- id: rich_text_editor-mantine-v2-T01 name: Bubble menu carryover in the correct blurb editor canonical_type: rich_text_editor implementation_source: mantine implementation_variant: bubble_menu implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: editor_operation secondary_template: disclose browsergym_goal: In the “Service blurbs” panel, edit only the rich text editor labeled “Incident blurb”. Make `urgent` bold and `today` italic in the sentence `This is urgent today.`, then click “Apply blurb”. Do not edit “Public blurb”. The task will finish automatically when done. ui_copy: In the “Service blurbs” panel, edit only the rich text editor labeled “Incident blurb”. Make `urgent` bold and `today` italic in the sentence `This is urgent today.`, then click “Apply blurb”. Do not edit “Public blurb”. The task will finish automatically when done. setup_description: |- Layout uses settings_panel with compact spacing, small controls, and medium clutter. A Mantine Card titled “Service blurbs” contains two same-size rich text editors labeled “Public blurb” and “Incident blurb”. The target is “Incident blurb”. The editors are configured without a full toolbar. Instead, selecting text reveals a small BubbleMenu near the selection with Bold and Italic controls. The panel footer contains a single “Apply blurb” button that commits both drafts. Initial contents: - Public blurb: `Routine update only.` - Incident blurb: `This is urgent today.` Only the Incident blurb should change. scene_context: theme: light spacing: compact layout: settings_panel 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: popover design_intent: active_factors: - instances - confirmation_model - secondary_supporting_primitive factor_rationale: The target editor is one of two visually similar instances, the formatting controls are conditional on text selection, and the panel-level Apply step forces the agent to preserve the non-target editor before committing. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 2 depth_layering: 4 feedback_dynamics: 3 semantic_observability: 3 disambiguation_load: 5 justification: This stays human-feasible but hard because the agent must activate the right editor, make two precise selections through a conditional BubbleMenu, and then commit only the intended draft. success_trigger: human_readable: - The committed content in “Incident blurb” remains `This is urgent today.` with only `urgent` bold and only `today` italicized. - The committed content in “Public blurb” remains `Routine update only.`. - The panel-level “Apply blurb” control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Incident blurb text: This is urgent today. marked_spans: - substring: urgent marks: - bold - substring: today marks: - italic forbidden_marks_outside_spans: - bold - italic - instance_label: Public blurb plaintext_equals: Routine update only. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Apply blurb require_correct_instance: true target_instance_label_or_id: Incident blurb terminal_condition: task ends when predicate holds negative_cases: - Editing “Public blurb” instead of “Incident blurb”. - Applying bold or italic to the wrong substring, or including surrounding spaces/punctuation. - Making the draft look correct but failing to click “Apply blurb”. - Changing the underlying text instead of only the formatting. expected_interaction_path: - Locate and focus “Incident blurb”. - Select `urgent`, wait for the BubbleMenu, and apply Bold. - Select `today`, reopen the BubbleMenu, and apply Italic. - Verify “Public blurb” is unchanged and click “Apply blurb”. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: verify persisted marks by substring offsets after the Apply action, and verify the non-target editor is unchanged.' - id: rich_text_editor-mantine-v2-T02 name: Floating-menu rollback snippet canonical_type: rich_text_editor implementation_source: mantine implementation_variant: floating_menu implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: editor_operation secondary_template: disclose browsergym_goal: In the rich text editor labeled “Release snippet”, use the empty-line quick insert menu to create a Heading 3 `Rollback` followed by a bulleted list with `Stop traffic` and `Restore backup`, then click “Save snippet”. The task will finish automatically when done. ui_copy: In the rich text editor labeled “Release snippet”, use the empty-line quick insert menu to create a Heading 3 `Rollback` followed by a bulleted list with `Stop traffic` and `Restore backup`, then click “Save snippet”. The task will finish automatically when done. setup_description: |- Layout uses dashboard_panel with compact spacing and medium clutter. A Mantine Card titled “Release snippet” sits beside static metric tiles and a read-only timeline. The rich text editor labeled “Release snippet” starts empty. It is configured with an empty-line FloatingMenu: when the caret is in an empty line, a quick insert menu appears with Heading and Bullet List controls. The panel footer contains “Save snippet”. The task should create a short two-block document from an empty state. scene_context: theme: light spacing: compact layout: dashboard_panel placement: off_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: discrete secondary_supporting_primitive: menu design_intent: active_factors: - layout - confirmation_model - secondary_supporting_primitive factor_rationale: 'The task pressures editor idiom knowledge rather than raw typing alone: the quick insert menu appears only in the right caret state, and the final save distinguishes a drafted structure from a committed one.' difficulty: difficulty_bucket: hard tier: L2 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 2 depth_layering: 4 feedback_dynamics: 3 semantic_observability: 3 disambiguation_load: 2 justification: This is hard because the control surface is conditional on an empty line, the target structure is exact, and the save step forces the agent to verify that a heading and bullet list—not plain paragraphs—were committed. success_trigger: human_readable: - The committed document in “Release snippet” consists of an H3 heading `Rollback` followed by a bullet list with `Stop traffic` and `Restore backup` in that order. - The `Save snippet` action has been completed. canonical_predicate: predicate_type: equals target_state: instance_label: Release snippet mode: block_structure_equals blocks: - block_type: heading level: 3 text: Rollback - block_type: bullet_list items: - Stop traffic - Restore backup saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save snippet require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Typing the right words as plain paragraphs instead of a heading plus list. - Using numbered items instead of bullets. - Creating the right draft but not clicking “Save snippet”. - Adding extra empty blocks before or after the target structure. expected_interaction_path: - Focus the empty editor and place the caret on an empty line. - Open the FloatingMenu and insert a heading block. - Enter `Rollback`, then insert a bullet list and add the two items. - Click `Save snippet` and verify the structure persisted. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: evaluate the saved document semantically as heading + bullet list, not by toolbar click history.' - id: rich_text_editor-mantine-v2-T03 name: Sticky-toolbar offscreen inline-code edit canonical_type: rich_text_editor implementation_source: mantine implementation_variant: sticky_toolbar implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: scroll_find secondary_template: editor_operation browsergym_goal: In the rich text editor labeled “How-to”, scroll within the editor to the `## Commands` section and format only `npm run build` as inline code in the sentence `Run npm run build before deployment.` Then click `Save how-to`. Do not change any other text. The task will finish automatically when done. ui_copy: In the rich text editor labeled “How-to”, scroll within the editor to the `## Commands` section and format only `npm run build` as inline code in the sentence `Run npm run build before deployment.` Then click `Save how-to`. Do not change any other text. The task will finish automatically when done. setup_description: |- Layout is nested_scroll with compact spacing and medium clutter. The page has an outer scrollable analytics column and an inner editor panel; the rich text editor is the inner scroll target. The editor labeled “How-to” is rendered inside a fixed-height Mantine ScrollArea and uses a sticky toolbar so the controls remain visible while its content scrolls. Only about 9 lines are visible at once. The `## Commands` section is offscreen at load. Initial document: # Runbook ## Summary Prepare the release carefully. Check the staging environment before proceeding. Coordinate with the on-call team if needed. ## Prerequisites Verify all feature flags are set correctly. Review the deployment checklist one more time. Confirm that monitoring dashboards are green. ## Commands Run npm run build before deployment. ## Notes Escalate if staging fails. A “Save how-to” button commits the editor state. scene_context: theme: light spacing: compact layout: nested_scroll placement: off_center scale: default instances: 1 guidance: text clutter: medium component_context: overlay_model: none internal_scroll_region: true confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - internal_scroll_region - layout - confirmation_model factor_rationale: The target line begins offscreen inside the editor rather than the page, and the final save step makes the task about precise offscreen localization plus exact local formatting rather than generic typing. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 5 target_acquisition: 3 density_choice_interference: 1 depth_layering: 3 feedback_dynamics: 3 semantic_observability: 5 disambiguation_load: 2 justification: The textual change is tiny, but it is buried in a long document inside an internal scroll region, and the agent must preserve every non-target block while adding only one inline-code mark. success_trigger: human_readable: - The committed document in `How-to` matches the target document after normalization, with only `npm run build` formatted as inline code. - The `Save how-to` control has been clicked. canonical_predicate: predicate_type: equals target_state: instance_label: How-to mode: mixed_structure_and_marks blocks: - block_type: heading level: 1 text: Runbook - block_type: heading level: 2 text: Summary - block_type: paragraph text: Prepare the release carefully. - block_type: heading level: 2 text: Commands - block_type: paragraph text: Run npm run build before deployment. marked_spans: - substring: npm run build marks: - code forbidden_marks_outside_spans: - code - block_type: heading level: 2 text: Notes - block_type: paragraph text: Escalate if staging fails. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save how-to require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Scrolling the page instead of the editor and never reaching the Commands section. - Changing the text itself instead of only adding inline-code formatting. - Formatting too much of the sentence as code. - Leaving the document in the correct draft state but not clicking `Save how-to`. expected_interaction_path: - Focus the editor and scroll within its content area until `## Commands` is visible. - Select exactly `npm run build` in the target sentence. - Apply the inline code mark with the sticky toolbar. - Click `Save how-to`. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: compare a semantic block tree so that only the intended inline-code mark may change.' - id: rich_text_editor-mantine-v2-T04 name: Exact-span link insertion in a drawer canonical_type: rich_text_editor implementation_source: mantine implementation_variant: null implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: open_overlay secondary_template: editor_operation browsergym_goal: Open the `Edit resources` drawer. In the rich text editor labeled `Resources`, make only `release notes` a link to `https://example.com/notes`. Confirm the link in the link popover, then click `Save resources`. Do not change any other text. The task will finish automatically when done. ui_copy: Open the `Edit resources` drawer. In the rich text editor labeled `Resources`, make only `release notes` a link to `https://example.com/notes`. Confirm the link in the link popover, then click `Save resources`. Do not change any other text. The task will finish automatically when done. setup_description: |- Layout is drawer_flow with compact spacing and medium clutter. Clicking “Edit resources” opens a Mantine Drawer anchored off-center. Inside the drawer is one rich text editor labeled “Resources” and a footer with “Cancel” and “Save resources”. The editor starts with one paragraph: `See the release notes for details.` A Link control in the toolbar opens a small URL popover with a field and a “Save link” action. The drawer should close only after “Save resources” is clicked. scene_context: theme: light spacing: compact layout: drawer_flow placement: off_center 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 - secondary_supporting_primitive factor_rationale: The target action requires a nested overlay inside a drawer plus a final drawer-level commit, so the task isolates selection-dependent linking and state verification in layered UI rather than plain text editing. difficulty: difficulty_bucket: hard tier: L2 axes_ratings: precision_requirement: 5 target_acquisition: 3 density_choice_interference: 2 depth_layering: 4 feedback_dynamics: 4 semantic_observability: 4 disambiguation_load: 1 justification: This is hard because the agent must select the exact substring, use a nested link popover correctly, and then complete a separate save step that commits the drawer draft. success_trigger: human_readable: - The committed content in `Resources` remains `See the release notes for details.` with only `release notes` linked to `https://example.com/notes`. - The nested link popover is no longer open, and the final `Save resources` action has been completed. canonical_predicate: predicate_type: equals target_state: instance_label: Resources mode: marked_substring text: See the release notes for details. marked_spans: - substring: release notes marks: - type: link href: https://example.com/notes saved: true overlay_open: false tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save resources require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Linking the whole sentence or the wrong substring. - Entering the right URL in the popover but not saving the drawer draft. - Typing or deleting additional text. - Leaving the nested popover open or cancelling it. expected_interaction_path: - Open the resources drawer. - Select exactly `release notes`. - Open the link popover, enter the URL, and confirm it. - Click `Save resources` to commit the drawer draft. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: validate link-mark attributes after the final drawer save, not only while the popover is open.' - id: rich_text_editor-mantine-v2-T05 name: Row-local clear formatting in the correct table row canonical_type: rich_text_editor implementation_source: mantine implementation_variant: null implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: table_operation secondary_template: clear_reset browsergym_goal: In the `Gateway` row of the table, edit only the `Runbook reply` rich text editor so that the text remains exactly `Deploy now.` but all formatting is removed. Leave the `Billing` row unchanged, then click `Save` in the Gateway row. The task will finish automatically when done. ui_copy: In the `Gateway` row of the table, edit only the `Runbook reply` rich text editor so that the text remains exactly `Deploy now.` but all formatting is removed. Leave the `Billing` row unchanged, then click `Save` in the Gateway row. The task will finish automatically when done. setup_description: |- Layout is table_cell with compact spacing, small controls, and medium clutter. A dense Mantine Table shows two rows: “Gateway” and “Billing”. Each row contains a compact rich text editor in the “Runbook reply” column and a row-local Save button. Target row initial content: - Gateway: the text `Deploy now.` with `Deploy` bold and `now` italic. - Billing: the text `Hold release.` with no requested changes. The task is row-scoped: only the Gateway row should change, and only its row-local Save should be used. scene_context: theme: light spacing: compact layout: table_cell placement: off_center scale: small instances: 2 guidance: text clutter: medium component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: discrete secondary_supporting_primitive: table design_intent: active_factors: - layout - instances - confirmation_model factor_rationale: 'The difficulty comes from row-level grounding and row-level save semantics: the agent must target the correct compact embedded editor, remove formatting without changing the text, and avoid committing the wrong row.' difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 2 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 3 disambiguation_load: 5 justification: Dense row-local editors plus separate row saves create high wrong-row risk, and clear-formatting must preserve the exact text rather than retyping a near-match. success_trigger: human_readable: - The Gateway row’s committed text is exactly `Deploy now.` with no inline formatting marks remaining. - The Billing row’s text remains `Hold release.` unchanged. - The Gateway row’s Save control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Gateway text: Deploy now. forbidden_marks: - bold - italic - underline - strike - highlight - code - link - instance_label: Billing plaintext_equals: Hold release. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: save-gateway-row require_correct_instance: true target_instance_label_or_id: Gateway terminal_condition: task ends when predicate holds negative_cases: - Clearing formatting in Billing instead of Gateway. - Changing the words or punctuation in `Deploy now.`. - Leaving any formatting marks on either word in Gateway. - Reaching the correct state in draft but clicking the wrong row’s Save or not saving at all. expected_interaction_path: - Locate the Gateway row and focus its embedded editor. - Select the full text and clear formatting. - Verify Billing is unchanged. - Click the Gateway row’s Save control. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: bind the save control and editor state to the row key rather than DOM order.' - id: rich_text_editor-mantine-v2-T06 name: Reconstruct a visual reference card and save canonical_type: rich_text_editor implementation_source: mantine implementation_variant: null implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: match_reference secondary_template: null browsergym_goal: In the rich text editor labeled `Sprint brief`, make the content match the visual Example exactly, then click `Save brief`. The task will finish automatically when done. ui_copy: In the rich text editor labeled `Sprint brief`, make the content match the visual Example exactly, then click `Save brief`. The task will finish automatically when done. setup_description: |- Layout is dashboard_panel with compact spacing and high clutter. A Mantine Card titled “Sprint brief” sits beside static charts, KPI chips, and an activity rail. The target editor starts empty. Next to the editor is a read-only Example card showing the target rich text appearance only: a centered Heading 2 `Sprint 12`, an ordered list with `Fix bugs` and `Polish UI`, and a final italic paragraph `Last updated: Feb 3`. A `Save brief` button commits the editor state. scene_context: theme: light spacing: compact layout: dashboard_panel placement: top_right scale: default instances: 1 guidance: visual clutter: high component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - guidance - layout - confirmation_model factor_rationale: The target is specified by rendered appearance rather than source text, so the agent must map visual structure onto the editor’s block and mark idioms, then explicitly commit the result in a cluttered dashboard scene. difficulty: difficulty_bucket: hard tier: L2 axes_ratings: precision_requirement: 5 target_acquisition: 3 density_choice_interference: 2 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 2 justification: 'This is hard because the agent cannot just copy a visible source string: it must infer heading level, alignment, list type, and italic scope from the reference card and then save the result.' success_trigger: human_readable: - 'The committed content in `Sprint brief` matches the Example card semantically: centered H2 `Sprint 12`, an ordered list [`Fix bugs`, `Polish UI`], and an italic paragraph `Last updated: Feb 3`.' - The `Save brief` control has been clicked. canonical_predicate: predicate_type: matches_reference target_state: instance_label: Sprint brief mode: block_structure_equals blocks: - block_type: heading level: 2 text: Sprint 12 attributes: textAlign: center - block_type: ordered_list items: - Fix bugs - Polish UI - block_type: paragraph text: 'Last updated: Feb 3' marks: - italic saved: true tolerance: normalize_whitespace: true allow_equivalent_markup: true require_confirm: true confirm_control: Save brief require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Using bullets instead of an ordered list. - Applying the italic mark to only part of the final line or the wrong line. - Creating the right blocks but not centering the heading. - Matching the reference in draft but not clicking `Save brief`. expected_interaction_path: - Inspect the Example card and infer the target block structure. - Build the heading, ordered list, and italic final line in the editor. - Verify the heading is centered. - Click `Save brief`. