Add DeepSeek V4 provider (#574)

jahooma · web-flow · commit d11df242cc56 · 2026-05-01T16:23:30.000-07:00
diff --git a/agents/base2/base2-free-deepseek-v4.ts b/agents/base2/base2-free-deepseek-v4.ts
@@ -0,0 +1,11 @@
+import { createBase2 } from './base2'
+
+const definition = {
+  ...createBase2('free', {
+    noAskUser: true,
+    model: 'deepseek/deepseek-v4-pro',
+  }),
+  id: 'base2-free-deepseek-v4',
+  displayName: 'Buffy the DeepSeek V4 Free Orchestrator',
+}
+export default definition
diff --git a/agents/types/agent-definition.ts b/agents/types/agent-definition.ts
@@ -415,6 +415,8 @@ export type ModelName =
   | 'qwen/qwen3-30b-a3b:nitro'
 
   // DeepSeek
+  | 'deepseek/deepseek-v4-pro'
+  | 'deepseek-v4-pro'
   | 'deepseek/deepseek-chat-v3-0324'
   | 'deepseek/deepseek-chat-v3-0324:nitro'
   | 'deepseek/deepseek-r1-0528'
diff --git a/common/src/constants/free-agents.ts b/common/src/constants/free-agents.ts
@@ -2,6 +2,7 @@ import { parseAgentId } from '../util/agent-id-parsing'
 
 import { FREEBUFF_GEMINI_THINKER_AGENT_ID } from './freebuff-gemini-thinker'
 import {
+  FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID,
   FREEBUFF_GEMINI_PRO_MODEL_ID,
   SUPPORTED_FREEBUFF_MODELS,
 } from './freebuff-models'
@@ -20,7 +21,10 @@ export const FREE_COST_MODE = 'free' as const
  * excluded — they're spawned by the root, so counting them would inflate
  * every user's apparent activity.
  */
-export const FREEBUFF_ROOT_AGENT_IDS = ['base2-free'] as const
+export const FREEBUFF_ROOT_AGENT_IDS = [
+  'base2-free',
+  'base2-free-deepseek-v4',
+] as const
 const FREEBUFF_ROOT_AGENT_ID_SET: ReadonlySet<string> = new Set(
   FREEBUFF_ROOT_AGENT_IDS,
 )
@@ -39,6 +43,7 @@ const FREEBUFF_ALLOWED_MODEL_IDS = SUPPORTED_FREEBUFF_MODELS.map(
 export const FREE_MODE_AGENT_MODELS: Record<string, Set<string>> = {
   // Root orchestrator
   'base2-free': new Set(FREEBUFF_ALLOWED_MODEL_IDS),
+  'base2-free-deepseek-v4': new Set([FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID]),
 
   // File exploration agents
   'file-picker': new Set(['google/gemini-2.5-flash-lite']),
diff --git a/common/src/constants/freebuff-models.ts b/common/src/constants/freebuff-models.ts
@@ -22,6 +22,7 @@ export interface FreebuffModelOption {
  *  `getFreebuffDeploymentAvailabilityLabel()` instead. */
 export const FREEBUFF_DEPLOYMENT_HOURS_LABEL = '9am ET-5pm PT every day'
 export const FREEBUFF_GEMINI_PRO_MODEL_ID = 'google/gemini-3.1-pro-preview'
+export const FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID = 'deepseek/deepseek-v4-pro'
 export const FREEBUFF_GLM_MODEL_ID = 'z-ai/glm-5.1'
 export const FREEBUFF_KIMI_MODEL_ID = 'moonshotai/kimi-k2.6'
 export const FREEBUFF_MINIMAX_MODEL_ID = 'minimax/minimax-m2.7'
diff --git a/common/src/constants/model-config.ts b/common/src/constants/model-config.ts
@@ -6,6 +6,7 @@ export const ALLOWED_MODEL_PREFIXES = [
   'openai',
   'google',
   'x-ai',
+  'deepseek',
 ] as const
 
