@vybestack/llxprt-code-core 0.5.0 → 0.6.0-nightly.251128.1049d5f2b

package/dist/src/core/baseLlmClient.js

@@ -0,0 +1,175 @@
+/**
+ * @license
+ * Copyright 2025 Google LLC
+ * SPDX-License-Identifier: Apache-2.0
+ */
+import { getResponseText } from '../utils/generateContentResponseUtilities.js';
+import { getErrorMessage } from '../utils/errors.js';
+/**
+ * Extracts JSON from a string that might be wrapped in markdown code blocks
+ * @param text - The raw text that might contain markdown-wrapped JSON
+ * @returns The extracted JSON string or the original text if no markdown found
+ */
+function extractJsonFromMarkdown(text) {
+    // Try to match ```json ... ``` or ``` ... ```
+    const markdownMatch = text.match(/```(?:json)?\s*\n?([\s\S]*?)\n?```/);
+    if (markdownMatch && markdownMatch[1]) {
+        return markdownMatch[1].trim();
+    }
+    // If no markdown found, return trimmed original text
+    return text.trim();
+}
+/**
+ * BaseLLMClient extracts stateless utility methods for LLM operations.
+ * Unlike the main Client class, this handles utility calls without conversation state.
+ *
+ * This implements the baseLlmClient pattern from upstream gemini-cli but adapted
+ * for llxprt's multi-provider architecture.
+ *
+ * Key features:
+ * - Multi-provider support (Anthropic, OpenAI, Gemini, Vertex AI)
+ * - Stateless operations (no conversation history)
+ * - Clean separation from GeminiClient
+ * - Dependency injection for testing
+ */
+export class BaseLLMClient {
+    contentGenerator;
+    constructor(contentGenerator) {
+        this.contentGenerator = contentGenerator;
+        if (!contentGenerator) {
+            throw new Error('ContentGenerator is required');
+        }
+    }
+    /**
+     * Generate structured JSON from a prompt with optional schema validation.
+     * Supports all providers through the ContentGenerator abstraction.
+     *
+     * @param options - Generation options including prompt, schema, model, etc.
+     * @returns Parsed JSON object
+     * @throws Error if generation fails or response cannot be parsed
+     */
+    async generateJson(options) {
+        const { prompt, schema, model, temperature = 0, systemInstruction, promptId = 'baseLlmClient-generateJson', } = options;
+        try {
+            const contents = [
+                {
+                    role: 'user',
+                    parts: [{ text: prompt }],
+                },
+            ];
+            const config = {
+                temperature,
+                topP: 1,
+            };
+            if (systemInstruction) {
+                config.systemInstruction = { text: systemInstruction };
+            }
+            if (schema) {
+                config.responseJsonSchema = schema;
+                config.responseMimeType = 'application/json';
+            }
+            const result = await this.contentGenerator.generateContent({
+                model,
+                config,
+                contents,
+            }, promptId);
+            let text = getResponseText(result);
+            if (!text) {
+                throw new Error('API returned an empty response for generateJson.');
+            }
+            // Handle markdown wrapping
+            const prefix = '```json';
+            const suffix = '```';
+            if (text.startsWith(prefix) && text.endsWith(suffix)) {
+                text = text
+                    .substring(prefix.length, text.length - suffix.length)
+                    .trim();
+            }
+            try {
+                // Extract JSON from potential markdown wrapper
+                const cleanedText = extractJsonFromMarkdown(text);
+                return JSON.parse(cleanedText);
+            }
+            catch (parseError) {
+                throw new Error(`Failed to parse API response as JSON: ${getErrorMessage(parseError)}`);
+            }
+        }
+        catch (error) {
+            throw new Error(`Failed to generate JSON content: ${getErrorMessage(error)}`);
+        }
+    }
+    /**
+     * Generate embeddings for text input.
+     * Supports single text string or array of strings.
