✨ feat: bedrock claude model thinking support (#10422)

2026-03-27 13:29:15 +07:00 · 2025-11-26 00:42:36 +08:00
parent 5bab1a4bcf
commit 8b41638755
13 changed files with 309 additions and 211 deletions
--- a/packages/const/src/index.ts
+++ b/packages/const/src/index.ts
@@ -6,6 +6,7 @@ export * from './discover';
 export * from './layoutTokens';
 export * from './message';
 export * from './meta';
+export * from './models';
 export * from './plugin';
 export * from './session';
 export * from './settings';
--- a/packages/const/src/models.ts
+++ b/packages/const/src/models.ts
@@ -43,51 +43,63 @@ export const responsesAPIModels = new Set([
 ]);

 /**
- * models support context caching
+ * Regex patterns for models that support context caching (3.5+)
 */
-export const contextCachingModels = new Set([
-  'claude-opus-4-5-20251101',
-  'claude-haiku-4-5-20251001',
-  'claude-sonnet-4-5-latest',
-  'claude-sonnet-4-5-20250929',
-  'anthropic/claude-sonnet-4.5',
-  'claude-opus-4-latest',
-  'claude-opus-4-20250514',
-  'claude-sonnet-4-latest',
-  'claude-sonnet-4-20250514',
-  'claude-3-7-sonnet-latest',
-  'claude-3-7-sonnet-20250219',
-  'claude-3-5-sonnet-latest',
-  'claude-3-5-sonnet-20241022',
-  'claude-3-5-sonnet-20240620',
-  'claude-3-5-haiku-latest',
-  'claude-3-5-haiku-20241022',
-  // Bedrock model IDs
-  'us.anthropic.claude-sonnet-4-5-20250929-v1:0',
-  'anthropic.claude-sonnet-4-5-20250929-v1:0',
-  'us.anthropic.claude-haiku-4-5-20251001-v1:0',
-  'anthropic.claude-haiku-4-5-20251001-v1:0',
-  'global.anthropic.claude-opus-4-5-20251101-v1:0',
-  'anthropic.claude-opus-4-5-20251101-v1:0',
-]);
+export const contextCachingModelPatterns: RegExp[] = [
+  // Claude 4.5 series - Anthropic API
+  /^claude-(opus|sonnet|haiku)-4-5-/,
+  // Claude 4 series - Anthropic API
+  /^claude-(opus|sonnet)-4-/,
+  // Claude 3.7 - Anthropic API
+  /^claude-3-7-sonnet-/,
+  // Claude 3.5 series - Anthropic API
+  /^claude-3-5-(sonnet|haiku)-/,
+  // OpenRouter format (3.5+)
+  /^anthropic\/claude-(opus|sonnet|haiku)-(4\.5|4|3\.7|3\.5)/,
+  /^anthropic\/claude-(4\.5|4|3\.7|3\.5)-(opus|sonnet|haiku)/,
+  // AWS Bedrock format: [region.]anthropic.claude-xxx
+  /anthropic\.claude-(opus|sonnet|haiku)-(4-5|4|3-7|3-5)-/,
+];

