Desktop Benchmark Starter

How the official desktop benchmark path fits the v0.5 QitOS story.

Desktop Benchmark Starter

QitOS v0.5 now has one canonical multimodal path:

• DesktopEnv
• the desktop-starter benchmark
• qitos.recipes.desktop.osworld_starter
• examples/real/openai_cua_agent.py
• qita visual replay

This is the path to use if you want a credible starting point for computer-use research.

Why this is the release path

The desktop starter benchmark is the first path where QitOS can say:

• one benchmark family is official
• one baseline agent is canonical
• one qita debugging workflow is documented
• one artifact (a persistent output from a run) schema is reused end to end

Completing one path end to end matters more than shipping several half-complete multimodal demos.

What the starter includes

• OSWorld-inspired desktop task shape
• screenshot-backed observations (the data returned by the environment after each action)
• optional a11y / OCR / DOM / UI candidates
• provider-neutral GUI actions
• planner + grounding + action selector + critic (a step-level validator that can allow, stop, or retry) baseline loop
• qita screenshot timeline, playback preview, and basic overlays

What it does not promise yet

• full official OSWorld runtime parity
• rich accessibility-tree execution across every provider
• enterprise-grade approval governance
• full visual replay depth planned for v0.6

The right mental model is:

v0.5 = one complete starter path

not

v0.5 = every possible multimodal path finished

When you need the real benchmark adapter rather than the starter path, move to osworld.