dsh-native-browser

When a DSH agent needs to drive the Chrome you actually use day to day, a generic browser MCP server falls short. It lacks conversation-level ownership, precise tab leases, approval policies, and instant human takeover—and it does not care whether your signed-in sessions and extensions survive. dsh-native-browser closes that product-layer gap. Rather than another click(x, y) wrapper, it is a DSH-native stateful browsing runtime: a Manifest V3 extension talks to a local Broker over Native Messaging, operating directly on your existing Chrome tabs while preserving logged-in state and regular extensions. Its core capabilities center on accessibility-tree-first observation with compact incremental updates, stable element references with pre-dispatch hit-target validation, a presentation-only virtual pointer plus click/wheel pulses so you can watch the agent work, automatic control release the moment the visible DSH conversation switches, short-lived tab leases that prevent cross-conversation authority transfer, and screenshots used only as a visual fallback rather than the default page-structure source. Frame discovery now maps same-process and recursive OOPIF documents, and approval policies can be scoped per action, per lease, or restricted to an explicit trusted-origin list. This project is aimed at developers and power users on macOS with Chrome who are running DSH 0.1.5 release candidates and want their agent to work inside their real browser environment while every step remains visible, interruptible, and auditable. It ships as a public alpha suited for opt-in Chrome testing; production hardening, broad page compatibility, and visual-model accuracy gates are still open acceptance work.