grov 0.5.2 → 0.5.3

package/dist/proxy/server.js

@@ -1,314 +1,26 @@
 // Grov Proxy Server - Fastify + undici
 // Intercepts Claude Code <-> Anthropic API traffic for drift detection and context injection
 import Fastify from 'fastify';
-import { config } from './config.js';
+import { config, buildSafeHeaders } from './config.js';
 import { forwardToAnthropic, isForwardError } from './forwarder.js';
+import { extendedCache, evictOldestCacheEntry, checkExtendedCache, log } from './extended-cache.js';
+import { setDebugMode, getNextRequestId, taskLog, proxyLog, logTokenUsage } from './utils/logging.js';
+import { detectKeyDecision, extractTextContent, extractProjectPath, extractGoalFromMessages, extractConversationHistory } from './utils/extractors.js';
+import { appendToLastUserMessage, injectIntoRawBody } from './injection/injectors.js';
+import { preProcessRequest, setPendingPlanClear } from './handlers/preprocess.js';
 import { parseToolUseBlocks, extractTokenUsage } from './action-parser.js';
-import { createSessionState, getSessionState, updateSessionState, createStep, updateTokenCount, logDriftEvent, getRecentSteps, getValidatedSteps, updateSessionMode, markWaitingForRecovery, incrementEscalation, updateLastChecked, markCleared, getActiveSessionForUser, deleteSessionState, deleteStepsForSession, updateRecentStepsReasoning, markSessionCompleted, getCompletedSessionForProject, cleanupOldCompletedSessions, getKeyDecisions, getEditedFiles, } from '../lib/store.js';
-import { smartTruncate } from '../lib/utils.js';
+import { createSessionState, getSessionState, updateSessionState, createStep, updateTokenCount, logDriftEvent, getRecentSteps, getValidatedSteps, updateSessionMode, markWaitingForRecovery, incrementEscalation, updateLastChecked, getActiveSessionForUser, deleteSessionState, deleteStepsForSession, updateRecentStepsReasoning, markSessionCompleted, getCompletedSessionForProject, cleanupOldCompletedSessions, cleanupStaleActiveSessions, } from '../lib/store.js';
 import { checkDrift, scoreToCorrectionLevel, shouldSkipSteps, isDriftCheckAvailable, checkRecoveryAlignment, generateForcedRecovery, } from '../lib/drift-checker-proxy.js';
 import { buildCorrection, formatCorrectionForInjection } from '../lib/correction-builder-proxy.js';
 import { generateSessionSummary, isSummaryAvailable, extractIntent, isIntentExtractionAvailable, analyzeTaskContext, isTaskAnalysisAvailable, } from '../lib/llm-extractor.js';
-import { buildTeamMemoryContext, extractFilesFromMessages } from './request-processor.js';
 import { saveToTeamMemory } from './response-processor.js';
 import { randomUUID } from 'crypto';
-import * as fs from 'fs';
-import * as path from 'path';
 // Store last drift result for recovery alignment check
 const lastDriftResults = new Map();
+// Server logger reference (set in startServer)
+let serverLog = null;
 // Track last messageCount per session to detect retries vs new turns
 const lastMessageCount = new Map();
-// Cache injection content per session (MUST be identical across requests for cache preservation)
-// Stored in memory because session DB state doesn't exist on first request
-const cachedInjections = new Map();
-const sessionInjectionTracking = new Map();
-function getOrCreateTracking(sessionId) {
-    if (!sessionInjectionTracking.has(sessionId)) {
-        sessionInjectionTracking.set(sessionId, {
-            files: new Set(),
-            decisionIds: new Set(),
-            reasonings: new Set(),
-        });
-    }
-    return sessionInjectionTracking.get(sessionId);
-}
-/**
- * Build dynamic injection content for user message (DELTA only)
- * Includes: edited files, key decisions, drift correction, forced recovery
- * Only injects NEW content that hasn't been injected before
- */
-function buildDynamicInjection(sessionId, sessionState, logger) {
-    const tracking = getOrCreateTracking(sessionId);
-    const parts = [];
-    const debugInfo = {};
-    // 1. Get edited files (delta - not already injected)
-    const allEditedFiles = getEditedFiles(sessionId);
-    const newFiles = allEditedFiles.filter(f => !tracking.files.has(f));
-    debugInfo.totalEditedFiles = allEditedFiles.length;
-    debugInfo.newEditedFiles = newFiles.length;
-    debugInfo.alreadyTrackedFiles = tracking.files.size;
-    if (newFiles.length > 0) {
-        // Track and add to injection
-        newFiles.forEach(f => tracking.files.add(f));
-        const fileNames = newFiles.slice(0, 5).map(f => f.split('/').pop());
-        parts.push(`[EDITED: ${fileNames.join(', ')}]`);
-        debugInfo.editedFilesInjected = fileNames;
-    }
-    // 2. Get key decisions with reasoning (delta - not already injected)
-    const keyDecisions = getKeyDecisions(sessionId, 5);
-    debugInfo.totalKeyDecisions = keyDecisions.length;
-    debugInfo.alreadyTrackedDecisions = tracking.decisionIds.size;
-    const newDecisions = keyDecisions.filter(d => !tracking.decisionIds.has(d.id) &&
-        d.reasoning &&
-        !tracking.reasonings.has(d.reasoning));
-    debugInfo.newKeyDecisions = newDecisions.length;
-    for (const decision of newDecisions.slice(0, 3)) {
-        tracking.decisionIds.add(decision.id);
-        tracking.reasonings.add(decision.reasoning);
-        const truncated = smartTruncate(decision.reasoning, 120);
-        parts.push(`[DECISION: ${truncated}]`);
-        // Log the original and truncated reasoning for debugging
-        if (logger) {
-            logger.info({
-                msg: 'Key decision reasoning extracted',
-                originalLength: decision.reasoning.length,
-                truncatedLength: truncated.length,
-                original: decision.reasoning.substring(0, 200) + (decision.reasoning.length > 200 ? '...' : ''),
-                truncated,
-            });
-        }
-    }
-    debugInfo.decisionsInjected = newDecisions.slice(0, 3).length;
-    // 3. Add drift correction if pending
-    if (sessionState?.pending_correction) {
-        parts.push(`[DRIFT: ${sessionState.pending_correction}]`);
-        debugInfo.hasDriftCorrection = true;
-        debugInfo.driftCorrectionLength = sessionState.pending_correction.length;
-    }
-    // 4. Add forced recovery if pending
-    if (sessionState?.pending_forced_recovery) {
-        parts.push(`[RECOVERY: ${sessionState.pending_forced_recovery}]`);
-        debugInfo.hasForcedRecovery = true;
-        debugInfo.forcedRecoveryLength = sessionState.pending_forced_recovery.length;
-    }
-    // Log debug info
-    if (logger) {
-        logger.info({
-            msg: 'Dynamic injection build details',
-            ...debugInfo,
-            partsCount: parts.length,
-        });
-    }
-    if (parts.length === 0) {
-        return null;
-    }
-    const injection = '---\n[GROV CONTEXT]\n' + parts.join('\n');
-    // Log final injection content
-    if (logger) {
-        logger.info({
-            msg: 'Dynamic injection content',
-            size: injection.length,
-            content: injection,
-        });
-    }
-    return injection;
-}
-/**
- * Append dynamic injection to the last user message in raw body string
- * This preserves cache for system + previous messages, only the last user msg changes
- */
-function appendToLastUserMessage(rawBody, injection) {
-    // Find the last occurrence of "role":"user" followed by content
-    // We need to find the content field of the last user message and append to it
-    // Strategy: Find all user messages, get the last one, append to its content
-    // This is tricky because content can be string or array
-    // Simpler approach: Find the last user message's closing content
-    // Look for pattern: "role":"user","content":"..." or "role":"user","content":[...]