-export const thinkingWithToolClaudeModels = new Set([
-  'claude-opus-4-5-20251101',
-  'claude-opus-4-latest',
-  'claude-opus-4-20250514',
-  'claude-sonnet-4-latest',
-  'claude-sonnet-4-20250514',
-  'claude-sonnet-4-5-latest',
-  'claude-sonnet-4-5-20250929',
-  'claude-haiku-4-5-20251001',
-  'anthropic/claude-sonnet-4.5',
-  'claude-3-7-sonnet-latest',
-  'claude-3-7-sonnet-20250219',
-  // Bedrock model IDs
-  'us.anthropic.claude-sonnet-4-5-20250929-v1:0',
-  'anthropic.claude-sonnet-4-5-20250929-v1:0',
-  'us.anthropic.claude-haiku-4-5-20251001-v1:0',
-  'anthropic.claude-haiku-4-5-20251001-v1:0',
-  'global.anthropic.claude-opus-4-5-20251101-v1:0',
-  'anthropic.claude-opus-4-5-20251101-v1:0',
-]);
+export const isContextCachingModel = (model: string): boolean => {
+  return contextCachingModelPatterns.some((pattern) => pattern.test(model));
+};
+
+/**
+ * Regex patterns for Claude models that support thinking with tools (3.7+)
+ */
+export const thinkingWithToolClaudeModelPatterns: RegExp[] = [
+  // Claude 4.5 series - Anthropic API
+  /^claude-(opus|sonnet|haiku)-4-5-/,
+  // Claude 4 series - Anthropic API
+  /^claude-(opus|sonnet)-4-/,
+  // Claude 3.7 - Anthropic API
+  /^claude-3-7-sonnet-/,
+  // OpenRouter format (3.7+)
+  /^anthropic\/claude-(opus|sonnet|haiku)-(4\.5|4|3\.7)/,
+  /^anthropic\/claude-(4\.5|4|3\.7)-(opus|sonnet|haiku)/,
+  // AWS Bedrock format: [region.]anthropic.claude-xxx
+  /anthropic\.claude-(opus|sonnet|haiku)-(4-5|4|3-7)-/,
+];
+
+export const isThinkingWithToolClaudeModel = (model: string): boolean => {
+  return thinkingWithToolClaudeModelPatterns.some((pattern) => pattern.test(model));
+};
+
+/**
+ * Regex patterns for Claude 4+ models that have temperature/top_p parameter conflict
+ * (cannot set both temperature and top_p at the same time)
+ */
+export const temperatureTopPConflictModelPatterns: RegExp[] = [
+  // Claude 4+ series - Anthropic API (4, 4.1, 4.5)
+  /^claude-(opus|sonnet|haiku)-4/,
+  // OpenRouter format
+  /^anthropic\/claude-(opus|sonnet|haiku)-(4\.5|4\.1|4)/,
+  /^anthropic\/claude-(4\.5|4\.1|4)-(opus|sonnet|haiku)/,
+  // AWS Bedrock format: [region.]anthropic.claude-xxx
+  /anthropic\.claude-(opus|sonnet|haiku)-4/,
+];
+
+export const hasTemperatureTopPConflict = (model: string): boolean => {
+  return temperatureTopPConflictModelPatterns.some((pattern) => pattern.test(model));
+};
--- a/packages/model-bank/src/aiModels/bedrock.ts
+++ b/packages/model-bank/src/aiModels/bedrock.ts
@@ -13,7 +13,7 @@ const bedrockChatModels: AIChatModelCard[] = [
      'Claude Opus 4.5 是 Anthropic 的旗舰模型，结合了卓越的智能与可扩展性能，适合需要最高质量回应和推理能力的复杂任务。',
    displayName: 'Claude Opus 4.5',
    enabled: true,
-    id: 'us.anthropic.claude-opus-4-5-20251101-v1:0',
+    id: 'global.anthropic.claude-opus-4-5-20251101-v1:0',
    maxOutput: 64_000,
    pricing: {
      units: [
@@ -23,6 +23,9 @@ const bedrockChatModels: AIChatModelCard[] = [
      ],
    },
    releasedAt: '2025-11-24',
+    settings: {
+      extendParams: ['disableContextCaching', 'enableReasoning', 'reasoningBudgetToken'],
+    },
    type: 'chat',
  },
  {
@@ -45,6 +48,9 @@ const bedrockChatModels: AIChatModelCard[] = [
      ],
    },
    releasedAt: '2025-09-29',
+    settings: {
+      extendParams: ['disableContextCaching', 'enableReasoning', 'reasoningBudgetToken'],
+    },
    type: 'chat',
  },
  {
@@ -68,6 +74,9 @@ const bedrockChatModels: AIChatModelCard[] = [
      ],
    },
    releasedAt: '2025-10-15',
+    settings: {
+      extendParams: ['disableContextCaching', 'enableReasoning', 'reasoningBudgetToken'],
+    },
    type: 'chat',
  },
  /*
@@ -103,7 +112,7 @@ const bedrockChatModels: AIChatModelCard[] = [
      'Claude 3.7 sonnet 是 Anthropic 最快的下一代模型。与 Claude 3 Haiku 相比，Claude 3.7 Sonnet 在各项技能上都有所提升，并在许多智力基准测试中超越了上一代最大的模型 Claude 3 Opus。',
    displayName: 'Claude 3.7 Sonnet',
    id: 'us.anthropic.claude-3-7-sonnet-20250219-v1:0',
-    maxOutput: 8192,
+    maxOutput: 64_000,
    pricing: {
      units: [
        { name: 'textInput', rate: 3, strategy: 'fixed', unit: 'millionTokens' },
@@ -111,6 +120,9 @@ const bedrockChatModels: AIChatModelCard[] = [
      ],
    },
    releasedAt: '2025-02-24',
+    settings: {
+      extendParams: ['disableContextCaching', 'enableReasoning', 'reasoningBudgetToken'],
+    },
    type: 'chat',
  },
  {
@@ -131,6 +143,9 @@ const bedrockChatModels: AIChatModelCard[] = [
      ],
    },
    releasedAt: '2024-10-22',
+    settings: {
+      extendParams: ['disableContextCaching'],
+    },
    type: 'chat',
  },
  {
@@ -151,6 +166,9 @@ const bedrockChatModels: AIChatModelCard[] = [
      ],
    },
    releasedAt: '2024-10-22',
+    settings: {
+      extendParams: ['disableContextCaching'],
+    },
    type: 'chat',
  },
  {
@@ -171,6 +189,9 @@ const bedrockChatModels: AIChatModelCard[] = [
      ],
    },
    releasedAt: '2024-06-20',
+    settings: {
+      extendParams: ['disableContextCaching'],
+    },
    type: 'chat',
  },
  {
@@ -191,6 +212,9 @@ const bedrockChatModels: AIChatModelCard[] = [
      ],
    },
    releasedAt: '2024-03-07',
+    settings: {
+      extendParams: ['disableContextCaching'],
+    },
    type: 'chat',
  },
  {
@@ -228,6 +252,9 @@ const bedrockChatModels: AIChatModelCard[] = [
      ],
    },
    releasedAt: '2024-02-29',
+    settings: {
+      extendParams: ['disableContextCaching'],
+    },
    type: 'chat',
  },
  {
@@ -264,6 +291,7 @@ const bedrockChatModels: AIChatModelCard[] = [
      '一款快速、经济且仍然非常有能力的模型，可以处理包括日常对话、文本分析、总结和文档问答在内的一系列任务。',
    displayName: 'Claude Instant',
    id: 'anthropic.claude-instant-v1',
+    maxOutput: 4096,
    pricing: {
      units: [
        { name: 'textInput', rate: 0.8, strategy: 'fixed', unit: 'millionTokens' },
--- a/packages/model-runtime/src/const/models.ts
+++ b/packages/model-runtime/src/const/models.ts
@@ -5,14 +5,14 @@ export const systemToUserModels = new Set([
  'o1-mini-2024-09-12',
 ]);