+     *
+     * @param options - Embedding options including text and model
+     * @returns Embedding vector(s) as number array(s)
+     * @throws Error if generation fails or response is invalid
+     */
+    async generateEmbedding(options) {
+        const { text, model } = options;
+        try {
+            const texts = Array.isArray(text) ? text : [text];
+            const embedContentResponse = await this.contentGenerator.embedContent({
+                model,
+                contents: texts,
+            });
+            if (!embedContentResponse.embeddings ||
+                embedContentResponse.embeddings.length === 0) {
+                throw new Error('No embeddings found in API response.');
+            }
+            if (embedContentResponse.embeddings.length !== texts.length) {
+                throw new Error(`API returned a mismatched number of embeddings. Expected ${texts.length}, got ${embedContentResponse.embeddings.length}.`);
+            }
+            const embeddings = embedContentResponse.embeddings.map((embedding, index) => {
+                const values = embedding.values;
+                if (!values || values.length === 0) {
+                    throw new Error(`API returned an empty embedding for input text at index ${index}: "${texts[index]}"`);
+                }
+                return values;
+            });
+            // Return single array if input was a single string
+            return Array.isArray(text) ? embeddings : embeddings[0];
+        }
+        catch (error) {
+            throw new Error(`Failed to generate embedding: ${getErrorMessage(error)}`);
+        }
+    }
+    /**
+     * Count tokens in text or contents without making an API call to generate.
+     * Useful for checking context limits before generation.
+     *
+     * @param options - Options including text/contents and model
+     * @returns Token count
+     * @throws Error if counting fails
+     */
+    async countTokens(options) {
+        const { text, contents, model } = options;
+        try {
+            let requestContents;
+            if (contents) {
+                requestContents = contents;
+            }
+            else if (text) {
+                requestContents = [
+                    {
+                        role: 'user',
+                        parts: [{ text }],
+                    },
+                ];
+            }
+            else {
+                throw new Error('Either text or contents must be provided');
+            }
+            const response = await this.contentGenerator.countTokens({
+                model,
+                contents: requestContents,
+            });
+            return response.totalTokens ?? 0;
+        }
+        catch (error) {
+            throw new Error(`Failed to count tokens: ${getErrorMessage(error)}`);
+        }
+    }
+}
+//# sourceMappingURL=baseLlmClient.js.map

