dsh-think-translate 1.2.5 → 1.2.8

package/README.de.md

@@ -69,6 +69,10 @@ Am Plugin ist nichts zu konfigurieren: Es deklariert keine Reihenfolge-Abhängig
 
 Für DSH **0.2.0-rc.2** verwenden Sie **dsh-think-translate 1.2.5 oder neuer**. Version 1.2.5 behebt den Registrierungskonflikt des Slots `todo_write` und deklariert den Peer-Versionsbereich 0.2. Führen Sie `dsh plugin --profile web add dsh-think-translate@latest` aus und starten Sie DSH neu; eine Versionsausnahme ist nicht nötig.
 
+**1.2.6** behebt doppelt angezeigte Gedanken und Antworten in DSH 0.2: Der Renderer und die Originaltext-Anzeige bei Fehlern beachten die Gruppen reasoning/response.
+
+**1.2.7 lokale Vorbereitung:** Aktiviertes Ollama an erster Stelle oder dessen Test findet vorhandene Installationen und startet/lädt das Modell. Fehlende Software und Modelle werden erst über die Vorbereitungsschaltfläche installiert/heruntergeladen, mit Fortschritt je Phase. Automatische Softwareinstallation unter Windows, sonst offizielle Anleitung. Der aktuelle Anbieter folgt der Priorität aktivierter, geprüfter Anbieter.
+
 ## 🚀 Verwendung
 
 1. **Einstellungen → Übersetzung der Gedankenkette** öffnen

package/README.es.md

@@ -69,6 +69,10 @@ No hay nada que configurar en el plugin: no declara dependencia de orden con los
 
 Para DSH **0.2.0-rc.2**, usa **dsh-think-translate 1.2.5 o posterior**. La versión 1.2.5 corrige el conflicto de registro del slot `todo_write` y declara el rango peer 0.2. Ejecuta `dsh plugin --profile web add dsh-think-translate@latest` y reinicia DSH; no hace falta una excepción de versión.
 
+**1.2.6** corrige la duplicación del razonamiento y las respuestas en DSH 0.2: el renderizador y su texto original de respaldo respetan los grupos reasoning/response.
+
+**1.2.7 preparación local:** Ollama habilitado en primera posición, o su botón de prueba, detecta la instalación existente y arranca/carga el modelo. La instalación y descarga requieren el botón de preparación y muestran progreso por etapas. La instalación automática del software admite Windows; otros sistemas enlazan la guía oficial. El proveedor en uso sigue la prioridad de los proveedores habilitados y verificados.
+
 ## 🚀 Uso
 
 1. Abre **Ajustes → Traducción de cadena de pensamiento**

package/README.fr.md

@@ -69,6 +69,10 @@ Rien à configurer côté plugin : il ne déclare aucune dépendance d'ordre vis
 
 Pour DSH **0.2.0-rc.2**, utilisez **dsh-think-translate 1.2.5 ou une version ultérieure**. La version 1.2.5 corrige le conflit du slot `todo_write` et déclare la plage de versions peer 0.2. Exécutez `dsh plugin --profile web add dsh-think-translate@latest`, puis redémarrez DSH ; aucune exception de version requise.
 
+**1.2.6** corrige les pensées et réponses affichées en double dans DSH 0.2 : le rendu et le texte original de secours respectent les groupes reasoning/response.
+
+**1.2.7 préparation locale :** Ollama activé en première position, ou son test, détecte une installation existante, démarre le service et charge le modèle. Le bouton de préparation autorise les installations et téléchargements manquants, avec progression par étapes. Installation automatique du logiciel sur Windows ; guide officiel ailleurs. Le fournisseur utilisé respecte la priorité des fournisseurs activés et vérifiés.
+
 ## 🚀 Utilisation
 
 1. Ouvrez **Réglages → Traduction de chaîne de réflexion**

package/README.ja.md

@@ -69,6 +69,10 @@ New-Item -ItemType Junction -Path "$HOME\.dsh\profiles\node_modules\dsh-think-tr
 
 DSH **0.2.0-rc.2** では **dsh-think-translate 1.2.5 以降**を使用してください。1.2.5 は `todo_write` の slot 登録競合を修正し、0.2 の peer バージョン範囲を宣言します。`dsh plugin --profile web add dsh-think-translate@latest` を実行して DSH を再起動してください。バージョンの例外設定は不要です。
 
+**1.2.6** は DSH 0.2 で思考と回答が重複表示される問題を修正します。助手の表示とエラー時の原文表示が reasoning/response の各グループに従います。
+
+**1.2.7 ローカル準備：**有効な Ollama を最優先にするかテストすると、既存のインストールを検出し起動・モデルロードします。不足するソフトやモデルは準備ボタンでインストールし、段階別の進捗を表示します。自動インストールは Windows 対応、他の OS は公式ガイドを表示します。使用中の表示は優先順位の高い検証済み提供元を維持します。
+
 ## 🚀 使い方
 
 1. **設定 → 思考チェーン翻訳** を開く

package/README.ko.md

@@ -69,6 +69,10 @@ New-Item -ItemType Junction -Path "$HOME\.dsh\profiles\node_modules\dsh-think-tr
 
 DSH **0.2.0-rc.2**에서는 **dsh-think-translate 1.2.5 이상**을 사용하세요. 1.2.5는 `todo_write` slot 등록 충돌을 수정하고 0.2 peer 버전 범위를 선언합니다. `dsh plugin --profile web add dsh-think-translate@latest`를 실행한 뒤 DSH를 다시 시작하세요. 버전 예외 설정은 필요 없습니다.
 
+**1.2.6**은 DSH 0.2에서 사고와 답변이 중복 표시되는 문제를 수정합니다. 어시스턴트 렌더링과 오류 시 원문 표시 모두 reasoning/response 그룹을 따릅니다.
+
+**1.2.7 로컬 준비:** 활성 Ollama를 첫 순위로 두거나 테스트하면 기존 설치를 찾아 시작하고 모델을 로드합니다. 누락된 소프트웨어와 모델은 준비 버튼을 눌러 설치/다운로드하며 단계별 진행 상황을 표시합니다. 자동 설치는 Windows에서 지원하며 다른 OS는 공식 안내를 제공합니다. 현재 사용 표시는 우선순위가 가장 높은 검증된 제공자를 유지합니다.
+
 ## 🚀 사용법
 
 1. **설정 → 사고 체인 번역** 열기

package/README.md

@@ -66,6 +66,10 @@ Nothing to configure on the plugin side: it declares no ordering dependency on D
 
 For DSH **0.2.0-rc.2**, use **dsh-think-translate 1.2.5 or later**. Version 1.2.5 fixes the `todo_write` slot collision and declares the 0.2 peer range. Run `dsh plugin --profile web add dsh-think-translate@latest`, then restart DSH; no version exemption is needed.
 
+Version **1.2.6** fixes duplicated thinking and answer text in DSH 0.2: the assistant renderer now respects the separate reasoning/response groups, including its original-text error fallback.
+
+**1.2.7 local setup:** enabled Ollama in first priority, or its Test button, finds an existing installation and starts/loads it automatically. Missing software/models require the labelled setup action, with staged download/install/load progress. Automatic software installation supports Windows; other systems link to the official guide. “In use” follows the first enabled, verified provider in priority order; testing a lower row does not override a verified higher row.
+
 ## 🚀 Usage
 
 1. Open **Settings → Think Translation**

package/README.ru.md

@@ -69,6 +69,10 @@ New-Item -ItemType Junction -Path "$HOME\.dsh\profiles\node_modules\dsh-think-tr
 
 Для DSH **0.2.0-rc.2** используйте **dsh-think-translate 1.2.5 или новее**. Версия 1.2.5 исправляет конфликт регистрации slot `todo_write` и объявляет диапазон peer-версий 0.2. Выполните `dsh plugin --profile web add dsh-think-translate@latest`, затем перезапустите DSH; исключение версии не требуется.
 
