@tiens.nguyen/gu-cli 1.0.795 → 1.0.796

package/client-expert-api.mjs

@@ -701,7 +701,7 @@ export async function probeLocalOllama() {
  */
 /** The shared Ollama every client machine can already reach. */
 export const SHARED_OLLAMA_URL =
-  process.env.GONEXT_CLIENT_AGENT_URL?.trim() || "https://ollama1.gomarsic.cc/v1";
+  process.env.GONEXT_CLIENT_AGENT_URL?.trim() || "https://ollama1.getmarsic.cc/v1";
 /**
  * Chat model — the one the agent routes and thinks with, on every step of every turn.
  *

package/coding-choice-tags.mjs

@@ -69,7 +69,7 @@ export function codingChoiceTags(choices) {
 }
 
 /**
- * WHERE AN OLLAMA CODER RUNS — "ollama.com", "127.0.0.1", "ollama1.gomarsic.cc" — or "" for any
+ * WHERE AN OLLAMA CODER RUNS — "ollama.com", "127.0.0.1", "ollama1.getmarsic.cc" — or "" for any
  * other kind.
  *
  * Reported: "/model ✓ coding model → kimi-k3 (ollama)". The tag was the API's provider label, the

package/coding-key-prompt.mjs

@@ -37,10 +37,10 @@ export const SECRET_KEY_HINT = "type or paste the key, then press Enter — it s
  * that does not exist and, worse, reappears on every start because it can never be satisfied.
  *
  * Reported live: a Client Expert machine whose coder is gemma4-quadro-49k on
- * ollama1.gomarsic.cc being asked, at every launch, for that model's "API key".
+ * ollama1.getmarsic.cc being asked, at every launch, for that model's "API key".
  *
- * Matched on the HOST, not a substring: `evil.com/?x=gomarsic.cc` and
- * `gomarsic.cc.attacker.net` must not be treated as ours — being wrong here would suppress a
+ * Matched on the HOST, not a substring: `evil.com/?x=getmarsic.cc` and
+ * `getmarsic.cc.attacker.net` must not be treated as ours — being wrong here would suppress a
  * genuine prompt and leave someone's coder quietly unusable.
  */
 export function isKeylessCodingHost(url) {
@@ -51,7 +51,7 @@ export function isKeylessCodingHost(url) {
     return false;
   }
   if (host === "127.0.0.1" || host === "localhost" || host === "::1" || host === "[::1]") return true;
-  return host === "gomarsic.cc" || host.endsWith(".gomarsic.cc");
+  return host === "getmarsic.cc" || host.endsWith(".getmarsic.cc");
 }
 
 /**

package/device-login.mjs

@@ -8,7 +8,7 @@ import { homedir, platform } from "node:os";
 import { join } from "node:path";
 import dotenv from "dotenv";
 
-export const DEFAULT_API_BASE = "https://chat-api-v2.gomarsic.cc";
+export const DEFAULT_API_BASE = "https://chat-api-v2.getmarsic.cc";
 const ENV_FILE = join(homedir(), ".gonext", "worker.env");
 const POLL_INTERVAL_MS = 2500;
 const POLL_TIMEOUT_MS = 10 * 60_000; // matches the server's pairing TTL

package/embed-model-command.mjs

@@ -186,3 +186,40 @@ export function describeEmbedChange(plan) {
   lines.push("They will be ignored at search time and rebuilt on the next search.");
   return lines;
 }
+
+/**
+ * IS RAG ACTUALLY ON, AND IF NOT, WHY? (task #347)
+ *
+ * Reported: "Choose embed model but exit and gu again, it keeps asking 'embed model not set',
+ * again and again." Two settings decide it and they were treated as one: WHICH embedder
+ * (`ragEmbedModel` + where it is served) and WHETHER RAG is on (`ragEnabled`). /embed-model saved
+ * the first and never the second, and the bar's warning is driven by the second — so the choice
+ * was stored, never took effect, and the warning came back at every start.
+ *
+ * → { state, model }
+ *     "on"      an embedder is chosen and RAG is on
+ *     "off"     an embedder IS chosen, RAG is switched off  — not "no embed model"
+ *     "unset"   no embedder was ever chosen
+ * Pure. `settings` is the worker-settings view (null/undefined = unknown → "unset").
+ */
+export function embedState(settings) {
+  const model = String(settings?.ragEmbedModel ?? "").trim();
+  const where = String(settings?.ragEmbedUrl ?? "").trim() || String(settings?.ragEmbedModelPath ?? "").trim();
+  if (!model || !where) return { state: "unset", model };
+  return { state: settings?.ragEnabled === true ? "on" : "off", model };
+}
+
+/**
+ * What saving an embedder must write: the model, its server — AND RAG on. Choosing an embedder
+ * is the act of turning RAG on; a choice that leaves it off is the bug above.
+ */
+export function embedSaveFields({ url, model }) {
+  return { ragEmbedUrl: url, ragEmbedModel: model, ragEmbedKeyId: "", ragEnabled: true };
+}
+
+/** The bar's row for a state that is not "on" — "" when RAG is on. One short row (#317). */
+export function embedNotice({ state, model } = {}) {
+  if (state === "on") return "";
+  if (state === "off" && model) return `⚠ RAG is switched off — ${model} is set · /embed-model turns it on`;
+  return "⚠ no embed model — RAG is off · /embed-model";
+}

package/gonext_agent_chat.py

@@ -1426,7 +1426,7 @@ def _resolve_coding_backend(kind, base_url, probe=None):
     budget). Every branch now keys off this one value.
 
