@saccolabs/tars 1.31.0 → 1.32.0

package/dist/supervisor/gemini-engine.js

@@ -1,1056 +0,0 @@
-import { Config as CoreConfig, GeminiEventType, AuthType, promptIdContext, Scheduler, ApprovalMode, PolicyDecision, SimpleExtensionLoader, MCPServerConfig, CompressionStatus, loadConversationRecord } from '@google/gemini-cli-core';
-import { EventEmitter } from 'events';
-import logger from '../utils/logger.js';
-import { v4 as uuidv4 } from 'uuid';
-import fs from 'fs';
-import path from 'path';
-import { SendNotificationTool } from '../tools/send-notification.js';
-import { GetQuotaTool } from '../tools/get-quota.js';
-import { LocalRateLimiter } from './rate-limiter.js';
-import { LlamaCppGenerator } from '../inference/LlamaCppGenerator.js';
-import { DLPService } from '../utils/dlp-service.js';
-/**
- * Detects the best authentication type based on environment variables.
- * (Local implementation since it's not exported from core index)
- */
-function getAuthTypeFromEnv() {
-    if (process.env['GOOGLE_GENAI_USE_GCA'] === 'true') {
-        return AuthType.LOGIN_WITH_GOOGLE;
-    }
-    if (process.env['GOOGLE_GENAI_USE_VERTEXAI'] === 'true') {
-        return AuthType.USE_VERTEX_AI;
-    }
-    if (process.env['GEMINI_API_KEY']) {
-        return AuthType.USE_GEMINI;
-    }
-    return undefined;
-}
-/**
- * GeminiEngine - Native replacement for GeminiCli subprocess
- *
- * Uses @google/gemini-cli-core directly to interact with Gemini models.
- * Operates within the ~/.tars isolated environment by overriding HOME.
- */
-export class GeminiEngine extends EventEmitter {
-    tarsConfig;
-    coreConfig;
-    client;
-    initialized = false;
-    initializedWithFallback = false;
-    currentSessionId = null;
-    channelManager;
-    sessionManager;
-    rateLimiter;
-    constructor(tarsConfig) {
-        super();
-        this.tarsConfig = tarsConfig;
-        this.rateLimiter = new LocalRateLimiter(tarsConfig.maxRPM || 14, tarsConfig.maxTPM || 900000);
-    }
-    /**
-     * Provide the ChannelManager instance to the engine so it can build proactive notification tools
-     */
-    setChannelManager(channelManager) {
-        this.channelManager = channelManager;
-    }
-    /**
-     * Provide the SessionManager instance to the engine for session-aware tools
-     */
-    setSessionManager(sessionManager) {
-        this.sessionManager = sessionManager;
-    }
-    /**
-     * Initializes the core Gemini client with proper auth and config.
-     */
-    async initialize(initialSessionId) {
-        if (this.initialized)
-            return;
-        logger.info('🚀 Initializing Gemini Engine (Native Core)...');
-        const savedHome = process.env.HOME;
-        try {
-            // Ensure home directory exists
-            if (!fs.existsSync(this.tarsConfig.homeDir)) {
-                fs.mkdirSync(this.tarsConfig.homeDir, { recursive: true });
-            }
-            // Isolating to ~/.tars
-            process.env.HOME = this.tarsConfig.homeDir;
-            process.env.GEMINI_CLI_HOME = this.tarsConfig.homeDir;
-            // Tell the Gemini Core PromptProvider to use our custom system.md
-            const systemMdPath = path.join(this.tarsConfig.homeDir, '.gemini', 'system.md');
-            process.env.GEMINI_SYSTEM_MD = systemMdPath;
-            let authType = getAuthTypeFromEnv() || AuthType.LOGIN_WITH_GOOGLE;
-            // Prevent interactive Google login prompt if using local inference
-            if (this.tarsConfig.inferenceBackend === 'llamacpp') {
-                authType = AuthType.USE_GEMINI;
-                if (!process.env.GEMINI_API_KEY) {
-                    process.env.GEMINI_API_KEY = 'dummy_llama_key_to_bypass_sdk_auth';
-                }
-            }
-            const discoveredExtensions = await this.discoverExtensions();
-            const extensionLoader = new SimpleExtensionLoader(discoveredExtensions);
-            this.coreConfig = new CoreConfig({
-                sessionId: initialSessionId || uuidv4(),
-                targetDir: this.tarsConfig.homeDir,
-                cwd: this.tarsConfig.homeDir,
-                model: this.tarsConfig.geminiModel,
-                debugMode: false,
-                approvalMode: ApprovalMode.YOLO,
-                disableModelRouterForAuth: this.tarsConfig.inferenceBackend === 'llamacpp'
-                    ? [AuthType.USE_GEMINI]
-                    : undefined,
-                policyEngineConfig: {
-                    defaultDecision: PolicyDecision.ALLOW
-                },
-                interactive: true,
-                enableHooks: true,
-                mcpEnabled: true,
-                extensionsEnabled: true,
-                enableAgents: true,
-                skillsSupport: true,
-                adminSkillsEnabled: true,
-                noBrowser: true,
-                folderTrust: true,
-                trustedFolder: true,
-                extensionLoader
-            });
-            await this.coreConfig.refreshAuth(authType);
-            await this.coreConfig.initialize();
-            // Handle Local Inference Override
-            if (this.tarsConfig.inferenceBackend === 'llamacpp') {
-                logger.info(`🔌 Overriding Gemini Core with Local Inference: ${this.tarsConfig.localInferenceUrl}`);
-                const localGenerator = new LlamaCppGenerator(this.tarsConfig.localInferenceUrl);
-                // Override the content generator at runtime to bypass the SDK's internal Gemini calls
-                this.coreConfig.contentGenerator = localGenerator;
-                // We'll apply more overrides after the client is created
-            }
-            // Register system prompt template for tars-request
-            const promptProvider = this.coreConfig.promptProvider;
-            if (promptProvider) {
-                promptProvider.registerPrompt('tars-request', {
-                    template: fs.readFileSync(systemMdPath, 'utf-8'),
-                    includeContext: true,
-                    includeTools: true,
