fix sampling params

packages/cli/src/config/config.ts

@@ -5,6 +5,7 @@
  */
 
 import type {
+  ContentGeneratorConfig,
   FileFilteringOptions,
   MCPServerConfig,
   OutputFormat,
@@ -123,6 +124,24 @@ export interface CliArgs {
   outputFormat: string | undefined;
 }
 
+type LegacySamplingSettings = {
+  sampling_params?: ContentGeneratorConfig['samplingParams'];
+};
+
+function getLegacySamplingParams(
+  settings: Settings,
+): ContentGeneratorConfig['samplingParams'] | undefined {
+  if (
+    typeof settings !== 'object' ||
+    settings === null ||
+    !('sampling_params' in (settings as Record<string, unknown>))
+  ) {
+    return undefined;
+  }
+
+  return (settings as Settings & LegacySamplingSettings).sampling_params;
+}
+
 export async function parseArguments(settings: Settings): Promise<CliArgs> {
   const rawArgv = hideBin(process.argv);
   const yargsInstance = yargs(rawArgv)
@@ -685,6 +704,7 @@ export async function loadCliConfig(
 
   const vlmSwitchMode =
     argv.vlmSwitchMode || settings.experimental?.vlmSwitchMode;
+  const legacySamplingParams = getLegacySamplingParams(settings);
   return new Config({
     sessionId,
     embeddingModel: DEFAULT_QWEN_EMBEDDING_MODEL,
@@ -745,6 +765,8 @@ export async function loadCliConfig(
         (typeof argv.openaiLogging === 'undefined'
           ? settings.model?.enableOpenAILogging
           : argv.openaiLogging) ?? false,
+      // Include sampling_params from root level settings
+      ...(legacySamplingParams ? { samplingParams: legacySamplingParams } : {}),
     },
     cliVersion: await getCliVersion(),
     tavilyApiKey:

packages/core/src/core/baseLlmClient.ts

@@ -65,10 +65,7 @@ export interface GenerateJsonOptions {
  */
 export class BaseLlmClient {
   // Default configuration for utility tasks
-  private readonly defaultUtilityConfig: GenerateContentConfig = {
-    temperature: 0,
-    topP: 1,
-  };
+  private readonly defaultUtilityConfig: GenerateContentConfig = {};
 
   constructor(
     private readonly contentGenerator: ContentGenerator,

packages/core/src/core/client.ts

@@ -149,10 +149,7 @@ const COMPRESSION_PRESERVE_THRESHOLD = 0.3;
 
 export class GeminiClient {
   private chat?: GeminiChat;
-  private readonly generateContentConfig: GenerateContentConfig = {
-    temperature: 0,
-    topP: 1,
-  };
+  private readonly generateContentConfig: GenerateContentConfig;
   private sessionTurnCount = 0;
 
   private readonly loopDetector: LoopDetectionService;
@@ -169,6 +166,44 @@ export class GeminiClient {
   constructor(private readonly config: Config) {
     this.loopDetector = new LoopDetectionService(config);
     this.lastPromptId = this.config.getSessionId();
+    this.generateContentConfig = this.buildDefaultGenerateContentConfig();
+  }
+
+  private buildDefaultGenerateContentConfig(): GenerateContentConfig {
+    const samplingParams =
+      this.config.getContentGeneratorConfig()?.samplingParams;
+
+    if (!samplingParams) {
+      return {};
+    }
+
+    const config: GenerateContentConfig = {};
+
+    if (samplingParams.temperature !== undefined) {
+      config.temperature = samplingParams.temperature;
+    }
+
+    if (samplingParams.top_p !== undefined) {
+      config.topP = samplingParams.top_p;
+    }
+
+    if (samplingParams.top_k !== undefined) {
+      config.topK = samplingParams.top_k;
+    }
+
+    if (samplingParams.max_tokens !== undefined) {
+      config.maxOutputTokens = samplingParams.max_tokens;
+    }
+
+    if (samplingParams.presence_penalty !== undefined) {
+      config.presencePenalty = samplingParams.presence_penalty;
+    }
+
+    if (samplingParams.frequency_penalty !== undefined) {
+      config.frequencyPenalty = samplingParams.frequency_penalty;
+    }
+
+    return config;
   }
 
   async initialize() {