erosolar-cli 2.1.191 → 2.1.192

package/dist/shell/interactiveShell.js

@@ -4,6 +4,7 @@ import { exec } from 'node:child_process';
 import { promisify } from 'node:util';
 import { existsSync, readFileSync, statSync, writeFileSync } from 'node:fs';
 import { join, resolve } from 'node:path';
+import { AgentOrchestrator } from '../core/agentOrchestrator.js';
 import { display } from '../ui/display.js';
 import { theme } from '../ui/theme.js';
 import { getTerminalColumns } from '../ui/layout.js';
@@ -12,13 +13,11 @@ import { ensureSecretForProvider, getSecretDefinitionForProvider, getSecretValue
 import { saveActiveProfilePreference, saveModelPreference, loadToolSettings, saveToolSettings, clearToolSettings, clearActiveProfilePreference, loadSessionPreferences, saveSessionPreferences, loadFeatureFlags, saveFeatureFlags, toggleFeatureFlag, FEATURE_FLAG_INFO, } from '../core/preferences.js';
 import { getLearningSummary, getRecentLearning, commitLearning, exportAllLearning, getLearningDir, } from '../core/learningPersistence.js';
 import { buildEnabledToolSet, evaluateToolPermissions, getToolToggleOptions, } from '../capabilities/toolRegistry.js';
-import { FlowOrchestrator } from '../runtime/flowOrchestrator.js';
 import { detectApiKeyError } from '../core/errors/apiKeyErrors.js';
 import { detectPromptBlockError, } from '../core/errors/promptBlockErrors.js';
 import { detectNetworkError } from '../core/errors/networkErrors.js';
 import { buildWorkspaceContext } from '../workspace.js';
 import { buildInteractiveSystemPrompt } from './systemPrompt.js';
-import { getTaskCompletionDetector, resetTaskCompletionDetector, WRITE_TOOLS, } from './taskCompletionDetector.js';
 import { discoverAllModels, quickCheckProviders, getCachedDiscoveredModels, sortModelsByPriority } from '../core/modelDiscovery.js';
 import { getModels, getSlashCommands, getProviders } from '../core/agentSchemaLoader.js';
 import { loadMcpServers } from '../mcp/config.js';
@@ -86,10 +85,6 @@ const CONTEXT_CLEANUP_SYSTEM_PROMPT = `Summarize earlier IDE collaboration so th
 - Respond in plain Markdown only (no tool or shell calls).`;
 const MAX_ATTACHMENT_BYTES = 200 * 1024; // 200KB per attachment
 const MAX_ATTACHMENT_CHARS = 16000; // Guardrail to avoid flooding context
-const WRITE_TOOL_NAMES = new Set(Array.from(WRITE_TOOLS)
-    .map((tool) => tool.toLowerCase())
-    // Bash/execute commands can be read-only; guard only on clear mutating tools
-    .filter((tool) => !tool.includes('bash') && !tool.startsWith('execute')));
 export class InteractiveShell {
     agent = null;
     profile;
@@ -124,7 +119,6 @@ export class InteractiveShell {
     uiUpdates;
     _fileChangeTracker = new FileChangeTracker(); // Reserved for future file tracking features
     alphaZeroMetrics; // Alpha Zero 2 performance tracking
-    flowOrchestrator = new FlowOrchestrator();
     maxNetworkRetries = 2;
     statusSubscription = null;
     followUpQueue = [];
@@ -140,6 +134,7 @@ export class InteractiveShell {
     lastContextWarningLevel = null;
     sessionPreferences;
     autosaveEnabled;
+    orchestrationNoticeShown = false;
     verificationEnabled = false;
     criticalApprovalMode = 'auto';
     editGuardMode = 'display-edits';
@@ -1997,9 +1992,8 @@ export class InteractiveShell {
             return;
         }
         const isReasoning = type === 'reasoning';
-        // Approximate token count (roughly 4 chars per token)
-        this.streamingTokenCount += Math.ceil(chunk.length / 4);
-        this.renderer?.updateStreamingTokens(this.streamingTokenCount);
+        // Token count is updated from real provider usage data in onAssistantMessage
+        // Do NOT estimate tokens from chunk length - wait for actual API response
         // Keep pinned status updated for all streaming chunks
         this.updateStreamingStatusFromChunk(chunk);
         // Handle <thinking> tags as separate events in the queue
@@ -2276,14 +2270,17 @@ export class InteractiveShell {
         // Check for continuous/infinite loop commands or auto-escalation to completion mode
         const explicitContinuous = this.isContinuousCommand(trimmed);
         const autoContinuous = this.shouldAutoRunToCompletion(trimmed);
-        if (explicitContinuous || autoContinuous) {
-            if (autoContinuous && !explicitContinuous) {
-                display.showSystemMessage('⚡ Actionable request detected; running continuously until complete (Ctrl+C to stop).');
-            }
+        if (explicitContinuous) {
             await this.processContinuousRequest(trimmed);
             this.syncRendererInput();
             return;
         }
+        if (autoContinuous) {
+            display.showSystemMessage('⚡ Actionable request detected; orchestrating until complete (Ctrl+C to stop).');
+            await this.processRequest(trimmed, { orchestrate: true });
+            this.syncRendererInput();
+            return;
+        }
         // Direct execution for all inputs, including multi-line pastes
         await this.processRequest(trimmed);
         this.syncRendererInput();
@@ -2320,19 +2317,113 @@ export class InteractiveShell {
         ];
         const strongMaintenanceIntent = maintenancePatterns.some((pattern) => pattern.test(normalized));
         // General action intent in a code context
-        const actionVerb = /\b(fix|resolve|address|refactor|implement|upgrade|migrate|optimi[sz]e|modernize|stabilize|harden|ship|complete|finish|clean|remove|delete|prune|tidy|audit|cleanup)\b/;
+        const actionVerb = /\b(fix|resolve|address|refactor|implement|upgrade|migrate|optimi[sz]e|modernize|stabilize|harden|ship|complete|finish|clean|remove|delete|prune|tidy|audit|cleanup|debug|investigate|triage|scan|check|diagnos(e|is))\b/;
         const codeContext = /\b(repo|codebase|project|app|service|package|module|component|workspace|cli|tests?|source|files?|artifacts?)\b/;
-        const strongActionIntent = actionVerb.test(normalized) && codeContext.test(normalized) && normalized.length > 20;
-        // Allow maintenance/refactor tasks to auto-run even if phrased as questions ("got any junk to remove?")
