What the agent does
The agent reads the Worker's one route before touching it. In a Worker like the llm-chat-app-template in cloudflare/templates, wrangler.jsonc sets a compatibility_date of 2025-10-08, the nodejs_compat and global_fetch_strictly_public flags, an ASSETS binding on ./public and an AI binding. src/index.ts sends anything outside /api/ to the assets binding, answers 405 for a non-POST on /api/chat and 404 elsewhere, and hands POSTs to handleChatRequest. That function does const { messages = [] } = (await request.json()) as { messages: ChatMessage[] }, a cast with no check, prepends SYSTEM_PROMPT when no system message is present, and streams env.AI.run(MODEL_ID, { messages, max_tokens: 1024, stream: true }). A body that is not JSON, or a messages that is a string, throws inside the try and comes back as 500 Failed to process request.
src/types.ts already says what a valid message is: role is "system" | "user" | "assistant" and content is a string. The agent turns that type into a runtime check in a new src/validate.ts, written by hand because the template has no runtime dependencies and the repository's check:deps and check:lockfiles scripts police package.json and the lockfiles across every template; adding zod would ripple into package-lock.json and the root pnpm-lock.yaml. The validator requires a JSON object with a non-empty messages array, a role from the set, a non-empty string content, and limits on the number of messages and total characters that the prompt sets, and returns either the typed value or a list of issues with a path such as messages[2].content.
One decision needs a product answer rather than a type: public/chat.js only ever sends user and assistant turns, yet the handler accepts a client-supplied system message and skips its own SYSTEM_PROMPT when it finds one, so any caller can replace the server's instructions. The prompt below tells the agent to reject system from the client with a 400, and the pull request explains the alternative. Malformed JSON becomes a 400 with { error: "invalid_json" }, a failed check becomes 400 with { error: "invalid_request", issues: [...] }, and the catch that returned 500 now covers only the AI call.
The template lists vitest 4.1.10 and @cloudflare/vitest-pool-workers 0.21.3 in devDependencies and has a test script, but no test file and no vitest.config.ts; its tsconfig.json even excludes a test directory that does not exist. The sibling to-do-list-kv-template shows the house pattern with its vitest.config.ts and test/tsconfig.json, so the agent follows it: a vitest.config.ts using defineWorkersConfig that points at wrangler.e2e.jsonc, the variant of the config without the AI binding, and test/chat.test.ts that imports SELF from cloudflare:test and POSTs a non-JSON body, an object with no messages, a message with a bad role, an empty content, a client system message and an oversized history, asserting 400 and the issue paths, plus a check that GET /api/chat is still 405. The agent runs npm test and npm run check, which is tsc --noEmit && wrangler deploy --dry-run, and pastes both.
Start it with one API call
POST https://api.phoenix.so/o/{org}/sessions with this body:
{
"agent": "codex",
"input": "Add request body validation to the /api/chat route of this Cloudflare Worker and open a pull request. Read wrangler.jsonc, wrangler.e2e.jsonc, package.json, tsconfig.json, src/index.ts, src/types.ts and public/chat.js first. Write src/validate.ts by hand with no new dependency: the body must be a JSON object with a non-empty messages array of at most 50 items, each with a role of user or assistant and a non-empty string content, and at most 32000 characters in total. Reject a client-supplied system message so the server's SYSTEM_PROMPT always applies, and explain that choice in the pull request. In src/index.ts answer 400 with { error: \"invalid_json\" } for a body that is not JSON and 400 with { error: \"invalid_request\", issues } naming each failing path for a body that fails the checks, and keep the 500 path for the AI call only. Add vitest.config.ts with defineWorkersConfig pointing at wrangler.e2e.jsonc and a test/tsconfig.json following the to-do-list-kv-template in the same repository, and write test/chat.test.ts using SELF from cloudflare:test covering each rejection and the existing 405 on GET. Run npm test and npm run check and paste both outputs. Do not change public/chat.js, the model id or the streaming response.",
"resources": [
{
"url": "https://github.com/your-org/your-app"
}
]
}- Get an API key. In the console, open Settings, then API keys, and create a key for your org. Send it in the Authorization header as a Bearer token.