     Order matters. The Ollama probe runs BEFORE the remote-host rule so a remote Ollama
-    (ollama1.gomarsic.cc) still resolves to "ollama" and keeps its own treatment.
+    (ollama1.getmarsic.cc) still resolves to "ollama" and keeps its own treatment.
 
         explicit kind                    → that value
         probe says Ollama                → "ollama"
@@ -1467,7 +1467,7 @@ def _sentinel_against_ollama(agent_base_url, agent_model_id, probe=None):
 
     — an error naming a model the user never chose, on a machine whose settings look fine. That
     is the whole defect: not that it fails, but that it fails illegibly and the reader has no
-    thread to pull. Verified live against ollama1.gomarsic.cc, 2026-08-28.
+    thread to pull. Verified live against ollama1.getmarsic.cc, 2026-08-28.
 
     WHY IT CANNOT BE FIXED WHERE IT IS PRODUCED. The API builds that payload with no idea which
     KIND of server the agent URL points at — there is no agentModelKind, only
@@ -1505,7 +1505,7 @@ def _resolve_coding_target(agent_base_url, agent_model_id, raw_coding_base, raw_
     Extracted from run_agent_chat unchanged (task #213). It decides the single most consequential
     thing about a turn — the model that writes every code block — and it lived in the middle of a
     400-line function, so no test could execute it. The Client Expert default puts the chat model,
-    the coder and the embedder on ONE Ollama box (ollama1.gomarsic.cc), which is exactly the shape
+    the coder and the embedder on ONE Ollama box (ollama1.getmarsic.cc), which is exactly the shape
     where "same server" and "different model" collide, and where a silent fall-back to the chat
     model is invisible: both answers are a real model on a real server and the turn still works,
     just not with the coder that was chosen.
@@ -3877,7 +3877,7 @@ def _is_request_timeout(err) -> bool:
         likely STILL RUNNING on the box. Re-sending the identical request — streamed or
         not — starts that same multi-minute generation again from zero.
 
-    Live (2026-08-30, qwen3.8:27b on ollama1.gomarsic.cc): a 7,251-token prompt produced no
+    Live (2026-08-30, qwen3.8:27b on ollama1.getmarsic.cc): a 7,251-token prompt produced no
     first token inside the 600s client timeout. We logged "streamed generate failed (Request
     timed out.) → falling back to non-streaming", told the user "Streaming unavailable", and
     re-sent it unstreamed — buying a second 600s wait with no first-token signal at all.
@@ -10088,7 +10088,7 @@ def run_agent_chat(cfg):
             # stream_agent: request the completion with stream=True and reassemble the
             # deltas into one message. Enabled for Ollama, where a long single-shot
             # (non-stream) generation sends nothing until the very end — a reverse proxy
-            # in front of the box (ollama1.gomarsic.cc) then hits its idle read-timeout
+            # in front of the box (ollama1.getmarsic.cc) then hits its idle read-timeout
             # and kills the request mid-generation. Streaming keeps tokens flowing so the
             # idle timer never fires. smolagents still receives the COMPLETE text, so its
             # <code>/final_answer parsing is unaffected.

package/gu-cli.mjs

@@ -1926,7 +1926,7 @@ const AGENT_WARMUP_IDLE_MS = 30 * 60_000;
 const AGENT_WARMUP_INTERVAL_MS = 240_000; // 4 min — under Ollama's 5-min default unload
 // OCR translation/correction is hardcoded to a managed Ollama service + model.
 // (The user-configurable "OCR translation model" chooser was removed.)
-const OCR_TRANSLATE_OLLAMA_URL = "https://ollama1.gomarsic.cc";
+const OCR_TRANSLATE_OLLAMA_URL = "https://ollama1.getmarsic.cc";
 const OCR_TRANSLATE_OLLAMA_MODEL = "qwen3:14b";
 // Ollama defaults num_ctx to 2048 and silently truncates longer inputs (keeping
 // only the tail), which would drop the start of a long OCR page. Pin the context
@@ -3068,7 +3068,7 @@ function normalizeCorrection(raw, originalText) {
 
 /**
  * Call an Ollama host's native chat endpoint (POST {base}/api/chat) for text
- * correction/translation, e.g. base "https://ollama1.gomarsic.cc" model "qwen3:14b".
+ * correction/translation, e.g. base "https://ollama1.getmarsic.cc" model "qwen3:14b".
  * `think:false` suppresses reasoning models' <think> preamble so we get clean text.
  */
 const OCR_CORRECT_SYSTEM_PROMPT =

