The approvals command group is the human side of workflow checkpoints — the “keep one door open” principle as a first-class primitive. A workflow that produces something risky or publishable (an AI-written article, a bulk data change) can gate the final action behind a human: it files an approval request and stops. You list what’s waiting, read it, and approve or reject. On approve, the stored action runs in-pod.
Usage
The flow
- A producer workflow composes a request with the
compose_approval node (title, preview, and what should run if approved — a workflow name + inputs), saves it as an approval-request document, and posts the ask to the pod’s Slack channel.
- A human sees it (Slack or
approvals list) and decides.
approve/reject execute the pod’s machina-approval-resolve workflow — approve dispatches the stored action; reject just records the decision. Double-resolution is guarded: a request resolves exactly once.
Resolution logic lives in the pod, not the CLI. approve/reject just execute the machina-approval-resolve workflow with {request_id, decision}. Any surface — this CLI, the Studio, an MCP agent — resolves through the same workflow, so the guard rails (double-resolution, dispatch recording) are shared.
Gating your own workflow
The request side is three tasks from the machina-nodes library (see the provisioned machina-approval-demo workflow for a copyable example):
compose_approval — inputs: title, preview, action_workflow, action_inputs.- A
document save task for the composed approval-request.
slack_notify with the composed ask.
An actionless request (empty action_workflow) is also valid — a pure human acknowledgment gate.
loop — the harness loop’s needs_review is the same principle at the conversation level; approvals brings it to any workflow.- Provisioning:
docs/harness-loop-kit/nodes.py in the machina-cli repo (python3 nodes.py against the pod).
