mstro-app 0.1.57 → 0.2.0

package/dist/server/cli/improvisation-session-manager.js

@@ -7,10 +7,18 @@
  * For complex multi-part prompts with parallel/sequential movements, use Compose tab instead.
  */
 import { EventEmitter } from 'node:events';
-import { existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs';
+import { existsSync, mkdirSync, readFileSync, rmSync, writeFileSync } from 'node:fs';
 import { join } from 'node:path';
 import { AnalyticsEvents, trackEvent } from '../services/analytics.js';
 import { HeadlessRunner } from './headless/index.js';
+import { assessBestResult, assessContextLoss } from './headless/stall-assessor.js';
+/** Score a run result for best-result tracking (higher = more productive) */
+function scoreRunResult(r) {
+    const toolCount = r.toolUseHistory?.filter(t => t.result !== undefined && !t.isError).length ?? 0;
+    const responseLen = Math.min((r.assistantResponse?.length ?? 0) / 50, 100);
+    const hasThinking = r.thinkingOutput ? 20 : 0;
+    return toolCount * 10 + responseLen + hasThinking;
+}
 export class ImprovisationSessionManager extends EventEmitter {
     sessionId;
     improviseDir;
@@ -27,6 +35,8 @@ export class ImprovisationSessionManager extends EventEmitter {
     accumulatedKnowledge = '';
     /** Whether a prompt is currently executing */
     _isExecuting = false;
+    /** Timestamp when current execution started (for accurate elapsed time across reconnects) */
+    _executionStartTimestamp;
     /** Buffered events during current execution, for replay on reconnect */
     executionEventLog = [];
     /**
@@ -35,10 +45,10 @@ export class ImprovisationSessionManager extends EventEmitter {
      * The first prompt will include context from the historical session.
      */
     static resumeFromHistory(workingDir, historicalSessionId, overrides) {
-        const improviseDir = join(workingDir, '.mstro', 'improvise');
+        const historyDir = join(workingDir, '.mstro', 'history');
         // Extract timestamp from session ID (format: improv-1234567890123 or just 1234567890123)
         const timestamp = historicalSessionId.replace('improv-', '');
-        const historyPath = join(improviseDir, `history-${timestamp}.json`);
+        const historyPath = join(historyDir, `${timestamp}.json`);
         if (!existsSync(historyPath)) {
             throw new Error(`Historical session not found: ${historicalSessionId}`);
         }
@@ -80,9 +90,9 @@ export class ImprovisationSessionManager extends EventEmitter {
             model: options.model,
         };
         this.sessionId = this.options.sessionId;
-        this.improviseDir = join(this.options.workingDir, '.mstro', 'improvise');
-        this.historyPath = join(this.improviseDir, `history-${this.sessionId.replace('improv-', '')}.json`);
-        // Ensure improvise directory exists
+        this.improviseDir = join(this.options.workingDir, '.mstro', 'history');
+        this.historyPath = join(this.improviseDir, `${this.sessionId.replace('improv-', '')}.json`);
+        // Ensure history directory exists
         if (!existsSync(this.improviseDir)) {
             mkdirSync(this.improviseDir, { recursive: true });
         }
@@ -117,24 +127,68 @@ export class ImprovisationSessionManager extends EventEmitter {
         }
     }
     /**
-     * Build prompt with text file attachments prepended
+     * Build prompt with text file attachments prepended and disk path references
      * Format: each text file is shown as @path followed by content in code block
      */
-    buildPromptWithAttachments(userPrompt, attachments) {
-        if (!attachments || attachments.length === 0) {
+    buildPromptWithAttachments(userPrompt, attachments, diskPaths) {
+        if ((!attachments || attachments.length === 0) && (!diskPaths || diskPaths.length === 0)) {
             return userPrompt;
         }
+        const parts = [];
         // Filter to text files only (non-images)
-        const textFiles = attachments.filter(a => !a.isImage);
-        if (textFiles.length === 0) {
+        if (attachments) {
+            const textFiles = attachments.filter(a => !a.isImage);
+            for (const file of textFiles) {
+                parts.push(`@${file.filePath}\n\`\`\`\n${file.content}\n\`\`\``);
+            }
+        }
+        // Add disk path references for all persisted files
+        if (diskPaths && diskPaths.length > 0) {
+            parts.push(`Attached files saved to disk:\n${diskPaths.map(p => `- ${p}`).join('\n')}`);
+        }
+        if (parts.length === 0) {
             return userPrompt;
         }
-        // Build file content blocks
-        const fileBlocks = textFiles.map(file => {
-            return `@${file.filePath}\n\`\`\`\n${file.content}\n\`\`\``;
-        }).join('\n\n');
-        // Prepend file content to user prompt
-        return `${fileBlocks}\n\n${userPrompt}`;
+        return `${parts.join('\n\n')}\n\n${userPrompt}`;
+    }
+    /**
+     * Write attachments to disk at .mstro/tmp/attachments/{sessionId}/
+     * Returns array of absolute file paths for each persisted attachment.