+**1.2.6** исправляет двойное отображение рассуждений и ответов в DSH 0.2: основной вывод и исходный текст при ошибке учитывают группы reasoning/response.
+
+**1.2.7 локальная подготовка:** включённый Ollama на первом месте или его тест находит существующую установку, запускает сервис и загружает модель. Отсутствующие компоненты устанавливаются после нажатия кнопки подготовки с отображением этапов и прогресса. Автоустановка программы поддерживает Windows; для других ОС доступна официальная инструкция. Текущий провайдер выбирается по приоритету включённых и проверенных провайдеров.
+
 ## 🚀 Использование
 
 1. Откройте **Настройки → Перевод цепочки размышлений**

package/README.zh-CN.md

@@ -69,6 +69,10 @@ New-Item -ItemType Junction -Path "$HOME\.dsh\profiles\node_modules\dsh-think-tr
 
 DSH **0.2.0-rc.2** 请使用 **dsh-think-translate 1.2.5 或更新版本**。1.2.5 修复了 `todo_write` 的 slot 注册冲突，并声明支持 0.2 版本范围。执行 `dsh plugin --profile web add dsh-think-translate@latest` 后重启 DSH，无需添加版本豁免。
 
+**1.2.6** 修复 DSH 0.2 中思考与回答重复显示的问题：助手渲染器现在遵守 reasoning/response 分组，出错时的原文回退也按组显示。
+
+**1.2.7 本地准备：**已启用的 Ollama 排到第一位，或点击其测试按钮时，会自动查找已有安装并启动、加载模型。缺少软件或模型时，点击首次准备入口后才安装/下载，并显示分阶段进度。自动安装软件支持 Windows，其他系统提供官方指南。“当前使用”按优先级选择第一个已启用且验证成功的提供方，测试后面的行不会覆盖已成功的前面一行。
+
 ## 🚀 使用
 
 1. 打开 **设置 → 思考链翻译**

package/lib/client.js

@@ -23,7 +23,7 @@ window.__ModuleLoader__.load({
 		// there yet (an older process still serving the port), and the contract test
 		// pins PLUGIN_VERSION to package.json so the two can never drift.
 		var PLUGIN_NAME = "dsh-think-translate";
-		var PLUGIN_VERSION = "1.2.5";
+		var PLUGIN_VERSION = "1.2.8";
 		var PLUGIN_REPO = "https://github.com/UncleK/dsh-think-translate";
 		// The official MarkdownText contract changed in 0.1.2-alpha: it now takes a
 		// nested `labels` object and dereferences it without defaults inside its code
@@ -665,7 +665,14 @@ window.__ModuleLoader__.load({
 		var MEM_KEY = "dsh-xlate-cache:";
 		function cacheMemGet(key) {
 			var v = mem.get(key);
-			if (v !== undefined) return v;
+			if (v !== undefined) {
+				if (v.ok === false) {
+					if (Date.now() < v.retryAt) return v;
+					mem.delete(key);
+					return undefined;
+				}
+				return v;
+			}
 			try {
 				var lv = localStorage.getItem(MEM_KEY + key);
 				if (lv !== null) {
@@ -705,6 +712,7 @@ window.__ModuleLoader__.load({
 		function cacheMemSet(key, entry) {
 			mem.set(key, entry);
 			if (mem.size > 800) mem.delete(mem.keys().next().value);
+			if (entry && !entry.ok) return;
 			try {
 				var stamped = entry && typeof entry === "object" ? Object.assign({ at: Date.now() }, entry) : entry;
 				localStorage.setItem(MEM_KEY + key, JSON.stringify(stamped));
@@ -714,6 +722,17 @@ window.__ModuleLoader__.load({
 		}
 		var pend = new Map();
 		var stats = { ok: 0, fail: 0, lastError: null, lastProvider: null, lastModel: null };
+		var providerChecks = {};
+		var providerTestSequence = {};
+		var statusListeners = new Set();
+		function providerFingerprint(p) {
+			return fnv(JSON.stringify({ type: p.type, baseURL: p.baseURL, model: p.model, apiKeyEnv: p.apiKeyEnv,
+				apiKey: p.source === "dsh" || p.apiKeyEnv ? undefined : p.apiKey }));
+		}
+		function recordCheck(id, ok, model, signature) {
+			providerChecks[id] = { ok: ok, model: model || null, signature: signature };
+			statusListeners.forEach(function (notify) { notify(); });
+		}
 		// Diagnostic ledger surfaced in the settings panel (2026-08-24): bundle
 		// version, apply execution, per-slot registration outcome, current lang.
 		var diag = { version: "v5-mode", applied: false, lang: "?", slots: {}, lastLangChange: 0 };
@@ -838,14 +857,21 @@ window.__ModuleLoader__.load({
 		}
 
 		async function hostAttempt(text, target) {
+			// The host may first load a local model (up to 120s), then infer or
+			// walk its fallback providers. Bound the full route, not each hop.
 			var res = await fetchWithTimeout("/_xlate/translate", {
 				method: "POST",
 				headers: { "Content-Type": "application/json" },
 				body: JSON.stringify({ text: text, target: target })
-			}, 30000);
+			}, 180000);
 			if (!res.ok) throw new Error("route HTTP " + res.status);
 			var j = await res.json();
-			if (!(j && j.ok)) throw new Error("route ok=false");
+			if (j && Array.isArray(j.attempts)) j.attempts.forEach(function (attempt) { recordCheck(attempt.provider, attempt.ok, attempt.model, attempt.signature); });
+			if (!(j && j.ok)) {
+				var error = new Error(j && j.error || "route ok=false");
+				if (j && typeof j.retryAt === "number") error.retryAt = j.retryAt;
+				throw error;
+			}
 			return j;
 		}
 
@@ -863,14 +889,13 @@ window.__ModuleLoader__.load({
 						return skipped;
 					}
 					// Host chain with bounded retries: a single transient google/proxy
-					// blip must not flip the row into its permanent failed state
-					// (2026-08-22: one blip used to kill translation for the whole block).
+					// blip must not flip the row into its permanent failed state.
 					var r = null, lastErr = null;
-					var backoff = [0, 400, 1500];
+					var backoff = [0, 400];
 					for (var i = 0; i < backoff.length; i++) {
 						if (backoff[i] > 0) await sleepMs(backoff[i]);
 						try { r = await hostAttempt(text, target); break; }
-						catch (e) { r = null; lastErr = e; }
+						catch (e) { r = null; lastErr = e; if (e && e.retryAt > Date.now()) break; }
 					}
 					if (!(r && r.ok)) {
 						try {
@@ -881,6 +906,7 @@ window.__ModuleLoader__.load({
 						} catch (e2) { r = null; if (e2 && e2.message) lastErr = e2; }
 					}
 					if (r && r.ok) {
+						if (r.providerSignature) recordCheck(r.provider, true, r.model, r.providerSignature);
 						stats.ok++; stats.lastError = null; stats.lastProvider = r.provider; stats.lastModel = r.model || null;
 						memSet(key, r);
 						return r;
@@ -889,7 +915,9 @@ window.__ModuleLoader__.load({
 				} catch (e) {
 					stats.fail++;
 					stats.lastError = e instanceof Error ? e.message : String(e);
-					return { ok: false, text: text, skipped: false };
+					var failedResult = { ok: false, text: text, skipped: false, retryAt: Math.max(Date.now() + 30000, lastErr && lastErr.retryAt || 0) };
+					memSet(key, failedResult);
+					return failedResult;
 				}
 			})();
 			pend.set(key, p);
@@ -910,21 +938,18 @@ window.__ModuleLoader__.load({
 			useEffect(function () {
 				if (!active || key === null) { setEntry(undefined); return undefined; }
 				var alive = true;
-				var hit = cacheMemGet(key);
-				if (hit !== undefined) { setEntry(hit); return undefined; }
-				requestTranslate(text, s.lang).then(function (r) {
+				var retryTimer = null;
+				function attempt() {
 					if (!alive) return;
-					setEntry(r);
-					if (!(r && r.ok)) {
-						// One delayed retry for answer text: failures are not cached, so
-						// this re-issues a real request after the backoff window.
-						setTimeout(function () {
-							if (!alive) return;
-							requestTranslate(text, s.lang).then(function (r2) { if (alive) setEntry(r2); });
-						}, 3000);
-					}
-				});
-				return function () { alive = false; };
+					retryTimer = null;
+					requestTranslate(text, s.lang).then(function (r) {
+						if (!alive) return;
+						setEntry(r);
+						if (!(r && r.ok)) retryTimer = setTimeout(attempt, Math.max(1, (r && r.retryAt || Date.now() + 30000) - Date.now()));
+					});
+				}
+				attempt();
+				return function () { alive = false; if (retryTimer !== null) clearTimeout(retryTimer); };
 			}, [key, active]);
 			return entry;
 		}
@@ -1013,7 +1038,7 @@ window.__ModuleLoader__.load({
 				(async function () {
 					while (alive) {
 						var cur = tRef.current;
-						if (cur.failed || cur.done) return;
+						if (cur.done) return;
 						if (cur.busy) { await sleepMs(150); continue; }
 						if (cur.retryAt && Date.now() < cur.retryAt) { await sleepMs(300); continue; }
 						var latest = stateRef.current.text;
@@ -1047,19 +1072,15 @@ window.__ModuleLoader__.load({
 						mutate(function (p) {
 							var next = Object.assign({}, p, { busy: false });
 							if (r && r.ok) {
+								next.failed = false;
+								next.retryAt = 0;
 								if (!r.skipped) next.transText = p.transText + r.text;
 								next.segIdx = Math.max(p.segIdx, nextIdx);
 							} else {
-								var fc = (p.failCount || 0) + 1;
-								if (fc >= 3) {
-									next.failed = true;
-								} else {
-									next.failCount = fc;
-									next.retryAt = Date.now() + 2500;
-									setTimeout(function () {
-										mutate(function (q) { return Object.assign({}, q, { retryAt: 0 }); });
-									}, 2600);
-								}
+								// A negative-cache hit is a cooldown, not another failed
+								// network attempt. Resume automatically after it expires.
+								next.failed = true;
+								next.retryAt = Math.max(Date.now() + 2500, r && r.retryAt || Date.now() + 30000);
 							}
 							return next;
 						});
@@ -1227,6 +1248,7 @@ window.__ModuleLoader__.load({
 		function AssistantView(props) {
 			var node = props.node;
 			var data = node.data;
+			var groupPart = props.groupPart;
 			var useTurnData = props.useTurnData;
 			var tt = UI_TEXT[useStore().lang] || UI_TEXT.en;
 			var tail = useTurnData("turn-tail");
@@ -1247,6 +1269,10 @@ window.__ModuleLoader__.load({
 			for (var i = 0; i < blocks.length; i++) {
 				var block = blocks[i];
 				if (!block) continue;
+				// DSH may mount this step separately for reasoning and response.
+				// Keep original block indexes so streaming-tail detection is unchanged.
+				if (groupPart === "reasoning" && block.kind !== "reasoning") continue;
+				if (groupPart === "response" && block.kind === "reasoning") continue;
 				if (block.kind === "text") {
 					rendered.push(createElement(TextBlockSafe, { key: "t" + i, text: block.text, streaming: streaming }));
 				} else if (block.kind === "reasoning") {
@@ -1263,15 +1289,19 @@ window.__ModuleLoader__.load({
 					rendered.push(createElement("pre", { key: "u" + i, className: "xl-pre" }, JSON.stringify(block, null, 2)));
 				}
 			}
+			var showStopped = interrupted && (groupPart === undefined || groupPart === "response" ||
+				!blocks.some(function (b) { return b && b.kind !== "reasoning" && b.kind !== "tool-call"; }));
 			return createElement("div", null, rendered,
-				interrupted ? createElement("span", { className: "xl-warn" }, (props.t && props.t("message.stopped")) || tt.stopped) : null);
+				showStopped ? createElement("span", { className: "xl-warn" }, (props.t && props.t("message.stopped")) || tt.stopped) : null);
 		}
 
