getaxonflow/axonflow ? reverse-engineered prompt

Build me a self hosted control layer for AI workflows that sits between my app and the model or tool calls.

I want it to let me decide if a step is allowed, blocked, paused, or resumed, and to keep a clear audit trail of why that decision happened. It should support simple gateway style checks for existing apps, a proxy mode that can sit in the request path, and a workflow mode for multi step agent runs with approvals, retries, and step level visibility.

Please include policies for things like sensitive data, unsafe actions, usage limits, and basic compliance rules, plus a way to run in observe only mode before turning on blocking. I also want human approval gates, decision logs, replay for debugging, and exports for audits.

Make it work well as a Docker based service, with APIs and SDK friendly structure, and use current docs online if you need to fill in any details.