AtlasCloudAI org-wide automation

Shared GitHub automation for AtlasCloudAI repositories.

Live model-catalog sync (update-models-readme)

Keeps the "Supported Models" sections of our repos in sync with the live model catalog (GET https://api.atlascloud.ai/api/v1/models) instead of hand-written, perpetually stale lists.

• Script: scripts/update-models-readme.mjs (zero-dependency Node ≥ 18)
• Reusable workflow: .github/workflows/update-models-readme.yml
• Currently onboarded: mcp-server, atlas-cloud-skills, cli

How it works

1. A thin caller workflow in each repo runs daily (cron) or on demand (workflow_dispatch).
2. It calls the reusable workflow here, which fetches the live catalog, filters display_console, groups models by platform category, and rewrites every ATLAS-MODELS marker block found in the repo's markdown files.
3. If (and only if) the catalog changed, it commits with the caller repo's own GITHUB_TOKEN. No secrets, no cross-repo PATs.

Onboarding a new repo (2 steps)

1. Wrap the section you want auto-updated in markers (attributes optional):

<!-- ATLAS-MODELS:START lang=en campaign=my-repo -->
(anything here is overwritten)
<!-- ATLAS-MODELS:END -->

Attributes: lang=en|zh-CN|ja|ko|es|fr · campaign=<utm_campaign> · groups=video,image,3d,llm,audio (subset/order) · featured=<n> (names per line). Multiple blocks per file and multiple files per repo are fine — every block is found automatically.

2. Add .github/workflows/sync-models.yml:

name: sync-models
on:
  schedule:
    - cron: "17 2 * * *"
  workflow_dispatch: {}
permissions:
  contents: write
jobs:
  sync:
    uses: AtlasCloudAI/.github/.github/workflows/update-models-readme.yml@main

Escape hatch

A markdown file that needs to QUOTE the marker syntax without being rewritten (like this README) can opt out entirely by containing the literal directive atlas-models-sync: skip-file anywhere in the file (put it in an HTML comment).

Design notes

• New platform categories are never dropped: categories the script doesn't recognize are auto-collected into a "More" group (opt-in via groups=...,more), so when the platform launches a new modality the READMEs pick it up without a script change.
• Featured names are data-driven: per category group, model families are ranked by the platform's own priority field — no hand-curated lists to rot.
• No date stamps inside blocks: reruns are byte-identical unless the catalog actually changed, so the cron stays silent on quiet days.
• Fail-closed: any API error aborts before touching files.