feat: handle Opus 4.7 unsupported sampling params (temperature/topP/topK)

- Add noSamplingParams flag to FeatureFlags
- Add CLAUDE_OPUS_4_7_DEFAULT_PARAMS without temperature
- Add model entries for global/us.anthropic.claude-opus-4-7
- Strip temperature/topP from inferenceConfig for noSamplingParams models
- Add prompt caching support for Opus 4.7

Author: kenicazu

packages/cdk/lambda/utils/models.ts

@@ -137,6 +137,12 @@ const CLAUDE_OPUS_4_6_DEFAULT_PARAMS: ConverseInferenceParams = {
   },
 };
 
+const CLAUDE_OPUS_4_7_DEFAULT_PARAMS: ConverseInferenceParams = {
+  inferenceConfig: {
+    maxTokens: 128000,
+  },
+};
+
 const CLAUDE_SONNET_4_6_DEFAULT_PARAMS: ConverseInferenceParams = {
   inferenceConfig: {
     maxTokens: 64000,
@@ -504,7 +510,9 @@ const createConverseCommandInput = (
     modelId: model.modelId,
     messages: conversationWithCache,
     system: systemContextWithCache,
-    inferenceConfig: params.inferenceConfig,
+    inferenceConfig: modelMetadata[model.modelId].flags.noSamplingParams
+      ? { maxTokens: params.inferenceConfig?.maxTokens }
+      : params.inferenceConfig,
     guardrailConfig,
   };
 
@@ -513,9 +521,10 @@ const createConverseCommandInput = (
     (model.modelParameters?.reasoningConfig?.type === 'enabled' ||
       model.modelParameters?.reasoningConfig?.type === 'adaptive')
   ) {
+    const noSampling = modelMetadata[model.modelId].flags.noSamplingParams;
     converseCommandInput.inferenceConfig = {
       ...params.inferenceConfig,
-      temperature: 1, // reasoning requires temperature to be 1
+      temperature: noSampling ? undefined : 1, // reasoning requires temperature to be 1, but some models don't support it
       topP: undefined, // reasoning does not require topP
       maxTokens: params.inferenceConfig?.maxTokens,
     };
@@ -989,6 +998,22 @@ export const BEDROCK_TEXT_GEN_MODELS: {
     extractConverseStreamOutput: (body: ConverseStreamOutput) => StreamingChunk;
   };
 } = {
+  'global.anthropic.claude-opus-4-7': {
+    defaultParams: CLAUDE_OPUS_4_7_DEFAULT_PARAMS,
+    usecaseParams: USECASE_DEFAULT_PARAMS,
+    createConverseCommandInput: createConverseCommandInput,
+    createConverseStreamCommandInput: createConverseStreamCommandInput,
+    extractConverseOutput: extractConverseOutput,
+    extractConverseStreamOutput: extractConverseStreamOutput,
+  },
+  'us.anthropic.claude-opus-4-7': {
+    defaultParams: CLAUDE_OPUS_4_7_DEFAULT_PARAMS,
+    usecaseParams: USECASE_DEFAULT_PARAMS,
+    createConverseCommandInput: createConverseCommandInput,
+    createConverseStreamCommandInput: createConverseStreamCommandInput,
+    extractConverseOutput: extractConverseOutput,
+    extractConverseStreamOutput: extractConverseStreamOutput,
+  },
   'global.anthropic.claude-opus-4-6-v1': {
     defaultParams: CLAUDE_OPUS_4_6_DEFAULT_PARAMS,
     usecaseParams: USECASE_DEFAULT_PARAMS,

packages/common/src/application/model.ts

@@ -48,6 +48,15 @@ const MODEL_FEATURE: Record<string, FeatureFlags> = {
     reasoning: true,
     adaptiveThinking: true,
   },
+  TEXT_DOC_IMAGE_ADAPTIVE_THINKING_NO_SAMPLING: {
+    text: true,
+    doc: true,
+    image: true,
+    video: false,
+    reasoning: true,
+    adaptiveThinking: true,
+    noSamplingParams: true,
+  },
   TEXT_DOC_IMAGE_VIDEO: { text: true, doc: true, image: true, video: true },
   IMAGE_GEN: { image_gen: true },
   VIDEO_GEN: { video_gen: true },
@@ -110,11 +119,11 @@ export const modelMetadata: Record<string, ModelMetadata> = {
     displayName: 'Claude Opus 4',
   },
   'global.anthropic.claude-opus-4-7': {
-    flags: MODEL_FEATURE.TEXT_DOC_IMAGE_ADAPTIVE_THINKING,
+    flags: MODEL_FEATURE.TEXT_DOC_IMAGE_ADAPTIVE_THINKING_NO_SAMPLING,
     displayName: 'Claude Opus 4.7',
   },
   'us.anthropic.claude-opus-4-7': {
-    flags: MODEL_FEATURE.TEXT_DOC_IMAGE_ADAPTIVE_THINKING,
+    flags: MODEL_FEATURE.TEXT_DOC_IMAGE_ADAPTIVE_THINKING_NO_SAMPLING,
     displayName: 'Claude Opus 4.7',
   },
   'global.anthropic.claude-opus-4-6-v1': {
@@ -735,6 +744,7 @@ export const BEDROCK_SPEECH_TO_SPEECH_MODELS = Object.keys(
 // Prompt caching
 // https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html
 export const SUPPORTED_CACHE_FIELDS: Record<string, PromptCacheField[]> = {
+  'anthropic.claude-opus-4-7': ['messages', 'system', 'tools'],
   'anthropic.claude-opus-4-6-v1': ['messages', 'system', 'tools'],
   'anthropic.claude-sonnet-4-6': ['messages', 'system', 'tools'],
   'anthropic.claude-opus-4-5-20251101-v1:0': ['messages', 'system', 'tools'],

packages/types/src/model.d.ts

@@ -7,6 +7,7 @@ export type FeatureFlags = {
   video?: boolean;
   reasoning?: boolean;
   adaptiveThinking?: boolean;
+  noSamplingParams?: boolean;
 
   image_gen?: boolean;
   video_gen?: boolean;