-                    includeHistory: true
-                });
-            }
-            // Inject native tools
-            const toolRegistry = this.coreConfig.getToolRegistry();
-            if (this.channelManager) {
-                const notifyTool = new SendNotificationTool(this.channelManager);
-                toolRegistry.registerTool(notifyTool);
-                logger.info('🔌 Registered native tool: send_notification');
-            }
-            const getQuotaTool = new GetQuotaTool(this.coreConfig, this.sessionManager, {
-                inferenceBackend: this.tarsConfig.inferenceBackend,
-                contextWindowTokens: this.tarsConfig.contextWindowTokens,
-                geminiModel: this.tarsConfig.geminiModel,
-                localInferenceUrl: this.tarsConfig.localInferenceUrl
-            });
-            toolRegistry.registerTool(getQuotaTool);
-            logger.info('🔌 Registered native tool: get_model_quota');
-            this.client = this.coreConfig.getGeminiClient();
-            this.applyClientOverrides(this.client);
-            // Deregister plan-mode tools — they require interactive user confirmation
-            // that Tars cannot provide (non-interactive agent). Without this, the model
-            // calls enter_plan_mode which switches ApprovalMode to PLAN, but exit_plan_mode
-            // silently fails ("Rejected (no feedback)"), leaving the agent permanently
-            // stuck in PLAN mode. This forces all requests through the rate-limited
-            // gemini-3.1-pro-preview model, exhausting quota instantly.
-            try {
-                toolRegistry.unregisterTool('enter_plan_mode');
-                toolRegistry.unregisterTool('exit_plan_mode');
-                logger.info('🔇 Deregistered plan-mode tools (non-interactive agent)');
-            }
-            catch (e) {
-                logger.debug(`Plan-mode tool deregistration skipped: ${e.message}`);
-            }
-            this.initialized = true;
-            this.currentSessionId = this.coreConfig.getSessionId();
-            logger.info('✨ Gemini Engine initialized successfully.');
-        }
-        catch (error) {
-            logger.error(`❌ Failed to initialize Gemini Engine: ${error.message}`);
-            throw error;
-        }
-        finally {
-            process.env.HOME = savedHome;
-        }
-    }
-    /**
-     * Discovers extensions from the ~/.tars/.gemini/extensions directory.
-     */
-    async discoverExtensions() {
-        const extensionsDir = path.join(this.tarsConfig.homeDir, '.gemini', 'extensions');
-        if (!fs.existsSync(extensionsDir))
-            return [];
-        const extensions = [];
-        try {
-            const entries = fs.readdirSync(extensionsDir, { withFileTypes: true });
-            for (const entry of entries) {
-                if (!entry.isDirectory() && !entry.isSymbolicLink())
-                    continue;
-                const extPath = path.resolve(extensionsDir, entry.name);
-                const configPath = path.join(extPath, 'gemini-extension.json');
-                if (fs.existsSync(configPath)) {
-                    try {
-                        const content = fs.readFileSync(configPath, 'utf-8');
-                        const config = JSON.parse(content);
-                        // Ensure mcpServers are converted to MCPServerConfig instances if they exist
-                        const mcpServers = {};
-                        if (config.mcpServers) {
-                            for (const [name, srv] of Object.entries(config.mcpServers)) {
-                                const s = srv;
-                                // Manually resolve ${extensionPath} since we are constructing configs early
-                                const resolvedArgs = s.args?.map((arg) => arg.replace(/\${extensionPath}/g, extPath));
-                                const resolvedEnv = s.env ? { ...s.env } : {};
-                                for (const key in resolvedEnv) {
-                                    resolvedEnv[key] = resolvedEnv[key].replace(/\${extensionPath}/g, extPath);
-                                }
-                                mcpServers[name] = new MCPServerConfig(s.command, resolvedArgs, resolvedEnv, s.cwd?.replace(/\${extensionPath}/g, extPath), s.url?.replace(/\${extensionPath}/g, extPath), s.httpUrl?.replace(/\${extensionPath}/g, extPath), s.headers, s.tcp, s.type, s.timeout, s.trust);
-                            }
-                        }
-                        extensions.push({
-                            ...config,
-                            id: config.name,
-                            path: extPath,
-                            isActive: true,
-                            mcpServers,
-                            contextFiles: config.contextFiles || []
-                        });
-                        logger.info(`🔌 Found extension: ${config.name}`);
-                    }
-                    catch (e) {
-                        logger.error(`Failed to parse extension at ${extPath}: ${e}`);
-                    }
-                }
-            }
-        }
-        catch (error) {
-            logger.error(`Error during extension discovery: ${error}`);
-        }
-        return extensions;
-    }
-    /**
-     * Executes a prompt and streams events back.
-     */
-    async run(prompt, onEvent, sessionId, attachments, onStatus) {
-        if (!this.initialized) {
-            await this.initialize(sessionId);
-        }
-        const sid = sessionId || this.coreConfig.getSessionId();
-        const savedHome = process.env.HOME;
-        try {
-            process.env.HOME = this.tarsConfig.homeDir;
-            process.env.GEMINI_CLI_HOME = this.tarsConfig.homeDir;
-            // Session Swapping Logic or First Run
-            // We must call startChat at least once to initialize the GeminiChat session,
-            // even if the sessionId matches the coreConfig's initial ID.
