Add DeepSeek V4 flash/pro support and DeepSeek thinking compatibility (#877)

* Add DeepSeek V4 support and thinking compatibility

* Fix DeepSeek profile persistence regression

* Align multi-model handling with openai-multi-model

src/utils/context.test.ts

@@ -31,25 +31,36 @@ afterEach(() => {
   }
 })
 
-test('deepseek-chat uses provider-specific context and output caps', () => {
+test('deepseek-v4-flash uses provider-specific context and output caps', () => {
+  process.env.CLAUDE_CODE_USE_OPENAI = '1'
+  delete process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS
+  delete process.env.OPENAI_MODEL
+
+  expect(getContextWindowForModel('deepseek-v4-flash')).toBe(1_048_576)
+  expect(getModelMaxOutputTokens('deepseek-v4-flash')).toEqual({
+    default: 262_144,
+    upperLimit: 262_144,
+  })
+  expect(getMaxOutputTokensForModel('deepseek-v4-flash')).toBe(262_144)
+})
+
+test('deepseek legacy aliases keep their documented provider caps', () => {
   process.env.CLAUDE_CODE_USE_OPENAI = '1'
   delete process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS
   delete process.env.OPENAI_MODEL
 
   expect(getContextWindowForModel('deepseek-chat')).toBe(128_000)
-  expect(getModelMaxOutputTokens('deepseek-chat')).toEqual({
-    default: 8_192,
-    upperLimit: 8_192,
-  })
+  expect(getContextWindowForModel('deepseek-reasoner')).toBe(128_000)
   expect(getMaxOutputTokensForModel('deepseek-chat')).toBe(8_192)
+  expect(getMaxOutputTokensForModel('deepseek-reasoner')).toBe(65_536)
 })
 
-test('deepseek-chat clamps oversized max output overrides to the provider limit', () => {
+test('deepseek-v4-flash clamps oversized max output overrides to the provider limit', () => {
   process.env.CLAUDE_CODE_USE_OPENAI = '1'
-  process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS = '32000'
+  process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS = '500000'
   delete process.env.OPENAI_MODEL
 
-  expect(getMaxOutputTokensForModel('deepseek-chat')).toBe(8_192)
+  expect(getMaxOutputTokensForModel('deepseek-v4-flash')).toBe(262_144)
 })
 
 test('gpt-4o uses provider-specific context and output caps', () => {

src/utils/model/openaiContextWindows.ts

@@ -96,7 +96,13 @@ const OPENAI_CONTEXT_WINDOWS: Record<string, number> = {
   'o3-mini':                  200_000,
   'o4-mini':                  200_000,
 
-  // DeepSeek (V3: 128k context per official docs)
+  // DeepSeek V4 coding-agent models. DeepSeek's official coding-agent guide
+  // publishes V4 Pro at 1,048,576 context / 262,144 output; Flash is treated
+  // as the same family for local budgeting until a dedicated public model card
+  // lands.
+  'deepseek-v4-flash':      1_048_576,
+  'deepseek-v4-pro':        1_048_576,
+  // Legacy DeepSeek API aliases documented in the public pricing/model pages.
   'deepseek-chat':            128_000,
   'deepseek-reasoner':        128_000,
 
@@ -316,9 +322,12 @@ const OPENAI_MAX_OUTPUT_TOKENS: Record<string, number> = {
   'o3-mini':                  100_000,
   'o4-mini':                  100_000,
 
-  // DeepSeek
+  // DeepSeek V4 coding-agent models. See context-window note above.
+  'deepseek-v4-flash':        262_144,
+  'deepseek-v4-pro':          262_144,
+  // Legacy DeepSeek API aliases documented in the public pricing/model pages.
   'deepseek-chat':              8_192,
-  'deepseek-reasoner':         32_768,
+  'deepseek-reasoner':         65_536,
 
   // Groq
   'llama-3.3-70b-versatile':  32_768,

src/utils/providerModels.test.ts

@@ -16,6 +16,21 @@ describe('parseModelList', () => {
     ])
   })
 
+  test('splits semicolon-separated models', () => {
+    expect(parseModelList('glm-4.7; glm-4.7-flash')).toEqual([
+      'glm-4.7',
+      'glm-4.7-flash',
+    ])
+  })
+
+  test('splits mixed comma- and semicolon-separated models', () => {
+    expect(parseModelList('gpt-5.4; gpt-5.4-mini, o3')).toEqual([
+      'gpt-5.4',
+      'gpt-5.4-mini',
+      'o3',
+    ])
+  })
+
   test('returns single model in an array', () => {
     expect(parseModelList('llama3.1:8b')).toEqual(['llama3.1:8b'])
   })

src/utils/providerProfile.test.ts

@@ -814,6 +814,22 @@ test('openai profiles ignore poisoned shell model and base url values', () => {
   })
 })
 
