What the agent does
The agent reads what the app already runs before it writes a workflow, because the wrong SDK version or a test that needs a simulator makes CI red on the first run. In an app shaped like simplezhli/flutter_deer, pubspec.yaml asks for Dart >=3.4.0 <4.0.0 and Flutter >=3.19.0, turns on generate: true for the l10n.yaml localizations in lib/l10n, and lists flutter_test, flutter_driver, integration_test and test under dev dependencies. .github/workflows/flutter-drive.yml already exists but runs only flutter drive on a macOS runner with an iPhone simulator, pinned to Flutter 3.47.4, and flutter-web-deploy.yml triggers on a file name that does not exist, so no push ever runs flutter analyze or flutter test.
The tests that can run on Linux are in test/: widget_test.dart pumps MessagePage, net/dio_test.dart calls the GitHub API through DioUtils and so depends on the network, and accessibility_test.dart gathers eight per-module files that check tap-target semantics with meetsGuideline(labeledTapTargetGuideline) under MyApp and the ThemeProvider theme. analysis_options.yaml enables strict-casts and strict-inference. test_driver/ and integration_test/ need a device and stay with the existing macOS workflow.
The agent adds .github/workflows/ci.yml on push and pull_request with one ubuntu-latest job: actions/checkout, subosito/flutter-action with the same 3.47.4 the drive workflow pins and cache: true, flutter pub get, flutter gen-l10n so deer_localizations.dart exists before analysis, flutter analyze, and flutter test --exclude-tags network after tagging dio_test.dart with a network tag in a new dart_test.yaml, so a GitHub API hiccup cannot fail the build. A second job runs flutter build web to prove the web target still compiles, since the repository deploys it. A concurrency group cancels superseded runs on the same branch.
Before pushing, the agent installs that Flutter version in the sandbox and runs the same commands, so an analyzer finding that was never checked in CI, or a widget test that needs a font or asset, is fixed or reported in the pull request rather than discovered on the first red run. It adds a badge for the new workflow next to the existing text in README.md and README-EN.md.
Start it with one API call
POST https://api.phoenix.so/o/{org}/sessions with this body:
{
"agent": "claude",
"input": "Add GitHub Actions CI to this Flutter app and open a pull request. Read pubspec.yaml, l10n.yaml, analysis_options.yaml, .github/workflows/flutter-drive.yml, .github/workflows/flutter-web-deploy.yml, test/widget_test.dart, test/accessibility_test.dart, test/net/dio_test.dart and README-EN.md first. Add .github/workflows/ci.yml on push and pull_request with an ubuntu-latest job that checks out, sets up Flutter with subosito/flutter-action at the version the drive workflow pins with caching, runs flutter pub get, flutter gen-l10n, flutter analyze and flutter test excluding a network tag, and a second job that runs flutter build web. Add dart_test.yaml and tag test/net/dio_test.dart as network. Add a concurrency group that cancels superseded runs. Leave the drive and web deploy workflows as they are. Install the same Flutter version in the sandbox, run the same commands, fix any analyzer finding or test setup problem you can and list what you could not, paste the output in the pull request, and add a badge to README.md and README-EN.md. Do not change application code beyond what the analyzer requires.",
"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 .github/workflows/ci.yml, a dart_test.yaml with the network tag, a one-line @Tags annotation on test/net/dio_test.dart, and a badge line in both READMEs. The description shows the output of flutter analyze and flutter test --exclude-tags network from the sandbox on Flutter 3.47.4, with the count of tests found and any analyzer finding it fixed.
Because the workflow runs on pull_request, the pull request that adds it is its first run, so the reviewer sees the analyze, test and web build jobs on the pull request itself. The existing flutter_deer driver workflow still runs on macOS beside it.
A reviewer checks that the pinned Flutter version matches the one the team develops on, that the network-tagged test is the only one excluded, that flutter gen-l10n runs before flutter analyze so the generated localizations resolve, and that the web build job does not need secrets. Then they make the new checks required on the default branch.
Run it on every Flutter repository without an analyze-and-test workflow
A script on your server lists repositories through GET /o/{org}/github, checks each pubspec.yaml for a flutter dependency and each .github/workflows for a job that runs flutter test, and sends one POST /o/{org}/sessions per repository that has none, with that repository in resources. Each run reads its own pubspec.yaml and existing workflows, so an app that pins a different SDK or has no tests yet gets a workflow that fits it, and each 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 use case this job follows, with its automation and trigger: Hand off a long task.
Why run it on Rush?
Flutter CI fails for setup reasons more than code reasons: the SDK version, a generated file that is not checked in, a test that quietly calls the network. In a Rush sandbox the agent installs the exact SDK the repository pins, runs gen-l10n, analyze and test before it pushes, and keeps every command and its output in the run's record, so the workflow it proposes has already passed once.
The existing drive workflow needs a macOS runner and a simulator, so nobody wanted to add a second slow job by hand. The agent adds the fast Linux job that catches analyzer and widget failures in minutes, and the same request works unchanged on every Flutter app a team owns.
Frequently asked questions
Will it change the existing driver workflow?
No. flutter-drive.yml keeps running flutter drive on macOS with the iPhone simulator. The new ci.yml only adds the Linux analyze, test and build web jobs that need no device.
Why exclude the dio test?
test/net/dio_test.dart sets the base URL to api.github.com and makes a real request, so it fails when the runner is rate limited or offline. Tagging it network in dart_test.yaml keeps it runnable locally with flutter test and out of the required check.
Which Flutter version does it pin?
The one the repository already pins in flutter-drive.yml, 3.47.4, so both workflows test the same SDK. If pubspec.yaml later raises the constraint, the version is one line to change in each file.
Do the accessibility tests run in CI?
Yes. test/accessibility_test.dart and its per-module files are widget tests that pump pages under MyApp and check semantics, so they run on ubuntu-latest with flutter test. The agent runs them in the sandbox first and reports any page that fails a guideline.