package/gu-repl.mjs

@@ -138,7 +138,8 @@ import { unapprovedProjectServers } from "./mcp-command.mjs";
 import { readHooks } from "./hooks-config.mjs";
 import { readSkillSource, sourceLabel, SKILL_SOURCE_KEY } from "./skill-source.mjs";
 import { codingBannerLabel, isOllamaCloudUrl, SECRET_KEY_HINT } from "./coding-key-prompt.mjs";
-import { thoughtPrintSeq } from "./terminal-layout.mjs";
+import { questionEchoText, thoughtPrintSeq } from "./terminal-layout.mjs";
+import { embedNotice, embedState } from "./embed-model-command.mjs";
 import {
   MODELFALLBACK_PREFIX, fallbackAfterTurn, fallbackHistoryLine, fallbackNotice, keyMissingNotice,
   parseModelFallback,
@@ -160,7 +161,7 @@ import {
   registeredRange,
   selectionKeyAction,
 } from "./terminal-copy.mjs";
-import { createHoverController, applyToggleShift } from "./terminal-hover.mjs";
+import { createHoverController, applyToggleShift, rowsDiffer } from "./terminal-hover.mjs";
 // The required pair, imported rather than re-typed: the startup "no models" message names them,
 // and a second copy of these strings would drift from the check that enforces them.
 import { ACCEPTED_EMBED_MODELS, REQUIRED_AGENT_MODEL, REQUIRED_AGENT_PREFIX } from "./ollama-setup.mjs";
@@ -847,7 +848,7 @@ if (!workerKey || !apiBase) {
       console.error(red(`gu: setup failed: ${e instanceof Error ? e.message : e}`));
       if (e && e.code === "PAIRING_UNSUPPORTED") {
         console.error(
-          "Configure manually: gu-cli set <workerKey> --api-base https://chat-api-v2.gomarsic.cc"
+          "Configure manually: gu-cli set <workerKey> --api-base https://chat-api-v2.getmarsic.cc"
         );
       }
       if (e && (e.code === "KEY_REJECTED" || e.code === "KEY_UNCHECKED")) {
@@ -862,7 +863,7 @@ if (!workerKey || !apiBase) {
     console.error(
       red("gu: no worker key / API base found.") +
         `\nRun ${accent("gu-cli login")} to sign in, or configure manually:` +
-        "\n  gu-cli set <workerKey> --api-base https://chat-api-v2.gomarsic.cc"
+        "\n  gu-cli set <workerKey> --api-base https://chat-api-v2.getmarsic.cc"
     );
     process.exit(1);
   }
@@ -1633,11 +1634,14 @@ let agentMissing = false; // no agent model chosen → agent falls back to the c
 // question to. Seeded from the startup payload; re-checked before a question is refused.
 let noModelAtAll = false;
 let embedMissing = false; // no embed model chosen → RAG is off until /embed-model picks one
+// #347: WHY it is off — "unset" (none chosen) or "off" (one IS chosen, RAG is switched off) — and
+// the chosen model's name, so the row never says "no embed model" about a model that is set.
+let embedWhy = { state: "unset", model: "" };
 // One footer row; the two halves are independent so a model that IS set is never named.
 // ONE SHORT ROW EACH (#317): joined, they ran past 80 columns and were cut mid-word.
 const missingNoticeRows = () => [
   ...(agentMissing ? ["⚠ agent model not set — using the coding model · /agent-model"] : []),
-  ...(embedMissing ? ["⚠ no embed model — RAG is off · /embed-model"] : []),
+  ...(embedMissing ? [embedNotice(embedWhy) || "⚠ no embed model — RAG is off · /embed-model"] : []),
 ];
 
 // ANOTHER MODEL STOOD IN (task #343). The agent reports it as a __MODELFALLBACK__:: step; it is
@@ -1662,6 +1666,30 @@ const fallbackNoticeRows = () => {
   ];
 };
 
+/**
+ * Set the embed warning from what the settings SAY (task #347) — the one place it is decided
+ * after startup's first guess.
+ *
+ * `repair`: an embedder is chosen but RAG is off. On a Local-mode machine the only way to choose
+ * one is /embed-model or setup, and until this fix /embed-model never switched RAG on — so that
+ * state is this bug's leftover, and it is finished here instead of asking the person to choose
+ * again. In Cloud mode the web Settings has its own RAG switch; "off" may be a decision, so it is
+ * reported, not changed.
+ */
+async function syncEmbedWarning({ repair = false } = {}) {
+  const settings = await fetchWorkerSettings();
+  if (!settings) return; // offline: keep what the bar already says
+  let st = embedState(settings);
+  if (st.state === "off" && repair && isClientExpert()) {
+    const md = await import("./model-doctor.mjs");
+    const saved = await md.saveWorkerSettings(apiBase, workerKey, { ragEnabled: true }).catch(() => false);
+    if (saved) st = { ...st, state: "on" };
+  }
+  embedWhy = st;
+  embedMissing = st.state !== "on";
+  if (barPinned()) paintBar();
+}
+
 // #307: the calm note after a quick change of mind — dim, not the ⚠ orange of a warning, and in the
 // bar rather than the transcript (which the whole point is to leave untouched).
 let stopNotice = "";
@@ -2577,12 +2605,18 @@ const tokenLineText = () => {
   );
   const total = dim(" · ") + dim("total ") + accent(`↓ ${fmtTokens(Math.max(globalOutputTokens, lastTurnTotalOutput))}`);
   // #323: what the last turn's prompt cost, LAST on the row after a "|" — "~" when the cached share
-  // was assumed; nothing when the model has no price (never a misleading "| $0"). It survives a
+  // was assumed; "no price" when the model has none (never a misleading "| $0"). It survives a
   // narrow window: "total ↓" is dropped before it is (tokenRowWithPrice).
   return tokenRowWithPrice({
     withTotal: head + total,
     withoutTotal: head,
-    price: lastTurnCost?.usd != null ? dim(" | ") + accent(fmtCost(lastTurnCost)) : "",
+    // #349: a model with NO price says so ("no price", dim) — the blank it used to leave read as
+    // the cost feature being broken. Still never "$0": unpriced is not free.
+    price: lastTurnCost?.usd != null
+      ? dim(" | ") + accent(fmtCost(lastTurnCost))
+      : lastTurnCost
+        ? dim(" | no price")
+        : "",
     room: (process.stdout.columns || 80) - 1 - visibleLen(GUTTER),
   });
 };
@@ -3348,6 +3382,7 @@ const hoverCtl = createHoverController({
   pinned: () => barPinned(),
   transcriptRows: () => transcriptRows(),
   blockAtRow: (row) => thoughtBlockAtRow(row),
+  canToggle: (b) => blockCanToggle(b),
   paintRows: (b, on) => paintThoughtRows(b, on),
   replaceRange: (from, to, rows) => transcript.replaceRange(from, to, rows),
   repaintSeq: (b) => thoughtBlockRepaintSeq(b),
@@ -3646,9 +3681,20 @@ function copyTargetAtRow(row) {
   return { block, renderedLines: lines.slice(from, to + 1) };
 }
 
