Skip to content

PATCHES.md

Latest commit

 

History

History
81 lines (62 loc) · 3.23 KB

PATCHES.md

File metadata and controls

81 lines (62 loc) · 3.23 KB
Raw

GrokPool patches on upstream Sub2API

All patches live under sub2api/ and are baked into image grokpool-sub2api:local.

1. Free-tier Grok scheduling

Files:

  • backend/internal/service/openai_gateway_scheduling.go
  • backend/internal/service/openai_account_scheduler.go
  • backend/internal/service/openai_gateway_grok.go
  • backend/internal/service/openai_gateway_grok_cache.go
  • backend/internal/service/grok_token_provider.go

Intent:

  • Prefer free accounts with more remaining tokens (~1M windows)
  • Soft fit / headroom scoring for Grok OAuth
  • Shorter free-account 429 cooldowns
  • Avoid queuing when free slots still exist (TopK overflow / free-slot scan / sticky escape)

Details: docs/SCHEDULER.md

2. Dead-account probe

Files:

  • backend/internal/service/grok_quota_service.go — ClassifyGrokAccountLiveness
  • backend/internal/handler/admin/grok_oauth_handler.go — ProbeDeadAccounts
  • backend/internal/handler/admin/grok_import_probe.go — import-time hybrid probe
  • backend/internal/server/routes/admin.go — POST /admin/grok/accounts/probe-dead
  • Frontend bulk actions + i18n (probeDead)

Rules:

  • Dead = real upstream chat ban / permanent credential death (e.g. permission-denied on chat)
  • Billing-only 403 is not auto-dead
  • Do not SetAccountError for dead marks in a way that breaks OAuth refresh; use schedulable=false + notes

3. OAuth SSO → Build token (critical)

Files:

  • backend/internal/pkg/xai/sso_device.go — Authorization Code + PKCE
  • backend/internal/pkg/xai/oauth.go — referrer=grok-build (not sub2api)
  • backend/internal/service/grok_oauth_service.go — credentials base_url=cli-chat-proxy.grok.com/v1

Hard requirements:

  1. SSO cookie is only an entry ticket
  2. Final credential is OAuth access/refresh
  3. JWT must contain referrer=grok-build
  4. base_url must be https://cli-chat-proxy.grok.com/v1
  5. Device-flow tokens without the claim are rejected

4. Free-usage exhausted (daily 1M window)

Files:

  • backend/internal/service/openai_gateway_grok.go
    • detect body subscription:free-usage-exhausted / tokens actual/limit
    • pin remaining≈0, observation_source=free_usage_exhausted_body
    • rate_limit_reset_at ≈ +24h (do not clamp to free RPM 2m window)
  • backend/internal/service/openai_account_runtime_block_fastpath.go
    • OpenAIOAuth429FailoverState.freeUsageExhaustedSeen
    • free-usage keeps multi-account switching (not early stop after 1–2)
  • handlers (openai_gateway_handler, chat/images/media/alpha_search):
    • pass upstream body into ShouldStopOpenAIOAuth429Failover
    • failover exhausted free-usage/429 → clean client 429 (no raw body / no 503 wrap)
  • tests: openai_gateway_grok_free_usage_test.go

Intent: free daily exhaust must cool the account ~24h and rotate pool; clients should see a normal 429, not a gateway 503 storm.

5. Admin UX

  • Groups route allowed in simple mode
  • Accounts bulk bar: probe dead / delete dead
  • Quick nav links as needed
  • Account page size can load all accounts when needed

Rebuild

powershell -ExecutionPolicy Bypass -File scripts\build-sub2api.ps1
powershell -ExecutionPolicy Bypass -File scripts\check-parity.ps1

Stock image weishaw/sub2api:latest does not include these patches.