-        if (strongMaintenanceIntent || strongActionIntent) {
-            // But still avoid purely informational questions like "what is this repo?"
-            if (startsWithInfoWord && !strongMaintenanceIntent) {
-                return false;
-            }
-            return true;
+        const hasActionVerb = actionVerb.test(normalized);
+        const hasCodeContext = codeContext.test(normalized);
+        // Bug/issue hunt or health-check intents, even when phrased as questions
+        const bugSweepIntent = /\b(got\s+any|any|open|known)\s+(bugs?|issues?|defects?)\b/;
+        const triageIntent = /\b(find|check|hunt|triage|detect|scan|review|look\s+for)\s+(bugs?|issues?|errors?|failures?)\b/;
+        const failingTestsIntent = /\b(failing|broken|red)\s+tests?\b|\btests?\s+(are\s+)?failing\b/;
+        const qualitySweepIntent = /\b(audit|health\s+check|stability\s+check|bug\s+scan|issue\s+scan|regression\s+pass)\b/;
+        const bugOrHealthIntent = [bugSweepIntent, triageIntent, failingTestsIntent, qualitySweepIntent].some((pattern) => pattern.test(normalized));
+        // Score-based intent detection keeps questions like "what is this repo?" out
+        let actionScore = 0;
+        if (strongMaintenanceIntent)
+            actionScore += 2;
+        if (bugOrHealthIntent)
+            actionScore += 2;
+        if (hasActionVerb)
+            actionScore += 1;
+        if (hasCodeContext)
+            actionScore += 1;
+        const decisiveAction = actionScore >= 2 || (actionScore === 1 && !isQuestion && normalized.length > 10);
+        if (!decisiveAction) {
+            return false;
         }
-        return false;
+        // Respect informational openers unless intent is clearly actionable
+        if (startsWithInfoWord && actionScore < 3 && !bugOrHealthIntent && !strongMaintenanceIntent) {
+            return false;
+        }
+        return true;
     }
+    shouldContinueOrchestrating(originalRequest, result) {
+        if (!result) {
+            return { shouldContinue: false, reason: null };
+        }
+        const actionable = this.shouldAutoRunToCompletion(originalRequest);
+        if (!actionable) {
+            const incomplete = result.exitReason !== 'complete';
+            return { shouldContinue: incomplete, reason: incomplete ? result.exitReason : null };
+        }
+        if (result.exitReason !== 'complete') {
+            return { shouldContinue: true, reason: result.exitReason };
+        }
+        const executedTools = result.passes.some(pass => (pass.toolsUsed?.length ?? 0) > 0);
+        const lastPass = result.passes[result.passes.length - 1];
+        const lastPlanOnly = Boolean(lastPass?.planOnly) && !lastPass?.tookAction && (!lastPass?.toolsUsed?.length);
+        const lastEmpty = !lastPass?.response?.trim();
+        if (!executedTools) {
+            return { shouldContinue: true, reason: 'no-action' };
+        }
+        if (lastPlanOnly) {
+            return { shouldContinue: true, reason: 'plan-only' };
+        }
+        if (lastEmpty) {
+            return { shouldContinue: true, reason: 'empty-response' };
+        }
+        return { shouldContinue: false, reason: null };
+    }
+    describeContinuationReason(reason) {
+        switch (reason) {
+            case 'no-action':
+                return 'No tools or concrete actions were executed; continuing until real work is done.';
+            case 'plan-only':
+                return 'Last pass was just planning; executing the next concrete step now.';
+            case 'empty-response':
+            case 'empty':
+                return 'Previous pass returned nothing; resuming with a concrete action.';
+            case 'stalled':
+                return 'Previous orchestration stalled; forcing continuation.';
+            case 'max-passes':
+                return 'Reached pass limit; extending the run to finish the task.';
+            default:
+                return 'Continuing orchestration until the task is actually finished.';
+        }
+    }
+    buildForcedContinuationPrompt(originalRequest, lastResult, continuationReason) {
+        const lastPass = lastResult.passes[lastResult.passes.length - 1];
+        const rawResponse = lastPass?.response?.trim() ?? '';
+        const truncatedResponse = rawResponse ? rawResponse.slice(0, 1200) : '';
+        const responseNote = rawResponse && rawResponse.length > truncatedResponse.length
+            ? '\n\n[Last response truncated]'
+            : '';
+        const usedTools = lastResult.passes.flatMap(pass => pass.toolsUsed ?? []);
+        const hasToolUsage = usedTools.length > 0;
+        const toolsUsed = hasToolUsage
+            ? `Tools used so far: ${usedTools.join(', ')}.`
+            : 'No tools have been used yet - start with a quick workspace scan (list_files + glob/grep/search) and then take concrete actions now.';
+        const responseSection = truncatedResponse
+            ? `Last response snapshot:\n${truncatedResponse}${responseNote}\n`
+            : '';
+        const exitReason = continuationReason ?? lastResult.exitReason ?? 'incomplete';
+        const reasonLine = this.describeContinuationReason(exitReason);
+        const missingActionLine = hasToolUsage
