moderado 0.3.4 → 0.3.8

package/dist/commands/chat.js

@@ -1,8 +1,9 @@
 import path from 'node:path';
 import fs from 'node:fs';
+import readline from 'node:readline';
 import { AgentLoop, PolicyManager, Router, cleanConversationalFiller, DEFAULT_MAX_OUTPUT_TOKENS } from '../vendor/core/index.js';
 import { NvidiaAdapter } from '../vendor/providers/index.js';
-import { createDefaultToolRegistry, canonicalizeRoot, createMcpTools, createWebSearchTool, discoverMcpServers, resolveInJail, WorkspaceCheckpointStore, WriteFileTool } from '../vendor/tools/index.js';
+import { createDefaultToolRegistry, ToolRegistry, canonicalizeRoot, createMcpTools, createWebSearchTool, discoverMcpServers, resolveInJail, WorkspaceCheckpointStore, WriteFileTool } from '../vendor/tools/index.js';
 import { McpServerConfigSchema } from '../vendor/contracts/index.js';
 import { getActiveConnection, loadConfig, removeMcpServer, resolveApiKey, resolveConnectionCredential, saveConnection, saveMcpServer, setMcpServerEnabled, storeConnectionCredential } from '../config.js';
 import { MemoryCredentialStore } from '../credentials.js';
@@ -13,13 +14,29 @@ import { connectProviderInteractive, isAuthenticationFailure, replaceProviderKey
 import { initWorkspace } from './init.js';
 import { expandMentions, createWorkspaceFileSource } from '../ui/file_mentions.js';
 import { listFiles } from '../vendor/tools/index.js';
-import { exitCleanly, promptInteractiveTurn, renderChatAnswerDelta, renderChatComposerCursor, renderChatThoughtTimeUpdate, renderFullWelcomeScreen, renderWelcomePopupLayer, terminalCleanExitDone, wrapText } from '../ui/welcome.js';
-import { calculateOutputTokenRate, calculateSessionCost, compactSessionMessages, createSession, exportSessionMarkdown, formatSessionCost, SessionStore } from '../sessions.js';
+import { exitCleanly, promptInteractiveTurn, renderChatAnswerDelta, renderChatComposerCursor, renderChatThoughtTimeUpdate, renderChatUsageUpdate, renderFullWelcomeScreen, renderWelcomePopupLayer, terminalCleanExitDone, wrapText } from '../ui/welcome.js';
+import { accumulateSessionUsage, compactSessionMessages, createSession, exportSessionMarkdown, formatSessionCost, SessionStore } from '../sessions.js';
 import { layerPromptBox, renderBoxLines, selectConfirmPopup, selectListPopup } from '../ui/popup.js';
 import { askModalChoice } from '../ui/prompt.js';
 import { findModelPricing } from '../model_pricing.js';
 import { inspectGitWorkspace, readGitDiff } from '../vendor/tools/index.js';
+import { buildWorkspaceMap } from '../repo_map.js';
+import { evaluateTokenBudget, parseBudgetCommand } from '../budget.js';
 import { discoverSkills, formatSkillContext } from '../skills.js';
+export function formatSessionExitSummary(session, modelId) {
+    const usage = session.usage;
+    const tokens = usage.available
+        ? `Input ${usage.promptTokens.toLocaleString('en-US')} | Output ${usage.completionTokens.toLocaleString('en-US')} | Total ${usage.totalTokens.toLocaleString('en-US')}${usage.estimated ? ' (estimated)' : ''}`
+        : 'Usage unavailable';
+    return [
+        '\x1b[1;38;5;25mSession Summary\x1b[0m',
+        `ID: ${session.id}`,
+        `Model: ${modelId ?? session.modelId ?? 'No model selected'}`,
+        `Tokens: ${tokens}`,
+        `Cost: ${formatSessionCost(usage)}`,
+        'Resume: /session resume',
+    ].join('\n');
+}
 function mutationPaths(toolName, parameters) {
     const value = parameters;
     if (toolName === 'apply_patch' && Array.isArray(value.edits)) {
@@ -32,6 +49,19 @@ export function createAgentTask(request, mode) {
         return request;
     return `Create a concise implementation checklist for this request. Inspect files as needed, but do not modify files or run commands.\n\nUser request:\n${request}`;
 }
+export function createReviewTask(changedFiles) {