-            if (this.currentSessionId !== sid || !this.client.isInitialized()) {
-                logger.debug(`🔄 Initializing/Swapping Gemini session to: ${sid}`);
-                const resumedData = await this.loadResumedSessionData(sid);
-                let history = undefined;
-                if (resumedData && resumedData.conversation) {
-                    history = this.convertRecordToHistory(resumedData.conversation);
-                }
-                // @ts-ignore - access private to swap session
-                await this.client.startChat(history, resumedData || undefined);
-                // Sync: Read back the actual session ID that Core assigned.
-                // Core may create a new session ID internally (e.g. if the project hash
-                // changed or the old session was not found). We must keep Tars's
-                // SessionManager in sync to prevent ID mismatch on next restart.
-                const recordingService = this.client.getChatRecordingService();
-                const actualCoreSessionId = recordingService?.sessionId || this.coreConfig.getSessionId();
-                if (actualCoreSessionId && actualCoreSessionId !== sid) {
-                    logger.warn(`⚠️ Session ID mismatch detected: Tars=${sid}, Core=${actualCoreSessionId}. Syncing to Core's ID.`);
-                    this.currentSessionId = actualCoreSessionId;
-                    // Update SessionManager so the correct ID is persisted to disk
-                    if (this.sessionManager) {
-                        await this.sessionManager.save(actualCoreSessionId);
-                    }
-                }
-                else {
-                    this.currentSessionId = sid;
-                }
-            }
-            let currentRequestParts = [{ text: prompt }];
-            const reqPromptId = `req-${Date.now()}-${Math.random().toString(36).slice(2, 6)}`;
-            // Handle Multimodal Attachments
-            if (attachments && attachments.length > 0) {
-                for (const attachment of attachments) {
-                    try {
-                        const data = fs.readFileSync(attachment.path).toString('base64');
-                        currentRequestParts.push({
-                            inlineData: {
-                                data,
-                                mimeType: attachment.mimeType
-                            }
-                        });
-                        logger.debug(`📎 Attached file to prompt: ${attachment.path} (${attachment.mimeType})`);
-                    }
-                    catch (err) {
-                        logger.error(`Failed to read attachment ${attachment.path}: ${err.message}`);
-                    }
-                }
-            }
-            let turnCount = 0;
-            const maxTurns = 100; // Increased to handle complex autonomous tasks
-            const abortController = new AbortController();
-            let finalUsageStats = undefined;
-            // Accumulate usage across multi-turn interactions.
-            // For local models, each turn reports only its own tokens. For Gemini Cloud,
-            // promptTokenCount is cumulative (total context). We keep the maximum
-            // promptTokenCount seen (last turn = largest context) and sum outputTokens.
-            let accumulatedInputTokens = 0;
-            let accumulatedOutputTokens = 0;
-            let accumulatedCachedTokens = 0;
-            let loopDetected = false;
-            let hasRealContent = false;
-            // Track recent tool calls for live status updates
-            const recentTools = [];
-            while (turnCount < maxTurns) {
-                turnCount++;
-                const toolRequests = [];
-                let stream;
-                let retryCount = 0;
-                const maxRetries = 8;
-                let lastError = null;
-                while (retryCount < maxRetries) {
-                    try {
-                        const estimatedTokens = Math.max(100, accumulatedInputTokens);
-                        const waitTime = this.rateLimiter.checkWaitTime(estimatedTokens);
-                        if (waitTime > 0) {
-                            logger.info(`⏳ Pre-emptive throttling: waiting ${Math.round(waitTime / 1000)}s to avoid rate limits...`);
-                            await new Promise((resolve) => setTimeout(resolve, waitTime));
-                        }
-                        this.rateLimiter.recordRequest(estimatedTokens);
-                        stream = await promptIdContext.run(sid, () => {
-                            // Combine manual abort signal with a 5-minute timeout to prevent deadlock
-                            const timeoutSignal = AbortSignal.timeout(5 * 60 * 1000);
-                            const combinedSignal = abortController.signal.aborted
-                                ? abortController.signal
-                                : timeoutSignal;
-                            // We listen for manual aborts to also abort the combined signal
-                            // (AbortSignal.any is available in Node 20+, but we can just pass timeoutSignal
-                            // and manual aborts are rare here, or use AbortSignal.any if supported)
-                            const signalToUse = typeof AbortSignal.any === 'function'
-                                ? AbortSignal.any([abortController.signal, timeoutSignal])
-                                : timeoutSignal;
-                            return this.client.sendMessageStream(currentRequestParts, signalToUse, reqPromptId);
-                        });
-                        break; // Success
-                    }
-                    catch (error) {
-                        retryCount++;
-                        lastError = error;
-                        const isTransient = error.message?.includes('429') ||
-                            error.message?.includes('503') ||
-                            error.message?.toLowerCase().includes('rate limit') ||
-                            error.message?.toLowerCase().includes('capacity') ||
-                            error.message?.toLowerCase().includes('quota') ||
-                            error.message?.toLowerCase().includes('overloaded');
-                        if (isTransient && retryCount < maxRetries) {
-                            const delay = Math.pow(2, retryCount) * 1000 + Math.random() * 1000;
-                            logger.warn(`⚠️ Gemini API transient error (attempt ${retryCount}/${maxRetries}): ${error.message}. Retrying in ${Math.round(delay)}ms...`);
-                            await new Promise((resolve) => setTimeout(resolve, delay));
-                            continue;
-                        }
-                        // Fallback logic for 'auto' model on permanent error or final retry
-                        if (this.tarsConfig.geminiModel === 'auto' &&
-                            !this.initializedWithFallback) {
-                            logger.warn(`🔄 'auto' model failed with error: ${error.message}. Attempting fallback to gemini-2.0-flash...`);
-                            this.initializedWithFallback = true;