+/**
+ * Would a click on this block change anything? (#350) — its collapsed and its expanded rows, laid
+ * out for THIS window, are not the same. Asked by the hover (may it light up?), by the click and
+ * by Ctrl+T, so the three can never disagree. See rowsDiffer in terminal-hover.mjs.
+ */
+function blockCanToggle(block) {
+  if (!block) return false;
+  return rowsDiffer(paintThoughtRows(block, false, true), paintThoughtRows(block, false, false));
+}
+
 function toggleThoughtAtRow(row) {
   const block = thoughtBlockAtRow(row);
-  return block ? toggleThoughtBlock(block) : false;
+  // A block shown whole has no other form: it did not light up, and the click is not for it.
+  return block && blockCanToggle(block) ? toggleThoughtBlock(block) : false;
 }
 
 /**
@@ -3718,8 +3764,12 @@ function toggleThoughtBlock(block) {
 
 /** Expand/collapse the most recently printed thought block. The keyboard path (Ctrl+T). */
 function toggleNewestThought() {
-  const block = thoughtBlocks[thoughtBlocks.length - 1];
-  if (!barPinned() || !block) return false;
+  if (!barPinned()) return false;
+  // The newest block that HAS another form (#350): Ctrl+T on a thought already shown whole did
+  // nothing visible, while an expandable one sat just above it.
+  let block = null;
+  for (let i = thoughtBlocks.length - 1; i >= 0 && !block; i--) if (blockCanToggle(thoughtBlocks[i])) block = thoughtBlocks[i];
+  if (!block) return false;
   return toggleThoughtBlock(block);
 }
 
@@ -4617,13 +4667,13 @@ const ask = async (q, { prefill = "", hint = [], atomic = false, noCancel = fals
       rl.line = String(prefill);
       rl.cursor = rl.line.length;
     }
-    // ELLIPSIS, so a truncation is VISIBLE. Silently cutting "…[https://ollama2.gomarsic.cc/v1]:"
+    // ELLIPSIS, so a truncation is VISIBLE. Silently cutting "…[https://ollama2.getmarsic.cc/v1]:"
     // to "…[https:/" hides the default Enter would accept, which is worse than not showing it —
     // the reader cannot tell whether the value is short or the row lied. Callers should keep the
     // final line short and put the long part above; this is the guard, not the plan.
     // "… " with the space: the label normally ends in ": ", and a truncation that eats the
     // separator leaves the typed text jammed against the last surviving word
-    // ("agent + embed ser…ollama2.gomarsic.cc"). Two columns for the ellipsis and its space.
+    // ("agent + embed ser…ollama2.getmarsic.cc"). Two columns for the ellipsis and its space.
     inlinePrompt =
       (visibleLen(label) > budget ? clipVisible(label, budget - 2) + "… " : label) || PROMPT;
     try {
@@ -5704,7 +5754,7 @@ async function reconfigureOllamaServer() {
       const typed = (await ask(`  ${short}: `)).trim();
       // "s" is the advertised skip (#301); the older words still work.
       if (ui.readStepAnswer(typed).kind === "skip") return null;
-      // EVERY READING OF A BARE HOSTNAME, in order. Typing "ollama2.gomarsic.cc" means
+      // EVERY READING OF A BARE HOSTNAME, in order. Typing "ollama2.getmarsic.cc" means
       // https://…:443 behind a proxy, while "192.168.1.40" means http://…:11434 — one guess
       // is right for one of them and produces a timeout that looks like a dead server for the
       // other. See urlCandidates; the load decides, not the heuristic.
@@ -5785,7 +5835,7 @@ async function reconfigureOllamaServer() {
   // ask again. Keep it, say where it is changed, and go on to the coder.
   const agentOnProvider = am.agentIsOnProvider(settings);
   const curAgent =
-    String(settings.agentModelUrl ?? "").trim() || "https://ollama1.gomarsic.cc/v1";
+    String(settings.agentModelUrl ?? "").trim() || "https://ollama1.getmarsic.cc/v1";
   const curCode = am.coderServerDefault(settings);
 