+    return `Review the current workspace changes. Inspect staged and unstaged Git diffs with git_diff, then inspect relevant surrounding code. Also review untracked files listed below with read_file. Do not modify files or run commands. Report only actionable bugs or security issues, ordered by severity, with file paths and line numbers. If none are found, say "No findings". Keep the report concise.\n\nChanged files:\n${changedFiles.join('\n') || '(none)'}`;
+}
+export function createReviewToolRegistry() {
+    const source = createDefaultToolRegistry();
+    const registry = new ToolRegistry();
+    for (const name of ['git_diff', 'read_file', 'list_files', 'search_files', 'get_definition', 'find_references']) {
+        const tool = source.get(name);
+        if (tool)
+            registry.register(tool);
+    }
+    return registry;
+}
 export const STANDARD_SLASH_COMMANDS = [
     '/connect',
     '/model',
@@ -41,6 +71,9 @@ export const STANDARD_SLASH_COMMANDS = [
     '/session',
     '/queue',
     '/workflow',
+    '/review',
+    '/map',
+    '/budget',
     '/skills',
     '/clear',
     '/help',
@@ -139,6 +172,10 @@ export function formatTurnFailureAnswer(status, errorEvent, modelId, otherConnec
     return `⚠️ **Model Error (${errorEvent?.code || 'ERR_INFERENCE_FAILED'}):**\n${cleanDetail || 'Model returned no response.'}\n\n👉 **Suggestion:** The model \`${modelId}\` encountered an error. Switch models with \`/model\` or switch providers with \`/connect\`${altSuggestion}.`;
 }
 export function resolveTurnAssistantAnswer(result, previousHistoryLength, lastErrorEvent, streamedAnswer = '', currentModel = 'unknown', otherConnections = []) {
+    if (result.status === 'cancelled') {
+        const partial = cleanConversationalFiller(streamedAnswer);
+        return { answer: [partial, 'Task stopped.'].filter(Boolean).join('\n\n'), isError: true };
+    }
     if (result.status === 'failed' || lastErrorEvent) {
         return {
             answer: formatTurnFailureAnswer(result.status, lastErrorEvent, result.selectedModel?.id || currentModel, otherConnections),
@@ -216,15 +253,9 @@ export function handleGenerationKeypress(str, key, queue, actions) {
         actions.abort();
         return;
     }
-    // 2. Escape: if draft has text, clear draft; if draft is empty, abort generation
+    // 2. Escape always stops the active task; the caller clears queued input.
     if (isGenerationCancelKey(key, str)) {
-        if (queue.currentDraft.length > 0) {
-            queue.setDraft('');
-            actions.onDraftChange();
-        }
-        else {
-            actions.abort();
-        }
+        actions.abort();
         return;
     }
     // 3. Return / Enter: commit draft or execute /btw immediately
@@ -422,9 +453,6 @@ export async function decideApproval(request, options, signal) {
     if (options.autoApprove && !request.toolName.startsWith('mcp.')) {
         return { requestId: request.requestId, status: 'approved' };
     }
-    if (['write_file', 'edit_file', 'apply_patch', 'run_command', 'run_diagnostics'].includes(request.toolName)) {
-        return { requestId: request.requestId, status: 'approved' };
-    }
     return options.requestInteractiveApproval(request, signal);
 }
 export async function createMcpToolRegistry(servers, webSearch = {}) {
@@ -697,7 +725,8 @@ export async function handleChatSession(args, version, signal) {
     if (activeConnection)
         activateConnection(activeConnection, args.model);
     let activeMode = args.readOnly ? 'Plan' : 'Execute';
-    let activeAutoApprove = Boolean(args.autoApprove);
+    let activeAutoApprove = !args.nonInteractive;
+    let tokenBudget;
     const sessionStore = new SessionStore();
     let activeSession = sessionStore.loadLatestSession(canonicalWorkspace) ??
