ActionBoxDOCS

Integration Recipes

Practical ActionBox patterns for agent interrupts, CI approval gates, callbacks, and run supervision.

Choose a recipe based on who owns the paused workflow state. In every recipe, ActionBox owns the durable human request and typed response; the calling system keeps its own state and performs the side effect.

Agent frameworks

Frameworks such as LangGraph and the OpenAI Agents SDK keep their own checkpoint or paused-run state. ActionBox supplies the durable human request and typed response.

CI and deployment workflows

Use the GitHub Actions API pattern or the blocking CLI when a runner must wait. Make the deployment a separate step that runs only after the recorded decision matches the approved option.

Asynchronous services

Create an Action with a stable idempotency key and an HTTPS callback URL when the service needs asynchronous notification. Verify the callback, deduplicate delivery, re-read current Action state when necessary, and report the real execution outcome after the operation finishes.

Background and scheduled work

Use Heartbeat Watches to detect missing progress. A Watch raises an ordinary Action when its signal is late; it does not restart your infrastructure or invent an execution result.

Safety rules shared by every recipe

  • Keep Source keys, callback secrets, framework state, and tool credentials in trusted runtime configuration, never in Action content.
  • Bind retries to the same logical operation with an idempotency key.
  • Treat rejection, expiry, cancellation, malformed responses, and timeouts as non-approval unless your explicit expiration interaction says otherwise.
  • Execute the side effect only after reading the authoritative terminal Action.
  • Record execution success or failure separately from the human decision.

On this page