-                            // @ts-ignore - modifying private config for fallback
-                            this.coreConfig.model = 'gemini-2.0-flash';
-                            // Re-initialize client with new model
-                            this.client = this.coreConfig.getGeminiClient();
-                            this.applyClientOverrides(this.client);
-                            await this.client.initialize();
-                            retryCount = 0; // Reset retries for the fallback model
-                            continue;
-                        }
-                        throw error;
-                    }
-                }
-                for await (const event of stream) {
-                    if (event.type === GeminiEventType.ToolCallRequest) {
-                        toolRequests.push(event.value);
-                    }
-                    if (event.type === GeminiEventType.LoopDetected) {
-                        loopDetected = true;
-                    }
-                    if (event.type === GeminiEventType.Finished) {
-                        const usage = event.value.usageMetadata;
-                        if (usage) {
-                            // promptTokenCount reflects total context size (cumulative)
-                            // so we always take the latest (highest) value
-                            if (usage.promptTokenCount) {
-                                accumulatedInputTokens = Math.max(accumulatedInputTokens, usage.promptTokenCount);
-                            }
-                            // candidatesTokenCount is per-turn, so we accumulate
-                            if (usage.candidatesTokenCount) {
-                                accumulatedOutputTokens += usage.candidatesTokenCount;
-                            }
-                            if (usage.cachedContentTokenCount) {
-                                accumulatedCachedTokens = Math.max(accumulatedCachedTokens, usage.cachedContentTokenCount);
-                            }
-                            finalUsageStats = usage;
-                        }
-                        continue; // Don't emit done yet
-                    }
-                    const normalized = this.normalizeEvent(event, sid);
-                    if (normalized) {
-                        if (normalized.type === 'text' && normalized.content?.trim()) {
-                            hasRealContent = true;
-                        }
-                        await onEvent(normalized);
-                    }
-                }
-                if (loopDetected) {
-                    logger.warn(`⚠️ Loop detected in Gemini Engine at turn ${turnCount}.`);
-                    break;
-                }
-                if (toolRequests.length === 0) {
-                    logger.debug(`✅ Interaction complete after ${turnCount} turns.`);
-                    break;
-                }
-                // ... (rest of the loop)
-                if (turnCount >= maxTurns) {
-                    logger.warn(`⚠️ Hit maxTurns (${maxTurns}) limit. Force terminating interaction.`);
-                    await onEvent({
-                        type: 'text',
-                        role: 'assistant',
-                        content: '\n\n⚠️ *Task was complex and reached the maximum turn limit. I have executed as much as I could.*',
-                        sessionId: sid
-                    });
-                    break;
-                }
-                // Runtime Safety Filter: Prevent self-destructive commands and unauthorized path access
-                const filteredToolRequests = [];
-                const blockedResponses = new Map();
-                const sensitiveCalls = new Set();
-                for (const req of toolRequests) {
-                    const toolName = req.name;
-                    const args = req.args || {};
-                    const commandLine = args.CommandLine || args.command || '';
-                    const filePath = args.file_path || args.path || args.dir_path || '';
-                    // 1. Block self-destructive commands
-                    if ((toolName.includes('run_command') ||
-                        toolName.includes('run_shell_command')) &&
-                        (commandLine.includes('tars stop') ||
-                            /\bpm2\s+(stop|kill|delete)\b/.test(commandLine))) {
-                        logger.warn(`🛑 INTERCEPTED self-destructive command: ${commandLine}`);
-                        await onEvent({
-                            type: 'text',
-                            role: 'assistant',
-                            content: `\n\n⚠️ **Safety Interruption**: I attempted to run a command that would stop my own supervisor process (${commandLine}). To prevent a loss of connection or state, I have blocked this action. If you really want me to stop, please run \`tars stop\` manually in your terminal.`,
-                            sessionId: sid
-                        });
-                        blockedResponses.set(req.callId, 'Execution blocked: Self-destructive command detected.');
-                        continue;
-                    }
-                    // 2. Block blacklisted path access
-                    if (filePath && DLPService.isPathBlacklisted(filePath)) {
-                        logger.warn(`🛑 INTERCEPTED unauthorized path access: ${filePath}`);
-                        await onEvent({
-                            type: 'text',
-                            role: 'assistant',
-                            content: `\n\n⚠️ **Security Interruption**: I attempted to access a protected file or directory (${filePath}). Access to this path is restricted by the Tars Data Loss Prevention (DLP) policy.`,
-                            sessionId: sid
-                        });
-                        blockedResponses.set(req.callId, `Access to ${filePath} is restricted by DLP policy.`);
-                        continue;
-                    }
-                    // 3. Mark sensitive paths for aggressive scrubbing
-                    if ((filePath && DLPService.isSensitivePath(filePath)) ||
-                        (commandLine && DLPService.isSensitivePath(commandLine))) {
-                        sensitiveCalls.add(req.callId);
-                    }
-                    filteredToolRequests.push(req);
-                }
-                // Execute tools using Scheduler
-                let completedCalls = [];
-                if (filteredToolRequests.length > 0) {
-                    logger.debug(`🛠️ Executing ${filteredToolRequests.length} tool calls...`);
-                    const scheduler = new Scheduler({
-                        context: this.coreConfig,
-                        messageBus: this.coreConfig.getMessageBus(),
-                        getPreferredEditor: () => undefined,
-                        schedulerId: sid
-                    });
-                    completedCalls = await scheduler.schedule(filteredToolRequests, abortController.signal);
-                    // Emit tool responses so the Supervisor can log them
-                    for (const call of completedCalls) {
-                        const normalized = this.normalizeEvent({