+     */
+    persistAttachments(attachments) {
+        if (attachments.length === 0)
+            return [];
+        const attachDir = join(this.options.workingDir, '.mstro', 'tmp', 'attachments', this.sessionId);
+        if (!existsSync(attachDir)) {
+            mkdirSync(attachDir, { recursive: true });
+        }
+        const paths = [];
+        for (const attachment of attachments) {
+            const filePath = join(attachDir, attachment.fileName);
+            try {
+                // All paste content arrives as base64 — decode to binary
+                writeFileSync(filePath, Buffer.from(attachment.content, 'base64'));
+                paths.push(filePath);
+            }
+            catch (err) {
+                console.error(`Failed to persist attachment ${attachment.fileName}:`, err);
+            }
+        }
+        return paths;
+    }
+    /**
+     * Clean up persisted attachments for this session
+     */
+    cleanupAttachments() {
+        const attachDir = join(this.options.workingDir, '.mstro', 'tmp', 'attachments', this.sessionId);
+        if (existsSync(attachDir)) {
+            try {
+                rmSync(attachDir, { recursive: true, force: true });
+            }
+            catch {
+                // Ignore cleanup errors
+            }
+        }
     }
     /**
      * Execute a user prompt directly (Improvise mode - no score decomposition)
@@ -142,144 +196,79 @@ export class ImprovisationSessionManager extends EventEmitter {
      * Each tab maintains its own claudeSessionId for proper isolation
      * Supports file attachments: text files prepended to prompt, images via stream-json multimodal
      */
-    async executePrompt(userPrompt, attachments) {
+    async executePrompt(userPrompt, attachments, options) {
         const _execStart = Date.now();
-        // Start execution event log for reconnect replay
         this._isExecuting = true;
+        this._executionStartTimestamp = _execStart;
         this.executionEventLog = [];
-        this.emit('onMovementStart', this.history.movements.length + 1, userPrompt);
+        const sequenceNumber = this.history.movements.length + 1;
+        this.emit('onMovementStart', sequenceNumber, userPrompt);
         trackEvent(AnalyticsEvents.IMPROVISE_PROMPT_RECEIVED, {
             prompt_length: userPrompt.length,
             has_attachments: !!(attachments && attachments.length > 0),
             attachment_count: attachments?.length || 0,
             image_attachment_count: attachments?.filter(a => a.isImage).length || 0,
-            sequence_number: this.history.movements.length + 1,
+            sequence_number: sequenceNumber,
             is_resumed_session: this.isResumedSession,
             model: this.options.model || 'default',
         });
         try {
-            const sequenceNumber = this.history.movements.length + 1;
-            // Log the movement start event
             this.executionEventLog.push({
                 type: 'movementStart',
-                data: { sequenceNumber, prompt: userPrompt, timestamp: Date.now() },
+                data: { sequenceNumber, prompt: userPrompt, timestamp: Date.now(), executionStartTimestamp: this._executionStartTimestamp },
                 timestamp: Date.now(),
             });
-            // DEBUG: Removed "Executing prompt..." message - it serves no purpose now that system responds fast
-            // this.queueOutput(`\n🎵 Executing prompt...\n`);
-            // this.flushOutputQueue();
-            // NOTE: Risk analysis removed - now handled by MCP bouncer at tool-use time
-            // The MCP bouncer intercepts ALL tool calls (Bash, Write, etc.) and uses
-            // Mstro's bouncer-integration.ts for AI-powered approval/denial.
-            // This is more effective than analyzing user prompts, which had false positives.
-            // Build prompt with text file attachments prepended
-            const promptWithAttachments = this.buildPromptWithAttachments(userPrompt, attachments);
-            // PERSISTENT SESSION: Use --resume <sessionId> to maintain conversation history per tab
-            // CRITICAL FIX: Using claudeSessionId ensures each tab resumes its own Claude session
-            // Previously used --continue which resumed the most recent global session, causing cross-tab contamination
-            const runner = new HeadlessRunner({
-                workingDir: this.options.workingDir,
-                tokenBudgetThreshold: this.options.tokenBudgetThreshold,
-                maxSessions: this.options.maxSessions,
-                verbose: this.options.verbose,
-                noColor: this.options.noColor,
-                model: this.options.model,
-                improvisationMode: true,
-                movementNumber: sequenceNumber,
-                continueSession: !this.isFirstPrompt, // Used as fallback only if claudeSessionId is missing
-                claudeSessionId: this.claudeSessionId, // Resume specific session for tab isolation
-                outputCallback: (text) => {
-                    this.executionEventLog.push({ type: 'output', data: { text, timestamp: Date.now() }, timestamp: Date.now() });
-                    this.queueOutput(text);
-                    this.flushOutputQueue();
-                },
-                thinkingCallback: (text) => {
-                    this.executionEventLog.push({ type: 'thinking', data: { text }, timestamp: Date.now() });
-                    this.emit('onThinking', text);
-                    this.flushOutputQueue();
-                },
-                toolUseCallback: (event) => {
-                    this.executionEventLog.push({ type: 'toolUse', data: { ...event, timestamp: Date.now() }, timestamp: Date.now() });
-                    this.emit('onToolUse', event);
-                    this.flushOutputQueue();
-                },
-                directPrompt: promptWithAttachments,
-                // Pass image attachments for multimodal handling via stream-json
-                imageAttachments: attachments?.filter(a => a.isImage),
-                // Inject historical context on first prompt of a resumed session
-                // This serves as both the primary context mechanism (no claudeSessionId)
-                // and a fallback if claudeSessionId is stale (client restarted since original session)