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: compare semantic block structure rather than pixels.' - id: rich_text_editor-mantine-v2-T07 name: Three-step carryover formatting with save canonical_type: rich_text_editor implementation_source: mantine implementation_variant: null implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: editor_operation secondary_template: null browsergym_goal: In the rich text editor labeled `Reminder`, leave the text exactly as `Please review the Q1 plan today.` but highlight `Q1`, italicize `today`, and format `plan` as inline code. Then click `Apply reminder`. The task will finish automatically when done. ui_copy: In the rich text editor labeled `Reminder`, leave the text exactly as `Please review the Q1 plan today.` but highlight `Q1`, italicize `today`, and format `plan` as inline code. Then click `Apply reminder`. The task will finish automatically when done. setup_description: |- Layout is settings_panel with compact spacing and high clutter. The right panel contains one rich text editor labeled “Reminder” beneath unrelated toggles, segmented controls, and static status chips. Initial content is a single paragraph: `Please review the Q1 plan today.` The toolbar includes Highlight, Italic, and Inline Code controls. A footer button labeled “Apply reminder” commits the change. scene_context: theme: light spacing: compact layout: settings_panel placement: off_center scale: small instances: 1 guidance: text clutter: high component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - spacing - clutter - confirmation_model factor_rationale: The pressure comes from maintaining selection state across three different formatting operations in a dense panel, then verifying that the exact text—not a rewritten approximation—was committed. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 5 target_acquisition: 3 density_choice_interference: 2 depth_layering: 3 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 2 justification: This is hard because three precise substring operations must be sequenced without drifting the selection or adding extra marks, and the final apply step punishes stopping after only the draft looks right. success_trigger: human_readable: - The committed text remains exactly `Please review the Q1 plan today.`. - Only `Q1` is highlighted, only `plan` is inline code, and only `today` is italicized. - The `Apply reminder` control has been clicked. canonical_predicate: predicate_type: equals target_state: instance_label: Reminder mode: multiple_marked_substrings text: Please review the Q1 plan today. marked_spans: - substring: Q1 marks: - highlight - substring: plan marks: - code - substring: today marks: - italic forbidden_marks_outside_spans: - highlight - code - italic saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Apply reminder require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Selecting the wrong substrings or including surrounding spaces/punctuation. - Applying one or two of the requested marks but not all three. - Rewriting the sentence instead of preserving it exactly. - Reaching the right draft state but not clicking `Apply reminder`. expected_interaction_path: - Select `Q1` and apply Highlight. - Select `plan` and apply Inline Code. - Select `today` and apply Italic. - Click `Apply reminder`. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: verify mark coverage by exact substring offsets after the apply action.' - id: rich_text_editor-mantine-v2-T08 name: Clear formatting in the correct editor among four canonical_type: rich_text_editor implementation_source: mantine implementation_variant: null implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: clear_reset secondary_template: null browsergym_goal: In the high-contrast `Partner communications` panel, edit only the rich text editor labeled `Partner summary` so that the sentence remains exactly `Beta is available now.` but all formatting is removed. Do not edit the other three editors. Then click `Save panel`. The task will finish automatically when done. ui_copy: In the high-contrast `Partner communications` panel, edit only the rich text editor labeled `Partner summary` so that the sentence remains exactly `Beta is available now.` but all formatting is removed. Do not edit the other three editors. Then click `Save panel`. The task will finish automatically when done. setup_description: |- Layout is settings_panel with high-contrast theme, compact spacing, small controls, and high clutter. Inside a Mantine Card titled “Partner communications” are four vertically stacked rich text editors: `Public summary`, `Partner summary`, `Rollback note`, and `Internal escalation`. The target editor is `Partner summary`. Its underlying text is `Beta is available now.` but `Beta` is bold and `available` is italic in the initial state. The other three editors contain distinct notes and must remain unchanged. A shared `Save panel` button commits the whole panel. scene_context: theme: high_contrast spacing: compact layout: settings_panel placement: bottom_right scale: small instances: 4_plus 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 - theme - clutter factor_rationale: 'This is a deliberate disambiguation-heavy task: four same-type editors are packed into one high-pressure panel, and the only successful path is clearing formatting in exactly one of them while preserving the rest through a shared save.' difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 3 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 5 justification: The text operation itself is simple, but the panel is intentionally dense and high-contrast, so the hard part is targeting the correct editor and preserving all non-target state through a shared commit. success_trigger: human_readable: - The committed content in `Partner summary` is exactly `Beta is available now.` with no inline marks remaining. - The committed contents of the other three editors are unchanged. - The `Save panel` control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Partner summary text: Beta is available now. forbidden_marks: - bold - italic - underline - strike - highlight - code - link - instance_label: Public summary plaintext_equals: No breaking changes. - instance_label: Rollback note plaintext_equals: Draft the rollback memo. - instance_label: Internal escalation plaintext_equals: Notify the on-call manager. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save panel require_correct_instance: true target_instance_label_or_id: Partner summary terminal_condition: task ends when predicate holds negative_cases: - Clearing or editing the wrong editor. - Changing the words or punctuation in `Partner summary`. - Leaving any formatting mark in the target editor. - Making the draft correct but failing to click `Save panel`. expected_interaction_path: - Locate the `Partner summary` editor among the four stacked editors. - Select all text in that editor and clear formatting. - Confirm the other three editors still contain their original text. - Click `Save panel`. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: evaluate non-target editors as invariants so accidental collateral edits fail cleanly.' - id: rich_text_editor-mantine-v2-T09 name: Publish after mixed block-plus-inline formatting canonical_type: rich_text_editor implementation_source: mantine implementation_variant: null implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: confirm_cancel secondary_template: editor_operation browsergym_goal: Open `Edit post`. In the rich text editor labeled `Post body`, format the first line `Launch` as a centered Heading 2 and format only `npm install` as inline code in the second line. Then click `Publish` and `Confirm publish`. The task will finish automatically when done. ui_copy: Open `Edit post`. In the rich text editor labeled `Post body`, format the first line `Launch` as a centered Heading 2 and format only `npm install` as inline code in the second line. Then click `Publish` and `Confirm publish`. The task will finish automatically when done. setup_description: |- Layout is modal_flow with compact spacing and low clutter. Clicking `Edit post` opens a Mantine Modal containing one rich text editor labeled `Post body`, initially prefilled with two plain paragraphs: 1) `Launch` 2) `Run npm install before release.` The modal footer has `Cancel` and `Publish`. Clicking `Publish` opens a second confirmation overlay with `Back` and `Confirm publish`. The post is committed only after the final confirm step. 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 - precision_surface factor_rationale: The task combines exact rich-text semantics with a two-stage commit boundary, so the agent must distinguish a correct draft from a correctly published result. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 5 target_acquisition: 3 density_choice_interference: 1 depth_layering: 5 feedback_dynamics: 5 semantic_observability: 4 disambiguation_load: 2 justification: The editor work is already multi-step, and the stacked confirmation overlay adds another point of failure for agents that stop once the modal draft looks correct. success_trigger: human_readable: - The committed `Post body` contains a centered H2 `Launch` followed by a paragraph where only `npm install` is inline code. - The `Publish` action and then `Confirm publish` action have both been completed. - The overlays are closed at terminal state. canonical_predicate: predicate_type: equals target_state: instance_label: Post body mode: mixed_structure_and_marks blocks: - block_type: heading level: 2 text: Launch attributes: textAlign: center - block_type: paragraph text: Run npm install before release. marked_spans: - substring: npm install marks: - code forbidden_marks_outside_spans: - code saved: true overlay_open: false tolerance: normalize_whitespace: true require_confirm: true confirm_control: Confirm publish require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Formatting the heading or inline code incorrectly. - Clicking `Publish` but not completing `Confirm publish`. - Creating the correct structure in draft and then cancelling or backing out. - Turning the whole second line into a code block instead of inline code. expected_interaction_path: - Open the edit modal. - Format `Launch` as centered H2. - Select exactly `npm install` and apply inline code. - Click `Publish`, then `Confirm publish`. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: key success off the committed post preview/state after the final confirmation, not the modal draft.' - id: rich_text_editor-mantine-v2-T10 name: Append an exact final block at the true end canonical_type: rich_text_editor implementation_source: mantine implementation_variant: sticky_toolbar implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: scroll_find secondary_template: editor_operation browsergym_goal: 'In the rich text editor labeled `Changelog`, scroll to the true end of the document and append a final paragraph `Checksum: `sha256-ready`` with `sha256-ready` formatted as inline code. Do not change any earlier content. Then click `Save changelog`. The task will finish automatically when done.' ui_copy: 'In the rich text editor labeled `Changelog`, scroll to the true end of the document and append a final paragraph `Checksum: `sha256-ready`` with `sha256-ready` formatted as inline code. Do not change any earlier content. Then click `Save changelog`. The task will finish automatically when done.' setup_description: |- Layout is nested_scroll with compact spacing and medium clutter. The rich text editor labeled `Changelog` sits in a fixed-height panel with an internal scrollbar and a sticky toolbar. A separate page-level sidebar contains unrelated release metrics. The document is long enough that its true end is offscreen at load. The last existing block is a paragraph `Monitoring continues overnight.` The footer button `Save changelog` commits the result. scene_context: theme: light spacing: compact layout: nested_scroll placement: bottom_right scale: default instances: 1 guidance: text clutter: medium component_context: overlay_model: none internal_scroll_region: true confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - internal_scroll_region - layout - confirmation_model factor_rationale: The task is hard because the insertion point is the real end of a long editor, not the visible viewport, and the final save step forces preservation of the existing document prefix. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 3 density_choice_interference: 1 depth_layering: 3 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 2 justification: Agents often append in the wrong place in long editors; here they also need exact inline-code formatting and a final commit without disturbing the existing prefix. success_trigger: human_readable: - 'The committed `Changelog` document preserves its original prefix and ends with a final paragraph `Checksum: sha256-ready` where only `sha256-ready` is inline code.' - The `Save changelog` control has been clicked. canonical_predicate: predicate_type: equals target_state: instance_label: Changelog mode: ends_with_blocks preserve_prefix: true final_blocks: - block_type: paragraph text: 'Checksum: sha256-ready' marked_spans: - substring: sha256-ready marks: - code forbidden_marks_outside_spans: - code saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save changelog require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Inserting the checksum paragraph anywhere except the true end of the document. - Appending the right words but without inline-code formatting. - Changing or deleting earlier changelog content. - Leaving the draft correct but never clicking `Save changelog`. expected_interaction_path: - Scroll within the editor to its true end. - Place the caret after the last existing block. - Append the checksum paragraph and format only the token as inline code. - Click `Save changelog`. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: compare a hash of the original prefix plus semantic inspection of the final block.' - id: rich_text_editor-mantine-v2-T11 name: Subscript and superscript in the correct note canonical_type: rich_text_editor implementation_source: mantine implementation_variant: null implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: editor_operation secondary_template: null browsergym_goal: In the review card, edit only the rich text editor labeled `Chemistry note`. Keep the text exactly `H2O uses m2 notation.`, but format the `2` in `H2O` as subscript and the `2` in `m2` as superscript. Do not edit `Meeting note`. Then click `Apply note`. The task will finish automatically when done. ui_copy: In the review card, edit only the rich text editor labeled `Chemistry note`. Keep the text exactly `H2O uses m2 notation.`, but format the `2` in `H2O` as subscript and the `2` in `m2` as superscript. Do not edit `Meeting note`. Then click `Apply note`. The task will finish automatically when done. setup_description: |- Layout is inline_surface with compact spacing, small controls, and medium clutter. A review card contains two side-by-side rich text editors labeled `Meeting note` and `Chemistry note`; only `Chemistry note` should change. Initial content in `Chemistry note` is `H2O uses m2 notation.` The toolbar includes Subscript and Superscript controls. A footer button `Apply note` commits the card state. `Meeting note` contains `Agenda locked.` and must remain unchanged. scene_context: theme: light spacing: compact layout: inline_surface placement: off_center scale: small instances: 2 guidance: text clutter: medium 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 - scale - confirmation_model factor_rationale: 'This is a localized but high-precision semantics task: two identical tiny editors sit side by side, and success depends on applying two different marks to two different occurrences of the same character before committing.' difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 5 target_acquisition: 4 density_choice_interference: 2 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 5 justification: The text does not change, so the only visible evidence is mark state; that makes exact occurrence tracking and correct-instance disambiguation substantially harder than ordinary typing. success_trigger: human_readable: - The committed text in `Chemistry note` remains exactly `H2O uses m2 notation.` with the first `2` subscripted and the second `2` superscripted. - The committed text in `Meeting note` remains `Agenda locked.`. - The `Apply note` control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Chemistry note text: H2O uses m2 notation. marked_occurrences: - substring: '2' occurrence_index: 1 marks: - subscript - substring: '2' occurrence_index: 2 marks: - superscript forbidden_marks_outside_occurrences: - subscript - superscript - instance_label: Meeting note plaintext_equals: Agenda locked. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Apply note require_correct_instance: true target_instance_label_or_id: Chemistry note terminal_condition: task ends when predicate holds negative_cases: - Applying both marks to the same `2`. - Applying subscript or superscript in the wrong editor. - Changing the text instead of only the mark state. - Leaving the right draft in place but failing to click `Apply note`. expected_interaction_path: - Focus `Chemistry note`. - Select the `2` in `H2O` and apply Subscript. - Select the `2` in `m2` and apply Superscript. - Verify `Meeting note` is unchanged and click `Apply note`. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: mark occurrence indices explicitly so the two identical characters are not conflated.' - id: rich_text_editor-mantine-v2-T12 name: Bubble-menu formatting in the correct service row canonical_type: rich_text_editor implementation_source: mantine implementation_variant: bubble_menu implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: table_operation secondary_template: disclose browsergym_goal: In the `Auth` row of the table, use the selection bubble menu to make only `blocked` bold in the sentence `Deploy is blocked today.`. Leave the `Billing` row unchanged, then click `Save` in the Auth row. The task will finish automatically when done. ui_copy: In the `Auth` row of the table, use the selection bubble menu to make only `blocked` bold in the sentence `Deploy is blocked today.`. Leave the `Billing` row unchanged, then click `Save` in the Auth row. The task will finish automatically when done. setup_description: |- Layout is table_cell with compact spacing, small controls, and high clutter. A compact Mantine Table shows two service rows, `Auth` and `Billing`. Each row contains a one-line rich text editor and a row-local Save button. The target row `Auth` starts with `Deploy is blocked today.`. The editor has no full toolbar; selecting text reveals a small BubbleMenu with Bold and Italic. The `Billing` row starts with `Ready to ship.` and must remain unchanged. scene_context: theme: light spacing: compact layout: table_cell placement: bottom_left scale: small 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: table design_intent: active_factors: - layout - instances - secondary_supporting_primitive factor_rationale: 'This combines two of the strongest v1 pressure sources: compact embedded editors inside a table row and a conditional BubbleMenu that only appears after a correct text selection.' difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 5 density_choice_interference: 2 depth_layering: 4 feedback_dynamics: 3 semantic_observability: 3 disambiguation_load: 5 justification: The agent must first enter the correct row-local editor, then make a precise substring selection under dense controls, then save the correct row and leave the sibling row unchanged. success_trigger: human_readable: - The committed `Auth` row text remains `Deploy is blocked today.` with only `blocked` bolded. - The committed `Billing` row text remains `Ready to ship.` unchanged. - The `Auth` row’s Save control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Auth text: Deploy is blocked today. marked_spans: - substring: blocked marks: - bold forbidden_marks_outside_spans: - bold - instance_label: Billing plaintext_equals: Ready to ship. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: save-auth-row require_correct_instance: true target_instance_label_or_id: Auth terminal_condition: task ends when predicate holds negative_cases: - Applying the bold mark in `Billing` instead of `Auth`. - Bolding extra spaces or punctuation around `blocked`. - Editing the right row but clicking the wrong Save button. - Changing the text content instead of only the formatting. expected_interaction_path: - Locate the `Auth` row editor. - Select exactly `blocked` to reveal the BubbleMenu. - Apply Bold and verify `Billing` is unchanged. - Click the `Auth` row’s Save control. