@tiens.nguyen/gu-cli 1.0.821 → 1.0.822

package/gu-repl.mjs

@@ -154,7 +154,7 @@ import {
   parseModelFallback,
 } from "./model-fallback.mjs";
 import { codingChoiceTags, coderDisplayName, coderWhere, whereOfActiveCoder } from "./coding-choice-tags.mjs";
-import { TITLE_POP, TITLE_PUSH, idleTitle, needsSummary, summariseForTitle, titleSeq, trimTitle } from "./terminal-title.mjs";
+import { TITLE_POP, TITLE_PUSH, idleTitle, needsSummary, summariseForTitle, titleModel, titleSeq, trimTitle } from "./terminal-title.mjs";
 import { IMAGE_NEEDS_CLAUDE, coderTakesImagesFor, clipboardImageToFile, fitImageNotice, imageMarkersForDisplay, createImageStore, imagePasteKey, imagePathFromPaste, isClaudeCoderUrl, isImagePasteKey, storeImageFile } from "./image-paste.mjs";
 import {
   isDoubleClick,
@@ -1633,7 +1633,17 @@ function showQuestionInTitle(question) {
   void (async () => {
     const settings = await fetchWorkerSettings();
     if (!settings || mine !== _titleSeq) return;
-    const title = await summariseForTitle(question, settings, { fetchImpl: doctorFetch });
+    // #358: the coder this turn runs on, with the key bound to its server — not the agent-model
+    // setting, which can name a model the turns never use, called with no key at all.
+    const target = titleModel(settings, { url: activeCodingUrl(), model: activeCodingModelId() });
+    if (!target) return;
+    const key = target.backendId ? await fetchKeyForBackend(target.backendId) : null;
+    if (mine !== _titleSeq) return;
+    const title = await summariseForTitle(question, settings, {
+      fetchImpl: doctorFetch,
+      target,
+      apiKey: key?.apiKey ?? "",
+    });
     if (title && mine === _titleSeq) writeTitle(title);
   })().catch(() => {});
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gu-cli",
-  "version": "1.0.821",
+  "version": "1.0.822",
   "description": "Gu CLI — the gu terminal plus the local worker that runs agent / OCR / PDF / embedding jobs on your Mac (Ollama / MLX / OpenAI-compatible).",
   "type": "module",
   "license": "MIT",

package/terminal-title.mjs

@@ -7,8 +7,8 @@
  * shortened, and keep it there until the next question.
  *
  * TWO STEPS, AND THE FIRST NEVER WAITS FOR THE SECOND. The trimmed question goes up the moment
- * the turn starts. Only a question too long to fit is then sent to the agent model for a
- * few-word summary, which replaces the trimmed one if — and only if — it arrives usable and the
+ * the turn starts. Only a question too long to fit is then sent to the coder the turn runs on
+ * (with its server's key — see titleModel) for a few-word summary, which replaces the trimmed one if — and only if — it arrives usable and the
  * user has not asked something else in the meantime. Every way that call can fail (no model
  * configured, a provider that wants a key this side does not hold, a timeout, a model that
  * answers with an essay) ends the same way: the trimmed question stays.
@@ -98,6 +98,30 @@ export function agentModelFromSettings(settings) {
   return { url, model: str(settings?.agentModelName) || "default_model" };
 }
 
+/** A server URL reduced to what identifies it: no case, no trailing slash, no trailing /v1. */
+function serverOf(url) {
+  return String(url ?? "").trim().toLowerCase().replace(/\/+$/, "").replace(/\/v1$/, "");
+}
+
+/**
+ * WHICH MODEL WRITES THE TITLE (#358): the coder this turn runs on, and the server row its key
+ * is bound to. → { url, model, backendId } or null.
+ *
+ * It used to be the account's agent-model setting, which on a real account was OpenAI o1 —
+ * not what the turns ran on (kimi-k3 on ollama.com), and called without a key, so every long
+ * question stayed trimmed. The coder is the model the user has already chosen and keyed; the
+ * agent-model setting is only the fallback for a session whose coder is not known yet.
+ */
+export function titleModel(settings, coder = {}) {
+  const rows = Array.isArray(settings?.agentCodingBackends) ? settings.agentCodingBackends : [];
+  const rowFor = (url) => rows.find((r) => r?.id && serverOf(r.url) === serverOf(url))?.id ?? "";
+  const url = String(coder?.url ?? "").trim();
+  const model = String(coder?.model ?? "").trim();
+  if (url && model) return { url, model, backendId: rowFor(url) };
+  const agent = agentModelFromSettings(settings);
+  return agent ? { ...agent, backendId: rowFor(agent.url) } : null;
+}
+
 export function summaryRequest(model, question) {
   const q = Array.from(cleanTitleText(question)).slice(0, SUMMARY_INPUT_MAX).join("");
   return {
@@ -140,20 +164,22 @@ export function pickSummary(body, max = TITLE_MAX) {
 }
 
 /**
- * Ask the agent model for a title. → the title, or "" for every reason there is not one.
+ * Ask a model for a title. → the title, or "" for every reason there is not one.
+ * `target` ({ url, model }, see titleModel) picks the model — the agent-model setting when absent —
+ * and `apiKey` is sent as a Bearer token when the server needs one.
  * Never throws: this runs beside a turn and must not be able to disturb it.
  */
 export async function summariseForTitle(
   question,
   settings,
-  { fetchImpl = fetch, timeoutMs = 20_000 } = {},
+  { fetchImpl = fetch, timeoutMs = 20_000, target = null, apiKey = "" } = {},
 ) {
-  const agent = agentModelFromSettings(settings);
+  const agent = target?.url && target?.model ? target : agentModelFromSettings(settings);
   if (!agent || !needsSummary(question)) return "";
   try {
     const res = await fetchImpl(chatEndpoint(agent.url), {
       method: "POST",
-      headers: { "Content-Type": "application/json" },
+      headers: { "Content-Type": "application/json", ...(apiKey ? { Authorization: `Bearer ${apiKey}` } : {}) },
       body: JSON.stringify(summaryRequest(agent.model, question)),
       signal: AbortSignal.timeout(timeoutMs),
     });