-    // Find last "role":"user"
-    const userRolePattern = /"role"\s*:\s*"user"/g;
-    let lastUserMatch = null;
-    let match;
-    while ((match = userRolePattern.exec(rawBody)) !== null) {
-        lastUserMatch = match;
-    }
-    if (!lastUserMatch) {
-        // No user message found, can't inject
-        return rawBody;
-    }
-    // From lastUserMatch position, find the content field
-    const afterRole = rawBody.slice(lastUserMatch.index);
-    // Find "content" field after role
-    const contentMatch = afterRole.match(/"content"\s*:\s*/);
-    if (!contentMatch || contentMatch.index === undefined) {
-        return rawBody;
-    }
-    const contentStartGlobal = lastUserMatch.index + contentMatch.index + contentMatch[0].length;
-    const afterContent = rawBody.slice(contentStartGlobal);
-    // Determine if content is string or array
-    if (afterContent.startsWith('"')) {
-        // String content - find closing quote (handling escapes)
-        let i = 1; // Skip opening quote
-        while (i < afterContent.length) {
-            if (afterContent[i] === '\\') {
-                i += 2; // Skip escaped char
-            }
-            else if (afterContent[i] === '"') {
-                // Found closing quote
-                const insertPos = contentStartGlobal + i;
-                // Insert before closing quote, escape the injection for JSON
-                const escapedInjection = injection
-                    .replace(/\\/g, '\\\\')
-                    .replace(/"/g, '\\"')
-                    .replace(/\n/g, '\\n');
-                return rawBody.slice(0, insertPos) + '\\n\\n' + escapedInjection + rawBody.slice(insertPos);
-            }
-            else {
-                i++;
-            }
-        }
-    }
-    else if (afterContent.startsWith('[')) {
-        // Array content - find last text block and append, or add new text block
-        // Find the closing ] of the content array
-        let depth = 1;
-        let i = 1;
-        while (i < afterContent.length && depth > 0) {
-            const char = afterContent[i];
-            if (char === '[')
-                depth++;
-            else if (char === ']')
-                depth--;
-            else if (char === '"') {
-                // Skip string
-                i++;
-                while (i < afterContent.length && afterContent[i] !== '"') {
-                    if (afterContent[i] === '\\')
-                        i++;
-                    i++;
-                }
-            }
-            i++;
-        }
-        if (depth === 0) {
-            // Found closing bracket at position i-1
-            const insertPos = contentStartGlobal + i - 1;
-            // Add new text block before closing bracket
-            const escapedInjection = injection
-                .replace(/\\/g, '\\\\')
-                .replace(/"/g, '\\"')
-                .replace(/\n/g, '\\n');
-            const newBlock = `,{"type":"text","text":"\\n\\n${escapedInjection}"}`;
-            return rawBody.slice(0, insertPos) + newBlock + rawBody.slice(insertPos);
-        }
-    }
-    // Fallback: couldn't parse, return unchanged
-    return rawBody;
-}
-// ============================================
-// DEBUG MODE - Controlled via --debug flag
-// ============================================
-let debugMode = false;
-export function setDebugMode(enabled) {
-    debugMode = enabled;
-}
-// ============================================
-// FILE LOGGER - Request/Response tracking (debug only)
-// ============================================
-const PROXY_LOG_PATH = path.join(process.cwd(), 'grov-proxy.log');
-let requestCounter = 0;
-function proxyLog(entry) {
-    if (!debugMode)
-        return; // Skip file logging unless --debug flag
-    const logEntry = {
-        timestamp: new Date().toISOString(),
-        ...entry,
-    };
-    const line = JSON.stringify(logEntry) + '\n';
-    fs.appendFileSync(PROXY_LOG_PATH, line);
-}
-/**
- * Log token usage to console (always shown, compact format)
- */
-function logTokenUsage(requestId, usage, latencyMs) {
-    const total = usage.cacheCreation + usage.cacheRead;
-    const hitRatio = total > 0 ? ((usage.cacheRead / total) * 100).toFixed(0) : '0';
-    console.log(`[${requestId}] ${hitRatio}% cache | in:${usage.inputTokens} out:${usage.outputTokens} | create:${usage.cacheCreation} read:${usage.cacheRead} | ${latencyMs}ms`);
-}
-/**
- * Helper to append text to system prompt (handles string or array format)
- */
-function appendToSystemPrompt(body, textToAppend) {
-    if (typeof body.system === 'string') {
-        body.system = body.system + textToAppend;
-    }
-    else if (Array.isArray(body.system)) {
-        // Append as new text block WITHOUT cache_control
-        // Anthropic allows max 4 cache blocks - Claude Code already uses 2+
-        // Grov's injections are small (~2KB) so uncached is fine
-        body.system.push({
-            type: 'text',
-            text: textToAppend,
-        });
-    }
-    else {
-        // No system prompt yet, create as string
-        body.system = textToAppend;
-    }
-}
-/**
- * Get system prompt as string (for reading)
- */
-function getSystemPromptText(body) {
-    if (typeof body.system === 'string') {
-        return body.system;
-    }
-    else if (Array.isArray(body.system)) {
-        return body.system
-            .filter(block => block.type === 'text')
-            .map(block => block.text)
-            .join('\n');
-    }
-    return '';
-}
-/**
- * Inject text into raw body string WITHOUT re-serializing
- * This preserves the original formatting/whitespace for cache compatibility
- *
- * Adds a new text block to the end of the system array
- */
-function injectIntoRawBody(rawBody, injectionText) {
-    // Find the system array in the raw JSON
-    // Pattern: "system": [....]