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: bind save semantics to the row key and verify non-target row invariance.' - id: rich_text_editor-mantine-v2-T13 name: High-contrast visual quote match with sibling preservation canonical_type: rich_text_editor implementation_source: mantine implementation_variant: null implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: match_reference secondary_template: null browsergym_goal: 'In the high-contrast quotes panel, edit only the rich text editor labeled `Customer-facing quote` so that it matches the Example exactly: a blockquote `Stay curious` followed by a normal paragraph `— Ada`. Do not edit the other two editors. Then click `Apply quotes`. The task will finish automatically when done.' ui_copy: 'In the high-contrast quotes panel, edit only the rich text editor labeled `Customer-facing quote` so that it matches the Example exactly: a blockquote `Stay curious` followed by a normal paragraph `— Ada`. Do not edit the other two editors. Then click `Apply quotes`. The task will finish automatically when done.' setup_description: |- Layout is settings_panel with high-contrast theme, compact spacing, and high clutter. Three rich text editors are stacked in a card: `Customer-facing quote`, `Internal quote`, and `Legal blurb`. Next to them is a small read-only Example card showing the target appearance only: a blockquote `Stay curious` followed by a normal paragraph `— Ada`. Only `Customer-facing quote` should be edited. A shared `Apply quotes` button commits all three drafts. scene_context: theme: high_contrast spacing: compact layout: settings_panel placement: off_center scale: small instances: 3 guidance: visual clutter: high component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - guidance - instances - theme factor_rationale: The target is purely visual and one of three similar editors must be chosen, so the agent has to map rendered semantics to editor structure while preserving sibling state in a visually dense panel. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 3 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 5 justification: This is hard because it combines visual-only structural inference with multi-instance disambiguation and a shared panel-level commit. success_trigger: human_readable: - 'The committed content in `Customer-facing quote` semantically matches the Example card: a blockquote `Stay curious` followed by a paragraph `— Ada`.' - The committed contents in `Internal quote` and `Legal blurb` remain unchanged. - The `Apply quotes` action has been completed. canonical_predicate: predicate_type: matches_reference target_state: mode: multi_instance targets: - instance_label: Customer-facing quote reference_semantics: blocks: - block_type: blockquote text: Stay curious - block_type: paragraph text: — Ada - instance_label: Internal quote plaintext_equals: Do not publish yet. - instance_label: Legal blurb plaintext_equals: Awaiting approval. saved: true tolerance: normalize_whitespace: true allow_equivalent_markup: true require_confirm: true confirm_control: Apply quotes require_correct_instance: true target_instance_label_or_id: Customer-facing quote terminal_condition: task ends when predicate holds negative_cases: - Editing `Internal quote` or `Legal blurb` instead of the customer-facing editor. - Using italics or plain text where a blockquote is required. - Producing the right draft in the target editor but not clicking `Apply quotes`. - Changing any non-target editor. expected_interaction_path: - Inspect the Example card. - Focus `Customer-facing quote` and create the blockquote plus paragraph structure. - Verify the other two editors are unchanged. - Click `Apply quotes`. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: use semantic reference matching plus non-target invariants.' - id: rich_text_editor-mantine-v2-T14 name: Two-editor drawer with exact heading-divider-list structure canonical_type: rich_text_editor implementation_source: mantine implementation_variant: null implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: open_overlay secondary_template: editor_operation browsergym_goal: Open the `Checklist drawer`. In the rich text editor labeled `Deployment note` (not `Internal note`), make the content exactly an H3 `Risks`, then a horizontal divider, then a bulleted list with `API drift` and `Manual rollback`. Then click `Save drawer`. The task will finish automatically when done. ui_copy: Open the `Checklist drawer`. In the rich text editor labeled `Deployment note` (not `Internal note`), make the content exactly an H3 `Risks`, then a horizontal divider, then a bulleted list with `API drift` and `Manual rollback`. Then click `Save drawer`. The task will finish automatically when done. setup_description: |- Layout is drawer_flow with compact spacing and medium clutter. Clicking `Open checklist drawer` opens a Mantine Drawer containing two rich text editors labeled `Deployment note` and `Internal note`. The target is `Deployment note`. The target editor starts empty. The toolbar includes Heading controls, Horizontal Rule, and Bullet List. The drawer footer contains `Cancel` and `Save drawer`. `Internal note` starts with `Leave untouched.` and must remain unchanged. scene_context: theme: light spacing: compact layout: drawer_flow 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: drawer design_intent: active_factors: - layout - instances - confirmation_model factor_rationale: The task couples drawer-local save semantics and same-type disambiguation inside the drawer, while the target itself requires exact multi-block structure rather than plain text. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 3 density_choice_interference: 2 depth_layering: 3 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 5 justification: This is hard because the agent must open the drawer, build an exact three-block structure in the right editor, preserve the sibling editor, and then commit through the drawer footer. success_trigger: human_readable: - The committed `Deployment note` editor contains exactly an H3 `Risks`, a horizontal rule, and a bullet list [`API drift`, `Manual rollback`]. - The committed `Internal note` remains `Leave untouched.` unchanged. - The drawer-local `Save drawer` control has been clicked and the drawer is closed at terminal state. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Deployment note mode: block_structure_equals blocks: - block_type: heading level: 3 text: Risks - block_type: horizontal_rule - block_type: bullet_list items: - API drift - Manual rollback - instance_label: Internal note plaintext_equals: Leave untouched. saved: true overlay_open: false tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save drawer require_correct_instance: true target_instance_label_or_id: Deployment note terminal_condition: task ends when predicate holds negative_cases: - Editing `Internal note` instead of `Deployment note`. - Using a paragraph or blockquote instead of a horizontal divider. - Building the right draft but not saving the drawer. - Changing the target list items or their order. expected_interaction_path: - Open the drawer. - Focus `Deployment note`. - Insert the H3 heading, horizontal rule, and bullet list in order. - Verify `Internal note` is unchanged and click `Save drawer`. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: evaluate saved block structure after the drawer closes.' - id: rich_text_editor-mantine-v2-T15 name: Below-the-fold signature update with exact bold second line canonical_type: rich_text_editor implementation_source: mantine implementation_variant: null implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: scroll_find secondary_template: editor_operation browsergym_goal: 'On the Settings page, scroll to the rich text editor labeled `Email signature`. Change it so it becomes exactly two lines: `Thanks,` on the first line and `Alex` in bold on the second line. Then click `Save signature`. The task will finish automatically when done.' ui_copy: 'On the Settings page, scroll to the rich text editor labeled `Email signature`. Change it so it becomes exactly two lines: `Thanks,` on the first line and `Alex` in bold on the second line. Then click `Save signature`. The task will finish automatically when done.' setup_description: |- Layout is settings_panel with compact spacing and high clutter. Above the fold are unrelated toggles, dropdowns, and notification preferences. The rich text editor labeled `Email signature` is below the fold. The editor initially contains two plain paragraphs: `Best,` and `Sam`. A `Save signature` button commits changes for that section only. The task should leave the first line plain and make the second line exactly `Alex` in bold. scene_context: theme: dark spacing: compact layout: settings_panel placement: center scale: default instances: 1 guidance: text clutter: high component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - layout - clutter - confirmation_model factor_rationale: The target editor is below the fold in a busy settings page, and the final state mixes exact line breaks with a mark on only the second line, so the task stresses both scroll localization and exact structural verification. difficulty: difficulty_bucket: hard tier: L2 axes_ratings: precision_requirement: 4 target_acquisition: 3 density_choice_interference: 1 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 2 justification: The editor is easy to manipulate once found, but the hard part is finding the correct section, preserving exact line structure, and applying bold only to the second line before saving. success_trigger: human_readable: - 'The committed `Email signature` document consists of exactly two paragraphs: `Thanks,` and a bold `Alex`.' - The `Save signature` control has been clicked. canonical_predicate: predicate_type: equals target_state: instance_label: Email signature mode: block_structure_equals blocks: - block_type: paragraph text: Thanks, - block_type: paragraph text: Alex marks: - bold saved: true tolerance: normalize_whitespace: false normalize_line_endings: true ignore_trailing_newline: true require_confirm: true confirm_control: Save signature require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Creating the right words with the wrong line breaks. - Leaving `Alex` plain or bolding both lines. - Editing the wrong settings field instead of the signature editor. - Leaving the section in the right draft state but not clicking `Save signature`. expected_interaction_path: - Scroll to the signature section. - Replace the existing two lines with `Thanks,` and `Alex`. - Bold only the second line. - Click `Save signature`. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: compare semantic paragraphs plus mark state, allowing only a trailing editor newline if the implementation emits one.' - id: rich_text_editor-mantine-v2-T16 name: Correct reply editor with list-plus-closing and send canonical_type: rich_text_editor implementation_source: mantine implementation_variant: null implementation_component: 'Mantine: @mantine/tiptap RichTextEditor' task_template: editor_operation secondary_template: null browsergym_goal: In the support card, edit only the rich text editor labeled `Customer reply`. Make it an ordered list with `Acknowledge issue` and `Ship fix`, followed by an italic paragraph `— Support team`. Do not edit `Internal note` or `Legal note`. Then click `Send reply`. The task will finish automatically when done. ui_copy: In the support card, edit only the rich text editor labeled `Customer reply`. Make it an ordered list with `Acknowledge issue` and `Ship fix`, followed by an italic paragraph `— Support team`. Do not edit `Internal note` or `Legal note`. Then click `Send reply`. The task will finish automatically when done. setup_description: |- Layout is inline_surface with compact spacing, medium clutter, and three same-size editors stacked in one support card: `Customer reply`, `Internal note`, and `Legal note`. Only `Customer reply` should change. The target card footer contains `Send reply`. Initial contents: - Customer reply: empty - Internal note: `Remember to tag billing.` - Legal note: `Do not promise compensation yet.` The rich text toolbar includes Ordered List and Italic. scene_context: theme: light spacing: compact layout: inline_surface placement: off_center scale: small instances: 3 guidance: text clutter: medium component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - instances - confirmation_model - layout factor_rationale: The agent has to choose the only editable target among three same-type editors, build a mixed list-plus-italic structure, preserve the two non-target notes, and then commit through a card-local send action. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 2 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 5 justification: 'The structure is short but exact, and the real challenge is disambiguation plus preservation: one wrong focus or one collateral edit invalidates the task even if the list itself is correct.' success_trigger: human_readable: - The committed `Customer reply` editor contains an ordered list [`Acknowledge issue`, `Ship fix`] followed by an italic paragraph `— Support team`. - The committed `Internal note` and `Legal note` remain unchanged. - The `Send reply` control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Customer reply mode: block_structure_equals blocks: - block_type: ordered_list items: - Acknowledge issue - Ship fix - block_type: paragraph text: — Support team marks: - italic - instance_label: Internal note plaintext_equals: Remember to tag billing. - instance_label: Legal note plaintext_equals: Do not promise compensation yet. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Send reply require_correct_instance: true target_instance_label_or_id: Customer reply terminal_condition: task ends when predicate holds negative_cases: - Editing `Internal note` or `Legal note` instead of `Customer reply`. - Using bullets instead of an ordered list. - Failing to italicize the closing line or including it inside the list. - Reaching the right draft state but not clicking `Send reply`. expected_interaction_path: - Focus `Customer reply`. - Create the ordered list with the two items. - Add a final paragraph `— Support team` and italicize it. - Verify the other two editors are unchanged and click `Send reply`. notes: 'Coverage note: uses Mantine’s native @mantine/tiptap RichTextEditor. Checker hint: evaluate target/non-target invariants together so wrong-instance edits fail deterministically.' - id: rich_text_editor-mui-v2-T01 name: Bubble menu carryover in the correct blurb editor canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap_bubble implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: editor_operation secondary_template: disclose browsergym_goal: In the “Service blurbs” panel, edit only the rich text editor labeled “Incident blurb”. Make `urgent` bold and `today` italic in the sentence `This is urgent today.`, then click “Apply blurb”. Do not edit “Public blurb”. The task will finish automatically when done. ui_copy: In the “Service blurbs” panel, edit only the rich text editor labeled “Incident blurb”. Make `urgent` bold and `today` italic in the sentence `This is urgent today.`, then click “Apply blurb”. Do not edit “Public blurb”. The task will finish automatically when done. setup_description: |- Layout uses settings_panel with compact spacing, small controls, and medium clutter. A MUI Paper/Card titled “Service blurbs” contains two same-size rich text editors labeled “Public blurb” and “Incident blurb”. The target is “Incident blurb”. The editors are configured without a full toolbar. Instead, selecting text reveals a small BubbleMenu near the selection with Bold and Italic controls. The panel footer contains a single “Apply blurb” button that commits both drafts. Initial contents: - Public blurb: `Routine update only.` - Incident blurb: `This is urgent today.` Only the Incident blurb should change. scene_context: theme: dark spacing: compact layout: settings_panel 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: popover design_intent: active_factors: - instances - confirmation_model - secondary_supporting_primitive factor_rationale: The target editor is one of two visually similar instances, the formatting controls are conditional on text selection, and the panel-level Apply step forces the agent to preserve the non-target editor before committing. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 2 depth_layering: 4 feedback_dynamics: 3 semantic_observability: 3 disambiguation_load: 5 justification: This stays human-feasible but hard because the agent must activate the right editor, make two precise selections through a conditional BubbleMenu, and then commit only the intended draft. success_trigger: human_readable: - The committed content in “Incident blurb” remains `This is urgent today.` with only `urgent` bold and only `today` italicized. - The committed content in “Public blurb” remains `Routine update only.`. - The panel-level “Apply blurb” control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Incident blurb text: This is urgent today. marked_spans: - substring: urgent marks: - bold - substring: today marks: - italic forbidden_marks_outside_spans: - bold - italic - instance_label: Public blurb plaintext_equals: Routine update only. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Apply blurb require_correct_instance: true target_instance_label_or_id: Incident blurb terminal_condition: task ends when predicate holds negative_cases: - Editing “Public blurb” instead of “Incident blurb”. - Applying bold or italic to the wrong substring, or including surrounding spaces/punctuation. - Making the draft look correct but failing to click “Apply blurb”. - Changing the underlying text instead of only the formatting. expected_interaction_path: - Locate and focus “Incident blurb”. - Select `urgent`, wait for the BubbleMenu, and apply Bold. - Select `today`, reopen the BubbleMenu, and apply Italic. - Verify “Public blurb” is unchanged and click “Apply blurb”. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: verify persisted marks by substring offsets after the Apply action, and verify the non-target editor is unchanged.' - id: rich_text_editor-mui-v2-T02 name: Floating-menu rollback snippet canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap_floating implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: editor_operation secondary_template: disclose browsergym_goal: In the rich text editor labeled “Release snippet”, use the empty-line quick insert menu to create a Heading 3 `Rollback` followed by a bulleted list with `Stop traffic` and `Restore backup`, then click “Save snippet”. The task will finish automatically when done. ui_copy: In the rich text editor labeled “Release snippet”, use the empty-line quick insert menu to create a Heading 3 `Rollback` followed by a bulleted list with `Stop traffic` and `Restore backup`, then click “Save snippet”. The task will finish automatically when done. setup_description: |- Layout uses dashboard_panel with compact spacing and medium clutter. A MUI Paper/Card titled “Release snippet” sits beside static metric tiles and a read-only timeline. The rich text editor labeled “Release snippet” starts empty. It is configured with an empty-line FloatingMenu: when the caret is in an empty line, a quick insert menu appears with Heading and Bullet List controls. The panel footer contains “Save snippet”. The task should create a short two-block document from an empty state. scene_context: theme: light spacing: compact layout: dashboard_panel placement: top_left 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: menu design_intent: active_factors: - layout - confirmation_model - secondary_supporting_primitive factor_rationale: 'The task pressures editor idiom knowledge rather than raw typing alone: the quick insert menu appears only in the right caret state, and the final save distinguishes a drafted structure from a committed one.' difficulty: difficulty_bucket: hard tier: L2 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 2 depth_layering: 4 feedback_dynamics: 3 semantic_observability: 3 disambiguation_load: 2 justification: This is hard because the control surface is conditional on an empty line, the target structure is exact, and the save step forces the agent to verify that a heading and bullet list—not plain paragraphs—were committed. success_trigger: human_readable: - The committed document in “Release snippet” consists of an H3 heading `Rollback` followed by a bullet list with `Stop traffic` and `Restore backup` in that order. - The `Save snippet` action has been completed. canonical_predicate: predicate_type: equals target_state: instance_label: Release snippet mode: block_structure_equals blocks: - block_type: heading level: 3 text: Rollback - block_type: bullet_list items: - Stop traffic - Restore backup saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save snippet require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Typing the right words as plain paragraphs instead of a heading plus list. - Using numbered items instead of bullets. - Creating the right draft but not clicking “Save snippet”. - Adding extra empty blocks before or after the target structure. expected_interaction_path: - Focus the empty editor and place the caret on an empty line. - Open the FloatingMenu and insert a heading block. - Enter `Rollback`, then insert a bullet list and add the two items. - Click `Save snippet` and verify the structure persisted. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: evaluate the saved document semantically as heading + bullet list, not by toolbar click history.' - id: rich_text_editor-mui-v2-T03 name: Sticky-toolbar offscreen inline-code edit canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap_sticky implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: scroll_find secondary_template: editor_operation browsergym_goal: In the rich text editor labeled “How-to”, scroll within the editor to the `## Commands` section and format only `npm run build` as inline code in the sentence `Run npm run build before deployment.` Then click `Save how-to`. Do not change any other text. The task will finish automatically when done. ui_copy: In the rich text editor labeled “How-to”, scroll within the editor to the `## Commands` section and format only `npm run build` as inline code in the sentence `Run npm run build before deployment.` Then click `Save how-to`. Do not change any other text. The task will finish automatically when done. setup_description: |- Layout is nested_scroll with compact spacing and medium clutter. The page has an outer scrollable analytics column and an inner editor panel; the rich text editor is the inner scroll target. The editor labeled “How-to” is rendered inside a fixed-height MUI Paper panel and uses a sticky toolbar so the controls remain visible while its content scrolls. Only about 9 lines are visible at once. The `## Commands` section is offscreen at load. Initial document: # Runbook ## Summary Prepare the release carefully. Check the staging environment before proceeding. Coordinate with the on-call team if needed. ## Prerequisites Verify all feature flags are set correctly. Review the deployment checklist one more time. Confirm that monitoring dashboards are green. ## Commands Run npm run build before deployment. ## Notes Escalate if staging fails. A “Save how-to” button commits the editor state. scene_context: theme: dark spacing: compact layout: nested_scroll placement: off_center scale: default instances: 1 guidance: text clutter: medium component_context: overlay_model: none internal_scroll_region: true confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - internal_scroll_region - layout - confirmation_model factor_rationale: The target line begins offscreen inside the editor rather than the page, and the final save step makes the task about precise offscreen localization plus exact local formatting rather than generic typing. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 5 target_acquisition: 3 density_choice_interference: 1 depth_layering: 3 feedback_dynamics: 3 semantic_observability: 5 disambiguation_load: 2 justification: The textual change is tiny, but it is buried in a long document inside an internal scroll region, and the agent must preserve every non-target block while adding only one inline-code mark. success_trigger: human_readable: - The committed document in `How-to` matches the target document after normalization, with only `npm run build` formatted as inline code. - The `Save how-to` control has been clicked. canonical_predicate: predicate_type: equals target_state: instance_label: How-to mode: mixed_structure_and_marks blocks: - block_type: heading level: 1 text: Runbook - block_type: heading level: 2 text: Summary - block_type: paragraph text: Prepare the release carefully. - block_type: heading level: 2 text: Commands - block_type: paragraph text: Run npm run build before deployment. marked_spans: - substring: npm run build marks: - code forbidden_marks_outside_spans: - code - block_type: heading level: 2 text: Notes - block_type: paragraph text: Escalate if staging fails. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save how-to require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Scrolling the page instead of the editor and never reaching the Commands section. - Changing the text itself instead of only adding inline-code formatting. - Formatting too much of the sentence as code. - Leaving the document in the correct draft state but not clicking `Save how-to`. expected_interaction_path: - Focus the editor and scroll within its content area until `## Commands` is visible. - Select exactly `npm run build` in the target sentence. - Apply the inline code mark with the sticky toolbar. - Click `Save how-to`. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: compare a semantic block tree so that only the intended inline-code mark may change.' - id: rich_text_editor-mui-v2-T04 name: Exact-span link insertion in a drawer canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: open_overlay secondary_template: editor_operation browsergym_goal: Open the `Edit resources` drawer. In the rich text editor labeled `Resources`, make only `release notes` a link to `https://example.com/notes`. Confirm the link in the link popover, then click `Save resources`. Do not change any other text. The task will finish automatically when done. ui_copy: Open the `Edit resources` drawer. In the rich text editor labeled `Resources`, make only `release notes` a link to `https://example.com/notes`. Confirm the link in the link popover, then click `Save resources`. Do not change any other text. The task will finish automatically when done. setup_description: |- Layout is drawer_flow with compact spacing and medium clutter. Clicking “Edit resources” opens a MUI Drawer anchored off-center. Inside the drawer is one rich text editor labeled “Resources” and a footer with “Cancel” and “Save resources”. The editor starts with one paragraph: `See the release notes for details.` A Link control in the toolbar opens a small URL popover with a field and a “Save link” action. The drawer should close only after “Save resources” is clicked. scene_context: theme: light spacing: compact layout: drawer_flow placement: off_center 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 - secondary_supporting_primitive factor_rationale: The target action requires a nested overlay inside a drawer plus a final drawer-level commit, so the task isolates selection-dependent linking and state verification in layered UI rather than plain text editing. difficulty: difficulty_bucket: hard tier: L2 axes_ratings: precision_requirement: 5 target_acquisition: 3 density_choice_interference: 2 depth_layering: 4 feedback_dynamics: 4 semantic_observability: 4 disambiguation_load: 1 justification: This is hard because the agent must select the exact substring, use a nested link popover correctly, and then complete a separate save step that commits the drawer draft. success_trigger: human_readable: - The committed content in `Resources` remains `See the release notes for details.` with only `release notes` linked to `https://example.com/notes`. - The nested link popover is no longer open, and the final `Save resources` action has been completed. canonical_predicate: predicate_type: equals target_state: instance_label: Resources mode: marked_substring text: See the release notes for details. marked_spans: - substring: release notes marks: - type: link href: https://example.com/notes saved: true overlay_open: false tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save resources require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Linking the whole sentence or the wrong substring. - Entering the right URL in the popover but not saving the drawer draft. - Typing or deleting additional text. - Leaving the nested popover open or cancelling it. expected_interaction_path: - Open the resources drawer. - Select exactly `release notes`. - Open the link popover, enter the URL, and confirm it. - Click `Save resources` to commit the drawer draft. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: validate link-mark attributes after the final drawer save, not only while the popover is open.' - id: rich_text_editor-mui-v2-T05 name: Row-local clear formatting in the correct table row canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: table_operation secondary_template: clear_reset browsergym_goal: In the `Gateway` row of the table, edit only the `Runbook reply` rich text editor so that the text remains exactly `Deploy now.` but all formatting is removed. Leave the `Billing` row unchanged, then click `Save` in the Gateway row. The task will finish automatically when done. ui_copy: In the `Gateway` row of the table, edit only the `Runbook reply` rich text editor so that the text remains exactly `Deploy now.` but all formatting is removed. Leave the `Billing` row unchanged, then click `Save` in the Gateway row. The task will finish automatically when done. setup_description: |- Layout is table_cell with compact spacing, small controls, and medium clutter. A dense MUI Table shows two rows: “Gateway” and “Billing”. Each row contains a compact rich text editor in the “Runbook reply” column and a row-local Save button. Target row initial content: - Gateway: the text `Deploy now.` with `Deploy` bold and `now` italic. - Billing: the text `Hold release.` with no requested changes. The task is row-scoped: only the Gateway row should change, and only its row-local Save should be used. scene_context: theme: light spacing: compact layout: table_cell placement: off_center scale: small instances: 2 guidance: text clutter: medium component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: discrete secondary_supporting_primitive: table design_intent: active_factors: - layout - instances - confirmation_model factor_rationale: 'The difficulty comes from row-level grounding and row-level save semantics: the agent must target the correct compact embedded editor, remove formatting without changing the text, and avoid committing the wrong row.' difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 2 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 3 disambiguation_load: 5 justification: Dense row-local editors plus separate row saves create high wrong-row risk, and clear-formatting must preserve the exact text rather than retyping a near-match. success_trigger: human_readable: - The Gateway row’s committed text is exactly `Deploy now.` with no inline formatting marks remaining. - The Billing row’s text remains `Hold release.` unchanged. - The Gateway row’s Save control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Gateway text: Deploy now. forbidden_marks: - bold - italic - underline - strike - highlight - code - link - instance_label: Billing plaintext_equals: Hold release. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: save-gateway-row require_correct_instance: true target_instance_label_or_id: Gateway terminal_condition: task ends when predicate holds negative_cases: - Clearing formatting in Billing instead of Gateway. - Changing the words or punctuation in `Deploy now.`. - Leaving any formatting marks on either word in Gateway. - Reaching the correct state in draft but clicking the wrong row’s Save or not saving at all. expected_interaction_path: - Locate the Gateway row and focus its embedded editor. - Select the full text and clear formatting. - Verify Billing is unchanged. - Click the Gateway row’s Save control. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: bind the save control and editor state to the row key rather than DOM order.' - id: rich_text_editor-mui-v2-T06 name: Reconstruct a visual reference card and save canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: match_reference secondary_template: null browsergym_goal: In the rich text editor labeled `Sprint brief`, make the content match the visual Example exactly, then click `Save brief`. The task will finish automatically when done. ui_copy: In the rich text editor labeled `Sprint brief`, make the content match the visual Example exactly, then click `Save brief`. The task will finish automatically when done. setup_description: |- Layout is dashboard_panel with compact spacing and high clutter. A MUI Paper/Card titled “Sprint brief” sits beside static charts, KPI chips, and an activity rail. The target editor starts empty. Next to the editor is a read-only Example card showing the target rich text appearance only: a centered Heading 2 `Sprint 12`, an ordered list with `Fix bugs` and `Polish UI`, and a final italic paragraph `Last updated: Feb 3`. A `Save brief` button commits the editor state. scene_context: theme: light spacing: compact layout: dashboard_panel placement: top_right scale: default instances: 1 guidance: visual clutter: high component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - guidance - layout - confirmation_model factor_rationale: The target is specified by rendered appearance rather than source text, so the agent must map visual structure onto the editor’s block and mark idioms, then explicitly commit the result in a cluttered dashboard scene. difficulty: difficulty_bucket: hard tier: L2 axes_ratings: precision_requirement: 5 target_acquisition: 3 density_choice_interference: 2 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 2 justification: 'This is hard because the agent cannot just copy a visible source string: it must infer heading level, alignment, list type, and italic scope from the reference card and then save the result.' success_trigger: human_readable: - 'The committed content in `Sprint brief` matches the Example card semantically: centered H2 `Sprint 12`, an ordered list [`Fix bugs`, `Polish UI`], and an italic paragraph `Last updated: Feb 3`.' - The `Save brief` control has been clicked. canonical_predicate: predicate_type: matches_reference target_state: instance_label: Sprint brief mode: block_structure_equals blocks: - block_type: heading level: 2 text: Sprint 12 attributes: textAlign: center - block_type: ordered_list items: - Fix bugs - Polish UI - block_type: paragraph text: 'Last updated: Feb 3' marks: - italic saved: true tolerance: normalize_whitespace: true allow_equivalent_markup: true require_confirm: true confirm_control: Save brief require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Using bullets instead of an ordered list. - Applying the italic mark to only part of the final line or the wrong line. - Creating the right blocks but not centering the heading. - Matching the reference in draft but not clicking `Save brief`. expected_interaction_path: - Inspect the Example card and infer the target block structure. - Build the heading, ordered list, and italic final line in the editor. - Verify the heading is centered. - Click `Save brief`. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: compare semantic block structure rather than pixels.' - id: rich_text_editor-mui-v2-T07 name: Three-step carryover formatting with save canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: editor_operation secondary_template: null browsergym_goal: In the rich text editor labeled `Reminder`, leave the text exactly as `Please review the Q1 plan today.` but highlight `Q1`, italicize `today`, and format `plan` as inline code. Then click `Apply reminder`. The task will finish automatically when done. ui_copy: In the rich text editor labeled `Reminder`, leave the text exactly as `Please review the Q1 plan today.` but highlight `Q1`, italicize `today`, and format `plan` as inline code. Then click `Apply reminder`. The task will finish automatically when done. setup_description: |- Layout is settings_panel with compact spacing and high clutter. The right panel contains one rich text editor labeled “Reminder” beneath unrelated toggles, segmented controls, and static status chips. Initial content is a single paragraph: `Please review the Q1 plan today.` The toolbar includes Highlight, Italic, and Inline Code controls. A footer button labeled “Apply reminder” commits the change. scene_context: theme: dark spacing: compact layout: settings_panel placement: off_center scale: small instances: 1 guidance: text clutter: high component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - spacing - clutter - confirmation_model factor_rationale: The pressure comes from maintaining selection state across three different formatting operations in a dense panel, then verifying that the exact text—not a rewritten approximation—was committed. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 5 target_acquisition: 3 density_choice_interference: 2 depth_layering: 3 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 2 justification: This is hard because three precise substring operations must be sequenced without drifting the selection or adding extra marks, and the final apply step punishes stopping after only the draft looks right. success_trigger: human_readable: - The committed text remains exactly `Please review the Q1 plan today.`. - Only `Q1` is highlighted, only `plan` is inline code, and only `today` is italicized. - The `Apply reminder` control has been clicked. canonical_predicate: predicate_type: equals target_state: instance_label: Reminder mode: multiple_marked_substrings text: Please review the Q1 plan today. marked_spans: - substring: Q1 marks: - highlight - substring: plan marks: - code - substring: today marks: - italic forbidden_marks_outside_spans: - highlight - code - italic saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Apply reminder require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Selecting the wrong substrings or including surrounding spaces/punctuation. - Applying one or two of the requested marks but not all three. - Rewriting the sentence instead of preserving it exactly. - Reaching the right draft state but not clicking `Apply reminder`. expected_interaction_path: - Select `Q1` and apply Highlight. - Select `plan` and apply Inline Code. - Select `today` and apply Italic. - Click `Apply reminder`. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: verify mark coverage by exact substring offsets after the apply action.' - id: rich_text_editor-mui-v2-T08 name: Clear formatting in the correct editor among four canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: clear_reset secondary_template: null browsergym_goal: In the high-contrast `Partner communications` panel, edit only the rich text editor labeled `Partner summary` so that the sentence remains exactly `Beta is available now.` but all formatting is removed. Do not edit the other three editors. Then click `Save panel`. The task will finish automatically when done. ui_copy: In the high-contrast `Partner communications` panel, edit only the rich text editor labeled `Partner summary` so that the sentence remains exactly `Beta is available now.` but all formatting is removed. Do not edit the other three editors. Then click `Save panel`. The task will finish automatically when done. setup_description: |- Layout is settings_panel with high-contrast theme, compact spacing, small controls, and high clutter. Inside a MUI Paper/Card titled “Partner communications” are four vertically stacked rich text editors: `Public summary`, `Partner summary`, `Rollback note`, and `Internal escalation`. The target editor is `Partner summary`. Its underlying text is `Beta is available now.` but `Beta` is bold and `available` is italic in the initial state. The other three editors contain distinct notes and must remain unchanged. A shared `Save panel` button commits the whole panel. scene_context: theme: high_contrast spacing: compact layout: settings_panel placement: bottom_right scale: small instances: 4_plus 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 - theme - clutter factor_rationale: 'This is a deliberate disambiguation-heavy task: four same-type editors are packed into one high-pressure panel, and the only successful path is clearing formatting in exactly one of them while preserving the rest through a shared save.' difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 3 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 5 justification: The text operation itself is simple, but the panel is intentionally dense and high-contrast, so the hard part is targeting the correct editor and preserving all non-target state through a shared commit. success_trigger: human_readable: - The committed content in `Partner summary` is exactly `Beta is available now.` with no inline marks remaining. - The committed contents of the other three editors are unchanged. - The `Save panel` control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Partner summary text: Beta is available now. forbidden_marks: - bold - italic - underline - strike - highlight - code - link - instance_label: Public summary plaintext_equals: No breaking changes. - instance_label: Rollback note plaintext_equals: Draft the rollback memo. - instance_label: Internal escalation plaintext_equals: Notify the on-call manager. