All jobs

Set Up GitHub Actions CI for a Dagster Project

Yes. An agent adds a GitHub Actions workflow that installs each Dagster code location with uv, loads its definitions, runs its pytest suite per PR, and opens a PR.

Run this on RushOpens the console with this prompt ready to send, in a Project that has your repository.

What the agent does

The agent reads the workspace layout first. In a repository like dagster-io/hooli-data-eng-pipelines, dg.toml declares a workspace of six projects, hooli-data-eng, hooli-data-ingest, hooli-bi, hooli-basics, hooli-snowflake-insights and hooli-batch-enrichment, with a seventh, hooli-airlift, commented out with a note that it currently fails on an import error. Each project has its own pyproject.toml built with hatchling, a dev extra, a uv.lock and a Dockerfile; hooli-data-eng, hooli-bi and hooli-basics declare requires-python >= 3.9, < 3.13 and list pytest in that extra, and hooli-data-eng/Dockerfile starts from the uv:python3.12-bookworm image and runs uv sync --frozen. dagster_cloud.yaml names each code location's module, such as hooli_data_eng.definitions, and the Makefile has dependencies, deps for dbt deps, manifest for dbt parse --project-dir=dbt_project --target BRANCH, and ruff.

Tests exist but nothing runs them. hooli-data-eng/hooli_data_eng_tests/test_assets.py materializes the orders and users partitioned assets with RawDataAPI(flaky=False) and a DuckDB IO manager writing test.duckdb; test_repo_loads.py asserts a job can be found; test_databricks_pipes.py calls materialize at import time and says in a comment that it needs DAGSTER_CLOUD_DEPLOYMENT_NAME, DATABRICKS_HOST and DATABRICKS_TOKEN to mean anything; hooli-bi and hooli-airlift have test_assets.py files of their own. The only workflow, deploy-dagster-cloud.yml, is a Dagster+ hybrid deploy: it runs ruff check --fix through astral-sh/ruff-action, detects which locations changed, builds a Docker image per changed location, pushes it to ECR and deploys a branch deployment, with a dagster-cloud ci check on the configuration. It never runs pytest, and a ruff check --fix step reformats rather than fails.

The agent writes .github/workflows/test.yml on pull_request and push to main with a matrix over the six active projects from dg.toml, so a failure names its location. Each job uses astral-sh/setup-uv with caching, Python 3.12 to match the Dockerfile, uv sync --frozen --extra dev in the project directory, then two checks: a definitions load, importing the module dagster_cloud.yaml names for that location and constructing its Definitions, and uv run pytest over the project's tests directory. For hooli-data-eng the dbt assets need a manifest before definitions load, so that job runs dbt deps and dbt parse first the way the Makefile does; the BRANCH and PROD targets in dbt_project/profiles.yml are Snowflake with credentials read from environment variables that default to empty strings, and the LOCAL target is DuckDB, so the agent checks in the sandbox whether dbt parse against BRANCH needs a connection and uses LOCAL if it does. test_databricks_pipes.py is deselected in CI with a pytest marker and a skip when DATABRICKS_HOST is unset, since it talks to a real workspace. A separate lint job runs ruff check and ruff format --check without --fix.

Before writing the YAML the agent runs every step itself: the sync for each project, the manifest build, the definitions load and pytest for the locations that have tests. Where test_repo_loads.py imports a name that no longer exists in the package, the agent fixes the test to load definitions the way dagster_cloud.yaml does rather than deleting it, and says so. Projects without a tests directory get the definitions load only, and the pull request names them so the team knows where tests are missing.

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 Dagster workspace. Read dg.toml, dagster_cloud.yaml, the Makefile, every project's pyproject.toml, uv.lock and Dockerfile, the dbt project's profiles.yml, every existing test file and the existing workflow first. Write .github/workflows/test.yml that runs on pull requests and pushes to the default branch with one matrix job per active code location: set up uv with caching and the Python version the Dockerfiles use, sync that project's locked dependencies with its dev extra, load the location's definitions module as dagster_cloud.yaml names it, and run pytest over its tests. For the location with dbt assets, build the dbt manifest first, using a target that needs no warehouse credentials. Skip any test that needs external credentials with a marker and a clear skip reason. Add a lint job that runs ruff check and ruff format --check without fixing. Run every step yourself before writing the workflow, fix tests that fail for reasons unrelated to credentials, and open a pull request that lists what each job runs and which locations have no tests.",
  "resources": [
    {
      "url": "https://github.com/your-org/your-app"
    }
  ]
}
  1. 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.
  2. 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.
  3. 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/test.yml, a conftest.py or marker change in hooli-data-eng/hooli_data_eng_tests/ that skips the Databricks test when its credentials are absent, a fix to test_repo_loads.py if its import was stale, and nothing else: deploy-dagster-cloud.yml, the Dockerfiles and dagster_cloud.yaml are unchanged, so deploys behave exactly as before.

The description has a table with one row per code location: the sync command, whether definitions loaded, the pytest summary line from the sandbox or the note that the project has no tests. The dbt manifest step and which profile target it used are called out. Every command and its output is in the transcript at GET /o/{org}/sessions/{id}/messages.

The workflow runs on the pull request itself, so a reviewer sees seven jobs on the checks tab before reading the YAML. They confirm the matrix matches dg.toml, read the skip reason on the Databricks test, and decide whether hooli-airlift should join the matrix once its import error is fixed, which a follow-up on the same session can attempt.

Run it on every terminal hand-off

A data engineer who has been running pytest by hand before each deploy starts this from the terminal with this body and goes back to the pipeline that needs attention. The run reads the real dg.toml, lock files and dagster_cloud.yaml, so the matrix, the Python version and the manifest step are taken from the repository rather than from a template.

Send a request_id, a UUID, with the body so a retried start returns the run it already began. The pull request is on the run's GET /o/{org}/sessions/{id} answer once it is open, and a key with a webhook_url hears run.completed within about 30 seconds of the turn ending.

The use case this job follows, with its automation and trigger: Hand off a long task.

Why run it on Rush?

A Dagster workspace with several code locations, a dbt project and a cloud deploy is exactly where a hand-written workflow fails on its first run: a manifest that is not there yet, a test that imports a Databricks client, a profile target that wants a warehouse. The agent runs the whole matrix in a sandbox first, with uv and the locked dependencies of each project, so the jobs it writes have each passed once before the pull request opens.

The run keeps every command and its output, so the pull request's table of what passed per location is backed by the transcript. Teams with several Dagster repositories can start one run each, in parallel, each in its own sandbox.

Frequently asked questions

Does CI need Snowflake or Databricks credentials?

No. The dbt manifest is built with a target that does not connect, the DuckDB-backed tests write a local test.duckdb, and the Databricks pipes test is skipped with a stated reason when DATABRICKS_HOST is unset. Adding the secrets later turns that skip into a real run.

Will this slow down or change the Dagster+ deploy?

No. deploy-dagster-cloud.yml is untouched and keeps building and deploying changed locations. The new workflow is separate and runs in parallel; if you want deploys to wait for green tests, that is a one-line needs change the pull request describes but does not make.

Why one job per code location?

Each project has its own uv.lock and Python dependencies, which is why each has its own Dockerfile. Installing them together would resolve a different set than production runs. A matrix also means a failure in hooli-bi is labelled as such rather than buried in one long log.

What about the commented-out hooli-airlift project?

It stays out of the matrix, because dg.toml says it fails to import today, and the pull request notes that. Once the import error is fixed, adding it is one line in the matrix.

More Dagster jobs

The same job on other stacks

Keep reading