Skip to content

Latest commit

 

History

History
45 lines (33 loc) · 1.64 KB

File metadata and controls

45 lines (33 loc) · 1.64 KB

Constraint Net Protocol

Constraint Net is a discovery, planning, consent, execution, and receipt protocol for agent-safe actions.

The core idea is simple: agents should not scrape pages or blindly call APIs for side-effectful work. They should discover publisher-declared capabilities, plan a coherent path under constraints, ask for consent when risk requires it, execute with replay protection, and return receipts that can be verified later.

Publisher Manifests

Publishers expose an actions manifest at:

https://<publisher-domain>/.well-known/constraint-net/actions.json

The manifest declares actions, schemas, planning metadata, risk tier, reversibility, confirmation policy, OpenAPI bindings, key discovery, and signatures.

Trust Rules

  • Manifests must validate against schema version 0.1.
  • Manifests must be active, unexpired, and signed.
  • The manifest publisher domain must match the discovered domain.
  • Tier 2 side effects must be reversible and require confirmation.
  • Side-effectful OpenAPI actions must require idempotency.
  • Execution must pin the manifest digest observed at preflight.

Agent Flow

  1. Discover a publisher manifest.
  2. Validate and ingest it.
  3. Plan a capability path from the user goal.
  4. Preflight a selected step.
  5. Ask the user to confirm Tier 2 side effects.
  6. Execute with an idempotency key.
  7. Verify intent, consent, and execution receipts.

Local Endpoints

  • POST /v1/manifests
  • POST /v1/manifests/ingest-url
  • POST /v1/actions/search
  • POST /v1/executions/preflight
  • POST /v1/confirmations/:id/decision
  • POST /v1/executions
  • GET /v1/receipts/:id
  • POST /v1/receipts/verify