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save panel require_correct_instance: true target_instance_label_or_id: Partner summary terminal_condition: task ends when predicate holds negative_cases: - Clearing or editing the wrong editor. - Changing the words or punctuation in `Partner summary`. - Leaving any formatting mark in the target editor. - Making the draft correct but failing to click `Save panel`. expected_interaction_path: - Locate the `Partner summary` editor among the four stacked editors. - Select all text in that editor and clear formatting. - Confirm the other three editors still contain their original text. - Click `Save panel`. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: evaluate non-target editors as invariants so accidental collateral edits fail cleanly.' - id: rich_text_editor-mui-v2-T09 name: Publish after mixed block-plus-inline formatting canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: confirm_cancel secondary_template: editor_operation browsergym_goal: Open `Edit post`. In the rich text editor labeled `Post body`, format the first line `Launch` as a centered Heading 2 and format only `npm install` as inline code in the second line. Then click `Publish` and `Confirm publish`. The task will finish automatically when done. ui_copy: Open `Edit post`. In the rich text editor labeled `Post body`, format the first line `Launch` as a centered Heading 2 and format only `npm install` as inline code in the second line. Then click `Publish` and `Confirm publish`. The task will finish automatically when done. setup_description: |- Layout is modal_flow with compact spacing and low clutter. Clicking `Edit post` opens a MUI Dialog containing one rich text editor labeled `Post body`, initially prefilled with two plain paragraphs: 1) `Launch` 2) `Run npm install before release.` The modal footer has `Cancel` and `Publish`. Clicking `Publish` opens a second confirmation overlay with `Back` and `Confirm publish`. The post is committed only after the final confirm step. scene_context: theme: dark 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 - precision_surface factor_rationale: The task combines exact rich-text semantics with a two-stage commit boundary, so the agent must distinguish a correct draft from a correctly published result. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 5 target_acquisition: 3 density_choice_interference: 1 depth_layering: 5 feedback_dynamics: 5 semantic_observability: 4 disambiguation_load: 2 justification: The editor work is already multi-step, and the stacked confirmation overlay adds another point of failure for agents that stop once the modal draft looks correct. success_trigger: human_readable: - The committed `Post body` contains a centered H2 `Launch` followed by a paragraph where only `npm install` is inline code. - The `Publish` action and then `Confirm publish` action have both been completed. - The overlays are closed at terminal state. canonical_predicate: predicate_type: equals target_state: instance_label: Post body mode: mixed_structure_and_marks blocks: - block_type: heading level: 2 text: Launch attributes: textAlign: center - block_type: paragraph text: Run npm install before release. marked_spans: - substring: npm install marks: - code forbidden_marks_outside_spans: - code saved: true overlay_open: false tolerance: normalize_whitespace: true require_confirm: true confirm_control: Confirm publish require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Formatting the heading or inline code incorrectly. - Clicking `Publish` but not completing `Confirm publish`. - Creating the correct structure in draft and then cancelling or backing out. - Turning the whole second line into a code block instead of inline code. expected_interaction_path: - Open the edit modal. - Format `Launch` as centered H2. - Select exactly `npm install` and apply inline code. - Click `Publish`, then `Confirm publish`. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: key success off the committed post preview/state after the final confirmation, not the modal draft.' - id: rich_text_editor-mui-v2-T10 name: Append an exact final block at the true end canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap_sticky implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: scroll_find secondary_template: editor_operation browsergym_goal: 'In the rich text editor labeled `Changelog`, scroll to the true end of the document and append a final paragraph `Checksum: `sha256-ready`` with `sha256-ready` formatted as inline code. Do not change any earlier content. Then click `Save changelog`. The task will finish automatically when done.' ui_copy: 'In the rich text editor labeled `Changelog`, scroll to the true end of the document and append a final paragraph `Checksum: `sha256-ready`` with `sha256-ready` formatted as inline code. Do not change any earlier content. Then click `Save changelog`. The task will finish automatically when done.' setup_description: |- Layout is nested_scroll with compact spacing and medium clutter. The rich text editor labeled `Changelog` sits in a fixed-height panel with an internal scrollbar and a sticky toolbar. A separate page-level sidebar contains unrelated release metrics. The document is long enough that its true end is offscreen at load. The last existing block is a paragraph `Monitoring continues overnight.` The footer button `Save changelog` commits the result. scene_context: theme: light spacing: compact layout: nested_scroll placement: bottom_right scale: default instances: 1 guidance: text clutter: medium component_context: overlay_model: none internal_scroll_region: true confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - internal_scroll_region - layout - confirmation_model factor_rationale: The task is hard because the insertion point is the real end of a long editor, not the visible viewport, and the final save step forces preservation of the existing document prefix. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 3 density_choice_interference: 1 depth_layering: 3 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 2 justification: Agents often append in the wrong place in long editors; here they also need exact inline-code formatting and a final commit without disturbing the existing prefix. success_trigger: human_readable: - 'The committed `Changelog` document preserves its original prefix and ends with a final paragraph `Checksum: sha256-ready` where only `sha256-ready` is inline code.' - The `Save changelog` control has been clicked. canonical_predicate: predicate_type: equals target_state: instance_label: Changelog mode: ends_with_blocks preserve_prefix: true final_blocks: - block_type: paragraph text: 'Checksum: sha256-ready' marked_spans: - substring: sha256-ready marks: - code forbidden_marks_outside_spans: - code saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save changelog require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Inserting the checksum paragraph anywhere except the true end of the document. - Appending the right words but without inline-code formatting. - Changing or deleting earlier changelog content. - Leaving the draft correct but never clicking `Save changelog`. expected_interaction_path: - Scroll within the editor to its true end. - Place the caret after the last existing block. - Append the checksum paragraph and format only the token as inline code. - Click `Save changelog`. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: compare a hash of the original prefix plus semantic inspection of the final block.' - id: rich_text_editor-mui-v2-T11 name: Subscript and superscript in the correct note canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: editor_operation secondary_template: null browsergym_goal: In the review card, edit only the rich text editor labeled `Chemistry note`. Keep the text exactly `H2O uses m2 notation.`, but format the `2` in `H2O` as subscript and the `2` in `m2` as superscript. Do not edit `Meeting note`. Then click `Apply note`. The task will finish automatically when done. ui_copy: In the review card, edit only the rich text editor labeled `Chemistry note`. Keep the text exactly `H2O uses m2 notation.`, but format the `2` in `H2O` as subscript and the `2` in `m2` as superscript. Do not edit `Meeting note`. Then click `Apply note`. The task will finish automatically when done. setup_description: |- Layout is inline_surface with compact spacing, small controls, and medium clutter. A review card contains two side-by-side rich text editors labeled `Meeting note` and `Chemistry note`; only `Chemistry note` should change. Initial content in `Chemistry note` is `H2O uses m2 notation.` The toolbar includes Subscript and Superscript controls. A footer button `Apply note` commits the card state. `Meeting note` contains `Agenda locked.` and must remain unchanged. scene_context: theme: light spacing: compact layout: inline_surface placement: off_center scale: small instances: 2 guidance: text clutter: medium 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 - scale - confirmation_model factor_rationale: 'This is a localized but high-precision semantics task: two identical tiny editors sit side by side, and success depends on applying two different marks to two different occurrences of the same character before committing.' difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 5 target_acquisition: 4 density_choice_interference: 2 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 5 justification: The text does not change, so the only visible evidence is mark state; that makes exact occurrence tracking and correct-instance disambiguation substantially harder than ordinary typing. success_trigger: human_readable: - The committed text in `Chemistry note` remains exactly `H2O uses m2 notation.` with the first `2` subscripted and the second `2` superscripted. - The committed text in `Meeting note` remains `Agenda locked.`. - The `Apply note` control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Chemistry note text: H2O uses m2 notation. marked_occurrences: - substring: '2' occurrence_index: 1 marks: - subscript - substring: '2' occurrence_index: 2 marks: - superscript forbidden_marks_outside_occurrences: - subscript - superscript - instance_label: Meeting note plaintext_equals: Agenda locked. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Apply note require_correct_instance: true target_instance_label_or_id: Chemistry note terminal_condition: task ends when predicate holds negative_cases: - Applying both marks to the same `2`. - Applying subscript or superscript in the wrong editor. - Changing the text instead of only the mark state. - Leaving the right draft in place but failing to click `Apply note`. expected_interaction_path: - Focus `Chemistry note`. - Select the `2` in `H2O` and apply Subscript. - Select the `2` in `m2` and apply Superscript. - Verify `Meeting note` is unchanged and click `Apply note`. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: mark occurrence indices explicitly so the two identical characters are not conflated.' - id: rich_text_editor-mui-v2-T12 name: Bubble-menu formatting in the correct service row canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap_bubble implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: table_operation secondary_template: disclose browsergym_goal: In the `Auth` row of the table, use the selection bubble menu to make only `blocked` bold in the sentence `Deploy is blocked today.`. Leave the `Billing` row unchanged, then click `Save` in the Auth row. The task will finish automatically when done. ui_copy: In the `Auth` row of the table, use the selection bubble menu to make only `blocked` bold in the sentence `Deploy is blocked today.`. Leave the `Billing` row unchanged, then click `Save` in the Auth row. The task will finish automatically when done. setup_description: |- Layout is table_cell with compact spacing, small controls, and high clutter. A compact MUI Table shows two service rows, `Auth` and `Billing`. Each row contains a one-line rich text editor and a row-local Save button. The target row `Auth` starts with `Deploy is blocked today.`. The editor has no full toolbar; selecting text reveals a small BubbleMenu with Bold and Italic. The `Billing` row starts with `Ready to ship.` and must remain unchanged. scene_context: theme: dark spacing: compact layout: table_cell placement: bottom_left scale: small 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: table design_intent: active_factors: - layout - instances - secondary_supporting_primitive factor_rationale: 'This combines two of the strongest v1 pressure sources: compact embedded editors inside a table row and a conditional BubbleMenu that only appears after a correct text selection.' difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 5 density_choice_interference: 2 depth_layering: 4 feedback_dynamics: 3 semantic_observability: 3 disambiguation_load: 5 justification: The agent must first enter the correct row-local editor, then make a precise substring selection under dense controls, then save the correct row and leave the sibling row unchanged. success_trigger: human_readable: - The committed `Auth` row text remains `Deploy is blocked today.` with only `blocked` bolded. - The committed `Billing` row text remains `Ready to ship.` unchanged. - The `Auth` row’s Save control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Auth text: Deploy is blocked today. marked_spans: - substring: blocked marks: - bold forbidden_marks_outside_spans: - bold - instance_label: Billing plaintext_equals: Ready to ship. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: save-auth-row require_correct_instance: true target_instance_label_or_id: Auth terminal_condition: task ends when predicate holds negative_cases: - Applying the bold mark in `Billing` instead of `Auth`. - Bolding extra spaces or punctuation around `blocked`. - Editing the right row but clicking the wrong Save button. - Changing the text content instead of only the formatting. expected_interaction_path: - Locate the `Auth` row editor. - Select exactly `blocked` to reveal the BubbleMenu. - Apply Bold and verify `Billing` is unchanged. - Click the `Auth` row’s Save control. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: bind save semantics to the row key and verify non-target row invariance.' - id: rich_text_editor-mui-v2-T13 name: High-contrast visual quote match with sibling preservation canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: match_reference secondary_template: null browsergym_goal: 'In the high-contrast quotes panel, edit only the rich text editor labeled `Customer-facing quote` so that it matches the Example exactly: a blockquote `Stay curious` followed by a normal paragraph `— Ada`. Do not edit the other two editors. Then click `Apply quotes`. The task will finish automatically when done.' ui_copy: 'In the high-contrast quotes panel, edit only the rich text editor labeled `Customer-facing quote` so that it matches the Example exactly: a blockquote `Stay curious` followed by a normal paragraph `— Ada`. Do not edit the other two editors. Then click `Apply quotes`. The task will finish automatically when done.' setup_description: |- Layout is settings_panel with high-contrast theme, compact spacing, and high clutter. Three rich text editors are stacked in a card: `Customer-facing quote`, `Internal quote`, and `Legal blurb`. Next to them is a small read-only Example card showing the target appearance only: a blockquote `Stay curious` followed by a normal paragraph `— Ada`. Only `Customer-facing quote` should be edited. A shared `Apply quotes` button commits all three drafts. scene_context: theme: high_contrast spacing: compact layout: settings_panel placement: off_center scale: small instances: 3 guidance: visual clutter: high component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - guidance - instances - theme factor_rationale: The target is purely visual and one of three similar editors must be chosen, so the agent has to map rendered semantics to editor structure while preserving sibling state in a visually dense panel. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 3 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 5 justification: This is hard because it combines visual-only structural inference with multi-instance disambiguation and a shared panel-level commit. success_trigger: human_readable: - 'The committed content in `Customer-facing quote` semantically matches the Example card: a blockquote `Stay curious` followed by a paragraph `— Ada`.' - The committed contents in `Internal quote` and `Legal blurb` remain unchanged. - The `Apply quotes` action has been completed. canonical_predicate: predicate_type: matches_reference target_state: mode: multi_instance targets: - instance_label: Customer-facing quote reference_semantics: blocks: - block_type: blockquote text: Stay curious - block_type: paragraph text: — Ada - instance_label: Internal quote plaintext_equals: Do not publish yet. - instance_label: Legal blurb plaintext_equals: Awaiting approval. saved: true tolerance: normalize_whitespace: true allow_equivalent_markup: true require_confirm: true confirm_control: Apply quotes require_correct_instance: true target_instance_label_or_id: Customer-facing quote terminal_condition: task ends when predicate holds negative_cases: - Editing `Internal quote` or `Legal blurb` instead of the customer-facing editor. - Using italics or plain text where a blockquote is required. - Producing the right draft in the target editor but not clicking `Apply quotes`. - Changing any non-target editor. expected_interaction_path: - Inspect the Example card. - Focus `Customer-facing quote` and create the blockquote plus paragraph structure. - Verify the other two editors are unchanged. - Click `Apply quotes`. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: use semantic reference matching plus non-target invariants.' - id: rich_text_editor-mui-v2-T14 name: Two-editor drawer with exact heading-divider-list structure canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: open_overlay secondary_template: editor_operation browsergym_goal: Open the `Checklist drawer`. In the rich text editor labeled `Deployment note` (not `Internal note`), make the content exactly an H3 `Risks`, then a horizontal divider, then a bulleted list with `API drift` and `Manual rollback`. Then click `Save drawer`. The task will finish automatically when done. ui_copy: Open the `Checklist drawer`. In the rich text editor labeled `Deployment note` (not `Internal note`), make the content exactly an H3 `Risks`, then a horizontal divider, then a bulleted list with `API drift` and `Manual rollback`. Then click `Save drawer`. The task will finish automatically when done. setup_description: |- Layout is drawer_flow with compact spacing and medium clutter. Clicking `Open checklist drawer` opens a MUI Drawer containing two rich text editors labeled `Deployment note` and `Internal note`. The target is `Deployment note`. The target editor starts empty. The toolbar includes Heading controls, Horizontal Rule, and Bullet List. The drawer footer contains `Cancel` and `Save drawer`. `Internal note` starts with `Leave untouched.` and must remain unchanged. scene_context: theme: light spacing: compact layout: drawer_flow 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: drawer design_intent: active_factors: - layout - instances - confirmation_model factor_rationale: The task couples drawer-local save semantics and same-type disambiguation inside the drawer, while the target itself requires exact multi-block structure rather than plain text. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 3 density_choice_interference: 2 depth_layering: 3 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 5 justification: This is hard because the agent must open the drawer, build an exact three-block structure in the right editor, preserve the sibling editor, and then commit through the drawer footer. success_trigger: human_readable: - The committed `Deployment note` editor contains exactly an H3 `Risks`, a horizontal rule, and a bullet list [`API drift`, `Manual rollback`]. - The committed `Internal note` remains `Leave untouched.` unchanged. - The drawer-local `Save drawer` control has been clicked and the drawer is closed at terminal state. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Deployment note mode: block_structure_equals blocks: - block_type: heading level: 3 text: Risks - block_type: horizontal_rule - block_type: bullet_list items: - API drift - Manual rollback - instance_label: Internal note plaintext_equals: Leave untouched. saved: true overlay_open: false tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save drawer require_correct_instance: true target_instance_label_or_id: Deployment note terminal_condition: task ends when predicate holds negative_cases: - Editing `Internal note` instead of `Deployment note`. - Using a paragraph or blockquote instead of a horizontal divider. - Building the right draft but not saving the drawer. - Changing the target list items or their order. expected_interaction_path: - Open the drawer. - Focus `Deployment note`. - Insert the H3 heading, horizontal rule, and bullet list in order. - Verify `Internal note` is unchanged and click `Save drawer`. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: evaluate saved block structure after the drawer closes.' - id: rich_text_editor-mui-v2-T15 name: Below-the-fold signature update with exact bold second line canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: scroll_find secondary_template: editor_operation browsergym_goal: 'On the Settings page, scroll to the rich text editor labeled `Email signature`. Change it so it becomes exactly two lines: `Thanks,` on the first line and `Alex` in bold on the second line. Then click `Save signature`. The task will finish automatically when done.' ui_copy: 'On the Settings page, scroll to the rich text editor labeled `Email signature`. Change it so it becomes exactly two lines: `Thanks,` on the first line and `Alex` in bold on the second line. Then click `Save signature`. The task will finish automatically when done.' setup_description: |- Layout is settings_panel with compact spacing and high clutter. Above the fold are unrelated toggles, dropdowns, and notification preferences. The rich text editor labeled `Email signature` is below the fold. The editor initially contains two plain paragraphs: `Best,` and `Sam`. A `Save signature` button commits changes for that section only. The task should leave the first line plain and make the second line exactly `Alex` in bold. scene_context: theme: dark spacing: compact layout: settings_panel placement: center scale: default instances: 1 guidance: text clutter: high component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - layout - clutter - confirmation_model factor_rationale: The target editor is below the fold in a busy settings page, and the final state mixes exact line breaks with a mark on only the second line, so the task stresses both scroll localization and exact structural verification. difficulty: difficulty_bucket: hard tier: L2 axes_ratings: precision_requirement: 4 target_acquisition: 3 density_choice_interference: 1 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 2 justification: The editor is easy to manipulate once found, but the hard part is finding the correct section, preserving exact line structure, and applying bold only to the second line before saving. success_trigger: human_readable: - 'The committed `Email signature` document consists of exactly two paragraphs: `Thanks,` and a bold `Alex`.' - The `Save signature` control has been clicked. canonical_predicate: predicate_type: equals target_state: instance_label: Email signature mode: block_structure_equals blocks: - block_type: paragraph text: Thanks, - block_type: paragraph text: Alex marks: - bold saved: true tolerance: normalize_whitespace: false normalize_line_endings: true ignore_trailing_newline: true require_confirm: true confirm_control: Save signature require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Creating the right words with the wrong line breaks. - Leaving `Alex` plain or bolding both lines. - Editing the wrong settings field instead of the signature editor. - Leaving the section in the right draft state but not clicking `Save signature`. expected_interaction_path: - Scroll to the signature section. - Replace the existing two lines with `Thanks,` and `Alex`. - Bold only the second line. - Click `Save signature`. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: compare semantic paragraphs plus mark state, allowing only a trailing editor newline if the implementation emits one.' - id: rich_text_editor-mui-v2-T16 name: Correct reply editor with list-plus-closing and send canonical_type: rich_text_editor implementation_source: mui implementation_variant: composite_mui_tiptap implementation_component: 'MUI: mui-tiptap rich text editor + MUI carrier components [composite]' task_template: editor_operation secondary_template: null browsergym_goal: In the support card, edit only the rich text editor labeled `Customer reply`. Make it an ordered list with `Acknowledge issue` and `Ship fix`, followed by an italic paragraph `— Support team`. Do not edit `Internal note` or `Legal note`. Then click `Send reply`. The task will finish automatically when done. ui_copy: In the support card, edit only the rich text editor labeled `Customer reply`. Make it an ordered list with `Acknowledge issue` and `Ship fix`, followed by an italic paragraph `— Support team`. Do not edit `Internal note` or `Legal note`. Then click `Send reply`. The task will finish automatically when done. setup_description: |- Layout is inline_surface with compact spacing, medium clutter, and three same-size editors stacked in one support card: `Customer reply`, `Internal note`, and `Legal note`. Only `Customer reply` should change. The target card footer contains `Send reply`. Initial contents: - Customer reply: empty - Internal note: `Remember to tag billing.` - Legal note: `Do not promise compensation yet.` The rich text toolbar includes Ordered List and Italic. scene_context: theme: light spacing: compact layout: inline_surface placement: off_center scale: small instances: 3 guidance: text clutter: medium component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - instances - confirmation_model - layout factor_rationale: The agent has to choose the only editable target among three same-type editors, build a mixed list-plus-italic structure, preserve the two non-target notes, and then commit through a card-local send action. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 2 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 5 justification: 'The structure is short but exact, and the real challenge is disambiguation plus preservation: one wrong focus or one collateral edit invalidates the task even if the list itself is correct.' success_trigger: human_readable: - The committed `Customer reply` editor contains an ordered list [`Acknowledge issue`, `Ship fix`] followed by an italic paragraph `— Support team`. - The committed `Internal note` and `Legal note` remain unchanged. - The `Send reply` control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Customer reply mode: block_structure_equals blocks: - block_type: ordered_list items: - Acknowledge issue - Ship fix - block_type: paragraph text: — Support team marks: - italic - instance_label: Internal note plaintext_equals: Remember to tag billing. - instance_label: Legal note plaintext_equals: Do not promise compensation yet. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Send reply require_correct_instance: true target_instance_label_or_id: Customer reply terminal_condition: task ends when predicate holds negative_cases: - Editing `Internal note` or `Legal note` instead of `Customer reply`. - Using bullets instead of an ordered list. - Failing to italicize the closing line or including it inside the list. - Reaching the right draft state but not clicking `Send reply`. expected_interaction_path: - Focus `Customer reply`. - Create the ordered list with the two items. - Add a final paragraph `— Support team` and italicize it. - Verify the other two editors are unchanged and click `Send reply`. notes: 'Coverage note: Material UI does not ship a core rich text editor in its main component library, so this task uses mui-tiptap inside MUI carrier components. Checker hint: evaluate target/non-target invariants together so wrong-instance edits fail deterministically.' - id: rich_text_editor-antd-v2-T01 name: Bubble menu carryover in the correct blurb editor canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap_bubble implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: editor_operation secondary_template: disclose browsergym_goal: In the “Service blurbs” panel, edit only the rich text editor labeled “Incident blurb”. Make `urgent` bold and `today` italic in the sentence `This is urgent today.`, then click “Apply blurb”. Do not edit “Public blurb”. The task will finish automatically when done. ui_copy: In the “Service blurbs” panel, edit only the rich text editor labeled “Incident blurb”. Make `urgent` bold and `today` italic in the sentence `This is urgent today.`, then click “Apply blurb”. Do not edit “Public blurb”. The task will finish automatically when done. setup_description: |- Layout uses settings_panel with compact spacing, small controls, and medium clutter. A Ant Design Card titled “Service blurbs” contains two same-size rich text editors labeled “Public blurb” and “Incident blurb”. The target is “Incident blurb”. The editors are configured without a full toolbar. Instead, selecting text reveals a small BubbleMenu near the selection with Bold and Italic controls. The panel footer contains a single “Apply blurb” button that commits both drafts. Initial contents: - Public blurb: `Routine update only.` - Incident blurb: `This is urgent today.` Only the Incident blurb should change. scene_context: theme: high_contrast spacing: compact layout: settings_panel 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: popover design_intent: active_factors: - instances - confirmation_model - secondary_supporting_primitive factor_rationale: The target editor is one of two visually similar instances, the formatting controls are conditional on text selection, and the panel-level Apply step forces the agent to preserve the non-target editor before committing. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 2 depth_layering: 4 feedback_dynamics: 3 semantic_observability: 3 disambiguation_load: 5 justification: This stays human-feasible but hard because the agent must activate the right editor, make two precise selections through a conditional BubbleMenu, and then commit only the intended draft. success_trigger: human_readable: - The committed content in “Incident blurb” remains `This is urgent today.` with only `urgent` bold and only `today` italicized. - The committed content in “Public blurb” remains `Routine update only.`. - The panel-level “Apply blurb” control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Incident blurb text: This is urgent today. marked_spans: - substring: urgent marks: - bold - substring: today marks: - italic forbidden_marks_outside_spans: - bold - italic - instance_label: Public blurb plaintext_equals: Routine update only. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Apply blurb require_correct_instance: true target_instance_label_or_id: Incident blurb terminal_condition: task ends when predicate holds negative_cases: - Editing “Public blurb” instead of “Incident blurb”. - Applying bold or italic to the wrong substring, or including surrounding spaces/punctuation. - Making the draft look correct but failing to click “Apply blurb”. - Changing the underlying text instead of only the formatting. expected_interaction_path: - Locate and focus “Incident blurb”. - Select `urgent`, wait for the BubbleMenu, and apply Bold. - Select `today`, reopen the BubbleMenu, and apply Italic. - Verify “Public blurb” is unchanged and click “Apply blurb”. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: verify persisted marks by substring offsets after the Apply action, and verify the non-target editor is unchanged.' - id: rich_text_editor-antd-v2-T02 name: Floating-menu rollback snippet canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap_floating implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: editor_operation secondary_template: disclose browsergym_goal: In the rich text editor labeled “Release snippet”, use the empty-line quick insert menu to create a Heading 3 `Rollback` followed by a bulleted list with `Stop traffic` and `Restore backup`, then click “Save snippet”. The task will finish automatically when done. ui_copy: In the rich text editor labeled “Release snippet”, use the empty-line quick insert menu to create a Heading 3 `Rollback` followed by a bulleted list with `Stop traffic` and `Restore backup`, then click “Save snippet”. The task will finish automatically when done. setup_description: |- Layout uses dashboard_panel with compact spacing and medium clutter. A Ant Design Card titled “Release snippet” sits beside static metric tiles and a read-only timeline. The rich text editor labeled “Release snippet” starts empty. It is configured with an empty-line FloatingMenu: when the caret is in an empty line, a quick insert menu appears with Heading and Bullet List controls. The panel footer contains “Save snippet”. The task should create a short two-block document from an empty state. scene_context: theme: light spacing: compact layout: dashboard_panel placement: off_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: discrete secondary_supporting_primitive: menu design_intent: active_factors: - layout - confirmation_model - secondary_supporting_primitive factor_rationale: 'The task pressures editor idiom knowledge rather than raw typing alone: the quick insert menu appears only in the right caret state, and the final save distinguishes a drafted structure from a committed one.' difficulty: difficulty_bucket: hard tier: L2 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 2 depth_layering: 4 feedback_dynamics: 3 semantic_observability: 3 disambiguation_load: 2 justification: This is hard because the control surface is conditional on an empty line, the target structure is exact, and the save step forces the agent to verify that a heading and bullet list—not plain paragraphs—were committed. success_trigger: human_readable: - The committed document in “Release snippet” consists of an H3 heading `Rollback` followed by a bullet list with `Stop traffic` and `Restore backup` in that order. - The `Save snippet` action has been completed. canonical_predicate: predicate_type: equals target_state: instance_label: Release snippet mode: block_structure_equals blocks: - block_type: heading level: 3 text: Rollback - block_type: bullet_list items: - Stop traffic - Restore backup saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save snippet require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Typing the right words as plain paragraphs instead of a heading plus list. - Using numbered items instead of bullets. - Creating the right draft but not clicking “Save snippet”. - Adding extra empty blocks before or after the target structure. expected_interaction_path: - Focus the empty editor and place the caret on an empty line. - Open the FloatingMenu and insert a heading block. - Enter `Rollback`, then insert a bullet list and add the two items. - Click `Save snippet` and verify the structure persisted. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: evaluate the saved document semantically as heading + bullet list, not by toolbar click history.' - id: rich_text_editor-antd-v2-T03 name: Sticky-toolbar offscreen inline-code edit canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap_sticky implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: scroll_find secondary_template: editor_operation browsergym_goal: In the rich text editor labeled “How-to”, scroll within the editor to the `## Commands` section and format only `npm run build` as inline code in the sentence `Run npm run build before deployment.` Then click `Save how-to`. Do not change any other text. The task will finish automatically when done. ui_copy: In the rich text editor labeled “How-to”, scroll within the editor to the `## Commands` section and format only `npm run build` as inline code in the sentence `Run npm run build before deployment.` Then click `Save how-to`. Do not change any other text. The task will finish automatically when done. setup_description: |- Layout is nested_scroll with compact spacing and medium clutter. The page has an outer scrollable analytics column and an inner editor panel; the rich text editor is the inner scroll target. The editor labeled “How-to” is rendered inside a fixed-height Ant Design side panel and uses a sticky toolbar so the controls remain visible while its content scrolls. Only about 9 lines are visible at once. The `## Commands` section is offscreen at load. Initial document: # Runbook ## Summary Prepare the release carefully. Check the staging environment before proceeding. Coordinate with the on-call team if needed. ## Prerequisites Verify all feature flags are set correctly. Review the deployment checklist one more time. Confirm that monitoring dashboards are green. ## Commands Run npm run build before deployment. ## Notes Escalate if staging fails. A “Save how-to” button commits the editor state. scene_context: theme: dark spacing: compact layout: nested_scroll placement: bottom_left scale: default instances: 1 guidance: text clutter: medium component_context: overlay_model: none internal_scroll_region: true confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - internal_scroll_region - layout - confirmation_model factor_rationale: The target line begins offscreen inside the editor rather than the page, and the final save step makes the task about precise offscreen localization plus exact local formatting rather than generic typing. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 5 target_acquisition: 3 density_choice_interference: 1 depth_layering: 3 feedback_dynamics: 3 semantic_observability: 5 disambiguation_load: 2 justification: The textual change is tiny, but it is buried in a long document inside an internal scroll region, and the agent must preserve every non-target block while adding only one inline-code mark. success_trigger: human_readable: - The committed document in `How-to` matches the target document after normalization, with only `npm run build` formatted as inline code. - The `Save how-to` control has been clicked. canonical_predicate: predicate_type: equals target_state: instance_label: How-to mode: mixed_structure_and_marks blocks: - block_type: heading level: 1 text: Runbook - block_type: heading level: 2 text: Summary - block_type: paragraph text: Prepare the release carefully. - block_type: heading level: 2 text: Commands - block_type: paragraph text: Run npm run build before deployment. marked_spans: - substring: npm run build marks: - code forbidden_marks_outside_spans: - code - block_type: heading level: 2 text: Notes - block_type: paragraph text: Escalate if staging fails. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save how-to require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Scrolling the page instead of the editor and never reaching the Commands section. - Changing the text itself instead of only adding inline-code formatting. - Formatting too much of the sentence as code. - Leaving the document in the correct draft state but not clicking `Save how-to`. expected_interaction_path: - Focus the editor and scroll within its content area until `## Commands` is visible. - Select exactly `npm run build` in the target sentence. - Apply the inline code mark with the sticky toolbar. - Click `Save how-to`. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: compare a semantic block tree so that only the intended inline-code mark may change.' - id: rich_text_editor-antd-v2-T04 name: Exact-span link insertion in a drawer canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: open_overlay secondary_template: editor_operation browsergym_goal: Open the `Edit resources` drawer. In the rich text editor labeled `Resources`, make only `release notes` a link to `https://example.com/notes`. Confirm the link in the link popover, then click `Save resources`. Do not change any other text. The task will finish automatically when done. ui_copy: Open the `Edit resources` drawer. In the rich text editor labeled `Resources`, make only `release notes` a link to `https://example.com/notes`. Confirm the link in the link popover, then click `Save resources`. Do not change any other text. The task will finish automatically when done. setup_description: |- Layout is drawer_flow with compact spacing and medium clutter. Clicking “Edit resources” opens a Ant Design Drawer anchored off-center. Inside the drawer is one rich text editor labeled “Resources” and a footer with “Cancel” and “Save resources”. The editor starts with one paragraph: `See the release notes for details.` A Link control in the toolbar opens a small URL popover with a field and a “Save link” action. The drawer should close only after “Save resources” is clicked. scene_context: theme: light spacing: compact layout: drawer_flow placement: off_center 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 - secondary_supporting_primitive factor_rationale: The target action requires a nested overlay inside a drawer plus a final drawer-level commit, so the task isolates selection-dependent linking and state verification in layered UI rather than plain text editing. difficulty: difficulty_bucket: hard tier: L2 axes_ratings: precision_requirement: 5 target_acquisition: 3 density_choice_interference: 2 depth_layering: 4 feedback_dynamics: 4 semantic_observability: 4 disambiguation_load: 1 justification: This is hard because the agent must select the exact substring, use a nested link popover correctly, and then complete a separate save step that commits the drawer draft. success_trigger: human_readable: - The committed content in `Resources` remains `See the release notes for details.