feat: 添加环境变量支持以覆盖 max_tokens 设置

src/services/api/openai/__tests__/queryModelOpenAI.test.ts

@@ -194,6 +194,16 @@ mock.module('../convertTools.js', () => ({
 mock.module('../../../../utils/context.js', () => ({
   getModelMaxOutputTokens: () => ({ upperLimit: 8192, default: 8192 }),
   getContextWindowForModel: () => 200_000,
+  modelSupports1M: () => false,
+  has1mContext: () => false,
+  is1mContextDisabled: () => false,
+  getSonnet1mExpTreatmentEnabled: () => false,
+  MODEL_CONTEXT_WINDOW_DEFAULT: 200_000,
+  COMPACT_MAX_OUTPUT_TOKENS: 20_000,
+  CAPPED_DEFAULT_MAX_TOKENS: 8_000,
+  ESCALATED_MAX_TOKENS: 64_000,
+  calculateContextPercentages: () => ({ used: null, remaining: null }),
+  getMaxThinkingTokensForModel: () => 8191,
 }))
 
 mock.module('../../../../utils/messages.js', () => ({
@@ -211,6 +221,22 @@ mock.module('../../../../utils/api.js', () => ({
   toolToAPISchema: async (t: any) => t,
 }))
 
+mock.module('../../../../Tool.js', () => ({
+  getEmptyToolPermissionContext: () => ({
+    alwaysAllow: [],
+    alwaysDeny: [],
+    needsPermission: [],
+    mode: 'default',
+    isBypassingPermissions: false,
+  }),
+  toolMatchesName: () => false,
+}))
+
+mock.module('../../../../utils/envUtils.js', () => ({
+  isEnvTruthy: (v: string | undefined) => v === '1' || v === 'true',
+  isEnvDefinedFalsy: (v: string | undefined) => v === '0' || v === 'false' || v === 'no' || v === 'off',
+}))
+
 mock.module('../../../../utils/toolSearch.js', () => ({
   isToolSearchEnabled: async () => false,
   extractDiscoveredToolNames: () => new Set(),
@@ -451,4 +477,83 @@ describe('queryModelOpenAI — max_tokens forwarded to request', () => {
     expect(_lastCreateArgs).not.toBeNull()
     expect(_lastCreateArgs!.max_tokens).toBe(8192)
   })
+
+  test('OPENAI_MAX_TOKENS env var overrides max_tokens', async () => {
+    const original = process.env.OPENAI_MAX_TOKENS
+    process.env.OPENAI_MAX_TOKENS = '4096'
+    try {
+      _nextEvents = [
+        makeMessageStart(),
+        makeContentBlockStart(0, 'text'),
+        makeTextDelta(0, 'hi'),
+        makeContentBlockStop(0),
+        makeMessageDelta('end_turn', 5),
+        makeMessageStop(),
+      ]
+
+      await runQueryModel(_nextEvents)
+
+      expect(_lastCreateArgs).not.toBeNull()
+      expect(_lastCreateArgs!.max_tokens).toBe(4096)
+    } finally {
+      if (original === undefined) {
+        delete process.env.OPENAI_MAX_TOKENS
+      } else {
+        process.env.OPENAI_MAX_TOKENS = original
+      }
+    }
+  })
+
+  test('CLAUDE_CODE_MAX_OUTPUT_TOKENS env var overrides max_tokens', async () => {
+    const original = process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS
+    process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS = '2048'
+    try {
+      _nextEvents = [
+        makeMessageStart(),
+        makeContentBlockStart(0, 'text'),
+        makeTextDelta(0, 'hi'),
+        makeContentBlockStop(0),
+        makeMessageDelta('end_turn', 5),
+        makeMessageStop(),
+      ]
+
+      await runQueryModel(_nextEvents)
+
+      expect(_lastCreateArgs).not.toBeNull()
+      expect(_lastCreateArgs!.max_tokens).toBe(2048)
+    } finally {
+      if (original === undefined) {
+        delete process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS
+      } else {
+        process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS = original
+      }
+    }
+  })
+
+  test('OPENAI_MAX_TOKENS takes priority over CLAUDE_CODE_MAX_OUTPUT_TOKENS', async () => {
+    const origOpenai = process.env.OPENAI_MAX_TOKENS
+    const origClaude = process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS
+    process.env.OPENAI_MAX_TOKENS = '4096'
+    process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS = '2048'
+    try {
+      _nextEvents = [
+        makeMessageStart(),
+        makeContentBlockStart(0, 'text'),
+        makeTextDelta(0, 'hi'),
+        makeContentBlockStop(0),
+        makeMessageDelta('end_turn', 5),
+        makeMessageStop(),
+      ]
+
+      await runQueryModel(_nextEvents)
+
+      expect(_lastCreateArgs).not.toBeNull()
+      expect(_lastCreateArgs!.max_tokens).toBe(4096)
+    } finally {
+      if (origOpenai === undefined) delete process.env.OPENAI_MAX_TOKENS
+      else process.env.OPENAI_MAX_TOKENS = origOpenai
+      if (origClaude === undefined) delete process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS
+      else process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS = origClaude
+    }
+  })
 })