-// TODO: 临时写法，后续要重构成 model card 展示配置
+// TODO: temporary implementation, needs to be refactored into model card display configuration
 export const disableStreamModels = new Set([
  'o1',
  'o1-2024-12-17',
  'o1-pro',
  'o1-pro-2025-03-19',
  /*
-  官网显示不支持，但是实际试下来支持 Streaming，暂时注释掉
+  Official documentation shows no support, but actual testing shows Streaming is supported, temporarily commented out
  'o3-pro',
  'o3-pro-2025-06-10',
  */
@@ -38,30 +38,68 @@ export const responsesAPIModels = new Set([
  'gpt-5-codex',
  'gpt-5-pro',
  'gpt-5-pro-2025-10-06',
+  'gpt-5.1-codex',
+  'gpt-5.1-codex-mini',
 ]);

 /**
- * models support context caching
+ * Regex patterns for models that support context caching (3.5+)
 */
-export const contextCachingModels = new Set([
-  'claude-opus-4-latest',
-  'claude-opus-4-20250514',
-  'claude-sonnet-4-latest',
-  'claude-sonnet-4-20250514',
-  'claude-3-7-sonnet-latest',
-  'claude-3-7-sonnet-20250219',
-  'claude-3-5-sonnet-latest',
-  'claude-3-5-sonnet-20241022',
-  'claude-3-5-sonnet-20240620',
-  'claude-3-5-haiku-latest',
-  'claude-3-5-haiku-20241022',
-]);
+export const contextCachingModelPatterns: RegExp[] = [
+  // Claude 4.5 series - Anthropic API
+  /^claude-(opus|sonnet|haiku)-4-5-/,
+  // Claude 4 series - Anthropic API
+  /^claude-(opus|sonnet)-4-/,
+  // Claude 3.7 - Anthropic API
+  /^claude-3-7-sonnet-/,
+  // Claude 3.5 series - Anthropic API
+  /^claude-3-5-(sonnet|haiku)-/,
+  // OpenRouter format (3.5+)
+  /^anthropic\/claude-(opus|sonnet|haiku)-(4\.5|4|3\.7|3\.5)/,
+  /^anthropic\/claude-(4\.5|4|3\.7|3\.5)-(opus|sonnet|haiku)/,
+  // AWS Bedrock format: [region.]anthropic.claude-xxx
+  /anthropic\.claude-(opus|sonnet|haiku)-(4-5|4|3-7|3-5)-/,
+];

-export const thinkingWithToolClaudeModels = new Set([
-  'claude-opus-4-latest',
-  'claude-opus-4-20250514',
-  'claude-sonnet-4-latest',
-  'claude-sonnet-4-20250514',
-  'claude-3-7-sonnet-latest',
-  'claude-3-7-sonnet-20250219',
-]);
+export const isContextCachingModel = (model: string): boolean => {
+  return contextCachingModelPatterns.some((pattern) => pattern.test(model));
+};
+
+/**
+ * Regex patterns for Claude models that support thinking with tools (3.7+)
+ */
+export const thinkingWithToolClaudeModelPatterns: RegExp[] = [
+  // Claude 4.5 series - Anthropic API
+  /^claude-(opus|sonnet|haiku)-4-5-/,
+  // Claude 4 series - Anthropic API
+  /^claude-(opus|sonnet)-4-/,
+  // Claude 3.7 - Anthropic API
+  /^claude-3-7-sonnet-/,
+  // OpenRouter format (3.7+)
+  /^anthropic\/claude-(opus|sonnet|haiku)-(4\.5|4|3\.7)/,
+  /^anthropic\/claude-(4\.5|4|3\.7)-(opus|sonnet|haiku)/,
+  // AWS Bedrock format: [region.]anthropic.claude-xxx
+  /anthropic\.claude-(opus|sonnet|haiku)-(4-5|4|3-7)-/,
+];
+
+export const isThinkingWithToolClaudeModel = (model: string): boolean => {
+  return thinkingWithToolClaudeModelPatterns.some((pattern) => pattern.test(model));
+};
+
+/**
+ * Regex patterns for Claude 4+ models that have temperature/top_p parameter conflict
+ * (cannot set both temperature and top_p at the same time)
+ */
+export const temperatureTopPConflictModelPatterns: RegExp[] = [
+  // Claude 4+ series - Anthropic API (4, 4.1, 4.5)
+  /^claude-(opus|sonnet|haiku)-4/,
+  // OpenRouter format
+  /^anthropic\/claude-(opus|sonnet|haiku)-(4\.5|4\.1|4)/,
+  /^anthropic\/claude-(4\.5|4\.1|4)-(opus|sonnet|haiku)/,
+  // AWS Bedrock format: [region.]anthropic.claude-xxx
+  /anthropic\.claude-(opus|sonnet|haiku)-4/,
+];
+
+export const hasTemperatureTopPConflict = (model: string): boolean => {
+  return temperatureTopPConflictModelPatterns.some((pattern) => pattern.test(model));
+};
--- a/packages/model-runtime/src/core/contextBuilders/anthropic.ts
+++ b/packages/model-runtime/src/core/contextBuilders/anthropic.ts
@@ -223,3 +223,17 @@ export const buildAnthropicTools = (
    }),
  );
 };
+
+export const buildSearchTool = (): Anthropic.WebSearchTool20250305 => {
+  const maxUses = process.env.ANTHROPIC_MAX_USES;
+
+  return {
+    name: 'web_search',
+    type: 'web_search_20250305',
+    ...(maxUses &&
+      Number.isInteger(Number(maxUses)) &&
+      Number(maxUses) > 0 && {
+        max_uses: Number(maxUses),
+      }),
+  };
+};
--- a/packages/model-runtime/src/core/parameterResolver.test.ts
+++ b/packages/model-runtime/src/core/parameterResolver.test.ts
@@ -1,10 +1,7 @@
 import { describe, expect, it } from 'vitest';