-                promptContext: (this.isResumedSession && this.isFirstPrompt)
-                    ? { accumulatedKnowledge: this.buildHistoricalContext(), filesModified: [] }
-                    : undefined
-            });
-            this.currentRunner = runner;
-            const result = await runner.run();
-            this.currentRunner = null;
-            // Capture Claude session ID for future prompts in this tab
-            // This is critical for tab isolation - each tab maintains its own Claude session
-            if (result.claudeSessionId) {
-                this.claudeSessionId = result.claudeSessionId;
-                this.history.claudeSessionId = result.claudeSessionId;
-            }
-            // Mark that we've executed at least one prompt
-            this.isFirstPrompt = false;
-            // Create movement record with accumulated output for persistence
-            const movement = {
-                id: `prompt-${sequenceNumber}`,
-                sequenceNumber,
-                userPrompt,
-                timestamp: new Date().toISOString(),
-                tokensUsed: result.totalTokens,
-                summary: '', // No summary needed - Claude session maintains context
-                filesModified: [],
-                // Persist accumulated output for history replay
-                assistantResponse: result.assistantResponse,
-                thinkingOutput: result.thinkingOutput,
-                toolUseHistory: result.toolUseHistory?.map(t => ({
-                    toolName: t.toolName,
-                    toolId: t.toolId,
-                    toolInput: t.toolInput,
-                    result: t.result,
-                    isError: t.isError,
-                    duration: t.duration
-                })),
-                errorOutput: result.error
+            const { prompt: promptWithAttachments, imageAttachments } = this.preparePromptAndAttachments(userPrompt, attachments);
+            const state = {
+                currentPrompt: promptWithAttachments,
+                retryNumber: 0,
+                checkpointRef: { value: null },
+                contextRecoverySessionId: undefined,
+                freshRecoveryMode: false,
+                accumulatedToolResults: [],
+                contextLost: false,
+                lastWatchdogCheckpoint: null,
+                timedOutTools: [],
+                bestResult: null,
             };
-            // Handle file conflicts if any
-            if (result.conflicts && result.conflicts.length > 0) {
-                this.queueOutput(`\n⚠ File conflicts detected: ${result.conflicts.length}`);
-                result.conflicts.forEach(c => {
-                    this.queueOutput(`  - ${c.filePath} (modified by: ${c.modifiedBy.join(', ')})`);
-                    if (c.backupPath) {
-                        this.queueOutput(`    Backup created: ${c.backupPath}`);
-                    }
-                });
-                this.flushOutputQueue();
+            const maxRetries = 3;
+            let result;
+            // eslint-disable-next-line no-constant-condition
+            while (true) {
+                this.resetIterationState(state);
+                const { useResume, resumeSessionId } = this.determineResumeStrategy(state);
+                const runner = this.createExecutionRunner(state, sequenceNumber, useResume, resumeSessionId, imageAttachments, options?.sandboxed);
+                this.currentRunner = runner;
+                result = await runner.run();
+                this.currentRunner = null;
+                this.updateBestResult(state, result);
+                const nativeTimeouts = result.nativeTimeoutCount ?? 0;
+                this.detectResumeContextLoss(result, state, useResume, maxRetries, nativeTimeouts);
+                await this.detectNativeTimeoutContextLoss(result, state, maxRetries, nativeTimeouts);
+                this.flushPostTimeoutOutput(result, state);
+                if (this.shouldRetryContextLoss(result, state, useResume, nativeTimeouts, maxRetries, promptWithAttachments))
+                    continue;
+                if (this.applyToolTimeoutRetry(state, maxRetries, promptWithAttachments))
+                    continue;
+                break;
             }
-            this.history.movements.push(movement);
-            this.history.totalTokens += movement.tokensUsed;
-            // Save history
-            this.saveHistory();
-            // Completion message is now handled by the client-side movementComplete event handler
-            // This prevents duplicate completion messages (one white, one green)
-            // this.queueOutput(`\n✓ Complete (tokens: ${result.totalTokens.toLocaleString()})\n`);
-            // this.flushOutputQueue();
+            if (state.contextLost)
+                this.claudeSessionId = undefined;
+            result = await this.selectBestResult(state, result, userPrompt);
+            this.captureSessionAndSurfaceErrors(result);
+            this.isFirstPrompt = false;
+            const movement = this.buildMovementRecord(result, userPrompt, sequenceNumber, _execStart);
+            this.handleConflicts(result);
+            this.persistMovement(movement);
             this._isExecuting = false;
+            this._executionStartTimestamp = undefined;
             this.executionEventLog = [];
-            this.emit('onMovementComplete', movement);
-            trackEvent(AnalyticsEvents.IMPROVISE_MOVEMENT_COMPLETED, {
-                tokens_used: movement.tokensUsed,
-                duration_ms: Date.now() - _execStart,
-                sequence_number: sequenceNumber,
-                tool_count: result.toolUseHistory?.length || 0,
-                model: this.options.model || 'default',
-            });
-            this.emit('onSessionUpdate', this.getHistory());
+            this.emitMovementComplete(movement, result, _execStart, sequenceNumber);
             return movement;
         }
         catch (error) {
             this._isExecuting = false;
+            this._executionStartTimestamp = undefined;
             this.executionEventLog = [];
             this.currentRunner = null;
             this.emit('onMovementError', error);
@@ -294,10 +283,375 @@ export class ImprovisationSessionManager extends EventEmitter {