-    const systemMatch = rawBody.match(/"system"\s*:\s*\[/);
-    if (!systemMatch || systemMatch.index === undefined) {
-        return { modified: rawBody, success: false };
-    }
-    // Find the matching closing bracket for the system array
-    const startIndex = systemMatch.index + systemMatch[0].length;
-    let bracketCount = 1;
-    let endIndex = startIndex;
-    for (let i = startIndex; i < rawBody.length && bracketCount > 0; i++) {
-        const char = rawBody[i];
-        if (char === '[')
-            bracketCount++;
-        else if (char === ']')
-            bracketCount--;
-        if (bracketCount === 0) {
-            endIndex = i;
-            break;
-        }
-    }
-    if (bracketCount !== 0) {
-        return { modified: rawBody, success: false };
-    }
-    // Escape the injection text for JSON
-    const escapedText = JSON.stringify(injectionText).slice(1, -1); // Remove outer quotes
-    // Create the new block (without cache_control - will be cache_creation)
-    const newBlock = `,{"type":"text","text":"${escapedText}"}`;
-    // Insert before the closing bracket
-    const modified = rawBody.slice(0, endIndex) + newBlock + rawBody.slice(endIndex);
-    return { modified, success: true };
-}
 // Session tracking (in-memory for active sessions)
 const activeSessions = new Map();
 /**
@@ -381,7 +93,7 @@ async function handleMessages(request, reply) {
         promptCount: sessionInfo.promptCount,
         projectPath: sessionInfo.projectPath,
     });
-    const currentRequestId = ++requestCounter;
+    const currentRequestId = getNextRequestId();
     logger.info({
         msg: 'Incoming request',
         sessionId: sessionInfo.sessionId.substring(0, 8),
@@ -405,7 +117,7 @@ async function handleMessages(request, reply) {
     // Process request to get injection text
     // __grovInjection = team memory (system prompt, cached)
     // __grovUserMsgInjection = dynamic content (user message, delta only)
-    const processedBody = await preProcessRequest(request.body, sessionInfo, logger);
+    const processedBody = await preProcessRequest(request.body, sessionInfo, logger, detectRequestType);
     const systemInjection = processedBody.__grovInjection;
     const userMsgInjection = processedBody.__grovUserMsgInjection;
     // Get raw body bytes
@@ -470,7 +182,12 @@ async function handleMessages(request, reply) {
         // FIRE-AND-FORGET: Don't block response to Claude Code
         // This prevents retry loops caused by Haiku calls adding latency
         if (result.statusCode === 200 && isAnthropicResponse(result.body)) {
-            postProcessResponse(result.body, sessionInfo, request.body, logger)
+            // Prepare extended cache data (only if enabled)
+            const extendedCacheData = config.EXTENDED_CACHE_ENABLED ? {
+                headers: buildSafeHeaders(request.headers),
+                rawBody: typeof finalBodyToSend === 'string' ? Buffer.from(finalBodyToSend, 'utf-8') : finalBodyToSend,
+            } : undefined;
+            postProcessResponse(result.body, sessionInfo, request.body, logger, extendedCacheData)
                 .catch(err => console.error('[GROV] postProcess error:', err));
         }
         const latency = Date.now() - startTime;
@@ -591,6 +308,7 @@ async function getOrCreateSession(request, logger) {
         projectPath,
     };
     activeSessions.set(tempSessionId, sessionInfo);
+    // Note: team memory is now GLOBAL (not per session), no propagation needed
     logger.info({ msg: 'No existing session, will create after task analysis' });
     return { ...sessionInfo, isNew: true, currentSession: null, completedSession };
 }
@@ -600,10 +318,10 @@ async function getOrCreateSession(request, logger) {
  * - continuation: tool result (messageCount changed, last msg has tool_result)
  * - retry: same messageCount as before
  */
-function detectRequestType(messages, sessionId) {
+function detectRequestType(messages, projectPath) {
     const currentCount = messages?.length || 0;
-    const lastCount = lastMessageCount.get(sessionId);
-    lastMessageCount.set(sessionId, currentCount);
+    const lastCount = lastMessageCount.get(projectPath);
+    lastMessageCount.set(projectPath, currentCount);
     // Same messageCount = retry
     if (lastCount !== undefined && currentCount === lastCount) {
         return 'retry';
@@ -623,103 +341,6 @@ function detectRequestType(messages, sessionId) {
     }
     return 'first';
 }
-/**
- * Pre-process request before forwarding
- * - Context injection (first request only)
- * - CLEAR operation (first request only)
- * - Drift correction (first request only)
- *
- * SKIP all injections on: retry, continuation
- */
-async function preProcessRequest(body, sessionInfo, logger) {
-    const modified = { ...body };
-    // Detect request type: first, continuation, or retry
-    const requestType = detectRequestType(modified.messages || [], sessionInfo.sessionId);
-    // === NEW ARCHITECTURE: Separate static and dynamic injection ===
-    //
-    // STATIC (system prompt, cached):
-    //   - Team memory from PAST sessions only
-    //   - CLEAR summary when triggered
-    //   -> Uses __grovInjection + injectIntoRawBody()
-    //
-    // DYNAMIC (user message, delta only):
-    //   - Files edited in current session
-    //   - Key decisions with reasoning
-    //   - Drift correction, forced recovery
-    //   -> Uses __grovUserMsgInjection + appendToLastUserMessage()
-    // Get session state
-    const sessionState = getSessionState(sessionInfo.sessionId);
-    // === CLEAR MODE (100% threshold) ===
-    // If token count exceeds threshold AND we have a pre-computed summary, apply CLEAR
-    if (sessionState) {
-        const currentTokenCount = sessionState.token_count || 0;
-        if (currentTokenCount > config.TOKEN_CLEAR_THRESHOLD &&
-            sessionState.pending_clear_summary) {
-            logger.info({
-                msg: 'CLEAR MODE ACTIVATED - resetting conversation',
-                tokenCount: currentTokenCount,
-                threshold: config.TOKEN_CLEAR_THRESHOLD,
-                summaryLength: sessionState.pending_clear_summary.length,
-            });
-            // 1. Empty messages array (fundamental reset)
-            modified.messages = [];
-            // 2. Inject summary into system prompt (this will cause cache miss - intentional)
-            appendToSystemPrompt(modified, sessionState.pending_clear_summary);
-            // 3. Mark session as cleared
-            markCleared(sessionInfo.sessionId);
-            // 4. Clear pending summary and invalidate team memory cache (new baseline)
-            updateSessionState(sessionInfo.sessionId, { pending_clear_summary: undefined });
-            cachedInjections.delete(sessionInfo.sessionId);
-            // 5. Clear tracking (fresh start after CLEAR)
-            sessionInjectionTracking.delete(sessionInfo.sessionId);
-            logger.info({ msg: 'CLEAR complete - conversation reset with summary' });
-            return modified; // Skip other injections - this is a complete reset
-        }
-    }
-    // === STATIC INJECTION: Team memory (PAST sessions only) ===
-    // Cached per session - identical across all requests for cache preservation
-    const cachedTeamMemory = cachedInjections.get(sessionInfo.sessionId);
-    if (cachedTeamMemory) {
-        // Reuse cached team memory (constant for this session)
