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

web/src/app/api/v1/chat/completions/_post.ts

@@ -67,6 +67,10 @@ import {
   handleDeepSeekStream,
   isDeepSeekModel,
 } from '@/llm-api/deepseek'
+import {
+  isLikelyDeepSeekOutage,
+  shouldBypassDeepSeek,
+} from '@/llm-api/deepseek-health'
 import {
   handleMoonshotNonStream,
   handleMoonshotStream,
@@ -118,6 +122,23 @@ import { withDefaultProperties } from '@codebuff/common/analytics'
 import { checkFreeModeRateLimit as defaultCheckFreeModeRateLimit } from './free-mode-rate-limiter'
 import { beginChatCompletionRequestMetrics } from './request-metrics'
 
+/**
+ * Decide whether a failed DeepSeek request should transparently fail over to
+ * Fireworks. Pre-stream errors (network/timeout/5xx) on a model that has a
+ * known Fireworks fallback are eligible. The circuit-breaker failure was
+ * already recorded inside the DeepSeek handler.
+ */
+function canFailoverDeepSeekToFireworks(
+  error: unknown,
+  model: string,
+): boolean {
+  if (!isFireworksModel(model)) return false
+  if (error instanceof DeepSeekError) {
+    return isLikelyDeepSeekOutage(undefined, error.statusCode)
+  }
+  return isLikelyDeepSeekOutage(error)
+}
+
 export const formatQuotaResetCountdown = (
   nextQuotaReset: string | null | undefined,
 ): string => {
@@ -814,10 +835,18 @@ export async function postChatCompletions(params: {
         const useMoonshot = !useOpenCodeZen && isMoonshotModel(typedBody.model)
         const useCanopyWave =
           !useMoonshot && !useOpenCodeZen && isCanopyWaveModel(typedBody.model)
+        const deepseekBypassed = shouldBypassDeepSeek(typedBody.model)
+        if (deepseekBypassed) {
+          providerLogger.info(
+            { model: typedBody.model },
+            'DeepSeek circuit open — routing to Fireworks fallback',
+          )
+        }
         const useDeepSeek =
           !useMoonshot &&
           !useOpenCodeZen &&
           !useCanopyWave &&
+          !deepseekBypassed &&
           isDeepSeekModel(typedBody.model)
         const useFireworks =
           !useMoonshot &&
@@ -841,6 +870,23 @@ export async function postChatCompletions(params: {
           logger: providerLogger,
           insertMessageBigquery,
         }
+        const callDeepSeekStream = async () => {
+          try {
+            return await handleDeepSeekStream(baseArgs)
+          } catch (error) {
+            if (canFailoverDeepSeekToFireworks(error, typedBody.model)) {
+              providerLogger.warn(
+                {
+                  model: typedBody.model,
+                  error: getErrorObject(error),
+                },
+                'DeepSeek failed pre-stream — falling back to Fireworks',
+              )
+              return await handleFireworksStream(baseArgs)
+            }
+            throw error
+          }
+        }
         const stream = useSiliconFlow
           ? await handleSiliconFlowStream(baseArgs)
           : useMoonshot
@@ -850,7 +896,7 @@ export async function postChatCompletions(params: {
               : useCanopyWave
                 ? await handleCanopyWaveStream(baseArgs)
                 : useDeepSeek
-                  ? await handleDeepSeekStream(baseArgs)
+                  ? await callDeepSeekStream()
                   : useFireworks
                     ? await handleFireworksStream(baseArgs)
                     : useOpenAIDirect
@@ -886,10 +932,18 @@ export async function postChatCompletions(params: {
         const useMoonshot = !useOpenCodeZen && isMoonshotModel(model)
         const useCanopyWave =
           !useMoonshot && !useOpenCodeZen && isCanopyWaveModel(model)
+        const deepseekBypassed = shouldBypassDeepSeek(model)
+        if (deepseekBypassed) {
+          providerLogger.info(
+            { model },
+            'DeepSeek circuit open — routing to Fireworks fallback',
+          )
+        }
         const useDeepSeek =
           !useMoonshot &&
           !useOpenCodeZen &&
           !useCanopyWave &&
+          !deepseekBypassed &&
           isDeepSeekModel(model)
         const useFireworks =
           !useMoonshot &&
@@ -914,6 +968,20 @@ export async function postChatCompletions(params: {
           logger: providerLogger,
           insertMessageBigquery,
         }
+        const callDeepSeekNonStream = async () => {
+          try {
+            return await handleDeepSeekNonStream(baseArgs)
+          } catch (error) {
+            if (canFailoverDeepSeekToFireworks(error, model)) {
+              providerLogger.warn(
+                { model, error: getErrorObject(error) },
+                'DeepSeek failed — falling back to Fireworks',
+              )
+              return await handleFireworksNonStream(baseArgs)
+            }
+            throw error
+          }
+        }
         const nonStreamRequest = useSiliconFlow
           ? handleSiliconFlowNonStream(baseArgs)
           : useMoonshot
@@ -923,7 +991,7 @@ export async function postChatCompletions(params: {
               : useCanopyWave
                 ? handleCanopyWaveNonStream(baseArgs)
                 : useDeepSeek
-                  ? handleDeepSeekNonStream(baseArgs)
+                  ? callDeepSeekNonStream()
                   : useFireworks
                     ? handleFireworksNonStream(baseArgs)
                     : shouldUseOpenAIEndpoint