-		function originalBlocksText(blocks) {
+		function originalBlocksText(blocks, groupPart) {
 			var parts = [];
 			for (var i = 0; i < blocks.length; i++) {
 				var b = blocks[i];
 				if (!b) continue;
+				if (groupPart === "reasoning" && b.kind !== "reasoning") continue;
+				if (groupPart === "response" && b.kind === "reasoning") continue;
 				if ((b.kind === "text" || b.kind === "reasoning") && typeof b.text === "string" && b.text.trim()) parts.push(b.text);
 			}
 			return parts;
@@ -1294,7 +1324,7 @@ window.__ModuleLoader__.load({
 			try {
 				var node = props && props.node;
 				var blocks = node && node.data && Array.isArray(node.data.blocks) ? node.data.blocks : [];
-				var parts = originalBlocksText(blocks);
+				var parts = originalBlocksText(blocks, props.groupPart);
 				if (parts.length) fb = createElement("div", null, createElement(Markdownish, { text: parts.join("\n\n") }));
 			} catch (e) { fb = null; }
 			return createElement(XBoundary, { fallback: fb }, createElement(AssistantView, props));
@@ -1479,11 +1509,16 @@ window.__ModuleLoader__.load({
 			// 缓存条目 counter kept the old number until some other state moved.
 			var tickPair = useState(0);
 			var bumpTick = tickPair[1];
-			// "当前使用" answers the question the block exists for: which provider and
-			// model the last successful translation actually went through.
-			var inUse = stats.lastProvider
-				? ((t.labels[stats.lastProvider] || stats.lastProvider) + (stats.lastModel ? " · " + stats.lastModel : ""))
-				: t.providerDesc;
+			useEffect(function () { var notify = function () { bumpTick(function (n) { return n + 1; }); }; statusListeners.add(notify); return function () { statusListeners.delete(notify); }; }, []);
+			// A lower-ranked provider test cannot displace a verified higher-ranked one.
+			var currentCfg = draft.cfg || {};
+			var ids = (currentCfg.chain || []).concat(currentCfg.fallback && currentCfg.fallback.enabled ? currentCfg.fallback.chain || [] : []);
+			var current = null;
+			for (var i = 0; i < ids.length; i++) {
+				var p = currentCfg.providers && currentCfg.providers[ids[i]], check = providerChecks[ids[i]];
+				if (p && (p.enabled || p.source === "dsh") && check && check.ok && check.signature === providerFingerprint(p)) { current = { id: ids[i], model: check.model }; break; }
+			}
+			var inUse = current ? ((t.labels[current.id] || current.id) + (current.model ? " · " + current.model : "")) : (LOCAL_EMPTY[s.lang] || LOCAL_EMPTY.en);
 			return createElement("div", null,
 				// The clear button rides the "当前使用" row, pushed to that row's end:
 				// one row per idea (what is in use, then the counters), no stray rows.
@@ -1591,6 +1626,24 @@ window.__ModuleLoader__.load({
 		// they belong to the instance that owns the in-flight requests.
 		// ------------------------------------------------------------------
 		var draft = {};
+		var LOCAL_EMPTY = { "zh-CN": "尚未验证可用提供方", en: "No provider verified yet", ja: "提供元は未確認です", ko: "검증된 제공자가 없습니다", es: "Sin proveedor verificado", fr: "Aucun fournisseur vérifié", de: "Noch kein Anbieter geprüft", ru: "Провайдер пока не проверен" };
+		var LOCAL_PHASES = ["checking", "needs_install", "downloading_ollama", "verifying", "installing", "starting", "needs_model", "downloading_model", "loading_model", "ready", "error", "setup", "retry", "manual"];
+		var LOCAL_LABELS = {
+			"zh-CN": ["检查已有 Ollama", "未找到 Ollama", "下载 Ollama", "验证安装包", "安装 Ollama", "启动 Ollama", "所选模型尚未下载", "下载模型", "加载模型", "本地模型已就绪", "准备失败", "一键安装并准备（含模型下载）", "重试准备", "打开官方安装指南"],
+			en: ["Checking existing Ollama", "Ollama not found", "Downloading Ollama", "Verifying installer", "Installing Ollama", "Starting Ollama", "Selected model is not downloaded", "Downloading model", "Loading model", "Local model is ready", "Preparation failed", "Install and prepare (includes model download)", "Retry preparation", "Official installation guide"],
+			ja: ["既存の Ollama を確認", "Ollama が見つかりません", "Ollama をダウンロード", "インストーラーを検証", "Ollama をインストール", "Ollama を起動", "モデルは未ダウンロード", "モデルをダウンロード", "モデルをロード", "ローカルモデル準備完了", "準備に失敗", "インストールして準備（モデルを含む）", "準備を再試行", "公式インストールガイド"],
+			ko: ["기존 Ollama 확인", "Ollama를 찾을 수 없음", "Ollama 다운로드", "설치 파일 검증", "Ollama 설치", "Ollama 시작", "모델이 다운로드되지 않음", "모델 다운로드", "모델 로드", "로컬 모델 준비 완료", "준비 실패", "설치 및 준비 (모델 다운로드 포함)", "준비 다시 시도", "공식 설치 안내"],
+			es: ["Comprobando Ollama", "Ollama no encontrado", "Descargando Ollama", "Verificando instalador", "Instalando Ollama", "Iniciando Ollama", "Modelo sin descargar", "Descargando modelo", "Cargando modelo", "Modelo local listo", "Preparación fallida", "Instalar y preparar (incluye el modelo)", "Reintentar preparación", "Guía oficial de instalación"],
+			fr: ["Vérification de Ollama", "Ollama introuvable", "Téléchargement de Ollama", "Vérification du programme", "Installation de Ollama", "Démarrage de Ollama", "Modèle non téléchargé", "Téléchargement du modèle", "Chargement du modèle", "Modèle local prêt", "Échec de préparation", "Installer et préparer (modèle inclus)", "Réessayer", "Guide officiel"],
+			de: ["Ollama prüfen", "Ollama nicht gefunden", "Ollama herunterladen", "Installationsdatei prüfen", "Ollama installieren", "Ollama starten", "Modell nicht heruntergeladen", "Modell herunterladen", "Modell laden", "Lokales Modell bereit", "Vorbereitung fehlgeschlagen", "Installieren und vorbereiten (mit Modell)", "Erneut vorbereiten", "Offizielle Installationsanleitung"],
+			ru: ["Проверка Ollama", "Ollama не найден", "Загрузка Ollama", "Проверка установщика", "Установка Ollama", "Запуск Ollama", "Модель не скачана", "Скачивание модели", "Загрузка модели", "Локальная модель готова", "Ошибка подготовки", "Установить и подготовить (с моделью)", "Повторить подготовку", "Официальная инструкция"],
+		};
+		function localLabel(phase, lang) { var index = LOCAL_PHASES.indexOf(phase); return (LOCAL_LABELS[lang] || LOCAL_LABELS.en)[index < 0 ? 0 : index]; }
+		function localOrigin(p) {
+			try { var u = new URL(p && p.baseURL); return p.type === "openai" && u.protocol === "http:" &&
+				["localhost", "127.0.0.1", "[::1]"].indexOf(u.hostname) >= 0 && u.port === "11434" && /^\/v1\/?$/.test(u.pathname) && !u.username && !u.password && !u.search && !u.hash ? u.origin : null; }
+			catch (e) { return null; }
+		}
 		function useDraft(key, initial) {
 			var pair = useState(key in draft ? draft[key] : initial);
 			if (!(key in draft)) draft[key] = pair[0];
@@ -1614,6 +1667,10 @@ window.__ModuleLoader__.load({
 			var cfg = pair[0], setCfg = pair[1];
 			var msgPair = useDraft("msg", "");
 			var msg = msgPair[0], setMsg = msgPair[1];
+			var localPair = useDraft("localJob", null);
+			var localJob = localPair[0];
+			var localJobRef = useState(function () { return { current: localJob && localJob.id }; })[0];
+			function setLocalJob(job) { localJobRef.current = job && job.id; localPair[1](job); }
 			var modelsPair = useDraft("models", null);
 			var models = modelsPair[0], setModels = modelsPair[1];
 			var modelMenuPair = useDraft("modelMenu", null); var modelMenu = modelMenuPair[0], setModelMenu = modelMenuPair[1];
@@ -1662,6 +1719,52 @@ window.__ModuleLoader__.load({
 			// not overwrite a newer optimistic state (Important-3/4).
 			var seqRefPair = useState(function () { return { current: 0 }; });
 			var seqRef = seqRefPair[0];
+			var localSeq = useState(function () { return { current: 0 }; })[0];
+			function testProvider(id, sequence) {
+				var testSequence = (providerTestSequence[id] || 0) + 1;
+				providerTestSequence[id] = testSequence;
+				var testedProvider = cfgRef.current.providers[id] || {};
+				var testedModel = testedProvider.model;
+				var testedSignature = providerFingerprint(testedProvider);
+				setMsg(t.testing);
+				return fetch("/_xlate/translate", { method: "POST", headers: { "Content-Type": "application/json" },
+					body: JSON.stringify({ text: "Hello world, connection test. [" + Date.now().toString(36) + "]", target: s.lang, provider: id }) })
+					.then(function (r) { if (!r.ok) throw new Error("HTTP " + r.status); return r.json(); })
+					.then(function (j) { if (sequence !== undefined && sequence !== localSeq.current) return;
