cawdex 1.35.68 → 1.35.73

package/dist/modes.js.map

@@ -1 +1 @@
-{"version":3,"file":"modes.js","sourceRoot":"","sources":["../src/modes.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAaH,MAAM,CAAC,MAAM,KAAK,GAA6B;IAC7C,GAAG,EAAE;QACH,IAAI,EAAE,KAAK;QACX,KAAK,EAAE,aAAa;QACpB,WAAW,EAAE,qDAAqD;QAClE,oBAAoB,EAAE;;;;;;;oCAOU;QAChC,cAAc,EAAE,CAAC,MAAM,EAAE,WAAW,EAAE,WAAW,EAAE,YAAY,EAAE,MAAM,EAAE,MAAM,CAAC;KACjF;IAED,MAAM,EAAE;QACN,IAAI,EAAE,QAAQ;QACd,KAAK,EAAE,aAAa;QACpB,WAAW,EAAE,oDAAoD;QACjE,oBAAoB,EAAE;;;;;;;;;;uFAU6D;QACnF,cAAc,EAAE,CAAC,WAAW,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC;KACtD;IAED,GAAG,EAAE;QACH,IAAI,EAAE,KAAK;QACX,KAAK,EAAE,yBAAyB;QAChC,WAAW,EAAE,wCAAwC;QACrD,oBAAoB,EAAE;;;;;;;;;;;4FAWkE;QACxF,cAAc,EAAE,CAAC,MAAM,EAAE,YAAY,EAAE,WAAW,EAAE,WAAW,CAAC;QAChE,WAAW,EAAE,GAAG;KACjB;IAED,QAAQ,EAAE;QACR,IAAI,EAAE,UAAU;QAChB,KAAK,EAAE,UAAU;QACjB,WAAW,EAAE,kDAAkD;QAC/D,oBAAoB,EAAE;;;;;;;;;gDASsB;QAC5C,cAAc,EAAE,CAAC,WAAW,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU,EAAE,WAAW,CAAC;QACtE,WAAW,EAAE,GAAG;KACjB;IAED,IAAI,EAAE;QACJ,IAAI,EAAE,MAAM;QACZ,KAAK,EAAE,UAAU;QACjB,WAAW,EAAE,2CAA2C;QACxD,oBAAoB,EAAE;;;;;;;;;;iEAUuC;QAC7D,cAAc,EAAE,CAAC,WAAW,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU,CAAC;QACzD,WAAW,EAAE,GAAG;KACjB;IAED,KAAK,EAAE;QACL,IAAI,EAAE,OAAO;QACb,KAAK,EAAE,OAAO;QACd,WAAW,EAAE,gCAAgC;QAC7C,oBAAoB,EAAE;;;;;;;;;;;2CAWiB;QACvC,cAAc,EAAE,CAAC,MAAM,EAAE,WAAW,EAAE,MAAM,EAAE,MAAM,EAAE,WAAW,CAAC;QAClE,WAAW,EAAE,GAAG;KACjB;IAED,SAAS,EAAE;QACT,IAAI,EAAE,WAAW;QACjB,KAAK,EAAE,WAAW;QAClB,WAAW,EAAE,iGAAiG;QAC9G,oBAAoB,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;sEAyD4C;QAClE,cAAc,EAAE,CAAC,MAAM,EAAE,mBAAmB,EAAE,YAAY,EAAE,WAAW,EAAE,WAAW,EAAE,aAAa,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU,EAAE,eAAe,EAAE,kBAAkB,EAAE,wBAAwB,EAAE,oBAAoB,CAAC;QACrN,WAAW,EAAE,GAAG;KACjB;IAED,SAAS,EAAE;QACT,IAAI,EAAE,WAAW;QACjB,KAAK,EAAE,WAAW;QAClB,WAAW,EAAE,0CAA0C;QACvD,oBAAoB,EAAE;;;;;;;;;;;;;0DAagC;QACtD,cAAc,EAAE,CAAC,WAAW,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU,EAAE,WAAW,CAAC;QACtE,WAAW,EAAE,GAAG;KACjB;IAED,MAAM,EAAE;QACN,IAAI,EAAE,QAAQ;QACd,KAAK,EAAE,wBAAwB;QAC/B,WAAW,EAAE,+JAA+J;QAC5K,oBAAoB,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;4FAoEkE;QACxF,cAAc,EAAE,CAAC,MAAM,EAAE,WAAW,EAAE,WAAW,EAAE,YAAY,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU,EAAE,WAAW,CAAC;QACzG,WAAW,EAAE,GAAG;KACjB;IAED,MAAM,EAAE;QACN,IAAI,EAAE,QAAQ;QACd,KAAK,EAAE,yBAAyB;QAChC,WAAW,EAAE,iJAAiJ;QAC9J,oBAAoB,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iEA+JuC;QAC7D,cAAc,EAAE,CAAC,MAAM,EAAE,WAAW,EAAE,WAAW,EAAE,YAAY,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU,EAAE,WAAW,EAAE,QAAQ,CAAC;QACnH,WAAW,EAAE,GAAG;KACjB;CACF,CAAC;AAEF,MAAM,UAAU,OAAO,CAAC,IAAY;IAClC,OAAO,KAAK,CAAC,IAAY,CAAC,CAAC;AAC7B,CAAC;AAED,MAAM,UAAU,qBAAqB,CAAC,IAAU;IAC9C,OAAO,KAAK,CAAC,IAAI,CAAC,EAAE,oBAAoB,IAAI,EAAE,CAAC;AACjD,CAAC;AAED,MAAM,UAAU,SAAS;IACvB,OAAO,MAAM,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;AAC9B,CAAC"}
