:
Exponential Backoff β progressively longer cooldowns after each failure
Header Support β respects upstream hints- Configurable Base/Max β tune cooldown duration per use case
- Auto-Recovery β connection automatically becomes available after cooldown expires
Connection-scoped rate limits stay in Connection Cooldown (don't trip the breaker)
- Provider-wide transient errors (5xx, network timeouts) increment the failure counter
- Breaker trips only after fallback is exhausted AND the provider still fails
- Recovery β breaker automatically moves to half-open state after timeout, tests with probe request
.
Automatic β server waits for the earliest cooldown to expire
- Transparent β client sees a slightly delayed response instead of an error
- Configurable β enable/disable per combo or globally
Mutex Protection β only one retry attempt at a time per connection
- Semaphore β limits concurrent retry storms across connections
- Deduplication β identical requests within 5s window are deduplicated
:
Combo: "always-on"
1. cc/claude-opus-4-7 β Primary (subscription)
2. cx/gpt-5.2-codex β Secondary (subscription)
3. glm/glm-4.7 β Cheap backup ($0.5/1M)
4. if/kimi-k2-thinking β Free fallback (unlimited)
Browser-like TLS via β prevents bot detection
- CLI Fingerprint Matching β reorders headers and body fields to match native CLI binary signatures (Claude Code, Codex, etc.)
- Proxy IP Preservation β stealth features work on top of proxy IP masking
:
Uptime β server uptime and last restart
- Provider Breaker States β open/closed/half-open per provider
- Connection Cooldowns β active cooldowns with expiry times
- Cache Stats β signature + semantic cache hit rates
- Lockouts β API key lockouts and IP bans
- Latency β p50/p95/p99 percentiles
Architecture Guide β System architecture and internals
- User Guide β Providers, combos, CLI integration
- Auto-Combo Engine β 6-factor scoring, mode packs