+						if (providerTestSequence[id] !== testSequence) return;
+						var currentProvider = cfgRef.current.providers[id] || {};
+						if (providerFingerprint(currentProvider) !== testedSignature || (j.model && testedModel && j.model !== testedModel)) return;
+						recordCheck(id, !!(j && j.ok), j.model || testedModel, j.providerSignature || testedSignature);
+						if (j && j.ok) stats.lastError = null;
+						setMsg(j && j.ok ? t.connOk + " ✓ (" + providerLabel(id, (cfgRef.current.providers || {})[id] || {}, t) + (j.model ? " · " + j.model : "") + ")" : (t.connFail + ": " + (j.error || "unknown"))); })
+					.catch(function (e) { if (sequence !== undefined && sequence !== localSeq.current) return;
+						if (providerTestSequence[id] !== testSequence || providerFingerprint(cfgRef.current.providers[id] || {}) !== testedSignature) return;
+						recordCheck(id, false, testedModel, testedSignature);
+						setMsg(t.connFail + ": " + (e && e.message ? e.message : e)); });
+			}
+			function prepareLocal(allowSetup, testAfter, downloadOnly) {
+				var sequence = ++localSeq.current;
+				return fetch("/_xlate/local/prepare", { method: "POST", headers: { "Content-Type": "application/json" },
+					body: JSON.stringify({ allowInstall: allowSetup === true, allowDownload: allowSetup === true || downloadOnly === true }) })
+					.then(function (r) { return r.json().then(function (j) { if (!r.ok || !j.ok) throw new Error(j.error || "HTTP " + r.status); return j; }); })
+					.then(function (j) { if (sequence !== localSeq.current) return; setLocalJob(j.job); if (testAfter) return testProvider("openai", sequence); })
+					.catch(function (e) { if (sequence === localSeq.current) setMsg(localLabel("error", s.lang) + ": " + (e.message || e)); });
+			}
+			var localProvider = cfg && cfg.providers && cfg.providers.openai;
+			var firstEnabled = cfg && (cfg.chain || []).filter(function (id) { var p = cfg.providers[id]; return p && (p.enabled || p.source === "dsh"); })[0];
+			var localKey = firstEnabled === "openai" && localOrigin(localProvider) && localProvider.model ? localProvider.baseURL + "|" + localProvider.model : "";
+			useEffect(function () { if (localKey) prepareLocal(false, false); }, [localKey]);
+			useEffect(function () {
+				if (!localJob || !localJob.running) return;
+				var alive = true;
+				var timer = setInterval(function () {
+					fetch("/_xlate/local/status?id=" + encodeURIComponent(localJob.id)).then(function (r) { return r.json(); }).then(function (j) {
+						if (alive && j.job && localJobRef.current === localJob.id) setLocalJob(j.job);
+					}).catch(function () {});
+				}, 750);
+				return function () { alive = false; clearInterval(timer); };
+			}, [localJob ? localJob.id + "|" + localJob.running : ""]);
+			useEffect(function () { return function () { localSeq.current++; }; }, []);
 			// Add/edit inline form state. Lives here (not in a nested component,
 			// and before the cfg guard) so re-renders — including the cfg=null →
 			// cfg=value transition — never reset or reorder what the user types;
@@ -1779,20 +1882,9 @@ window.__ModuleLoader__.load({
 			var startPull = function (m) {
 				if (!m || !m.trim()) return;
 				var name = m.trim();
-				setPull({ model: name, status: "starting", percent: 0, done: false, error: null });
-				// The POST only starts the job (the 1 s poll reports progress), so a
-				// rejected start must surface here — otherwise the panel sits in
-				// "downloading" forever and the user is never told why.
-				fetch("/_xlate/model/pull", { method: "POST", headers: { "Content-Type": "application/json" }, body: JSON.stringify({ model: name }) })
-					.then(function (r) {
-						return r.json().catch(function () { return {}; }).then(function (j) {
-							if (!r.ok || (j && j.ok === false)) throw new Error((j && j.error) || ("HTTP " + r.status));
-							return j;
-						});
-					})
-					.catch(function (e) {
-						setPull({ model: name, status: "error", percent: 0, done: false, error: e && e.message ? e.message : "network" });
-					});
+				// Download intent prepares an existing service first. Software install
+				// still needs the separately labelled setup button if Ollama is absent.
+				pickModel(name).then(function (saved) { if (saved) prepareLocal(false, false, true); });
 			};
 			var pickModel = function (m) {
 				if (!m) return;
@@ -1800,9 +1892,9 @@ window.__ModuleLoader__.load({
 				var openai = (cur.providers && cur.providers.openai) || {};
 				applyLocal({ providers: { openai: { model: m } } });
 				setModelMenu(null);
-				fetch("/_xlate/config", { method: "POST", headers: { "Content-Type": "application/json" }, body: JSON.stringify({ providers: { openai: { model: m } } }) })
+				return fetch("/_xlate/config", { method: "POST", headers: { "Content-Type": "application/json" }, body: JSON.stringify({ providers: { openai: { model: m } } }) })
 					.then(function (r) { if (!r.ok) throw new Error("HTTP " + r.status); return r.json(); })
-					.then(function (c) { cfgRef.current = c; setCfg(c); })
+					.then(function (c) { cfgRef.current = c; setCfg(c); return c; })
 					.catch(function (e) {
 						// Do not leave the chip showing a model the host never stored.
 						setMsg((t.providerSaveError || t.connFail) + ": " + (e && e.message ? e.message : e));
@@ -1870,12 +1962,8 @@ window.__ModuleLoader__.load({
 					saveCfg({ chain: chain });
 				},
 				test: function (id) {
-					setMsg(t.testing);
-					fetch("/_xlate/translate", { method: "POST", headers: { "Content-Type": "application/json" },
-						body: JSON.stringify({ text: "Hello world, connection test. [" + Date.now().toString(36) + "]", target: s.lang, provider: id }) })
-						.then(function (r) { return r.json(); })
-						.then(function (j) { setMsg(j && j.ok ? t.connOk + " ✓ (" + providerLabel(id, (cfgRef.current.providers || {})[id] || {}, t) + (j.model ? " · " + j.model : "") + ")" : (t.connFail + ": " + (j.error || "unknown"))); })
-						.catch(function (e) { setMsg(t.connFail + ": " + (e && e.message ? e.message : e)); });
+					if (id === "openai" && localOrigin(cfgRef.current.providers.openai)) prepareLocal(false, true);
+					else testProvider(id);
 				},
 				edit: function (id) {
 					var p = cfgRef.current.providers && cfgRef.current.providers[id];
@@ -2229,6 +2317,17 @@ window.__ModuleLoader__.load({
 					: createElement("button", { className: "xl-btn", type: "button", onClick: rescanDsh, style: { marginTop: "4px", marginLeft: "6px" } },
 						"↻ " + t.dshImport),
 				unusedArea,
+				localJob && localProvider && localJob.model === localProvider.model && localJob.origin === localOrigin(localProvider)
+					? createElement("div", { className: "xl-local-status", role: "status", style: { marginTop: "8px" } },
+						createElement("div", { className: "xl-note" }, localLabel(localJob.phase, s.lang) + " · " + localJob.model +
+							(localJob.percent !== null && localJob.phase.indexOf("downloading") === 0 ? " · " + localJob.percent + "%" : "") +
+							(localJob.completed ? " · " + (localJob.completed / 1048576).toFixed(1) + " MB" + (localJob.total ? " / " + (localJob.total / 1048576).toFixed(1) + " MB" : "") : "")),
+						localJob.running ? createElement("div", { className: "xl-progress" }, createElement("div", { className: "xl-progress-bar", style: { width: (localJob.percent === null ? 15 : localJob.percent) + "%" } })) : null,
+						localJob.error ? createElement("div", { className: "xl-warn" }, localJob.error) : null,
+						!localJob.running && (localJob.phase === "needs_model" || localJob.phase === "error" || (localJob.phase === "needs_install" && localJob.canInstall))
+							? createElement("button", { className: "xl-btn", onClick: function () { prepareLocal(true, true); } }, localLabel(localJob.phase === "error" ? "retry" : "setup", s.lang)) : null,
+						localJob.phase === "needs_install" && !localJob.canInstall ? createElement("a", { href: "https://ollama.com/download", target: "_blank", rel: "noreferrer" }, localLabel("manual", s.lang)) : null)
+					: null,
 				msg ? createElement("div", { className: "xl-note", style: { marginTop: "6px" } }, msg) : null);
 		}
 

