Before serving users, make sure each part of the request path has a responsible team and a practical health check.
Assign operational responsibility
| Layer | Responsibility | Readiness evidence |
|---|---|---|
| Application server | User authentication, trusted context, public routes | Allowed and denied request tests |
| Runner service | Authenticated execution and public contracts | Health, capabilities, command, and model checks |
| Runtime and store | Durable sessions, events, terminal outcomes, replay | Known-session completion and later inspection |
| Model access | Approved provider, model, credential, and deployment | Qualification and short test run |
| Observability | Correlation across application and runner | Searchable request, session, and run identifiers |
| Operations | Incident action, recovery, and escalation | Repeated user-path verification |
Release a change
- Record the versions and intended behavior.
- Verify authentication and denial paths.
- Run one terminal and one streaming request through the production route.
- Confirm that the result, artifacts, and traces are inspectable after completion.
- Exercise cancellation or waiting behavior if the application supports it.
- Run the relevant Ruhroh evaluation and product quality gates.
Do not declare readiness from a health endpoint alone. A healthy process can still have broken user context, model access, persistence, or terminal handling.