@tiens.nguyen/gonext-local-worker 1.0.315 → 1.0.317

package/gonext-repl.mjs

@@ -1182,6 +1182,10 @@ let jobIdForApprovalRace = "";
 // Per-session coding-model override chosen via /model (empty = use the account default).
 // Sent to /agent-ask as codingModelOverride; the API validates it against the whitelist.
 let sessionCodingModel = "";
+// The default coding model id (from the startup agent-payload probe) — shown in the
+// per-turn token footer next to the GLOBAL output total so it's clear which model that
+// lifetime count belongs to. The /model override (sessionCodingModel) wins when set.
+let defaultCodingModel = "";
 // Auto-test mode (/test-auto): when on, the agent verifies its work (build/run/curl/…)
 // and fixes → re-tests until it passes before finishing. Default off; remembered per
 // folder in the session file; sent to /agent-ask as autoTest so python drives the loop.
@@ -1937,13 +1941,20 @@ async function runAgentTurn(history) {
           }
         }
         clearStatus();
-        // Task #104: expose this turn's code-model INPUT + OUTPUT token totals so the caller
-        // can print a persistent footer at the bottom of the screen.
+        // Task #104: expose this turn's code-model INPUT + the CUMULATIVE output-token
+        // total (all turns for this user + coding server) so the caller can print a
+        // persistent footer. Output is the running total (globalOutputTokens already holds
+        // every prior turn; latestOutputTokens is THIS turn, not yet folded in because the
+        // addOutputTokens POST above is async) — the same figure the live "↓" showed during
+        // the turn, so it doesn't drop back to just this turn's count at the prompt (#112).
+        // `turnOutputTokens` is kept separately so the caller's footer guard reflects whether
+        // THIS turn used the code model (a plain-reply greeting = 0 → no footer).
         return {
           text: answerFrom(text),
           shown: answerShownLive,
           inputTokens: latestCodeTokens,
-          outputTokens: latestOutputTokens,
+          outputTokens: globalOutputTokens + latestOutputTokens,
+          turnOutputTokens: latestOutputTokens,
         };
       }
       if (text.length > shownChars) {
@@ -2053,6 +2064,8 @@ async function main() {
   try {
     const probe = await fetchAgentPayload([{ role: "user", content: "ping" }]);
     const p = probe?.payload ?? {};
+    // Remember the default coder id for the token footer's global-total label.
+    if (p.codingModelId) defaultCodingModel = p.codingModelId;
     // RAG storage choice (task #97): ask ONCE per folder, and only when RAG is actually
     // enabled in Settings. Default Yes = LOCAL (kept on this machine, no S3 needed).
     if (p.ragEnabled && sessionRagMode === null) {
@@ -2270,7 +2283,7 @@ async function main() {
     }
     history.push({ role: "user", content: taskLine });
     try {
-      const { text: answer, shown, cancelled, inputTokens, outputTokens } =
+      const { text: answer, shown, cancelled, inputTokens, outputTokens, turnOutputTokens } =
         await runAgentTurn(history);
       if (answer) {
         history.push({ role: "assistant", content: answer });
@@ -2281,15 +2294,21 @@ async function main() {
         // A plain-reply answer already streamed live in the loop above — printing it
         // again here would just duplicate it; a blank line is enough for spacing.
         console.log(shown ? "" : `\n\n${answer}\n`);
-        // Task #104: persistent per-turn token footer at the bottom of the screen — the
-        // last turn's code-model INPUT + OUTPUT tokens, kept visible above the next prompt.
-        // Only for turns that actually used the code model (plain-reply greetings = 0/0).
-        if ((inputTokens ?? 0) > 0 || (outputTokens ?? 0) > 0) {
+        // Task #104: persistent token footer at the bottom of the screen — this turn's
+        // code-model INPUT + this turn's OUTPUT (↓ out), then the model's GLOBAL lifetime
+        // output total (all turns for this user + coding server), labelled with the model
+        // name so it's clear which coder that running total belongs to. Guard on THIS turn's
+        // activity so a plain-reply greeting that did no code-model work shows nothing (#112).
+        if ((inputTokens ?? 0) > 0 || (turnOutputTokens ?? 0) > 0) {
+          const coder = sessionCodingModel || defaultCodingModel;
           console.log(
             dim("tokens · ↑ in ") +
               fmtTokens(inputTokens ?? 0) +
               dim(" · ") +
-              cyan(`↓ out ${fmtTokens(outputTokens ?? 0)}`)
+              cyan(`↓ out ${fmtTokens(turnOutputTokens ?? 0)}`) +
+              dim(" · ") +
+              dim(`${coder ? coder + " " : ""}total `) +
+              cyan(`↓ ${fmtTokens(outputTokens ?? 0)}`)
           );
         }
       } else {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.315",
+  "version": "1.0.317",
   "description": "Polls GoNext cloud API for async local LLM jobs and runs them against Ollama/OpenAI-compatible servers on this Mac",
   "type": "module",
   "license": "MIT",