Add drag-and-drop, set_slider, hover persistence, and scroll-clip fix#58
Merged
softpudding merged 9 commits intomainfrom Apr 13, 2026
Merged
Add drag-and-drop, set_slider, hover persistence, and scroll-clip fix#58softpudding merged 9 commits intomainfrom
softpudding merged 9 commits intomainfrom
Conversation
Implement end-to-end drag-and-drop support: element discovery via draggable/droppable element types and interaction hints, 2PC confirmation flow with container preview, and precise drop placement using relative_to/position. This addresses the 0% pass rate on taskflow_drag_and_edit eval tasks where the agent had no way to discover or execute DnD operations. Key changes: - Add draggable/droppable as element_types and interactionHints - Detection heuristics: explicit attrs, cursor:grab, parent-of-draggable - HighlightDropPreviewCommand: crops container, highlights inner elements - confirm_drag_and_drop with relative_to/position for precise placement - Fix drag script rAF hang on hidden tabs (setTimeout fallback) - Fix post-drag occlusion false positive (DOM mutates during drag) - Remove misleading offset_x/offset_y from LLM-facing tool schema - Harden eval SSE streaming with retry and configurable timeouts Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…lider Three features addressing eval failures and interaction gaps: 1. Hover persistence: store last-hovered element per conversation/tab and replay hover events before confirmation screenshots, so hover-revealed UI (video controls) stays visible during click confirmation. 2. Slidable interaction hint: detect slider-like elements across three tiers (native range, ARIA role=slider, custom progress bars via structural heuristics) and annotate them with a "slidable" hint. Ancestor walk ensures leaf elements inside slider containers inherit the hint. 3. Universal set_slider: extend from native-only to three paths — native range (write value), ARIA slider (position click via aria-valuemin/max), and generic custom sliders (percentage-based position click with ancestor walk to find full-width container). Also fixes: ancestor opacity walk for visibility detection, large interactive region detection (video players), and slider/draggable exclusion to prevent conflicting hints. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Updates VISUAL_GROUNDING, INTERACTION_MODEL, and DISCOVERY_STRATEGY sections in the system prompt to cover the new DnD 2PC flow, slidable interaction hints, and droppable element discovery. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds a Vite plugin that starts a WebSocket server (port 8767) during `npm run dev`. The extension's background script connects to it in dev builds and calls chrome.runtime.reload() on each rebuild, eliminating the need to manually reload on chrome://extensions. The reload code is tree-shaken out of production builds via a __DEV__ compile-time constant. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The dynamic import via Vite's __vitePreload polyfill silently failed in MV3 service workers (polyfill references `document` which doesn't exist). Switch to static import and disable the modulepreload polyfill. Also change `npm run dev` from watch mode to one-shot build+reload. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The instruction said "board-packet" (hyphenated) but the mock thread subject uses "Board Packet" (space-separated), causing the agent to use a search term that doesn't match. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…eval results - Add isElementVisibleInScrollParent() to filter elements scrolled out of view inside overflow containers, preventing phantom labels from confusing the agent (fixes Drive Bulk Release Assets flash regression) - Add API-stall detection and retry logic to evaluate_browser_agent.py so transient API timeouts trigger automatic re-queue - Add qwen3.6-plus to the large model profile - Update evaluation_report.json with merged results: 89/105 (84.76%) flash 27/35, plus 31/35, 3.6-plus 31/35 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The agent-sdk prompt changed from "Only click, keyboard_input, and select use the YELLOW stage" to "click, keyboard_input, and select return a YELLOW preview screenshot before execution." Update the test assertion to match. Also includes auto-formatting from black and prettier. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
isElementVisibleInScrollParent()filters out elements scrolled outsideoverflow:auto/scrollcontainers, preventing phantom highlight labels (fixes Drive Bulk Release flash regression)Eval Results (84.76% vs 82.86% main baseline)
Test plan
🤖 Generated with Claude Code