1. /api/heicode-auth/* proxy: CriticalRateLimit (20/20min) → GlobalAPIRateLimit
(180/180s). The Heicode external-identity proxy is hit on every page
render for /me + /refresh plus the login burst — CriticalRateLimit is
sized for sensitive ops (password reset, 2FA) and trips at ~5 quick
page loads, returning 429 to a normal user. APIM upstream rate-limits
itself, so a second tight layer here adds no security and just
manufactures 429s.
2. JIT-create user group: seed "default" instead of me.Data.ChannelID.
Companion to 578a68f which only patched the every-login overwrite
path. New users (yj2824269760@gmail.com et al, JIT-created after
578a68f) still landed in a UUID group → empty /v1/models response →
desktop client showed the static 3-Claude fallback list.
Both fixes were applied on the production VM directly today (sed +
python patch) — committing them so the next docker rebuild on VM keeps
them instead of reverting to the buggy file via git checkout.
DB hot-fix already applied: 6 affected users moved to group=default.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>