@tiens.nguyen/gu-cli 1.0.782 → 1.0.784

package/client-expert-run.mjs

@@ -1111,7 +1111,7 @@ export async function runClientExpertSetup({
       console.log("");
       console.log("  " + cyan("✓ Model setup skipped") + dim(loadPreset
         ? " — the preset picker opens as soon as gu starts"
-        : " — set them any time with /model (or /open-presets)"));
+        : " — set them any time with /model (or /use-presets)"));
     } else if (Object.keys(written).length === 0) {
       console.log("  " + cyan("✓ Models and RAG") + dim(" — already configured, left alone"));
     } else {
@@ -1159,7 +1159,7 @@ export async function runClientExpertSetup({
     dim("\n  machine. The API restarts with `gu-cli api start`; its log is") +
     dim(`\n  ${API_LOG_FILE}\n`)
   );
-  // #301: the caller opens /load-presets once the terminal exists — the picker, the plan and the
+  // #301: the caller opens /use-presets once the terminal exists — the picker, the plan and the
   // confirmation are that command's, not rebuilt here.
   return { complete: true, stoppedAt: null, ...(loadPreset ? { loadPreset: true } : {}) };
 }

package/gu-cli.mjs

@@ -20,7 +20,7 @@ import { execFile as execFileCallback, spawn } from "node:child_process";
 import { createHash, randomUUID } from "node:crypto";
 import { writeFileSync, unlinkSync } from "node:fs";
 import { createTaskPool, createFolderQueue } from "./job-pools.mjs";
-import { NO_STEP, foldStepTokens, withStepTokens } from "./step-tokens.mjs";
+import { NO_STEP, foldStepTokens, withStepTokens, NO_CACHE, foldTurnCache, withTurnCache } from "./step-tokens.mjs";
 import { agentStoppedLine, jobEnded, jobStatusFrom409, stopGracefully } from "./agent-stop.mjs";
 import { ragJobConfig } from "./rag-job-config.mjs";
 import { codingKeyFallbackLine, isOllamaCloudUrl } from "./coding-key-prompt.mjs";
@@ -4138,6 +4138,9 @@ async function runAgentChatJob(job) {
     // Task #298: the CURRENT model call's own ↑ in / ↓ out, for the thinking line (the bar
     // already shows the totals). Rides on the breakdown — see step-tokens.mjs.
     let stepTokens = NO_STEP;
+    // #323: the turn's cached prompt tokens (when the provider reports them) and its first step's
+    // prompt — what the terminal needs to say what the turn COST. Rides on the breakdown too.
+    let turnCache = NO_CACHE;
     // Raw model-stream tokens (event.type === "stream") are Python code + `Thought:`
     // prose + <code>/<|"|> fragments. The web Thought panel renders reasoning as
     // markdown, which turns `#`-prefixed code comments into <h1> headings and
@@ -4305,17 +4308,20 @@ async function runAgentChatJob(job) {
         if (breakdown) latestPromptBreakdown = breakdown;
         const prevStep = stepTokens;
         stepTokens = foldStepTokens(stepTokens, event);
+        const prevCache = turnCache;
+        turnCache = foldTurnCache(turnCache, event);
+        const cacheChanged = turnCache.cached !== prevCache.cached || turnCache.firstIn !== prevCache.firstIn || turnCache.cacheReported !== prevCache.cacheReported;
         const stepChanged =
           stepTokens.step !== prevStep.step ||
           stepTokens.stepIn !== prevStep.stepIn ||
           stepTokens.stepOut !== prevStep.stepOut;
-        if (latestPromptBreakdown) latestPromptBreakdown = withStepTokens(latestPromptBreakdown, stepTokens);
+        if (latestPromptBreakdown) latestPromptBreakdown = withTurnCache(withStepTokens(latestPromptBreakdown, stepTokens), turnCache);
         if (
           nextTotal !== latestCodeTokens ||
           nextMax !== latestMaxCodeTokens ||
           nextOut !== latestCodeOutputTokens ||
           breakdown ||
-          (stepChanged && latestPromptBreakdown)
+          ((stepChanged || cacheChanged) && latestPromptBreakdown)
         ) {
           latestCodeTokens = nextTotal;
           latestMaxCodeTokens = nextMax;