v2.9.3

What's Changed

• Compressed UIA tree: capture-tree compress=true reduces tree payload by 85-95% on complex desktops
• MemoryPack serialization: capture-tree-binary is 48% faster, 30% smaller payload
• LLM-driven agent loop: LlmPlanner generates steps via Hermes gateway
• Watchdog as default: install.ps1 launches watchdog by default with crash recovery
• CI badge fixed: shields.io for private repo support
• Version bump: 2.9.2 → 2.9.3

Full changelog: https://github.com/lessenings-prog/OpenGUI/commits/v2.9.3

v2.9.2

What's new since v2.7.5

• CacheRequest in GetActionable — 4x speedup (14ms avg). All UIA tree walks use batch property fetching.
• WinRT OCR — Windows.Media.Ocr with Tesseract fallback. 2-5x faster, zero-install on Windows 10+.
• Next-action overlay — real-time desktop label showing what the agent plans to do next.
• Windows Service support — watchdog can now run as a proper Windows Service (--install, --uninstall flags).
• No console windows — all three processes compiled as WinExe to suppress background console flashes.
• Predicate-based step advancement — plan steps can specify wait_for text before executing.
• Semver policy — VERSIONING.md added. Going forward: MAJOR=breaking, MINOR=features, PATCH=fixes.

Test suite

• Phase 2 C# tests: 36/36 passing
• Phase 3 agent loop tests: 7/7 passing (automated, no operator input)

Install

powershell -ExecutionPolicy Bypass -File dist\OpenGUI\install.ps1

Connect

from reliable_bridge import send
print(send('status'))

v2.7.5

OpenGUI — Windows-native desktop automation for LLM agents

Full documentation →

OpenGUI is a .NET 9 Windows service that gives LLM agents (Claude, GPT, DeepSeek, Hermes) direct control over desktop applications. Communicate over a named pipe with JSON commands.

Key features:

• Named-pipe IPC — sub-millisecond dispatch, no TCP overhead
• UIA tree perception — CacheRequest-optimized, ~11ms for modal scan
• SendInput execution — keyboard, mouse, hotkeys with timing control
• Divergence detection — ActionTruth verification after every action
• Phase 3 agent loop — 7/7 automated tests (open Notepad → type → save → verify → close)
• Crash recovery — external watchdog with challenge-response heartbeat + Job Object isolation
• WinRT OCR — Windows-native OCR with Tesseract fallback
• Overlay feedback — real-time desktop highlights showing agent intent

v2.7.5 updates

• CacheRequest in DetectModals — 12,000ms timeout → 11ms (1080x improvement)
• All UIA tree walks use el.Cached instead of el.Current
• detect-modals bounded test: 5/5 PASS, 11ms

Test suite: 36/36 passing

• Hardening: 14/14
• Timeout Governor: 5/5
• Phase 2 stress: 17/17

Install

powershell -ExecutionPolicy Bypass -File dist\OpenGUI\install.ps1

Connect

from reliable_bridge import send
print(send('status'))