Codex Guard

Standards nobody reads are just decoration. Codex Guard makes them read your code back.

Paste a Dockerfile, CI workflow, Kubernetes manifest, or TypeScript / Go / shell file. Get a scored review in seconds, with the exact lines that broke the rules, a fix for each one, and a chat assistant that remembers your team and your exceptions.

Built on Cloudflare Agents, Workflows, Durable Objects and Workers AI (Llama 3.3), for the Developer Productivity assignment. It's a miniature of the job itself: turn engineering standards into automated guardrails.

The rules in src/codex.ts are examples I wrote for this project. They are not Cloudflare's real Engineering Codex.

Try it in 60 seconds

1. Open the live demo, choose Load a sample → Dockerfile, and click Run review.
2. Watch the pipeline light up step by step. You should end at 38 / 100 with three failures: a hard-coded token, node:latest, and a container running as root.
3. Click Request a waiver on the root finding, type a reason, and save it.
4. Run the review again. The finding now reads Waived and the score jumps to 63.
5. Ask the chat "what waivers do we have?" It knows, and it knows your team and stack too.
6. Refresh the page. Everything is still there.

The one rule that makes it trustworthy

Evidence beats opinion.

LLMs are great at explaining a problem and unreliable at deciding whether it exists. So every rule has a deterministic check (policy-as-code) that collects hard evidence, with line numbers. The model explains and judges only the fuzzy cases. If the check finds a violation, the model cannot talk its way to a pass. A prompt injection hidden in your Dockerfile changes nothing.

Rule type	Who decides	Example
Authoritative	The check alone	Unpinned :latest, missing USER, no permissions: block
Fuzzy	Check first, then the LLM	Timeouts on outbound calls, swallowed errors, secrets in odd shapes

What happens when you click Run review

• One flaky model call doesn't sink the review. Each rule is its own workflow step with retries and a timeout. A garbled model reply downgrades one rule to "needs review" instead of failing everything.
• State lives with the user. Each browser gets a room, and each room is its own Durable Object. There are no accounts in this demo, so anyone with a room id can open that room.

Checklist

Requirement	Where it lives
LLM (Llama 3.3 on Workers AI)	src/llm.ts, using @cf/meta/llama-3.3-70b-instruct-fp8-fast; streaming for chat, one-shot for reviews
Workflow / coordination	src/workflow.ts (Cloudflare Workflows), started and observed by the agent in src/agent.ts
User input via chat	public/index.html: a WebSocket chat plus a review form
Memory or state	Agents SDK state (profile, waivers, active review) plus SQLite history, inside one Durable Object per room

Run it yourself

You need Node 22+ and a free Cloudflare account. These are the exact commands I used, in order.

npm install                    # dependencies
npx wrangler login             # authenticate with Cloudflare (opens your browser)
npm run dev                    # local server at http://localhost:8787

npm run typecheck              # generates binding types, then tsc
npm test                       # unit tests for the policy-as-code checks
npx wrangler deploy --dry-run  # bundle everything without deploying

npm run deploy                 # ship it to *.workers.dev
npx wrangler tail              # stream live production logs (Ctrl+C to stop)

Good to know:

• npm run dev runs the Worker, Durable Object and Workflow locally, but the Workers AI binding always calls your Cloudflare account, so LLM calls need wrangler login and count toward your usage.
• wrangler tail only shows the deployed app. Start it, then use the live URL.
• Local and live have separate data. A waiver saved on localhost won't exist on workers.dev.
• .github/workflows/ci.yml runs typecheck, tests and a deploy dry run on every push. It follows the same standards the app enforces: pinned versions, least-privilege permissions, and tests before deploy.

Field notes

Local dev sometimes printed The Workers runtime canceled this request because it detected that your Worker's code had hung. It appeared even on a fresh start with a single tab, while every feature kept working. In production, wrangler tail showed Ok on every event and no errors, so I treated it as a local dev-server quirk and moved on. But still Working on it ..

Built with AI

I built this with Claude and kept the full record in PROMPTS.md. I tested it locally and in production, and I can walk through every file.

Project layout

src/index.ts       Worker entry: routes /agents/* to the agent
src/agent.ts       CodexAgent: chat, memory, workflow orchestration
src/workflow.ts    ReviewWorkflow: the durable review pipeline
src/codex.ts       Rules, file detection, deterministic checks, scoring
src/llm.ts         Workers AI helpers (streaming, JSON extraction)
public/index.html  Chat and review UI (no build step)
test/              Unit tests for src/codex.ts

To try another model, change MODEL in src/llm.ts.

Screenshots