cawdex 1.35.74 → 1.35.76

package/dist/query.js

@@ -9,7 +9,7 @@ import { buildSystemPrompt } from './system-prompt.js';
 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 { shouldCompact, compactMessages, quickCompact, estimateTokens, buildCompactionConfig, contextCapTokens, enforceContextCap, inferContextWindowTokens, } from './compaction.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';
@@ -20,6 +20,7 @@ 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 { applyAgentToolInstructions } from './agents-md.js';
 import { buildBenchmarkCompletionReminder, buildBenchmarkTrajectorySystemBlock, makeBenchmarkInvalidToolActionEvent, makeBenchmarkTraceEvent, writeBenchmarkTrace, } from './benchmark-trace.js';
 import { buildTodoStateBlock } from './tools/todo.js';
 import { buildRuntimeInfoBlock } from './runtime-info.js';
@@ -30,11 +31,11 @@ import { archiveLargeToolOutput } from './tool-output-archive.js';
 // not persisted — restart, see hint again. Keyed by sessionId so different
 // sessions get fresh hints.
 const _thinkingHintShownForSession = new Set();
-const INTERACTIVE_FIRST_TOKEN_TIMEOUT_MS = 12_000;
+const INTERACTIVE_FIRST_TOKEN_TIMEOUT_MS = 8_000;
 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 FAST_DIRECT_FIRST_TOKEN_TIMEOUT_MS = 5_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;
@@ -84,13 +85,6 @@ function fallbackModelForTurn(config, usedFallbackModel) {
         return null;
     return fallback;
 }
-export function fallbackModelForKnownFlakyTurn(config, usedFallbackModel = false) {
-    if (process.env.CAWDEX_ALLOW_FLAKY_MODELS === '1')
-        return null;
-    if (!isKnownFlakyOpenRouterModel(config))
-        return null;
-    return fallbackModelForTurn(config, usedFallbackModel);
-}
 export function isTurnCancelKeySequence(chunk) {
     const seq = chunk.toString('utf8');
     return (seq === '\x1b' ||
@@ -154,9 +148,10 @@ function printInteractiveTurnAccepted(config) {
     }
 }
 export function formatWorkingIndicatorFrame(elapsedMs, frameIndex = 0, message = 'Working') {
-    const frames = ['\u25e6', '\u25c7', '\u25c6', '\u25c7'];
+    void elapsedMs;
+    const frames = ['\u25dc', '\u25dd', '\u25de', '\u25df'];
     const frame = frames[Math.abs(frameIndex) % frames.length];
-    return `  ${frame} ${message} (${formatDuration(elapsedMs)} \u2022 esc to interrupt)`;
+    return `  ${frame} ${message} \u2022 Esc/F5 to interrupt`;
 }
 function startWorkingIndicator(startedAtMs, screenReader, turn = 0) {
     if (screenReader)
@@ -166,11 +161,11 @@ function startWorkingIndicator(startedAtMs, screenReader, turn = 0) {
     if (!process.stdout.isTTY)
         return null;
     const messages = [
-        'Working hard on your request',
-        'Sumi ink settling',
-        'Edo lanterns lit',
-        'Kamon crest aligned',
-        'Neon alley quiet',
+        'Sumi ink moving',
+        'Edo lanterns cycling',
+        'Kamon crest pulsing',
+        'Neon shoji breathing',
+        'Signal blade drawn',
     ];
     let frame = 0;
     let stopped = false;
@@ -1129,18 +1124,6 @@ export async function runQuery(ctx) {
     // user's configured fallbackModel. After we use it, this latches so we
     // don't bounce back and forth between failing models in a single chain.
     let usedFallbackModel = false;
-    const immediateFallback = fallbackModelForKnownFlakyTurn(ctx.config, usedFallbackModel);
-    if (immediateFallback) {
-        usedFallbackModel = true;
-        const failedModel = ctx.config.model;
-        ctx.config.model = immediateFallback;
-        resetClient();
-        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) {
-        printInteractiveTurnAccepted(ctx.config);
-    }
     // Tracks whether ANY reasoning tokens arrived across the entire chain.