             throw error;
         }
         finally {
-            // Ensure final flush
             this.flushOutputQueue();
         }
     }
+    // ========== Extracted helpers for executePrompt ==========
+    /** Prepare prompt with attachments and limit image count */
+    preparePromptAndAttachments(userPrompt, attachments) {
+        const diskPaths = attachments ? this.persistAttachments(attachments) : [];
+        const prompt = this.buildPromptWithAttachments(userPrompt, attachments, diskPaths);
+        const MAX_IMAGE_ATTACHMENTS = 20;
+        const allImages = attachments?.filter(a => a.isImage);
+        let imageAttachments = allImages;
+        if (allImages && allImages.length > MAX_IMAGE_ATTACHMENTS) {
+            imageAttachments = allImages.slice(-MAX_IMAGE_ATTACHMENTS);
+            this.queueOutput(`\n[[MSTRO_ERROR:TOO_MANY_IMAGES]] ${allImages.length} images attached, limit is ${MAX_IMAGE_ATTACHMENTS}. Using the ${MAX_IMAGE_ATTACHMENTS} most recent.\n`);
+            this.flushOutputQueue();
+        }
+        return { prompt, imageAttachments };
+    }
+    /** Determine whether to use --resume and which session ID */
+    determineResumeStrategy(state) {
+        if (state.freshRecoveryMode) {
+            state.freshRecoveryMode = false;
+            return { useResume: false, resumeSessionId: undefined };
+        }
+        if (state.contextRecoverySessionId) {
+            const id = state.contextRecoverySessionId;
+            state.contextRecoverySessionId = undefined;
+            return { useResume: true, resumeSessionId: id };
+        }
+        if (state.retryNumber === 0) {
+            return { useResume: !this.isFirstPrompt, resumeSessionId: this.claudeSessionId };
+        }
+        if (state.lastWatchdogCheckpoint?.inProgressTools.length === 0 && state.lastWatchdogCheckpoint.claudeSessionId) {
+            return { useResume: true, resumeSessionId: state.lastWatchdogCheckpoint.claudeSessionId };
+        }
+        return { useResume: false, resumeSessionId: undefined };
+    }
+    /** Create HeadlessRunner for one retry iteration */
+    createExecutionRunner(state, sequenceNumber, useResume, resumeSessionId, imageAttachments, sandboxed) {
+        return new HeadlessRunner({
+            workingDir: this.options.workingDir,
+            tokenBudgetThreshold: this.options.tokenBudgetThreshold,
+            maxSessions: this.options.maxSessions,
+            verbose: this.options.verbose,
+            noColor: this.options.noColor,
+            model: this.options.model,
+            improvisationMode: true,
+            movementNumber: sequenceNumber,
+            continueSession: useResume,
+            claudeSessionId: resumeSessionId,
+            outputCallback: (text) => {
+                this.executionEventLog.push({ type: 'output', data: { text, timestamp: Date.now() }, timestamp: Date.now() });
+                this.queueOutput(text);
+                this.flushOutputQueue();
+            },
+            thinkingCallback: (text) => {
+                this.executionEventLog.push({ type: 'thinking', data: { text }, timestamp: Date.now() });
+                this.emit('onThinking', text);
+                this.flushOutputQueue();
+            },
+            toolUseCallback: (event) => {
+                this.executionEventLog.push({ type: 'toolUse', data: { ...event, timestamp: Date.now() }, timestamp: Date.now() });
+                this.emit('onToolUse', event);
+                this.flushOutputQueue();
+            },
+            directPrompt: state.currentPrompt,
+            imageAttachments,
+            promptContext: (state.retryNumber === 0 && this.isResumedSession && this.isFirstPrompt)
+                ? { accumulatedKnowledge: this.buildHistoricalContext(), filesModified: [] }
+                : undefined,
+            onToolTimeout: (checkpoint) => {
+                state.checkpointRef.value = checkpoint;
+            },
+            sandboxed,
+        });
+    }
+    /** Save checkpoint and reset per-iteration state before each retry loop pass. */
+    resetIterationState(state) {
+        if (state.checkpointRef.value)
+            state.lastWatchdogCheckpoint = state.checkpointRef.value;
+        state.checkpointRef.value = null;
+        state.contextLost = false;
+    }
+    /** Update best result tracking */
+    updateBestResult(state, result) {
+        if (!state.bestResult || scoreRunResult(result) > scoreRunResult(state.bestResult)) {
+            state.bestResult = result;
+        }
+    }
+    /** Detect resume context loss (Path 1): session expired on --resume */
+    detectResumeContextLoss(result, state, useResume, maxRetries, nativeTimeouts) {
+        if (!useResume || state.checkpointRef.value || state.retryNumber >= maxRetries || nativeTimeouts > 0) {
+            return;
+        }
+        if (!result.assistantResponse || result.assistantResponse.trim().length === 0) {
+            state.contextLost = true;
+            if (this.options.verbose)
+                console.log('[CONTEXT-RECOVERY] Resume context loss: null/empty response');
+        }
+        else if (result.resumeBufferedOutput !== undefined) {
+            state.contextLost = true;
+            if (this.options.verbose)
+                console.log('[CONTEXT-RECOVERY] Resume context loss: buffer never flushed (no thinking/tools)');
+        }
+        else if ((!result.toolUseHistory || result.toolUseHistory.length === 0) &&
+            !result.thinkingOutput &&
+            result.assistantResponse.length < 500) {
+            state.contextLost = true;
+            if (this.options.verbose)
+                console.log('[CONTEXT-RECOVERY] Resume context loss: no tools, no thinking, short response');
+        }
+    }
+    /** Detect native timeout context loss (Path 2): tool timeouts caused confusion */
+    async detectNativeTimeoutContextLoss(result, state, maxRetries, nativeTimeouts) {
+        if (state.contextLost)
+            return;
+        const toolsWithoutResult = result.toolUseHistory?.filter(t => t.result === undefined).length ?? 0;
+        const effectiveTimeouts = Math.max(nativeTimeouts, toolsWithoutResult);
+        if (effectiveTimeouts === 0 || !result.assistantResponse || state.checkpointRef.value || state.retryNumber >= maxRetries) {
