Platform / Review
The gate the runtime accepts. Routed, prioritized, due-dated, and written down.
Human Review is where a pipeline pauses for a person. A review block files an item and waits; an agent that is not sure files one instead of guessing; a colleague files a handoff from NoVA. The item carries a queue, a reason — policy, missing evidence, low confidence, returned, handoff, gate — a priority, an assignee and a due date.
Routing rules decide where it lands: first match wins, the rule is recorded on the item so the queue can say why. A person with the reviewer capability approves, edits and approves, returns with guidance, escalates or rejects. The pipeline continues on the approve port with the edited content, or leaves on the reject port where you wired a fix-up path.
A decided item can become a draft knowledge document or a test case, so the correction lands where the next run will find it. Every decision is an audit event with the decider’s name.
The approver who should not have to poll a log, the lead who decides who reviews what, and the auditor who asks who approved it.
Platform, on every tier; a human review meters at one Runtime Credit. The review block and the queue are the same on an App Kit instance.
Who may ask for a person. Items are created only on the server.
A control.review block in nQube files the item and waits, keeping the job alive and honoring cancel; the same in batch and streaming.
An answer under the guardrail’s confidence files a review with the app and agent on it instead of answering.
Any member, or NoVA on their behalf, can file “a person should look at this.” Asking for a human is not deciding.
What a reviewer can do, and who is one.
Approve, edit and approve, return with guidance, escalate, reject. Return and escalate keep the item open.
Set per member under Administration; without it the queue shows the item but hides the decision buttons and says why.
Each item has a link; the header bell counts unread items and opens the row.
Where it lands and what remains.
Match on queue, reason, app or priority; set queue, assignee, priority and a due time. Edits are numbered revisions with restore.
review.decided is written when the block sees the decision, with the decider as actor.
A decided item becomes a draft knowledge document or a test case, recorded on the item.
A review is one of the two gate kinds agentOS accepts in front of a write. A model is not.
Queue, reason, priority, assignee and due date from rules you set; escalation after a time you set.
A decision becomes a knowledge draft or a test case in one click.
A review block, an uncertain agent, or a person filing a handoff.
Queue, assignee, priority and due date from the first matching rule; the rule name rides on the item.
Approve, edit and approve, return with guidance, escalate, reject. Only the reviewer capability or an admin may decide; anyone may triage.
Approve leaves on one port with the edited text, reject on the other. A timeout you set rejects on its own.
| The usual way | Here |
|---|---|
| Approval is a Slack message and a thumbs-up emoji | A routed item with a reason, an assignee, a due date and an audit event with the decider’s name |
| Rejection ends the run silently | Reject is a port on the block; wire a fix-up path or a notice |
| The correction lives in someone’s head | One click makes it a knowledge draft or a test case |
No. A review is a person. The other gate the runtime accepts is a keyword filter. A filter that asks a model is not a gate.
One Runtime Credit per human review, on the plan’s credit rate.
Anyone can triage and file. Deciding needs the reviewer capability or an admin role.
Platform Start is the self-serve canvas. A first pipeline we stand up with you is delivery, priced apart from the subscription.