package/lib/index.js

@@ -1,16 +1,17 @@
 // dsh-think-translate host half: same-origin translation proxy with a pluggable
 // provider chain. The browser calls POST /_xlate/translate (no CORS involved);
 // this side owns provider adapters, API keys (local config file), caching,
-// and the system-proxy-aware curl transport.
+// and the system-proxy-aware native transport.
 //
-// Providers: google (gtx via Node CONNECT tunnel / curl), bing (ttranslatev3
-// via curl), openai-compatible (any /chat/completions endpoint — usually the
-// local Ollama).
-import { readFile, writeFile, mkdir, copyFile } from 'node:fs/promises'
-import { dirname } from 'node:path'
+// Providers: google (gtx), bing (ttranslatev3), openai-compatible (any
+// /chat/completions endpoint — usually the local Ollama).
+import { readFile, writeFile, mkdir, copyFile, access, unlink, stat } from 'node:fs/promises'
+import { dirname, join } from 'node:path'
+import { tmpdir } from 'node:os'
+import { randomUUID } from 'node:crypto'
+import { createLocalPreparation, ollamaOrigin, localFirst, validOllamaSignature } from './ollama.js'
 import { fileURLToPath } from 'node:url'
-import http from 'node:http'
-import https from 'node:https'
+import { requestText, isLoopback } from './transport.js'
 
 export const name = 'dsh-think-translate'
 
@@ -85,6 +86,133 @@ const VALID_TYPES = ['google', 'bing', 'openai', 'anthropic']
 let subprocessRef = null
 let timerRef = null
 
+async function localJson(url, options = {}, timeoutMs = 10000) {
+  const response = await fetch(url, { ...options, redirect: 'error', signal: AbortSignal.timeout(timeoutMs) })
+  const body = await response.json()
+  if (!response.ok || body.error) throw new Error(body.error || 'Ollama HTTP ' + response.status)
+  return body
+}
+
+async function findOllamaExecutable() {
+  if (!subprocessRef) throw new Error('subprocess unavailable')
+  try { return await subprocessRef.resolveExecutable('ollama') } catch { /* Try installed-app locations when PATH is stale. */ }
+  const candidates = process.platform === 'win32'
+    ? [join(process.env.LOCALAPPDATA || '', 'Programs', 'Ollama', 'ollama.exe')]
+    : ['/usr/local/bin/ollama', '/usr/bin/ollama', '/Applications/Ollama.app/Contents/Resources/ollama']
+  if (process.platform === 'win32') for (const hive of ['HKCU', 'HKLM']) {
+    try {
+      const key = hive + ':\\Software\\Microsoft\\Windows\\CurrentVersion\\Uninstall\\{44E83376-CE68-45EB-8FC1-393500EB558C}_is1'
+      const result = await runProgram('powershell.exe', ['-NoProfile', '-NonInteractive', '-Command',
+        '[Console]::OutputEncoding=[Text.UTF8Encoding]::new($false);$p=Get-ItemProperty -LiteralPath ' + "'" + key + "'" + ' -ErrorAction Stop;[Console]::Write($p.InstallLocation)'], 8000)
+      if (result.trim()) candidates.push(join(result.trim(), 'ollama.exe'))
+    } catch { /* This hive may not contain an Ollama installation. */ }
+  }
+  for (const candidate of candidates) {
+    try { await access(candidate); return candidate } catch { /* Probe the next known installation location. */ }
+  }
+  return null
+}
+
+async function startOllamaServer(executable, origin) {
+  const child = subprocessRef.spawn({
+    argv: [executable, 'serve'], cwd: await workdir(), graceMs: 2000,
+    stdio: { stdin: 'ignore', stdout: 'ignore', stderr: 'pipe' },
+    env: { OLLAMA_HOST: new URL(origin).host },
+  })
+  // Drain the long-lived pipe; the subprocess provider owns shutdown and joins.
+  if (child.stderr) (async () => { for await (const chunk of child.stderr) { void chunk } })().catch(() => {})
+  child.done.catch(() => {})
+}
+
+async function streamModelPull(origin, model, progress) {
+  const response = await fetch(origin + '/api/pull', {
+    method: 'POST', headers: { 'Content-Type': 'application/json' }, redirect: 'error',
+    body: JSON.stringify({ model, stream: true }), signal: AbortSignal.timeout(60 * 60 * 1000),
+  })
+  if (!response.ok) throw new Error('Ollama download HTTP ' + response.status)
+  const decoder = new TextDecoder()
+  const layers = new Map()
+  let buffer = '', success = false
+  const consume = line => {
+    if (!line.trim()) return
+    const event = JSON.parse(line)
+    if (event.error) throw new Error(event.error)
+    if (event.digest && typeof event.total === 'number') layers.set(event.digest, {
+      total: event.total, completed: typeof event.completed === 'number' ? event.completed : 0,
+    })
+    let total = 0, completed = 0
+    for (const layer of layers.values()) { total += layer.total; completed += layer.completed }
+    progress({ total, completed, percent: total ? Math.min(100, Math.round(completed / total * 100)) : null })
+    if (event.status === 'success') success = true
+  }
+  for await (const bytes of response.body) {
+    buffer += decoder.decode(bytes, { stream: true })
+    let index
+    while ((index = buffer.indexOf('\n')) >= 0) { consume(buffer.slice(0, index)); buffer = buffer.slice(index + 1) }
+  }
+  consume(buffer + decoder.decode())
+  if (!success) throw new Error('Model download ended before Ollama reported success')
+}
+
+async function installOllama(progress) {
+  const file = join(tmpdir(), 'dsh-ollama-' + randomUUID() + '.exe')
+  try {
+    progress({ phase: 'downloading_ollama', percent: null, completed: 0, total: 0 })
+    let totalBytes = 0
+    try {
+      const headers = await runProgram('curl', await curlArgs(['--fail', '--head', '--location', 'https://ollama.com/download/OllamaSetup.exe']), 30000)
+      const lengths = [...headers.matchAll(/content-length:\s*(\d+)/gi)]
+      if (lengths.length) totalBytes = Number(lengths[lengths.length - 1][1])
+    } catch { /* A missing HEAD size does not prevent a signed download. */ }
+    const args = await curlArgs(['--fail', '--location', '--proto', '=https', '--proto-redir', '=https',
+      '--progress-bar', '--retry', '2', '--max-time', '1800', '--output', file, 'https://ollama.com/download/OllamaSetup.exe'])
+    // Replace curlArgs' silent/25-second prefix with the download's own limit.
+    const filtered = args.slice(3)
+    const executable = await subprocessRef.resolveExecutable('curl')
+    const child = subprocessRef.spawn({ argv: [executable, ...filtered], cwd: await workdir(), graceMs: 2000,
+      signal: AbortSignal.timeout(30 * 60 * 1000), stdio: { stdin: 'ignore', stdout: 'ignore', stderr: 'pipe' } })
+    let tail = ''
+    if (child.stderr) for await (const bytes of child.stderr) {
+      tail = (tail + new TextDecoder().decode(bytes)).slice(-2000)
+      const matches = [...tail.matchAll(/(\d+(?:\.\d+)?)%/g)]
+      const downloaded = await stat(file).catch(() => null)
+      progress({ completed: downloaded ? downloaded.size : 0, total: totalBytes,
+        percent: matches.length ? Math.min(100, Number(matches[matches.length - 1][1])) : null })
+    }
+    const outcome = await child.done
+    if (outcome.exitCode !== 0) throw new Error('Ollama installer download failed: ' + tail.trim().slice(-300))
+    progress({ phase: 'verifying', percent: null })
+    const quoted = "'" + file.replace(/'/g, "''") + "'"
+    const signature = await runProgram('powershell.exe', ['-NoProfile', '-NonInteractive', '-Command',
+      '$s=Get-AuthenticodeSignature -LiteralPath ' + quoted + '; @{status=[string]$s.Status;subject=$s.SignerCertificate.Subject}|ConvertTo-Json -Compress'], 30000)
+    if (!validOllamaSignature(JSON.parse(signature))) throw new Error('Ollama installer signature verification failed')
+    progress({ phase: 'installing', percent: null })
+    const appState = join(process.env.LOCALAPPDATA, 'Ollama')
+    await mkdir(appState, { recursive: true })
+    await writeFile(join(appState, 'upgraded'), '', 'utf8') // Official installer convention: launch the app hidden.
+    // The official per-user installer supports these unattended flags.
+    const controller = new AbortController()
+    const deadline = setTimeout(() => controller.abort(), 10 * 60 * 1000)
+    try {
+      // Ignore inherited pipes: a background app launched by the installer can
+      // retain them even after the installer itself exits successfully.
+      const child = subprocessRef.spawn({ argv: [file, '/VERYSILENT', '/NORESTART', '/SUPPRESSMSGBOXES'],
+        cwd: await workdir(), graceMs: 2000, signal: controller.signal,
+        stdio: { stdin: 'ignore', stdout: 'ignore', stderr: 'ignore' } })
+      const outcome = await child.done
+      if (outcome.exitCode !== 0) throw new Error('Ollama installation failed with exit code ' + outcome.exitCode)
+    } finally { clearTimeout(deadline) }
+  } finally {
+    await unlink(file).catch(() => {}) // Only this operation's unique installer download.
+  }
+}
+
+const localPreparation = createLocalPreparation({
+  platform: process.platform, json: localJson, findExecutable: findOllamaExecutable,
+  startServer: startOllamaServer, install: installOllama, pull: streamModelPull,
+  wait: milliseconds => new Promise(resolve => setTimeout(resolve, milliseconds)),
+})
+
 async function workdir() {
   return process.cwd() || '.'
 }
@@ -197,84 +325,49 @@ async function getProxy() {
   return value
 }
 