-import {
-  MODEL_PARAMETER_CONFLICTS,
-  createParameterResolver,
-  resolveParameters,
-} from './parameterResolver';
+import { hasTemperatureTopPConflict } from '../const/models';
+import { createParameterResolver, resolveParameters } from './parameterResolver';

 describe('resolveParameters', () => {
  describe('Basic functionality', () => {
@@ -247,54 +244,41 @@ describe('createParameterResolver', () => {
  });
 });

-describe('MODEL_PARAMETER_CONFLICTS', () => {
-  describe('ANTHROPIC_CLAUDE_4_PLUS', () => {
-    it('should contain expected Claude 4+ models', () => {
-      expect(MODEL_PARAMETER_CONFLICTS.ANTHROPIC_CLAUDE_4_PLUS.has('claude-opus-4-1')).toBe(true);
-      expect(
-        MODEL_PARAMETER_CONFLICTS.ANTHROPIC_CLAUDE_4_PLUS.has('claude-opus-4-1-20250805'),
-      ).toBe(true);
-      expect(
-        MODEL_PARAMETER_CONFLICTS.ANTHROPIC_CLAUDE_4_PLUS.has('claude-sonnet-4-5-20250929'),
-      ).toBe(true);
+describe('hasTemperatureTopPConflict', () => {
+  describe('Anthropic Claude 4+ models', () => {
+    it('should return true for Claude 4+ models', () => {
+      expect(hasTemperatureTopPConflict('claude-opus-4-1-20250805')).toBe(true);
+      expect(hasTemperatureTopPConflict('claude-sonnet-4-5-20250929')).toBe(true);
+      expect(hasTemperatureTopPConflict('claude-haiku-4-5-20251001')).toBe(true);
    });

-    it('should not contain Claude 3.x models', () => {
-      expect(MODEL_PARAMETER_CONFLICTS.ANTHROPIC_CLAUDE_4_PLUS.has('claude-3-opus-20240229')).toBe(
-        false,
-      );
-      expect(
-        MODEL_PARAMETER_CONFLICTS.ANTHROPIC_CLAUDE_4_PLUS.has('claude-3.5-sonnet-20240620'),
-      ).toBe(false);
+    it('should return false for Claude 3.x models', () => {
+      expect(hasTemperatureTopPConflict('claude-3-opus-20240229')).toBe(false);
+      expect(hasTemperatureTopPConflict('claude-3-5-sonnet-20240620')).toBe(false);
    });
  });

-  describe('BEDROCK_CLAUDE_4_PLUS', () => {
-    it('should contain both standard and Bedrock-specific model IDs', () => {
-      expect(MODEL_PARAMETER_CONFLICTS.BEDROCK_CLAUDE_4_PLUS.has('claude-opus-4-1')).toBe(true);
-      expect(
-        MODEL_PARAMETER_CONFLICTS.BEDROCK_CLAUDE_4_PLUS.has(
-          'anthropic.claude-opus-4-1-20250805-v1:0',
-        ),
-      ).toBe(true);
-      expect(
-        MODEL_PARAMETER_CONFLICTS.BEDROCK_CLAUDE_4_PLUS.has(
-          'us.anthropic.claude-sonnet-4-5-20250929-v1:0',
-        ),
-      ).toBe(true);
+  describe('OpenRouter Claude 4+ models', () => {
+    it('should return true for OpenRouter Claude 4+ models', () => {
+      expect(hasTemperatureTopPConflict('anthropic/claude-opus-4.5')).toBe(true);
+      expect(hasTemperatureTopPConflict('anthropic/claude-sonnet-4.1')).toBe(true);
+      expect(hasTemperatureTopPConflict('anthropic/claude-4.5-opus')).toBe(true);
    });