+            ? ''
+            : 'No tool calls or file edits have been made yet. Start with list_files + glob/grep/search to sweep the repo, then run a command (read/search/test/edit) and show the output before summarizing.';
+        return `${originalRequest.trim()}
+
+The previous orchestration stopped early (reason: ${exitReason}). Continue working until the task is truly finished. Use tools, run checks, and only declare completion when nothing remains.
+
+${reasonLine}
+${toolsUsed}
+${responseSection}
+${missingActionLine ? `${missingActionLine}\n` : ''}Resume with the next concrete action now.`;
+    }
+    /**
+     * Decide whether to automatically continue execution when the model stops after a plan/summary
+     * without taking actions. This keeps flows moving toward completion instead of stalling on planning.
+     */
     isExitCommand(input) {
         const lower = input.trim().toLowerCase();
         return (lower === 'exit' ||
@@ -5780,22 +5871,20 @@ export class InteractiveShell {
         this.clearInlinePanel();
         this.syncRendererInput();
     }
-    async runFlowControlledTask(initialRequest, options) {
-        const { mode } = options;
-        const followUpType = mode === 'continuous' ? 'continuous' : 'request';
+    async processRequest(request, options) {
         if (this.isProcessing) {
-            this.enqueueFollowUpAction({ type: followUpType, text: initialRequest });
-            return null;
+            this.enqueueFollowUpAction({ type: 'request', text: request });
+            return;
         }
         if (!this.agent && !this.rebuildAgent()) {
             display.showWarning('Configure an API key via /secrets before sending requests.');
-            return null;
+            return;
         }
         this.inlinePanelScopeActive = false;
         this.clearInlinePanel();
         const agent = this.agent;
         if (!agent) {
-            return null;
+            return;
         }
         this.toolsUsedThisRun = [];
         this.currentToolCalls = [];
@@ -5806,169 +5895,166 @@ export class InteractiveShell {
         else {
             this.resetNetworkRetryState();
         }
+        // Reset per-request render tracking
         this.responseRendered = false;
-        if (this.shouldLogPrompt(initialRequest)) {
-            this.logUserPrompt(initialRequest);
+        if (this.shouldLogPrompt(request)) {
+            this.logUserPrompt(request);
         }
         this.isProcessing = true;
         this.uiUpdates.setMode('processing');
-        this.streamingTokenCount = 0;
+        this.streamingTokenCount = 0; // Reset token counter for new request
         this.terminalInput.setStreaming(true);
+        // Keep the persistent input/control bar active as we transition into streaming.
         this.syncRendererInput();
         this.renderer?.render();
-        const overallStartTime = Date.now();
+        const requestStartTime = Date.now(); // Alpha Zero 2 timing
+        // Clear previous parallel agents and start fresh for new request
         const parallelManager = getParallelAgentManager();
         parallelManager.clear();
         parallelManager.startBatch();
-        this.lastUserQuery = initialRequest;
-        this.currentTaskType = classifyTaskType(initialRequest);
+        // AlphaZero: Track task for learning
+        this.lastUserQuery = request;
+        this.currentTaskType = classifyTaskType(request);
         this.currentToolCalls = [];
         this.clearToolUsageMeta();
         this.renderer?.setActivity('Starting...');
-        this.uiAdapter.startProcessing(mode === 'continuous' ? 'Continuous execution mode' : 'Working on your request');
+        this.uiAdapter.startProcessing('Working on your request');
         this.setProcessingStatus();
         this.beginAiRuntime();
-        this.startStreamingHeartbeat(mode === 'continuous' ? 'Streaming' : 'Streaming response');
-        const completionDetector = getTaskCompletionDetector();
-        completionDetector.reset();
-        this.flowOrchestrator.start(initialRequest);
-        this.updateStatusMessage('Orchestrating request...');
-        if (mode === 'continuous') {
-            display.showSystemMessage('Continuous mode active. Ctrl+C to stop.');
-        }
-        else {
-            display.showSystemMessage('Flow orchestrator engaged; running until the request is satisfied.');
+        let responseText = '';
+        let orchestratorResult = null;
+        const orchestrate = options?.orchestrate ?? true;
+        const orchestratorPassLimit = options?.maxPasses ?? 10;
+        if (orchestrate && !this.orchestrationNoticeShown) {
+            display.showSystemMessage('⚡ Orchestrating every prompt until completion. Press Ctrl+C to stop a run early.');
+            this.orchestrationNoticeShown = true;
         }
-        const informationalGuardrail = this.flowOrchestrator.isInformationalRequest(initialRequest)
-            ? 'This is an informational request. Keep actions lightweight: avoid running full test/build/lint suites unless explicitly required to answer the question or to confirm a change you made. Prefer quick reads/searches and summarize findings instead of long validations.'