package/dist/src/core/baseLlmClient.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"baseLlmClient.js","sourceRoot":"","sources":["../../../src/core/baseLlmClient.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AASH,OAAO,EAAE,eAAe,EAAE,MAAM,8CAA8C,CAAC;AAC/E,OAAO,EAAE,eAAe,EAAE,MAAM,oBAAoB,CAAC;AA+BrD;;;;GAIG;AACH,SAAS,uBAAuB,CAAC,IAAY;IAC3C,8CAA8C;IAC9C,MAAM,aAAa,GAAG,IAAI,CAAC,KAAK,CAAC,oCAAoC,CAAC,CAAC;IACvE,IAAI,aAAa,IAAI,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC;QACtC,OAAO,aAAa,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC;IACjC,CAAC;IAED,qDAAqD;IACrD,OAAO,IAAI,CAAC,IAAI,EAAE,CAAC;AACrB,CAAC;AAED;;;;;;;;;;;;GAYG;AACH,MAAM,OAAO,aAAa;IACK;IAA7B,YAA6B,gBAAyC;QAAzC,qBAAgB,GAAhB,gBAAgB,CAAyB;QACpE,IAAI,CAAC,gBAAgB,EAAE,CAAC;YACtB,MAAM,IAAI,KAAK,CAAC,8BAA8B,CAAC,CAAC;QAClD,CAAC;IACH,CAAC;IAED;;;;;;;OAOG;IACH,KAAK,CAAC,YAAY,CAAc,OAA4B;QAC1D,MAAM,EACJ,MAAM,EACN,MAAM,EACN,KAAK,EACL,WAAW,GAAG,CAAC,EACf,iBAAiB,EACjB,QAAQ,GAAG,4BAA4B,GACxC,GAAG,OAAO,CAAC;QAEZ,IAAI,CAAC;YACH,MAAM,QAAQ,GAAc;gBAC1B;oBACE,IAAI,EAAE,MAAM;oBACZ,KAAK,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC;iBAC1B;aACF,CAAC;YAEF,MAAM,MAAM,GAA4B;gBACtC,WAAW;gBACX,IAAI,EAAE,CAAC;aACR,CAAC;YAEF,IAAI,iBAAiB,EAAE,CAAC;gBACtB,MAAM,CAAC,iBAAiB,GAAG,EAAE,IAAI,EAAE,iBAAiB,EAAE,CAAC;YACzD,CAAC;YAED,IAAI,MAAM,EAAE,CAAC;gBACX,MAAM,CAAC,kBAAkB,GAAG,MAAM,CAAC;gBACnC,MAAM,CAAC,gBAAgB,GAAG,kBAAkB,CAAC;YAC/C,CAAC;YAED,MAAM,MAAM,GACV,MAAM,IAAI,CAAC,gBAAiB,CAAC,eAAe,CAC1C;gBACE,KAAK;gBACL,MAAM;gBACN,QAAQ;aACT,EACD,QAAQ,CACT,CAAC;YAEJ,IAAI,IAAI,GAAG,eAAe,CAAC,MAAM,CAAC,CAAC;YACnC,IAAI,CAAC,IAAI,EAAE,CAAC;gBACV,MAAM,IAAI,KAAK,CAAC,kDAAkD,CAAC,CAAC;YACtE,CAAC;YAED,2BAA2B;YAC3B,MAAM,MAAM,GAAG,SAAS,CAAC;YACzB,MAAM,MAAM,GAAG,KAAK,CAAC;YACrB,IAAI,IAAI,CAAC,UAAU,CAAC,MAAM,CAAC,IAAI,IAAI,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,CAAC;gBACrD,IAAI,GAAG,IAAI;qBACR,SAAS,CAAC,MAAM,CAAC,MAAM,EAAE,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC,MAAM,CAAC;qBACrD,IAAI,EAAE,CAAC;YACZ,CAAC;YAED,IAAI,CAAC;gBACH,+CAA+C;gBAC/C,MAAM,WAAW,GAAG,uBAAuB,CAAC,IAAI,CAAC,CAAC;gBAClD,OAAO,IAAI,CAAC,KAAK,CAAC,WAAW,CAAM,CAAC;YACtC,CAAC;YAAC,OAAO,UAAU,EAAE,CAAC;gBACpB,MAAM,IAAI,KAAK,CACb,yCAAyC,eAAe,CACtD,UAAU,CACX,EAAE,CACJ,CAAC;YACJ,CAAC;QACH,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,IAAI,KAAK,CACb,oCAAoC,eAAe,CAAC,KAAK,CAAC,EAAE,CAC7D,CAAC;QACJ,CAAC;IACH,CAAC;IAED;;;;;;;OAOG;IACH,KAAK,CAAC,iBAAiB,CACrB,OAAiC;QAEjC,MAAM,EAAE,IAAI,EAAE,KAAK,EAAE,GAAG,OAAO,CAAC;QAEhC,IAAI,CAAC;YACH,MAAM,KAAK,GAAG,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC;YAElD,MAAM,oBAAoB,GACxB,MAAM,IAAI,CAAC,gBAAiB,CAAC,YAAY,CAAC;gBACxC,KAAK;gBACL,QAAQ,EAAE,KAAK;aAChB,CAAC,CAAC;YAEL,IACE,CAAC,oBAAoB,CAAC,UAAU;gBAChC,oBAAoB,CAAC,UAAU,CAAC,MAAM,KAAK,CAAC,EAC5C,CAAC;gBACD,MAAM,IAAI,KAAK,CAAC,sCAAsC,CAAC,CAAC;YAC1D,CAAC;YAED,IAAI,oBAAoB,CAAC,UAAU,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,