Skip to main content

Policy Service

What it does​

The policy service is the authorization authority in Manasvi. Every proposed action — every tool call, every side effect — must be evaluated by the policy service before it can proceed.

The service takes in a description of the proposed action and its context, evaluates it against configured rules, and returns a decision.

Policy evaluation request​

When the agent runtime wants to authorize an action, it sends a policy evaluation request containing:

  • Actor — who is making the request (user ID, trust level, channel)
  • Tool — which tool is being called
  • Action class — the category (read, write, network access, shell execution, etc.)
  • Resource — what specific resource is targeted
  • Parameters — the full parameters of the proposed action
  • Session context — current session state, previous approvals in this conversation

Policy decision​

The service returns:

{
  decision: "allow" | "allow_with_approval" | "deny",
  decisionId: string,        // unique ID recorded in audit trail
  reasonCodes: string[],     // why this decision was made
  requiredApprovalLevel: string | null
}

The decisionId is embedded in the execution intent that's issued after an allow decision. This creates an unbreakable link between the authorization decision and the action that was executed.

Policy rule layers​

Rules are evaluated in priority order:

  1. Hard denials — actions that can never be permitted regardless of configuration
  2. System defaults — built-in rules for known risk levels
  3. Operator rules — deployment-specific configuration
  4. User rules — per-user permissions

A deny at any layer is final. A lower-priority layer cannot override a higher-priority denial.

Fail-closed behavior​

The policy service is designed to fail safely:

  • If the service is unreachable, the default decision is deny
  • If a rule evaluation throws an error, the default decision is deny
  • If required context is missing, the default decision is deny

The agent runtime treats any failure to obtain a valid policy decision as a denial and will not proceed with the action.

Audit integration​

Every policy decision is recorded in the audit trail with:

  • The full request parameters (redacted for sensitive fields)
  • The decision and all reason codes
  • The timestamp
  • The evaluating policy version

This means you can always reconstruct exactly why any action was authorized or denied.