What the agent does
Starlette 1.0 removed two things apps written against 0.x still use. The @app.route decorator is gone, so importing the module raises AttributeError: 'Starlette' object has no attribute 'route'. Did you mean: 'router'?. And Jinja2Templates.TemplateResponse no longer accepts the template name first and the context second: that order had warned since 0.29 (The 'name' is not the first parameter anymore) and now fails as TypeError: unhashable type: 'dict'. The run starts when the upgrade lands, and the agent clones the repository into its own cloud sandbox.
In a Starlette app like natanlao/yaas, pyproject.toml requires Python 3.14 and starlette>=0.52.1, and uv.lock holds 0.52.1, so the lock is what keeps the app working. yaas.py builds app = Starlette(exception_handlers=...) and decorates index, fetch, fetch_json and robots_txt with @app.route, two paths for index. It calls templates.TemplateResponse five times in the old order, including the HTTP exception handler at line 35. tests.py has a SmokeTest with test_index against a TestClient, and the Makefile's test target runs unittest, then ruff format, ruff check and ty check. The Dockerfile installs with uv sync --locked.
The agent reproduces the break by moving the lock: uv lock --upgrade-package starlette takes it to 1.10.0, uv sync, and python -m unittest tests fails at import on line 54, the first @app.route. It replaces the decorators with a routes list of starlette.routing.Route entries, one per path, keeping both paths for index and the function names as route names. Importing now works and test_index fails at line 56 with the TypeError, which is the second removal.
Each TemplateResponse(name, {"request": request, ...}) becomes TemplateResponse(request, name, {...}) with request out of the context, five calls in all. test_index passes on 1.10.0. The agent runs ruff format, ruff check and ty check as the Makefile does, and leaves test_video and test_playlist to the reviewer's note, since they fetch from Vimeo and SoundCloud and need network access the sandbox may not have.
Start it with one API call
POST https://api.phoenix.so/o/{org}/sessions with this body:
{
"agent": "claude",
"input": "Upgrade starlette to the current 1.x release in uv.lock and fix what breaks: replace the removed @app.route decorators with Route entries that keep every path and name, move the Request to the first argument of every TemplateResponse call, run the unittest suite, ruff and ty as the Makefile does, and open a pull request that quotes the AttributeError and the TypeError and lists each call you changed.",
"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 updates uv.lock, rewrites the routing in yaas.py and the five template calls, and raises the lower bound in pyproject.toml so a fresh lock cannot pick 0.x again. Its description quotes both errors with their line numbers and shows test_index passing on the new version.
A reviewer checks that every path is still served, including /index.html beside /, that request.url_for("index") still resolves because the route keeps its name, and that the exception handler renders error.html with the request. The Docker image builds from the same lock with uv sync --locked, so the deployed app gets exactly what the tests ran against.
If a test needs the network, the pull request says which ones ran and which were skipped in the sandbox, so the reviewer knows what proof is present and what to run on merge.
Run it on every Monday morning
Create an automation in the Project with this prompt and a schedule trigger, for example cron set to 0 9 * * 1 in your timezone. Each week the run refreshes the lock, runs the suite, and opens a pull request only when something changed; a week with no new release ends with a note and no pull request.
The automation's max_runs_per_hour and the Project's branch rules apply as usual. A release that breaks something the agent cannot fix, such as a test that needs network access, ends in the pull request's description as an open item rather than a silent skip.
The use case this job follows, with its automation and trigger: Upgrade dependencies weekly.
Why run it on Rush?
A lockfile hides this break until someone regenerates it, and the two removals surface one after the other: the import error first, the template error only once the import works. A Rush run moves the lock in its own sandbox, hits both, fixes both, and keeps each traceback and the passing test in the session as the record.
Teams with several Starlette or FastAPI services face the same two edits in each. Runs go in parallel, one per repository, each with its own pull request and its own test output, and the pull request for each names the exact calls changed so review is a read, not a hunt.
Frequently asked questions
Does this affect FastAPI apps too?
Yes, the TemplateResponse change does: FastAPI's Jinja2Templates is Starlette's, and current FastAPI releases resolve Starlette 1.x, so the old argument order fails the same way. @app.route is a Starlette decorator; FastAPI's @app.get is unaffected.
Why Route entries instead of app.add_route?
A routes list passed to Starlette(...) is the documented form and keeps every path visible in one place. add_route works too; the agent uses the list and keeps the function names as route names so url_for calls keep resolving.
Can the lock stay on 0.52 and the code change later?
It can, and nothing forces the upgrade today. The pull request makes the code valid on both lines, since Route and the request-first TemplateResponse exist in 0.52.1 as well, so merging it first and moving the lock later is a safe order.
Will ruff or ty complain about the new code?
The agent runs both as the Makefile does and fixes what they report before opening the pull request. A ty finding about the new routes list, if any, is addressed in the same change rather than left for the reviewer.
More jobs like this: Upgrade dependencies weekly
- Fix "TypeError: stub(obj, 'meth', fn) has been removed" After Upgrading Sinon (Sinon)
- Fix $.ajax(...).success Is Not a Function After a jQuery Upgrade (jQuery)
- Fix "Textual SQL expression should be explicitly declared as text()" in SQLAlchemy 2 (SQLAlchemy)
- Fix "TypeError: this.getOptions is not a function" From sass-loader on webpack 4 (sass-loader)
- Fix THREE.Geometry Is Not a Constructor After a three.js Upgrade (three.js)
- Fix TypeScript Errors After an Upgrade (NestJS)