         createSession(canonicalWorkspace, {
@@ -742,8 +771,10 @@ export async function handleChatSession(args, version, signal) {
     let lastAnswer = '';
     let lastThoughtTime = 0;
     let lastOutputTokenRate;
+    let lastTokenUsage;
     let activePlan;
     const costLabel = () => formatSessionCost(activeSession.usage);
+    const exitMessage = () => `\x1b[32mGoodbye! Stay Tuned with Moderado!\x1b[0m\n\n${formatSessionExitSummary(activeSession, currentModel)}`;
     const commandQueue = new TurnCommandQueue();
     const btwHistory = [];
     while (!signal?.aborted) {
@@ -755,8 +786,9 @@ export async function handleChatSession(args, version, signal) {
         else {
             const turn = await promptInteractiveTurn({
                 model: currentModel ?? 'No model connected — use /connect', tokens: Math.round(sessionTokens), cost: costLabel(), workspace: canonicalWorkspace, version,
-                usageAvailable: activeSession.usage.available, initialMode: activeMode, initialAutoApprove: activeAutoApprove, isFirstTurn: isFirst, signal, chatQuestion: lastQuestion || undefined, chatAnswer: lastAnswer || undefined, chatThoughtTime: lastThoughtTime, outputTokenRate: lastOutputTokenRate,
+                usageAvailable: activeSession.usage.available, usageEstimated: activeSession.usage.estimated, initialMode: activeMode, initialAutoApprove: activeAutoApprove, isFirstTurn: isFirst, signal, chatQuestion: lastQuestion || undefined, chatAnswer: lastAnswer || undefined, chatThoughtTime: lastThoughtTime, outputTokenRate: lastOutputTokenRate, tokenUsage: lastTokenUsage,
                 queuedCommands: commandQueue.items,
+                onExit: exitMessage,
                 questionHistory: conversationHistory.flatMap((message) => message.role === 'user' && message.content?.trim() ? [message.content] : []),
                 onModelSelect: async (drawFrame) => {
                     if (!activeConnection) {
@@ -770,7 +802,6 @@ export async function handleChatSession(args, version, signal) {
                             providerId: activeConnection.id,
                             providerName: activeConnection.displayName,
                             currentModel,
-                            allModelsFree: activeConnection.id === 'agnes-ai',
                             signal,
                             drawFrame,
                         });
@@ -792,7 +823,7 @@ export async function handleChatSession(args, version, signal) {
                     return selection.modelId;
                 },
                 onConnect: async (drawFrame) => {
-                    const connection = await connectProviderInteractive({ signal, drawFrame, savedConnections: config.connections, resolveSavedConnection: (saved) => resolveConnectionCredential(saved, credentialStore) });
+                    const connection = await connectProviderInteractive({ signal, drawFrame, savedConnections: config.connections, connectProviders: config.connectProviders, resolveSavedConnection: (saved) => resolveConnectionCredential(saved, credentialStore) });
                     if (!connection)
                         return currentModel;
                     const runtimeConnection = process.platform === 'win32' ? await storeConnectionCredential(connection, credentialStore) : connection;
@@ -808,6 +839,7 @@ export async function handleChatSession(args, version, signal) {
                     lastAnswer = '';
                     lastThoughtTime = 0;
                     lastOutputTokenRate = undefined;
+                    lastTokenUsage = undefined;
                     activeSession.messages = [];
                     sessionStore.save(activeSession);
                 },
@@ -926,6 +958,7 @@ export async function handleChatSession(args, version, signal) {
                         lastAnswer = '';
                         lastThoughtTime = 0;
                         lastOutputTokenRate = undefined;
+                        lastTokenUsage = undefined;
                         sessionStore.save(activeSession);
                         return;
                     }
@@ -1026,11 +1059,46 @@ export async function handleChatSession(args, version, signal) {
                 continue;
             }
         }
+        if (trimmed === '/budget' || trimmed.startsWith('/budget ')) {
+            const budget = parseBudgetCommand(trimmed, tokenBudget);
+            tokenBudget = budget.limit;
+            lastQuestion = trimmed;
+            lastAnswer = budget.message;
+            lastThoughtTime = 0.001;
+            lastOutputTokenRate = undefined;
+            lastTokenUsage = undefined;
+            process.stdout.write('\x1b[H\x1b[J');
+            process.stdout.write(renderFullWelcomeScreen({
+                model: currentModel ?? 'No model connected — use /connect', tokens: Math.round(sessionTokens), cost: costLabel(),
+                usageAvailable: activeSession.usage.available, usageEstimated: activeSession.usage.estimated, workspace: canonicalWorkspace, mode: activeMode,
+                autoApprove: activeAutoApprove, chatQuestion: lastQuestion, chatAnswer: lastAnswer,
+                chatThoughtTime: lastThoughtTime, queuedCommands: commandQueue.items,
+            }, process.stdout.rows));
+            process.stdout.write(renderChatComposerCursor({ width: process.stdout.columns }, 0));
+            continue;
+        }
+        if (trimmed === '/map') {
+            lastQuestion = trimmed;
+            lastAnswer = await buildWorkspaceMap(canonicalWorkspace);
+            lastThoughtTime = 0.001;
+            lastOutputTokenRate = undefined;
+            lastTokenUsage = undefined;
+            process.stdout.write('\x1b[H\x1b[J');
+            process.stdout.write(renderFullWelcomeScreen({
+                model: currentModel ?? 'No model connected — use /connect', tokens: Math.round(sessionTokens), cost: costLabel(),
+                usageAvailable: activeSession.usage.available, usageEstimated: activeSession.usage.estimated, workspace: canonicalWorkspace, mode: activeMode,
+                autoApprove: activeAutoApprove, chatQuestion: lastQuestion, chatAnswer: lastAnswer,
+                chatThoughtTime: lastThoughtTime, queuedCommands: commandQueue.items,
+            }, process.stdout.rows));
+            process.stdout.write(renderChatComposerCursor({ width: process.stdout.columns }, 0));
+            continue;
+        }
         if (trimmed.startsWith('/queue')) {
             const args = trimmed.slice('/queue'.length).trim();
             lastQuestion = trimmed;
             lastThoughtTime = 0.001;
             lastOutputTokenRate = undefined;
+            lastTokenUsage = undefined;
             if (!args || args === 'list') {
                 lastAnswer = commandQueue.length === 0
                     ? 'Command queue is empty. Use `/queue <command>` to add commands.'