-                            type: GeminiEventType.ToolCallResponse,
-                            value: call
-                        }, sid);
-                        if (normalized)
-                            await onEvent(normalized);
-                    }
-                    // Record results in chat recording service for persistence/memory
-                    const model = this.tarsConfig.geminiModel;
-                    this.client.getChat().recordCompletedToolCalls(model, completedCalls);
-                }
-                // Build tool status for live progress updates
-                const turnToolStatuses = [];
-                for (const call of completedCalls) {
-                    const req = toolRequests.find((r) => r.callId === call.request?.callId || r.callId === call.callId);
-                    if (!req)
-                        continue;
-                    const part = call.response?.responseParts?.find((p) => 'functionResponse' in p);
-                    const response = part?.functionResponse?.response;
-                    const responseStr = typeof response === 'string'
-                        ? response
-                        : typeof response === 'object'
-                            ? JSON.stringify(response)
-                            : String(response || '');
-                    turnToolStatuses.push({
-                        name: req.name,
-                        responsePreview: responseStr.substring(0, 120),
-                        responseSize: responseStr.length
-                    });
-                }
-                // Also include blocked tools
-                for (const [callId, reason] of blockedResponses) {
-                    const req = toolRequests.find((r) => r.callId === callId);
-                    if (req) {
-                        turnToolStatuses.push({
-                            name: req.name,
-                            responsePreview: `⛔ ${reason}`,
-                            responseSize: reason.length
-                        });
-                    }
-                }
-                recentTools.push(...turnToolStatuses);
-                const isMilestone = turnCount > 0 && turnCount % 20 === 0;
-                // Fire status update after each tool batch or on a milestone
-                if (onStatus && (turnToolStatuses.length > 0 || isMilestone)) {
-                    if (isMilestone) {
-                        logger.info(`[GeminiEngine] Milestone ${turnCount} — firing status update...`);
-                    }
-                    try {
-                        await onStatus(turnCount, recentTools.slice(-10), isMilestone);
-                    }
-                    catch (e) {
-                        logger.warn(`[GeminiEngine] Status update failed: ${e.message}`);
-                    }
-                }
-                // Prepare next request with tool results (Scrubbed via DLP and mapped back to 1:1)
-                currentRequestParts = GeminiEngine.buildToolResponseParts(toolRequests, completedCalls, blockedResponses, sensitiveCalls, this.tarsConfig.inferenceBackend === 'llamacpp');
-                if (currentRequestParts.length === 0) {
-                    logger.warn('⚠️ No tool responses generated after execution.');
-                    break;
-                }
-                // -------------------------------------------------------------------------
-                // PROACTIVE COMPRESSION & USER CHECK-IN
-                // -------------------------------------------------------------------------
-                // 1. Check if we need to compress mid-loop to avoid context window crashes
-                if (this.sessionManager && this.tarsConfig) {
-                    const stats = this.sessionManager.getStats();
-                    const threshold = this.tarsConfig.compressionThreshold || 0.8;
-                    const limit = this.tarsConfig.contextWindowTokens || 128000;
-                    if (stats && stats.lastInputTokens > limit * threshold) {
-                        logger.info(`[GeminiEngine] Mid-loop compression triggered (${stats.lastInputTokens}/${limit} tokens)`);
-                        try {
-                            const didCompress = await this.compressSession();
-                            if (didCompress) {
-                                await onEvent({
-                                    type: 'text',
-                                    role: 'assistant',
-                                    content: '\n\n✨ *Mid-task memory compacted to optimally save context space.*',
-                                    sessionId: sid
-                                });
-                            }
-                        }
-                        catch (e) {
-                            logger.warn(`[GeminiEngine] Mid-loop compression failed: ${e.message}`);
-                        }
-                    }
-                }
-                // 2. Max turns safety (handled by loop condition, but we break here if needed)
-            }
-            // If the loop finished without producing any content, notify the user
-            if (!hasRealContent) {
-                let fallbackMsg = '\n\n⚠️ **Model Interaction Issue**: The Gemini model failed to produce a valid text response.';
-                if (loopDetected) {
-                    fallbackMsg +=
-                        ' A repetitive output loop was detected and terminated. This can sometimes happen with complex prompts or transient API glitches.';
-                }
-                fallbackMsg += '\n\nPlease try rephrasing your request or starting a new session.';
-                await onEvent({
-                    type: 'text',
-                    role: 'assistant',
-                    content: fallbackMsg,
-                    sessionId: sid
-                });
-            }
-            // Always emit final done event when exiting the loop
-            await onEvent({
-                type: 'done',
-                usageStats: accumulatedInputTokens > 0 || accumulatedOutputTokens > 0
-                    ? {
-                        inputTokens: accumulatedInputTokens,
-                        outputTokens: accumulatedOutputTokens,
-                        cachedTokens: accumulatedCachedTokens
-                    }
-                    : finalUsageStats
-                        ? {
-                            inputTokens: finalUsageStats.promptTokenCount || 0,
-                            outputTokens: finalUsageStats.candidatesTokenCount || 0,
-                            cachedTokens: finalUsageStats.cachedContentTokenCount || 0
-                        }
-                        : undefined,
-                sessionId: sid
-            });
-        }
-        catch (error) {
-            const errorMsg = error.message ||
-                (typeof error === 'object' ? JSON.stringify(error) : String(error));
-            logger.error(`❌ Gemini Engine run error: ${errorMsg}`);
-            await onEvent({ type: 'error', error: errorMsg });
-            throw error;
-        }
-        finally {
-            process.env.HOME = savedHome;
-        }
-    }
-    /**
-     * Synchronous-style run for background tasks.
