AgentTwin 0.2.2 · discovery and hosted workflow execution · Node.js 22.12+
Fidelity contract
These concepts describe the broader product contract. The current runtime uses explicit fixtures, supported operation mappings and reprovisioning. Automatic shared-identity compilation, coordinated snapshots and comprehensive provider conformance are not shipped. See current capabilities.
A twin that silently diverges from production can create false confidence. A replica should publish a versioned manifest that declares its limits, and every operation needs one of three labels.
Verified
Behavior compared against a real provider's allowed test environment.
Modeled
Behavior implemented from documentation or observed examples, with assumptions visible.
Unsupported
Behavior unknown or outside the scenario. Tests must not silently pass through it.
What the manifest records
- Provider version the twin was built against.
- Fixture source for starting state and seed data.
- Last verification date — when the behavior was last compared with reality.
- Configuration fields that were captured, and which were not.
- State boundaries — what the replica covers and where it ends.
- Known differences from the real provider.
Evidence, not just a response
A passing response cannot hide a bad write. Every run should show:
- A trace of provider calls and state transitions, in causal order.
- A before/after state diff across the systems involved.
- Assertions against final state and invariants, not only the agent's explanation of what it did.
The three labels and the manifest are a proposed contract. No provider twin has been verified yet, and this repository contains no adapters or verification evidence. Until then, results shown in the prototype are scripted.