+{"version":3,"file":"modes.js","sourceRoot":"","sources":["../src/modes.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAaH,MAAM,CAAC,MAAM,KAAK,GAA6B;IAC7C,GAAG,EAAE;QACH,IAAI,EAAE,KAAK;QACX,KAAK,EAAE,aAAa;QACpB,WAAW,EAAE,qDAAqD;QAClE,oBAAoB,EAAE;;;;;;;oCAOU;QAChC,cAAc,EAAE,CAAC,MAAM,EAAE,WAAW,EAAE,WAAW,EAAE,YAAY,EAAE,MAAM,EAAE,MAAM,CAAC;KACjF;IAED,MAAM,EAAE;QACN,IAAI,EAAE,QAAQ;QACd,KAAK,EAAE,aAAa;QACpB,WAAW,EAAE,oDAAoD;QACjE,oBAAoB,EAAE;;;;;;;;;;uFAU6D;QACnF,cAAc,EAAE,CAAC,WAAW,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC;KACtD;IAED,GAAG,EAAE;QACH,IAAI,EAAE,KAAK;QACX,KAAK,EAAE,yBAAyB;QAChC,WAAW,EAAE,wCAAwC;QACrD,oBAAoB,EAAE;;;;;;;;;;;4FAWkE;QACxF,cAAc,EAAE,CAAC,MAAM,EAAE,YAAY,EAAE,WAAW,EAAE,WAAW,CAAC;QAChE,WAAW,EAAE,GAAG;KACjB;IAED,QAAQ,EAAE;QACR,IAAI,EAAE,UAAU;QAChB,KAAK,EAAE,UAAU;QACjB,WAAW,EAAE,kDAAkD;QAC/D,oBAAoB,EAAE;;;;;;;;;gDASsB;QAC5C,cAAc,EAAE,CAAC,WAAW,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU,EAAE,WAAW,CAAC;QACtE,WAAW,EAAE,GAAG;KACjB;IAED,IAAI,EAAE;QACJ,IAAI,EAAE,MAAM;QACZ,KAAK,EAAE,UAAU;QACjB,WAAW,EAAE,2CAA2C;QACxD,oBAAoB,EAAE;;;;;;;;;;iEAUuC;QAC7D,cAAc,EAAE,CAAC,WAAW,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU,CAAC;QACzD,WAAW,EAAE,GAAG;KACjB;IAED,KAAK,EAAE;QACL,IAAI,EAAE,OAAO;QACb,KAAK,EAAE,OAAO;QACd,WAAW,EAAE,gCAAgC;QAC7C,oBAAoB,EAAE;;;;;;;;;;;2CAWiB;QACvC,cAAc,EAAE,CAAC,MAAM,EAAE,WAAW,EAAE,MAAM,EAAE,MAAM,EAAE,WAAW,CAAC;QAClE,WAAW,EAAE,GAAG;KACjB;IAED,SAAS,EAAE;QACT,IAAI,EAAE,WAAW;QACjB,KAAK,EAAE,WAAW;QAClB,WAAW,EAAE,iGAAiG;QAC9G,oBAAoB,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;sEAyD4C;QAClE,cAAc,EAAE,CAAC,MAAM,EAAE,mBAAmB,EAAE,YAAY,EAAE,WAAW,EAAE,WAAW,EAAE,aAAa,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU,EAAE,eAAe,EAAE,kBAAkB,EAAE,wBAAwB,EAAE,oBAAoB,CAAC;QACrN,WAAW,EAAE,GAAG;KACjB;IAED,SAAS,EAAE;QACT,IAAI,EAAE,WAAW;QACjB,KAAK,EAAE,WAAW;QAClB,WAAW,EAAE,0CAA0C;QACvD,oBAAoB,EAAE;;;;;;;;;;;;;0DAagC;QACtD,cAAc,EAAE,CAAC,WAAW,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU,EAAE,WAAW,CAAC;QACtE,WAAW,EAAE,GAAG;KACjB;IAED,SAAS,EAAE;QACT,IAAI,EAAE,WAAW;QACjB,KAAK,EAAE,2BAA2B;QAClC,WAAW,EAAE,wHAAwH;QACrI,oBAAoB,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;4FAoEkE;QACxF,cAAc,EAAE,CAAC,MAAM,EAAE,WAAW,EAAE,WAAW,EAAE,YAAY,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU,EAAE,WAAW,CAAC;QACzG,WAAW,EAAE,GAAG;KACjB;IAED,MAAM,EAAE;QACN,IAAI,EAAE,QAAQ;QACd,KAAK,EAAE,yBAAyB;QAChC,WAAW,EAAE,iJAAiJ;QAC9J,oBAAoB,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iEA+JuC;QAC7D,cAAc,EAAE,CAAC,MAAM,EAAE,WAAW,EAAE,WAAW,EAAE,YAAY,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU,EAAE,WAAW,EAAE,QAAQ,CAAC;QACnH,WAAW,EAAE,GAAG;KACjB;CACF,CAAC;AAEF,MAAM,UAAU,iBAAiB,CAAC,IAAY;IAC5C,MAAM,UAAU,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;IAC7C,IAAI,UAAU,KAAK,QAAQ;QAAE,OAAO,WAAW,CAAC;IAChD,OAAO,UAAU,IAAI,KAAK,CAAC,CAAC,CAAC,UAAkB,CAAC,CAAC,CAAC,IAAI,CAAC;AACzD,CAAC;AAED,MAAM,UAAU,OAAO,CAAC,IAAY;IAClC,MAAM,IAAI,GAAG,iBAAiB,CAAC,IAAI,CAAC,CAAC;IACrC,OAAO,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;AACxC,CAAC;AAED,MAAM,UAAU,qBAAqB,CAAC,IAAU;IAC9C,OAAO,OAAO,CAAC,IAAI,CAAC,EAAE,oBAAoB,IAAI,EAAE,CAAC;AACnD,CAAC;AAED,MAAM,UAAU,SAAS;IACvB,OAAO,MAAM,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;AAC9B,CAAC"}

package/dist/query.d.ts

@@ -2,6 +2,7 @@ import * as readline from 'node:readline/promises';
 import type { Message, CawdexConfig } from './types.js';
 import type { Mode } from './modes.js';
 export declare function resolveFirstTokenTimeoutMs(config: Pick<CawdexConfig, 'model' | 'provider'>): number;