-     */
-    async runSync(prompt, sessionId) {
-        let fullContent = '';
-        await this.run(prompt, (event) => {
-            if (event.content && event.role === 'assistant') {
-                fullContent += event.content;
-            }
-        }, sessionId);
-        return fullContent;
-    }
-    /**
-     * Rebuilds the response array for Gemini ensuring 1:1 parity with function calls.
-     */
-    static buildToolResponseParts(toolRequests, completedCalls, blockedResponses, sensitiveCalls, isLocalInference = false) {
-        return toolRequests
-            .map((req) => {
-            const callId = req.callId;
-            if (blockedResponses.has(callId)) {
-                return {
-                    functionResponse: {
-                        name: req.name,
-                        response: { error: blockedResponses.get(callId) }
-                    }
-                };
-            }
-            const completedCall = completedCalls.find((c) => (c.request?.callId || c.callId) === callId);
-            if (!completedCall)
-                return null;
-            const part = completedCall.response?.responseParts?.find((p) => 'functionResponse' in p);
-            if (part?.functionResponse?.response) {
-                let scrubbedResponse = DLPService.scrubDeep(part.functionResponse.response);
-                if (sensitiveCalls.has(callId) &&
-                    typeof scrubbedResponse === 'object' &&
-                    scrubbedResponse !== null) {
-                    if (scrubbedResponse.content &&
-                        typeof scrubbedResponse.content === 'string') {
-                        scrubbedResponse.content = DLPService.scrubEnvContent(scrubbedResponse.content);
-                    }
-                    if (scrubbedResponse.stdout &&
-                        typeof scrubbedResponse.stdout === 'string') {
-                        scrubbedResponse.stdout = DLPService.scrubEnvContent(scrubbedResponse.stdout);
-                    }
-                }
-                part.functionResponse.response = scrubbedResponse;
-            }
-            // Inject the original callId so LlamaCppGenerator can map it to tool_call_id.
-            // We ONLY do this for local inference as the Cloud Gemini API rejects unknown fields.
-            if (isLocalInference) {
-                part.id = callId;
-            }
-            return part;
-        })
-            .filter(Boolean);
-    }
-    /**
-     * Proactively compress the session history to reclaim context window space.
-     * Delegates to Gemini Core's built-in compression.
-     */
-    async compressSession(force = false) {
-        if (!this.initialized || !this.client)
-            return false;
-        const sid = this.currentSessionId || 'unknown';
-        logger.info(`🗜️ Triggering session compression (force=${force})...`);
-        try {
-            if (this.tarsConfig.inferenceBackend === 'llamacpp') {
-                const history = this.client.getHistory();
-                // We keep the most recent ~60% and ensure the boundary lands on a 'user' role
-                // to maintain proper turn alternation (no orphaned tool responses).
-                if (history && history.length > 20) {
-                    const keepCount = Math.ceil(history.length * 0.6);
-                    let cutIndex = history.length - keepCount;
-                    // Walk forward to find a 'user' role entry for clean boundary
-                    while (cutIndex < history.length && history[cutIndex]?.role !== 'user') {
-                        cutIndex++;
-                    }
-                    if (cutIndex < history.length) {
-                        const historyToCompress = history.slice(0, cutIndex);
-                        const tail = history.slice(cutIndex);
-                        logger.info(`🗜️ Local inference compaction: Summarizing oldest ${historyToCompress.length} turns...`);
-                        // Use the local generator non-streamed to summarize the truncated chunk
-                        const generator = this.coreConfig.contentGenerator;
-                        const hasPreviousSnapshot = historyToCompress.some((c) => c.parts?.some((p) => p.text?.includes('<state_snapshot>')));
-                        const anchorInstruction = hasPreviousSnapshot
-                            ? 'A previous <state_snapshot> exists in the history. You MUST integrate all still-relevant information from that snapshot into the new one, updating it with the more recent events.'
-                            : 'Generate a new <state_snapshot> based on the provided history.';
-                        const summaryPrompt = `${anchorInstruction}\nExtract all important constraints, configs, details and tool results from this chunk of history. Format your response cleanly.`;
-                        let summaryContent = '';
-                        try {
-                            const response = await generator.generateContent({
-                                model: this.tarsConfig.geminiModel,
-                                contents: [
-                                    ...historyToCompress,
-                                    { role: 'user', parts: [{ text: summaryPrompt }] }
-                                ]
-                            }, sid);
-                            summaryContent =
-                                response?.candidates?.[0]?.content?.parts?.[0]?.text || '';
-                        }
-                        catch (err) {
-                            logger.warn(`Semantic compression inference failed: ${err.message}`);
-                        }
-                        if (!summaryContent) {
-                            summaryContent =
-                                '*(Summary generation failed, falling back to raw truncation)*';
-                        }
-                        const newHistory = [
-                            {
-                                role: 'user',
-                                parts: [
-                                    {
-                                        text: `<state_snapshot>\n${summaryContent.trim()}\n</state_snapshot>`
-                                    }
-                                ]
-                            },
-                            {
-                                role: 'model',
-                                parts: [
-                                    { text: 'Got it. I will keep this historical context in mind.' }
-                                ]
-                            },
-                            ...tail
-                        ];
-                        this.client.setHistory(newHistory);
-                        logger.info(`🗜️ Local inference context compacted: retained tail of ${tail.length} turns + snapshot.`);
-                        return true;
-                    }
-                }
-                return false;
-            }
-            const result = await this.client.tryCompressChat(sid, force);
-            logger.info(`🗜️ Compression result: status=${result.compressionStatus}, ` +
-                `${result.originalTokenCount} → ${result.newTokenCount} tokens`);
-            return String(result.compressionStatus) === 'COMPRESSED';
-        }
-        catch (e) {
-            logger.warn(`⚠️ Compression failed: ${e.message}`);
-            return false;
-        }
-    }
-    /**
-     * Refreshes the system instruction in-place without destroying the session.