EAAE,CAAC;gBAC5D,MAAM,IAAI,KAAK,CACb,4DAA4D,KAAK,CAAC,MAAM,SAAS,oBAAoB,CAAC,UAAU,CAAC,MAAM,GAAG,CAC3H,CAAC;YACJ,CAAC;YAED,MAAM,UAAU,GAAG,oBAAoB,CAAC,UAAU,CAAC,GAAG,CACpD,CAAC,SAAS,EAAE,KAAK,EAAE,EAAE;gBACnB,MAAM,MAAM,GAAG,SAAS,CAAC,MAAM,CAAC;gBAChC,IAAI,CAAC,MAAM,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;oBACnC,MAAM,IAAI,KAAK,CACb,2DAA2D,KAAK,MAAM,KAAK,CAAC,KAAK,CAAC,GAAG,CACtF,CAAC;gBACJ,CAAC;gBACD,OAAO,MAAM,CAAC;YAChB,CAAC,CACF,CAAC;YAEF,mDAAmD;YACnD,OAAO,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC;QAC1D,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,IAAI,KAAK,CACb,iCAAiC,eAAe,CAAC,KAAK,CAAC,EAAE,CAC1D,CAAC;QACJ,CAAC;IACH,CAAC;IAED;;;;;;;OAOG;IACH,KAAK,CAAC,WAAW,CAAC,OAA2B;QAC3C,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE,KAAK,EAAE,GAAG,OAAO,CAAC;QAE1C,IAAI,CAAC;YACH,IAAI,eAA0B,CAAC;YAE/B,IAAI,QAAQ,EAAE,CAAC;gBACb,eAAe,GAAG,QAAQ,CAAC;YAC7B,CAAC;iBAAM,IAAI,IAAI,EAAE,CAAC;gBAChB,eAAe,GAAG;oBAChB;wBACE,IAAI,EAAE,MAAM;wBACZ,KAAK,EAAE,CAAC,EAAE,IAAI,EAAE,CAAC;qBAClB;iBACF,CAAC;YACJ,CAAC;iBAAM,CAAC;gBACN,MAAM,IAAI,KAAK,CAAC,0CAA0C,CAAC,CAAC;YAC9D,CAAC;YAED,MAAM,QAAQ,GACZ,MAAM,IAAI,CAAC,gBAAiB,CAAC,WAAW,CAAC;gBACvC,KAAK;gBACL,QAAQ,EAAE,eAAe;aAC1B,CAAC,CAAC;YAEL,OAAO,QAAQ,CAAC,WAAW,IAAI,CAAC,CAAC;QACnC,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,IAAI,KAAK,CAAC,2BAA2B,eAAe,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;QACvE,CAAC;IACH,CAAC;CACF"}

package/dist/src/core/client.d.ts

@@ -17,7 +17,7 @@ import type { AgentRuntimeState } from '../runtime/AgentRuntimeState.js';
  *
  * Exported for testing purposes.
  */
-export declare function findIndexAfterFraction(history: Content[], fraction: number): number;
+export declare function findCompressSplitPoint(contents: Content[], fraction: number): number;
 export declare class GeminiClient {
     private readonly config;
     private chat?;
@@ -30,6 +30,7 @@ export declare class GeminiClient {
     private _pendingConfig?;
     private _previousHistory?;
     private _storedHistoryService?;
+    private currentSequenceModel;
     private readonly loopDetector;
     private lastPromptId?;
     private readonly complexityAnalyzer;
@@ -59,6 +60,11 @@ export declare class GeminiClient {
     private readonly runtimeState;
     private _historyService?;
     private _unsubscribe?;
+    /**
+     * BaseLLMClient for stateless utility operations (generateJson, embeddings, etc.)
+     * Lazily initialized when needed
+     */
+    private _baseLlmClient?;
     /**
      * @plan PLAN-20251027-STATELESS5.P10
      * @requirement REQ-STAT5-003.1
@@ -74,6 +80,11 @@ export declare class GeminiClient {
     private lazyInitialize;
     getContentGenerator(): ContentGenerator;
     getUserTier(): UserTierId | undefined;
+    /**
+     * Get or create the BaseLLMClient for stateless utility operations.
+     * This is lazily initialized to avoid creating it when not needed.
