| name | conda-forge |
| description | Performs conda-forge operations. Fixes failing builds by analyzing CI logs, creates new packages via staged-recipes, adds native ARM CI support or explicit cross-compilation to feedstocks, and migrates recipes from v0 to v1 format. Use when working with conda-forge feedstocks, staged-recipes, build failures, recipe migrations, or when the user mentions conda-forge. |
| license | BSD-3-Clause |
| compatibility | Requires gh (GitHub CLI), git, curl, jq, tar, and pixi with rattler-build, conda-smithy, and cf-job-logs. Requires network access. |
Conda-Forge Skill
First, determine the workflow by checking the current working directory:
- staged-recipes repo → you are creating a new feedstock
- <name>-feedstock repo → you are updating an existing feedstock
Staged Recipes: Creating New Feedstocks
New packages are submitted via staged-recipes. After merge, conda-forge auto-creates a dedicated feedstock.
Fork and branch:
gh repo fork conda-forge/staged-recipes --clone=true
cd staged-recipes
git switch -c <PACKAGE_NAME>
Generate a recipe for Python packages with pixi exec grayskull pypi --use-v1-format --strict-conda-forge <PACKAGE_NAME>. Since grayskull creates the my-package/recipe.yaml in the CWD, run this command in the recipes/ directory of your staged-recipes fork.
For Go or Rust, adapt the templates from example-recipe-go.md or example-recipe-rust.md. See also example-recipe-python.md.
Place the recipe in recipes/<PACKAGE_NAME>/recipe.yaml.
Recipe rules:
- Use source tarballs with SHA256 checksums (compute with
curl -Ls "<URL>" | sha256sum -)
- Use SPDX license identifiers and include
license_file
- Tests are mandatory (at minimum: import test for Python,
--help for CLI)
- All dependencies must already exist on conda-forge
- Remove any template/placeholder comments
Test locally:
rattler-build build -r recipes/<PACKAGE_NAME> -m .ci_support/<VARIANT>.yaml
For noarch: python packages, provide a python_min variant override since staged-recipes doesn't define it:
rattler-build build -r recipes/<PACKAGE_NAME> -m .ci_support/<VARIANT>.yaml --variant python_min=3.10
Submit a draft PR, always use the PR template for the description. Watch CI until green. Don't mark as ready for review — let the human do that. To skip a platform, add skip: win in the build section.
For submitting multiple related packages, place each in a separate directory under recipes/. The build system resolves dependency order within staged-recipes.
Existing Feedstock: Updating
Work on recipe/recipe.yaml. If you only find a meta.yaml, you must convert it to recipe.yaml first — see rattler-build-migration.md.
Critical: Always Use a Personal Fork
conda-forge requires all PRs to come from forks, NOT from branches in the main feedstock repository.
If you push directly to conda-forge/<feedstock>, it will try to upload to conda-forge directly which is not allowed!
You MUST:
- Fork the feedstock to your personal GitHub account (or use an existing fork)
- Push your fix branch to YOUR fork
- Create a PR from <YOUR_USERNAME>: to conda-forge/:main
Use gh repo fork conda-forge/<feedstock> --clone to fork the feedstock.
This will clone the forked repository with the origin remote pointing to your fork and the upstream remote pointing to the original feedstock.
Important: If the fork already existed, its main branch may be out of date with upstream. Always sync it before creating your fix branch. First back up the fork's current main in case it has useful commits, then reset to upstream:
git fetch upstream
git checkout main
git branch main-fork-backup
git push origin main-fork-backup
git reset --hard upstream/main
git push origin main --force
Fixing Failing Builds
Start by reproducing the issue locally — run a local build (see Test Locally below) and iterate from there. If rattler-build fails, it keeps the work directory at output/bld/rattler-build_.../work — you can debug with cd <work> && source build_env.sh.
If the user explicitly references CI failures or pastes a link, use cf-job-logs to investigate:
- Wait for CI:
pixi exec cf-job-logs wait-for-ci --json <PR_URL>
- Polls until all checks complete, then reports which passed/failed
- Exits 0 if all pass, 1 if any fail — use this to gate further investigation
- List failed jobs:
pixi exec cf-job-logs list-jobs --json <PR_URL>
- Returns a JSON array of jobs with
id, result, platform, and name fields
- The output only contains failed jobs by default; use
--all to include successful jobs if needed
- Pipe to
jq for filtering if needed
- Download a specific job's log:
pixi exec cf-job-logs download-log <PR_URL> <JOB_ID>
- Use the
id from the list-jobs output as <JOB_ID>
- Logs are sanitized by default (timestamps and known boilerplate removed;
--no-sanitize is available but rarely needed)
- Redirect to a file with
> log.txt for large logs
- Read the error log — understand the root cause before making any changes
Recommended workflow: If CI is still running, start with wait-for-ci to block until completion. Then use list-jobs to get the job IDs for any failures, and download-log to fetch the actual logs.
Run pixi exec cf-job-logs --help or pixi exec cf-job-logs <command> --help for the full list of options.
Apply the minimal fix needed. Only modify files in the recipe/ directory.
Test Locally
Always test with a variant config that matches your local platform:
rattler-build build --recipe recipe -m .ci_support/<VARIANT>.yaml
Finalize Changes
After all recipe changes, always run:
pixi exec conda-smithy rerender --no-check-uptodate --commit=auto
pixi exec conda-smithy lint --conda-forge .
Commit your recipe changes first, then rerender (which auto-commits if there are changes). Only modify files in the recipe/ directory — everything else is autogenerated by rerender.
Push to your fork and create a PR to conda-forge/<FEEDSTOCK>:main.
General
If a tool isn't available locally, use pixi exec <tool> to run tools (rattler-build, conda-smithy, grayskull, feedrattler) without installing them. Use mktemp -d for temporary working directories.
Python Version Pinning (noarch recipes)
For noarch: python recipes, use conda-forge's pinning conventions:
- host:
python ${{ python_min }}.*
- run:
python >=${{ python_min }}
- tests:
python_version: ${{ python_min }}.*
If a newer Python minimum is required than conda-forge's default (3.10), override python_min in the context section of the recipe.
Common Errors
no candidates were found: Wrong dependency name. Use pixi search <name> (supports globs like pixi search 'lib*').- Import errors in test phase: Missing runtime dependency → add to
requirements.run.
- Compilation failure: Missing library → add to
requirements.host.
- Command not found (exit 127): Missing build tool → add to
requirements.build.
Key Decisions
noarch: python: Use for pure Python packages with no compiled extensions or platform-specific code.noarch: generic: Use for packages without any compiled code (e.g., shell-only recipes).- Build backend: Match
pyproject.toml's [build-system].requires.
- For ARM enablement and explicit cross-compilation details, see arm-builds-and-cross-compilation.md.
References