-            : '';
-        const applyInformationalGuardrail = (prompt) => {
-            if (!informationalGuardrail) {
-                return prompt;
-            }
-            return prompt.includes(informationalGuardrail)
-                ? prompt
-                : `${prompt}\n\n${informationalGuardrail}`;
-        };
-        let currentPrompt = applyInformationalGuardrail(initialRequest);
-        if (this.isSelfImprovementRequest(initialRequest)) {
-            currentPrompt = applyInformationalGuardrail(`${initialRequest}
-
-IMPORTANT: You have full git access. After making improvements:
-1. Use bash to run: git status (see changes)
-2. Use bash to run: git add -A (stage changes)
-3. Use bash to run: git commit -m "descriptive message" (commit)
-4. Use bash to run: git push (when milestone reached)
-
-Commit frequently with descriptive messages. Push when ready.
-When truly finished with ALL tasks, explicitly state "TASK_FULLY_COMPLETE".`);
-        }
-        let iteration = 0;
-        let lastResponseText = '';
-        let lastToolsUsed = [];
-        let result = null;
         try {
-            while (true) {
-                iteration++;
-                this.toolsUsedThisRun = [];
-                this.updateStatusMessage('Orchestrating request...');
-                try {
-                    display.showThinking('Responding...');
-                    this.refreshStatusLine(true);
-                    const response = await agent.send(currentPrompt, true);
-                    lastResponseText = response ?? '';
-                    this.finishStreamingFormatter(undefined, { refreshPrompt: false, mode: 'complete' });
-                    await this.awaitPendingCleanup();
-                    this.captureHistorySnapshot();
-                    this.autosaveIfEnabled();
-                    const elapsedMs = Date.now() - overallStartTime;
-                    this.alphaZeroMetrics.recordMessage(elapsedMs);
-                    if (!response?.trim()) {
-                        display.showWarning('Model returned an empty response. Continuing orchestration...');
-                        currentPrompt = applyInformationalGuardrail(`${initialRequest}
-
-The previous reply was empty. Resume the task now: take the next action, call the necessary tools, and report progress.`);
-                        continue;
-                    }
-                    const toolsUsed = this.getExecutedTools(response);
-                    lastToolsUsed = toolsUsed;
-                    toolsUsed.forEach(tool => completionDetector.recordToolCall(tool, true, true));
-                    const completionAnalysis = completionDetector.analyzeCompletion(response, toolsUsed);
-                    display.showSystemMessage(`📈 Completion confidence: ${(completionAnalysis.confidence * 100).toFixed(0)}%`);
-                    const decision = this.flowOrchestrator.decide({
-                        iteration,
-                        response,
-                        toolsUsed,
-                        completionAnalysis,
-                        verificationConfirmed: this.flowOrchestrator.isVerificationPending()
-                            ? completionDetector.isVerificationConfirmed(response)
-                            : false,
-                    });
-                    if (decision.type === 'stop') {
-                        display.showSystemMessage(decision.message);
-                        break;
-                    }
-                    if (decision.type === 'stagnation-stop') {
-                        display.showWarning(decision.message);
+            // Start streaming - no header needed, the input area already provides context
+            this.startStreamingHeartbeat('Streaming response');
+            if (orchestrate) {
+                const orchestrator = new AgentOrchestrator(agent);
+                orchestratorResult = await orchestrator.runToCompletion(request, {
+                    streaming: true,
+                    maxPasses: orchestratorPassLimit,
+                    maxStagnantPasses: 3,
+                    verificationMode: 'auto',
+                    enforceActions: true,
+                });
+                const MAX_CONTINUATIONS = 2;
+                let continuationRuns = 0;
+                while (orchestratorResult && continuationRuns < MAX_CONTINUATIONS) {
+                    const continuationDecision = this.shouldContinueOrchestrating(request, orchestratorResult);
+                    if (!continuationDecision.shouldContinue) {
                         break;
                     }
-                    if (decision.type === 'execute-plan') {
-                        display.showSystemMessage(decision.message);
-                        currentPrompt = applyInformationalGuardrail(decision.prompt);
-                        await new Promise(resolve => setTimeout(resolve, 500));
-                        continue;
-                    }
-                    if (decision.type === 'verify') {
-                        display.showSystemMessage(decision.message);
-                        currentPrompt = applyInformationalGuardrail(decision.prompt);
-                        await new Promise(resolve => setTimeout(resolve, 500));
-                        continue;
-                    }
-                    if (decision.type === 'continue') {
-                        if (decision.message) {
-                            display.showSystemMessage(decision.message);
-                        }
-                        currentPrompt = applyInformationalGuardrail(decision.prompt);
-                    }
