v0.38.0
QwenLM/Qwen3-VLv0.38.0Sep 16, 2026by github-actions[bot]
Upgrading?
AI Summary
A major feature release introducing conditional screenshots, snapshot deltas, human-like pointer movement, and advanced recording tools.
Key Highlights
- Added conditional screenshots with change detection thresholds to save tokens.
- Introduced automatic snapshot deltas for compact structural changes.
- Added human-like pointer movement and new recording presentation tools like cursor ripples.
New Features
- Conditional screenshots with `screenshot --if-changed`
- Automatic snapshot deltas with `snapshot --delta`
- Persistent snapshot refs for surviving DOM elements
- Human-like pointer movement with reproducible curved paths
- Recording presentation tools: `--cursor` and `--contact-sheet`
- Stateful auth vault login with `auth login --no-navigate`
- WebMCP catalog updates for browser responses
Full Release Notes
### New Features - Added **conditional screenshots** with `screenshot --if-changed` to skip unchanged captures and `--threshold <0-1>` to tolerate small pixel differences. Screenshot history is scoped to each tab and capture mode, and unchanged captures omit the image path to save tokens (#1813) - Added **automatic snapshot deltas** with `snapshot --delta`, which returns a full baseline followed by an unchanged revision or compact structural changes. Use `--full` to refresh the baseline (#1811) - Added **persistent snapshot refs** for surviving DOM elements across same-document changes. Replaced elements and navigated pages or iframes invalidate their refs without recycling identifiers (#1812) - Added **human-like pointer movement** with session `--input-mode`, per-action `--human` for clicks and drags, and reproducible curved `mouse move` paths (#1810) - Added **recording presentation tools**: `--cursor` renders an animated pointer and click ripple, while `--contact-sheet` saves a timestamped PNG summary of visual changes (#1807, #1808) - Added **stateful auth vault login** with `auth login --no-navigate`, which fills an already prepared login page after verifying its origin against the credential URL (#1771) - Added **WebMCP catalog updates** to normal browser responses. Concise, content-bounded summaries appear on initial discovery and catalog changes, while full schemas remain opt-in (#1842) ### Improvements - Improved **recording frame timing** by holding the latest Chrome frame between repaints so the output duration follows the requested frame rate (#1806) ### Bug Fixes - Fixed **CDP transport recovery and shutdown** so malformed transport messages do not strand unrelated commands and failed initialization disconnects cleanly (#1739) - Fixed **dropdown label matching** to normalize whitespace, including non-breaking spaces, when selecting options by visible label (#1736) ### Documentation - Clarified **reliable browser wait strategies** by recommending page-specific selectors, text, URLs, or JavaScript conditions and reserving `networkidle` for pages known to become quiet (#1834) ### Contributors - @ctate - @Railly - @mauricioantolin - @mecampbellsoup - @evrenverse - @judegao - @petehunt
Related Repositories
rasbt/LLMs-from-scratch★ 107,072
Implement a ChatGPT-like LLM in PyTorch from scratch, step by step
lyogavin/airllm★ 40,206
AirLLM 2.8T LLM inference with single 4GB GPU
NirDiamant/RAG_Techniques★ 29,960
This repository showcases various advanced techniques for Retrieval-Augmented Generation (RAG) systems. Each technique has a detailed notebook tutorial.
microsoft/OmniParser★ 25,539
A simple screen parsing tool towards pure vision based GUI agent
karpathy/nn-zero-to-hero★ 24,907
Neural Networks: Zero to Hero