+            return;
+        }
+        const writeToolNames = new Set(['Edit', 'Write', 'MultiEdit', 'NotebookEdit']);
+        const contextLossCtx = {
+            assistantResponse: result.assistantResponse,
+            effectiveTimeouts,
+            nativeTimeoutCount: nativeTimeouts,
+            successfulToolCalls: result.toolUseHistory?.filter(t => t.result !== undefined && !t.isError).length ?? 0,
+            thinkingOutputLength: result.thinkingOutput?.length ?? 0,
+            hasSuccessfulWrite: result.toolUseHistory?.some(t => writeToolNames.has(t.toolName) && t.result !== undefined && !t.isError) ?? false,
+        };
+        const claudeCmd = process.env.CLAUDE_COMMAND || 'claude';
+        const verdict = await assessContextLoss(contextLossCtx, claudeCmd, this.options.verbose);
+        state.contextLost = verdict.contextLost;
+        if (this.options.verbose) {
+            console.log(`[CONTEXT-RECOVERY] Haiku verdict: ${state.contextLost ? 'LOST' : 'OK'} — ${verdict.reason}`);
+        }
+    }
+    /** Flush post-timeout output if context wasn't lost */
+    flushPostTimeoutOutput(result, state) {
+        if (!state.contextLost && result.postTimeoutOutput) {
+            this.queueOutput(result.postTimeoutOutput);
+            this.flushOutputQueue();
+        }
+    }
+    /** Check if context loss recovery should trigger a retry. Returns true if loop should continue. */
+    shouldRetryContextLoss(result, state, useResume, nativeTimeouts, maxRetries, promptWithAttachments) {
+        if (state.checkpointRef.value || state.retryNumber >= maxRetries || !state.contextLost) {
+            return false;
+        }
+        this.accumulateToolResults(result, state);
+        state.retryNumber++;
+        if (useResume && nativeTimeouts === 0) {
+            this.applyInterMovementRecovery(state, promptWithAttachments);
+        }
+        else {
+            this.applyNativeTimeoutRecovery(result, state, promptWithAttachments);
+        }
+        return true;
+    }
+    /** Accumulate completed tool results from a run into the retry state */
+    accumulateToolResults(result, state) {
+        if (!result.toolUseHistory)
+            return;
+        for (const t of result.toolUseHistory) {
+            if (t.result !== undefined) {
+                state.accumulatedToolResults.push({
+                    toolName: t.toolName,
+                    toolId: t.toolId,
+                    toolInput: t.toolInput,
+                    result: t.result,
+                    isError: t.isError,
+                    duration: t.duration,
+                });
+            }
+        }
+    }
+    /** Handle inter-movement context loss recovery (resume session expired) */
+    applyInterMovementRecovery(state, promptWithAttachments) {
+        this.claudeSessionId = undefined;
+        const historicalResults = this.extractHistoricalToolResults();
+        const allResults = [...historicalResults, ...state.accumulatedToolResults];
+        this.emit('onAutoRetry', {
+            retryNumber: state.retryNumber,
+            maxRetries: 3,
+            toolName: 'InterMovementRecovery',
+            completedCount: allResults.length,
+        });
+        this.queueOutput(`\n[[MSTRO_CONTEXT_RECOVERY]] Session context expired — continuing with ${allResults.length} preserved results from prior work (retry ${state.retryNumber}/3).\n`);
+        this.flushOutputQueue();
+        state.freshRecoveryMode = true;
+        state.currentPrompt = this.buildInterMovementRecoveryPrompt(promptWithAttachments, allResults);
+    }
+    /** Handle native-timeout context loss recovery (tool timeouts caused confusion) */
+    applyNativeTimeoutRecovery(result, state, promptWithAttachments) {
+        const completedCount = state.accumulatedToolResults.length;
+        this.emit('onAutoRetry', {
+            retryNumber: state.retryNumber,
+            maxRetries: 3,
+            toolName: 'ContextRecovery',
+            completedCount,
+        });
+        if (result.claudeSessionId && state.retryNumber === 1) {
+            this.queueOutput(`\n[[MSTRO_CONTEXT_RECOVERY]] Context loss detected — resuming session with ${completedCount} preserved results (retry ${state.retryNumber}/3).\n`);
+            this.flushOutputQueue();
+            state.contextRecoverySessionId = result.claudeSessionId;
+            this.claudeSessionId = result.claudeSessionId;
+            state.currentPrompt = this.buildContextRecoveryPrompt(promptWithAttachments);
+        }
+        else {
+            this.queueOutput(`\n[[MSTRO_CONTEXT_RECOVERY]] Continuing with fresh context — ${completedCount} preserved results injected (retry ${state.retryNumber}/3).\n`);
+            this.flushOutputQueue();
+            state.freshRecoveryMode = true;
+            state.currentPrompt = this.buildFreshRecoveryPrompt(promptWithAttachments, state.accumulatedToolResults);
+        }
+    }
+    /** Handle tool timeout checkpoint. Returns true if loop should continue. */
+    applyToolTimeoutRetry(state, maxRetries, promptWithAttachments) {
+        if (!state.checkpointRef.value || state.retryNumber >= maxRetries) {
+            return false;
+        }
+        const cp = state.checkpointRef.value;
+        state.retryNumber++;
+        state.timedOutTools.push({
+            toolName: cp.hungTool.toolName,
+            input: cp.hungTool.input ?? {},
+            timeoutMs: cp.hungTool.timeoutMs,
+        });
+        const canResumeSession = cp.inProgressTools.length === 0 && !!cp.claudeSessionId;
+        this.emit('onAutoRetry', {
+            retryNumber: state.retryNumber,
+            maxRetries,
+            toolName: cp.hungTool.toolName,
+            url: cp.hungTool.url,
+            completedCount: cp.completedTools.length,
+        });
+        trackEvent(AnalyticsEvents.IMPROVISE_AUTO_RETRY, {
+            retry_number: state.retryNumber,
+            hung_tool: cp.hungTool.toolName,
+            hung_url: cp.hungTool.url?.slice(0, 200),
+            completed_tools: cp.completedTools.length,
+            elapsed_ms: cp.elapsedMs,
+            resume_attempted: canResumeSession,
+        });
+        state.currentPrompt = canResumeSession
+            ? this.buildResumeRetryPrompt(cp, state.timedOutTools)