@@ -1053,7 +1121,7 @@ export async function handleChatSession(args, version, signal) {
             process.stdout.write('\x1b[H\x1b[J');
             process.stdout.write(renderFullWelcomeScreen({
                 model: currentModel ?? 'No model connected — use /connect', tokens: Math.round(sessionTokens), cost: costLabel(),
-                usageAvailable: activeSession.usage.available, workspace: canonicalWorkspace, mode: activeMode,
+                usageAvailable: activeSession.usage.available, usageEstimated: activeSession.usage.estimated, workspace: canonicalWorkspace, mode: activeMode,
                 autoApprove: activeAutoApprove, chatQuestion: lastQuestion, chatAnswer: lastAnswer,
                 chatThoughtTime: lastThoughtTime,
                 queuedCommands: commandQueue.items,
@@ -1068,6 +1136,7 @@ export async function handleChatSession(args, version, signal) {
             lastAnswer = '';
             lastThoughtTime = 0;
             lastOutputTokenRate = undefined;
+            lastTokenUsage = undefined;
             activeSession.messages = [];
             sessionStore.save(activeSession);
             continue;
@@ -1077,6 +1146,7 @@ export async function handleChatSession(args, version, signal) {
             lastAnswer = 'To leave Moderado, type /exit.';
             lastThoughtTime = 0;
             lastOutputTokenRate = undefined;
+            lastTokenUsage = undefined;
             continue;
         }
         // Non-standard or misspelled slash command rejection with advice
@@ -1086,10 +1156,11 @@ export async function handleChatSession(args, version, signal) {
             lastAnswer = slashAdvice.advice;
             lastThoughtTime = 0.001;
             lastOutputTokenRate = undefined;
+            lastTokenUsage = undefined;
             process.stdout.write('\x1b[H\x1b[J');
             process.stdout.write(renderFullWelcomeScreen({
                 model: currentModel ?? 'No model connected — use /connect', tokens: Math.round(sessionTokens), cost: costLabel(),
-                usageAvailable: activeSession.usage.available, workspace: canonicalWorkspace, mode: activeMode,
+                usageAvailable: activeSession.usage.available, usageEstimated: activeSession.usage.estimated, workspace: canonicalWorkspace, mode: activeMode,
                 autoApprove: activeAutoApprove, chatQuestion: lastQuestion, chatAnswer: lastAnswer,
                 chatThoughtTime: lastThoughtTime,
                 queuedCommands: commandQueue.items,
@@ -1105,13 +1176,14 @@ export async function handleChatSession(args, version, signal) {
                 : 'No valid user skills found in ~/.moderado/skills.';
             lastThoughtTime = 0.001;
             lastOutputTokenRate = undefined;
+            lastTokenUsage = undefined;
             process.stdout.write('\x1b[H\x1b[J');
-            process.stdout.write(renderFullWelcomeScreen({ model: currentModel ?? 'No model connected — use /connect', tokens: Math.round(sessionTokens), cost: costLabel(), usageAvailable: activeSession.usage.available, workspace: canonicalWorkspace, mode: activeMode, autoApprove: activeAutoApprove, chatQuestion: lastQuestion, chatAnswer: lastAnswer, chatThoughtTime: lastThoughtTime, queuedCommands: commandQueue.items }, process.stdout.rows));
+            process.stdout.write(renderFullWelcomeScreen({ model: currentModel ?? 'No model connected — use /connect', tokens: Math.round(sessionTokens), cost: costLabel(), usageAvailable: activeSession.usage.available, usageEstimated: activeSession.usage.estimated, workspace: canonicalWorkspace, mode: activeMode, autoApprove: activeAutoApprove, chatQuestion: lastQuestion, chatAnswer: lastAnswer, chatThoughtTime: lastThoughtTime, queuedCommands: commandQueue.items }, process.stdout.rows));
             process.stdout.write(renderChatComposerCursor({ width: process.stdout.columns }, 0));
             continue;
         }
         if (trimmed === '/exit' || trimmed === '/quit') {
-            exitCleanly('\x1b[32mGoodbye! Stay Tuned with Moderado!\x1b[0m');
+            exitCleanly(exitMessage());
         }
         if (trimmed === '/help') {
             lastQuestion = trimmed;
@@ -1123,16 +1195,20 @@ export async function handleChatSession(args, version, signal) {
                 '  /session   - Create, resume, undo, redo, share, export, or compact sessions\n' +
                 '  /queue     - Add, inspect, or clear queued follow-up commands\n' +
                 '  /workflow  - Inspect Git, build plans, or undo agent changes\n' +