+     */
+    private getBaseLlmClient;
     private processComplexityAnalysis;
     private shouldEscalateReminder;
     private isTodoToolCall;
@@ -113,6 +124,7 @@ export declare class GeminiClient {
     setTools(): Promise<void>;
     clearTools(): void;
     resetChat(): Promise<void>;
+    getCurrentSequenceModel(): string | null;
     addDirectoryContext(): Promise<void>;
     generateDirectMessage(params: SendMessageParameters, promptId: string): Promise<GenerateContentResponse>;
     startChat(extraHistory?: Content[]): Promise<GeminiChat>;

package/dist/src/core/client.js

@@ -7,7 +7,6 @@ import { getDirectoryContextString, getEnvironmentContext, } from '../utils/envi
 import { Turn, GeminiEventType, DEFAULT_AGENT_ID, } from './turn.js';
 import { CompressionStatus } from './turn.js';
 import { getCoreSystemPromptAsync, getCompressionPrompt } from './prompts.js';
-import { getResponseText } from '../utils/generateContentResponseUtilities.js';
 import { reportError } from '../utils/errorReporting.js';
 import { GeminiChat } from './geminiChat.js';
 import { DebugLogger } from '../debug/index.js';
@@ -25,10 +24,12 @@ import { LoopDetectionService } from '../services/loopDetectionService.js';
 import { ideContext } from '../ide/ideContext.js';
 import { ComplexityAnalyzer, } from '../services/complexity-analyzer.js';
 import { TodoReminderService } from '../services/todo-reminder-service.js';
+import { uiTelemetryService } from '../telemetry/uiTelemetry.js';
 import { TodoStore } from '../tools/todo-store.js';
 import { isFunctionResponse } from '../utils/messageInspectors.js';
 import { estimateTokens as estimateTextTokens } from '../utils/toolOutputLimiter.js';
 import { subscribeToAgentRuntimeState } from '../runtime/AgentRuntimeState.js';
+import { BaseLLMClient } from './baseLlmClient.js';
 const COMPLEXITY_ESCALATION_TURN_THRESHOLD = 3;
 const TODO_PROMPT_SUFFIX = 'Use TODO List to organize this effort.';
 function isThinkingSupported(model) {
@@ -36,40 +37,37 @@ function isThinkingSupported(model) {
         return true;
     return false;
 }
-/**
- * Extracts JSON from a string that might be wrapped in markdown code blocks
- * @param text - The raw text that might contain markdown-wrapped JSON
- * @returns The extracted JSON string or the original text if no markdown found
- */
-function extractJsonFromMarkdown(text) {
-    // Try to match ```json ... ``` or ``` ... ```
-    const markdownMatch = text.match(/```(?:json)?\s*\n?([\s\S]*?)\n?```/);
-    if (markdownMatch && markdownMatch[1]) {
-        return markdownMatch[1].trim();
-    }
-    // If no markdown found, return trimmed original text
-    return text.trim();
-}
 /**
  * Returns the index of the content after the fraction of the total characters in the history.
  *
  * Exported for testing purposes.