+test('openai profiles normalize multi-model profile values to the primary model', () => {
+  const env = buildOpenAIProfileEnv({
+    goal: 'balanced',
+    apiKey: 'sk-live',
+    model: 'deepseek-v4-flash, deepseek-v4-pro, deepseek-chat',
+    baseUrl: 'https://api.deepseek.com/v1',
+    processEnv: {},
+  })
+
+  assert.deepEqual(env, {
+    OPENAI_BASE_URL: 'https://api.deepseek.com/v1',
+    OPENAI_MODEL: 'deepseek-v4-flash',
+    OPENAI_API_KEY: 'sk-live',
+  })
+})
+
 test('startup env ignores poisoned persisted openai model and base url', async () => {
   const env = await buildStartupEnvFromProfile({
     persisted: profile('openai', {

src/utils/providerProfile.ts

@@ -15,6 +15,7 @@ import {
 } from './providerRecommendation.js'
 import { readGeminiAccessToken } from './geminiCredentials.js'
 import { getOllamaChatBaseUrl } from './providerDiscovery.js'
+import { getPrimaryModel } from './providerModels.js'
 import { getProviderValidationError } from './providerValidation.js'
 import {
   maskSecretForDisplay,

src/utils/providerProfiles.test.ts

@@ -590,6 +590,20 @@ describe('getProviderPresetDefaults', () => {
     expect(defaults.baseUrl).toBe('http://127.0.0.1:1337/v1')
     expect(defaults.requiresApiKey).toBe(false)
   })
+
+  test('deepseek preset defaults to DeepSeek V4 flash and exposes flash/pro aliases', async () => {
+    const { getProviderPresetDefaults } = await importFreshProviderProfileModules()
+
+    const defaults = getProviderPresetDefaults('deepseek')
+
+    expect(defaults.provider).toBe('openai')
+    expect(defaults.name).toBe('DeepSeek')
+    expect(defaults.baseUrl).toBe('https://api.deepseek.com/v1')
+    expect(defaults.model).toBe(
+      'deepseek-v4-flash, deepseek-v4-pro, deepseek-chat, deepseek-reasoner',
+    )
+    expect(defaults.requiresApiKey).toBe(true)
+  })
 })
 
 describe('setActiveProviderProfile', () => {
@@ -659,6 +673,45 @@ describe('setActiveProviderProfile', () => {
     }
   })
 
+  test('persists primary model for keyed openai-compatible multi-model profiles', async () => {
+    const tempDir = mkdtempSync(join(tmpdir(), 'openclaude-provider-'))
+    process.chdir(tempDir)
+
+    try {
+      const { setActiveProviderProfile } =
+        await importFreshProviderProfileModules()
+      const deepSeekProfile = buildProfile({
+        id: 'deepseek_prof',
+        name: 'DeepSeek',
+        provider: 'openai',
+        baseUrl: 'https://api.deepseek.com/v1',
+        model: 'deepseek-v4-flash, deepseek-v4-pro, deepseek-chat',
+        apiKey: 'sk-deepseek-live',
+      })
+
+      saveMockGlobalConfig(current => ({
+        ...current,
+        providerProfiles: [deepSeekProfile],
+      }))
+
+      const result = setActiveProviderProfile('deepseek_prof')
+      const persisted = JSON.parse(
+        readFileSync(join(tempDir, '.openclaude-profile.json'), 'utf8'),
+      )
+
+      expect(result?.id).toBe('deepseek_prof')
+      expect(persisted.profile).toBe('openai')
+      expect(persisted.env).toEqual({
+        OPENAI_BASE_URL: 'https://api.deepseek.com/v1',
+        OPENAI_MODEL: 'deepseek-v4-flash',
+        OPENAI_API_KEY: 'sk-deepseek-live',
+      })
+    } finally {
+      process.chdir(originalCwd)
+      rmSync(tempDir, { recursive: true, force: true })
+    }
+  })
+
   test('sets ANTHROPIC_MODEL env var when switching to an anthropic-type provider', async () => {
     const { setActiveProviderProfile } =
       await importFreshProviderProfileModules()

src/utils/providerProfiles.ts

@@ -174,7 +174,7 @@ export function getProviderPresetDefaults(
         provider: 'openai',
         name: 'DeepSeek',
         baseUrl: 'https://api.deepseek.com/v1',
-        model: 'deepseek-chat',
+        model: 'deepseek-v4-flash, deepseek-v4-pro, deepseek-chat, deepseek-reasoner',
         apiKey: '',
         requiresApiKey: true,
       }
@@ -839,7 +839,7 @@ export function persistActiveProviderProfileModel(
 
 /**
  * Generate model options from a provider profile's model field.
- * Each comma-separated model becomes a separate option in the picker.
+ * Each parsed model becomes a separate option in the picker.
  */
 export function getProfileModelOptions(profile: ProviderProfile): ModelOption[] {
   const models = parseModelList(profile.model)

src/utils/thinking.ts

@@ -105,6 +105,12 @@ export function modelSupportsThinking(model: string): boolean {
   if (provider === 'foundry' || provider === 'firstParty') {
     return !canonical.includes('claude-3-')
   }
+  if (
+    canonical.startsWith('deepseek-v4-') ||
+    canonical === 'deepseek-reasoner'
+  ) {
+    return true
+  }
   // 3P (Bedrock/Vertex): only Opus 4+ and Sonnet 4+
   return canonical.includes('sonnet-4') || canonical.includes('opus-4')
 }