 export const costModes = [
@@ -55,6 +56,8 @@ export type openrouterModel =
 export const deepseekModels = {
   deepseekChat: 'deepseek-chat',
   deepseekReasoner: 'deepseek-reasoner',
+  deepseekV4ProDirect: 'deepseek-v4-pro',
+  deepseekV4Pro: 'deepseek/deepseek-v4-pro',
 } as const
 export type DeepseekModel = (typeof deepseekModels)[keyof typeof deepseekModels]
 
diff --git a/common/src/templates/initial-agents-dir/types/agent-definition.ts b/common/src/templates/initial-agents-dir/types/agent-definition.ts
@@ -415,6 +415,8 @@ export type ModelName =
   | 'qwen/qwen3-30b-a3b:nitro'
 
   // DeepSeek
+  | 'deepseek/deepseek-v4-pro'
+  | 'deepseek-v4-pro'
   | 'deepseek/deepseek-chat-v3-0324'
   | 'deepseek/deepseek-chat-v3-0324:nitro'
   | 'deepseek/deepseek-r1-0528'
diff --git a/evals/buffbench/main-single-eval.ts b/evals/buffbench/main-single-eval.ts
@@ -7,7 +7,7 @@ async function main() {
 
   await runBuffBench({
     evalDataPaths: [path.join(__dirname, 'eval-codebuff.json')],
-    agents: ['base2-free-evals'],
+    agents: ['base2-free-deepseek-v4'],
     taskIds: ['server-agent-validation'],
     saveTraces,
   })
diff --git a/packages/agent-runtime/src/constants.ts b/packages/agent-runtime/src/constants.ts
@@ -9,3 +9,7 @@ export const globalStopSequence = `${JSON.stringify(endsAgentStepParam)}`
  * to diff sequential requests and find what's breaking prompt caching.
  */
 export const CACHE_DEBUG_FULL_LOGGING = false
+
+// Keep disabled by default to preserve mainline behavior until reasoning-token
+// replay has been tested more thoroughly.
+export const INCLUDE_REASONING_IN_MESSAGE_HISTORY = false
diff --git a/packages/agent-runtime/src/tools/stream-parser.ts b/packages/agent-runtime/src/tools/stream-parser.ts
@@ -8,6 +8,7 @@ import {
 import { generateCompactId } from '@codebuff/common/util/string'
 
 import { processStreamWithTools } from '../tool-stream-parser'
+import { INCLUDE_REASONING_IN_MESSAGE_HISTORY } from '../constants'
 import {
   executeCustomToolCall,
   executeToolCall,
@@ -276,6 +277,11 @@ export async function processStream(
       }
 
       if (chunk.type === 'reasoning') {
+        if (INCLUDE_REASONING_IN_MESSAGE_HISTORY && chunk.text) {
+          assistantMessages.push(
+            assistantMessage({ type: 'reasoning', text: chunk.text }),
+          )
+        }
         onResponseChunk({
           type: 'reasoning_delta',
           text: chunk.text,
diff --git a/packages/internal/src/env-schema.ts b/packages/internal/src/env-schema.ts
@@ -8,6 +8,7 @@ export const serverEnvSchema = clientEnvSchema.extend({
   ANTHROPIC_API_KEY: z.string().min(1),
   FIREWORKS_API_KEY: z.string().min(1),
   CANOPYWAVE_API_KEY: z.string().min(1).optional(),
+  DEEPSEEK_API_KEY: z.string().min(1).optional(),
   SILICONFLOW_API_KEY: z.string().min(1).optional(),
   LINKUP_API_KEY: z.string().min(1),
   CONTEXT7_API_KEY: z.string().optional(),
@@ -87,6 +88,7 @@ export const serverProcessEnv: ServerInput = {
   ANTHROPIC_API_KEY: process.env.ANTHROPIC_API_KEY,
   FIREWORKS_API_KEY: process.env.FIREWORKS_API_KEY,
   CANOPYWAVE_API_KEY: process.env.CANOPYWAVE_API_KEY,
+  DEEPSEEK_API_KEY: process.env.DEEPSEEK_API_KEY,
   SILICONFLOW_API_KEY: process.env.SILICONFLOW_API_KEY,
   LINKUP_API_KEY: process.env.LINKUP_API_KEY,
   CONTEXT7_API_KEY: process.env.CONTEXT7_API_KEY,
diff --git a/packages/internal/src/env.ts b/packages/internal/src/env.ts
@@ -18,6 +18,7 @@ if (isCI) {
   ensureEnvDefault('ANTHROPIC_API_KEY', 'test')
   ensureEnvDefault('FIREWORKS_API_KEY', 'test')
   ensureEnvDefault('CANOPYWAVE_API_KEY', 'test')
+  ensureEnvDefault('DEEPSEEK_API_KEY', 'test')
   ensureEnvDefault('LINKUP_API_KEY', 'test')
   ensureEnvDefault('GRAVITY_API_KEY', 'test')
   ensureEnvDefault('IPINFO_TOKEN', 'test')
diff --git a/packages/internal/src/openai-compatible/chat/convert-to-openai-compatible-chat-messages.test.ts b/packages/internal/src/openai-compatible/chat/convert-to-openai-compatible-chat-messages.test.ts
@@ -509,6 +509,43 @@ describe('provider-specific metadata merging', () => {
     ])
   })
 