+export declare function resolveStreamIdleTimeoutMs(fastDirect?: boolean): number;
 export declare function isKnownFlakyOpenRouterModel(config: Pick<CawdexConfig, 'model' | 'provider'>): boolean;
 export declare function fallbackModelForKnownFlakyTurn(config: CawdexConfig, usedFallbackModel?: boolean): string | null;
 export declare function isTurnCancelKeySequence(chunk: Buffer): boolean;

package/dist/query.js

@@ -10,13 +10,14 @@ import { runHooks } from './hooks.js';
 import { scanToolCall, printSecurityWarning } from './security.js';
 import { trackUsage } from './cost-tracker.js';
 import { shouldCompact, compactMessages, quickCompact, buildCompactionConfig, contextCapTokens, enforceContextCap, inferContextWindowTokens, } from './compaction.js';
-import { theme, sym, printToolRun, printToolResult, printThinkingOpen, printThinkingText, printThinkingClose, printCost, printApiError, formatDuration, categorizeApiError } from './theme.js';
+import { assistantTranscriptPrefix, theme, sym, printToolRun, printToolResult, printThinkingOpen, printThinkingText, printThinkingClose, printCost, printApiError, formatDuration, categorizeApiError } from './theme.js';
 import { isVoiceEnabled, getTtsConfig, getAccessibilityConfig, speakAssistantResponse, speak, speakUserEcho, } from './voice.js';
 import { isLikelyDestructive, describeDestructive, countWords, summarize } from './accessibility.js';
 import { audioCue } from './audio.js';
 import { setStatus } from './status.js';
 import { collapseCompletedTurns } from './turn-context.js';
 import * as liveQueue from './live-queue.js';
+import { isFooterActive, setFooterActivity, setFooterCost, writeScrollableLine } from './fixed-footer.js';
 import { applyQueuedInputChunk, drainQueuedInputBytes, queuedInputBytesToText } from './prompt-buffer.js';
 import { emit as dbgEmit } from './debug.js';
 import { buildBenchmarkCompletionReminder, buildBenchmarkTrajectorySystemBlock, makeBenchmarkInvalidToolActionEvent, makeBenchmarkTraceEvent, writeBenchmarkTrace, } from './benchmark-trace.js';
@@ -34,6 +35,9 @@ const INTERACTIVE_FLAKY_FIRST_TOKEN_TIMEOUT_MS = 6_000;
 const NON_INTERACTIVE_FIRST_TOKEN_TIMEOUT_MS = 60_000;
 const NON_INTERACTIVE_FLAKY_FIRST_TOKEN_TIMEOUT_MS = 20_000;
 const FAST_DIRECT_FIRST_TOKEN_TIMEOUT_MS = 8_000;
+const INTERACTIVE_STREAM_IDLE_TIMEOUT_MS = 45_000;
+const NON_INTERACTIVE_STREAM_IDLE_TIMEOUT_MS = 120_000;
+const FAST_DIRECT_STREAM_IDLE_TIMEOUT_MS = 20_000;
 const KNOWN_FLAKY_OPENROUTER_MODEL_PATTERNS = [
     'owl-alpha',
     'horizon-alpha',
@@ -41,6 +45,7 @@ const KNOWN_FLAKY_OPENROUTER_MODEL_PATTERNS = [
     'optimus-alpha',
     'quasar-alpha',
     'deepseek-v4-flash',
+    'deepseek-v4-pro',
 ];
 function envTimeoutMs(name, fallback) {
     const raw = process.env[name];
@@ -59,6 +64,14 @@ export function resolveFirstTokenTimeoutMs(config) {
         : (flaky ? INTERACTIVE_FLAKY_FIRST_TOKEN_TIMEOUT_MS : INTERACTIVE_FIRST_TOKEN_TIMEOUT_MS);
     return envTimeoutMs('CAWDEX_FIRST_TOKEN_TIMEOUT_MS', fallback);
 }
+export function resolveStreamIdleTimeoutMs(fastDirect = false) {
+    const fallback = process.env.CAWDEX_NON_INTERACTIVE === '1'
+        ? NON_INTERACTIVE_STREAM_IDLE_TIMEOUT_MS
+        : fastDirect
+            ? FAST_DIRECT_STREAM_IDLE_TIMEOUT_MS