+            : this.buildRetryPrompt(cp, promptWithAttachments, state.timedOutTools);
+        this.queueOutput(`\n[[MSTRO_AUTO_RETRY]] Auto-retry ${state.retryNumber}/${maxRetries}: ${canResumeSession ? 'Resuming session' : 'Continuing'} with ${cp.completedTools.length} successful results, skipping failed ${cp.hungTool.toolName}.\n`);
+        this.flushOutputQueue();
+        return true;
+    }
+    /** Select the best result across retries using Haiku assessment */
+    async selectBestResult(state, result, userPrompt) {
+        if (!state.bestResult || state.bestResult === result || state.retryNumber === 0) {
+            return result;
+        }
+        const claudeCmd = process.env.CLAUDE_COMMAND || 'claude';
+        const bestToolCount = state.bestResult.toolUseHistory?.filter(t => t.result !== undefined && !t.isError).length ?? 0;
+        const currentToolCount = result.toolUseHistory?.filter(t => t.result !== undefined && !t.isError).length ?? 0;
+        try {
+            const verdict = await assessBestResult({
+                originalPrompt: userPrompt,
+                resultA: {
+                    successfulToolCalls: bestToolCount,
+                    responseLength: state.bestResult.assistantResponse?.length ?? 0,
+                    hasThinking: !!state.bestResult.thinkingOutput,
+                    responseTail: (state.bestResult.assistantResponse ?? '').slice(-500),
+                },
+                resultB: {
+                    successfulToolCalls: currentToolCount,
+                    responseLength: result.assistantResponse?.length ?? 0,
+                    hasThinking: !!result.thinkingOutput,
+                    responseTail: (result.assistantResponse ?? '').slice(-500),
+                },
+            }, claudeCmd, this.options.verbose);
+            if (verdict.winner === 'A') {
+                if (this.options.verbose)
+                    console.log(`[BEST-RESULT] Haiku picked earlier attempt: ${verdict.reason}`);
+                return this.mergeResultSessionId(state.bestResult, result.claudeSessionId);
+            }
+            if (this.options.verbose)
+                console.log(`[BEST-RESULT] Haiku picked final attempt: ${verdict.reason}`);
+            return result;
+        }
+        catch {
+            return this.fallbackBestResult(state.bestResult, result);
+        }
+    }
+    /** Fallback best result selection using numeric scoring */
+    fallbackBestResult(bestResult, result) {
+        if (scoreRunResult(bestResult) > scoreRunResult(result)) {
+            if (this.options.verbose) {
+                console.log(`[BEST-RESULT] Haiku unavailable, numeric fallback: earlier attempt (score ${scoreRunResult(bestResult)} vs ${scoreRunResult(result)})`);
+            }
+            return this.mergeResultSessionId(bestResult, result.claudeSessionId);
+        }
+        return result;
+    }
+    /** Replace a result's claudeSessionId with a newer one */
+    mergeResultSessionId(result, sessionId) {
+        if (sessionId)
+            return { ...result, claudeSessionId: sessionId };
+        return result;
+    }
+    /** Capture Claude session ID and surface execution failures */
+    captureSessionAndSurfaceErrors(result) {
+        if (result.claudeSessionId) {
+            this.claudeSessionId = result.claudeSessionId;
+            this.history.claudeSessionId = result.claudeSessionId;
+        }
+        if (!result.completed && result.error) {
+            this.queueOutput(`\n[[MSTRO_ERROR:EXECUTION_FAILED]] ${result.error}\n`);
+            this.flushOutputQueue();
+        }
+    }
+    /** Build a MovementRecord from execution result */
+    buildMovementRecord(result, userPrompt, sequenceNumber, execStart) {
+        return {
+            id: `prompt-${sequenceNumber}`,
+            sequenceNumber,
+            userPrompt,
+            timestamp: new Date().toISOString(),
+            tokensUsed: result.totalTokens,
+            summary: '',
+            filesModified: [],
+            assistantResponse: result.assistantResponse,
+            thinkingOutput: result.thinkingOutput,
+            toolUseHistory: result.toolUseHistory?.map(t => ({
+                toolName: t.toolName,
+                toolId: t.toolId,
+                toolInput: t.toolInput,
+                result: t.result,
+                isError: t.isError,
+                duration: t.duration
+            })),
+            errorOutput: result.error,
+            durationMs: Date.now() - execStart,
+        };
+    }
+    /** Handle file conflicts from execution result */
+    handleConflicts(result) {
+        if (!result.conflicts || result.conflicts.length === 0)
+            return;
+        this.queueOutput(`\n⚠ File conflicts detected: ${result.conflicts.length}`);
+        result.conflicts.forEach(c => {
+            this.queueOutput(`  - ${c.filePath} (modified by: ${c.modifiedBy.join(', ')})`);
+            if (c.backupPath) {
+                this.queueOutput(`    Backup created: ${c.backupPath}`);
+            }
+        });
+        this.flushOutputQueue();
+    }
+    /** Persist movement to history */
+    persistMovement(movement) {
+        this.history.movements.push(movement);
+        this.history.totalTokens += movement.tokensUsed;
+        this.saveHistory();
+    }
+    /** Emit movement completion events and analytics */
+    emitMovementComplete(movement, result, execStart, sequenceNumber) {
+        this.emit('onMovementComplete', movement);
+        trackEvent(AnalyticsEvents.IMPROVISE_MOVEMENT_COMPLETED, {
+            tokens_used: movement.tokensUsed,
+            duration_ms: Date.now() - execStart,
+            sequence_number: sequenceNumber,
+            tool_count: result.toolUseHistory?.length || 0,
+            model: this.options.model || 'default',
+        });
+        this.emit('onSessionUpdate', this.getHistory());
+    }
     /**
      * Build historical context for resuming a session.
      * This creates a summary of the previous conversation that will be injected
@@ -339,6 +693,250 @@ export class ImprovisationSessionManager extends EventEmitter {
         contextParts.push('');
         return contextParts.join('\n');
     }
+    /**
+     * Build a retry prompt from a tool timeout checkpoint.
+     * Injects completed tool results and instructs Claude to skip the failed resource.
