feedback_dynamics: 4
      semantic_observability: 4
      disambiguation_load: 2
    justification: The chip set is small but every target chip is a replacement, so the agent must carefully remove the old set, rebuild the new one, and then save.
  success_trigger:
    human_readable:
    - Billing tags equals exactly {billing, overdue, escalation}.
    - Apply billing tags has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        values:
        - billing
        - overdue
        - escalation
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Apply billing tags
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Leaving invoice, payment, or refund selected.
  - Rebuilding only part of the new set.
  - Using clear-all and then forgetting one required chip.
  - Not clicking Apply billing tags.
  expected_interaction_path:
  - Open the drawer.
  - Remove the initial three chips.
  - Add billing, overdue, and escalation as the new exact set.
  - Click Apply billing tags.
- id: combobox_editable_multi-mantine-v2-T03
  name: Top-position multiselect from visual framework target
  canonical_type: combobox_editable_multi
  implementation_source: mantine
  implementation_variant: MultiSelect (searchable, maxValues=4, comboboxProps position=top)
  implementation_component: 'Mantine: MultiSelect (searchable, dropdown above input) in compact card'
  task_template: match_reference
  secondary_template: confirm_cancel
  browsergym_goal: In the compact stack card, make the “Frameworks” field match the chips shown in “Target stack”, then click “Save stack”. The task will finish automatically when done.
  ui_copy: In the compact stack card, make the “Frameworks” field match the chips shown in “Target stack”, then click “Save stack”. The task will finish automatically when done.
  setup_description: |-
    Layout is inline_surface with compact spacing, bottom-right placement, small scale, and medium clutter. One searchable Mantine MultiSelect labeled “Frameworks” sits in a small card. The dropdown is forced above the input with comboboxProps position=top, and maxValues=4 is enabled.

    A visual target row labeled “Target stack” shows exactly four chips: React, Next.js, TypeScript, Tailwind.
    Options include React, Vue, Next.js, Nuxt, TypeScript, JavaScript, Tailwind, Bootstrap.
    “Save stack” commits the selected values.
  scene_context:
    theme: light
    spacing: compact
    layout: inline_surface
    placement: bottom_right
    scale: small
    instances: 1
    guidance: visual
    clutter: medium
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: popover
  design_intent:
    active_factors:
    - guidance
    - placement
    - scale
    factor_rationale: The target set is visual and the dropdown opens above a small corner-anchored input, increasing acquisition pressure while still keeping the underlying task a realistic MultiSelect comparison.
  difficulty:
    difficulty_bucket: hard
    tier: L2
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 3
      density_choice_interference: 3
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 4
      disambiguation_load: 2
    justification: Four exact values, a constrained max selection count, and an above-input dropdown make this meaningfully harder than a baseline multi-select card.
  success_trigger:
    human_readable:
    - Frameworks equals exactly {React, Next.js, TypeScript, Tailwind}.
    - Save stack has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        values:
        - React
        - Next.js
        - TypeScript
        - Tailwind
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save stack
      require_correct_instance: false
      target_instance_label_or_id: null
      terminal_condition: true
  negative_cases:
  - Selecting Vue, Nuxt, JavaScript, or Bootstrap.
  - Choosing only some of the target chips.
  - Exceeding the maxValues limit and forcing an unintended replacement.
  - Not clicking Save stack.
  expected_interaction_path:
  - Inspect the Target stack chips.
  - Open Frameworks and select the four required options.
  - Verify the exact set.
  - Click Save stack.
- id: combobox_editable_multi-mantine-v2-T04
  name: Exclude tags only in dual filter card
  canonical_type: combobox_editable_multi
  implementation_source: mantine
  implementation_variant: MultiSelect (searchable, clearable, withCheckIcon=false)
  implementation_component: 'Mantine: MultiSelect (searchable) with two instances in dashboard panel'
  task_template: open_and_select
  secondary_template: confirm_cancel
  browsergym_goal: In the filter card, set only the “Exclude tags” field to exactly `beta`, `deprecated`, and `noisy`, then click “Save filters”. Do not change “Include tags”. The task will finish automatically when done.
  ui_copy: In the filter card, set only the “Exclude tags” field to exactly `beta`, `deprecated`, and `noisy`, then click “Save filters”. Do not change “Include tags”. The task will finish automatically when done.
  setup_description: |-
    Layout is dashboard_panel with compact spacing, off-center placement, and high clutter. The target card contains two searchable Mantine MultiSelect controls:
    - Include tags — initial chips: stable
    - Exclude tags — initial chips: beta

    withCheckIcon={false} and clearable are enabled for both fields. Suggestions include stable, beta, deprecated, noisy, preview, internal. The card footer contains “Save filters”.
  scene_context:
    theme: light
    spacing: compact
    layout: dashboard_panel
    placement: off_center
    scale: default
    instances: 2
    guidance: text
    clutter: high
  component_context:
    overlay_model: single_overlay
    internal_scroll_region: false
    confirmation_model: explicit_apply
    precision_surface: mixed
    secondary_supporting_primitive: none
  design_intent:
    active_factors:
    - instances
    - clutter
    - confirmation_model
    factor_rationale: Two same-type multiselects with overlapping vocabularies make correct-instance targeting and final-set auditing the core challenge, especially once the check icon is removed.
  difficulty:
    difficulty_bucket: hard
    tier: L3
    axes_ratings:
      precision_requirement: 4
      target_acquisition: 3
      density_choice_interference: 3
      depth_layering: 2
      feedback_dynamics: 3
      semantic_observability: 3
      disambiguation_load: 5
    justification: The field-level semantics are opposed (include vs exclude), but the widgets are visually similar enough that the task is still primarily about correct-instance editing and exact chip-set verification.
  success_trigger:
    human_readable:
    - Exclude tags equals exactly {beta, deprecated, noisy}.
    - Include tags remains exactly {stable}.
    - Save filters has been clicked.
    canonical_predicate:
      predicate_type: set_equals
      target_state:
        target_instance: Exclude tags
        values:
        - beta
        - deprecated
        - noisy
        non_target_instances_must_remain:
          Include tags:
          - stable
        saved: true
      tolerance: null
      require_confirm: true
      confirm_control: Save filters
      require_correct_instance: true
      target_instance_label_or_id: Exclude tags
      terminal_condition: true
  negative_cases:
  - Editing Include tags instead of Exclude tags.
  - Leaving out deprecated or noisy.
  - Adding preview or internal as extra chips.
  - Not clicking Save filters.
  expected_interaction_path:
  - Focus Exclude tags.
  - Add deprecated and noisy while keeping beta.
  - Leave Include tags unchanged.
  - Click Save filters.