-        modified.__grovInjection = cachedTeamMemory;
-        modified.__grovInjectionCached = true;
-        logger.info({ msg: 'Using cached team memory', size: cachedTeamMemory.length });
-    }
-    else {
-        // First request: compute team memory from PAST sessions only
-        const mentionedFiles = extractFilesFromMessages(modified.messages || []);
-        // Pass currentSessionId to exclude current session data
-        const teamContext = buildTeamMemoryContext(sessionInfo.projectPath, mentionedFiles, sessionInfo.sessionId // Exclude current session
-        );
-        if (teamContext) {
-            modified.__grovInjection = teamContext;
-            modified.__grovInjectionCached = false;
-            // Cache for future requests (stays constant)
-            cachedInjections.set(sessionInfo.sessionId, teamContext);
-            logger.info({ msg: 'Computed and cached team memory', size: teamContext.length });
-        }
-    }
-    // SKIP dynamic injection for retries and continuations
-    if (requestType !== 'first') {
-        return modified;
-    }
-    // === DYNAMIC INJECTION: User message (delta only) ===
-    // Includes: edited files, key decisions, drift correction, forced recovery
-    // This goes into the LAST user message, not system prompt
-    const dynamicInjection = buildDynamicInjection(sessionInfo.sessionId, sessionState, logger);
-    if (dynamicInjection) {
-        modified.__grovUserMsgInjection = dynamicInjection;
-        logger.info({ msg: 'Dynamic injection ready for user message', size: dynamicInjection.length });
-        // Clear pending corrections after building injection
-        if (sessionState?.pending_correction || sessionState?.pending_forced_recovery) {
-            updateSessionState(sessionInfo.sessionId, {
-                pending_correction: undefined,
-                pending_forced_recovery: undefined,
-            });
-        }
-    }
-    return modified;
-}
 /**
  * Post-process response after receiving from Anthropic
  * - Task orchestration (new/continue/subtask/complete)
@@ -730,7 +351,7 @@ async function preProcessRequest(body, sessionInfo, logger) {
  * - Recovery alignment check (Section 4.4)
  * - Team memory triggers (Section 4.6)
  */
-async function postProcessResponse(response, sessionInfo, requestBody, logger) {
+async function postProcessResponse(response, sessionInfo, requestBody, logger, extendedCacheData) {
     // Parse tool_use blocks
     const actions = parseToolUseBlocks(response);
     // Extract text content for analysis
@@ -752,6 +373,29 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
     if (isWarmup) {
         return;
     }
+    // === EXTENDED CACHE: Capture for keep-alive ===
+    // Only capture on end_turn (user idle starts now, not during tool_use loops)
+    if (isEndTurn && extendedCacheData) {
+        const rawStr = extendedCacheData.rawBody.toString('utf-8');
+        const hasSystem = rawStr.includes('"system"');
+        const hasTools = rawStr.includes('"tools"');
+        const hasCacheCtrl = rawStr.includes('"cache_control"');
+        const msgMatch = rawStr.match(/"messages"\s*:\s*\[/);
+        const msgPos = msgMatch?.index ?? -1;
+        // Use projectPath as key (one entry per conversation, not per task)
+        const cacheKey = sessionInfo.projectPath;
+        // Evict oldest if at capacity (only for NEW entries, not updates)
+        if (!extendedCache.has(cacheKey)) {
+            evictOldestCacheEntry();
+        }
+        extendedCache.set(cacheKey, {
+            headers: extendedCacheData.headers,
+            rawBody: extendedCacheData.rawBody,
+            timestamp: Date.now(),
+            keepAliveCount: 0,
+        });
+        log(`Extended cache: CAPTURE project=${cacheKey.split('/').pop()} size=${rawStr.length} sys=${hasSystem} tools=${hasTools} cache_ctrl=${hasCacheCtrl} msg_pos=${msgPos}`);
+    }
     // If not end_turn (tool_use in progress), skip task orchestration but keep session
     if (!isEndTurn) {
         // Use existing session or create minimal one without LLM calls
@@ -774,23 +418,44 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
                 promptCount: 1,
                 projectPath: sessionInfo.projectPath,
             });
+            // Note: team memory is now GLOBAL (not per session), no propagation needed
         }
     }
     else if (isTaskAnalysisAvailable()) {
         // Use completed session for comparison if no active session
         const sessionForComparison = sessionInfo.currentSession || sessionInfo.completedSession;
+        // Extract conversation history for context-aware task analysis
+        const conversationHistory = extractConversationHistory(requestBody.messages || []);
         try {
-            const taskAnalysis = await analyzeTaskContext(sessionForComparison, latestUserMessage, recentSteps, textContent);
+            const taskAnalysis = await analyzeTaskContext(sessionForComparison, latestUserMessage, recentSteps, textContent, conversationHistory);
             logger.info({
                 msg: 'Task analysis',
                 action: taskAnalysis.action,
-                topic_match: taskAnalysis.topic_match,
+                task_type: taskAnalysis.task_type,
                 goal: taskAnalysis.current_goal?.substring(0, 50),
                 reasoning: taskAnalysis.reasoning,
             });
+            // TASK LOG: Analysis result
+            taskLog('TASK_ANALYSIS', {
+                sessionId: sessionInfo.sessionId,
+                action: taskAnalysis.action,
+                task_type: taskAnalysis.task_type,
+                goal: taskAnalysis.current_goal || '',
+                reasoning: taskAnalysis.reasoning || '',
+                userMessage: latestUserMessage.substring(0, 80),
+                hasCurrentSession: !!sessionInfo.currentSession,
+                hasCompletedSession: !!sessionInfo.completedSession,
+            });
             // Update recent steps with reasoning (backfill from end_turn response)
             if (taskAnalysis.step_reasoning && activeSessionId) {
                 const updatedCount = updateRecentStepsReasoning(activeSessionId, taskAnalysis.step_reasoning);
+                // TASK LOG: Step reasoning update
+                taskLog('STEP_REASONING', {
+                    sessionId: activeSessionId,
+                    stepsUpdated: updatedCount,
+                    reasoningEntries: Object.keys(taskAnalysis.step_reasoning).length,
+                    stepIds: Object.keys(taskAnalysis.step_reasoning).join(','),
+                });
             }
             // Handle task orchestration based on analysis
             switch (taskAnalysis.action) {
@@ -809,6 +474,13 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
                             });
                             activeSession.original_goal = taskAnalysis.current_goal;
                         }
+                        // TASK LOG: Continue existing session
+                        taskLog('ORCHESTRATION_CONTINUE', {
+                            sessionId: activeSessionId,
+                            source: 'current_session',
+                            goal: activeSession.original_goal,
+                            goalUpdated: taskAnalysis.current_goal !== activeSession.original_goal,
+                        });
                     }
                     else if (sessionInfo.completedSession) {
