[janeza2]

Summary

Clear stale transient rate-limit state on a provider connection after a successful quota refresh returns usable quota, so priority combos pick the primary connection again without manual Retest.

Problem

When a provider account hits a transient rate limit and recovers later, the provider-limits path fetches live quota and caches it, but does not clear the connection's transient cooldown/error DB fields (testStatus: "unavailable", rateLimitedUntil, lastError*, errorCode, backoffLevel). Priority combos still skip the primary because credential selection (getProviderCredentials) still sees the stale cooldown state, forcing the operator to hit Retest in the dashboard.

This affects every provider that goes through the provider-limits quota refresh path, not just GLM.

Fix

Add a generic recovery step in src/lib/usage/providerLimits.ts after a successful quota fetch:

• hasUsableQuota(usage) returns true when any quota window has remaining > 0, remainingPercentage > 0, or unlimited: true.
• maybeClearRecoveredQuotaState(connection, usage) calls clearRecoveredProviderState from src/sse/services/auth.ts only when:
  • quota is usable,
  • the connection is NOT in a terminal status (credits_exhausted, banned, expired, deactivated),
  • the connection currently has transient state worth clearing.
• Wired into both API-key and OAuth branches of fetchLiveProviderLimitsWithOptions.

Side fix: syncExpiredStatusIfNeeded now returns the updated connection so a recovered-quota path cannot overwrite a freshly written expired terminal status (race between auth-error message and usable quota).

Test plan

• New tests/unit/provider-limits-recovery.test.ts covers:
  • GLM transient cooldown state cleared on usable quota
  • terminal credits_exhausted status NOT cleared
  • terminal banned status NOT cleared
  • terminal expired status NOT cleared
  • error-only quota response (HTTP 429) does NOT clear transient state
• ReadLints clean on changed files
• git diff --check clean
• pre-push hooks passed (check:any-budget:t11, check:tracked-artifacts)
• CI test-unit (local env cannot resolve tsx package)

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[diegosouzapw]

Merged — thank you, @janeza2! 🙏 Solid resilience fix: after a successful quota refresh returns usable quota, maybeClearRecoveredQuotaState clears the stale transient rate-limit state (testStatus/rateLimitedUntil/error fields/backoffLevel) so priority combos re-pick the recovered primary connection without a manual Retest — while correctly leaving terminal statuses (credits_exhausted/banned/expired) untouched. I made one small test adjustment: the live GLM usage path throws on a 429 (it doesn't return an error envelope), so the 'error-only' case now asserts the rejection with assert.rejects — the transient-state guard still confirms an errored refresh never clears the cooldown (the throw precedes the clear). 5/5 tests green, lint + typecheck clean. Base redirected main→release/v3.8.44. Ships in the next release.