-                    await new Promise(resolve => setTimeout(resolve, 500));
+                    const reasonMessage = this.describeContinuationReason(continuationDecision.reason);
+                    display.showSystemMessage(`🔁 ${reasonMessage}`);
+                    const continuationPrompt = this.buildForcedContinuationPrompt(request, orchestratorResult, continuationDecision.reason ?? undefined);
+                    const continuationResult = await orchestrator.runToCompletion(continuationPrompt, {
+                        streaming: true,
+                        maxPasses: Math.max(orchestratorPassLimit, 12),
+                        maxStagnantPasses: 3,
+                        verificationMode: 'auto',
+                        enforceActions: true,
+                    });
+                    orchestratorResult = {
+                        finalResponse: continuationResult.finalResponse,
+                        passes: [...orchestratorResult.passes, ...continuationResult.passes],
+                        exitReason: continuationResult.exitReason,
+                    };
+                    continuationRuns++;
                 }
-                catch (error) {
-                    display.stopThinking(false);
-                    if (this.isContextOverflowError(error)) {
-                        display.showSystemMessage(`⚡ Context overflow handled. Continuing with reduced context...`);
-                        continue;
-                    }
-                    const handled = this.handleProviderError(error, () => this.runFlowControlledTask(initialRequest, options));
-                    if (!handled) {
-                        display.showError(error instanceof Error ? error.message : String(error), error);
-                        break;
+                responseText = orchestratorResult.finalResponse;
+            }
+            else {
+                responseText = await agent.send(request, true);
+            }
+            this.finishStreamingFormatter(undefined, { refreshPrompt: false, mode: 'complete' });
+            await this.awaitPendingCleanup();
+            this.captureHistorySnapshot();
+            this.autosaveIfEnabled();
+            // Track metrics with Alpha Zero 2
+            const elapsedMs = Date.now() - requestStartTime;
+            this.alphaZeroMetrics.recordMessage(elapsedMs);
+            if (!responseText?.trim()) {
+                display.showWarning('The provider returned an empty response. Check your API key/provider selection or retry the prompt.');
+            }
+            // AlphaZero: Extract and track tool calls from response
+            const toolsUsed = orchestratorResult
+                ? orchestratorResult.passes.flatMap(pass => pass.toolsUsed)
+                : this.getExecutedTools(responseText);
+            this.currentToolCalls = toolsUsed.map(name => ({
+                name,
+                arguments: {},
+                success: true, // Assume success if we got here
+                duration: 0,
+            }));
+            if (orchestratorResult && orchestratorResult.exitReason !== 'complete') {
+                const exitDetail = (() => {
+                    switch (orchestratorResult?.exitReason) {
+                        case 'max-passes':
+                            return 'Reached orchestrator pass limit; showing last response.';
+                        case 'empty-response':
+                            return 'Received empty replies while orchestrating; showing last response.';
+                        case 'stalled':
+                            return 'Orchestrator detected stagnation; showing last response.';
+                        default:
+                            return null;
                     }
+                })();
+                if (exitDetail) {
+                    display.showSystemMessage(`⚠️ ${exitDetail}`);
                 }
             }
-            if (!this.responseRendered && lastResponseText.trim()) {
-                const finalText = lastResponseText.trim();
+            // AlphaZero: Check for failure in response
+            const failure = detectFailure(responseText, {
+                toolCalls: this.currentToolCalls,
+                userMessage: request,
+            });
+            if (failure) {
+                this.lastFailure = failure;
+                // Check if we have a recovery strategy
+                const strategy = findRecoveryStrategy(failure);
+                if (strategy) {
+                    display.showSystemMessage(`🔄 Found recovery strategy for this type of issue (success rate: ${Math.round(strategy.successRate * 100)}%)`);
+                }
+            }
+            else {
+                // Success - record the tool pattern for this task type
+                if (this.currentToolCalls.length > 0) {
+                    const toolPattern = {
+                        taskType: this.currentTaskType,
+                        toolSequence: this.currentToolCalls.map(t => t.name),
+                        successRate: 1.0,
+                        avgDuration: elapsedMs,
+                        occurrences: 1,
+                    };
+                    addToolPattern(this.currentTaskType, toolPattern);
+                }
+                // Clear action history on success
+                clearActionHistory();
+                this.lastFailure = null;
+            }
+        }
+        catch (error) {
+            const handled = this.handleProviderError(error, () => this.processRequest(request, options));
+            if (!handled) {
+                // Pass full error object for enhanced formatting with stack trace