                         // Reactivate completed session (user wants to continue/add to it)
@@ -824,6 +496,13 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
                             promptCount: 1,
                             projectPath: sessionInfo.projectPath,
                         });
+                        // Note: team memory is now GLOBAL (not per session), no propagation needed
+                        // TASK LOG: Reactivate completed session
+                        taskLog('ORCHESTRATION_CONTINUE', {
+                            sessionId: activeSessionId,
+                            source: 'reactivated_completed',
+                            goal: activeSession.original_goal,
+                        });
                     }
                     break;
                 case 'new_task': {
@@ -843,9 +522,24 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
                         try {
                             intentData = await extractIntent(latestUserMessage);
                             logger.info({ msg: 'Intent extracted for new task', scopeCount: intentData.expected_scope.length });
+                            // TASK LOG: Intent extraction for new_task
+                            taskLog('INTENT_EXTRACTION', {
+                                sessionId: sessionInfo.sessionId,
+                                context: 'new_task',
+                                goal: intentData.goal,
+                                scopeCount: intentData.expected_scope.length,
+                                scope: intentData.expected_scope.join(', '),
+                                constraints: intentData.constraints.join(', '),
+                                keywords: intentData.keywords.join(', '),
+                            });
                         }
                         catch (err) {
                             logger.info({ msg: 'Intent extraction failed, using basic goal', error: String(err) });
+                            taskLog('INTENT_EXTRACTION_FAILED', {
+                                sessionId: sessionInfo.sessionId,
+                                context: 'new_task',
+                                error: String(err),
+                            });
                         }
                     }
                     const newSessionId = randomUUID();
@@ -865,6 +559,42 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
                         projectPath: sessionInfo.projectPath,
                     });
                     logger.info({ msg: 'Created new task session', sessionId: newSessionId.substring(0, 8) });
+                    // TASK LOG: New task created
+                    taskLog('ORCHESTRATION_NEW_TASK', {
+                        sessionId: newSessionId,
+                        goal: intentData.goal,
+                        scopeCount: intentData.expected_scope.length,
+                        keywordsCount: intentData.keywords.length,
+                    });
+                    // Q&A AUTO-SAVE: If this is an information request with a substantive answer
+                    // AND no tool calls, save immediately since pure Q&A completes in a single turn.
+                    // If there ARE tool calls (e.g., Read for "Analyze X"), wait for them to complete
+                    // so steps get captured properly before saving.
+                    if (taskAnalysis.task_type === 'information' && textContent.length > 100 && actions.length === 0) {
+                        logger.info({ msg: 'Q&A detected (pure text) - saving immediately', sessionId: newSessionId.substring(0, 8) });
+                        taskLog('QA_AUTO_SAVE', {
+                            sessionId: newSessionId,
+                            goal: intentData.goal,
+                            responseLength: textContent.length,
+                            toolCalls: 0,
+                        });
+                        // Store the response for reasoning extraction
+                        updateSessionState(newSessionId, {
+                            final_response: textContent.substring(0, 10000),
+                        });
+                        // Save to team memory and mark complete
+                        await saveToTeamMemory(newSessionId, 'complete');
+                        markSessionCompleted(newSessionId);
+                    }
+                    else if (taskAnalysis.task_type === 'information' && actions.length > 0) {
+                        // Q&A with tool calls - don't auto-save, let it continue until task_complete
+                        logger.info({ msg: 'Q&A with tool calls - waiting for completion', sessionId: newSessionId.substring(0, 8), toolCalls: actions.length });
+                        taskLog('QA_DEFERRED', {
+                            sessionId: newSessionId,
+                            goal: intentData.goal,
+                            toolCalls: actions.length,
+                        });
+                    }
                     break;
                 }
                 case 'subtask': {
@@ -878,8 +608,17 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
                     if (isIntentExtractionAvailable() && latestUserMessage.length > 10) {
                         try {
                             intentData = await extractIntent(latestUserMessage);
+                            taskLog('INTENT_EXTRACTION', {
+                                sessionId: sessionInfo.sessionId,
+                                context: 'subtask',
+                                goal: intentData.goal,
+                                scope: intentData.expected_scope.join(', '),
+                                keywords: intentData.keywords.join(', '),
+                            });
+                        }
+                        catch (err) {
+                            taskLog('INTENT_EXTRACTION_FAILED', { sessionId: sessionInfo.sessionId, context: 'subtask', error: String(err) });
                         }
-                        catch { /* use fallback */ }
                     }
                     const parentId = sessionInfo.currentSession?.session_id || taskAnalysis.parent_task_id;
                     const subtaskId = randomUUID();
@@ -900,6 +639,12 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
                         projectPath: sessionInfo.projectPath,
                     });
                     logger.info({ msg: 'Created subtask session', sessionId: subtaskId.substring(0, 8), parent: parentId?.substring(0, 8) });
+                    // TASK LOG: Subtask created
+                    taskLog('ORCHESTRATION_SUBTASK', {
+                        sessionId: subtaskId,
+                        parentId: parentId || 'none',
+                        goal: intentData.goal,
+                    });
                     break;
                 }
                 case 'parallel_task': {
@@ -913,8 +658,17 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
                     if (isIntentExtractionAvailable() && latestUserMessage.length > 10) {
                         try {
                             intentData = await extractIntent(latestUserMessage);
+                            taskLog('INTENT_EXTRACTION', {
+                                sessionId: sessionInfo.sessionId,
+                                context: 'parallel_task',
+                                goal: intentData.goal,