-    it('should contain all Bedrock regional variants', () => {
-      expect(
-        MODEL_PARAMETER_CONFLICTS.BEDROCK_CLAUDE_4_PLUS.has(
-          'anthropic.claude-opus-4-20250514-v1:0',
-        ),
-      ).toBe(true);
-      expect(
-        MODEL_PARAMETER_CONFLICTS.BEDROCK_CLAUDE_4_PLUS.has(
-          'us.anthropic.claude-opus-4-20250514-v1:0',
-        ),
-      ).toBe(true);
+    it('should return false for OpenRouter Claude 3.x models', () => {
+      expect(hasTemperatureTopPConflict('anthropic/claude-3.5-sonnet')).toBe(false);
+      expect(hasTemperatureTopPConflict('anthropic/claude-3.7-sonnet')).toBe(false);
+    });
+  });
+
+  describe('Bedrock Claude 4+ models', () => {
+    it('should return true for Bedrock Claude 4+ models', () => {
+      expect(hasTemperatureTopPConflict('anthropic.claude-opus-4-1-20250805-v1:0')).toBe(true);
+      expect(hasTemperatureTopPConflict('us.anthropic.claude-sonnet-4-5-20250929-v1:0')).toBe(true);
+    });
+
+    it('should return false for Bedrock Claude 3.x models', () => {
+      expect(hasTemperatureTopPConflict('anthropic.claude-3-5-sonnet-20240620-v1:0')).toBe(false);
    });
  });
 });
--- a/packages/model-runtime/src/core/parameterResolver.ts
+++ b/packages/model-runtime/src/core/parameterResolver.ts
@@ -239,44 +239,3 @@ export const createParameterResolver = (options: ParameterResolverOptions) => {
    return resolveParameters(config, options);
  };
 };
-
-/**
- * Common model sets that have parameter conflicts
- */
-export const MODEL_PARAMETER_CONFLICTS = {
-  /**
-   * Claude models after Opus 4.1 that don't allow both temperature and top_p
-   */
-  ANTHROPIC_CLAUDE_4_PLUS: new Set([
-    'claude-opus-4-1',
-    'claude-opus-4-1-20250805',
-    'claude-sonnet-4-5-20250929',
-    'claude-haiku-4-5-20251001',
-    'claude-opus-4-5-20251101',
-  ]),
-
-  /**
-   * Bedrock Claude 4+ models (including Bedrock-specific model IDs)
-   */
-  BEDROCK_CLAUDE_4_PLUS: new Set([
-    'claude-opus-4-1',
-    'claude-opus-4-1-20250805',
-    'claude-opus-4-20250514',
-    'claude-sonnet-4-20250514',
-    'claude-sonnet-4-5-20250929',
-    'claude-haiku-4-5-20251001',
-    // Bedrock model IDs
-    'anthropic.claude-opus-4-1-20250805-v1:0',
-    'us.anthropic.claude-opus-4-1-20250805-v1:0',
-    'anthropic.claude-opus-4-20250514-v1:0',
-    'us.anthropic.claude-opus-4-20250514-v1:0',
-    'anthropic.claude-sonnet-4-20250514-v1:0',
-    'us.anthropic.claude-sonnet-4-20250514-v1:0',
-    'anthropic.claude-sonnet-4-5-20250929-v1:0',
-    'us.anthropic.claude-sonnet-4-5-20250929-v1:0',
-    'anthropic.claude-haiku-4-5-20251001-v1:0',
-    'us.anthropic.claude-haiku-4-5-20251001-v1:0',
-    'global.anthropic.claude-opus-4-5-20251101-v1:0',
-    'anthropic.claude-opus-4-5-20251101-v1:0',
-  ]),
-};
--- a/packages/model-runtime/src/index.ts
+++ b/packages/model-runtime/src/index.ts
@@ -1,3 +1,4 @@
+export * from './const/models';
 export * from './core/BaseAI';
 export { pruneReasoningPayload } from './core/contextBuilders/openai';
 export { ModelRuntime } from './core/ModelRuntime';
--- a/packages/model-runtime/src/providers/anthropic/index.ts
+++ b/packages/model-runtime/src/providers/anthropic/index.ts
@@ -1,9 +1,14 @@
 import Anthropic, { ClientOptions } from '@anthropic-ai/sdk';
 import { ModelProvider } from 'model-bank';

+import { hasTemperatureTopPConflict } from '../../const/models';
 import { LobeRuntimeAI } from '../../core/BaseAI';
-import { buildAnthropicMessages, buildAnthropicTools } from '../../core/contextBuilders/anthropic';
-import { MODEL_PARAMETER_CONFLICTS, resolveParameters } from '../../core/parameterResolver';
+import {
+  buildAnthropicMessages,
+  buildAnthropicTools,
+  buildSearchTool,
+} from '../../core/contextBuilders/anthropic';
+import { resolveParameters } from '../../core/parameterResolver';
 import { AnthropicStream } from '../../core/streams';
 import {
  type ChatCompletionErrorPayload,
@@ -22,6 +27,7 @@ import { StreamingResponse } from '../../utils/response';
 import { createAnthropicGenerateObject } from './generateObject';
 import { handleAnthropicError } from './handleAnthropicError';
 import { resolveCacheTTL } from './resolveCacheTTL';
+import { resolveMaxTokens } from './resolveMaxTokens';

 export interface AnthropicModelCard {
  created_at: string;
@@ -31,8 +37,6 @@ export interface AnthropicModelCard {

 type anthropicTools = Anthropic.Tool | Anthropic.WebSearchTool20250305;