+                display.showError(error instanceof Error ? error.message : String(error), error);
+            }
+        }
+        finally {
+            // Fallback: if no assistant message was rendered (e.g., streaming hiccup), show the full response
+            if (!this.responseRendered && responseText.trim()) {
+                const finalText = responseText.trim();
                 display.showAssistantMessage(finalText, { isFinal: true });
                 this.ui.controller.recordAssistantResponse(finalText, {
                     source: 'final',
                 });
                 this.responseRendered = true;
             }
-            result = {
-                finalResponse: lastResponseText,
-                toolsUsed: lastToolsUsed,
-                elapsedMs: Date.now() - overallStartTime,
-            };
-        }
-        finally {
             this.finishStreamingFormatter(undefined, { refreshPrompt: false, mode: 'complete' });
-            const totalElapsed = Date.now() - overallStartTime;
-            const minutes = Math.floor(totalElapsed / 60000);
-            const seconds = Math.floor((totalElapsed % 60000) / 1000);
-            const completionLabel = mode === 'continuous'
-                ? `\n🏁 Continuous execution completed in ${minutes}m ${seconds}s total`
-                : `\n🏁 Flow orchestration completed in ${minutes}m ${seconds}s total`;
-            display.showSystemMessage(completionLabel);
-            resetTaskCompletionDetector();
             display.stopThinking(false);
             this.uiUpdates.setMode('processing');
             this.stopStreamingHeartbeat('complete', { quiet: true });
@@ -5981,60 +6067,20 @@ The previous reply was empty. Resume the task now: take the next action, call th
             this.updateStatusMessage(null);
             this.toolsUsedThisRun = [];
             queueMicrotask(() => this.uiUpdates.setMode('idle'));
+            // CRITICAL: Ensure readline prompt is active for user input
+            // Erosolar-CLI style: New prompt naturally appears at bottom
             this.ensureReadlineReady();
             this.scheduleQueueProcessing();
             this.maybeProcessPromptInbox();
             this.refreshQueueIndicators();
         }
-        return result;
-    }
-    handleFlowRunOutcome(request, result) {
-        this.currentToolCalls = result.toolsUsed.map((name) => ({
-            name,
-            arguments: {},
-            success: true,
-            duration: 0,
-        }));
-        const failure = detectFailure(result.finalResponse, {
-            toolCalls: this.currentToolCalls,
-            userMessage: request,
-        });
-        if (failure) {
-            this.lastFailure = failure;
-            const strategy = findRecoveryStrategy(failure);
-            if (strategy) {
-                display.showSystemMessage(`🔄 Found recovery strategy for this type of issue (success rate: ${Math.round(strategy.successRate * 100)}%)`);
-            }
-            return;
-        }
-        if (this.currentToolCalls.length > 0) {
-            const toolPattern = {
-                taskType: this.currentTaskType,
-                toolSequence: this.currentToolCalls.map((t) => t.name),
-                successRate: 1.0,
-                avgDuration: result.elapsedMs,
-                occurrences: 1,
-            };
-            addToolPattern(this.currentTaskType, toolPattern);
-        }
-        clearActionHistory();
-        this.lastFailure = null;
-    }
-    async processRequest(request) {
-        const result = await this.runFlowControlledTask(request, {
-            mode: 'standard',
-        });
-        if (!result) {
-            return;
-        }
-        this.handleFlowRunOutcome(request, result);
     }
     /**
      * Process a continuous/infinite loop request.
      * Runs the agent in a loop until:
      * 1. The agent indicates completion (verified by AI confirmation)
      * 2. User interrupts (Ctrl+C)
-     * 3. The orchestrator halts due to stagnation/verification exhaustion
+     * 3. Maximum iterations reached (safety limit)
      *
      * Uses intelligent task completion detection with AI verification
      * to ensure tasks are truly complete before stopping.
@@ -6042,13 +6088,29 @@ The previous reply was empty. Resume the task now: take the next action, call th
      * Context is automatically managed - overflow errors trigger auto-recovery.
      */
     async processContinuousRequest(initialRequest) {
-        const result = await this.runFlowControlledTask(initialRequest, {
-            mode: 'continuous',
-        });
-        if (!result) {
+        const MAX_PASSES = 100; // Safety limit to prevent truly infinite loops
+        if (this.isProcessing) {
+            this.enqueueFollowUpAction({ type: 'continuous', text: initialRequest });
             return;
         }
-        this.handleFlowRunOutcome(initialRequest, result);
+        if (!this.agent && !this.rebuildAgent()) {
+            display.showWarning('Configure an API key via /secrets before sending requests.');
+            return;
+        }
+        display.showSystemMessage(`Continuous mode active. Ctrl+C to stop.`);
+        const preparedRequest = this.isSelfImprovementRequest(initialRequest)
+            ? `${initialRequest}
+
+IMPORTANT: You have full git access. After making improvements:
+1. Use bash to run: git status (see changes)
+2. Use bash to run: git add -A (stage changes)
+3. Use bash to run: git commit -m "descriptive message" (commit)
+4. Use bash to run: git push (when milestone reached)
+
+Commit frequently with descriptive messages. Push when ready.