-     * Used after memory mutations so the model sees updated facts.
-     */
-    refreshSystemInstruction() {
-        if (!this.initialized || !this.client)
-            return;
-        this.client.updateSystemInstruction();
-        logger.debug('🔄 System instruction refreshed in-place');
-    }
-    /**
-     * Maps native core events to Tars-compatible event format.
-     */
-    normalizeEvent(event, sessionId) {
-        switch (event.type) {
-            case GeminiEventType.Content:
-                return {
-                    type: 'text',
-                    role: 'assistant',
-                    content: event.value,
-                    sessionId
-                };
-            case GeminiEventType.Thought:
-                // ThoughtSummary has subject and description
-                const thoughtText = event.value.subject
-                    ? `**${event.value.subject}** ${event.value.description}`
-                    : event.value.description;
-                return {
-                    type: 'thought',
-                    content: DLPService.scrub(thoughtText),
-                    sessionId
-                };
-            case GeminiEventType.ToolCallRequest:
-                return {
-                    type: 'tool_call',
-                    toolName: event.value.name,
-                    toolArgs: DLPService.scrubDeep(event.value.args),
-                    callId: event.value.callId,
-                    sessionId
-                };
-            case GeminiEventType.ToolCallResponse:
-                // resultDisplay can be string | FileDiff | AnsiOutput | TodoList
-                // Support both ToolCallResponseInfo and CompletedToolCall payloads
-                const val = event.value;
-                const callInfo = val.response ? val.response : val;
-                const display = callInfo.resultDisplay;
-                let content = '';
-                if (typeof display === 'string') {
-                    content = display;
-                }
-                else if (display) {
-                    content = JSON.stringify(display);
-                }
-                else if (callInfo.error) {
-                    content = callInfo.error.message;
-                }
-                return {
-                    type: 'tool_response',
-                    toolName: val.request?.callId || callInfo.callId,
-                    content: DLPService.scrub(content),
-                    sessionId
-                };
-            case GeminiEventType.Finished:
-                return {
-                    type: 'done',
-                    usageStats: event.value.usageMetadata
-                        ? {
-                            inputTokens: event.value.usageMetadata.promptTokenCount || 0,
-                            outputTokens: event.value.usageMetadata.candidatesTokenCount || 0,
-                            cachedTokens: event.value.usageMetadata.cachedContentTokenCount || 0
-                        }
-                        : undefined,
-                    sessionId
-                };
-            case GeminiEventType.Error:
-                let errorDetails = '';
-                if (event.value instanceof Error) {
-                    errorDetails = event.value.message;
-                }
-                else if (typeof event.value === 'object' && event.value !== null) {
-                    // Try to extract nested error message if it exists (common in Google API errors)
-                    const val = event.value;
-                    errorDetails = val.message || val.error?.message || JSON.stringify(event.value);
-                }
-                else {
-                    errorDetails = String(event.value);
-                }
-                return {
-                    type: 'error',
-                    error: errorDetails,
-                    sessionId
-                };
-            case GeminiEventType.LoopDetected:
-                return {
-                    type: 'loop_detected',
-                    sessionId
-                };
-            case GeminiEventType.ChatCompressed: {
-                const info = event.value;
-                if (info && info.compressionStatus === CompressionStatus.COMPRESSED) {
-                    return {
-                        type: 'compressed',
-                        content: `🗜️ Session compressed: ${info.originalTokenCount.toLocaleString()} → ${info.newTokenCount.toLocaleString()} tokens`,
-                        sessionId
-                    };
-                }
-                return null;
-            }
-            case GeminiEventType.ContextWindowWillOverflow:
-                logger.warn(`⚠️ Context window near overflow: ${event.value.estimatedRequestTokenCount} tokens, ${event.value.remainingTokenCount} remaining`);
-                return {
-                    type: 'context_warning',
-                    content: `⚠️ Context window near capacity (${event.value.remainingTokenCount.toLocaleString()} tokens remaining)`,
-                    sessionId
-                };
-            case GeminiEventType.MaxSessionTurns:
-                return {
-                    type: 'max_turns',
-                    content: '⚠️ Maximum session turns reached. Consider compressing the session.',
-                    sessionId
-                };
-            default:
-                return null;
-        }
-    }
-    /**
-     * Attempts to find and load session history from the Core's history directory.
-     */
-    async loadResumedSessionData(sessionId) {
-        try {
-            // Core history is usually in ~/.gemini/tmp/<hash>/chats/
-            // But we isolated HOME to ~/.tars, so it's in ~/.tars/.gemini/...