+                '  /review    - Review current workspace changes for bugs\n' +
+                '  /map       - Show a compact workspace file map\n' +
+                '  /budget    - Set a per-task token limit (/budget 5000, /budget off)\n' +
                 '  /skills    - List installed user skills and reload them\n' +
                 '  /clear     - Reset conversation memory\n' +
                 '  /help      - Display commands, shortcuts & version\n' +
                 '  /exit      - Exit Moderado';
             lastThoughtTime = 0.001;
             lastOutputTokenRate = undefined;
+            lastTokenUsage = undefined;
             process.stdout.write('\x1b[H\x1b[J');
             process.stdout.write(renderFullWelcomeScreen({
                 model: currentModel ?? 'No model connected — use /connect', tokens: Math.round(sessionTokens), cost: costLabel(),
-                usageAvailable: activeSession.usage.available, workspace: canonicalWorkspace, mode: activeMode,
+                usageAvailable: activeSession.usage.available, usageEstimated: activeSession.usage.estimated, workspace: canonicalWorkspace, mode: activeMode,
                 autoApprove: activeAutoApprove, chatQuestion: lastQuestion, chatAnswer: lastAnswer,
                 chatThoughtTime: lastThoughtTime,
                 queuedCommands: commandQueue.items,
@@ -1158,13 +1234,14 @@ export async function handleChatSession(args, version, signal) {
             lastAnswer = localMetaAnswer;
             lastThoughtTime = 0.001;
             lastOutputTokenRate = undefined;
+            lastTokenUsage = undefined;
             conversationHistory = [...conversationHistory, { role: 'user', content: trimmed }, { role: 'assistant', content: localMetaAnswer }];
             activeSession.messages = conversationHistory;
             sessionStore.save(activeSession);
             process.stdout.write('\x1b[H\x1b[J');
             process.stdout.write(renderFullWelcomeScreen({
                 model: currentModel ?? 'No model connected — use /connect', tokens: Math.round(sessionTokens), cost: costLabel(),
-                usageAvailable: activeSession.usage.available, workspace: canonicalWorkspace, mode: activeMode,
+                usageAvailable: activeSession.usage.available, usageEstimated: activeSession.usage.estimated, workspace: canonicalWorkspace, mode: activeMode,
                 autoApprove: activeAutoApprove, chatQuestion: lastQuestion, chatAnswer: lastAnswer,
                 chatThoughtTime: lastThoughtTime,
                 queuedCommands: commandQueue.items,
@@ -1177,13 +1254,14 @@ export async function handleChatSession(args, version, signal) {
             lastQuestion = trimmed;
             lastThoughtTime = 0.001;
             lastOutputTokenRate = undefined;
+            lastTokenUsage = undefined;
             conversationHistory = [...conversationHistory, { role: 'user', content: trimmed }, { role: 'assistant', content: lastAnswer }];
             activeSession.messages = conversationHistory;
             sessionStore.save(activeSession);
             process.stdout.write('\x1b[H\x1b[J');
             process.stdout.write(renderFullWelcomeScreen({
                 model: currentModel ?? 'No model connected — use /connect', tokens: Math.round(sessionTokens), cost: costLabel(),
-                usageAvailable: activeSession.usage.available, workspace: canonicalWorkspace, mode: activeMode,
+                usageAvailable: activeSession.usage.available, usageEstimated: activeSession.usage.estimated, workspace: canonicalWorkspace, mode: activeMode,
                 autoApprove: activeAutoApprove, chatQuestion: lastQuestion, chatAnswer: lastAnswer,
                 chatThoughtTime: lastThoughtTime,
                 queuedCommands: commandQueue.items,
@@ -1200,10 +1278,11 @@ export async function handleChatSession(args, version, signal) {
             lastAnswer = 'Searching the web...';
             lastThoughtTime = 0;
             lastOutputTokenRate = undefined;
+            lastTokenUsage = undefined;
             process.stdout.write('\x1b[H\x1b[J');
             process.stdout.write(renderFullWelcomeScreen({
                 model: currentModel ?? 'Web search', tokens: Math.round(sessionTokens), cost: costLabel(),
-                usageAvailable: activeSession.usage.available, workspace: canonicalWorkspace, mode: activeMode,
+                usageAvailable: activeSession.usage.available, usageEstimated: activeSession.usage.estimated, workspace: canonicalWorkspace, mode: activeMode,