+     */
+    buildRetryPrompt(checkpoint, originalPrompt, allTimedOut) {
+        const urlSuffix = checkpoint.hungTool.url ? ` while fetching: ${checkpoint.hungTool.url}` : '';
+        const parts = [
+            '## AUTOMATIC RETRY -- Previous Execution Interrupted',
+            '',
+            `The previous execution was interrupted because ${checkpoint.hungTool.toolName} timed out after ${Math.round(checkpoint.hungTool.timeoutMs / 1000)}s${urlSuffix}.`,
+            '',
+        ];
+        if (allTimedOut && allTimedOut.length > 0) {
+            parts.push(...this.formatTimedOutTools(allTimedOut), '');
+        }
+        else {
+            parts.push('This URL/resource is unreachable. DO NOT retry the same URL or query.', '');
+        }
+        if (checkpoint.completedTools.length > 0) {
+            parts.push(...this.formatCompletedTools(checkpoint.completedTools), '');
+        }
+        if (checkpoint.inProgressTools && checkpoint.inProgressTools.length > 0) {
+            parts.push(...this.formatInProgressTools(checkpoint.inProgressTools), '');
+        }
+        if (checkpoint.assistantText) {
+            const preview = checkpoint.assistantText.length > 8000
+                ? `${checkpoint.assistantText.slice(0, 8000)}...\n(truncated — full response was ${checkpoint.assistantText.length} chars)`
+                : checkpoint.assistantText;
+            parts.push('### Your response before interruption:', preview, '');
+        }
+        parts.push('### Original task (continue from where you left off):');
+        parts.push(originalPrompt);
+        parts.push('');
+        parts.push('INSTRUCTIONS:');
+        parts.push('1. Use the results above -- do not re-fetch content you already have');
+        parts.push('2. Find ALTERNATIVE sources for the content that timed out (different URL, different approach)');
+        parts.push('3. Re-run any in-progress tools that were lost (listed above) if their results are needed');
+        parts.push('4. If no alternative exists, proceed with the results you have and note what was unavailable');
+        return parts.join('\n');
+    }
+    /**
+     * Build a short retry prompt for --resume sessions.
+     * The session already has full conversation context, so we only need to
+     * explain what timed out and instruct Claude to continue.
+     */
+    buildResumeRetryPrompt(checkpoint, allTimedOut) {
+        const parts = [];
+        parts.push(`Your previous ${checkpoint.hungTool.toolName} call timed out after ${Math.round(checkpoint.hungTool.timeoutMs / 1000)}s${checkpoint.hungTool.url ? ` fetching: ${checkpoint.hungTool.url}` : ''}.`);
+        // List all timed-out tools across retries so Claude avoids repeating them
+        if (allTimedOut && allTimedOut.length > 1) {
+            parts.push('');
+            parts.push('All timed-out tools/resources (DO NOT retry any of these):');
+            for (const t of allTimedOut) {
+                const inputSummary = this.summarizeToolInput(t.input);
+                parts.push(`- ${t.toolName}(${inputSummary})`);
+            }
+        }
+        else {
+            parts.push('This URL/resource is unreachable. DO NOT retry the same URL or query.');
+        }
+        parts.push('Continue your task — find an alternative source or proceed with the results you already have.');
+        return parts.join('\n');
+    }
+    // Context loss detection is now handled by assessContextLoss() in stall-assessor.ts
+    // using Haiku assessment instead of brittle regex patterns.
+    /**
+     * Build a recovery prompt for --resume after context loss.
+     * Since we're resuming the same session, Claude has full conversation history
+     * (including all preserved tool results). We just need to redirect it back to the task.
+     */
+    buildContextRecoveryPrompt(originalPrompt) {
+        const parts = [];
+        parts.push('Your previous response indicated you lost context due to tool timeouts, but your full conversation history is preserved — including all successful tool results.');
+        parts.push('');
+        parts.push('Review your conversation history above. You already have results from many successful tool calls. Use those results to continue the task.');
+        parts.push('');
+        parts.push('Original task:');
+        parts.push(originalPrompt);
+        parts.push('');
+        parts.push('INSTRUCTIONS:');
+        parts.push('1. Review your conversation history — all your previous tool results are still available');
+        parts.push('2. Continue from where you left off using the results you already gathered');
+        parts.push('3. If specific tool calls timed out, skip those and work with what you have');
+        parts.push('4. Do NOT start over — build on the work already done');
+        parts.push('5. Do NOT spawn Task subagents for work that previously timed out — do it inline instead');
+        parts.push('6. Prefer multiple small, focused tool calls over single large ones to avoid further timeouts');
+        return parts.join('\n');
+    }
+    /**
+     * Build a recovery prompt for a fresh session (no --resume) after repeated context loss.
+     * Injects all accumulated tool results from previous attempts so Claude can continue
+     * the task without re-fetching data it already gathered.
+     */
+    buildFreshRecoveryPrompt(originalPrompt, toolResults) {
+        const parts = [
+            '## CONTINUING LONG-RUNNING TASK',
+            '',
+            'The previous execution encountered tool timeouts and lost context.',
+            'Below are all results gathered before the interruption. Continue the task using these results.',
+            '',
+        ];
+        parts.push(...this.formatToolResults(toolResults));
+        parts.push('### Original task:');
+        parts.push(originalPrompt);
+        parts.push('');
+        parts.push('INSTRUCTIONS:');
+        parts.push('1. Use the preserved results above \u2014 do NOT re-fetch data you already have');
+        parts.push('2. Continue the task from where it was interrupted');
+        parts.push('3. If you need additional data, fetch it (but try alternative sources if the original timed out)');
+        parts.push('4. Complete the original task fully');
+        parts.push('5. Do NOT spawn Task subagents for work that previously timed out \u2014 do it inline instead');
+        parts.push('6. Prefer multiple small, focused tool calls over single large ones to avoid further timeouts');
+        return parts.join('\n');
+    }
+    /**
+     * Extract tool results from the last N movements in history.
+     * Used for inter-movement recovery to provide context from prior work
+     * when a resume session is corrupted/expired.
+     */
+    extractHistoricalToolResults(maxMovements = 3) {
+        const results = [];
+        const recentMovements = this.history.movements.slice(-maxMovements);
+        for (const movement of recentMovements) {
+            if (!movement.toolUseHistory)
+                continue;
+            for (const tool of movement.toolUseHistory) {
+                if (tool.result !== undefined && !tool.isError) {
+                    results.push({
+                        toolName: tool.toolName,
+                        toolId: tool.toolId,
+                        toolInput: tool.toolInput,
+                        result: tool.result,
+                        isError: tool.isError,
+                        duration: tool.duration,
+                    });
+                }
+            }
+        }
+        return results;
+    }
+    /**
+     * Build a recovery prompt for inter-movement context loss.