+// curl remains necessary for the signed installer download and legacy SOCKS
+// proxies. Ordinary translation requests use Node directly without subprocesses.
 async function curlArgs(extra) {
   const args = ['-sS', '--max-time', '25']
-  // Loopback targets (local LLM endpoints) must bypass the system proxy.
-  const last = extra.length ? String(extra[extra.length - 1]) : ''
-  const mh = last.match(/^https?:\/\/([^\/?#]+)/i)
-  const host = mh ? mh[1].replace(/^\[/, '').split(':')[0].toLowerCase() : ''
-  const isLocal = host === 'localhost' || host === '127.0.0.1' || host === '::1' || host === '[::1]' || host === '::ffff:127.0.0.1'
-  if (!isLocal) {
-    const px = await getProxy()
-    if (px) args.push('-x', px)
+  const url = extra.length ? String(extra[extra.length - 1]) : ''
+  if (!isLoopback(url)) {
+    const proxy = await getProxy()
+    if (proxy) args.push('-x', proxy)
   }
-  for (const a of extra) args.push(a)
-  return args
+  return args.concat(extra)
 }
 
-async function curlGet(url) {
-  return runProgram('curl', await curlArgs(['-L', url]), 30000)
+async function httpFetch(url, options = {}, timeoutMs = 25000) {
+  const proxy = isLoopback(url) ? null : await getProxy()
+  if (proxy && /^socks/i.test(proxy)) {
+    const extra = ['--fail', '-L', '--max-time', String(Math.ceil(timeoutMs / 1000)), '-X', options.method || 'GET']
+    for (const [key, value] of Object.entries(options.headers || {})) extra.push('-H', key + ': ' + value)
+    if (options.body !== undefined) extra.push('--data-binary', options.body)
+    extra.push(url)
+    return runProgram('curl', await curlArgs(extra), timeoutMs + 5000)
+  }
+  // A failed proxy request is a provider failure. Do not silently bypass the
+  // configured proxy or send a second copy of a potentially billable POST.
+  return requestText(url, { ...options, proxy, timeoutMs })
 }
 
-async function curlPostForm(url, form) {
-  const extra = ['-X', 'POST', '-L', '-H', 'Content-Type: application/x-www-form-urlencoded', '--data-binary', form, url]
-  return runProgram('curl', await curlArgs(extra), 30000)
+async function curlGet(url, timeoutMs = 25000) {
+  return httpFetch(url, { method: 'GET' }, timeoutMs)
 }
 
-async function curlPostJson(url, headerPairs, body) {
-  const extra = ['-X', 'POST']
-  for (const h of headerPairs) extra.push('-H', h)
-  extra.push('--data-binary', body, url)
-  return runProgram('curl', await curlArgs(extra), 40000)
+async function curlPostForm(url, form, timeoutMs = 25000) {
+  return httpFetch(url, {
+    method: 'POST', headers: { 'Content-Type': 'application/x-www-form-urlencoded' }, body: form,
+  }, timeoutMs)
 }
 
-// HTTPS GET through an HTTP proxy via a CONNECT tunnel, using Node's own TLS
-// stack. curl is blocked by google gtx's anti-bot (TLS-fingerprint based:
-// curl gets a "Sorry..." HTML page while .NET/Node/real browsers pass), so the
-// google provider must not go through curl. SOCKS proxies are not supported
-// here — callers keep the curl path for those.
-function httpsViaProxy(urlStr, proxyHost, proxyPort, timeoutMs) {
-  return new Promise((resolve, reject) => {
-    const url = new URL(urlStr)
-    const connectReq = http.request({
-      host: proxyHost,
-      port: proxyPort,
-      method: 'CONNECT',
-      path: url.host + ':443',
-      headers: { Host: url.host + ':443', 'Proxy-Connection': 'Keep-Alive' },
-      timeout: timeoutMs,
-    })
-    connectReq.on('connect', (res, socket) => {
-      if (res.statusCode !== 200) {
-        socket.destroy()
-        reject(new Error('proxy CONNECT ' + res.statusCode))
-        return
-      }
-      const req2 = https.request({
-        host: url.host,
-        path: url.pathname + url.search,
-        method: 'GET',
-        headers: {
-          'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/126.0.0.0 Safari/537.36',
-          'Accept': 'application/json,text/plain,*/*',
-          'Accept-Language': 'zh-CN,zh;q=0.9',
-        },
-        socket,
-        servername: url.host,
-        timeout: timeoutMs,
-      }, (res2) => {
-        const chunks = []
-        res2.on('data', (c) => chunks.push(c))
-        res2.on('end', () => resolve(Buffer.concat(chunks).toString('utf8')))
-      })
-      req2.on('error', reject)
-      req2.on('timeout', () => req2.destroy(new Error('https timeout')))
-      req2.end()
-    })
-    connectReq.on('error', reject)
-    connectReq.on('timeout', () => connectReq.destroy(new Error('proxy connect timeout')))
-    connectReq.end()
-  })
+async function curlPostJson(url, headerPairs, body, timeoutMs = 25000) {
+  const headers = {}
+  for (const header of headerPairs) {
+    const colon = header.indexOf(':')
+    if (colon > 0) headers[header.slice(0, colon).trim()] = header.slice(colon + 1).trim()
+  }
+  return httpFetch(url, { method: 'POST', headers, body }, timeoutMs)
 }
 
 // ---------------------------------------------------------------- providers
@@ -282,38 +375,22 @@ function httpsViaProxy(urlStr, proxyHost, proxyPort, timeoutMs) {
 let bingCache = null
 
 async function viaGoogle(text, to) {
-  // translate.googleapis.com gets TLS-Reset by Clash/GFW in CN networks even
-  // through the proxy; translate.google.com serves the same gtx JSON (with
-  // client=dict-chrome-ex, the Chrome dictionary extension identifier, which
-  // bypasses the 429 anti-bot that now blocks client=gtx) and is reachable
-  // there.  The reverse may hold on other networks, so try both hosts.
   const qs = 'translate_a/single?client=dict-chrome-ex&sl=auto&tl=' + encodeURIComponent(to) + '&dt=t&q=' + encodeURIComponent(text)
   const hosts = ['https://translate.google.com/', 'https://translate.googleapis.com/']
-  const px = await getProxy()
   let lastErr = null
   for (const host of hosts) {
     const url = host + qs
     let raw = null
-    if (px && /^http/i.test(px)) {
-      // curl is fingerprint-blocked by google ("Sorry..." page); Node's TLS
-      // stack passes. Only plain HTTP proxies support CONNECT here; SOCKS falls
-      // back to curl.
-      try {
-        const p = new URL(px)
-        raw = await httpsViaProxy(url, p.hostname, Number(p.port) || 80, 30000)
-        // httpsViaProxy resolves the body even for anti-bot HTML (429
-        // "Sorry..."); treat non-JSON as a failure so we fall through to
-        // curl, which passes through the system proxy (2026-09-01).
-        if (raw !== null) {
-          try { JSON.parse(raw) } catch (e) { raw = null }
-        }
-      } catch (e) {
-        raw = null
+    try {
+      raw = await curlGet(url)
+      if (raw) {
+        try { JSON.parse(raw) } catch { raw = null }
       }
+    } catch (e) {
+      lastErr = e
+      continue
     }
-    if (raw === null) {
-      try { raw = await curlGet(url) } catch (e) { lastErr = e; continue }
-    }
+    if (!raw) continue
     try {
       const arr = JSON.parse(raw)
       const segs = arr && arr[0]
@@ -949,6 +1026,12 @@ function fnv(s) {
 function cacheGet(k) {
   if (!cache.has(k)) return undefined
   const v = cache.get(k)
+  if (v && v.ok === false) {
+    if (Date.now() >= v.retryAt) {
+      cache.delete(k)
+      return undefined
+    }
+  }
   cache.delete(k)
   cache.set(k, v)
   return v
@@ -1029,6 +1112,7 @@ function chunkText(text) {
 // Run one ordered provider-id list; returns { ok } or { ok: false, error }.
 async function runChainFor(ids, text, target, cfg) {
   let lastErr = null
+  const attempts = []
   for (const id of ids) {
     const p = cfg.providers && cfg.providers[id]
     // DSH-sourced providers are activated by chain membership (they cannot be
@@ -1039,6 +1123,7 @@ async function runChainFor(ids, text, target, cfg) {
     const adapter = ADAPTERS[p.type || id]
     if (!adapter) continue
     try {
+      if (ollamaOrigin(id, p)) await localPreparation.ensure(p)
       const chunks = chunkText(text)
       const outs = []
       for (const c of chunks) outs.push(await adapter(c, target, p))
@@ -1046,13 +1131,20 @@ async function runChainFor(ids, text, target, cfg) {
       state.lastProvider = id
       state.lastError = null
       // Carry the concrete model name so the UI can show e.g. "openai: qwen2.5:14b".
-      return { ok: true, text: outs.join('\n\n'), provider: id, ...(p.model ? { model: p.model } : {}) }
+      attempts.push({ provider: id, signature: providerFingerprint(p), ok: true, model: p.model || null })
+      return { ok: true, text: outs.join('\n\n'), provider: id, providerSignature: providerFingerprint(p), attempts, ...(p.model ? { model: p.model } : {}) }
     } catch (e) {
       lastErr = e
+      attempts.push({ provider: id, signature: providerFingerprint(p), ok: false, model: p.model || null })
       if (id === 'bing') bingCache = null
     }
   }
-  return { ok: false, error: lastErr instanceof Error ? lastErr.message : (lastErr || 'no enabled provider in chain') }
+  return { ok: false, attempts, error: lastErr instanceof Error ? lastErr.message : (lastErr || 'no enabled provider in chain') }
+}
+
+function providerFingerprint(p) {
+  return fnv(JSON.stringify({ type: p.type, baseURL: p.baseURL, model: p.model, apiKeyEnv: p.apiKeyEnv,
+    apiKey: p.source === 'dsh' || p.apiKeyEnv ? undefined : p.apiKey }))
 }
 