  */
-export function findIndexAfterFraction(history, fraction) {
+export function findCompressSplitPoint(contents, fraction) {
     if (fraction <= 0 || fraction >= 1) {
         throw new Error('Fraction must be between 0 and 1');
     }
-    const contentLengths = history.map((content) => JSON.stringify(content).length);
-    const totalCharacters = contentLengths.reduce((sum, length) => sum + length, 0);
-    const targetCharacters = totalCharacters * fraction;
-    let charactersSoFar = 0;
-    for (let i = 0; i < contentLengths.length; i++) {
-        charactersSoFar += contentLengths[i];
-        if (charactersSoFar >= targetCharacters) {
-            return i;
+    const charCounts = contents.map((content) => JSON.stringify(content).length);
+    const totalCharCount = charCounts.reduce((sum, length) => sum + length, 0);
+    const targetCharCount = totalCharCount * fraction;
+    let lastSplitPoint = 0;
+    let cumulativeCharCount = 0;
+    for (let i = 0; i < contents.length; i++) {
+        const content = contents[i];
+        const hasFunctionResponse = content.parts?.some((part) => !!part.functionResponse);
+        if (content.role === 'user' && !hasFunctionResponse) {
+            if (cumulativeCharCount >= targetCharCount) {
+                return i;
+            }
+            lastSplitPoint = i;
         }
+        cumulativeCharCount += charCounts[i];
+    }
+    const lastContent = contents[contents.length - 1];
+    if (lastContent?.role === 'model' &&
+        !lastContent?.parts?.some((part) => part.functionCall)) {
+        return contents.length;
     }
-    return contentLengths.length;
+    return lastSplitPoint;
 }
 export class GeminiClient {
     config;
@@ -86,6 +84,7 @@ export class GeminiClient {
     _pendingConfig;
     _previousHistory;
     _storedHistoryService;
+    currentSequenceModel = null;
     loopDetector;
     lastPromptId;
     complexityAnalyzer;
@@ -115,6 +114,11 @@ export class GeminiClient {
     runtimeState;
     _historyService;
     _unsubscribe;
+    /**
+     * BaseLLMClient for stateless utility operations (generateJson, embeddings, etc.)
+     * Lazily initialized when needed
+     */
+    _baseLlmClient;
     /**
      * @plan PLAN-20251027-STATELESS5.P10
      * @requirement REQ-STAT5-003.1
@@ -210,6 +214,16 @@ export class GeminiClient {
     getUserTier() {
         return this.contentGenerator?.userTier;
     }
+    /**
+     * Get or create the BaseLLMClient for stateless utility operations.
+     * This is lazily initialized to avoid creating it when not needed.
+     */
+    getBaseLlmClient() {
+        if (!this._baseLlmClient) {
+            this._baseLlmClient = new BaseLLMClient(this.getContentGenerator());
+        }
+        return this._baseLlmClient;
+    }
     processComplexityAnalysis(analysis) {
         if (!this.todoToolsAvailable) {
             this.consecutiveComplexTurns = 0;
@@ -514,6 +528,9 @@ export class GeminiClient {
         // Clear the stored history as well
         this._previousHistory = [];
     }
+    getCurrentSequenceModel() {
+        return this.currentSequenceModel;
+    }
     async addDirectoryContext() {
         if (!this.chat) {
             return;
@@ -827,6 +844,7 @@ export class GeminiClient {
         if (this.lastPromptId !== prompt_id) {
             this.loopDetector.reset(prompt_id);
             this.lastPromptId = prompt_id;
+            this.currentSequenceModel = null;
         }
         this.sessionTurnCount++;
         this.toolActivityCount = 0;
@@ -846,7 +864,7 @@ export class GeminiClient {
             const providerName = providerManager?.getActiveProviderName() || 'backend';
             return new Turn(this.getChat(), prompt_id, DEFAULT_AGENT_ID, providerName);
         }
-        const compressed = await this.tryCompressChat(prompt_id);
+        const compressed = await this.tryCompressChat(prompt_id, false);
         if (compressed.compressionStatus === CompressionStatus.COMPRESSED) {
             yield { type: GeminiEventType.ChatCompressed, value: compressed };
         }
@@ -1070,78 +1088,61 @@ export class GeminiClient {
     }
     async generateJson(contents, schema, abortSignal, model, config = {}) {
         await this.lazyInitialize();
-        // Use the provided model parameter directly
         const modelToUse = model;
         try {
             const userMemory = this.config.getUserMemory();
-            // Provider name removed from prompt call signature
             const systemInstruction = await getCoreSystemPromptAsync(userMemory, modelToUse, this.getEnabledToolNamesForPrompt());
-            const requestConfig = {
-                abortSignal,
-                ...this.generateContentConfig,
-                ...config,
+            // Convert Content[] to a single prompt for BaseLLMClient
+            // This preserves the conversation context in the prompt
+            const prompt = contents