   let agent = null;
@@ -6704,7 +6754,9 @@ async function addServer({ role = "coding", dialect = "openai", url: presetUrl =
   }
   if (role === "embed") {
     // ragEmbedKeyId "" — the server's own keys decide now (its default, or this folder's pick).
-    const fields = { ragEmbedUrl: target.roleUrl("embed"), ragEmbedModel: model, ragEmbedKeyId: "" };
+    // …and ragEnabled: true (#347): choosing an embedder IS turning RAG on. Without it the choice
+    // was stored and never took effect, and the "no embed model" warning came back at every start.
+    const fields = ec.embedSaveFields({ url: target.roleUrl("embed"), model });
     if (!(await saveEmbedSettings(fields))) return;
     if (plan) announceEmbedSaved({ chosen: model, plan, url: target.roleUrl("embed") });
     else {
@@ -7849,6 +7901,7 @@ async function noteTurnCost(inputTokens, breakdown, coder = null) {
     const c = mp.turnCost({ input: inputTokens, ...cache, price });
     if (!c) {
       lastTurnCost = { usd: null, model, host: mp.hostOf(url) };
+      if (barPinned()) paintBar(); // #349: "no price" is shown, so it is painted like a cost is
       return;
     }
     lastTurnCost = { ...c, input: inputTokens, cacheReported: cache.cacheReported, model, host: mp.hostOf(url) };
@@ -7891,7 +7944,7 @@ async function runPricesCommand(line = "/prices") {
   console.log(GUTTER + dim(pricePublishedAt ? `from gu's tables published ${new Date(pricePublishedAt).toLocaleString()}` : "gu's built-in tables (none published applied yet)"));
   if (priceUpdate) console.log(GUTTER + accent(`${mp.priceUpdateLabel(priceUpdate.publishedAt)} — click it below, or /prices update`));
   if (!mp.findPrice(rows, url, model)) {
-    console.log(GUTTER + yellow(`no price for ${model || "the coder"} at ${mp.hostOf(url) || "its server"} — no cost is shown for its turns.`));
+    for (const r of wrapBlock(mp.noPriceLine({ model, host: mp.hostOf(url) }), GUTTER, width)) console.log(GUTTER + yellow(r));
   }
   if (lastTurnCost?.usd != null) {
     const share = lastTurnCost.input ? Math.round((100 * lastTurnCost.cachedTokens) / lastTurnCost.input) : 0;
@@ -11088,6 +11141,7 @@ async function runAgentTurn(history, { attachJobId = "", ownJob = false } = {})
   // whether the user has clicked to expand it. Both reset when the step's Thought changes.
   let fullLiveThought = "";
   let thoughtExpanded = false;
+  let liveExpandable = false; // the live thought line has more to show than fits (set each tick)
   // Task #148: the CURRENT step's raw model output, accumulated in full (fenceThought stops at
   // 400 chars — enough for a one-line label, not for the paragraphs we now print). Bounded so a
   // runaway generation cannot grow without limit; the prose we keep is far shorter anyway.
@@ -11298,6 +11352,9 @@ async function runAgentTurn(history, { attachJobId = "", ownJob = false } = {})
     // "almost done" > "Thinking". When clickable we drive line 1 from the SAME fuller string, so
     // the grey expansion CONTINUES where line 1 was cut off instead of repeating the head.
     const clickable = !!fullLiveThought && !runningCmd && fullLiveThought.length > max;
+    // #350: the click and Ctrl+T ask the SAME question the highlight does. They used to accept
+    // any thought at all, so a line that was already whole took the click and redrew for nothing.
+    liveExpandable = clickable;
     let label, rest;
     if (clickable) {
       // Split at a WORD boundary near the line-1 width so nothing is cut mid-word: the last
@@ -11450,14 +11507,14 @@ async function runAgentTurn(history, { attachJobId = "", ownJob = false } = {})
     if (row < top || row > top + statusContentLines - LEAD_ROWS - WORD_GAP - 2) return;
     // Only meaningful when there IS a fuller thought to reveal; otherwise ignore the click
     // so we don't flicker an empty expand.
-    if (!thoughtExpanded && !(fullLiveThought && !runningCmd)) return;
+    if (!thoughtExpanded && !liveExpandable) return;
     toggleThought();
   };
   // Same expand/collapse, reachable without the mouse: Ctrl+T while a turn runs (see
   // toggleActiveThought). Ignored when there's nothing more to reveal, so it never flickers an
   // empty expansion.
   const toggleThought = () => {
-    if (!thoughtExpanded && !(fullLiveThought && !runningCmd)) return;
+    if (!thoughtExpanded && !liveExpandable) return;
     thoughtExpanded = !thoughtExpanded;
     clearStatus(); // next tick (≤120ms) redraws at the new size; keeps cursor math correct
   };
@@ -12457,7 +12514,10 @@ async function main() {
     // reaches the model over HTTP), so it runs HERE, where the files are, and calls the
     // model machine for tokens. That needs a worker here like any other machine. The earlier
     // build suppressed the daemon entirely, which is why this comment exists.
-    if (!status.running) {
+    // GU_NO_WORKER_AUTOSTART=1 (task #351): the terminal UI tests play a scripted turn from a fake
+    // API — no worker is needed, and one started for a throwaway home outlived the test (on
+    // Windows it would be a scheduled task). Nothing a user sets; without it, as always:
+    if (!status.running && process.env.GU_NO_WORKER_AUTOSTART !== "1") {
       // NO LONGER A QUESTION (task #159, user 2026-08-09: "it need to run otherwise cannot work,
       // so, just notify"). It used to ask, and a "no" was then remembered per folder — but the
       // only workable answer was yes: without a worker nothing claims your turns, so a question
@@ -12613,7 +12673,12 @@ async function main() {
     // private mode's local API can itself be unconfigured.
     if (p.agentMissing === true) agentMissing = true;
     noModelAtAll = nothingToAnswerWith(p);
-    if (p.ragEnabled !== true) embedMissing = true;
+    if (p.ragEnabled !== true) {
+      embedMissing = true;
+      // …but "off" has two causes, and only the settings can tell them apart (#347). In the
+      // background: startup never waits on it, and the bar is repainted when the answer is in.
+      void syncEmbedWarning({ repair: true });
+    }
     // A RESTORED PICK MUST STILL EXIST. The stored id names a model on the box this folder was
     // last pointed at — and the box can change (`/model` → change the Ollama server), a model
     // can be removed, or the folder can be opened on another machine entirely. Sending a name
@@ -12797,7 +12862,10 @@ async function main() {
         // check used to sit INSIDE the "chat model is down" branch, so a client pointing at
         // a healthy remote model was never offered them and every turn died with
         // "No module named 'smolagents'". Independent now.
-        if (!(await md.hasAgentLibs())) {
+        // GU_NO_AGENT_DEPS_PROMPT=1 (task #351): the terminal UI tests run no agent — on a machine
+        // without the framework every test's throwaway home answered "yes" and sat through a pip
+        // install (found on the Ubuntu server: 21 seconds of downloads, then a timeout).
+        if (process.env.GU_NO_AGENT_DEPS_PROMPT !== "1" && !(await md.hasAgentLibs())) {
           const skipDeps = await loadSessionSkip(resolve(process.cwd()), "skipDepsPrompt");
           if (!skipDeps) {
             console.log(
@@ -13338,7 +13406,12 @@ async function main() {
       // TRANSCRIPT: the echo is written exactly once per submitted line and never repainted,
       // so there is nothing here for a blank to accumulate against.
       // As typed, not as sent: "[Image #1]", never the file path the agent needs (#51).
-      const echo = PROMPT_PAST + imageMarkersForDisplay(line) + "\n\n";
+      // Every row under the first one's text (#348): a pasted multi-line question used to be
+      // echoed with only its first line indented, the rest hard against the left edge.
+      const echo = questionEchoText(imageMarkersForDisplay(line), {
+        width: process.stdout.columns || 80,
+        paintPrompt: () => PROMPT_PAST,
+      }) + "\n\n";
       if (stopNotice) {
         stopNotice = "";
         if (stopNoticeTimer) clearTimeout(stopNoticeTimer);
@@ -13418,7 +13491,10 @@ async function main() {
     }
     if (line === "/embed-model") {
       await chooseEmbedModel();
-      embedMissing = false; // Task #14: an embed model was just chosen → clear its warning
+      // Read back what is SAVED (#347). The warning used to be cleared here whatever happened —
+      // a cancelled menu or a failed save removed it for the session, and it returned at the
+      // next start as if the choice had been forgotten.
+      await syncEmbedWarning({ repair: false });
       continue;
     }
     if (line === "/server") {