web/src/llm-api/__tests__/deepseek-health.test.ts

@@ -0,0 +1,113 @@
+import { afterEach, beforeEach, describe, expect, it } from 'bun:test'
+
+import { deepseekModels } from '@codebuff/common/constants/model-config'
+
+import {
+  DEEPSEEK_CIRCUIT_CONFIG,
+  __resetDeepSeekCircuitForTests,
+  isDeepSeekCircuitOpen,
+  isLikelyDeepSeekOutage,
+  recordDeepSeekFailure,
+  recordDeepSeekSuccess,
+  shouldBypassDeepSeek,
+} from '../deepseek-health'
+
+describe('DeepSeek circuit breaker', () => {
+  beforeEach(() => {
+    __resetDeepSeekCircuitForTests()
+  })
+  afterEach(() => {
+    __resetDeepSeekCircuitForTests()
+  })
+
+  it('starts closed', () => {
+    expect(isDeepSeekCircuitOpen()).toBe(false)
+    expect(shouldBypassDeepSeek(deepseekModels.deepseekV4Flash)).toBe(false)
+  })
+
+  it('stays closed after fewer failures than threshold', () => {
+    for (let i = 0; i < DEEPSEEK_CIRCUIT_CONFIG.FAILURE_THRESHOLD - 1; i++) {
+      recordDeepSeekFailure()
+    }
+    expect(isDeepSeekCircuitOpen()).toBe(false)
+  })
+
+  it('opens after threshold failures in the window', () => {
+    for (let i = 0; i < DEEPSEEK_CIRCUIT_CONFIG.FAILURE_THRESHOLD; i++) {
+      recordDeepSeekFailure()
+    }
+    expect(isDeepSeekCircuitOpen()).toBe(true)
+  })
+
+  it('only bypasses v4-flash variants, not v4-pro', () => {
+    for (let i = 0; i < DEEPSEEK_CIRCUIT_CONFIG.FAILURE_THRESHOLD; i++) {
+      recordDeepSeekFailure()
+    }
+    expect(shouldBypassDeepSeek(deepseekModels.deepseekV4Flash)).toBe(true)
+    expect(shouldBypassDeepSeek(deepseekModels.deepseekV4FlashDirect)).toBe(
+      true,
+    )
+    expect(shouldBypassDeepSeek(deepseekModels.deepseekV4Pro)).toBe(false)
+    expect(shouldBypassDeepSeek(deepseekModels.deepseekV4ProDirect)).toBe(false)
+    expect(shouldBypassDeepSeek('anthropic/claude-sonnet-4.5')).toBe(false)
+  })
+
+  it('resets on success', () => {
+    for (let i = 0; i < DEEPSEEK_CIRCUIT_CONFIG.FAILURE_THRESHOLD; i++) {
+      recordDeepSeekFailure()
+    }
+    expect(isDeepSeekCircuitOpen()).toBe(true)
+    recordDeepSeekSuccess()
+    expect(isDeepSeekCircuitOpen()).toBe(false)
+    expect(shouldBypassDeepSeek(deepseekModels.deepseekV4Flash)).toBe(false)
+  })
+})
+
+describe('isLikelyDeepSeekOutage', () => {
+  it('treats 5xx, 408, 429 as outages', () => {
+    expect(isLikelyDeepSeekOutage(undefined, 500)).toBe(true)
+    expect(isLikelyDeepSeekOutage(undefined, 502)).toBe(true)
+    expect(isLikelyDeepSeekOutage(undefined, 503)).toBe(true)
+    expect(isLikelyDeepSeekOutage(undefined, 504)).toBe(true)
+    expect(isLikelyDeepSeekOutage(undefined, 408)).toBe(true)
+    expect(isLikelyDeepSeekOutage(undefined, 429)).toBe(true)
+  })
+
+  it('does not treat 4xx (other than 408/429) as outages', () => {
+    expect(isLikelyDeepSeekOutage(undefined, 400)).toBe(false)
+    expect(isLikelyDeepSeekOutage(undefined, 401)).toBe(false)
+    expect(isLikelyDeepSeekOutage(undefined, 403)).toBe(false)
+    expect(isLikelyDeepSeekOutage(undefined, 404)).toBe(false)
+  })
+
+  it('classifies undici header-timeout errors as outages', () => {
+    const error = Object.assign(new Error('Headers Timeout Error'), {
+      code: 'UND_ERR_HEADERS_TIMEOUT',
+    })
+    expect(isLikelyDeepSeekOutage(error)).toBe(true)
+  })
+
+  it('classifies common network errors as outages', () => {
+    for (const code of [
+      'ECONNRESET',
+      'ECONNREFUSED',
+      'ENOTFOUND',
+      'ETIMEDOUT',
+    ]) {
+      const error = Object.assign(new Error('boom'), { code })
+      expect(isLikelyDeepSeekOutage(error)).toBe(true)
+    }
+  })
+
+  it('classifies AbortError as outage', () => {
+    const error = new Error('aborted')
+    error.name = 'AbortError'
+    expect(isLikelyDeepSeekOutage(error)).toBe(true)
+  })
+
+  it('treats generic non-network errors as non-outage', () => {
+    expect(isLikelyDeepSeekOutage(new Error('bad json'))).toBe(false)
+    expect(isLikelyDeepSeekOutage(undefined)).toBe(false)
+    expect(isLikelyDeepSeekOutage('string')).toBe(false)
+  })
+})