     // Used at chain-end to print a one-time "/thinking is ON but this model
     // doesn't emit reasoning" hint. Hoisted to chain scope (not per-turn)
@@ -1210,8 +1193,11 @@ export async function runQuery(ctx) {
     // as new text, drowning the actual response).
     const isScreenReader = ctx.config.voice?.accessibility?.screenReader === true;
     const inputGuard = startInputSuppression(isScreenReader);
+    let earlyWorkingIndicator = null;
     try {
         if (!chainFastDirect) {
+            printInteractiveTurnAccepted(ctx.config);
+            earlyWorkingIndicator = startWorkingIndicator(chainStart, isScreenReader, 0);
             // Turn-boundary collapse runs BEFORE compaction. Every completed prior
             // turn becomes [user, "<final text>\n[Completed: used X, Y]"] — the
             // model no longer sees stale tool_calls that it might mistake for
@@ -1263,9 +1249,12 @@ export async function runQuery(ctx) {
             if (!fastDirect) {
                 replaceMessagesInPlace(ctx.messages, quickCompact(ctx.messages));
             }
+            const requestTools = fastDirect
+                ? []
+                : applyAgentToolInstructions(ALL_TOOLS, ctx.cwd, ctx.config.model);
             const systemPrompt = fastDirect
                 ? FAST_DIRECT_SYSTEM_PROMPT
-                : buildSystemPrompt(ctx.config, ctx.cwd, ctx.mode, userQuery);
+                : buildSystemPrompt(ctx.config, ctx.cwd, ctx.mode, userQuery, requestTools);
             let visibleMessages = fastDirect
                 ? (userQuery ? [{ role: 'user', content: userQuery }] : [])
                 : maskOldToolResults(ctx.messages);
@@ -1292,7 +1281,9 @@ export async function runQuery(ctx) {
             // override.
             const stateBlock = fastDirect ? null : buildStateBlock(visibleMessages);
             const runtimeInfoBlock = fastDirect ? null : buildRuntimeInfoBlock(ctx.cwd);
-            const repoMapBlock = fastDirect ? null : buildAutoRepoMapBlock(ctx.cwd, userQuery);
+            const repoMapBlock = fastDirect || ctx.mode === 'design'
+                ? null
+                : buildAutoRepoMapBlock(ctx.cwd, userQuery);
             const globalPlanBlock = fastDirect ? null : buildGlobalPlanBlock(visibleMessages);
             const todoStateBlock = fastDirect ? null : buildTodoStateBlock(ctx.cwd);
             const benchmarkTrajectoryBlock = !fastDirect && ctx.mode === 'benchmark'
@@ -1326,6 +1317,7 @@ export async function runQuery(ctx) {
             let lastCharWasNewline = false; // collapse 3+ consecutive newlines down to 2
             let consecutiveNewlines = 0;
             const turnStart = Date.now();
+            let usageRecorded = false;
             // Loop detection state: a stuck model can stream the SAME N-char
             // window of text 50+ times in a single API call (observed in the
             // wild with openrouter/owl-alpha emitting tool-call JSON as text).
@@ -1446,16 +1438,22 @@ export async function runQuery(ctx) {
             // line and then announce every subsequent token as "after the
             // waiting line", which is noisier than helpful).
             let firstTokenSeen = false;
+            let firstTokenLatencyMs = null;
             // Note: the outer `isScreenReader` declared at the top of runQuery
             // (line ~340) is in scope here via closure — no need for a second
             // declaration. Previously this re-declared inside the while loop
             // and TypeScript tolerated it as a different block scope, but it
             // was confusing and the audit flagged it as bug-bait.
             //
-            // 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, turns);
+            // Live waiting indicator on the response line. It keeps motion and the
+            // interrupt hint visible while still clearing itself before the first
+            // model event writes real output.
+            let workingIndicator = turns === 1 ? earlyWorkingIndicator : null;
+            if (turns === 1)
+                earlyWorkingIndicator = null;
+            if (!workingIndicator) {
+                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.
