Core System

Security Model

OpenCorpo's design goal is to let AI do useful operational work without giving unrestricted power.

1) Tool-Scoped Execution

The agent does not receive implicit full system access. It operates through registered tools.

• Tool definitions specify risk and required capabilities.
• Inputs are validated before invocation.
• Unknown tools are rejected.

2) Capability Grants

Runtime calls use capability sessions (x-oc-session).

• Calls without a valid grant are denied.
• Missing capabilities are denied and audited.
• Grants are revocable.

3) Policy Decisions

config/policy.json rules control action handling:

• allow: execute directly
• deny: block action
• approve: require human approval

High-risk tools are forced through approval paths unless policy explicitly allows.

4) Approval Workflow

Approvals exist for high-risk tool calls and certain code/config changes.

• Pending approvals can be approved or denied.
• Approved requests are executed by approval workers.
• Failed approvals are captured with error context.

5) Auditing and Diagnostics

OpenCorpo logs meaningful events, including:

• actor
• action
• policy rule metadata
• timestamps and runtime metadata

Diagnostics include audit integrity checks and repair endpoints.

Last updated from repo snapshot: 2026-02-20

Last updated: 2026-02-20