 async function translateViaChain(text, target, cfg) {
@@ -1061,11 +1153,11 @@ async function translateViaChain(text, target, cfg) {
   let fb = null
   if (cfg.fallback && cfg.fallback.enabled && Array.isArray(cfg.fallback.chain)) {
     fb = await runChainFor(cfg.fallback.chain, text, target, cfg)
-    if (fb.ok) return fb
+    if (fb.ok) return { ...fb, attempts: [...primary.attempts, ...fb.attempts] }
   }
   state.fail++
   state.lastError = primary.error || (fb && fb.error) || 'translate unavailable'
-  return { ok: false, text: text, error: state.lastError }
+  return { ok: false, text: text, attempts: [...primary.attempts, ...(fb ? fb.attempts : [])], error: state.lastError }
 }
 
 async function handleTranslate(body) {
@@ -1086,12 +1178,13 @@ async function handleTranslate(body) {
       const p = cfg.providers && cfg.providers[forced]
       if (!p) return { ok: false, text, error: 'unknown provider: ' + forced }
       // DSH providers are activated by chain membership, not the enabled flag.
-      if (!p.enabled && p.source !== 'dsh') return { ok: false, text, error: 'provider disabled: ' + forced }
+      if (!p.enabled && p.source !== 'dsh' && !ollamaOrigin(forced, p)) return { ok: false, text, error: 'provider disabled: ' + forced }
       const adapter = ADAPTERS[p.type || forced]
       if (!adapter) return { ok: false, text, error: 'unsupported type: ' + p.type }
       try {
+        if (ollamaOrigin(forced, p)) await localPreparation.ensure(p)
         const out = await adapter(text, target, p)
-        return { ok: true, text: out, provider: forced, ...(p.model ? { model: p.model } : {}) }
+        return { ok: true, text: out, provider: forced, providerSignature: providerFingerprint(p), ...(p.model ? { model: p.model } : {}) }
       } catch (e) {
         return { ok: false, text, error: e instanceof Error ? e.message : String(e) }
       }
@@ -1102,7 +1195,8 @@ async function handleTranslate(body) {
       return r
     }
     const r = await translateViaChain(text, target, cfg)
-    if (r.ok) cacheSet(key, r)
+    if (!r.ok) r.retryAt = Date.now() + 30000
+    cacheSet(key, r)
     return r
   })()
   inflight.set(key, job)
@@ -1175,8 +1269,7 @@ async function handleModels() {
       if (!res.ok) throw new Error('models HTTP ' + res.status)
       raw = await res.text()
     } else {
-      const extra = ['-H', 'Authorization: Bearer ' + key, url]
-      raw = await runProgram('curl', await curlArgs(extra), 30000)
+      raw = await httpFetch(url, { headers: { 'Authorization': 'Bearer ' + key } }, 10000)
     }
     const j = JSON.parse(raw)
     const list = Array.isArray(j && j.data)
@@ -1362,6 +1455,7 @@ export function apply(ctx) {
             // provider's output (and its name in the status row) until restart.
             cache.clear()
             inflight.clear()
+            if (localFirst(merged)) localPreparation.start(merged.providers.openai)
             sendJson(res, 200, stripResolvedKeys(merged))
             return
           }
@@ -1374,6 +1468,30 @@ export function apply(ctx) {
     return dispose
   }, 'dsh-think-translate: config route')
 
+  ctx.effect(function () {
+    return webServer.register({ kind: 'exact', path: '/_xlate/local/prepare', handler: async function (req, res) {
+      try {
+        if (req.method !== 'POST') { sendJson(res, 405, { ok: false, error: 'POST only' }); return }
+        const origin = req.headers.origin
+        if (origin && new URL(origin).host !== req.headers.host) { sendJson(res, 403, { ok: false, error: 'same-origin only' }); return }
+        const body = await readBody(req)
+        const cfg = await loadConfig()
+        sendJson(res, 202, { ok: true, job: localPreparation.start(cfg.providers.openai, {
+          allowInstall: body.allowInstall === true, allowDownload: body.allowDownload === true,
+        }) })
+      } catch (error) { sendJson(res, 400, { ok: false, error: error instanceof Error ? error.message : String(error) }) }
+    } })
+  }, 'dsh-think-translate: local preparation route')
+  ctx.effect(function () {
+    return webServer.register({ kind: 'exact', path: '/_xlate/local/status', handler: function (req, res) {
+      if (req.method !== 'GET') { sendJson(res, 405, { ok: false, error: 'GET only' }); return }
+      const id = new URL(req.url, 'http://localhost').searchParams.get('id')
+      sendJson(res, 200, { ok: true, job: localPreparation.snapshot(id) })
+    } })
+  }, 'dsh-think-translate: local preparation status')
+  // Restoring a profile with Ollama already first must prepare it too.
+  loadConfig().then(cfg => { if (localFirst(cfg)) localPreparation.start(cfg.providers.openai) }).catch(() => {})
+
   ctx.effect(function () {
     const dispose = webServer.register({
       kind: 'exact',
@@ -1475,4 +1593,5 @@ export {
   structuredCloneSafe, ADAPTERS,
   resolveApiKey, stripResolvedKeys, missingApiKey, clearNullFields, handleVersion,
   harnessHomeCandidates, pickHarnessHome,
+  streamModelPull,
 }

