openadapt-types

Pydantic models for computer-use agents: what is on the screen, what to do to it, and what happened when you did. Use them when you build an agent, a screen recorder, or a benchmark harness and want one validated format for observations and actions, plus JSON Schema for code that isn't in Python.

It needs Python 3.10 or later, and its only dependency is Pydantic 2.

Quick start

pip install openadapt-types

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

# What is on the 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)),
    ],
)
print(state.to_text_tree())
# [n0] window: My App
#   [n1] button: Submit

# What to do to it
action = Action(type=ActionType.CLICK, target=ActionTarget(node_id="n1"))
print(action.model_dump_json(exclude_none=True))
# {"type":"click","target":{"node_id":"n1","is_normalized":false}}

to_text_tree() returns the element tree as text you can put straight into a prompt. Every model checks its fields when you create it, round-trips through JSON, and gives you a JSON Schema from model_json_schema().

OpenAdapt's benchmark harness (openadapt-evals) and training code (openadapt-ml) use these types. If you want to show a task and have it run again on its own, install OpenAdapt (pip install openadapt) instead.

What's in it

The same screen model covers web, Windows, macOS, and Linux. It keeps the screenshot and the element tree together, because each holds information the other lacks. raw, attributes, and metadata fields let you carry your own data through without forking the schema.

Targeting

ActionTarget accepts three kinds of answer to "which element". Prefer them in this order:

from openadapt_types import ActionTarget

ActionTarget(node_id="n1")                          # an element in the tree
ActionTarget(description="the blue submit button")  # resolved later by a grounding model
ActionTarget(x=550, y=420)                          # pixel coordinates, last resort
ActionTarget(x=0.29, y=0.39, is_normalized=True)    # coordinates as a fraction of the screen

Have your agent produce a node_id or a description and work out the pixels when the action runs. Coordinates break when a window moves.

Read an action from a model's reply

from openadapt_types import parse_action

action = parse_action("Thought: the form is complete.\nAction: CLICK(x=0.29, y=0.39)")
print(action.type.value, action.target.x, action.target.y)  # click 0.29 0.39
print(action.reasoning)                                     # the form is complete.

print(parse_action("I'm not sure what to do").type.value)  # fail

parse_action() reads the NAME(arg=value) form shown here and JSON, including JSON inside a code fence. When x and y are both between 0 and 1, it reads them as fractions of the screen; otherwise it reads them as pixels. A reply it can't read becomes a fail action with the reason in reasoning, and the parser logs a warning. A harness can't mistake an unreadable reply for a finished task.

Report how a run ended

plain_outcome() turns the status values from an OpenAdapt run into one result a person can act on:

from openadapt_types import plain_outcome

result = plain_outcome(
    transaction_outcome="RECONCILIATION_REQUIRED",
    execution_outcome="HALTED",
)
print(result.label)          # Check the record
print(result.next_action)    # A save may have gone through. Check the record before anyone runs this again.
print(result.safe_to_retry)  # False

An unknown or contradictory status reads "Check the record", never "nothing was written". The table that maps status values to results also ships as JSON for other languages, and the openadapt-vocab-lint command checks user-facing text against it. See docs/VOCABULARY.md.

JSON Schema for other languages

import json
from openadapt_types import ComputerState

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

The messages that OpenAdapt components exchange ship under openadapt_types/schemas/: forty files, including execute-v1-openapi.json, the OpenAPI document for OpenAdapt Execute, the API that starts runs and reports their results. Two files hold test vectors, and the rest are JSON Schema files. docs/CONTRACTS.md lists each contract, its Python model, and what it may carry.

Convert from older formats

from openadapt_types import from_benchmark_action

action = from_benchmark_action({"type": "click", "x": 0.5, "y": 0.3})
print(action.type.value, action.target.is_normalized)          # click True
print(from_benchmark_action({"type": "teleport"}).type.value)  # fail

from_benchmark_action() reads an openadapt-evals action dict. The other converters are in openadapt_types._compat: from_benchmark_observation() for openadapt-evals observations, from_ml_* for openadapt-ml, and from_omnimcp_* for omnimcp. Each converter takes a plain dict, such as obs.__dict__. An action type it doesn't recognize becomes a fail action, never done.

Limits

• The API can still change between minor versions before 1.0. Pin the minor version you test against.
• ComputerState.screenshot_png holds raw bytes, which don't convert to JSON: model_dump_json() and Episode.to_json() raise an error. To save a state or an episode as JSON, put the screenshot in screenshot_path or screenshot_base64 instead.
• In the NAME(arg=value) form, parse_action() reads only x and y as a target. CLICK(node_id="n1") returns a click with no target. Use the JSON form for node_id and description targets.
• Some rules compare one field with another, such as "a result that says nothing changed can't also say something may have changed". The Python models enforce them. The JSON Schema, from the files or from model_json_schema(), doesn't always, so code in another language has to check them itself.

License

MIT