[tylerdotai]

What

Publish the global edition of MiniMax Code (agent.minimax.io) next to the
CN edition (agent.minimax.cn).

Why

MiniMax ships two separate desktop apps, with different backends, accounts,
bundle ids and UI languages. The packages published here are the CN edition
only. From the published minimax-code_3.1.4_amd64.deb:

$ grep -A2 '"app_id"' build-info.json
  "app_id": "com.minimax.agent.cn",

$ cat resources/app/.env.local
NEXT_PUBLIC_DOMAIN_URL="https://agent.minimax.cn"
NEXT_PUBLIC_LOCALE="zh"
OSS_TYPE=matrix_oss_cn

So anyone on agent.minimax.io who installs the released package gets the
Chinese edition talking to the CN backend — and there is no global package to
install instead.

Upstream publishes both editions (their download config exposes
cnX64MacosDownloadUrl → filecdn.minimax.chat and
overseasX64MacosDownloadUrl → file.cdn.minimax.io), and
packaging/resolve-version.sh already knew how to choose between them — but
nothing ever set MMX_REGION, so the default (cn) was the only edition that
was ever built.

Changes

• scripts/lib_common.sh — MMX_REGION (cn | overseas) selects an
  edition profile: app id, package name, install prefix, desktop entry, icon
  name, deep-link scheme, electron-updater manifest and release-tag prefix.
  Each edition gets its own package identity, so both can be installed side by
  side instead of overwriting each other.
• install.sh — verify that the extracted payload really is the requested
  edition (read NEXT_PUBLIC_DOMAIN_URL from the payload's .env.local) and
  fail the build on a mismatch, instead of shipping a package that quietly
  talks to the wrong backend. build-info.json now records region/domain.
• packaging/build.sh — stage the icon tree under the package's own name,
  so the two editions never claim the same files.
• packaging/templates/desktop.tmpl — register only the deep-link scheme
  the edition actually uses (minimax-cn vs minimax); the payload derives it
  from the locale.
• packaging/resolve-version.sh — the electron-updater fallback pointed at
  …/minimax-agent/release, which 404s; it now uses the edition's -prod
  manifest (verified 200 for both editions).
• .github/workflows/build.yml — build both editions (matrix) and publish
  minimax-code-v<version> and minimax-code-global-v<version>. A manual run
  can build one edition or both, and takes a DMG URL per edition.
• tests/run.sh, tests/fixtures/fake_asar — 5 new cases: each edition
  builds with its own identity, a payload from the other edition is refused,
  the per-edition profile is checked, an unknown region is rejected.
• README.md — document both editions. Also corrects the claim that
  packages are built "for both architectures": only x86-64 is built today
  (resolve_arch rejects anything else) — see the arm64 issue.

Verification

• tests/run.sh → 31 passed, 0 failed (26 before, +5 new).
• Built the global edition from the official global DMG (3.1.4,
  file.cdn.minimax.io):
  • edition check reports payload edition OK: https://agent.minimax.io
  • build-info.json → "region": "overseas", "app_id": "com.minimax.agent"
  • packaged app package.json name → minimax-code-global
  • the app starts and reports readiness ("outcome":"ready") on Linux
• Packaged it end to end: minimax-code-global_3.1.4_amd64.deb and
  minimax-code-global-3.1.4.x86_64.rpm, with
  • /opt/minimax-code-global, /usr/bin/minimax-code-global,
    minimax-code-global.desktop (Icon=minimax-code-global,
    StartupWMClass=minimax-code-global, MimeType=x-scheme-handler/minimax;)
  • icons installed as minimax-code-global.png — no file clash with the CN
    package, so both can be installed side by side
• The CN path is unchanged: same package name, install prefix, release tag and
  asset names as before.

Note

This touches .github/workflows/build.yml, which #2 and #3 also touch — happy
to rebase on top of whichever lands first.

[tylerdotai]

Closing this in favour of a single consolidated issue (#6) — four findings in one place, so there is one thing to read and reply to rather than three overlapping PRs. The work is not lost: the branch is still up in the fork and I will open a PR for any part of it you want.