@@ -1482,7 +1480,6 @@ export async function runQuery(ctx) {
                 const requestConfig = fastDirect
                     ? { ...ctx.config, maxTokens: Math.min(ctx.config.maxTokens ?? 700, 700) }
                     : ctx.config;
-                const requestTools = fastDirect ? [] : ALL_TOOLS;
                 const stream = streamChat(requestConfig, apiMessages, requestTools, streamAbort.signal);
                 const iterator = stream[Symbol.asyncIterator]();
                 while (true) {
@@ -1529,6 +1526,7 @@ export async function runQuery(ctx) {
                     // model warning timer; subsequent events are normal streaming.
                     if (!firstTokenSeen) {
                         firstTokenSeen = true;
+                        firstTokenLatencyMs = Date.now() - turnStart;
                         clearTimeout(slowTimer);
                         if (streamWaitTimer)
                             clearTimeout(streamWaitTimer);
@@ -1573,7 +1571,12 @@ export async function runQuery(ctx) {
                     else if (event.type === 'done') {
                         if (event.usage) {
                             const u = event.usage;
-                            const { cost, warning } = trackUsage(ctx.sessionId, ctx.config.model, u.prompt, u.completion);
+                            const turnDurationMs = Date.now() - turnStart;
+                            const { cost, warning } = trackUsage(ctx.sessionId, ctx.config.model, u.prompt, u.completion, {
+                                provider: ctx.config.provider,
+                                firstTokenMs: firstTokenLatencyMs,
+                                durationMs: turnDurationMs,
+                            });
                             chainStats.benchmarkUsageEvents.push({
                                 model: ctx.config.model,
                                 promptTokens: u.prompt,
@@ -1581,12 +1584,16 @@ export async function runQuery(ctx) {
                                 totalTokens: u.total || u.prompt + u.completion,
                                 estimatedCostUsd: cost,
                             });
-                            setFooterCost(cost, u.prompt, u.completion);
+                            usageRecorded = true;
+                            setFooterCost(cost, u.prompt, u.completion, {
+                                firstTokenMs: firstTokenLatencyMs,
+                                durationMs: turnDurationMs,
+                            });
                             // 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);
+                            printCost(u.prompt, u.completion, cost, warning, turnDurationMs);
                         }
                         try {
                             streamAbort.abort();
@@ -1596,6 +1603,28 @@ export async function runQuery(ctx) {
                         break;
                     }
                 }
+                if (!usageRecorded && (hasOutput || (toolCalls && toolCalls.length > 0))) {
+                    const promptEstimate = Math.max(1, estimateTokens(apiMessages));
+                    const completionEstimate = Math.max(1, Math.ceil(((fullText || '') + (toolCalls ? JSON.stringify(toolCalls) : '')).length / 3.5));
+                    const turnDurationMs = Date.now() - turnStart;
+                    const { cost } = trackUsage(ctx.sessionId, ctx.config.model, promptEstimate, completionEstimate, {
+                        provider: ctx.config.provider,
+                        firstTokenMs: firstTokenLatencyMs,
+                        durationMs: turnDurationMs,
+                    });
+                    chainStats.benchmarkUsageEvents.push({
+                        model: ctx.config.model,
+                        promptTokens: promptEstimate,
+                        completionTokens: completionEstimate,
+                        totalTokens: promptEstimate + completionEstimate,
+                        estimatedCostUsd: cost,
+                    });
+                    setFooterCost(cost, promptEstimate, completionEstimate, {
+                        firstTokenMs: firstTokenLatencyMs,
+                        durationMs: turnDurationMs,
+                    });
+                    usageRecorded = true;
+                }
                 clearTimeout(slowTimer);
                 if (streamWaitTimer)
                     clearTimeout(streamWaitTimer);
@@ -1873,6 +1902,8 @@ export async function runQuery(ctx) {
         }
         // Chain ended; back to idle so F1 reports the correct state.
         setStatus({ state: 'idle' });
+        if (isFooterActive())
+            setFooterActivity('Ready', 0, null);
         // ── Voice: read the assistant's final response ────────────
         // Off the hot path — fire-and-forget so the next prompt appears
         // immediately. The playback runs in background; F2 pauses, F4 skips.
@@ -1972,6 +2003,8 @@ export async function runQuery(ctx) {
         }
     }
     finally {
+        earlyWorkingIndicator?.stop();
+        earlyWorkingIndicator = null;
         // Drain any queued user input typed during streaming. Stash on
         // globalThis for the REPL loop in index.ts to restore into the
         // next editable prompt. Enter typed mid-stream is preserved as
@@ -1985,6 +2018,8 @@ export async function runQuery(ctx) {
         // can't be aborted between turns by Shift+F5 (soft-cancel).
         globalThis.__turnAbortCtl = null;
         globalThis.__turnCancelCurrent = null;
+        if (isFooterActive())
+            setFooterActivity('Ready', 0, null);
     }
 }
 const TOOL_CALL_LOOP_THRESHOLD = 3;