openadapt-types

Important

Status: Experimental. Interoperability schemas, not the product. This package publishes shared schemas for computer-use agents as an optional component, with no production support promise.

The OpenAdapt product is the demonstration compiler, openadapt-flow, installed via the OpenAdapt launcher (pip install openadapt): it compiles a demonstrated GUI workflow into a deterministic, locally executable program. Healthy runs make no model calls, and it halts instead of guessing when verification fails. Lifecycle labels for every repository are in the repository lifecycle registry.

Canonical Pydantic schemas for computer-use agents.

pip install openadapt-types

These are the shared Action and UI-state types used across the OpenAdapt stack: recorders emit them, the compiler stores and replays them, the grounding package resolves their targets, and the privacy package scrubs them. Defining the schema once keeps every substrate (web, Windows, macOS, Linux, RDP, Citrix/VDI) on the same contract.

The OpenAdapt stack

OpenAdapt is a governed demonstration compiler: record a workflow once, compile the recording into a deterministic program, and replay that program with zero model calls on the healthy path. When the live screen does not match what was demonstrated it halts instead of guessing, using identity gates and independent effect verification. Every substrate is first-class: web and desktop recording are validated, RDP and Windows replay are early, and Citrix is exploratory.

Package	Role
openadapt	Launcher and installer (pip install openadapt)
openadapt-flow	Records, compiles, verifies, and replays workflows
openadapt-capture	Cross-platform local desktop recording
openadapt-types	Canonical action and UI-state schema (this package)
openadapt-grounding	Local OCR text-anchoring plus optional model grounding
openadapt-privacy	PHI/PII detection and redaction

Documentation for the whole stack lives at docs.openadapt.ai.

What's in the box

Schema	Purpose
ComputerState	Screen state: screenshot + UI element graph + window context
UINode	Single UI element with role, bbox, hierarchy, platform anchors
Action	Agent action with typed action space + flexible targeting
ActionTarget	Where to act: node_id > description > (x, y) coordinates
ActionResult	Execution outcome with error taxonomy + state delta
Episode / Step	Complete task trajectory (observation → action → result)
FailureRecord	Classified failure for dataset pipelines
ControlOverlayFrameV1 / ControlOverlayTimelineV1	PHI-safe execution overlay state bound to exact evidence media
ControlOverlayFrameV2 / ControlOverlayTimelineV2	Exact, privacy-safe target geometry for sibling overlays and media composition

Quick start

from openadapt_types import (
    Action, ActionTarget, ActionType,
    ComputerState, UINode, BoundingBox,
)

# Describe what's on screen
state = ComputerState(
    viewport=(1920, 1080),
    nodes=[
        UINode(node_id="n0", role="window", name="My App", children_ids=["n1"]),
        UINode(node_id="n1", role="button", name="Submit", parent_id="n0",
               bbox=BoundingBox(x=500, y=400, width=100, height=40)),
    ],
)

# Agent decides what to do
action = Action(
    type=ActionType.CLICK,
    target=ActionTarget(node_id="n1"),
    reasoning="Click Submit to proceed",
)

# Render element tree for LLM prompts
print(state.to_text_tree())
# [n0] window: My App
#   [n1] button: Submit

Action targeting

ActionTarget supports three grounding strategies (in priority order):

# 1. Element-based (preferred, most robust)
ActionTarget(node_id="n1")

# 2. Description-based (resolved by grounding module)
ActionTarget(description="the blue submit button")

# 3. Coordinate-based (fallback)
ActionTarget(x=550, y=420)
ActionTarget(x=0.29, y=0.39, is_normalized=True)

Agents SHOULD produce node_id or description. The runtime resolves to coordinates.

Compatibility with existing schemas

Converters for three existing OpenAdapt schema formats:

from openadapt_types._compat import (
    from_benchmark_observation,   # openadapt-evals BenchmarkObservation
    from_benchmark_action,        # openadapt-evals BenchmarkAction
    from_ml_observation,          # openadapt-ml Observation
    from_ml_action,               # openadapt-ml Action
    from_omnimcp_screen_state,    # omnimcp ScreenState
    from_omnimcp_action_decision, # omnimcp ActionDecision
)

# Convert existing data
state = from_benchmark_observation(obs.__dict__)
action = from_benchmark_action(act.__dict__)

JSON Schema

Export for language-agnostic tooling:

import json
from openadapt_types import ComputerState, Action, Episode

# Get JSON Schema
schema = ComputerState.model_json_schema()
print(json.dumps(schema, indent=2))

The same API exports the versioned cross-surface overlay contracts:

from openadapt_types import (
    ControlOverlayFrameV1,
    ControlOverlayFrameV2,
    ControlOverlayTimelineV1,
    ControlOverlayTimelineV2,
)

frame_schema = ControlOverlayFrameV1.model_json_schema()
timeline_schema = ControlOverlayTimelineV1.model_json_schema()
tracking_frame_schema = ControlOverlayFrameV2.model_json_schema()
tracking_timeline_schema = ControlOverlayTimelineV2.model_json_schema()

The same schemas ship under openadapt_types/schemas/ for TypeScript, Rust, and other consumers. Version 1 remains the control-state contract. Version 2 adds an optional normalized top-level viewport rectangle, the exact source viewport and DPR, and an exact observation or decoded-media-frame binding without changing V1.

Overlay schemas reject unknown fields and contain only closed presentation labels and canonical statuses. Screenshot payloads, action-target selectors, accessible names, text and values, typed input, identities, URLs, logs, report bodies, and user-authored workflow names remain outside this public contract. V2 may carry only normalized target geometry from a browser top-level CSS viewport. Native and RDP device-pixel geometry is not part of this V2 schema.

Target geometry never carries locators, accessible names, values, URLs, or screenshots. A private live observation uses a run/export-scoped HMAC reference instead of a linkable raw frame hash. Published media uses the exact media SHA-256 and decoded frame index. A renderer draws tracking only when that binding matches; it omits the rectangle rather than replaying selectors, interpolating movement, or inferring a missing target from adjacent events. The runtime does not guess a future viewer transform. Desktop, Cloud, and media renderers map the normalized rectangle through their actual content box. If multiple runtime states land in one decoded media frame, the producer must coalesce them deterministically; it must not invent extra media frames or approximate their timing.

Design principles

• Pydantic v2: runtime validation, JSON Schema export, fast serialization
• Pixels and structure: always capture both visual and semantic UI state
• Node graph: full element tree, not just the focused element
• Platform-agnostic: same schema for web, Windows, macOS, Linux, RDP, Citrix/VDI
• Extension-friendly: raw, attributes, metadata fields everywhere
• Backward compatible: _compat converters for gradual migration

Dependencies

Just pydantic>=2.0. No ML libraries, no heavy deps.

License

MIT