` with only `release notes` linked to `https://example.com/notes`. - The nested link popover is no longer open, and the final `Save resources` action has been completed. canonical_predicate: predicate_type: equals target_state: instance_label: Resources mode: marked_substring text: See the release notes for details. marked_spans: - substring: release notes marks: - type: link href: https://example.com/notes saved: true overlay_open: false tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save resources require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Linking the whole sentence or the wrong substring. - Entering the right URL in the popover but not saving the drawer draft. - Typing or deleting additional text. - Leaving the nested popover open or cancelling it. expected_interaction_path: - Open the resources drawer. - Select exactly `release notes`. - Open the link popover, enter the URL, and confirm it. - Click `Save resources` to commit the drawer draft. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: validate link-mark attributes after the final drawer save, not only while the popover is open.' - id: rich_text_editor-antd-v2-T05 name: Row-local clear formatting in the correct table row canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: table_operation secondary_template: clear_reset browsergym_goal: In the `Gateway` row of the table, edit only the `Runbook reply` rich text editor so that the text remains exactly `Deploy now.` but all formatting is removed. Leave the `Billing` row unchanged, then click `Save` in the Gateway row. The task will finish automatically when done. ui_copy: In the `Gateway` row of the table, edit only the `Runbook reply` rich text editor so that the text remains exactly `Deploy now.` but all formatting is removed. Leave the `Billing` row unchanged, then click `Save` in the Gateway row. The task will finish automatically when done. setup_description: |- Layout is table_cell with compact spacing, small controls, and medium clutter. A dense Ant Design Table shows two rows: “Gateway” and “Billing”. Each row contains a compact rich text editor in the “Runbook reply” column and a row-local Save button. Target row initial content: - Gateway: the text `Deploy now.` with `Deploy` bold and `now` italic. - Billing: the text `Hold release.` with no requested changes. The task is row-scoped: only the Gateway row should change, and only its row-local Save should be used. scene_context: theme: dark spacing: compact layout: table_cell placement: off_center scale: small instances: 2 guidance: text clutter: medium component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: discrete secondary_supporting_primitive: table design_intent: active_factors: - layout - instances - confirmation_model factor_rationale: 'The difficulty comes from row-level grounding and row-level save semantics: the agent must target the correct compact embedded editor, remove formatting without changing the text, and avoid committing the wrong row.' difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 2 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 3 disambiguation_load: 5 justification: Dense row-local editors plus separate row saves create high wrong-row risk, and clear-formatting must preserve the exact text rather than retyping a near-match. success_trigger: human_readable: - The Gateway row’s committed text is exactly `Deploy now.` with no inline formatting marks remaining. - The Billing row’s text remains `Hold release.` unchanged. - The Gateway row’s Save control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Gateway text: Deploy now. forbidden_marks: - bold - italic - underline - strike - highlight - code - link - instance_label: Billing plaintext_equals: Hold release. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: save-gateway-row require_correct_instance: true target_instance_label_or_id: Gateway terminal_condition: task ends when predicate holds negative_cases: - Clearing formatting in Billing instead of Gateway. - Changing the words or punctuation in `Deploy now.`. - Leaving any formatting marks on either word in Gateway. - Reaching the correct state in draft but clicking the wrong row’s Save or not saving at all. expected_interaction_path: - Locate the Gateway row and focus its embedded editor. - Select the full text and clear formatting. - Verify Billing is unchanged. - Click the Gateway row’s Save control. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: bind the save control and editor state to the row key rather than DOM order.' - id: rich_text_editor-antd-v2-T06 name: Reconstruct a visual reference card and save canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: match_reference secondary_template: null browsergym_goal: In the rich text editor labeled `Sprint brief`, make the content match the visual Example exactly, then click `Save brief`. The task will finish automatically when done. ui_copy: In the rich text editor labeled `Sprint brief`, make the content match the visual Example exactly, then click `Save brief`. The task will finish automatically when done. setup_description: |- Layout is dashboard_panel with compact spacing and high clutter. A Ant Design Card titled “Sprint brief” sits beside static charts, KPI chips, and an activity rail. The target editor starts empty. Next to the editor is a read-only Example card showing the target rich text appearance only: a centered Heading 2 `Sprint 12`, an ordered list with `Fix bugs` and `Polish UI`, and a final italic paragraph `Last updated: Feb 3`. A `Save brief` button commits the editor state. scene_context: theme: light spacing: compact layout: dashboard_panel placement: top_right scale: default instances: 1 guidance: visual clutter: high component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - guidance - layout - confirmation_model factor_rationale: The target is specified by rendered appearance rather than source text, so the agent must map visual structure onto the editor’s block and mark idioms, then explicitly commit the result in a cluttered dashboard scene. difficulty: difficulty_bucket: hard tier: L2 axes_ratings: precision_requirement: 5 target_acquisition: 3 density_choice_interference: 2 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 2 justification: 'This is hard because the agent cannot just copy a visible source string: it must infer heading level, alignment, list type, and italic scope from the reference card and then save the result.' success_trigger: human_readable: - 'The committed content in `Sprint brief` matches the Example card semantically: centered H2 `Sprint 12`, an ordered list [`Fix bugs`, `Polish UI`], and an italic paragraph `Last updated: Feb 3`.' - The `Save brief` control has been clicked. canonical_predicate: predicate_type: matches_reference target_state: instance_label: Sprint brief mode: block_structure_equals blocks: - block_type: heading level: 2 text: Sprint 12 attributes: textAlign: center - block_type: ordered_list items: - Fix bugs - Polish UI - block_type: paragraph text: 'Last updated: Feb 3' marks: - italic saved: true tolerance: normalize_whitespace: true allow_equivalent_markup: true require_confirm: true confirm_control: Save brief require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Using bullets instead of an ordered list. - Applying the italic mark to only part of the final line or the wrong line. - Creating the right blocks but not centering the heading. - Matching the reference in draft but not clicking `Save brief`. expected_interaction_path: - Inspect the Example card and infer the target block structure. - Build the heading, ordered list, and italic final line in the editor. - Verify the heading is centered. - Click `Save brief`. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: compare semantic block structure rather than pixels.' - id: rich_text_editor-antd-v2-T07 name: Three-step carryover formatting with save canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: editor_operation secondary_template: null browsergym_goal: In the rich text editor labeled `Reminder`, leave the text exactly as `Please review the Q1 plan today.` but highlight `Q1`, italicize `today`, and format `plan` as inline code. Then click `Apply reminder`. The task will finish automatically when done. ui_copy: In the rich text editor labeled `Reminder`, leave the text exactly as `Please review the Q1 plan today.` but highlight `Q1`, italicize `today`, and format `plan` as inline code. Then click `Apply reminder`. The task will finish automatically when done. setup_description: |- Layout is settings_panel with compact spacing and high clutter. The right panel contains one rich text editor labeled “Reminder” beneath unrelated toggles, segmented controls, and static status chips. Initial content is a single paragraph: `Please review the Q1 plan today.` The toolbar includes Highlight, Italic, and Inline Code controls. A footer button labeled “Apply reminder” commits the change. scene_context: theme: light spacing: compact layout: settings_panel placement: off_center scale: small instances: 1 guidance: text clutter: high component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - spacing - clutter - confirmation_model factor_rationale: The pressure comes from maintaining selection state across three different formatting operations in a dense panel, then verifying that the exact text—not a rewritten approximation—was committed. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 5 target_acquisition: 3 density_choice_interference: 2 depth_layering: 3 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 2 justification: This is hard because three precise substring operations must be sequenced without drifting the selection or adding extra marks, and the final apply step punishes stopping after only the draft looks right. success_trigger: human_readable: - The committed text remains exactly `Please review the Q1 plan today.`. - Only `Q1` is highlighted, only `plan` is inline code, and only `today` is italicized. - The `Apply reminder` control has been clicked. canonical_predicate: predicate_type: equals target_state: instance_label: Reminder mode: multiple_marked_substrings text: Please review the Q1 plan today. marked_spans: - substring: Q1 marks: - highlight - substring: plan marks: - code - substring: today marks: - italic forbidden_marks_outside_spans: - highlight - code - italic saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Apply reminder require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Selecting the wrong substrings or including surrounding spaces/punctuation. - Applying one or two of the requested marks but not all three. - Rewriting the sentence instead of preserving it exactly. - Reaching the right draft state but not clicking `Apply reminder`. expected_interaction_path: - Select `Q1` and apply Highlight. - Select `plan` and apply Inline Code. - Select `today` and apply Italic. - Click `Apply reminder`. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: verify mark coverage by exact substring offsets after the apply action.' - id: rich_text_editor-antd-v2-T08 name: Clear formatting in the correct editor among four canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: clear_reset secondary_template: null browsergym_goal: In the high-contrast `Partner communications` panel, edit only the rich text editor labeled `Partner summary` so that the sentence remains exactly `Beta is available now.` but all formatting is removed. Do not edit the other three editors. Then click `Save panel`. The task will finish automatically when done. ui_copy: In the high-contrast `Partner communications` panel, edit only the rich text editor labeled `Partner summary` so that the sentence remains exactly `Beta is available now.` but all formatting is removed. Do not edit the other three editors. Then click `Save panel`. The task will finish automatically when done. setup_description: |- Layout is settings_panel with high-contrast theme, compact spacing, small controls, and high clutter. Inside a Ant Design Card titled “Partner communications” are four vertically stacked rich text editors: `Public summary`, `Partner summary`, `Rollback note`, and `Internal escalation`. The target editor is `Partner summary`. Its underlying text is `Beta is available now.` but `Beta` is bold and `available` is italic in the initial state. The other three editors contain distinct notes and must remain unchanged. A shared `Save panel` button commits the whole panel. scene_context: theme: high_contrast spacing: compact layout: settings_panel placement: bottom_right scale: small instances: 4_plus 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 - theme - clutter factor_rationale: 'This is a deliberate disambiguation-heavy task: four same-type editors are packed into one high-pressure panel, and the only successful path is clearing formatting in exactly one of them while preserving the rest through a shared save.' difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 3 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 5 justification: The text operation itself is simple, but the panel is intentionally dense and high-contrast, so the hard part is targeting the correct editor and preserving all non-target state through a shared commit. success_trigger: human_readable: - The committed content in `Partner summary` is exactly `Beta is available now.` with no inline marks remaining. - The committed contents of the other three editors are unchanged. - The `Save panel` control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Partner summary text: Beta is available now. forbidden_marks: - bold - italic - underline - strike - highlight - code - link - instance_label: Public summary plaintext_equals: No breaking changes. - instance_label: Rollback note plaintext_equals: Draft the rollback memo. - instance_label: Internal escalation plaintext_equals: Notify the on-call manager. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save panel require_correct_instance: true target_instance_label_or_id: Partner summary terminal_condition: task ends when predicate holds negative_cases: - Clearing or editing the wrong editor. - Changing the words or punctuation in `Partner summary`. - Leaving any formatting mark in the target editor. - Making the draft correct but failing to click `Save panel`. expected_interaction_path: - Locate the `Partner summary` editor among the four stacked editors. - Select all text in that editor and clear formatting. - Confirm the other three editors still contain their original text. - Click `Save panel`. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: evaluate non-target editors as invariants so accidental collateral edits fail cleanly.' - id: rich_text_editor-antd-v2-T09 name: Publish after mixed block-plus-inline formatting canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: confirm_cancel secondary_template: editor_operation browsergym_goal: Open `Edit post`. In the rich text editor labeled `Post body`, format the first line `Launch` as a centered Heading 2 and format only `npm install` as inline code in the second line. Then click `Publish` and `Confirm publish`. The task will finish automatically when done. ui_copy: Open `Edit post`. In the rich text editor labeled `Post body`, format the first line `Launch` as a centered Heading 2 and format only `npm install` as inline code in the second line. Then click `Publish` and `Confirm publish`. The task will finish automatically when done. setup_description: |- Layout is modal_flow with compact spacing and low clutter. Clicking `Edit post` opens a Ant Design Modal containing one rich text editor labeled `Post body`, initially prefilled with two plain paragraphs: 1) `Launch` 2) `Run npm install before release.` The modal footer has `Cancel` and `Publish`. Clicking `Publish` opens a second confirmation overlay with `Back` and `Confirm publish`. The post is committed only after the final confirm step. scene_context: theme: dark 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 - precision_surface factor_rationale: The task combines exact rich-text semantics with a two-stage commit boundary, so the agent must distinguish a correct draft from a correctly published result. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 5 target_acquisition: 3 density_choice_interference: 1 depth_layering: 5 feedback_dynamics: 5 semantic_observability: 4 disambiguation_load: 2 justification: The editor work is already multi-step, and the stacked confirmation overlay adds another point of failure for agents that stop once the modal draft looks correct. success_trigger: human_readable: - The committed `Post body` contains a centered H2 `Launch` followed by a paragraph where only `npm install` is inline code. - The `Publish` action and then `Confirm publish` action have both been completed. - The overlays are closed at terminal state. canonical_predicate: predicate_type: equals target_state: instance_label: Post body mode: mixed_structure_and_marks blocks: - block_type: heading level: 2 text: Launch attributes: textAlign: center - block_type: paragraph text: Run npm install before release. marked_spans: - substring: npm install marks: - code forbidden_marks_outside_spans: - code saved: true overlay_open: false tolerance: normalize_whitespace: true require_confirm: true confirm_control: Confirm publish require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Formatting the heading or inline code incorrectly. - Clicking `Publish` but not completing `Confirm publish`. - Creating the correct structure in draft and then cancelling or backing out. - Turning the whole second line into a code block instead of inline code. expected_interaction_path: - Open the edit modal. - Format `Launch` as centered H2. - Select exactly `npm install` and apply inline code. - Click `Publish`, then `Confirm publish`. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: key success off the committed post preview/state after the final confirmation, not the modal draft.' - id: rich_text_editor-antd-v2-T10 name: Append an exact final block at the true end canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap_sticky implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: scroll_find secondary_template: editor_operation browsergym_goal: 'In the rich text editor labeled `Changelog`, scroll to the true end of the document and append a final paragraph `Checksum: `sha256-ready`` with `sha256-ready` formatted as inline code. Do not change any earlier content. Then click `Save changelog`. The task will finish automatically when done.' ui_copy: 'In the rich text editor labeled `Changelog`, scroll to the true end of the document and append a final paragraph `Checksum: `sha256-ready`` with `sha256-ready` formatted as inline code. Do not change any earlier content. Then click `Save changelog`. The task will finish automatically when done.' setup_description: |- Layout is nested_scroll with compact spacing and medium clutter. The rich text editor labeled `Changelog` sits in a fixed-height panel with an internal scrollbar and a sticky toolbar. A separate page-level sidebar contains unrelated release metrics. The document is long enough that its true end is offscreen at load. The last existing block is a paragraph `Monitoring continues overnight.` The footer button `Save changelog` commits the result. scene_context: theme: dark spacing: compact layout: nested_scroll placement: bottom_right scale: default instances: 1 guidance: text clutter: medium component_context: overlay_model: none internal_scroll_region: true confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - internal_scroll_region - layout - confirmation_model factor_rationale: The task is hard because the insertion point is the real end of a long editor, not the visible viewport, and the final save step forces preservation of the existing document prefix. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 3 density_choice_interference: 1 depth_layering: 3 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 2 justification: Agents often append in the wrong place in long editors; here they also need exact inline-code formatting and a final commit without disturbing the existing prefix. success_trigger: human_readable: - 'The committed `Changelog` document preserves its original prefix and ends with a final paragraph `Checksum: sha256-ready` where only `sha256-ready` is inline code.' - The `Save changelog` control has been clicked. canonical_predicate: predicate_type: equals target_state: instance_label: Changelog mode: ends_with_blocks preserve_prefix: true final_blocks: - block_type: paragraph text: 'Checksum: sha256-ready' marked_spans: - substring: sha256-ready marks: - code forbidden_marks_outside_spans: - code saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save changelog require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Inserting the checksum paragraph anywhere except the true end of the document. - Appending the right words but without inline-code formatting. - Changing or deleting earlier changelog content. - Leaving the draft correct but never clicking `Save changelog`. expected_interaction_path: - Scroll within the editor to its true end. - Place the caret after the last existing block. - Append the checksum paragraph and format only the token as inline code. - Click `Save changelog`. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: compare a hash of the original prefix plus semantic inspection of the final block.' - id: rich_text_editor-antd-v2-T11 name: Subscript and superscript in the correct note canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: editor_operation secondary_template: null browsergym_goal: In the review card, edit only the rich text editor labeled `Chemistry note`. Keep the text exactly `H2O uses m2 notation.