AutoGluon Conda Feedstock Upgrade Workflow

CRITICAL: DO NOT MERGE PULL REQUESTS during the PR-creation workflow (Steps 1–8). Only create PRs. The user will review them.

Exception: Step 9 (Automated Merge Chain) is an opt-in workflow that DOES merge PRs, but ONLY when the user explicitly asks you to watch and merge the chain. Never merge otherwise.

Step 1: Prerequisites Check

1.1 Check GitHub CLI

gh --version

If not installed, stop and tell the user to install from https://github.com/cli/cli#installation and run gh auth login.

1.2 Check Authentication

gh auth status

If not authenticated, ask user to run gh auth login.

1.3 Gather User Input

Ask for:

• New AutoGluon version number (e.g., 1.5.0)
• Working directory (default: ~/autogluon-feedstock-upgrade)

Step 2: Setup Working Directory and Fork/Clone Repos

mkdir -p {WORKING_DIR}
cd {WORKING_DIR}

gh repo fork conda-forge/autogluon.common-feedstock --clone=true --remote=true
gh repo fork conda-forge/autogluon.features-feedstock --clone=true --remote=true
gh repo fork conda-forge/autogluon.core-feedstock --clone=true --remote=true
gh repo fork conda-forge/autogluon.tabular-feedstock --clone=true --remote=true
gh repo fork conda-forge/autogluon.multimodal-feedstock --clone=true --remote=true
gh repo fork conda-forge/autogluon.timeseries-feedstock --clone=true --remote=true
gh repo fork conda-forge/autogluon-feedstock --clone=true --remote=true

Step 3: Compute SHA256 Hash

curl -sL "https://github.com/autogluon/autogluon/archive/refs/tags/v{NEW_VERSION}.tar.gz" -o /tmp/autogluon-{NEW_VERSION}.tar.gz
openssl sha256 /tmp/autogluon-{NEW_VERSION}.tar.gz | awk '{print $2}'
rm /tmp/autogluon-{NEW_VERSION}.tar.gz

If curl fails with 404, ask user to verify the version number.

Step 4: Fetch and Analyze Dependencies

4.1 Get Current (Old) Version

Read from {WORKING_DIR}/autogluon.common-feedstock/recipe/meta.yaml:

{% set version = "X.Y.Z" %}

4.2 Fetch Version Bounds

Fetch _setup_utils.py for both versions:

• New: https://raw.githubusercontent.com/autogluon/autogluon/refs/tags/v{NEW_VERSION}/core/src/autogluon/core/_setup_utils.py
• Old: https://raw.githubusercontent.com/autogluon/autogluon/refs/tags/v{OLD_VERSION}/core/src/autogluon/core/_setup_utils.py

Extract DEPENDENT_PACKAGES dictionary and PYTHON_REQUIRES string.

4.3 Fetch Package-Specific Setup Files

For each subpackage (common, features, core, tabular, multimodal, timeseries, autogluon), fetch: https://raw.githubusercontent.com/autogluon/autogluon/refs/tags/v{NEW_VERSION}/{SUBPACKAGE}/setup.py

The install_requires shows which DEPENDENT_PACKAGES each subpackage needs.

4.4 Create Dependency Change Summary

Compare old vs new. Summarize:

1. Changed version bounds
2. Added dependencies
3. Removed dependencies
4. Python version changes

Present summary to user and ask for confirmation before proceeding.

Step 5: Update Each Feedstock

Process in dependency order:

Order	Feedstock	Dependencies
1	autogluon.common-feedstock	(none)
2	autogluon.features-feedstock	common
3	autogluon.core-feedstock	common
4	autogluon.tabular-feedstock	core, features
5	autogluon.multimodal-feedstock	core
6	autogluon.timeseries-feedstock	core, tabular
7	autogluon-feedstock	all subpackages

For Each Feedstock:

5.1 Sync Fork and Create Branch (DO THIS FIRST)

cd {WORKING_DIR}/{FEEDSTOCK_NAME}
git fetch upstream
git checkout main
git reset --hard upstream/main
git checkout -b {NEW_VERSION}

5.2 Read Current meta.yaml

After creating branch, read recipe/meta.yaml to understand current structure.