+            : INTERACTIVE_STREAM_IDLE_TIMEOUT_MS;
+    return envTimeoutMs('CAWDEX_STREAM_IDLE_TIMEOUT_MS', fallback);
+}
 export function isKnownFlakyOpenRouterModel(config) {
     const model = String(config.model || '').toLowerCase();
     const provider = String(config.provider || '').toLowerCase();
@@ -80,7 +93,10 @@ export function fallbackModelForKnownFlakyTurn(config, usedFallbackModel = false
 }
 export function isTurnCancelKeySequence(chunk) {
     const seq = chunk.toString('utf8');
-    return (seq === '\x1b[15~' ||
+    return (seq === '\x1b' ||
+        seq === '\x1b\x1b' ||
+        /^\x1b\[27(?:;\d+)*~$/.test(seq) ||
+        seq === '\x1b[15~' ||
         seq === '\x1b[15;2~' ||
         seq === '\x1b[15;5~' ||
         seq === '\x1b[15;6~' ||
@@ -90,6 +106,10 @@ const FAST_DIRECT_SYSTEM_PROMPT = 'You are Cawdex. Answer the user directly and
     'Do not mention tools, repositories, or implementation steps unless the user asks.';
 const FAST_DIRECT_POSITIVE = [
     /^(hi|hello|hey|thanks|thank you)\b/i,
+    /^(?:i\s+)?(?:need|want)\s+(?:your\s+)?help\b/i,
+    /^(?:help me|can you help|could you help|would you help)\b/i,
+    /^(?:i|we)\s+(?:like|love|hate|prefer|think|feel|am|are)\b/i,
+    /^(?:sorry,?\s*)?(?:make|rewrite|revise|adjust|redo|try)\s+(?:it|that|this)\b/i,
     /^(what|who|when|where|why|how|which)\b/i,
     /^(?:please\s+)?(?:can|could|would|will)\s+you\s+(?:please\s+)?(?:explain|define|summarize|translate|calculate|compute|write|draft|make|create|tell|give)\b/i,
     /^(?:please\s+)?(?:tell|give)\s+me\b/i,
@@ -125,14 +145,20 @@ function printInteractiveTurnAccepted(config) {
         return;
     if (!process.stdout.isTTY)
         return;
-    console.log(theme.dim(`  submitted to ${config.provider} · ${config.model}. Waiting for the first model event; Esc or F5 cancels.`));
+    const line = theme.dim(`  submitted to ${config.provider} · ${config.model}. Waiting for model events; Esc or F5 cancels.`);
+    if (isFooterActive()) {
+        writeScrollableLine(line);
+    }
+    else {
+        console.log(line);
+    }
 }
 export function formatWorkingIndicatorFrame(elapsedMs, frameIndex = 0, message = 'Working') {
-    const frames = ['◦', '◇', '◆', '◇'];
+    const frames = ['\u25e6', '\u25c7', '\u25c6', '\u25c7'];
     const frame = frames[Math.abs(frameIndex) % frames.length];
-    return `  ${frame} ${message} (${formatDuration(elapsedMs)} • esc to interrupt)`;
+    return `  ${frame} ${message} (${formatDuration(elapsedMs)} \u2022 esc to interrupt)`;
 }
-function startWorkingIndicator(startedAtMs, screenReader) {
+function startWorkingIndicator(startedAtMs, screenReader, turn = 0) {
     if (screenReader)
         return null;
     if (process.env.CAWDEX_NON_INTERACTIVE === '1')
@@ -140,7 +166,7 @@ function startWorkingIndicator(startedAtMs, screenReader) {
     if (!process.stdout.isTTY)
         return null;
     const messages = [
-        'Working',
+        'Working hard on your request',
         'Sumi ink settling',
         'Edo lanterns lit',
         'Kamon crest aligned',
@@ -149,7 +175,13 @@ function startWorkingIndicator(startedAtMs, screenReader) {
     let frame = 0;
     let stopped = false;
     const paint = () => {
-        const line = formatWorkingIndicatorFrame(Date.now() - startedAtMs, frame, messages[Math.floor(frame / 8) % messages.length]);