package/model-prices.mjs

@@ -177,6 +177,33 @@ export function turnCost({ input = 0, cached = 0, cacheReported = false, firstIn
 }
 
 /** "$1.23", "$0.123", "$0.0042"; "~" in front when the cached share was assumed. */
+/**
+ * WHAT THE BAR SAYS ABOUT THE LAST TURN'S COST (task #349).
+ *
+ * Reported: "price and cost is not shown on this gu terminal on this Mac." The coder there was a
+ * self-hosted Ollama model, which is in no price table — and for a model with no price the bar
+ * showed NOTHING, by design ("never a misleading | $0"). Right about the $0, wrong about the
+ * silence: a missing figure looked like a broken feature. So the three cases are told apart:
+ *
+ *   a price was found      → { kind: "cost", text: "$0.334" | "~$0.334" }
+ *   the turn's model has none → { kind: "none", text: "no price" }   — never "$0": it is not free,
+ *                                                                     it is not priced
+ *   no turn has run yet    → null (there is nothing to say)
+ *
+ * `last` is the terminal's lastTurnCost: { usd, estimated } or { usd: null, model, host }.
+ */
+export function costSegment(last) {
+  if (!last) return null;
+  if (last.usd == null) return { kind: "none", text: "no price" };
+  return { kind: "cost", text: formatUsd(last.usd, { estimated: last.estimated === true }) };
+}
+
+/** The line /prices prints when the coder in use has no price — why, and that nothing is broken. */
+export function noPriceLine({ model = "", host = "" } = {}) {
+  const what = `${model || "the coder"}${host ? ` at ${host}` : ""}`;
+  return `no price for ${what} — its turns show "no price" on the bar instead of a cost. The table prices hosted providers; a model on your own server has none.`;
+}
+
 export function formatUsd(usd, { estimated = false } = {}) {
   if (!Number.isFinite(usd)) return "";
   const v = Math.max(0, usd);

