What the agent does
The agent reads the one test that exists before it writes more. In an app like trpc/examples-next-prisma-starter, package.json runs vitest as test-unit, Playwright as test-e2e, and prisma generate plus prisma migrate deploy as prebuild; dependencies are the canary tRPC packages, @prisma/client ^6.7.0, zod ^4.2.1, superjson and Next.js. vitest.config.mts sets globals: true, excludes playwright/, aliases ~/ to src/ and loads dotenv/config, and tsconfig.json adds vitest/globals to types. src/server/routers/post.test.ts holds a single test: it builds a context with createContextInner({}), gets a caller from createCaller, calls post.add then post.byId, and asserts toMatchObject on the input. .github/workflows/main.yml runs that in a unit job with a postgres service, NODE_ENV: test, pnpm prebuild and then pnpm test-unit and pnpm tsc.
What is untested is most of src/server/routers/post.ts. list takes limit between 1 and 100 defaulting to 50 and a nullable cursor, fetches limit + 1 rows ordered by createdAt descending with a cursor on id, pops the extra row to produce nextCursor, and returns the page reversed; byId throws a TRPCError with code NOT_FOUND and the message No post with id '…'; add validates title between 1 and 32 characters, non-empty text and an optional id that must be a UUID. The root router in _app.ts also has healthcheck, which returns the string yay!. None of the branches, the pagination arithmetic, the NOT_FOUND path or the zod rejections has a test, and the seed in prisma/seed.ts upserts one post with a fixed id that any ordering test has to account for.
The agent writes src/server/routers/post.list.test.ts and post.validation.test.ts next to the existing file, in the same style: createContextInner, createCaller, inferProcedureInput for typed inputs. For pagination it inserts posts with createdAt values a second apart through prisma.post.createMany, asks for limit: 2, asserts the two newest items come back in ascending order with a nextCursor equal to the id of the next-older row, follows the cursor and asserts the last page has nextCursor set to undefined, and checks that limit: 0 and limit: 101 reject with a BAD_REQUEST zod error. For byId it asserts the NOT_FOUND code and the message with the missing id. For add it asserts a 33-character title, an empty text and a non-UUID id each reject with BAD_REQUEST, and that a valid UUID id is stored as given. A beforeEach deletes from post with prisma.post.deleteMany() so the seeded row and earlier tests do not leak into ordering assertions, and a healthcheck test asserts yay!.
The tests need the database the existing test already assumes. The agent installs PostgreSQL in its sandbox with apt, starts it, creates the database DATABASE_URL in .env names, runs pnpm install, pnpm prebuild and pnpm test-unit, then pnpm tsc and pnpm lint, because eslint.config.mjs enables type-checked rules that a careless test file trips. It keeps the output and mentions in the pull request that the unit job in main.yml already provides this database, so nothing in CI changes.
Start it with one API call
POST https://api.phoenix.so/o/{org}/sessions with this body:
{
"agent": "claude",
"input": "Write Vitest tests for the tRPC routers in this repository. Read package.json, vitest.config.mts, tsconfig.json, .env, .github/workflows/main.yml, prisma/schema.prisma, prisma/seed.ts, src/server/trpc.ts, src/server/context.ts, src/server/routers/_app.ts, src/server/routers/post.ts and src/server/routers/post.test.ts first. Install and start PostgreSQL in the sandbox, create the database that DATABASE_URL names, and run pnpm install and pnpm prebuild. Add tests beside post.test.ts in the same style, using createContextInner, createCaller and inferProcedureInput, with a beforeEach that clears the post table: list with limit 2 over posts created a second apart, asserting ascending order within a page, nextCursor pointing at the next row, and no nextCursor on the last page; limit 0 and 101 rejected as BAD_REQUEST; byId with a missing id rejected as NOT_FOUND with its message; add rejecting a 33-character title, empty text and a non-UUID id, and storing a given UUID id; and healthcheck returning the literal string _app.ts declares. Run pnpm test-unit, pnpm tsc and pnpm lint, fix what they report without changing router code, and open a pull request that includes the test output and notes that the unit job in main.yml already has the Postgres service.",
"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 that adds src/server/routers/post.list.test.ts and src/server/routers/post.validation.test.ts, and nothing else: no router, schema or workflow changes. The existing post.test.ts stays as it is.
The description lists each case and the branch of post.ts it covers, and quotes the vitest summary, the tsc result and the eslint result from the sandbox. The transcript at GET /o/{org}/sessions/{id}/messages keeps the PostgreSQL setup and every command.
Before merging, a reviewer checks that deleteMany in beforeEach is acceptable against the shared test database the CI service provides, decides whether the Playwright smoke test should keep covering add through the UI, and may ask for a test of the createdAt transformer path through superjson as a follow-up in the same session.
Run it on every Sunday night
Create an automation in the Project with a schedule trigger, a cron such as 0 2 * * 0 with your timezone, and a prompt that asks for tests on the router procedures changed that week that have no test file beside them. Each run opens one pull request on its own branch, so Monday starts with cases to review.
For one repository now, send this body to POST /o/{org}/sessions with a request_id that is a UUID v5 of the repository's full name, so a retry returns the run it already started. The key's webhook_url reports run.completed with session_url when the turn finishes.
The use case this job follows, with its automation and trigger: Write the missing tests.
Why run it on Rush?
Testing a tRPC router honestly means a Postgres with the migrations applied, a Prisma client generated for it, and attention to a seed row that changes the ordering of every pagination assertion; on a laptop that is a database left in whatever state the last run abandoned. The agent builds that environment from scratch in its own sandbox, runs the suite there and keeps the output, so the pull request arrives with the cases proven rather than described.
Several services can be covered at once, each in its own sandbox, and the reviewer reads the vitest, tsc and eslint results in the pull request instead of reproducing them.
Frequently asked questions
Why go through createCaller instead of HTTP?
It is how the existing post.test.ts works: createCaller from _app.ts invokes procedures with a context built by createContextInner, so input validation, the router and Prisma all run while Next.js and the HTTP adapter stay out of the test. The Playwright smoke test in playwright/ covers the HTTP path.
Does this need Docker for Postgres?
No. The sandbox has no Docker, so the agent installs PostgreSQL with apt and starts it directly, then runs pnpm prebuild to apply prisma/migrations. In CI the postgres service in main.yml already does this job.
Will the tests fight with the seeded post?
The beforeEach clears the Post table, so ordering and cursor assertions see only the rows a test created. The seed's fixed UUID is left for the development database.
How are zod rejections asserted?
tRPC wraps an input failure in a TRPCError with code BAD_REQUEST, so each test uses rejects.toMatchObject({ code: 'BAD_REQUEST' }); the NOT_FOUND test also matches the message post.ts builds with the id.
More tRPC jobs
The same job on other stacks
- Write Vitest Tests for a Turborepo Monorepo (Turborepo)
- Write Vitest Tests for a Vite React App (Vite React)
- Write Vitest Tests for a Vue App (Vue)