circleci

Use when writing, editing, or reviewing CircleCI configuration for the Astronomer APC repository. Covers script organization, inline vs external scripts, and…

npx skills add https://github.com/astronomer/astronomer --skill circleci

CircleCI Configuration Guide

Critical Rules

  1. No long inline scripts — script logic for any language must not be written inline in .circleci/config.yml if the script has complicated flow control. Complicated scripts belong in bin/.
  2. Scripts live in bin/ — every script called from CircleCI must exist as a file in the bin/ directory with an appropriate extension (e.g. bin/my-script.sh, bin/my-script.py).
  3. Pin all versions — never use latest or unpinned tags for Docker images or installed tools. Always specify an exact version to prevent supply chain vulnerabilities and ensure reproducible builds.

Script Organization

Scripts invoked by CircleCI jobs must be committed to the repository under bin/ so they can be:

  • Linted and reviewed like any other source file
  • Tested and run locally without needing CI
  • Reused across multiple jobs or workflows
# ✅ CORRECT — call a script from bin/
steps:
  - run:
      name: Build Helm chart
      command: bin/build-helm-chart.sh
# ❌ WRONG — inline shell logic in the CircleCI config
steps:
  - run:
      name: Build Helm chart
      command: |
        helm package .
        mv astronomer-*.tgz /tmp/chart/

Config Generation Pipeline

.circleci/config.yml is never edited directly. It is a generated file produced by rendering the Jinja2 template .circleci/config.yml.j2 via bin/generate_circleci_config.py:

# Regenerate config.yml from the template
uv run bin/generate_circleci_config.py

The generator injects a small set of computed variables (e.g. ci_runner_version, kube_versions, machine_image_version, docker_images) into the template at render time. Always edit .circleci/config.yml.j2, then regenerate.


Version Pinning

Always pin exact versions for Docker images and any tools installed during a job. Using latest or loose tags introduces supply chain risk and makes builds non-reproducible.

All pinned versions must be declared as Jinja2 variables at the top of .circleci/config.yml.j2, not scattered inline throughout the file. This makes them easy to audit and update in one place. All version declarations must include a link to where the list of released versions can be found, so that updating them is straightforward and doesn't require searching online to find more recent releases.

{# ✅ CORRECT — versions declared at top of config.yml.j2 #}
{#- https://circleci.com/docs/guides/execution-managed/building-docker-images/#docker-version -#}
{%- set circleci_docker_version = 'docker23' -%}

{#- https://circleci.com/developer/machine/image/ubuntu-2404 -#}
{%- set machine_image_version = 'ubuntu-2404:2025.09.1' -%}
# Then referenced inline:
docker:
  - image: cimg/python:{{ python_image_version }}
# ❌ WRONG — version hardcoded inline, not declared at top
docker:
  - image: cimg/python:3.8.1
# ❌ WRONG — unpinned image
docker:
  - image: cimg/python:latest
# ✅ CORRECT — pinned tool version installed in a step
- run:
    name: Install helm
    command: bin/install-ci-tools.py 3.17.2

# ❌ WRONG — unversioned tool install
- run:
    name: Install helm
    command: curl https://raw.githubusercontent.com/helm/helm/main/scripts/get-helm-3 | bash

More skills from astronomer

airflow-state-store
astronomer
Persists task and asset state across retries and DAG runs using Airflow 3.3's AIP-103 key/value stores (`task_state_store`, `asset_state_store`) and the…
creating-openlineage-extractors
astronomer
Custom OpenLineage extractors for unsupported Airflow operators and complex lineage scenarios. Two approaches: add OpenLineage methods directly to operators you own (recommended), or create custom extractors for third-party operators you cannot modify Extractors intercept operator execution at three points: before execution for static lineage, after success for runtime-determined outputs, and optionally after failure for partial lineage Register extractors via airflow.cfg or environment...
debugging-dags
astronomer
Systematic root cause analysis and remediation for failed Airflow DAGs with structured investigation workflows. Guides through four-step diagnosis process: identify the failure, extract error details, gather contextual information, and deliver actionable remediation steps Categorizes failures into four types (data, code, infrastructure, dependency) to focus investigation and suggest appropriate fixes Provides ready-to-use CLI commands for log retrieval, run comparison, task clearing, and DAG...
delegating-to-otto
astronomer
Drives Astronomer's Otto agent (`astro otto`) as a delegated sub-agent for Airflow, dbt, and data-engineering work. Use when the user explicitly asks to "use…
deploying-airflow
astronomer
Deploy Airflow DAGs and projects. Use when the user wants to deploy code, push DAGs, set up CI/CD, deploy to production, or asks about deployment strategies…
deploying-go-sdk-bundles
astronomer
Builds, packs, and deploys compiled Airflow Go SDK bundles so the ExecutableCoordinator can run them. Use when the user wants to compile a Go task bundle, asks…
testing-dags
astronomer
Iterative test-debug-fix cycles for Airflow DAGs with comprehensive failure diagnosis. Start with af runs trigger-wait <dag_id> to run a DAG and wait for completion; no pre-flight checks needed On failure, use af runs diagnose for comprehensive failure summary and af tasks logs to inspect error details from specific tasks Supports custom configuration, timeouts, and retry attempts; handles success, failure, and timeout scenarios with clear response interpretation Quick validation available...
tracing-downstream-lineage
astronomer
Trace downstream data lineage to assess change impact before modifying tables or DAGs. Identifies direct consumers of a target table or DAG through source code search, view dependencies, and BI tool connections Builds a full dependency tree mapping all downstream impacts, from tables to dashboards to ML models Categorizes dependencies by criticality (critical, high, medium, low) to prioritize stakeholder communication and testing Generates an impact report with risk assessment, affected...