-const modelsWithSmallContextWindow = new Set(['claude-3-opus-20240229', 'claude-3-haiku-20240307']);
-
 const DEFAULT_BASE_URL = 'https://api.anthropic.com';

 interface AnthropicAIParams extends ClientOptions {
@@ -140,15 +144,13 @@ export class LobeAnthropicAI implements LobeRuntimeAI {
    } = payload;

    const { anthropic: anthropicModels } = await import('model-bank');
-    const modelConfig = anthropicModels.find((m) => m.id === model);
-    const defaultMaxOutput = modelConfig?.maxOutput;

-    // 配置优先级：用户设置 > 模型配置 > 硬编码默认值
-    const getMaxTokens = () => {
-      if (max_tokens) return max_tokens;
-      if (defaultMaxOutput) return defaultMaxOutput;
-      return undefined;
-    };
+    const resolvedMaxTokens = await resolveMaxTokens({
+      max_tokens,
+      model,
+      providerModels: anthropicModels,
+      thinking,
+    });

    const system_message = messages.find((m) => m.role === 'system');
    const user_messages = messages.filter((m) => m.role !== 'system');
@@ -170,20 +172,8 @@ export class LobeAnthropicAI implements LobeRuntimeAI {
    });

    if (enabledSearch) {
-      // Limit the number of searches per request
-      const maxUses = process.env.ANTHROPIC_MAX_USES;
+      const webSearchTool = buildSearchTool();

-      const webSearchTool: Anthropic.WebSearchTool20250305 = {
-        name: 'web_search',
-        type: 'web_search_20250305',
-        ...(maxUses &&
-          Number.isInteger(Number(maxUses)) &&
-          Number(maxUses) > 0 && {
-            max_uses: Number(maxUses),
-          }),
-      };
-
-      // 如果已有工具，则添加到现有工具列表中；否则创建新的工具列表
      if (postTools && postTools.length > 0) {
        postTools = [...postTools, webSearchTool];
      } else {
@@ -192,19 +182,17 @@ export class LobeAnthropicAI implements LobeRuntimeAI {
    }

    if (!!thinking && thinking.type === 'enabled') {
-      const maxTokens = getMaxTokens() || 32_000; // Claude Opus 4 has minimum maxOutput
-
      // `temperature` may only be set to 1 when thinking is enabled.
      // `top_p` must be unset when thinking is enabled.
      return {
-        max_tokens: maxTokens,
+        max_tokens: resolvedMaxTokens,
        messages: postMessages,
        model,
        system: systemPrompts,
        thinking: {
          ...thinking,
          budget_tokens: thinking?.budget_tokens
-            ? Math.min(thinking.budget_tokens, maxTokens - 1) // `max_tokens` must be greater than `thinking.budget_tokens`.
+            ? Math.min(thinking.budget_tokens, resolvedMaxTokens - 1) // `max_tokens` must be greater than `thinking.budget_tokens`.
            : 1024,
        },
        tools: postTools,
@@ -212,7 +200,7 @@ export class LobeAnthropicAI implements LobeRuntimeAI {
    }

    // Resolve temperature and top_p parameters based on model constraints
-    const hasConflict = MODEL_PARAMETER_CONFLICTS.ANTHROPIC_CLAUDE_4_PLUS.has(model);
+    const hasConflict = hasTemperatureTopPConflict(model);
    const resolvedParams = resolveParameters(
      { temperature, top_p },
      { hasConflict, normalizeTemperature: true, preferTemperature: true },
@@ -221,7 +209,7 @@ export class LobeAnthropicAI implements LobeRuntimeAI {
    return {
      // claude 3 series model hax max output token of 4096, 3.x series has 8192
      // https://docs.anthropic.com/en/docs/about-claude/models/all-models#:~:text=200K-,Max%20output,-Normal%3A
-      max_tokens: getMaxTokens() || (modelsWithSmallContextWindow.has(model) ? 4096 : 8192),
+      max_tokens: resolvedMaxTokens,
      messages: postMessages,
      model,
      system: systemPrompts,
--- a/packages/model-runtime/src/providers/anthropic/resolveMaxTokens.ts
+++ b/packages/model-runtime/src/providers/anthropic/resolveMaxTokens.ts
@@ -0,0 +1,35 @@
+import type { ChatStreamPayload } from '../../types';
+
+const smallContextWindowPatterns = [
+  /claude-3-opus-20240229/,
+  /claude-3-haiku-20240307/,
+  /claude-v2(:1)?$/,
+];
+
+/**
+ * Resolve the max_tokens value to align Anthropic and Bedrock behavior.
+ * Priority: user input > model-bank default maxOutput > hardcoded fallback (context-window aware).
+ */
+export const resolveMaxTokens = async ({
+  max_tokens,
+  model,
+  thinking,
+  providerModels,
+}: {
+  max_tokens?: number;
+  model: string;
+  providerModels: { id: string; maxOutput?: number }[];
+  thinking?: ChatStreamPayload['thinking'];
+}) => {
+  const defaultMaxOutput = providerModels.find((m) => m.id === model)?.maxOutput;
+
+  const preferredMaxTokens = max_tokens ?? defaultMaxOutput;
+
+  if (preferredMaxTokens) return preferredMaxTokens;
+
+  if (thinking?.type === 'enabled') return 32_000;
+
+  const hasSmallContextWindow = smallContextWindowPatterns.some((pattern) => pattern.test(model));
+
+  return hasSmallContextWindow ? 4096 : 8192;
+};
--- a/packages/model-runtime/src/providers/bedrock/index.test.ts
+++ b/packages/model-runtime/src/providers/bedrock/index.test.ts
@@ -10,8 +10,6 @@ import { AgentRuntimeErrorType } from '../../types/error';
 import * as debugStreamModule from '../../utils/debugStream';
 import { LobeBedrockAI, experimental_buildLlama2Prompt } from './index';