+                                scope: intentData.expected_scope.join(', '),
+                                keywords: intentData.keywords.join(', '),
+                            });
+                        }
+                        catch (err) {
+                            taskLog('INTENT_EXTRACTION_FAILED', { sessionId: sessionInfo.sessionId, context: 'parallel_task', error: String(err) });
                         }
-                        catch { /* use fallback */ }
                     }
                     const parentId = sessionInfo.currentSession?.session_id || taskAnalysis.parent_task_id;
                     const parallelId = randomUUID();
@@ -935,22 +689,89 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
                         projectPath: sessionInfo.projectPath,
                     });
                     logger.info({ msg: 'Created parallel task session', sessionId: parallelId.substring(0, 8), parent: parentId?.substring(0, 8) });
+                    // TASK LOG: Parallel task created
+                    taskLog('ORCHESTRATION_PARALLEL', {
+                        sessionId: parallelId,
+                        parentId: parentId || 'none',
+                        goal: intentData.goal,
+                    });
                     break;
                 }
                 case 'task_complete': {
                     // Save to team memory and mark as completed (don't delete yet - keep for new_task detection)
                     if (sessionInfo.currentSession) {
                         try {
+                            // Set final_response BEFORE saving so reasoning extraction has the data
+                            updateSessionState(sessionInfo.currentSession.session_id, {
+                                final_response: textContent.substring(0, 10000),
+                            });
                             await saveToTeamMemory(sessionInfo.currentSession.session_id, 'complete');
                             markSessionCompleted(sessionInfo.currentSession.session_id);
                             activeSessions.delete(sessionInfo.currentSession.session_id);
                             lastDriftResults.delete(sessionInfo.currentSession.session_id);
+                            // TASK LOG: Task completed
+                            taskLog('ORCHESTRATION_TASK_COMPLETE', {
+                                sessionId: sessionInfo.currentSession.session_id,
+                                goal: sessionInfo.currentSession.original_goal,
+                            });
+                            // PLANNING COMPLETE: Trigger CLEAR-like reset for implementation phase
+                            // This ensures next request starts fresh with planning context from team memory
+                            if (taskAnalysis.task_type === 'planning' && isSummaryAvailable()) {
+                                try {
+                                    const allSteps = getValidatedSteps(sessionInfo.currentSession.session_id);
+                                    const planSummary = await generateSessionSummary(sessionInfo.currentSession, allSteps, 2000);
+                                    // Store for next request to trigger CLEAR
+                                    setPendingPlanClear({
+                                        projectPath: sessionInfo.projectPath,
+                                        summary: planSummary,
+                                    });
+                                    // Cache invalidation happens in response-processor.ts after syncTask completes
+                                    logger.info({
+                                        msg: 'PLANNING_CLEAR triggered',
+                                        sessionId: sessionInfo.currentSession.session_id.substring(0, 8),
+                                        summaryLen: planSummary.length,
+                                    });
+                                }
+                                catch {
+                                    // Silent fail - planning CLEAR is optional enhancement
+                                }
+                            }
                             logger.info({ msg: 'Task complete - saved to team memory, marked completed' });
                         }
                         catch (err) {
                             logger.info({ msg: 'Failed to save completed task', error: String(err) });
                         }
                     }
+                    else if (textContent.length > 100) {
+                        // NEW: Handle "instant complete" - task that's new AND immediately complete
+                        // This happens for simple Q&A when Haiku says task_complete without existing session
+                        // Example: user asks clarification question, answer is provided in single turn
+                        try {
+                            const newSessionId = randomUUID();
+                            const instantSession = createSessionState({
+                                session_id: newSessionId,
+                                project_path: sessionInfo.projectPath,
+                                original_goal: taskAnalysis.current_goal || latestUserMessage.substring(0, 500),
+                                task_type: 'main',
+                            });
+                            // Set final_response for reasoning extraction
+                            updateSessionState(newSessionId, {
+                                final_response: textContent.substring(0, 10000),
+                            });
+                            await saveToTeamMemory(newSessionId, 'complete');
+                            markSessionCompleted(newSessionId);
+                            logger.info({ msg: 'Instant complete - new task saved immediately', sessionId: newSessionId.substring(0, 8) });
+                            // TASK LOG: Instant complete (new task that finished in one turn)
+                            taskLog('ORCHESTRATION_TASK_COMPLETE', {
+                                sessionId: newSessionId,
+                                goal: taskAnalysis.current_goal || latestUserMessage.substring(0, 80),
+                                source: 'instant_complete',
+                            });
+                        }
+                        catch (err) {
+                            logger.info({ msg: 'Failed to save instant complete task', error: String(err) });
+                        }
+                    }
                     return; // Done, no more processing needed
                 }
                 case 'subtask_complete': {
@@ -969,6 +790,12 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
                                     activeSessionId = parentId;
                                     activeSession = parentSession;
                                     logger.info({ msg: 'Subtask complete - returning to parent', parent: parentId.substring(0, 8) });
+                                    // TASK LOG: Subtask completed
+                                    taskLog('ORCHESTRATION_SUBTASK_COMPLETE', {
+                                        sessionId: sessionInfo.currentSession.session_id,
+                                        parentId: parentId,
+                                        goal: sessionInfo.currentSession.original_goal,
+                                    });
                                 }
