openadapt-types

Pydantic schemas for describing a screen and an action on it. Recorders emit them, the compiler stores and replays them, the grounding package resolves their targets, and the privacy package scrubs them. One definition, so nothing in the stack has to translate.

pip install openadapt-types

Optional, and not the product. If you want to record and replay a workflow, you want openadapt-flow instead; this is the contract underneath it, published separately for anyone building their own computer-use agent. The API is not stable across minor versions yet.

Describe a screen, then act on it

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

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)),
    ],
)

print(state.to_text_tree())
# [n0] window: My App
#   [n1] button: Submit

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

to_text_tree() exists so you can drop the element tree straight into a prompt.

Targeting

ActionTarget takes three kinds of answer to "which thing", and the runtime prefers them in this order:

ActionTarget(node_id="n1")                          # an element the recorder saw
ActionTarget(description="the blue submit button")  # resolved by the grounder
ActionTarget(x=550, y=420)                          # coordinates, last resort
ActionTarget(x=0.29, y=0.39, is_normalized=True)

An agent should produce a node_id or a description and let the runtime work out the pixels. Coordinates are the thing that breaks when a window moves.

The rest of the types

Plus the versioned wire contracts: ControlOverlayFrameV1/V2 and ControlOverlayTimelineV1/V2 for PHI-safe execution overlays, ExecuteRequestV1 / ExecuteStatusV1 / ExecuteEvidenceReceiptV1 for asynchronous qualified execution, EffectStrengthV1, and the BusinessDecision*V1 family for signed, finite human choices. What those contracts may and may not carry is in docs/CONTRACTS.md.

JSON Schema for everything else

import json
from openadapt_types import ComputerState

print(json.dumps(ComputerState.model_json_schema(), indent=2))

The same schemas ship as JSON under openadapt_types/schemas/ for TypeScript, Rust, and anything else that isn't Python. Seventeen files, including execute-v1-openapi.json, the public OpenAdapt Execute contract.

Converting from the older 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
)

state = from_benchmark_observation(obs.__dict__)

OpenAdapt Execute

A partner sends an authorized request, gets an execution ID back, and then either reads a terminal receipt or waits for a signed webhook. The contract exposes no runner, no customer data, no evidence bytes, and no control-plane internals.

from openadapt_types import ExecuteClient

client = ExecuteClient(
    base_url="https://app.openadapt.ai/api",
    bearer_token="partner-provisioned-token",
)

The client requires HTTPS and refuses to follow a redirect before it would send the bearer token somewhere else. It also doesn't poll forever, deliberately: a workflow can sit waiting for a human decision or a reconciliation, so treat the terminal receipt or the signed webhook as the completion signal, not a timeout.

Reference Python and TypeScript clients: examples/execute.

Design

Pydantic v2, so you get runtime validation, JSON Schema export, and fast serialization. Pixels and structure are both captured, always, because either one alone loses information the other has. The node graph is the full element tree rather than the focused element. The same schema covers web, Windows, macOS, Linux, RDP, and Citrix/VDI. raw, attributes, and metadata fields exist everywhere so you can carry your own data through without forking.

The only dependency is pydantic>=2.0. No ML libraries.

License

MIT