Actionbox Python SDK

Actionbox gives backend services a durable, server-authoritative way to ask a human for a decision and continue when that decision is available. This package is the typed Python client for creating, resolving, and waiting on Actions, plus managing source-scoped heartbeat Watches.

Documentation

• Actionbox documentation

Requirements

• Python 3.11 or newer
• An Actionbox Source API key, supplied through ACTIONBOX_API_KEY

Keep API keys and Watch capability URLs on trusted servers, workers, or CI jobs. Do not put this SDK or its credentials in browser code.

Install

pip install actionbox-sdk

The distribution is named actionbox-sdk so it does not conflict with the Actionbox CLI distribution on PyPI. The Python import remains actionbox.

import os
from actionbox import Actionbox

with Actionbox(os.environ["ACTIONBOX_API_KEY"]) as client:
    decision = client.ask(
        title="Deploy to production?",
        options=["Approve", "Reject"],
        callback_url="https://ci.example.com/actionbox",
    )
    print(decision)

The SDK uses the hosted production API at https://api.actionbox.cloud by default. Customer integrations should use that default and only pass base_url in maintainer-controlled test environments.

ask(..., wait=False) returns an Action; Action.wait() polls the server and leaves the Action open when the local timeout expires. The concise single-choice API returns the selected option ID as a string.

Typed interactions and responses

The SDK exports typed interaction and response contracts that match the REST API. Use an explicit typed interaction with create or ask when the human response is more than a single choice:

from actionbox import Actionbox, BooleanInteraction

with Actionbox(os.environ["ACTIONBOX_API_KEY"]) as client:
    action = client.create(
        title="Deploy configuration",
        interaction=BooleanInteraction(
            type="boolean",
            label="Deploy now?",
            true_label="Deploy",
            false_label="Hold",
        ),
    )
    resolved = client.resolve(
        action.id,
        response={"type": "boolean", "value": True},
        reason="Approved by release manager",
    )
    print(resolved.response)  # {"type": "boolean", "value": True}

The available interaction types are boolean, single_choice, multi_choice, text, integer, number, rating, and form. Form fields use the same typed field shapes and are returned as {"type": "form", "values": {...}}. Create inputs also accept bounded developer context blocks and an explicit typed on_expire fallback; omitting it returns expired without inventing a response.

resolve supports concise single-choice syntax and generic typed input:

client.resolve(action.id, "approve")  # single-choice shorthand
client.resolve(action.id, {"response": {"type": "text", "value": "ship"}})
client.actions.resolve(action.id, response={"type": "number", "value": 4.5})

Action.interaction and Action.response expose the canonical typed wire values. options, option_id, ask(..., options=[...]), and string decision results are first-class single-choice conveniences.

Optional decision context

Keep simple Actions unchanged. For higher-impact reviews, use a helper that builds the same generic structured context accepted by the REST API:

from actionbox import deployment_decision_context

action = client.create(
    title="Deploy 2.18.0?",
    decision_class="production_deployment",
    decision_context=deployment_decision_context(
        reason="Release passed staging.",
        proposed_change="Deploy 2.18.0 to production.",
        risk_level="high",
        reversibility="reversible",
        rollback_plan="Restore the previous image.",
    ),
)

The generic, refund, database-change, and access-request helpers emit this same wire shape; they do not create server-side template types.

Agent framework integrations

The OpenAI Agents SDK and LangGraph keep their own paused run state; Actionbox supplies the durable human request and typed response. The repository includes tested examples for both patterns:

• OpenAI Agents: map result.interruptions to Actions, apply each decision to result.to_state(), then resume the original agent.
• LangGraph: create Actions after interrupts surface to the graph driver, then resume the same checkpoint and thread_id with Command(resume=...).

See integrations/agent_frameworks/ in the Actionbox repository. No additional Actionbox endpoint or framework-owned state migration is required.

Execution outcomes

After carrying out an approved operation, report its real result from the resolved Action snapshot:

outcome = resolved.report_outcome(
    "success",
    duration_ms=48_312,
    rollback=False,
)

The SDK sends the Action's exact version and fingerprint. Exact retries are safe; Actionbox rejects a conflicting second outcome.

Action Controls

Paid plans can attach a few secondary operations to an Action. A control does not answer or close the Action. It asks the Source to do something, such as retrying a job, while the reviewer keeps the original decision open.

from actionbox import control, link

action = client.create(
    title="Deployment failed",
    callback_url="https://ci.example.com/actionbox",
    controls=[
        control("retry", "Retry deployment"),
        control("rollback", "Roll back", destructive=True),
        link("logs", "Open logs", "https://ci.example.com/runs/4821"),
    ],
)

The signed action.control_requested webhook includes a control_request_id. Report the operation state with the Source client:

client.report_control_result(
    control_request_id,
    "running",
    message="Retry started",
)
client.report_control_result(
    control_request_id,
    "succeeded",
    message="Deployment recovered",
)

Use failed when the operation does not complete. ActionBox records the result without resolving the parent Action. During staged rollout, the API may return CONTROLS_DISABLED until the feature is enabled for the paid workspace. ActionBox never runs an infrastructure command itself. Your callback handler maps each control key to an operation and reports the result. A later Watch heartbeat or meaningful Agent Run update can add independent recovery evidence. That evidence does not replace the Source's reported result.

Agent Runs

Group an agent task and its Actions without managing another framework:

run = client.runs.start(
    external_id="checkout-fix-42",
    agent_name="codex",
    title="Fix checkout deadlock",
    stall_after_seconds=900,
)
run.progress(stage="tests", checkpoint="test-184")
action = client.create(
    title="Approve staging migration",
    run_id=run.id,
    callback_url="https://agent.example.com/actionbox",
    controls=[control("resume", "Resume run")],
)
run.complete()

The SDK handles progress sequence numbers. Runs are optional; standalone Actions continue to work exactly as before. When stall_after_seconds is set, unchanged status/stage/checkpoint updates do not reset the timer. Actionbox creates one ordinary Action if progress stalls and resolves it when progress changes or the Run completes; waiting pauses the timer. Meaningful progress or completion also verifies the latest delivered control on an Action linked to that Run.

Heartbeat Watches

Source credentials can create and list Watches scoped to that Source. The raw heartbeat URL is returned only by creation:

from actionbox import Actionbox, send_heartbeat

with Actionbox(os.environ["ACTIONBOX_API_KEY"]) as client:
    watch = client.watches.create(
        name="Nightly backup",
        schedule_type="interval",
        interval_seconds=3600,
        grace_seconds=60,
        signal_method="post",
    )
    send_heartbeat(watch.heartbeat_url, "start")

Use send_heartbeat for ping, start, success, or fail; it always sends POST. signal_method="post" prevents link previewers and security scanners from accidentally recording a heartbeat with GET. The source-scoped resource also exposes client.watches.pause(id), resume(id), rotate_token(id), and archive(id); only create/rotate return a raw URL. Store capability URLs in a secret manager; Watch details and exports never return them. When Action Controls are enabled for a paid workspace, a Watch incident also offers Skip this occurrence and Pause monitoring. Skipping closes only the current incident and advances the Watch schedule. A later healthy signal can independently verify a delivered control request.

License

MIT. See LICENSE.