                 autoApprove: activeAutoApprove, chatQuestion: lastQuestion, chatAnswer: lastAnswer,
             }, process.stdout.rows));
             process.stdout.write(renderChatComposerCursor({ width: process.stdout.columns }, 0));
@@ -1224,7 +1303,7 @@ export async function handleChatSession(args, version, signal) {
                 process.stdout.write('\x1b[H\x1b[J');
                 process.stdout.write(renderFullWelcomeScreen({
                     model: currentModel ?? 'Web search', tokens: Math.round(sessionTokens), cost: costLabel(),
-                    usageAvailable: activeSession.usage.available, workspace: canonicalWorkspace, mode: activeMode,
+                    usageAvailable: activeSession.usage.available, usageEstimated: activeSession.usage.estimated, workspace: canonicalWorkspace, mode: activeMode,
                     autoApprove: activeAutoApprove, chatQuestion: lastQuestion, chatAnswer: lastAnswer,
                     chatThoughtTime: lastThoughtTime,
                 }, process.stdout.rows));
@@ -1234,7 +1313,7 @@ export async function handleChatSession(args, version, signal) {
         }
         if (!provider) {
             process.stdout.write('\nNo provider is connected. Let\'s connect one before sending this task.\n');
-            const connection = await connectProviderInteractive({ signal, savedConnections: config.connections, resolveSavedConnection: (saved) => resolveConnectionCredential(saved, credentialStore) });
+            const connection = await connectProviderInteractive({ signal, savedConnections: config.connections, connectProviders: config.connectProviders, resolveSavedConnection: (saved) => resolveConnectionCredential(saved, credentialStore) });
             if (!connection) {
                 process.stdout.write('No provider connected. Use /connect whenever you are ready.\n\n');
                 continue;
@@ -1244,10 +1323,22 @@ export async function handleChatSession(args, version, signal) {
             config = loadConfig();
             activateConnection(runtimeConnection);
         }
-        const policy = new PolicyManager({ maxSteps: args.maxSteps, readOnly: activeMode === 'Plan' || args.readOnly, nonInteractive: args.nonInteractive, timeoutSeconds: args.timeout });
+        const isReview = trimmed === '/review';
+        const reviewFiles = isReview ? (await inspectGitWorkspace(canonicalWorkspace)).files.map(file => `${file.status} ${file.path}`) : [];
+        const policy = new PolicyManager({ maxSteps: args.maxSteps, readOnly: isReview || activeMode === 'Plan' || args.readOnly, nonInteractive: args.nonInteractive, timeoutSeconds: args.timeout });
         const startedAt = Date.now();
         let thinkingTimer;
         let removeGenerationListener;
+        let turnUsage;
+        let usageRecorded = false;
+        lastTokenUsage = undefined;
+        const recordTurnUsage = (pricing) => {
+            if (!turnUsage || usageRecorded)
+                return;
+            activeSession.usage = accumulateSessionUsage(activeSession.usage, turnUsage.usage, turnUsage.estimated, pricing);
+            sessionTokens = activeSession.usage.totalTokens;
+            usageRecorded = true;
+        };
         try {
             const requestAbort = new AbortController();
             const requestSignal = signal ? AbortSignal.any([signal, requestAbort.signal]) : requestAbort.signal;
@@ -1255,6 +1346,10 @@ export async function handleChatSession(args, version, signal) {
             lastAnswer = '';
             let firstAssistantDeltaAt;
             let outputTokenRate;
+            let budgetWarningIssued = false;
+            let budgetStopped = false;
+            let lastUsagePaint = 0;
+            const usedModels = new Set();
             let answerPosition = { row: 15, column: 1 };
             let lastErrorEvent;
             const redrawChatFrame = () => {
@@ -1265,7 +1360,7 @@ export async function handleChatSession(args, version, signal) {
                     model: currentModel ?? 'No model connected — use /connect',
                     tokens: Math.round(sessionTokens),
                     cost: costLabel(),
-                    usageAvailable: activeSession.usage.available,
+                    usageAvailable: activeSession.usage.available, usageEstimated: activeSession.usage.estimated,
                     workspace: canonicalWorkspace,
                     mode: activeMode,
                     autoApprove: activeAutoApprove,
@@ -1273,6 +1368,8 @@ export async function handleChatSession(args, version, signal) {
