Production Capstone: Build the Complete SupportOps Harness
The final SupportOps system accepts tenant-scoped requests, routes to an eligible OpenAI or Claude adapter, exposes only policy-approved tools, persists events and checkpoints, pauses exact side effects for approval, survives retries and restarts, verifies outcomes, exports sanitized telemetry, and blocks release unless deterministic, behavioral, adversarial, and operational gates pass.
What will you be able to do?
You will implement and demonstrate a complete vertical slice whose behavior, authority, cost, and recovery can be explained from evidence.
What is the system map?
Rendering diagram...
What should the repository contain?
In which order should you build it?
- Domain schemas and invariants.
- Fake adapter and deterministic harness tests.
- Read-only tools and policy gateway.
- Event store, budget, and checkpoint reducer.
- OpenAI adapter contract.
- Claude adapter contract.
- Proposal/approval/idempotent commit.
- Queue/worker recovery.
- API and reconnectable events.
- Telemetry/redaction and operator views.
- Evaluations and security tests.
- Sandbox deployment and progressive release.
What are the acceptance scenarios?
- Read-only billing and technical tasks on both providers.
- Correct grounding and citations.
- Denied cross-tenant lookup.
- Approval pause, authorized resume, exact-once mutation, verification.
- Rejected and expired approval.
- Worker crash before and after mutation.
- Provider outage and eligible pre-action fallback.
- Budget, cycle, fan-out, cancellation, and deadline stops.
- Prompt-injected document and compromised-tool simulation.
- Redaction canary absent from telemetry.
- Rollback to prior release manifest.
Production checkpoint: Record a demo evidence pack: threat model, policy matrix, architecture decision records, event timeline, trace, evaluation report, load/chaos report, SBOM/image scan, approval record, reconciliation proof, dashboard, and runbook exercise.
How does it fail safely?
If required evidence, policy, capability, provider, storage, or approval is unavailable, the run becomes blocked, queued, limited, or failed. It never silently downgrades to ungrounded prose or unapproved action.
What mistakes should you avoid?
- Integrating live providers before fake-adapter contracts pass.
- Treating the happy path as the capstone.
- Building one universal tool with internal string commands.
- Shipping without operator cancellation and quarantine switches.
- Declaring success before recovery and adversarial evidence exists.
Check your understanding
- Why is the fake adapter built first?
- When may the router switch providers safely?
- Which artifacts prove production readiness?
Expert extension
Present the evidence pack to a reviewer who did not build the system. Ask them to trace one side effect and one failure end-to-end. Any unanswered question becomes a backlog item.
Official references