feat: rate limit utilization extraction, scheduling, and passthrough

[AmazingAng]

Summary

• Extract upstream anthropic-ratelimit-* headers from every Claude API response (success and failure) in proxyWithRetry()
• Store parsed RateLimitInfo per account in AccountManager, with console warnings at 80%/95%/throttled thresholds
• Passthrough all anthropic-ratelimit-* headers to downstream clients (prefix-match, not whitelist), plus x-auth2api-account for account identification
• Schedule utilization-aware account selection: sticky accounts release at >= 0.8 utilization, prefer lowest-utilization account, fall back to round-robin on cold start. Stale utilization data (> 30 min) decays to 0 to avoid permanently deprioritizing accounts after window resets
• Enrich /admin/accounts snapshots with rateLimit and anonymousId (SHA-256 based, stable across restarts)

Compatibility with auth2api-router

Fully supports header_filter.lua which reads:

• anthropic-ratelimit-unified-5h-utilization / 7d-utilization (routing decisions)
• anthropic-ratelimit-requests-remaining / tokens-remaining (capacity checks)
• anthropic-ratelimit-requests-limit / tokens-limit (admin display)
• anthropic-ratelimit-unified-reset (admin display)

/admin/accounts enrichment is backward-compatible (additive fields only).

Changed files

File	Change
src/upstream/ratelimit.ts	New — extractRateLimitHeaders(), buildDownstreamRateLimitHeaders(), getEffectiveUtilization()
src/accounts/manager.ts	RateLimitInfo type, recordRateLimit(), getAnonymousId(), utilization-aware getNextAccount(), enriched getSnapshots()
src/utils/http.ts	ProxyOptions.success signature extended, centralized extraction/injection in proxyWithRetry()
src/handlers/anthropic.ts	Inject rateLimitHeaders in success callback
src/handlers/openai.ts	Same

Test plan

• Verify rate limit headers appear in non-streaming responses
• Verify rate limit headers appear in streaming responses (before SSE body)
• Verify rate limit headers appear on final failure (429/5xx) responses
• Verify /admin/accounts includes rateLimit and anonymousId fields
• Verify account rotation prefers lower-utilization accounts
• Verify stale utilization (> 30 min) decays to 0 and account re-enters rotation

Made with Cursor