package/model-setup-ui.mjs

@@ -1,7 +1,7 @@
 /**
  * THE WAY OUT OF THE MODEL QUESTIONS, and the key question that follows a typed server (#301).
  *
- * Reported: at "Agent + embedding server [https://ollama1.gomarsic.cc]:" there was no visible way
+ * Reported: at "Agent + embedding server [https://ollama1.getmarsic.cc]:" there was no visible way
  * to NOT choose. Enter takes the default server; a decline existed ("n", "no", "skip", "-") but no
  * prompt said so, and one hint even claimed "press Enter with no address to skip", which was the
  * opposite of what Enter does. Some people want no model setup at all and to load a saved preset.

package/ollama-setup.mjs

@@ -1,7 +1,7 @@
 /**
  * Pointing a Client Expert machine at ITS OWN Ollama box (task #220).
  *
- * Setup used to seed https://ollama1.gomarsic.cc and three model names, so anyone whose box is
+ * Setup used to seed https://ollama1.getmarsic.cc and three model names, so anyone whose box is
  * somewhere else — a company LAN, http://192.168.1.40:11434, their own dnsname — had to know to
  * override environment variables they had never heard of. This module is the pure half of asking
  * instead: normalise whatever URL they type, and turn the list that box actually reports into the
@@ -180,8 +180,8 @@ export function describeNotReady(host, missing) {
  *   192.168.1.40:11434           → http://192.168.1.40:11434/v1
  *   ollama.local                 → http://ollama.local:11434/v1
  *   http://box:11434/v1          → http://box:11434/v1
- *   https://ollama1.gomarsic.cc  → https://ollama1.gomarsic.cc/v1
- *   https://ollama1.gomarsic.cc/v1/ → https://ollama1.gomarsic.cc/v1
+ *   https://ollama1.getmarsic.cc  → https://ollama1.getmarsic.cc/v1
+ *   https://ollama1.getmarsic.cc/v1/ → https://ollama1.getmarsic.cc/v1
  *
  * TWO DEFAULTS, AND THEY ARE NOT SYMMETRICAL:
  *
@@ -243,7 +243,7 @@ export function normalizeOllamaUrl(raw) {
  * Every address a bare hostname could reasonably mean, in the order worth trying.
  *
  * ONE GUESS IS NOT ENOUGH, and this was found by typing a real hostname at a real prompt:
- * `ollama2.gomarsic.cc` became `http://ollama2.gomarsic.cc:11434` and timed out, because that
+ * `ollama2.getmarsic.cc` became `http://ollama2.getmarsic.cc:11434` and timed out, because that
  * box is behind a reverse proxy on HTTPS:443 — exactly like ollama1. The single default is
  * right for a LAN box and wrong for a public name, and the reader has no way to know which
  * assumption we made beyond a timeout that looks like their server is down.
@@ -390,7 +390,7 @@ export async function localOllamaDefault(fallbackUrl, fetchTags, { log } = {}) {
  *
  * DISPLAY ONLY. Every settings field in gu stores the /v1 base, because that is the
  * OpenAI-compatible endpoint the API actually calls; nothing about what is SAVED changes here.
- * What changes is what we show: a default of "https://ollama1.gomarsic.cc/v1" reads as an
+ * What changes is what we show: a default of "https://ollama1.getmarsic.cc/v1" reads as an
  * instruction, and the reader is left wondering whether their own address needs the suffix too
  * (reported 2026-09-04: "it makes user confused if he/she should input v1 to the url").
  *

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gu-cli",
-  "version": "1.0.795",
+  "version": "1.0.796",
   "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/proxy-config.mjs

@@ -126,8 +126,8 @@ export function isLoopbackUrl(url) {
 /**
  * The origin (scheme + host + port) of a URL, which is what "the same endpoint" means here.
  *
- * A configured model is https://ollama1.gomarsic.cc/v1 while the actual call is
- * https://ollama1.gomarsic.cc/v1/chat/completions — comparing full URLs would match neither, and
+ * A configured model is https://ollama1.getmarsic.cc/v1 while the actual call is
+ * https://ollama1.getmarsic.cc/v1/chat/completions — comparing full URLs would match neither, and
  * comparing hostnames alone would proxy a DIFFERENT service that happens to share the host.
  * Origin is the level that means "this endpoint", and URL normalizes the default port for us.
  */
@@ -141,7 +141,7 @@ export function originOf(url) {
 