                             }
                         }
@@ -993,8 +820,16 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
                 if (isIntentExtractionAvailable() && latestUserMessage.length > 10) {
                     try {
                         intentData = await extractIntent(latestUserMessage);
+                        taskLog('INTENT_EXTRACTION', {
+                            sessionId: sessionInfo.sessionId,
+                            context: 'fallback_analysis_failed',
+                            goal: intentData.goal,
+                            scope: intentData.expected_scope.join(', '),
+                        });
+                    }
+                    catch (err) {
+                        taskLog('INTENT_EXTRACTION_FAILED', { sessionId: sessionInfo.sessionId, context: 'fallback_analysis_failed', error: String(err) });
                     }
-                    catch { /* use fallback */ }
                 }
                 const newSessionId = randomUUID();
                 activeSession = createSessionState({
@@ -1012,6 +847,11 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
     }
     else {
         // No task analysis available - fallback with intent extraction
+        taskLog('TASK_ANALYSIS_UNAVAILABLE', {
+            sessionId: sessionInfo.sessionId,
+            hasCurrentSession: !!sessionInfo.currentSession,
+            userMessage: latestUserMessage.substring(0, 80),
+        });
         if (!sessionInfo.currentSession) {
             let intentData = {
                 goal: latestUserMessage.substring(0, 500),
@@ -1023,8 +863,16 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
                 try {
                     intentData = await extractIntent(latestUserMessage);
                     logger.info({ msg: 'Intent extracted (fallback)', scopeCount: intentData.expected_scope.length });
+                    taskLog('INTENT_EXTRACTION', {
+                        sessionId: sessionInfo.sessionId,
+                        context: 'no_analysis_available',
+                        goal: intentData.goal,
+                        scope: intentData.expected_scope.join(', '),
+                    });
+                }
+                catch (err) {
+                    taskLog('INTENT_EXTRACTION_FAILED', { sessionId: sessionInfo.sessionId, context: 'no_analysis_available', error: String(err) });
                 }
-                catch { /* use fallback */ }
             }
             const newSessionId = randomUUID();
             activeSession = createSessionState({
@@ -1043,19 +891,12 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
             activeSessionId = sessionInfo.currentSession.session_id;
         }
     }
-    // AUTO-SAVE on every end_turn (for all task types: new_task, continue, subtask, parallel)
-    // task_complete and subtask_complete already save and return early, so they won't reach here
-    if (isEndTurn && activeSession && activeSessionId) {
-        try {
-            await saveToTeamMemory(activeSessionId, 'complete');
-            markSessionCompleted(activeSessionId);
-            activeSessions.delete(activeSessionId);
-            logger.info({ msg: 'Auto-saved task on end_turn', sessionId: activeSessionId.substring(0, 8) });
-        }
-        catch (err) {
-            logger.info({ msg: 'Auto-save failed', error: String(err) });
-        }
-    }
+    // NOTE: Auto-save on every end_turn was REMOVED
+    // Task saving is now controlled by Haiku's task analysis:
+    // - task_complete: Haiku detected task is done (Q&A answered, implementation verified, planning confirmed)
+    // - subtask_complete: Haiku detected subtask is done
+    // This ensures we only save when work is actually complete, not on every Claude response.
+    // See analyzeTaskContext() in llm-extractor.ts for the decision logic.
     // Extract token usage
     const usage = extractTokenUsage(response);
     // Use cache metrics as actual context size (cacheCreation + cacheRead)
@@ -1106,18 +947,8 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
         });
     }
     if (actions.length === 0) {
-        // Pure Q&A (no tool calls) - auto-save as task
-        if (isEndTurn && activeSessionId && activeSession) {
-            try {
-                await saveToTeamMemory(activeSessionId, 'complete');
-                markSessionCompleted(activeSessionId);
-                activeSessions.delete(activeSessionId);
-                logger.info({ msg: 'Task saved on final answer', sessionId: activeSessionId.substring(0, 8) });
-            }
-            catch (err) {
-                logger.info({ msg: 'Task save failed', error: String(err) });
-            }
-        }
+        // Final response (no tool calls)
+        // NOTE: Task saving is controlled by Haiku's task analysis (see switch case 'task_complete' above)
         return;
     }
     logger.info({
@@ -1242,20 +1073,36 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
         }
     }
     // Save each action as a step (with reasoning from Claude's text)
+    // When multiple actions come from the same Claude response, they share identical reasoning.
+    // We store reasoning only on the first action and set NULL for subsequent ones to avoid duplication.
+    // At query time, we group steps by reasoning (non-NULL starts a group, NULLs continue it)
+    // and reconstruct the full context: reasoning + all associated files/actions.
+    let previousReasoning = null;
+    logger.info({ msg: 'DEDUP_DEBUG', actionsCount: actions.length, textContentLen: textContent.length });
     for (const action of actions) {
+        const currentReasoning = textContent.substring(0, 1000);
+        const isDuplicate = currentReasoning === previousReasoning;
+        logger.info({
+            msg: 'DEDUP_STEP',
+            actionType: action.actionType,
+            isDuplicate,
+            prevLen: previousReasoning?.length || 0,
+            currLen: currentReasoning.length
+        });
         // Detect key decisions based on action type and reasoning content
-        const isKeyDecision = detectKeyDecision(action, textContent);
+        const isKeyDecision = !isDuplicate && detectKeyDecision(action, textContent);
         createStep({
             session_id: activeSessionId,
             action_type: action.actionType,
             files: action.files,
             folders: action.folders,
             command: action.command,
-            reasoning: textContent.substring(0, 1000), // Claude's explanation (truncated)
+            reasoning: isDuplicate ? undefined : currentReasoning,
             drift_score: driftScore,
             is_validated: !skipSteps,
             is_key_decision: isKeyDecision,
         });
+        previousReasoning = currentReasoning;
         if (isKeyDecision) {
             logger.info({
                 msg: 'Key decision detected',
@@ -1265,138 +1112,6 @@ async function postProcessResponse(response, sessionInfo, requestBody, logger) {
         }
     }
 }
-/**
- * Detect if an action represents a key decision worth injecting later
- * Key decisions are:
- * - Edit/write actions (code modifications)
- * - Actions with decision-related keywords in reasoning
- * - Actions with substantial reasoning content
- */
-function detectKeyDecision(action, reasoning) {
-    // Code modifications are always key decisions