+                .map((c) => c.parts
+                ?.map((p) => ('text' in p ? p.text : ''))
+                .filter(Boolean)
+                .join('\n'))
+                .filter(Boolean)
+                .join('\n\n');
+            // Use BaseLLMClient for the core JSON generation
+            // This delegates to the stateless utility layer
+            const baseLlmClient = this.getBaseLlmClient();
+            const apiCall = async () => {
+                try {
+                    return await baseLlmClient.generateJson({
+                        prompt,
+                        schema,
+                        model: modelToUse,
+                        temperature: config.temperature ?? this.generateContentConfig.temperature,
+                        systemInstruction,
+                        promptId: this.lastPromptId || this.config.getSessionId(),
+                    });
+                }
+                catch (error) {
+                    // Preserve abort signal behavior
+                    if (abortSignal.aborted) {
+                        throw error;
+                    }
+                    throw error;
+                }
             };
-            const apiCall = () => this.getContentGenerator().generateContent({
-                model: modelToUse,
-                config: {
-                    ...requestConfig,
-                    systemInstruction,
-                    responseJsonSchema: schema,
-                    responseMimeType: 'application/json',
-                },
-                contents,
-            }, this.lastPromptId || this.config.getSessionId());
             const result = await retryWithBackoff(apiCall);
-            let text = getResponseText(result);
-            if (!text) {
-                const error = new Error('API returned an empty response for generateJson.');
-                await reportError(error, 'Error in generateJson: API returned an empty response.', contents, 'generateJson-empty-response');
-                throw error;
-            }
-            const prefix = '```json';
-            const suffix = '```';
-            if (text.startsWith(prefix) && text.endsWith(suffix)) {
-                // Note: upstream added logMalformedJsonResponse here but our telemetry doesn't have it
-                text = text
-                    .substring(prefix.length, text.length - suffix.length)
-                    .trim();
-            }
-            try {
-                // Extract JSON from potential markdown wrapper
-                const cleanedText = extractJsonFromMarkdown(text);
-                // Special case: Gemini sometimes returns just "user" or "model" for next speaker checks
-                // This happens particularly with non-ASCII content in the conversation
-                if ((cleanedText === 'user' || cleanedText === 'model') &&
-                    contents.some((c) => c.parts?.some((p) => 'text' in p && p.text?.includes('next_speaker')))) {
-                    this.logger.warn(() => `[generateJson] Gemini returned plain text "${cleanedText}" instead of JSON for next speaker check. Converting to valid response.`);
-                    return {
-                        reasoning: 'Gemini returned plain text response',
-                        next_speaker: cleanedText,
-                    };
-                }
-                return JSON.parse(cleanedText);
-            }
-            catch (parseError) {
-                // Log both the original and cleaned text for debugging
-                await reportError(parseError, 'Failed to parse JSON response from generateJson.', {
-                    responseTextFailedToParse: text,
-                    cleanedTextFailedToParse: extractJsonFromMarkdown(text),
-                    originalRequestContents: contents,
-                }, 'generateJson-parse');
-                throw new Error(`Failed to parse API response as JSON: ${getErrorMessage(parseError)}`);
+            // Special case: Gemini sometimes returns just "user" or "model" for next speaker checks
+            // This happens particularly with non-ASCII content in the conversation
+            if (typeof result === 'string' &&
+                (result === 'user' || result === 'model') &&
+                contents.some((c) => c.parts?.some((p) => 'text' in p && p.text?.includes('next_speaker')))) {
+                this.logger.warn(() => `[generateJson] Gemini returned plain text "${result}" instead of JSON for next speaker check. Converting to valid response.`);
+                return {
+                    reasoning: 'Gemini returned plain text response',
+                    next_speaker: result,
+                };
             }
+            return result;
         }
         catch (error) {
             if (abortSignal.aborted) {
                 throw error;
             }
-            // Avoid double reporting for the empty response case handled above