src/services/api/openai/index.ts

@@ -71,6 +71,28 @@ export function isOpenAIThinkingEnabled(model: string): boolean {
   return modelLower.includes('deepseek-reasoner') || modelLower.includes('deepseek-v3.2')
 }
 
+/**
+ * Resolve max output tokens for the OpenAI-compatible path.
+ *
+ * Override priority:
+ * 1. maxOutputTokensOverride (programmatic, from query pipeline)
+ * 2. OPENAI_MAX_TOKENS env var (OpenAI-specific, useful for local models
+ *    with small context windows, e.g. RTX 3060 12GB running 65536-token models)
+ * 3. CLAUDE_CODE_MAX_OUTPUT_TOKENS env var (generic override)
+ * 4. upperLimit default (64000)
+ *
+ * @internal Exported for testing purposes only
+ */
+export function resolveOpenAIMaxTokens(
+  upperLimit: number,
+  maxOutputTokensOverride?: number,
+): number {
+  return maxOutputTokensOverride
+    ?? (process.env.OPENAI_MAX_TOKENS ? parseInt(process.env.OPENAI_MAX_TOKENS, 10) || undefined : undefined)
+    ?? (process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS ? parseInt(process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS, 10) || undefined : undefined)
+    ?? upperLimit
+}
+
 /**
  * Build the request body for OpenAI chat.completions.create().
  * Extracted for testability — the thinking mode params are injected here.
@@ -165,7 +187,7 @@ function assembleFinalAssistantOutputs(params: {
   if (stopReason === 'max_tokens') {
     outputs.push(createAssistantAPIErrorMessage({
       content: `Output truncated: response exceeded the ${maxTokens} token limit. ` +
-        `Set CLAUDE_CODE_MAX_OUTPUT_TOKENS to override.`,
+        `Set OPENAI_MAX_TOKENS or CLAUDE_CODE_MAX_OUTPUT_TOKENS to override.`,
       apiError: 'max_output_tokens',
       error: 'max_output_tokens',
     }))
@@ -286,8 +308,15 @@ export async function* queryModelOpenAI(
     //     auto-retry at 64k in query.ts. The OpenAI path has no such retry, so
     //     using the capped 8k default would silently truncate responses in
     //     multi-turn conversations where thinking consumes most of the budget.
+    //
+    //     Override priority:
+    //     1. options.maxOutputTokensOverride (programmatic)
+    //     2. OPENAI_MAX_TOKENS env var (OpenAI-specific, useful for local models
+    //        with small context windows, e.g. RTX 3060 12GB running 65536-token models)
+    //     3. CLAUDE_CODE_MAX_OUTPUT_TOKENS env var (generic override)
+    //     4. upperLimit default (64000)
     const { upperLimit } = getModelMaxOutputTokens(openaiModel)
-    const maxTokens = options.maxOutputTokensOverride ?? upperLimit
+    const maxTokens = resolveOpenAIMaxTokens(upperLimit, options.maxOutputTokensOverride)
 
     // 11. Get client
     const client = getOpenAIClient({