5.3 Update meta.yaml

1. Update version: {% set version = "{NEW_VERSION}" %}
2. Update sha256: Use computed hash
3. Reset build number: number: 0
4. Update Python version (if changed): python >={{ python_min }},<{NEW_PYTHON_MAX}
5. Update dependency version bounds: Match DEPENDENT_PACKAGES

Rules:

• Keep autogluon.* dependencies as =={{ version }}
• Only include dependencies from that package's setup.py
• Preserve existing comments
• Use conda naming (see Package Name Mappings below)

5.4 Handle python_min Changes

If minimum Python changed, update .ci_support/linux_64_.yaml:

python_min:
- '{NEW_PYTHON_MIN}'

Step 6: Commit and Push

For each feedstock:

cd {WORKING_DIR}/{FEEDSTOCK_NAME}
git add recipe/meta.yaml
git commit -m "Update to v{NEW_VERSION}"
git push -u origin {NEW_VERSION}

Step 7: Create Pull Requests

For each feedstock:

cd {WORKING_DIR}/{FEEDSTOCK_NAME}
gh pr create \
  --repo conda-forge/{FEEDSTOCK_NAME} \
  --title "Update to v{NEW_VERSION}" \
  --body "$(cat <<'EOF'
## Summary
- Update {PACKAGE_NAME} to version {NEW_VERSION}
- Updated dependency version bounds from upstream

## Dependency Changes
{LIST_RELEVANT_CHANGES}

## Checklist
* [x] Used a personal fork of the feedstock to propose changes
* [x] Reset the build number to `0`
* [ ] Re-rendered (Use `@conda-forge-admin, please rerender` in a comment)
EOF
)"

Step 8: Final Summary

8.1 Provide PR Links

List all 7 created PRs with clickable links.

8.2 Merge Order Reminder

Merge PRs in dependency order:

1. autogluon.common (no dependencies)
2. autogluon.features and autogluon.core (parallel)
3. autogluon.tabular and autogluon.multimodal (parallel)
4. autogluon.timeseries
5. autogluon (meta-package)

8.3 Post-Merge Instructions

After each PR's CI passes:

1. Comment: @conda-forge-admin, please rerender
2. Wait for rerender bot to update
3. Once CI passes again, merge
4. Wait for package to be published before merging dependent PRs

Step 9: Automated Merge Chain (opt-in)

Only run this when the user explicitly asks you to watch CI and merge the PRs.

The 7 PRs must merge in dependency order because each feedstock's CI installs its autogluon.* dependencies from the conda-forge channel — so a dependent PR's CI cannot pass until every dependency it needs has been built and published to conda-forge (which happens only after that dependency's PR is merged, and takes ~30 min to a few hours after merge). "CI green" is therefore gated on "deps live", not just on merging the upstream PR.

9.1 Merge Order and Dependency Gates

Step	Feedstock(s) to merge	conda-forge deps that must be LIVE first
A	autogluon.common	(none)
B	autogluon.features	autogluon.common
C	autogluon.core	autogluon.common, autogluon.features
D	autogluon.tabular	autogluon.core, autogluon.features
D	autogluon.multimodal	autogluon.common, autogluon.core, autogluon.features
E	autogluon.timeseries	autogluon.common, autogluon.core, autogluon.features, autogluon.tabular
F	autogluon	autogluon.core, autogluon.features, autogluon.tabular, autogluon.multimodal, autogluon.timeseries

Feedstocks in the same step can be processed in parallel. Gates are derived from the autogluon.* =={{ version }} run-deps in each recipe — always re-verify them from the recipes (grep 'autogluon\.' recipe/meta.yaml); do not trust the Appendix A tree, which historically understated deps (e.g. core actually needs features).

9.2 Per-PR Loop

For the next unmerged PR whose dependency gate is satisfied:

1. Check deps are live on conda-forge (anaconda.org API):

   curl -s "https://api.anaconda.org/package/conda-forge/{DEP_PACKAGE}" \
     | python3 -c "import sys,json; print('{NEW_VERSION}' in json.load(sys.stdin).get('versions',[]))"
   

   {DEP_PACKAGE} is the conda name (e.g. autogluon.common). Repeat for every dep in the gate. If any dep is NOT live → sleep 10 min and retry (do nothing else).