                     chatAnswer: displayAnswer,
                     chatThoughtTime: thoughtTime,
                     outputTokenRate,
+                    tokenUsage: turnUsage,
+                    budgetStatus: budgetStopped ? 'Budget reached' : budgetWarningIssued ? 'Budget ≥80%' : undefined,
                     input: commandQueue.currentDraft,
                     queuedCommands: commandQueue.items,
                 }, process.stdout.rows));
@@ -1311,7 +1408,7 @@ export async function handleChatSession(args, version, signal) {
                                     model: currentModel ?? 'No model connected',
                                     tokens: Math.round(sessionTokens),
                                     cost: costLabel(),
-                                    usageAvailable: activeSession.usage.available,
+                                    usageAvailable: activeSession.usage.available, usageEstimated: activeSession.usage.estimated,
                                     workspace: canonicalWorkspace,
                                     mode: activeMode,
                                     autoApprove: activeAutoApprove,
@@ -1344,9 +1441,8 @@ export async function handleChatSession(args, version, signal) {
                 });
             };
             if (process.stdin.isTTY) {
-                const readlineModule = await import('node:readline');
                 const attachGenerationListener = () => {
-                    readlineModule.emitKeypressEvents(process.stdin);
+                    readline.emitKeypressEvents(process.stdin);
                     process.stdin.resume();
                     process.stdin.setRawMode(true);
                     process.stdin.on('keypress', onGenerationKeypress);
@@ -1370,17 +1466,36 @@ export async function handleChatSession(args, version, signal) {
             // Move to the chat layout before the provider can emit its first event.
             redrawChatFrame();
             thinkingTimer = setInterval(() => {
-                if (!firstAssistantDeltaAt) {
+                if (!firstAssistantDeltaAt && !turnUsage) {
                     process.stdout.write(renderChatThoughtTimeUpdate((Date.now() - startedAt) / 1000));
                 }
             }, 400);
             const evidenceTask = liveSearch ? buildSearchAnswerTask(effectivePrompt, liveSearch) : undefined;
-            const runAgent = () => loop.run(evidenceTask ?? createAgentTask(effectivePrompt, activeMode), {
-                workspaceRoot: canonicalWorkspace, provider: provider, tools, approvalHandler, router, policy,
+            const runAgent = () => loop.run(isReview ? createReviewTask(reviewFiles) : evidenceTask ?? createAgentTask(effectivePrompt, activeMode), {
+                workspaceRoot: canonicalWorkspace, provider: provider, tools: isReview ? createReviewToolRegistry() : tools, approvalHandler, router, policy,
                 maxOutputTokens: args.maxTokens ?? config.maxOutputTokens ?? DEFAULT_MAX_OUTPUT_TOKENS,
                 skillContext: skillContext(),
                 routeOptions: { pinnedModelId: currentModel === 'auto' ? undefined : currentModel, allowPaid: config.allowPaid ?? args.allowPaid, allowUnknown: config.allowUnknown ?? args.allowUnknown, isLocalProfile: args.profile.includes('local') },
                 eventListener: (event) => {
+                    if (event.type === 'model_change')
+                        usedModels.add(event.newModelId);
+                    if (event.type === 'usage') {
+                        turnUsage = event;
+                        lastTokenUsage = event;
+                        outputTokenRate = event.generationMs > 0 ? event.outputTokensPerSecond : undefined;
+                        const budgetState = evaluateTokenBudget(event.usage.totalTokens, tokenBudget);
+                        const firstBudgetWarning = budgetState === 'warning' && !budgetWarningIssued;
+                        if (budgetState === 'warning')
+                            budgetWarningIssued = true;
+                        if (budgetState === 'stop' && !budgetStopped) {
+                            budgetStopped = true;
+                            requestAbort.abort();
+                        }
+                        if (budgetStopped || firstBudgetWarning || event.final || event.timestamp - lastUsagePaint >= 400) {
+                            process.stdout.write(renderChatUsageUpdate(event, process.stdout.columns || 80, budgetStopped ? 'Budget reached' : budgetWarningIssued ? 'Budget ≥80%' : undefined));