+When truly finished with ALL tasks, explicitly state "TASK_FULLY_COMPLETE".`
+            : initialRequest;
+        await this.processRequest(preparedRequest, { orchestrate: true, maxPasses: MAX_PASSES });
     }
     /**
      * Resolve executed tools for the current turn. Prefer the actual tool
@@ -6070,6 +6132,129 @@ The previous reply was empty. Resume the task now: take the next action, call th
         }
         return this.extractToolsFromResponse(responseText);
     }
+    /**
+     * Detect plan-only responses that narrate intent without executing actions.
+     */
+    isPlanOnlyResponse(response) {
+        const normalized = response.trim().toLowerCase();
+        if (!normalized) {
+            return false;
+        }
+        // If the assistant is clearly declaring completion, don't treat it as plan-only
+        const completionGuards = [
+            /\bnothing\s+(left|else)\s+(to\s+do|pending)\b/i,
+            /\b(already|now)\s+(clean|complete|done)\b/i,
+            /\b(no\s+(junk|issues?|changes?)\s+found)\b/i,
+        ];
+        if (completionGuards.some((pattern) => pattern.test(response))) {
+            return false;
+        }
+        const planIndicators = [
+            /\bplan\b/i,
+            /\bapproach\b/i,
+            /\bsteps?:\b/i,
+            /\bstep\s+1\b/i,
+            /\bstart by\b/i,
+            /\bfirst[, ]/i,
+            /\bthen\b/i,
+            /\bnext\b/i,
+            /\bi['’]?\s*will\b/i,
+            /\bi['’]?\s*ll\b/i,
+            /\bi['’]?\s*can\b.{0,40}\bthen\b/i,
+            /\bi['’]?\s*(?:will|ll)\s+begin\b/i,
+        ];
+        return planIndicators.some((pattern) => pattern.test(response));
+    }
+    /**
+     * Check if a response contains indicators that work is actually incomplete,
+     * even if it also contains TASK_FULLY_COMPLETE marker.
+     * This catches contradictory responses where the AI says "done" but also "not integrated yet".
+     */
+    responseIndicatesIncompleteWork(response) {
+        // Patterns that indicate work isn't actually complete
+        // Organized by category for maintainability
+        const incompletePatterns = [
+            // === INTEGRATION/DEPLOYMENT STATE ===
+            // "hasn't been integrated/implemented/connected yet"
+            /hasn'?t\s+been\s+(integrated|implemented|connected|deployed|added|completed|tested|verified)\s*(yet|still)?/i,
+            // "not yet integrated/implemented" or "not integrated"
+            /not\s+(yet\s+)?(integrated|implemented|connected|deployed|functional|working|complete|tested|verified)/i,
+            // "ready for integration" = NOT integrated
+            /ready\s+(for|to\s+be)\s+(integration|integrated|connected|deployed|testing|review)/i,
+            // "needs to be integrated"
+            /needs?\s+to\s+be\s+(integrated|connected|deployed|added|hooked|wired|tested|reviewed|merged)/i,
+            // Passive voice: "was not performed/completed"
+            /was\s+not\s+(performed|completed|implemented|deployed|integrated|tested)/i,
+            // "the [X] service hasn't been"
+            /the\s+\w+\s+(service|module|component|feature)\s+hasn'?t\s+been/i,
+            // === PARTIAL/INCOMPLETE STATE ===
+            // "still stores/uses/has" (current bad state persists)
+            /still\s+(stores?|uses?|has|contains?|needs?|requires?|missing|lacks?|broken)/i,
+            // Partial completion: "partially", "mostly", "almost"
+            /\b(partially|mostly|almost|nearly|not\s+fully)\s+(complete|done|finished|implemented|working)/i,
+            // Explicit partial: "part of", "some of", "half of"
+            /\b(only\s+)?(part|some|half|portion)\s+of\s+(the\s+)?(task|work|feature|implementation)/i,
+            // === QUALIFIER WORDS (uncertain completion) ===
+            // "should be complete", "appears complete", "theoretically"
+            /\b(should|might|may|could|appears?\s+to)\s+be\s+(complete|done|working|functional)/i,
+            /\btheoretically\s+(complete|done|working|functional)/i,
+            // "assuming", "if everything works"
+            /\b(assuming|provided|if)\s+(everything|it|this|that)\s+(works?|is\s+correct)/i,
+            // === SELF-CONTRADICTION PHRASES ===
+            // "done but...", "complete except...", "finished however..."
