Fisfzy/dsh-ego-browser

## Core Features
- The plugin integrates CitroLabs/ego-lite into DeepSeek Harness and exposes 32 structured ego_* tools for spaces, page reading, navigation, waiting, interaction, execution, output, session handling, and diagnostics.
- It provides a real-time observation window that streams the page currently operated by an agent and includes a tab bar, a history drawer, and direct mouse and keyboard control over the same browser instance.
- It supports CDP JPEG screencast and optional FFmpeg H.264 capture, with CDP as the default backend and FFmpeg selectable only after a compatible installation passes capability checks.
- It bundles an ego runtime and a --no-sandbox wrapper, automatically detects Chrome, Chromium, Brave, or Edge, and handles root, Docker, and headless fallbacks without requiring a manual build.
- It includes operational tools such as ego_download, ego_captcha, ego_auth_flush, ego_http, and ego_doctor, plus worker single-instance guarding, crash recovery, cold-start retry, and front-end frame-cache limits.

## Use Cases
- Agents can perform authenticated literature or data collection while an operator watches scrolling, pagination, and PDF downloads in the observation window.
- Agents can fill forms or handle login flows, and a user can take over directly when a captcha or other blocking interaction appears.
- Agents can run QA or smoke tests on web products while the history drawer allows review of visited trajectories.
- Agents can operate the DSH interface itself while the same observation path shows session management, task boards, or settings changes.

## Technical Details
- Each tool converts parameters into a JavaScript script, runs it through ctx.subprocess with ego-browser nodejs, drives a shared Chromium through CDP, and parses results from an @@DSH_RESULT@@ sentinel line.
- All ego_* actions are serialized by an in-process mutex, and errors are normalized before being returned to the host.
- The observation stack uses lib/client.js for watcher leases, JPEG images, and MediaSource video, while lib/cast-server.js proxies metadata SSE, watch APIs, and binary video with backpressure.
- CDP capture uses Page.startScreencast JPEG at a default 20 FPS with frame-ID acknowledgment, while FFmpeg capture uses Windows gfxcapture(hwnd) or display-source cropping on other platforms and delivers fragmented MP4 chunks.

## Notes
- The observation window requires a graphical DSH web session; headless sessions can still use ego_* tools without live viewing.
- Windows support is adapted at the plugin layer, but the underlying ego-lite host remains a community port and may be less stable for complex multi-step flows.
- Runtime cookies are written to disk only during graceful shutdown, so forced restarts require logging in again, and multi-task spaces keep cookies isolated.