+                            lastUsagePaint = event.timestamp;
+                        }
+                    }
                     if (event.type === 'assistant_delta') {
                         if (!firstAssistantDeltaAt && event.delta.trim().length > 0) {
                             firstAssistantDeltaAt = Date.now();
@@ -1421,28 +1536,38 @@ export async function handleChatSession(args, version, signal) {
                 const replacement = await replaceProviderKeyInteractive(activeConnection, { signal });
                 if (!replacement)
                     throw error;
+                recordTurnUsage();
+                turnUsage = undefined;
+                usageRecorded = false;
                 const secured = process.platform === 'win32' ? await storeConnectionCredential(replacement, credentialStore) : replacement;
                 saveConnection(secured);
                 config = loadConfig();
                 activateConnection(secured);
                 result = await runAgent();
             }
-            if (result.usage) {
+            if (turnUsage && !turnUsage.estimated) {
                 let pricing;
                 try {
-                    pricing = findModelPricing(await provider.discoverModels(signal), result.selectedModel.id);
+                    if (usedModels.size <= 1)
+                        pricing = findModelPricing(await provider.discoverModels(signal), result.selectedModel.id);
                 }
                 catch {
                     // A valid answer remains usable when the provider catalog is unavailable.
                 }
-                activeSession.usage = { ...result.usage, ...calculateSessionCost(result.usage, pricing), available: true };
-                sessionTokens = result.usage.totalTokens;
-                if (firstAssistantDeltaAt)
-                    outputTokenRate = calculateOutputTokenRate(result.usage.completionTokens, Date.now() - firstAssistantDeltaAt);
+                recordTurnUsage(pricing);
             }
+            recordTurnUsage();
+            // Failed/cancelled tasks can still consume tokens, even without a saved answer.
+            if (turnUsage)
+                sessionStore.save(activeSession);
             const otherConnections = Object.keys(config.connections ?? {}).filter((id) => id !== activeConnection?.id);
             const resolution = resolveTurnAssistantAnswer(result, conversationHistory.length, lastErrorEvent, streamedAnswer, currentModel, otherConnections);
             lastAnswer = resolution.answer;
+            if (budgetStopped) {
+                const note = `Token budget reached (${tokenBudget?.toLocaleString()} per task). Task stopped; usage may exceed the limit because providers report in chunks.`;
+                lastAnswer = [lastAnswer, note].filter(Boolean).join('\n\n');
+                streamedAnswer = lastAnswer;
+            }
             if (resolution.isError) {
                 streamedAnswer = lastAnswer;
             }
@@ -1469,6 +1594,9 @@ export async function handleChatSession(args, version, signal) {
             if (thinkingTimer)
                 clearInterval(thinkingTimer);
             removeGenerationListener?.();
+            recordTurnUsage();
+            if (turnUsage)
+                sessionStore.save(activeSession);
             const otherConnections = Object.keys(config.connections ?? {}).filter((id) => id !== activeConnection?.id);
             const altSuggestion = otherConnections.length > 0
                 ? ` (other configured providers: ${otherConnections.join(', ')})`
@@ -1486,13 +1614,14 @@ export async function handleChatSession(args, version, signal) {
                 model: currentModel ?? 'No model connected — use /connect',
                 tokens: Math.round(sessionTokens),
                 cost: costLabel(),
-                usageAvailable: activeSession.usage.available,
+                usageAvailable: activeSession.usage.available, usageEstimated: activeSession.usage.estimated,
                 workspace: canonicalWorkspace,
                 mode: activeMode,
                 autoApprove: activeAutoApprove,
                 chatQuestion: lastQuestion,
                 chatAnswer: lastAnswer,
                 chatThoughtTime: lastThoughtTime,
+                tokenUsage: turnUsage,
                 queuedCommands: commandQueue.items,
             }, process.stdout.rows));
             process.stdout.write(renderChatComposerCursor({ width: process.stdout.columns }, 0));