`, but format the `2` in `H2O` as subscript and the `2` in `m2` as superscript. Do not edit `Meeting note`. Then click `Apply note`. The task will finish automatically when done. ui_copy: In the review card, edit only the rich text editor labeled `Chemistry note`. Keep the text exactly `H2O uses m2 notation.`, but format the `2` in `H2O` as subscript and the `2` in `m2` as superscript. Do not edit `Meeting note`. Then click `Apply note`. The task will finish automatically when done. setup_description: |- Layout is inline_surface with compact spacing, small controls, and medium clutter. A review card contains two side-by-side rich text editors labeled `Meeting note` and `Chemistry note`; only `Chemistry note` should change. Initial content in `Chemistry note` is `H2O uses m2 notation.` The toolbar includes Subscript and Superscript controls. A footer button `Apply note` commits the card state. `Meeting note` contains `Agenda locked.` and must remain unchanged. scene_context: theme: light spacing: compact layout: inline_surface placement: off_center scale: small instances: 2 guidance: text clutter: medium 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 - scale - confirmation_model factor_rationale: 'This is a localized but high-precision semantics task: two identical tiny editors sit side by side, and success depends on applying two different marks to two different occurrences of the same character before committing.' difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 5 target_acquisition: 4 density_choice_interference: 2 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 5 justification: The text does not change, so the only visible evidence is mark state; that makes exact occurrence tracking and correct-instance disambiguation substantially harder than ordinary typing. success_trigger: human_readable: - The committed text in `Chemistry note` remains exactly `H2O uses m2 notation.` with the first `2` subscripted and the second `2` superscripted. - The committed text in `Meeting note` remains `Agenda locked.`. - The `Apply note` control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Chemistry note text: H2O uses m2 notation. marked_occurrences: - substring: '2' occurrence_index: 1 marks: - subscript - substring: '2' occurrence_index: 2 marks: - superscript forbidden_marks_outside_occurrences: - subscript - superscript - instance_label: Meeting note plaintext_equals: Agenda locked. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Apply note require_correct_instance: true target_instance_label_or_id: Chemistry note terminal_condition: task ends when predicate holds negative_cases: - Applying both marks to the same `2`. - Applying subscript or superscript in the wrong editor. - Changing the text instead of only the mark state. - Leaving the right draft in place but failing to click `Apply note`. expected_interaction_path: - Focus `Chemistry note`. - Select the `2` in `H2O` and apply Subscript. - Select the `2` in `m2` and apply Superscript. - Verify `Meeting note` is unchanged and click `Apply note`. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: mark occurrence indices explicitly so the two identical characters are not conflated.' - id: rich_text_editor-antd-v2-T12 name: Bubble-menu formatting in the correct service row canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap_bubble implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: table_operation secondary_template: disclose browsergym_goal: In the `Auth` row of the table, use the selection bubble menu to make only `blocked` bold in the sentence `Deploy is blocked today.`. Leave the `Billing` row unchanged, then click `Save` in the Auth row. The task will finish automatically when done. ui_copy: In the `Auth` row of the table, use the selection bubble menu to make only `blocked` bold in the sentence `Deploy is blocked today.`. Leave the `Billing` row unchanged, then click `Save` in the Auth row. The task will finish automatically when done. setup_description: |- Layout is table_cell with compact spacing, small controls, and high clutter. A compact Ant Design Table shows two service rows, `Auth` and `Billing`. Each row contains a one-line rich text editor and a row-local Save button. The target row `Auth` starts with `Deploy is blocked today.`. The editor has no full toolbar; selecting text reveals a small BubbleMenu with Bold and Italic. The `Billing` row starts with `Ready to ship.` and must remain unchanged. scene_context: theme: light spacing: compact layout: table_cell placement: bottom_left scale: small 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: table design_intent: active_factors: - layout - instances - secondary_supporting_primitive factor_rationale: 'This combines two of the strongest v1 pressure sources: compact embedded editors inside a table row and a conditional BubbleMenu that only appears after a correct text selection.' difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 5 density_choice_interference: 2 depth_layering: 4 feedback_dynamics: 3 semantic_observability: 3 disambiguation_load: 5 justification: The agent must first enter the correct row-local editor, then make a precise substring selection under dense controls, then save the correct row and leave the sibling row unchanged. success_trigger: human_readable: - The committed `Auth` row text remains `Deploy is blocked today.` with only `blocked` bolded. - The committed `Billing` row text remains `Ready to ship.` unchanged. - The `Auth` row’s Save control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Auth text: Deploy is blocked today. marked_spans: - substring: blocked marks: - bold forbidden_marks_outside_spans: - bold - instance_label: Billing plaintext_equals: Ready to ship. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: save-auth-row require_correct_instance: true target_instance_label_or_id: Auth terminal_condition: task ends when predicate holds negative_cases: - Applying the bold mark in `Billing` instead of `Auth`. - Bolding extra spaces or punctuation around `blocked`. - Editing the right row but clicking the wrong Save button. - Changing the text content instead of only the formatting. expected_interaction_path: - Locate the `Auth` row editor. - Select exactly `blocked` to reveal the BubbleMenu. - Apply Bold and verify `Billing` is unchanged. - Click the `Auth` row’s Save control. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: bind save semantics to the row key and verify non-target row invariance.' - id: rich_text_editor-antd-v2-T13 name: High-contrast visual quote match with sibling preservation canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: match_reference secondary_template: null browsergym_goal: 'In the high-contrast quotes panel, edit only the rich text editor labeled `Customer-facing quote` so that it matches the Example exactly: a blockquote `Stay curious` followed by a normal paragraph `— Ada`. Do not edit the other two editors. Then click `Apply quotes`. The task will finish automatically when done.' ui_copy: 'In the high-contrast quotes panel, edit only the rich text editor labeled `Customer-facing quote` so that it matches the Example exactly: a blockquote `Stay curious` followed by a normal paragraph `— Ada`. Do not edit the other two editors. Then click `Apply quotes`. The task will finish automatically when done.' setup_description: |- Layout is settings_panel with high-contrast theme, compact spacing, and high clutter. Three rich text editors are stacked in a card: `Customer-facing quote`, `Internal quote`, and `Legal blurb`. Next to them is a small read-only Example card showing the target appearance only: a blockquote `Stay curious` followed by a normal paragraph `— Ada`. Only `Customer-facing quote` should be edited. A shared `Apply quotes` button commits all three drafts. scene_context: theme: high_contrast spacing: compact layout: settings_panel placement: off_center scale: small instances: 3 guidance: visual clutter: high component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - guidance - instances - theme factor_rationale: The target is purely visual and one of three similar editors must be chosen, so the agent has to map rendered semantics to editor structure while preserving sibling state in a visually dense panel. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 3 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 5 justification: This is hard because it combines visual-only structural inference with multi-instance disambiguation and a shared panel-level commit. success_trigger: human_readable: - 'The committed content in `Customer-facing quote` semantically matches the Example card: a blockquote `Stay curious` followed by a paragraph `— Ada`.' - The committed contents in `Internal quote` and `Legal blurb` remain unchanged. - The `Apply quotes` action has been completed. canonical_predicate: predicate_type: matches_reference target_state: mode: multi_instance targets: - instance_label: Customer-facing quote reference_semantics: blocks: - block_type: blockquote text: Stay curious - block_type: paragraph text: — Ada - instance_label: Internal quote plaintext_equals: Do not publish yet. - instance_label: Legal blurb plaintext_equals: Awaiting approval. saved: true tolerance: normalize_whitespace: true allow_equivalent_markup: true require_confirm: true confirm_control: Apply quotes require_correct_instance: true target_instance_label_or_id: Customer-facing quote terminal_condition: task ends when predicate holds negative_cases: - Editing `Internal quote` or `Legal blurb` instead of the customer-facing editor. - Using italics or plain text where a blockquote is required. - Producing the right draft in the target editor but not clicking `Apply quotes`. - Changing any non-target editor. expected_interaction_path: - Inspect the Example card. - Focus `Customer-facing quote` and create the blockquote plus paragraph structure. - Verify the other two editors are unchanged. - Click `Apply quotes`. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: use semantic reference matching plus non-target invariants.' - id: rich_text_editor-antd-v2-T14 name: Two-editor drawer with exact heading-divider-list structure canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: open_overlay secondary_template: editor_operation browsergym_goal: Open the `Checklist drawer`. In the rich text editor labeled `Deployment note` (not `Internal note`), make the content exactly an H3 `Risks`, then a horizontal divider, then a bulleted list with `API drift` and `Manual rollback`. Then click `Save drawer`. The task will finish automatically when done. ui_copy: Open the `Checklist drawer`. In the rich text editor labeled `Deployment note` (not `Internal note`), make the content exactly an H3 `Risks`, then a horizontal divider, then a bulleted list with `API drift` and `Manual rollback`. Then click `Save drawer`. The task will finish automatically when done. setup_description: |- Layout is drawer_flow with compact spacing and medium clutter. Clicking `Open checklist drawer` opens a Ant Design Drawer containing two rich text editors labeled `Deployment note` and `Internal note`. The target is `Deployment note`. The target editor starts empty. The toolbar includes Heading controls, Horizontal Rule, and Bullet List. The drawer footer contains `Cancel` and `Save drawer`. `Internal note` starts with `Leave untouched.` and must remain unchanged. scene_context: theme: light spacing: compact layout: drawer_flow 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: drawer design_intent: active_factors: - layout - instances - confirmation_model factor_rationale: The task couples drawer-local save semantics and same-type disambiguation inside the drawer, while the target itself requires exact multi-block structure rather than plain text. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 3 density_choice_interference: 2 depth_layering: 3 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 5 justification: This is hard because the agent must open the drawer, build an exact three-block structure in the right editor, preserve the sibling editor, and then commit through the drawer footer. success_trigger: human_readable: - The committed `Deployment note` editor contains exactly an H3 `Risks`, a horizontal rule, and a bullet list [`API drift`, `Manual rollback`]. - The committed `Internal note` remains `Leave untouched.` unchanged. - The drawer-local `Save drawer` control has been clicked and the drawer is closed at terminal state. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Deployment note mode: block_structure_equals blocks: - block_type: heading level: 3 text: Risks - block_type: horizontal_rule - block_type: bullet_list items: - API drift - Manual rollback - instance_label: Internal note plaintext_equals: Leave untouched. saved: true overlay_open: false tolerance: normalize_whitespace: true require_confirm: true confirm_control: Save drawer require_correct_instance: true target_instance_label_or_id: Deployment note terminal_condition: task ends when predicate holds negative_cases: - Editing `Internal note` instead of `Deployment note`. - Using a paragraph or blockquote instead of a horizontal divider. - Building the right draft but not saving the drawer. - Changing the target list items or their order. expected_interaction_path: - Open the drawer. - Focus `Deployment note`. - Insert the H3 heading, horizontal rule, and bullet list in order. - Verify `Internal note` is unchanged and click `Save drawer`. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: evaluate saved block structure after the drawer closes.' - id: rich_text_editor-antd-v2-T15 name: Below-the-fold signature update with exact bold second line canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: scroll_find secondary_template: editor_operation browsergym_goal: 'On the Settings page, scroll to the rich text editor labeled `Email signature`. Change it so it becomes exactly two lines: `Thanks,` on the first line and `Alex` in bold on the second line. Then click `Save signature`. The task will finish automatically when done.' ui_copy: 'On the Settings page, scroll to the rich text editor labeled `Email signature`. Change it so it becomes exactly two lines: `Thanks,` on the first line and `Alex` in bold on the second line. Then click `Save signature`. The task will finish automatically when done.' setup_description: |- Layout is settings_panel with compact spacing and high clutter. Above the fold are unrelated toggles, dropdowns, and notification preferences. The rich text editor labeled `Email signature` is below the fold. The editor initially contains two plain paragraphs: `Best,` and `Sam`. A `Save signature` button commits changes for that section only. The task should leave the first line plain and make the second line exactly `Alex` in bold. scene_context: theme: light spacing: compact layout: settings_panel placement: center scale: default instances: 1 guidance: text clutter: high component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - layout - clutter - confirmation_model factor_rationale: The target editor is below the fold in a busy settings page, and the final state mixes exact line breaks with a mark on only the second line, so the task stresses both scroll localization and exact structural verification. difficulty: difficulty_bucket: hard tier: L2 axes_ratings: precision_requirement: 4 target_acquisition: 3 density_choice_interference: 1 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 2 justification: The editor is easy to manipulate once found, but the hard part is finding the correct section, preserving exact line structure, and applying bold only to the second line before saving. success_trigger: human_readable: - 'The committed `Email signature` document consists of exactly two paragraphs: `Thanks,` and a bold `Alex`.' - The `Save signature` control has been clicked. canonical_predicate: predicate_type: equals target_state: instance_label: Email signature mode: block_structure_equals blocks: - block_type: paragraph text: Thanks, - block_type: paragraph text: Alex marks: - bold saved: true tolerance: normalize_whitespace: false normalize_line_endings: true ignore_trailing_newline: true require_confirm: true confirm_control: Save signature require_correct_instance: false target_instance_label_or_id: null terminal_condition: task ends when predicate holds negative_cases: - Creating the right words with the wrong line breaks. - Leaving `Alex` plain or bolding both lines. - Editing the wrong settings field instead of the signature editor. - Leaving the section in the right draft state but not clicking `Save signature`. expected_interaction_path: - Scroll to the signature section. - Replace the existing two lines with `Thanks,` and `Alex`. - Bold only the second line. - Click `Save signature`. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: compare semantic paragraphs plus mark state, allowing only a trailing editor newline if the implementation emits one.' - id: rich_text_editor-antd-v2-T16 name: Correct reply editor with list-plus-closing and send canonical_type: rich_text_editor implementation_source: antd implementation_variant: composite_antd_tiptap implementation_component: 'AntD: Tiptap-based rich text editor + Ant Design carrier components [composite]' task_template: editor_operation secondary_template: null browsergym_goal: In the support card, edit only the rich text editor labeled `Customer reply`. Make it an ordered list with `Acknowledge issue` and `Ship fix`, followed by an italic paragraph `— Support team`. Do not edit `Internal note` or `Legal note`. Then click `Send reply`. The task will finish automatically when done. ui_copy: In the support card, edit only the rich text editor labeled `Customer reply`. Make it an ordered list with `Acknowledge issue` and `Ship fix`, followed by an italic paragraph `— Support team`. Do not edit `Internal note` or `Legal note`. Then click `Send reply`. The task will finish automatically when done. setup_description: |- Layout is inline_surface with compact spacing, medium clutter, and three same-size editors stacked in one support card: `Customer reply`, `Internal note`, and `Legal note`. Only `Customer reply` should change. The target card footer contains `Send reply`. Initial contents: - Customer reply: empty - Internal note: `Remember to tag billing.` - Legal note: `Do not promise compensation yet.` The rich text toolbar includes Ordered List and Italic. scene_context: theme: light spacing: compact layout: inline_surface placement: off_center scale: small instances: 3 guidance: text clutter: medium component_context: overlay_model: none internal_scroll_region: false confirmation_model: explicit_apply precision_surface: mixed secondary_supporting_primitive: none design_intent: active_factors: - instances - confirmation_model - layout factor_rationale: The agent has to choose the only editable target among three same-type editors, build a mixed list-plus-italic structure, preserve the two non-target notes, and then commit through a card-local send action. difficulty: difficulty_bucket: hard tier: L3 axes_ratings: precision_requirement: 4 target_acquisition: 4 density_choice_interference: 2 depth_layering: 2 feedback_dynamics: 3 semantic_observability: 4 disambiguation_load: 5 justification: 'The structure is short but exact, and the real challenge is disambiguation plus preservation: one wrong focus or one collateral edit invalidates the task even if the list itself is correct.' success_trigger: human_readable: - The committed `Customer reply` editor contains an ordered list [`Acknowledge issue`, `Ship fix`] followed by an italic paragraph `— Support team`. - The committed `Internal note` and `Legal note` remain unchanged. - The `Send reply` control has been clicked. canonical_predicate: predicate_type: equals target_state: mode: multi_instance targets: - instance_label: Customer reply mode: block_structure_equals blocks: - block_type: ordered_list items: - Acknowledge issue - Ship fix - block_type: paragraph text: — Support team marks: - italic - instance_label: Internal note plaintext_equals: Remember to tag billing. - instance_label: Legal note plaintext_equals: Do not promise compensation yet. saved: true tolerance: normalize_whitespace: true require_confirm: true confirm_control: Send reply require_correct_instance: true target_instance_label_or_id: Customer reply terminal_condition: task ends when predicate holds negative_cases: - Editing `Internal note` or `Legal note` instead of `Customer reply`. - Using bullets instead of an ordered list. - Failing to italicize the closing line or including it inside the list. - Reaching the right draft state but not clicking `Send reply`. expected_interaction_path: - Focus `Customer reply`. - Create the ordered list with the two items. - Add a final paragraph `— Support team` and italicize it. - Verify the other two editors are unchanged and click `Send reply`. notes: 'Coverage note: Ant Design does not ship a core rich text editor and points users to third-party rich text editors, so this task uses a Tiptap-based editor inside Ant Design carrier components. Checker hint: evaluate target/non-target invariants together so wrong-instance edits fail deterministically.'