-            const projectRoot = this.tarsConfig.homeDir;
-            const geminiDir = path.join(this.tarsConfig.homeDir, '.gemini');
-            const tmpDir = path.join(geminiDir, 'tmp');
-            if (!fs.existsSync(tmpDir))
-                return null;
-            // 1. Try to find the exact project identifier from projects.json
-            let projectIdentifier = null;
-            const registryPath = path.join(geminiDir, 'projects.json');
-            if (fs.existsSync(registryPath)) {
-                try {
-                    const registry = JSON.parse(fs.readFileSync(registryPath, 'utf-8'));
-                    projectIdentifier = registry.projects[projectRoot] || null;
-                }
-                catch (e) {
-                    logger.warn(`⚠️ Failed to read projects.json: ${e}`);
-                }
-            }
-            // 2. Fallback: MD5 hash (used in some versions)
-            if (!projectIdentifier) {
-                const crypto = await import('node:crypto');
-                projectIdentifier = crypto.createHash('md5').update(projectRoot).digest('hex');
-            }
-            // 3. Search for the session file in candidate directories
-            // We search projectIdentifier first, then scan all if not found
-            const searchDirs = [projectIdentifier];
-            try {
-                const allDirs = fs.readdirSync(tmpDir);
-                for (const d of allDirs) {
-                    if (d !== projectIdentifier)
-                        searchDirs.push(d);
-                }
-            }
-            catch (e) { }
-            const shortId = sessionId.slice(0, 8);
-            for (const dir of searchDirs) {
-                if (!dir)
-                    continue;
-                const chatsDir = path.join(tmpDir, dir, 'chats');
-                if (!fs.existsSync(chatsDir))
-                    continue;
-                const files = fs.readdirSync(chatsDir);
-                const sessionFile = files.find((f) => f.includes(`-${shortId}.json`));
-                if (sessionFile) {
-                    const filePath = path.join(chatsDir, sessionFile);
-                    const content = await loadConversationRecord(filePath);
-                    logger.info(`📂 Resumed session from exact match: ${sessionFile}`);
-                    return {
-                        conversation: content,
-                        filePath
-                    };
-                }
-                // Fallback: If no exact session ID match, use the most recently
-                // modified chat file. This prevents a blank cold start when the
-                // Tars session ID has drifted from Core's internal session ID.
-                // We also check for .jsonl files used in newer Core versions.
-                const jsonFiles = files.filter((f) => f.endsWith('.json') || f.endsWith('.jsonl'));
-                if (jsonFiles.length > 0) {
-                    const sorted = jsonFiles
-                        .map((f) => ({
-                        name: f,
-                        mtime: fs.statSync(path.join(chatsDir, f)).mtimeMs
-                    }))
-                        .sort((a, b) => b.mtime - a.mtime);
-                    const latestFile = sorted[0].name;
-                    logger.warn(`⚠️ No exact session match for ${shortId}. Falling back to latest: ${latestFile}`);
-                    const filePath = path.join(chatsDir, latestFile);
-                    const content = await loadConversationRecord(filePath);
-                    return {
-                        conversation: content,
-                        filePath
-                    };
-                }
-            }
-            return null;
-        }
-        catch (e) {
-            logger.warn(`⚠️ Failed to load resumed session data: ${e}`);
-            return null;
-        }
-    }
-    /**
-     * Converts a ConversationRecord from Core to Gemini API Content[] history.
-     */
-    convertRecordToHistory(conversation) {
-        const history = [];
-        if (!conversation || !conversation.messages)
-            return history;
-        for (const msg of conversation.messages) {
-            if (msg.type === 'user') {
-                let parts = [];
-                if (typeof msg.content === 'string') {
-                    parts = [{ text: msg.content }];
-                }
-                else if (Array.isArray(msg.content)) {
-                    parts = msg.content;
-                }
-                history.push({ role: 'user', parts });
-            }
-            else if (msg.type === 'gemini') {
-                let parts = [];
-                if (typeof msg.content === 'string' && msg.content !== '') {
-                    parts.push({ text: msg.content });
-                }
-                else if (Array.isArray(msg.content)) {
-                    parts.push(...msg.content);
-                }
-                // Add function calls if any
-                const functionResponseParts = [];
-                if (msg.toolCalls) {
-                    for (const tc of msg.toolCalls) {
-                        parts.push({
-                            functionCall: {
-                                name: tc.name,
-                                args: tc.args
-                            }
-                        });
-                        // If the tool call has a result, we need to add a function response
-                        if (tc.status === 'done' || tc.result) {
-                            let responseObj = tc.result;
-                            if (typeof responseObj === 'string') {
-                                try {
-                                    responseObj = JSON.parse(responseObj);
-                                }
-                                catch (e) {
-                                    responseObj = { result: responseObj };
-                                }
-                            }
-                            else if (responseObj === null || responseObj === undefined) {
-                                responseObj = { result: 'Success' };
-                            }
-                            functionResponseParts.push({
-                                functionResponse: {
-                                    name: tc.name,
-                                    response: responseObj
-                                }
-                            });
-                        }
-                    }
-                }
-                history.push({ role: 'model', parts });
-                // If we generated function response parts, they belong to the NEXT turn as 'user'
-                if (functionResponseParts.length > 0) {
-                    history.push({ role: 'user', parts: functionResponseParts });
-                }
-            }
-        }
-        return history;
-    }
-    /**
-     * Applies runtime overrides to the Gemini client to ensure smooth operation
-     * in specific environments (e.g., local inference).
-     */
-    applyClientOverrides(client) {
-        if (this.tarsConfig.inferenceBackend === 'llamacpp') {
-            // The loop detector runs concurrently in the background and causes 400 crashes
-            // for local-only setups that don't have a valid Google API key.
-            const loopService = client.getLoopDetectionService();
-            if (loopService) {
-                logger.debug('🔇 Silencing LoopDetectionService for local inference...');
-                loopService.queryLoopDetectionModel = async () => {
-                    logger.debug('Background loop verification skipped (Local Mode).');
-                    return null;
-                };
-            }
-        }
-    }
-}
-//# sourceMappingURL=gemini-engine.js.map