+  it('should preserve assistant reasoning content with tool calls', () => {
+    const result = convertToOpenAICompatibleChatMessages([
+      {
+        role: 'assistant',
+        content: [
+          { type: 'reasoning', text: 'Need the date first. ' },
+          { type: 'reasoning', text: 'Then call weather.' },
+          { type: 'text', text: 'Checking that now...' },
+          {
+            type: 'tool-call',
+            toolCallId: 'call1',
+            toolName: 'get_weather',
+            input: { location: 'Hangzhou' },
+          },
+        ],
+      },
+    ])
+
+    expect(result).toEqual([
+      {
+        role: 'assistant',
+        content: 'Checking that now...',
+        reasoning_content: 'Need the date first. Then call weather.',
+        tool_calls: [
+          {
+            id: 'call1',
+            type: 'function',
+            function: {
+              name: 'get_weather',
+              arguments: JSON.stringify({ location: 'Hangzhou' }),
+            },
+          },
+        ],
+      },
+    ])
+  })
+
   it('should handle a single tool role message with multiple tool-result parts', () => {
     const result = convertToOpenAICompatibleChatMessages([
       {
diff --git a/packages/internal/src/openai-compatible/chat/convert-to-openai-compatible-chat-messages.ts b/packages/internal/src/openai-compatible/chat/convert-to-openai-compatible-chat-messages.ts
@@ -65,6 +65,7 @@ export function convertToOpenAICompatibleChatMessages(
 
       case 'assistant': {
         let text = ''
+        let reasoningContent = ''
         const toolCalls: Array<{
           id: string
           type: 'function'
@@ -78,6 +79,10 @@ export function convertToOpenAICompatibleChatMessages(
               text += part.text
               break
             }
+            case 'reasoning': {
+              reasoningContent += part.text
+              break
+            }
             case 'tool-call': {
               toolCalls.push({
                 id: part.toolCallId,
@@ -96,6 +101,8 @@ export function convertToOpenAICompatibleChatMessages(
         messages.push({
           role: 'assistant',
           content: text,
+          reasoning_content:
+            reasoningContent.length > 0 ? reasoningContent : undefined,
           tool_calls: toolCalls.length > 0 ? toolCalls : undefined,
           ...metadata,
         })
diff --git a/web/src/app/api/v1/chat/completions/__tests__/completions.test.ts b/web/src/app/api/v1/chat/completions/__tests__/completions.test.ts
diff --git a/web/src/app/api/v1/chat/completions/_post.ts b/web/src/app/api/v1/chat/completions/_post.ts
diff --git a/web/src/llm-api/deepseek.ts b/web/src/llm-api/deepseek.ts
diff --git a/web/src/server/free-session/config.ts b/web/src/server/free-session/config.ts
