Route deepseek-v4-flash to Fireworks when DeepSeek API is unhealthy

[jahooma]

Summary

• Adds Fireworks as a transparent fallback for deepseek-v4-flash (accounts/fireworks/models/deepseek-v4-flash).
• New passive circuit breaker (deepseek-health.ts): 3 failures / 60s window opens the circuit for 5 min, then the next request probes DeepSeek and resets on success. No background polling — every user request is the probe, so all pods converge naturally.
• Tighter 60s headersTimeout for the Flash undici agent so dead-API requests fail fast instead of hanging on the existing 30-min default (kept for reasoning models on v4-pro).
• _post.ts routes to Fireworks when the circuit is open, plus inline pre-stream failover so the first user to hit an outage also gets a Fireworks response instead of an error.
• Pricing entry in FIREWORKS_PRICING_MAP: 0.14 / 0.03 / 0.28 per M tokens (input / cached / output).

How it works

1. createDeepSeekRequestTracked wraps the DeepSeek fetch. Network errors, timeouts, 5xx/408/429 → recordDeepSeekFailure(). 2xx → recordDeepSeekSuccess() (clears state).
2. When recentFailures.length >= 3 within the 60s window, openUntil = now + 5min.
3. Routing in _post.ts calls shouldBypassDeepSeek(model) and, if true, sets useDeepSeek = false so the existing Fireworks branch picks up the same model id (now in FIREWORKS_MODEL_MAP).
4. After cooldown expires, the next request retries DeepSeek directly. Success resets; another failure re-opens.

Test plan

• Unit tests for circuit breaker and outage classifier (11 new tests in deepseek-health.test.ts, all passing).
• Existing fireworks-deployment and fireworks-health test suites still pass (75 llm-api tests total).
• bun run typecheck clean for changed files (pre-existing SDK errors unrelated).
• Manual verification once deployed: confirm deepseek/deepseek-v4-flash calls succeed via DeepSeek when healthy and via Fireworks once the breaker opens (induce by temporarily pointing DEEPSEEK_BASE_URL at a sink, or watch real outage logs).

🤖 Generated with Claude Code