Deploy a governed workflow to production

Pass the full quality gate, make your webhook receiver idempotent, run non-mutating deployment readiness, and wire observability so agent quality is evaluated from evidence, not vibes. Prerequisites: Railway and Cloudflare access for the services you own.

$ pnpm hooks:install$ pnpm format:check$ pnpm lint$ pnpm typecheck$ pnpm testpnpm quality:full

Handle webhook deliveries safely

• Verify the signature against the raw request body before you parse it as JSON.
• Deduplicate by delivery identifier before applying any side effect.
• Return a successful response only after the delivery is durably accepted on your side.
• Expect exponential retry and out-of-order delivery; reconcile from canonical resource state, not delivery order.

Run deployment readiness

$ pnpm readiness:railway$ pnpm readiness:railway:variables$ pnpm readiness:railway:source$ pnpm readiness:cloudflare$ pnpm cloudflare:inventory

Wire observability and gate on evaluations

OpenTelemetry

Cross-service spans for API, tools, sandboxes, and on-chain execution.

Langfuse

Prompt and model traces where policy permits.

Braintrust

Datasets, prompt operations, evaluation, and improvement cycles.

Sentry

Exceptions and production failure diagnosis.

Run ledger

Canonical event sequence for status, approvals, artifacts, and receipts.

$ pnpm eval:agent-harness$ pnpm test:policy$ pnpm readiness:provider-smokes$ pnpm braintrust:improvement-cycle