- Send the request. POST the body in request.json to https://api.phoenix.so/o/{org}/sessions, with your repository as the resources URL. The run starts in its own cloud sandbox.
- Read what came back. Follow the run on its live stream or wait for the run.completed webhook, then read the result and its trajectory.
What comes back
What comes back is a pull request with a new src/validate.ts, a narrower ChatRequest type beside ChatMessage in src/types.ts, a handleChatRequest in src/index.ts that parses, validates and only then calls the model, and three new files for tests: vitest.config.ts, test/tsconfig.json and test/chat.test.ts. The description pastes the Vitest run with each rejection case listed, and the tsc --noEmit and wrangler deploy --dry-run output from npm run check.
The tests run against the Worker in workerd through @cloudflare/vitest-pool-workers, using wrangler.e2e.jsonc so no AI binding is required; validation happens before the model is reached, which is why every case can be exercised without an account. In the templates monorepo the pull request also passes pnpm run check, since nothing in package.json or a lockfile moved.
A reviewer checks that the limits match what the UI in public/chat.js can send and what the model's context allows, that rejecting client system messages is the behaviour the product wants, and that the error JSON is something the frontend could show. Then they merge, and a bad request is a 400 with a reason instead of a 500 and a log line.
Run it on every new Worker with a POST route
A team that stamps out Workers from templates sends one POST /o/{org}/sessions per repository with this prompt and that repository in resources. Each run reads its own handler and types, so a Worker whose body is a form or a webhook payload gets a validator for that shape, and one that already depends on a schema library gets the schema written in it. Each run opens its own pull request.
Give the API key a webhook_url and Rush POSTs run.completed when each turn ends, with the session_url to read the pull request from. The pasted test output is in the session too, so the reviewer sees which malformed bodies were tried before opening the diff.
The use case this job follows, with its automation and trigger: Hand off a long task.
Why run it on Rush?
Validation code is easy to write and easy to leave untested, and a Worker has no local server to poke at by hand without wrangler dev and a model binding. In a Rush sandbox the agent runs the Worker under workerd through the Vitest pool, sends each malformed body for real and records the status and JSON that came back, so the pull request shows the 400s rather than asserting them.
The judgment calls are documented in the same place: why no dependency in a monorepo that lints lockfiles, why a client system message is refused, what the limits are and where they came from. The same request runs unchanged on the next Worker, in its own sandbox, and a reviewer reads a diff with its evidence attached.
Frequently asked questions
Why not zod or valibot?
The template has no runtime dependencies, and the repository's check:deps and check:lockfiles scripts verify every template's package.json and lockfile. A hand-written validator for one object with two fields is a few dozen lines and leaves both untouched. In a standalone Worker that already uses a schema library the agent writes the schema in it instead.
Why reject a system message from the client?
Because handleChatRequest skips SYSTEM_PROMPT when the body already has a system turn, so any caller could replace the server's instructions. public/chat.js never sends one, so nothing legitimate breaks. If your product wants callers to set the system prompt, the prompt says so and the validator allows it.
How do the tests run without Workers AI?
vitest.config.ts points the Workers pool at wrangler.e2e.jsonc, the config the template already keeps without the AI binding. Every rejection happens before the model is called, so the cases run in workerd with no account, and GET /api/chat still answers 405.
Does the streaming response change?
No. A valid body goes through the same env.AI.run call with stream: true and the same text/event-stream headers. Only the path before it changes, and the 500 handler now covers the AI call alone.
Where do the limits come from?
From the prompt. Fifty messages and 32,000 characters are the values this request sets, chosen to sit well under the model's context while keeping a long chat working; the validator reads them from two constants at the top of src/validate.ts so a reviewer can change them in one place.
More Cloudflare Workers jobs
The same job on other stacks
- Add Input Validation to a Fiber API (Fiber)
- Add Input Validation to Firebase Cloud Functions (Firebase Cloud Functions)
- Add Input Validation to a Hono API (Hono)