+            /\b(done|complete|finished)\s+(but|except|however|although|though)/i,
+            // "however" followed by incomplete indicator
+            /however[,\s].{0,50}?(hasn'?t|not\s+yet|still\s+needs?|pending|remains?|missing|broken|failing)/i,
+            // "but" followed by negative state
+            /\bbut\s+.{0,30}?(not|hasn'?t|won'?t|can'?t|doesn'?t|isn'?t|wasn'?t)/i,
+            // === FUTURE TENSE / DEFERRED WORK ===
+            // "will need to", "will require"
+            /will\s+(need\s+to|require|have\s+to)\s+(integrate|connect|deploy|complete|implement|test|fix)/i,
+            // Deferred: "left as", "deferred", "postponed", "out of scope"
+            /\b(left\s+as|deferred|postponed|out\s+of\s+scope|for\s+later|in\s+a\s+future)/i,
+            // Time-dependent: "after restart", "takes effect after", "once you"
+            /\b(after\s+(restart|reboot|redeploy)|takes?\s+effect\s+after|once\s+you)/i,
+            // === REMAINING WORK INDICATORS ===
+            // "remaining tasks", "outstanding items"
+            /\b(remaining|outstanding|pending|leftover)\s+(tasks?|items?|work|issues?|steps?)/i,
+            // "X more to do", "still have to"
+            /\b(more\s+to\s+do|still\s+have\s+to|yet\s+to\s+be\s+done)/i,
+            // Explicit blockers
+            /\b(blocker|blocked\s+by|waiting\s+(for|on)|depends?\s+on)/i,
+            // === ERROR/FAILURE STATE ===
+            // "failing tests", "build errors"
+            /\b(failing|broken|erroring)\s+(tests?|builds?|checks?|validations?)/i,
+            // "tests? (are )?(still )?failing"
+            /\btests?\s+(are\s+)?(still\s+)?failing/i,
+            // "errors? to (address|fix)"
+            /\b(errors?|warnings?|issues?)\s+to\s+(address|fix|resolve)/i,
+            // "doesn't work", "isn't working", "not working"
+            /\b(doesn'?t|isn'?t|not)\s+(work|working|functional|functioning)/i,
+            // === MANUAL STEPS REQUIRED ===
+            // "you'll need to", "manually run", "requires user"
+            /\b(you('ll|\s+will)\s+need\s+to|manually\s+(run|configure|set|update)|requires?\s+user)/i,
+            // "run this command", "execute the following"
+            /\b(run\s+this|execute\s+the\s+following|apply\s+the\s+migration)/i,
+            // === TODO/FIXME IN PROSE ===
+            // TODO or FIXME mentioned as remaining work (not in code blocks)
+            /\b(todo|fixme|hack|xxx):\s/i,
+            // "need to add", "should implement"
+            /\b(need\s+to|should|must)\s+(add|implement|create|write|build|fix)\b/i,
+            // === SCOPE LIMITATIONS ===
+            // "didn't have time", "ran out of time"
+            /\b(didn'?t|did\s+not)\s+have\s+(time|chance|opportunity)/i,
+            // "beyond scope", "outside scope"
+            /\b(beyond|outside)\s+(the\s+)?scope/i,
+            // "for now" (temporary state)
+            /\b(for\s+now|at\s+this\s+point|currently)\s*.{0,20}?(not|without|lacks?|missing)/i,
+        ];
+        for (const pattern of incompletePatterns) {
+            if (pattern.test(response)) {
+                return true;
+            }
+        }
+        return false;
+    }
     /**
      * Extract tool names from a response by looking for tool call patterns
      */
@@ -6562,19 +6747,12 @@ Return ONLY JSON array:
                     const activity = normalized ? `Working: ${normalized}` : 'Working';
                     this.renderer?.setActivity(activity);
                 },
-                onBeforeFirstToolCall: (toolNames, hasNarration) => {
+                onBeforeFirstToolCall: (toolNames) => {
                     const primaryTool = toolNames[0];
                     if (primaryTool) {
                         this.renderer?.setActivity(`Running ${primaryTool}`);
                     }
-                    if (hasNarration) {
-                        return undefined;
-                    }
-                    const toolList = toolNames.length ? toolNames.join(', ') : 'tools';
-                    const ack = `💭 Plan before tools:\n- Goal: ${this.lastUserQuery || 'address request'}\n- Tools: ${toolList}\n- Executing now.`;
-                    display.showNarrative(ack);
-                    this.ui.controller.recordAssistantThought(ack);
-                    return ack;
+                    return undefined;
                 },
                 onStreamChunk: (chunk, type) => {
                     this.handleStreamChunk(chunk, type ?? 'content');
@@ -6722,14 +6900,10 @@ Return ONLY JSON array:
                             activity = `Reading ${path}`;
                         }
                         this.renderer?.setActivity(activity);
-                        // Estimate tokens for tool call (~50 tokens per call)
-                        this.streamingTokenCount += 50;
-                        this.renderer?.updateStreamingTokens(this.streamingTokenCount);
+                        // Token count updated from real provider usage - do not estimate
                     }
                     else {
-                        // Tool finished - estimate result tokens (~100 per result)
-                        this.streamingTokenCount += 100;
-                        this.renderer?.updateStreamingTokens(this.streamingTokenCount);
+                        // Tool finished - token count updated from real provider usage
                         // Reset to thinking state while model generates next response
                         this.renderer?.setActivity('Thinking');
                     }
@@ -6738,6 +6912,14 @@ Return ONLY JSON array:
                     this.lastAssistantResponse = response;
                     void this.runAutoQualityChecks('verification', response, context);
                 },
+                // Real token usage from provider during streaming
+                onUsage: (usage) => {
+                    const totalTokens = this.totalTokens(usage);
+                    if (totalTokens !== null) {
+                        this.streamingTokenCount = totalTokens;
+                        this.renderer?.updateStreamingTokens(this.streamingTokenCount);
+                    }
+                },
                 // Retry notification for transient errors
                 onRetrying: (attempt, maxAttempts, error) => {
                     const shortError = error.message.slice(0, 100);