 /**
  * The three model endpoints, read from the settings row rather than hardcoded — the user can
- * change any of them at any time (/model, the web app), and a hardcoded ollama1.gomarsic.cc
+ * change any of them at any time (/model, the web app), and a hardcoded ollama1.getmarsic.cc
  * would keep proxying an endpoint they have already moved away from.
  */
 export function modelOrigins(settings = {}) {

package/terminal-hover.mjs

@@ -63,10 +63,38 @@ export function applyToggleShift(blocks, block, newEnd) {
  *   write(seq)          → put bytes on screen.
  *   afterPaint()        → put the caret back where the user is typing.
  */
+/**
+ * DOES A CLICK ON THIS BLOCK CHANGE ANYTHING? (task #350) — the ONE question hover and click
+ * both ask.
+ *
+ * Reported: "hover an item and it becomes whiter; some items, you click and nothing expands — so
+ * the user is confused. Only allow hover to turn white if it is expandable." A highlight is a
+ * promise that a click does something. Every thought and every look-up row made that promise;
+ * a short thought that is already shown whole could not keep it.
+ *
+ * A block has two forms, collapsed and expanded. It can be toggled exactly when they are not the
+ * same rows — which answers every case without a list of kinds:
+ *   · a long thought, collapsed       → different rows → highlight, click expands
+ *   · the same thought, expanded      → different rows → highlight, click collapses
+ *   · a thought that fits as it is    → identical rows → no highlight, click does nothing
+ *   · one look-up that fits its row   → identical rows → no highlight
+ * …and it follows the WINDOW: the rows are laid out for the current width, so a thought that
+ * fits at 160 columns and is cut at 80 is expandable at 80 only.
+ */
+export function rowsDiffer(collapsedRows, expandedRows) {
+  const a = collapsedRows ?? [];
+  const b = expandedRows ?? [];
+  if (a.length !== b.length) return true;
+  for (let i = 0; i < a.length; i++) if (a[i] !== b[i]) return true;
+  return false;
+}
+
 export function createHoverController({
   pinned = () => true,
   transcriptRows = () => 1,
   blockAtRow = () => null,
+  // #350: may this block be highlighted — i.e. would a click on it change anything?
+  canToggle = () => true,
   paintRows = () => [],
   replaceRange = () => null,
   repaintSeq = () => "",
@@ -101,7 +129,10 @@ export function createHoverController({
     if (key === lastKey) return false;
     lastKey = key;
 
-    const block = blockAtRow(row);
+    // ONLY WHAT A CLICK WOULD CHANGE (#350). A block that is already shown whole is treated as
+    // no block at all: nothing lights up, and whatever was lit before is put back.
+    const under = blockAtRow(row);
+    const block = under && canToggle(under) ? under : null;
     if (block === hovered) return false;
     const previous = hovered;
     hovered = block;

package/terminal-layout.mjs

@@ -528,6 +528,61 @@ export function buildFooterRows({
   ];
 }
 
+/**
+ * A SUBMITTED QUESTION, LAID OUT FOR THE TRANSCRIPT (task #348): every row at the column where
+ * the first one's text starts — after ">> ".
+ *
+ * Reported with a screenshot: a pasted question of many lines was echoed as ">> Build the first
+ * playable…" and then every other line hard against the left edge. The echo was the prompt plus
+ * the text exactly as typed, so only the first line had anything in front of it; a long line was
+ * left for the terminal to soft-wrap, back to column 0 as well.
+ *
+ * → rows WITHOUT the prompt: the caller puts ">> " before the first and `indent` spaces before
+ * the rest. Blank lines stay empty (no trailing spaces). A line's own leading spaces are kept —
+ * pasted code and nested lists keep their shape — and its wrapped rows start under its first
+ * word. A word longer than the row is cut rather than left to overflow.
+ */
+export function questionEchoRows(text, { width = 80, indent = 3 } = {}) {
+  const rows = [];
+  for (const line of String(text ?? "").replace(/\r/g, "").split("\n")) {
+    if (line.trim() === "") {
+      rows.push("");
+      continue;
+    }
+    const lead = /^[ \t]*/.exec(line)[0].replace(/\t/g, "  ");
+    const avail = Math.max(8, (width || 80) - 1 - indent - visibleLen(lead));
+    let cur = "";
+    for (let word of line.trim().split(/\s+/)) {
+      while (visibleLen(word) > avail) {
+        if (cur) {
+          rows.push(lead + cur);
+          cur = "";
+        }
+        const head = clipVisible(word, avail);
+        rows.push(lead + head);
+        word = word.slice(head.length);
+      }
+      if (!word) continue;
+      const next = cur ? `${cur} ${word}` : word;
+      if (visibleLen(next) <= avail) cur = next;
+      else {
+        rows.push(lead + cur);
+        cur = word;
+      }
+    }
+    if (cur) rows.push(lead + cur);
+  }
+  return rows;
+}
+
+/** The echo itself: ">> " on the first row, the same width of spaces on every other. */
+export function questionEchoText(text, { width = 80, prompt = ">> ", paintPrompt = (s) => s } = {}) {
+  const indent = visibleLen(prompt);
+  const rows = questionEchoRows(text, { width, indent });
+  if (rows.length === 0) return paintPrompt(prompt);
+  return rows.map((r, i) => (i === 0 ? paintPrompt(prompt) + r : r === "" ? "" : " ".repeat(indent) + r)).join("\n");
+}
+
 /**
  * The bytes that print a NEW thought block into the transcript (task #344): its painted rows,
  * with ONE empty line above when the row above is not already blank (`gap`).