Warden catches a production error, finds the root cause, writes the fix, proves it, and ships a pull request. The founder just taps approve.
Warden is an autonomous on-call engineer for the millions of people shipping apps they can't debug. When production breaks, a non-technical founder can't read a stack trace, so Warden runs the whole loop for them:
- Detect: a Sentry webhook delivers the exception (HMAC-verified).
- Investigate: Warden reads the error, the stack, and the implicated code, read-only.
- Fix: a model writes a patch on a fresh branch in a per-incident workspace.
- Review: an independent panel of models from different labs cross-checks the diff, and they have to agree.
- Verify: Warden replays the exact failing request to confirm the original error is gone, then re-runs the target's existing tests as a regression check so the fix can't break what was passing. Nothing ships on a model's say-so.
- Approve: the founder gets a one-tap approval (web, push, or Slack). Approval is consent to ship, not a code review.
- Deliver: Warden opens a pull request (or merges) on the linked repo, and the team's existing CI/CD ships it.
Every step is written to an append-only audit log, and Warden recognizes errors it has seen before via vector search. It runs in simulation mode (fully offline and deterministic) or live mode against real services.
The app and its APIs run on Vercel; the pipeline engine runs on an AWS EC2 worker; and both coordinate through a single Amazon Aurora database, which holds the job queue, the incident state machine, the audit log, vector memory, and runtime settings.
| Layer | Choice |
|---|---|
| Framework / Language | Next.js 16 (App Router), React 19, TypeScript 6 |
| Styling | Tailwind CSS v4 |
| Database | Amazon Aurora PostgreSQL Serverless v2 with pgvector |
| Compute (engine) | AWS EC2, the always-on pipeline worker |
| Inference | OpenRouter (managed gateway: one key for Claude, GPT, Gemini, and more) via an OpenAI-compatible seam |
| Error source | Sentry (HMAC-verified webhook ingestion) |
| Delivery | GitHub (pull request or merge, OAuth or token) |
| Billing | Managed inference, metered from a prepaid wallet |
| Auth | Clerk (env-gated, optional) |
| Notifications | Web Push (VAPID), Slack (optional) |
| Testing | Vitest, gated on every build |
| Hosting | Vercel (app) plus AWS (Aurora and the EC2 worker) |
Aurora isn't just storage here; it's the product. A single Aurora PostgreSQL Serverless v2 cluster does quadruple duty:
- State machine: the incident lifecycle, enforced with transactional integrity.
- Append-only audit log: every agent action, for a complete trail.
- pgvector memory: Warden recognizes an error it has seen before.
- Agent scorecard and runtime settings: learning, plus live configuration.
We chose Aurora because the system needs relational integrity and vector search in one place (Postgres with pgvector), and Serverless v2 scales down between incidents, which is the right shape for bursty, event-driven work.
Prerequisites: Node.js 20+ LTS, Docker (for the local pgvector Postgres).
npm install
npm run setup # start Postgres (Docker), run migrations, seed data
npm run dev # http://localhost:3000
npm run worker # (optional) the background pipeline workerWarden defaults to simulation mode, so no external accounts or keys are needed. It runs the full detect, fix, verify, approve, and deliver loop offline against the bundled sample app. Open /dashboard, fire a sample incident, and watch the pipeline run.
app/
├── dashboard/ # Incident feed and detail, metrics, audit, settings, API keys, usage
├── api/ # Route handlers (Sentry ingest, approve, settings, oauth, cron, ...)
├── approve/[id]/ # Mobile one-tap approval screen
├── report/[id]/ # Shareable incident report
└── _components/ # Shared UI primitives
lib/
├── orchestrator/ # Pipeline state machine and job runner
├── agents/ # Investigator, fixer, reviewer panel (OpenAI-compatible seam)
├── adapters/ # Sentry, GitHub, Vercel, per-incident git workspace, request-replay
├── repo/ # Data access (incidents, events, jobs, scorecard, wallet, settings)
├── statemachine/ # Incident transitions
├── policy/ # Guardrails (blast radius, reversibility, SQL guard)
├── auth/ # OAuth registry, API-secret gate, sessions
├── memory/ db/ sim/ # Embeddings, Postgres client, seeded sandbox bugs
migrations/ # Aurora / Postgres SQL migrations (0001 to 0010)
scripts/ # migrate, seed, worker, demo
sample-app/ # Bundled, zero-dependency target app for the offline demo
certs/ # Vendored Amazon RDS CA bundle (Aurora TLS)
Dockerfile # The EC2 worker image
Set the following as environment variables (Vercel for the app, the worker's own environment for the engine). Full guidance is in docs/operations/go-live.md.
| Variable | Required | Purpose |
|---|---|---|
DATABASE_URL |
yes | Aurora connection string. TLS is verified against the vendored RDS CA. |
WARDEN_MODE |
yes | simulation (default, offline) or live. |
WARDEN_API_SECRET |
live | Gates approve, rollback, and cron mutations; fails closed in live mode. |
OPENROUTER_API_KEY |
live | Managed inference. The engine routes every model through it (backend-only). |
SENTRY_CLIENT_SECRET |
live | Verifies inbound Sentry webhook signatures. |
GITHUB_TOKEN / GITHUB_OAUTH_CLIENT_ID / ..._SECRET |
live | Clone the linked repo and open pull requests. |
BILLING_MODE |
optional | managed (metered wallet) or byok. |
VAPID_*, SLACK_*, CLERK_* |
optional | Push approvals, Slack approvals, dashboard auth. |
Never commit real keys.
.envis gitignored, and the worker reads its secrets from its own environment.
npm run build # type-check and Vitest, then next build
npm start # serve the production build- App on Vercel: push to
maindeploys production; pull requests get preview URLs. - Engine on AWS EC2: the worker runs
npm run workeron an always-on instance, since it needs git and a real filesystem to clone, boot, and replay. Seedocs/operations/deploy-aws.mdand theDockerfile. - Migrations in
migrations/are applied withnpm run migrate, not on deploy. A daily Vercel cron hits/api/orchestrator/tickas a recovery sweep; the EC2 worker is the real driver.
| Command | Does |
|---|---|
npm run dev |
Dev server (http://localhost:3000) |
npm run worker |
The background pipeline worker (drains the job queue) |
npm run build |
Type-check and tests, then production build |
npm test |
Run the Vitest suite once |
npm run lint / npm run typecheck |
ESLint / tsc --noEmit |
npm run migrate / npm run seed |
Apply migrations / seed data |
npm run setup |
Postgres up, migrate, and seed (one-shot local bootstrap) |
npm run demo |
Scripted end-to-end run (CLI) |
- Agents have no standing deploy authority: only a human-issued approval moves a verified fix toward shipping.
- Verification, not review, is the gate: the test passes and the original error stops reproducing, deterministically. Reviewer agreement is a filter; disagreement escalates and never overrides a failed check.
- Least privilege: agents read the database through a read-only path, and delivery credentials never reach a model.
- Every change is reversible, and when uncertain, Warden escalates rather than guesses.
Proprietary. © 2026 Arifur Rahman Akash. All rights reserved. No use, copying, modification, or distribution without written permission.