+        const message = messages[Math.floor(frame / 8) % messages.length];
+        if (isFooterActive()) {
+            setFooterActivity(message, turn, startedAtMs);
+            frame++;
+            return;
+        }
+        const line = formatWorkingIndicatorFrame(Date.now() - startedAtMs, frame, message);
         process.stdout.write('\r\x1b[K' + theme.dim(line));
         frame++;
     };
@@ -163,7 +195,10 @@ function startWorkingIndicator(startedAtMs, screenReader) {
                 return;
             stopped = true;
             clearInterval(timer);
-            process.stdout.write('\r\x1b[K');
+            if (isFooterActive())
+                setFooterActivity('Receiving response', turn, startedAtMs);
+            else
+                process.stdout.write('\r\x1b[K');
         },
     };
 }
@@ -278,6 +313,15 @@ function startInputSuppression(screenReader = false) {
             }
             return;
         }
+        if (detached && isTurnCancelKeySequence(chunk)) {
+            if (steerHandler) {
+                try {
+                    steerHandler();
+                }
+                catch { /* never break input on a cancel error */ }
+            }
+            return;
+        }
         // Esc (0x1B) → Steer trigger, but only for a BARE Esc (chunk of
         // exactly one byte). Multi-byte chunks starting with 0x1B are ANSI
         // escape sequences for arrow keys / function keys / Alt+letter, and
@@ -438,7 +482,7 @@ export function buildStateBlock(messages) {
     const olderCount = actions.length - recent.length;
     const lines = [
         '<task_state>',
-        `Original goal: ${goal}${goal.length >= STATE_BLOCK_GOAL_MAX_CHARS ? '…' : ''}`,
+        `Original goal: ${goal}${goal.length >= STATE_BLOCK_GOAL_MAX_CHARS ? '\u2026' : ''}`,
         `Actions completed: ${actions.length}`,
     ];
     if (olderCount > 0) {
@@ -448,7 +492,7 @@ export function buildStateBlock(messages) {
         lines.push(`Actions:`);
     }
     recent.forEach((a, i) => {
-        lines.push(`  ${i + 1}. ${a.tool}(${a.argsPreview}${a.argsPreview.length >= 80 ? '…' : ''})`);
+        lines.push(`  ${i + 1}. ${a.tool}(${a.argsPreview}${a.argsPreview.length >= 80 ? '\u2026' : ''})`);
     });
     lines.push('');
     lines.push('Stay focused on the goal. Do not re-issue actions you have already completed — refer to their results in the conversation above.');
@@ -1091,7 +1135,7 @@ export async function runQuery(ctx) {
         const failedModel = ctx.config.model;
         ctx.config.model = immediateFallback;
         resetClient();
-        console.log(theme.warning(`  ${sym.warn} ${failedModel} is a known-stuck OpenRouter preview model; switching this turn to ${immediateFallback}.`));
+        console.log(theme.warning(`  ${sym.warn} ${failedModel} is known to stall in interactive OpenRouter sessions; switching this turn to ${immediateFallback}.`));
         console.log(theme.dim('    Override only if you really want it: CAWDEX_ALLOW_FLAKY_MODELS=1'));
     }
     if (!chainFastDirect) {
@@ -1103,12 +1147,12 @@ export async function runQuery(ctx) {
     // because we only care about "in this whole exchange did we see any
     // reasoning at all".