-const provider = 'bedrock';
-
 // Mock the console.error to avoid polluting test output
 vi.spyOn(console, 'error').mockImplementation(() => {});

@@ -478,7 +476,7 @@ describe('LobeBedrockAI', () => {
            accept: 'application/json',
            body: JSON.stringify({
              anthropic_version: 'bedrock-2023-05-31',
-              max_tokens: 4096,
+              max_tokens: 8192,
              messages: [
                {
                  content: [
@@ -521,7 +519,7 @@ describe('LobeBedrockAI', () => {
            accept: 'application/json',
            body: JSON.stringify({
              anthropic_version: 'bedrock-2023-05-31',
-              max_tokens: 4096,
+              max_tokens: 8192,
              messages: [
                {
                  content: [
@@ -565,7 +563,7 @@ describe('LobeBedrockAI', () => {
            accept: 'application/json',
            body: JSON.stringify({
              anthropic_version: 'bedrock-2023-05-31',
-              max_tokens: 4096,
+              max_tokens: 8192,
              messages: [
                {
                  content: [
@@ -610,7 +608,7 @@ describe('LobeBedrockAI', () => {
            accept: 'application/json',
            body: JSON.stringify({
              anthropic_version: 'bedrock-2023-05-31',
-              max_tokens: 4096,
+              max_tokens: 64_000,
              messages: [
                {
                  content: [
@@ -654,7 +652,7 @@ describe('LobeBedrockAI', () => {
            accept: 'application/json',
            body: JSON.stringify({
              anthropic_version: 'bedrock-2023-05-31',
-              max_tokens: 4096,
+              max_tokens: 8192,
              messages: [
                {
                  content: [
--- a/packages/model-runtime/src/providers/bedrock/index.ts
+++ b/packages/model-runtime/src/providers/bedrock/index.ts
@@ -1,3 +1,4 @@
+import type Anthropic from '@anthropic-ai/sdk';
 import {
  BedrockRuntimeClient,
  InvokeModelCommand,
@@ -5,9 +6,10 @@ import {
 } from '@aws-sdk/client-bedrock-runtime';
 import { ModelProvider } from 'model-bank';

+import { hasTemperatureTopPConflict } from '../../const/models';
 import { LobeRuntimeAI } from '../../core/BaseAI';
 import { buildAnthropicMessages, buildAnthropicTools } from '../../core/contextBuilders/anthropic';
-import { MODEL_PARAMETER_CONFLICTS, resolveParameters } from '../../core/parameterResolver';
+import { resolveParameters } from '../../core/parameterResolver';
 import {
  AWSBedrockClaudeStream,
  AWSBedrockLlamaStream,
@@ -26,6 +28,7 @@ import { debugStream } from '../../utils/debugStream';
 import { getModelPricing } from '../../utils/getModelPricing';
 import { StreamingResponse } from '../../utils/response';
 import { resolveCacheTTL } from '../anthropic/resolveCacheTTL';
+import { resolveMaxTokens } from '../anthropic/resolveMaxTokens';

 /**
 * A prompt constructor for HuggingFace LLama 2 chat models.
@@ -62,19 +65,24 @@ export function experimental_buildLlama2Prompt(messages: { content: string; role
 export interface LobeBedrockAIParams {
  accessKeyId?: string;
  accessKeySecret?: string;
+  id?: string;
  region?: string;
  sessionToken?: string;
 }

 export class LobeBedrockAI implements LobeRuntimeAI {
  private client: BedrockRuntimeClient;
+  private id: string;

  region: string;