+     * The Claude session expired between movements (not due to native timeouts).
+     * Includes prior conversation summary + preserved tool results + anti-timeout guidance.
+     */
+    buildInterMovementRecoveryPrompt(originalPrompt, toolResults) {
+        const parts = [
+            '## SESSION RECOVERY — Prior Session Expired',
+            '',
+            'Your previous session expired between prompts. Below is a summary of the conversation so far and all preserved tool results.',
+            '',
+        ];
+        parts.push(...this.formatConversationHistory(this.history.movements));
+        parts.push(...this.formatToolResults(toolResults));
+        parts.push('### Current user prompt:');
+        parts.push(originalPrompt);
+        parts.push('');
+        parts.push('INSTRUCTIONS:');
+        parts.push('1. Use the preserved results above — do NOT re-fetch data you already have');
+        parts.push('2. Continue the conversation naturally based on the history above');
+        parts.push('3. If you need additional data, fetch it with small focused tool calls');
+        parts.push('4. Do NOT spawn Task subagents — do work inline to avoid further timeouts');
+        parts.push('5. Prefer multiple small, focused tool calls over single large ones');
+        return parts.join('\n');
+    }
+    /** Summarize a tool input for display in retry prompts */
+    summarizeToolInput(input) {
+        if (input.url)
+            return String(input.url).slice(0, 100);
+        if (input.query)
+            return String(input.query).slice(0, 100);
+        if (input.command)
+            return String(input.command).slice(0, 100);
+        if (input.prompt)
+            return String(input.prompt).slice(0, 100);
+        return JSON.stringify(input).slice(0, 100);
+    }
+    /** Format a list of timed-out tools for retry prompts */
+    formatTimedOutTools(tools) {
+        const lines = [];
+        lines.push('### Tools/resources that have timed out (DO NOT retry these):');
+        for (const t of tools) {
+            const inputSummary = this.summarizeToolInput(t.input);
+            lines.push(`- **${t.toolName}**(${inputSummary}) — timed out after ${Math.round(t.timeoutMs / 1000)}s`);
+        }
+        return lines;
+    }
+    /** Format completed checkpoint tools for retry prompts */
+    formatCompletedTools(tools, maxLen = 2000) {
+        const lines = [];
+        lines.push('### Results already obtained:');
+        for (const tool of tools) {
+            const inputSummary = this.summarizeToolInput(tool.input);
+            const preview = tool.result.length > maxLen ? `${tool.result.slice(0, maxLen)}...` : tool.result;
+            lines.push(`- **${tool.toolName}**(${inputSummary}): ${preview}`);
+        }
+        return lines;
+    }
+    /** Format in-progress tools for retry prompts */
+    formatInProgressTools(tools) {
+        const lines = [];
+        lines.push('### Tools that were still running (lost when process was killed):');
+        for (const tool of tools) {
+            const inputSummary = this.summarizeToolInput(tool.input);
+            lines.push(`- **${tool.toolName}**(${inputSummary}) — was in progress, may need re-running`);
+        }
+        return lines;
+    }
+    /** Format tool results from ToolUseRecord[] for recovery prompts */
+    formatToolResults(toolResults, maxLen = 3000) {
+        const completed = toolResults.filter(t => t.result !== undefined && !t.isError);
+        if (completed.length === 0)
+            return [];
+        const lines = [`### ${completed.length} preserved results from prior work:`, ''];
+        for (const tool of completed) {
+            const inputSummary = this.summarizeToolInput(tool.toolInput);
+            const preview = tool.result && tool.result.length > maxLen
+                ? `${tool.result.slice(0, maxLen)}...\n(truncated, ${tool.result.length} chars total)`
+                : tool.result || '';
+            lines.push(`**${tool.toolName}**(${inputSummary}):`);
+            lines.push(preview);
+            lines.push('');
+        }
+        return lines;
+    }
+    /** Format conversation history for recovery prompts */
+    formatConversationHistory(movements, maxMovements = 5) {
+        const recent = movements.slice(-maxMovements);
+        if (recent.length === 0)
+            return [];
+        const lines = ['### Conversation so far:'];
+        for (const movement of recent) {
+            const promptText = movement.userPrompt.length > 300 ? `${movement.userPrompt.slice(0, 300)}...` : movement.userPrompt;
+            lines.push(`**User (prompt ${movement.sequenceNumber}):** ${promptText}`);
+            if (movement.assistantResponse) {
+                const response = movement.assistantResponse.length > 1000
+                    ? `${movement.assistantResponse.slice(0, 1000)}...\n(truncated, ${movement.assistantResponse.length} chars)`
+                    : movement.assistantResponse;
+                lines.push(`**Your response:** ${response}`);
+            }
+            lines.push('');
+        }
+        return lines;
+    }
     /**
      * Load history from disk
      */
@@ -404,6 +1002,7 @@ export class ImprovisationSessionManager extends EventEmitter {
         this.accumulatedKnowledge = '';
         this.isFirstPrompt = true; // Reset to start fresh Claude session
         this.claudeSessionId = undefined; // Clear Claude session ID to start new conversation
+        this.cleanupAttachments();
         this.saveHistory();
         this.emit('onSessionUpdate', this.getHistory());
     }
@@ -445,6 +1044,12 @@ export class ImprovisationSessionManager extends EventEmitter {
     get isExecuting() {
         return this._isExecuting;
     }
+    /**
+     * Timestamp when current execution started (undefined when not executing)
+     */
+    get executionStartTimestamp() {
+        return this._executionStartTimestamp;
+    }
     /**
      * Get buffered execution events for replay on reconnect.
      * Only meaningful while isExecuting is true.