feat(core): add QuotaManager with exponential backoff and persistence

[iceteaSA]

Unified quota cache and API gateway for main + fallback quota state. All consumers share one QuotaManager instance for consistent caching.

Features:

• Inflight deduplication prevents concurrent API calls
• Exponential backoff (60s–15min) for 429/5xx errors with escalating retry counts
• Persists main quota and backoff state to disk via callbacks
• Cross-process file lock guard for quota API dedup
• Seeds from persisted storage on construction (avoids cold-start API call)
• Integrates with FallbackAccountManager for shared staleness checks
• Captures storage path at init to prevent test config corruption
• Request-count-based refresh trigger (refreshEveryNRequests)

Files:

• packages/core/src/quota-manager.ts — new (391 lines)
• packages/core/src/accounts.ts — adds backoff helpers, FallbackAccountManager integration
• packages/opencode/src/index.ts — QuotaManager construction, persistence callbacks, routing
• packages/opencode/src/tests/quota-manager.test.ts — new (294 lines)

---

Summary by cubic

Adds a unified QuotaManager for main and fallback quotas with shared in-memory cache, a serial API gate, exponential backoff, and on-disk persistence. Integrates it into packages/opencode for consistent routing, shared staleness with fallbacks, and safer behavior when rate-limited.

• New Features

  • Single QuotaManager with inflight dedup, a serialized API gate, and a cross‑process file lock.
  • Exponential backoff for 429/5xx (60s up to 15m), with persisted backoff state and callbacks (onMainQuotaFetched, onApiError).
  • Seeds main quota and backoff from disk on startup; adds request-count refresh via quota.refreshEveryNRequests.
  • Integrates with FallbackAccountManager: shared staleness checks, seeding from persisted fallback quotas, and gated refreshAllFallbacks.
  • Optional fail-closed when the quota API is backed off and no cached quota (failClosedOnUnknownQuota returns 429 with retry-after).

• Migration

  • @cortexkit/anthropic-auth-core now exports QuotaManager.
  • If consuming FallbackAccountManager, pass a shared QuotaManager to unify cache and staleness across processes.
  • To persist main quota/backoff, wire QuotaManager callbacks to save quota.mainQuota, quota.mainQuotaCheckedAt, and quota.mainLastQuotaApiError (already implemented in packages/opencode).

^{Written for commit 731431a. Summary will update on new commits. Review in cubic}

[cubic-dev-ai]

2 issues found across 6 files

<sub>Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic</sub>