-            if (error instanceof Error &&
-                error.message === 'API returned an empty response for generateJson.') {
-                throw error;
-            }
             await reportError(error, 'Error generating JSON content via API.', contents, 'generateJson-api');
-            throw new Error(`Failed to generate JSON content: ${getErrorMessage(error)}`);
+            throw error;
         }
     }
     async generateContent(contents, generationConfig, abortSignal, model) {
@@ -1184,25 +1185,14 @@ export class GeminiClient {
         if (!texts || texts.length === 0) {
             return [];
         }
-        const embedModelParams = {
+        // Delegate to BaseLLMClient for stateless embedding generation
+        const baseLlmClient = this.getBaseLlmClient();
+        const result = await baseLlmClient.generateEmbedding({
+            text: texts,
             model: this.embeddingModel,
-            contents: texts,
-        };
-        const embedContentResponse = await this.getContentGenerator().embedContent(embedModelParams);
-        if (!embedContentResponse.embeddings ||
-            embedContentResponse.embeddings.length === 0) {
-            throw new Error('No embeddings found in API response.');
-        }
-        if (embedContentResponse.embeddings.length !== texts.length) {
-            throw new Error(`API returned a mismatched number of embeddings. Expected ${texts.length}, got ${embedContentResponse.embeddings.length}.`);
-        }
-        return embedContentResponse.embeddings.map((embedding, index) => {
-            const values = embedding.values;
-            if (!values || values.length === 0) {
-                throw new Error(`API returned an empty embedding for input text at index ${index}: "${texts[index]}"`);
-            }
-            return values;
         });
+        // Result is already validated by BaseLLMClient
+        return result;
     }
     /**
      * Manually trigger chat compression
@@ -1227,24 +1217,12 @@ export class GeminiClient {
                 compressionStatus: CompressionStatus.NOOP,
             };
         }
-        // Note: chat variable used later in method
+        // Use lastPromptTokenCount from telemetry service as the source of truth
+        // This is more accurate than estimating from history
+        const originalTokenCount = uiTelemetryService.getLastPromptTokenCount();
         // @plan PLAN-20251027-STATELESS5.P10
         // @requirement REQ-STAT5-003.1
         const model = this.runtimeState.model;
-        // Get the ACTUAL token count from the history service, not the curated subset
-        const historyService = this.getChat().getHistoryService();
-        const originalTokenCount = historyService
-            ? historyService.getTotalTokens()
-            : 0;
-        if (originalTokenCount === undefined) {
-            console.warn(`Could not determine token count for model ${model}.`);
-            this.hasFailedCompressionAttempt = !force && true;
-            return {
-                originalTokenCount: 0,
-                newTokenCount: 0,
-                compressionStatus: CompressionStatus.COMPRESSION_FAILED_TOKEN_COUNT_ERROR,
-            };
-        }
         const contextPercentageThreshold = this.config.getChatCompression()?.contextPercentageThreshold;
         // Don't compress if not forced and we are under the limit.
         if (!force) {
@@ -1259,7 +1237,7 @@ export class GeminiClient {
                 };
             }
         }
-        let compressBeforeIndex = findIndexAfterFraction(curatedHistory, 1 - COMPRESSION_PRESERVE_THRESHOLD);
+        let compressBeforeIndex = findCompressSplitPoint(curatedHistory, 1 - COMPRESSION_PRESERVE_THRESHOLD);
         // Find the first user message after the index. This is the start of the next turn.
         while (compressBeforeIndex < curatedHistory.length &&
             (curatedHistory[compressBeforeIndex]?.role === 'model' ||
@@ -1310,7 +1288,6 @@ export class GeminiClient {
         // TODO: Add proper telemetry logging once available
         console.debug(`Chat compression: ${originalTokenCount} -> ${newTokenCount} tokens`);
         if (newTokenCount > originalTokenCount) {
-            this.getChat().setHistory(curatedHistory);
             this.hasFailedCompressionAttempt = !force && true;
             return {
                 originalTokenCount,
@@ -1320,6 +1297,8 @@ export class GeminiClient {
         }
         else {
             this.chat = compressedChat; // Chat compression successful, set new state.
+            // Update telemetry service with new token count
+            uiTelemetryService.setLastPromptTokenCount(newTokenCount);
             // Emit token update event for the new compressed chat
             // This ensures the UI updates with the new token count
             // Only emit if compression was successful