-  constructor({ region, accessKeyId, accessKeySecret, sessionToken }: LobeBedrockAIParams = {}) {
+  constructor(options: LobeBedrockAIParams = {}) {
+    const { id, region, accessKeyId, accessKeySecret, sessionToken } = options;
+
    if (!(accessKeyId && accessKeySecret))
      throw AgentRuntimeError.createError(AgentRuntimeErrorType.InvalidBedrockCredentials);
    this.region = region ?? 'us-east-1';
+    this.id = id ?? ModelProvider.Bedrock;
    this.client = new BedrockRuntimeClient({
      credentials: {
        accessKeyId: accessKeyId,
@@ -158,18 +166,28 @@ export class LobeBedrockAI implements LobeRuntimeAI {
      temperature,
      top_p,
      tools,
+      thinking,
    } = payload;
    const inputStartAt = Date.now();
    const system_message = messages.find((m) => m.role === 'system');
    const user_messages = messages.filter((m) => m.role !== 'system');

    // Resolve temperature and top_p parameters based on model constraints
-    const hasConflict = MODEL_PARAMETER_CONFLICTS.BEDROCK_CLAUDE_4_PLUS.has(model);
+    const hasConflict = hasTemperatureTopPConflict(model);
    const resolvedParams = resolveParameters(
      { temperature, top_p },
      { hasConflict, normalizeTemperature: true, preferTemperature: true },
    );

+    const { bedrock: bedrockModels } = await import('model-bank');
+
+    const resolvedMaxTokens = await resolveMaxTokens({
+      max_tokens,
+      model,
+      providerModels: bedrockModels,
+      thinking,
+    });
+
    const systemPrompts = !!system_message?.content
      ? ([
          {
@@ -177,19 +195,40 @@ export class LobeBedrockAI implements LobeRuntimeAI {
            text: system_message.content as string,
            type: 'text',
          },
-        ] as any)
+        ] as Anthropic.TextBlockParam[])
      : undefined;

-    const anthropicPayload = {
+    const postTools = buildAnthropicTools(tools, {
+      enabledContextCaching,
+    });
+
+    const anthropicBase = {
      anthropic_version: 'bedrock-2023-05-31',
-      max_tokens: max_tokens || 4096,
+      max_tokens: resolvedMaxTokens,
      messages: await buildAnthropicMessages(user_messages, { enabledContextCaching }),
      system: systemPrompts,
-      temperature: resolvedParams.temperature,
-      tools: buildAnthropicTools(tools, { enabledContextCaching }),
-      top_p: resolvedParams.top_p,
+      tools: postTools,
    };

+    const anthropicPayload =
+      thinking?.type === 'enabled'
+        ? {
+            ...anthropicBase,
+            thinking: {
+              ...thinking,
+              // `max_tokens` must be greater than `budget_tokens`
+              budget_tokens: Math.max(
+                1,
+                Math.min(thinking.budget_tokens || 1024, resolvedMaxTokens - 1),
+              ),
+            },
+          }
+        : {
+            ...anthropicBase,
+            temperature: resolvedParams.temperature,
+            top_p: resolvedParams.top_p,
+          };
+
    const command = new InvokeModelWithResponseStreamCommand({
      accept: 'application/json',
      body: JSON.stringify(anthropicPayload),
@@ -209,7 +248,7 @@ export class LobeBedrockAI implements LobeRuntimeAI {
        debugStream(debug).catch(console.error);
      }

-      const pricing = await getModelPricing(payload.model, ModelProvider.Bedrock);
+      const pricing = await getModelPricing(payload.model, this.id);
      const cacheTTL = resolveCacheTTL({ ...payload, enabledContextCaching }, anthropicPayload);
      const pricingOptions = cacheTTL ? { lookupParams: { ttl: cacheTTL } } : undefined;

@@ -218,7 +257,7 @@ export class LobeBedrockAI implements LobeRuntimeAI {
        AWSBedrockClaudeStream(prod, {
          callbacks: options?.callback,
          inputStartAt,
-          payload: { model, pricing, pricingOptions, provider: ModelProvider.Bedrock },
+          payload: { model, pricing, pricingOptions, provider: this.id },
        }),
        {
          headers: options?.headers,
--- a/src/store/agent/slices/chat/selectors/chatConfig.ts
+++ b/src/store/agent/slices/chat/selectors/chatConfig.ts
@@ -1,4 +1,5 @@
-import { contextCachingModels, thinkingWithToolClaudeModels } from '@/const/models';
+import { isContextCachingModel, isThinkingWithToolClaudeModel } from '@lobechat/model-runtime';
+
 import { DEFAULT_AGENT_CHAT_CONFIG, DEFAULT_AGENT_SEARCH_FC_MODEL } from '@/const/settings';
 import { AgentStoreState } from '@/store/agent/initialState';
 import { LobeAgentChatConfig } from '@/types/agent';
@@ -24,12 +25,12 @@ const enableHistoryCount = (s: AgentStoreState) => {
  // 如果开启了上下文缓存，且当前模型类型匹配，则不开启历史记录
  const enableContextCaching = !chatConfig.disableContextCaching;

-  if (enableContextCaching && contextCachingModels.has(config.model)) return false;
+  if (enableContextCaching && isContextCachingModel(config.model)) return false;

  // 当开启搜索时，针对 claude 3.7 sonnet 模型不开启历史记录
  const enableSearch = isAgentEnableSearch(s);

-  if (enableSearch && thinkingWithToolClaudeModels.has(config.model)) return false;
+  if (enableSearch && isThinkingWithToolClaudeModel(config.model)) return false;

  return chatConfig.enableHistoryCount;
 };