-    if (action.actionType === 'edit' || action.actionType === 'write') {
-        return true;
-    }
-    // Check for decision-related keywords in reasoning
-    const decisionKeywords = [
-        'decision', 'decided', 'chose', 'chosen', 'selected', 'picked',
-        'approach', 'strategy', 'solution', 'implementation',
-        'because', 'reason', 'rationale', 'trade-off', 'tradeoff',
-        'instead of', 'rather than', 'prefer', 'opted',
-        'conclusion', 'determined', 'resolved'
-    ];
-    const reasoningLower = reasoning.toLowerCase();
-    const hasDecisionKeyword = decisionKeywords.some(kw => reasoningLower.includes(kw));
-    // Substantial reasoning (>200 chars) with decision keyword = key decision
-    if (hasDecisionKeyword && reasoning.length > 200) {
-        return true;
-    }
-    return false;
-}
-/**
- * Extract text content from response for analysis
- */
-function extractTextContent(response) {
-    return response.content
-        .filter((block) => block.type === 'text')
-        .map(block => block.text)
-        .join('\n');
-}
-/**
- * Detect task completion from response text
- * Returns trigger type or null
- */
-function detectTaskCompletion(text) {
-    const lowerText = text.toLowerCase();
-    // Strong completion indicators
-    const completionPhrases = [
-        'task is complete',
-        'task complete',
-        'implementation is complete',
-        'implementation complete',
-        'successfully implemented',
-        'all changes have been made',
-        'finished implementing',
-        'completed the implementation',
-        'done with the implementation',
-        'completed all the',
-        'all tests pass',
-        'build succeeds',
-    ];
-    for (const phrase of completionPhrases) {
-        if (lowerText.includes(phrase)) {
-            return 'complete';
-        }
-    }
-    // Subtask completion indicators
-    const subtaskPhrases = [
-        'step complete',
-        'phase complete',
-        'finished this step',
-        'moving on to',
-        'now let\'s',
-        'next step',
-    ];
-    for (const phrase of subtaskPhrases) {
-        if (lowerText.includes(phrase)) {
-            return 'subtask';
-        }
-    }
-    return null;
-}
-/**
- * Extract project path from request body
- */
-function extractProjectPath(body) {
-    // Try to extract from system prompt or messages
-    // Handle both string and array format for system prompt
-    let systemPrompt = '';
-    if (typeof body.system === 'string') {
-        systemPrompt = body.system;
-    }
-    else if (Array.isArray(body.system)) {
-        // New API format: system is array of {type: 'text', text: '...'}
-        systemPrompt = body.system
-            .filter((block) => block && typeof block === 'object' && block.type === 'text' && typeof block.text === 'string')
-            .map(block => block.text)
-            .join('\n');
-    }
-    const cwdMatch = systemPrompt.match(/Working directory:\s*([^\n]+)/);
-    if (cwdMatch) {
-        return cwdMatch[1].trim();
-    }
-    return null;
-}
-/**
- * Extract goal from FIRST user message with text content
- * Skips tool_result blocks, filters out system-reminder tags
- */
-function extractGoalFromMessages(messages) {
-    const userMessages = messages?.filter(m => m.role === 'user') || [];
-    for (const userMsg of userMessages) {
-        let rawContent = '';
-        // Handle string content
-        if (typeof userMsg.content === 'string') {
-            rawContent = userMsg.content;
-        }
-        // Handle array content - look for text blocks (skip tool_result)
-        if (Array.isArray(userMsg.content)) {
-            const textBlocks = userMsg.content
-                .filter((block) => block && typeof block === 'object' && block.type === 'text' && typeof block.text === 'string')
-                .map(block => block.text);
-            rawContent = textBlocks.join('\n');
-        }
-        // Remove <system-reminder>...</system-reminder> tags
-        const cleanContent = rawContent
-            .replace(/<system-reminder>[\s\S]*?<\/system-reminder>/g, '')
-            .trim();
-        // If we found valid text content, return it
-        if (cleanContent && cleanContent.length >= 5) {
-            return cleanContent.substring(0, 500);
-        }
-    }
-    return undefined;
-}
 /**
  * Filter response headers for forwarding to client
  */
@@ -1446,13 +1161,74 @@ export async function startServer(options = {}) {
         console.log('[DEBUG] Logging to grov-proxy.log');
     }
     const server = createServer();
+    // Set server logger for background tasks
+    serverLog = server.log;
     // Cleanup old completed sessions (older than 24 hours)
     cleanupOldCompletedSessions();
+    // Cleanup stale active sessions (no activity for 1 hour)
+    // Prevents old sessions from being reused in fresh Claude sessions
+    const staleCount = cleanupStaleActiveSessions();
+    if (staleCount > 0) {
+        log(`Cleaned up ${staleCount} stale active session(s)`);
+    }
+    // Start extended cache timer if enabled
+    let extendedCacheTimer = null;
+    // Track active connections for graceful shutdown
+    const activeConnections = new Set();
+    let isShuttingDown = false;
+    // Graceful shutdown handler (works with or without extended cache)
+    const gracefulShutdown = () => {
+        if (isShuttingDown)
+            return;
+        isShuttingDown = true;
+        log('Shutdown initiated...');
+        // 1. Stop extended cache timer if running
+        if (extendedCacheTimer) {
+            clearInterval(extendedCacheTimer);
+            extendedCacheTimer = null;
+            log('Extended cache: timer stopped');
+        }
+        // 2. Clear sensitive cache data
+        if (extendedCache.size > 0) {
+            log(`Extended cache: clearing ${extendedCache.size} entries`);
+            for (const entry of extendedCache.values()) {
+                for (const key of Object.keys(entry.headers)) {
+                    entry.headers[key] = '';
+                }
+                entry.rawBody = Buffer.alloc(0);
+            }
+            extendedCache.clear();
+        }
+        // 3. Stop accepting new connections
+        server.close();
+        // 4. Grace period (500ms) then force close remaining connections
+        setTimeout(() => {
+            if (activeConnections.size > 0) {
+                log(`Force closing ${activeConnections.size} connection(s)`);
+                for (const socket of activeConnections) {
+                    socket.destroy();
+                }
+            }
+            log('Goodbye!');
+            process.exit(0);
+        }, 500);
+    };
+    process.on('SIGTERM', gracefulShutdown);
+    process.on('SIGINT', gracefulShutdown);
+    if (config.EXTENDED_CACHE_ENABLED) {
+        extendedCacheTimer = setInterval(checkExtendedCache, 60_000);
+        log('Extended cache: enabled (keep-alive timer started)');
+    }
     try {
         await server.listen({
             host: config.HOST,
             port: config.PORT,
         });
+        // Track connections for graceful shutdown
+        server.server.on('connection', (socket) => {
+            activeConnections.add(socket);
+            socket.on('close', () => activeConnections.delete(socket));
+        });
         console.log(`Grov Proxy: http://${config.HOST}:${config.PORT} -> ${config.ANTHROPIC_BASE_URL}`);
         return server;
     }