     let sawAnyThinking = false;
-    // Skill-graduation telemetry. The Hermes audit's deterministic rule
+    // Skill-graduation telemetry. The Sentience audit's deterministic rule
     // for "this work was worth remembering" — the model is a bad judge of
     // its own complexity, so the dispatcher counts and decides. Thresholds:
     //   - 5+ tool calls in this chain   → complex task
     //   - any tool errored then recovered → learned-from-failure
-    // Only used to inform a chain-end suggestion in hermes mode; we don't
+    // Only used to inform a chain-end suggestion in sentience mode; we don't
     // auto-create skills (which would burn an extra LLM call). We surface
     // the opportunity with a one-line nudge so the user can /skill-create
     // or /learn if they want.
@@ -1278,6 +1322,7 @@ export async function runQuery(ctx) {
             let hasOutput = false;
             let thinkingActive = false;
             let leadingTrimmed = false; // strip leading whitespace from the model's first text chunk
+            let assistantPrefixed = false;
             let lastCharWasNewline = false; // collapse 3+ consecutive newlines down to 2
             let consecutiveNewlines = 0;
             const turnStart = Date.now();
@@ -1320,6 +1365,10 @@ export async function runQuery(ctx) {
                 if (out.length === 0)
                     return;
                 lastCharWasNewline = out.endsWith('\n');
+                if (!assistantPrefixed) {
+                    process.stdout.write(assistantTranscriptPrefix());
+                    assistantPrefixed = true;
+                }
                 process.stdout.write(theme.primary(out));
                 fullText += out;
                 // ── Loop detection ────────────────────────────────────
@@ -1406,7 +1455,7 @@ export async function runQuery(ctx) {
             // Live waiting indicator on the response line. It keeps elapsed time
             // and the interrupt hint visible while still clearing itself before
             // the first model event writes real output.
-            let workingIndicator = startWorkingIndicator(turnStart, isScreenReader);
+            let workingIndicator = startWorkingIndicator(turnStart, isScreenReader, turns);
             // Slow-model warning and first-token watchdog. The warning is a
             // UX hint; the watchdog is the hard recovery path for providers
             // that accept a request but then never produce a stream event.
@@ -1418,7 +1467,7 @@ export async function runQuery(ctx) {
                     workingIndicator = null;
                     console.log(chalk.yellow(`  ⏳ model is taking longer than 30s. Shift+F5 cancels, Ctrl+C exits. Often means the model returned no tokens (try /model <other> if this hangs).`));
                     if (!firstTokenSeen) {
-                        workingIndicator = startWorkingIndicator(turnStart, isScreenReader);
+                        workingIndicator = startWorkingIndicator(turnStart, isScreenReader, turns);
                     }
                 }
             }, 30_000);
@@ -1426,30 +1475,63 @@ export async function runQuery(ctx) {
             const firstTokenTimeoutMs = fastDirect
                 ? Math.min(resolveFirstTokenTimeoutMs(ctx.config), FAST_DIRECT_FIRST_TOKEN_TIMEOUT_MS)
                 : resolveFirstTokenTimeoutMs(ctx.config);
-            const firstTokenTimer = firstTokenTimeoutMs > 0
-                ? setTimeout(() => {
-                    if (!firstTokenSeen) {
-                        firstTokenTimedOut = true;
-                        try {
-                            streamAbort.abort();
-                        }
-                        catch { /* noop */ }
-                    }
-                }, firstTokenTimeoutMs)
-                : null;
+            let streamIdleTimedOut = false;
+            const streamIdleTimeoutMs = resolveStreamIdleTimeoutMs(fastDirect);
+            let streamWaitTimer = null;
             try {
                 const requestConfig = fastDirect
                     ? { ...ctx.config, maxTokens: Math.min(ctx.config.maxTokens ?? 700, 700) }
                     : ctx.config;
                 const requestTools = fastDirect ? [] : ALL_TOOLS;
-                for await (const event of streamChat(requestConfig, apiMessages, requestTools, streamAbort.signal)) {
+                const stream = streamChat(requestConfig, apiMessages, requestTools, streamAbort.signal);
+                const iterator = stream[Symbol.asyncIterator]();
+                while (true) {