   Important — repodata lag (observed up to ~45+ min for v1.6.2): the anaconda.org API lists a version as soon as the artifact is uploaded, but conda-build solves against the channel's regenerated repodata, which lags upload significantly. So a dep can show "live" in the API (and even in the /files endpoint on label main) yet still fail CI with autogluon.common=X.Y.Z does not exist.

   Authoritative resolvability check — the version must appear in the trimmed current_repodata.json (small, fast; do NOT parse the full repodata.json — it is hundreds of MB and truncates, giving false negatives):

   curl -s "https://conda.anaconda.org/conda-forge/noarch/current_repodata.json" \
     | python3 -c "import sys,json; d=json.load(sys.stdin); p={**d.get('packages',{}),**d.get('packages.conda',{})}; print('{VER}' in {v['version'] for v in p.values() if v.get('name')=='{PKG}'})"
   

   Only expect a dependent PR's CI to pass once its deps show True here. If CI fails with "does not exist" while this still says False, it's repodata lag — wait ~10–15 min and recheck; rerun CI only once it flips to True. Do not treat lag as a real failure.

2. Check CI status:

   gh pr checks {PR_NUM} --repo conda-forge/{FEEDSTOCK}
   

   • All required checks pass → go to step 4 (merge).
   • Any check in_progress / pending / queued → sleep 10 min and retry.
   • A check failed but deps just became live (CI ran before publish) → step 3 (restart CI).

3. Restart failed CI (deps are live now, so a rerun should pass):

   sha=$(gh pr view {PR_NUM} --repo conda-forge/{FEEDSTOCK} --json headRefOid --jq '.headRefOid')
   runid=$(gh run list --repo conda-forge/{FEEDSTOCK} --commit "$sha" \
     --json databaseId,name --jq '.[0].databaseId')
   gh run rerun "$runid" --repo conda-forge/{FEEDSTOCK} --failed
   

   Then sleep 10 min and retry from step 2. (Alternative if gh run rerun is unavailable: comment @conda-forge-admin, please restart ci.)

4. Merge — ONLY if CI is fully green:

   gh pr merge {PR_NUM} --repo conda-forge/{FEEDSTOCK} --merge
   

   NEVER merge on failed or pending CI. After merging, that package must publish (poll step 1 for it) before its dependents' gates open.

9.3 Polling Cadence

Use a 10-minute poll (ScheduleWakeup / /loop 10m). Each tick: advance every PR whose gate is satisfied, merge the green ones, restart the stale-red ones, and report what changed. Stop when all 7 are merged (or the user says stop).

Safety: merging is outward-facing. Do it only under the user's explicit direction (they triggered this step), strictly in dependency order, and never on red/pending CI.

---

Appendix A: Dependency Tree

Verified from the autogluon.* =={{ version }} run-deps in each recipe (2026-09, v1.6.2):

autogluon.common   (no AG deps)
autogluon.features (common)
autogluon.core     (common, features)
autogluon.tabular  (core, features)
autogluon.multimodal (common, core, features)
autogluon.timeseries (common, core, features, tabular)
autogluon [meta]   (core, features, tabular, multimodal, timeseries)

Merge order: common → features → core → {tabular, multimodal} → timeseries → autogluon.

Appendix B: URL Patterns

Resource	URL
Release tarball	https://github.com/autogluon/autogluon/archive/refs/tags/v{VERSION}.tar.gz
Version bounds file	https://raw.githubusercontent.com/autogluon/autogluon/refs/tags/v{VERSION}/core/src/autogluon/core/_setup_utils.py
Package setup.py	https://raw.githubusercontent.com/autogluon/autogluon/refs/tags/v{VERSION}/{SUBPACKAGE}/setup.py

Appendix C: Sample PRs

• autogluon.common PR #6
• autogluon.features PR #5
• autogluon.core PR #8
• autogluon.tabular PR #15
• autogluon.multimodal PR #16
• autogluon.timeseries PR #7
• autogluon PR #6

Appendix D: Package Name Mappings

PyPI Name	Conda-Forge Name
torch	pytorch
Pillow	pillow
scikit-learn	scikit-learn
PyYAML	pyyaml
opencv-python	opencv
tensorflow	tensorflow