Approvals

Human guardrails for high-impact Consuelo OS actions.

Approvals

Approvals let skills pause before high-impact actions.

The skill prepares the action, OS records the request, a human makes the decision, and the skill continues or stops. The approval belongs to OS, so the agent process does not need to stay alive while the human decides.

When approvals are useful

Use approvals when a skill is about to do something that changes shared state, publishes something, affects an external system, or creates a result other people will trust.

Examples:

• publish or share an artifact
• create campaign drafts in a connected system
• update important workspace records
• send a customer-facing message
• trigger a workflow another person depends on

Approval flow

skill prepares action
  -> OS creates approval request
  -> OS delivers request to a human surface
  -> human approves or rejects
  -> OS records the decision
  -> skill continues or stops

Start simple. The first approval surface should support approve, reject, and details. Editing can come later once action-specific forms exist.

Approval request shape

Field	Purpose
id	Stable approval identifier.
workspaceId	Workspace that owns the request.
skillExecutionId	Execution that requested the decision.
skillName	Skill asking for approval.
title	Human-readable decision title.
summary	What will happen if approved.
riskLevel	Internal impact level for sorting and policy.
status	Pending, approved, rejected, expired, or cancelled.
proposedPayload	Structured action details.
traceId	Link back to execution logs.
createdAt	Request creation time.
resolvedAt	Human decision time.

Delivery surfaces

Approval delivery should be adapter-based. Skills should ask OS for approval. Skills should not know whether the request goes to the app, Slack, Discord, or another surface.

skill
  -> approval service
    -> approval record
    -> delivery adapter

Resume model

The simplest useful model is explicit resume. A skill returns approval_required. The human decides. The agent or user then continues the skill with the approval reference.

Automatic wake-up can come later after OS has durable execution checkpoints.

Product rules

Approvals should be clear and specific. A request should explain what will happen, which workspace it affects, which skill asked for it, and what the human can do.

Approvals should be rare enough to keep OS useful and common enough to keep high-impact actions safe.

Related

• Skills
• OS portal
• Observability