+                    const nextPromise = iterator.next();
+                    let next;
+                    const waitTimeoutMs = !firstTokenSeen ? firstTokenTimeoutMs : streamIdleTimeoutMs;
+                    if (waitTimeoutMs > 0) {
+                        const timeoutPromise = new Promise((_, reject) => {
+                            streamWaitTimer = setTimeout(() => {
+                                try {
+                                    streamAbort.abort();
+                                }
+                                catch { /* noop */ }
+                                if (!firstTokenSeen) {
+                                    firstTokenTimedOut = true;
+                                    reject(new Error('__CAWDEX_FIRST_TOKEN_TIMEOUT__'));
+                                }
+                                else {
+                                    streamIdleTimedOut = true;
+                                    reject(new Error('__CAWDEX_STREAM_IDLE_TIMEOUT__'));
+                                }
+                            }, waitTimeoutMs);
+                        });
+                        try {
+                            next = await Promise.race([nextPromise, timeoutPromise]);
+                        }
+                        catch (err) {
+                            nextPromise.catch(() => { });
+                            throw err;
+                        }
+                        finally {
+                            if (streamWaitTimer)
+                                clearTimeout(streamWaitTimer);
+                            streamWaitTimer = null;
+                        }
+                    }
+                    else {
+                        next = await nextPromise;
+                    }
+                    if (next.done)
+                        break;
+                    const event = next.value;
                     // First event of any kind — model is alive. Cancel the slow-
                     // model warning timer; subsequent events are normal streaming.
                     if (!firstTokenSeen) {
                         firstTokenSeen = true;
                         clearTimeout(slowTimer);
-                        if (firstTokenTimer)
-                            clearTimeout(firstTokenTimer);
+                        if (streamWaitTimer)
+                            clearTimeout(streamWaitTimer);
                         // Tear down the live waiting indicator so the next print
                         // (thinking header, response text, or tool call) lands cleanly.
                         workingIndicator?.stop();
@@ -1499,17 +1581,24 @@ export async function runQuery(ctx) {
                                 totalTokens: u.total || u.prompt + u.completion,
                                 estimatedCostUsd: cost,
                             });
+                            setFooterCost(cost, u.prompt, u.completion);
                             // Single newline separator if we just streamed text, then the
                             // compact telemetry line.
                             if (hasOutput && !lastCharWasNewline)
                                 process.stdout.write('\n');
                             printCost(u.prompt, u.completion, cost, warning, Date.now() - turnStart);
                         }
+                        try {
+                            streamAbort.abort();
+                        }
+                        catch { /* close any provider socket left open after done */ }
+                        void iterator.return?.(undefined).catch(() => { });
+                        break;
                     }
                 }
                 clearTimeout(slowTimer);
-                if (firstTokenTimer)
-                    clearTimeout(firstTokenTimer);
+                if (streamWaitTimer)
+                    clearTimeout(streamWaitTimer);
                 workingIndicator?.stop();
                 workingIndicator = null;
             }
@@ -1518,8 +1607,8 @@ export async function runQuery(ctx) {
                 // timer so its 30s callback doesn't fire after the error is
                 // already on stdout (would look like a false positive).
                 clearTimeout(slowTimer);
-                if (firstTokenTimer)
-                    clearTimeout(firstTokenTimer);
+                if (streamWaitTimer)
+                    clearTimeout(streamWaitTimer);
                 // Tear down the waiting indicator if it's still ticking — error
                 // print below shouldn't trail an animated row.
