Architecture

The service simulation platform plan is the authoritative design.

xEntropy stores service definitions, workflow DAGs, test cases, and run history. A shared browser-safe execution library runs the DAG in either the browser or the user's Node.js process. It calls real endpoints directly and submits simulated inputs to the hosted graph API. It never uploads the workflow to a VM for command execution.

One workflow run is one session. Each simulated service instance owns isolated state within that run. Repeated calls to that instance are ordered; independent branches can execute concurrently. The runner checks assertions against actual outputs. Expected results never enter the agent context.

The hosted API verifies account and credential scope, freezes run inputs, admits idempotent simulation requests within published budgets, and delegates each invocation to durable Eve execution. Agent output and new service state are validated and committed together. Client-reported real-service observations are labeled separately from server-owned simulation evidence.

Neon/PostgreSQL transactions coordinate requests across API instances. The new runtime does not use the previous whole-ledger process ownership lock. Clerk retains account identity. The production Eve runtime is independent of benchmark agents.

Requests can outlive HTTP connections. Polling resumes observation without re-executing admitted calls. Uncertain dispatch is recorded explicitly, never retried blindly. Ending a run clears active simulation state; history is retained as evidence rather than resumed in the next run. Interrupted runs may resume; abandoned runs expire after 24 hours.

The old command-workspace execution paths have been removed. Existing database tables and historical evidence are preserved without enabling fallback execution.