                 workingIndicator?.stop();
@@ -1542,10 +1631,40 @@ export async function runQuery(ctx) {
                     }
                     const timeoutMsg = `${failedModel} produced no stream events for ${formatDuration(firstTokenTimeoutMs)}`;
                     console.log(theme.error(`  ${sym.warn} ${timeoutMsg}.`));
-                    console.log(theme.dim('    The request was cancelled so the prompt can recover. Try /openrouter-free or /model <known-good-model>.'));
+                    console.log(theme.dim('    No fallback is configured, so the request was cancelled and the prompt can recover.'));
+                    console.log(theme.dim('    Use /fallback <model-id> to enable automatic retry, or /model <known-good-model> to switch primary models.'));
                     ctx.messages.push({ role: 'assistant', content: `[Provider timeout: ${timeoutMsg}]` });
                     break;
                 }
+                if (streamIdleTimedOut) {
+                    const failedModel = ctx.config.model;
+                    const timeoutMsg = `${failedModel} stream stalled for ${formatDuration(streamIdleTimeoutMs)} after the first event`;
+                    const fallback = !hasOutput && !toolCalls
+                        ? fallbackModelForTurn(ctx.config, usedFallbackModel)
+                        : null;
+                    if (fallback) {
+                        usedFallbackModel = true;
+                        ctx.config.model = fallback;
+                        resetClient();
+                        console.log(theme.warning(`  ${sym.warn} ${timeoutMsg} — retrying once with fallback model ${fallback}.`));
+                        console.log(theme.dim('    Configure with /fallback <model-id>, disable with /fallback off, or switch now with /openrouter-free.'));
+                        turns--;
+                        continue;
+                    }
+                    if (fullText.trim()) {
+                        const partial = `${fullText.trimEnd()}\n[Provider timeout: stream stalled before completion]`;
+                        console.log(theme.warning(`  ${sym.warn} stream stalled before completion; returning the partial response.`));
+                        accumulatedAssistantText += (accumulatedAssistantText ? '\n\n' : '') + partial;
+                        ctx.messages.push({ role: 'assistant', content: partial });
+                    }
+                    else {
+                        console.log(theme.error(`  ${sym.warn} ${timeoutMsg}.`));
+                        console.log(theme.dim('    The request was cancelled so the prompt can recover instead of hanging indefinitely.'));
+                        console.log(theme.dim('    Use /fallback <model-id> to enable automatic retry, or /model <known-good-model> to switch primary models.'));
+                        ctx.messages.push({ role: 'assistant', content: `[Provider timeout: ${timeoutMsg}]` });
+                    }
+                    break;
+                }
                 // ── User cancel path (graceful — not an error) ─────────
                 // If the user pressed Ctrl+G / Esc / F5 during streaming, the
                 // AbortController fired and the OpenAI SDK threw something like
@@ -1648,7 +1767,8 @@ export async function runQuery(ctx) {
                 }
                 const emptyMsg = `${failedModel} returned an empty response`;
                 console.log(theme.error(`  ${sym.warn} ${emptyMsg}.`));
-                console.log(theme.dim('    Try /openrouter-free or /model <known-good-model>.'));
+                console.log(theme.dim('    No fallback is configured, so this empty response was not retried.'));
+                console.log(theme.dim('    Use /fallback <model-id> to enable automatic retry, or /model <known-good-model> to switch primary models.'));
                 ctx.messages.push({ role: 'assistant', content: `[Provider empty response: ${emptyMsg}]` });
                 break;
             }
@@ -1810,10 +1930,10 @@ export async function runQuery(ctx) {
             console.log(theme.dim(`         Reasoning models (DeepSeek-R1, o1, Claude with extended thinking) emit them; most general-purpose models don't.`));
             console.log(theme.dim(`         Hide this hint with /thinking (toggles off).`));
         }
-        // ── Skill graduation hint (Hermes audit, M2 item 2) ─────────
+        // ── Skill graduation hint (Sentience audit, M2 item 2) ─────────
         // Deterministic "this work was worth remembering" trigger. The model
-        // is a bad judge of its own complexity (Hermes audit's exact wording);
-        // we count instead. Fires at most once per chain in hermes mode, and
+        // is a bad judge of its own complexity (Sentience audit's exact wording);
+        // we count instead. Fires at most once per chain in sentience mode, and
         // only when a clear threshold was crossed:
         //
         //   - 5+ tool calls   → complex multi-step task
@@ -1821,9 +1941,9 @@ export async function runQuery(ctx) {
         //
         // We don't auto-execute /skill-create (it would burn another LLM call
         // and might extract noise); we surface the opportunity so the user can
-        // decide. Outside hermes mode, this is silent — keeps the noise floor
+        // decide. Outside sentience mode, this is silent — keeps the noise floor
         // low for regular dev/review/debug sessions.
-        if (ctx.mode === 'hermes') {
+        if (ctx.mode === 'sentience' || ctx.mode === 'hermes') {
             const complex = chainStats.toolCallCount >= 5;
             const learnedFromFailure = chainStats.sawToolError && chainStats.sawToolRecovery;
             if (complex || learnedFromFailure) {
@@ -1832,7 +1952,7 @@ export async function runQuery(ctx) {
                     : complex
                         ? `${chainStats.toolCallCount} tools — complex enough that a learned pattern might save time next time`
                         : `recovered from a tool error — the working path is worth banking`;
-                console.log(theme.dim(`  [hermes] graduation candidate: ${reason}.`));
+                console.log(theme.dim(`  [sentience] graduation candidate: ${reason}.`));
                 console.log(theme.dim(`           Run /skill-create or /learn to bank it. Skip if it was one-off.`));
             }
         }