package/lib/ollama.js

@@ -0,0 +1,132 @@
+// Local Ollama preparation: asynchronous, observable, and reusable by config
+// changes and explicit provider tests. Dependencies are injected for offline tests.
+export function ollamaOrigin(id, provider) {
+  if (id !== 'openai' || !provider || provider.type !== 'openai') return null
+  try {
+    const url = new URL(provider.baseURL)
+    if (url.protocol !== 'http:' || !['localhost', '127.0.0.1', '[::1]'].includes(url.hostname) ||
+        url.port !== '11434' || !/^\/v1\/?$/.test(url.pathname) || url.username || url.password || url.search || url.hash) return null
+    return url.origin
+  } catch { return null }
+}
+
+export function localFirst(cfg) {
+  const id = (cfg.chain || []).find(id => {
+    const p = cfg.providers && cfg.providers[id]
+    return p && (p.enabled || p.source === 'dsh')
+  })
+  const provider = cfg.providers && cfg.providers[id]
+  return ollamaOrigin(id, provider) !== null && typeof provider.model === 'string' && !!provider.model.trim()
+}
+
+export function validOllamaSignature(report) {
+  return report && report.status === 'Valid' && typeof report.subject === 'string' &&
+    /(^|, )O=Ollama Inc\.(,|$)/.test(report.subject)
+}
+
+function modelKey(model) { return model.includes(':') ? model : model + ':latest' }
+
+export function createLocalPreparation(runtime) {
+  const jobs = new Map()
+  const servers = new Map()
+  let sequence = 0
+  let latest = null
+  const snapshot = id => {
+    const job = id ? jobs.get(id) : latest
+    if (!job) return null
+    const { promise, options, ...out } = job
+    return { ...out, platform: runtime.platform, canInstall: runtime.platform === 'win32' }
+  }
+  async function online(origin) {
+    try { const version = await runtime.json(origin + '/api/version', {}, 2000); return typeof version.version === 'string' }
+    catch { return false }
+  }
+  async function server(job, options, progress) {
+    if (await online(job.origin)) return
+    if (servers.has(job.origin)) return servers.get(job.origin)
+    const pending = (async () => {
+      let executable = await runtime.findExecutable()
+      if (!executable) {
+        if (!options.allowInstall || runtime.platform !== 'win32') {
+          progress({ phase: 'needs_install', running: false })
+          return false
+        }
+        await runtime.install(progress)
+        executable = await runtime.findExecutable()
+        if (!executable) throw new Error('Ollama installation finished but the executable was not found')
+      }
+      progress({ phase: 'starting', percent: null })
+      await runtime.startServer(executable, job.origin)
+      for (let n = 0; n < 30; n++) {
+        if (await online(job.origin)) return true
+        await runtime.wait(500)
+      }
+      throw new Error('Ollama did not become reachable after startup')
+    })()
+    servers.set(job.origin, pending)
+    try { return await pending } finally { servers.delete(job.origin) }
+  }
+  async function run(job, options) {
+    const progress = patch => Object.assign(job, patch, { updatedAt: Date.now() })
+    try {
+      if (await server(job, options, progress) === false) {
+        progress({ phase: 'needs_install', running: false })
+        return
+      }
+      const tags = await runtime.json(job.origin + '/api/tags', {}, 10000)
+      const installed = (tags.models || []).some(m => typeof m.name === 'string' && modelKey(m.name) === modelKey(job.model))
+      if (!installed) {
+        if (!options.allowDownload) { progress({ phase: 'needs_model', running: false }); return }
+        progress({ phase: 'downloading_model', completed: 0, total: 0, percent: null })
+        await runtime.pull(job.origin, job.model, progress)
+      }
+      const loaded = await runtime.json(job.origin + '/api/ps', {}, 10000)
+      if (!(loaded.models || []).some(m => modelKey(m.name || m.model || '') === modelKey(job.model))) {
+        progress({ phase: 'loading_model', percent: null, completed: 0, total: 0 })
+        await runtime.json(job.origin + '/api/generate', {
+          method: 'POST', headers: { 'Content-Type': 'application/json' },
+          body: JSON.stringify({ model: job.model, prompt: '', stream: false, keep_alive: '10m' }),
+        }, 120000)
+        const verify = await runtime.json(job.origin + '/api/ps', {}, 10000)
+        if (!(verify.models || []).some(m => modelKey(m.name || m.model || '') === modelKey(job.model))) {
+          throw new Error('Ollama did not report the selected model as loaded')
+        }
+      }
+      progress({ phase: 'ready', running: false, percent: 100, error: null })
+    } catch (error) {
+      progress({ phase: 'error', running: false, error: error instanceof Error ? error.message : String(error) })
+    }
+  }
+  function start(provider, options = {}) {
+    const origin = ollamaOrigin('openai', provider)
+    const model = typeof provider?.model === 'string' ? provider.model.trim() : ''
+    if (!origin || !model) throw new Error('A local Ollama endpoint and model are required')
+    // No automatic installation or model download: the setup button grants both.
+    const allowed = { allowInstall: options.allowInstall === true, allowDownload: options.allowDownload === true }
+    const busy = [...jobs.values()].find(j => j.running && j.origin === origin && j.model === model)
+    if (busy) {
+      busy.options.allowInstall ||= allowed.allowInstall
+      busy.options.allowDownload ||= allowed.allowDownload
+      return snapshot(busy.id)
+    }
+    const job = { id: String(++sequence), origin, model, phase: 'checking', running: true,
+      percent: null, completed: 0, total: 0, error: null, updatedAt: Date.now() }
+    job.options = allowed
+    latest = job
+    jobs.set(job.id, job)
+    job.promise = Promise.resolve().then(() => run(job, allowed))
+    if (jobs.size > 30) for (const [id, old] of jobs) {
+      if (!old.running && old !== latest) { jobs.delete(id); break }
+    }
+    return snapshot(job.id)
+  }
+  async function ensure(provider) {
+    const result = start(provider)
+    const job = jobs.get(result.id)
+    await job.promise
+    if (job.phase !== 'ready') throw new Error(job.error || (job.phase === 'needs_install'
+      ? 'Ollama is not installed; use the local setup button' : 'Selected model is not downloaded; use the local setup button'))
+    return snapshot(job.id)
+  }
+  return { start, snapshot, ensure }
+}

package/lib/transport.js

@@ -0,0 +1,135 @@
+import http from 'node:http'
+import https from 'node:https'
+import tls from 'node:tls'
+import { isIP } from 'node:net'
+
+export function isLoopback(url) {
+  const host = new URL(url).hostname.replace(/^\[|\]$/g, '').toLowerCase()
+  return ['localhost', '127.0.0.1', '::1', '::ffff:7f00:1'].includes(host)
+}
+
+function proxyAuthorization(proxy) {
+  if (!proxy.username && !proxy.password) return {}
+  const credentials = decodeURIComponent(proxy.username) + ':' + decodeURIComponent(proxy.password)
+  return { 'Proxy-Authorization': 'Basic ' + Buffer.from(credentials).toString('base64') }
+}
+
+// One native request, with one deadline covering CONNECT, TLS, and the body.
+// HTTP targets use an ordinary forward proxy; HTTPS targets use its CONNECT
+// socket through a dedicated TLS agent, never a second direct connection.
+function requestOnce(url, options) {
+  return new Promise((resolve, reject) => {
+    const proxy = options.proxy ? new URL(options.proxy) : null
+    if (proxy && !['http:', 'https:'].includes(proxy.protocol)) {
+      reject(new Error('Unsupported native proxy protocol: ' + proxy.protocol))
+      return
+    }
+    let request, connectRequest, tunnel, agent, settled = false
+    const deadline = setTimeout(() => finish(new Error('request timeout')), options.timeoutMs)
+    function finish(error, value) {
+      if (settled) return
+      settled = true
+      clearTimeout(deadline)
+      // Destroy unfinished requests and the operation's private tunnel only.
+      // Direct requests keep Node's shared connection pool available.
+      if (error) request?.destroy()
+      connectRequest?.destroy()
+      tunnel?.destroy()
+      agent?.destroy()
+      if (error) reject(error)
+      else resolve(value)
+    }
+    function send(client, requestOptions) {
+      if (settled) return
+      try { request = client.request(requestOptions, response => {
+        const chunks = []
+        response.on('data', chunk => chunks.push(chunk))
+        response.on('error', error => finish(error))
+        response.on('aborted', () => finish(new Error('response aborted')))
+        response.on('end', () => finish(null, {
+          status: response.statusCode, headers: response.headers,
+          text: Buffer.concat(chunks).toString('utf8'),
+        }))
+      }) } catch (error) { finish(error); return }
+      request.on('error', error => finish(error))
+      if (options.body !== undefined) request.write(options.body)
+      request.end()
+    }
+    const targetHost = url.hostname.replace(/^\[|\]$/g, '')
+    const targetOptions = {
+      hostname: targetHost, port: url.port || (url.protocol === 'https:' ? 443 : 80),
+      path: url.pathname + url.search, method: options.method || 'GET',
+      headers: { ...options.headers }, ca: options.ca,
+    }
+    // Bing and some model gateways reject chunked POST bodies with HTTP 411.
+    if (options.body !== undefined && !Object.keys(targetOptions.headers).some(key =>
+      ['content-length', 'transfer-encoding'].includes(key.toLowerCase()))) {
+      targetOptions.headers['Content-Length'] = Buffer.byteLength(options.body)
+    }
+    if (!proxy) {
+      send(url.protocol === 'https:' ? https : http, targetOptions)
+    } else {
+      const proxyClient = proxy.protocol === 'https:' ? https : http
+      const proxyOptions = {
+        hostname: proxy.hostname.replace(/^\[|\]$/g, ''),
+        port: proxy.port || (proxy.protocol === 'https:' ? 443 : 80), ca: options.ca,
+      }
+      if (url.protocol === 'http:') {
+        send(proxyClient, { ...proxyOptions, method: targetOptions.method, path: url.href,
+          headers: { ...targetOptions.headers, Host: url.host, ...proxyAuthorization(proxy) } })
+      } else {
+        const authority = url.host + (url.port ? '' : ':443')
+        connectRequest = proxyClient.request({ ...proxyOptions, method: 'CONNECT', path: authority,
+          headers: { Host: authority, ...proxyAuthorization(proxy) } })
+        connectRequest.on('error', error => finish(error))
+        connectRequest.on('connect', (response, socket, head) => {
+          if (settled) { socket.destroy(); return }
+          tunnel = socket
+          tunnel.on('error', error => finish(error))
+          if (response.statusCode !== 200) { finish(new Error('proxy CONNECT ' + response.statusCode)); return }
+          if (head.length) socket.unshift(head)
+          agent = new https.Agent({ keepAlive: false })
+          agent.createConnection = () => tls.connect({ socket,
+            servername: isIP(targetHost) ? undefined : targetHost, ca: options.ca,
+            ALPNProtocols: ['http/1.1'],
+          })
+          send(https, { ...targetOptions, agent })
+        })
+        connectRequest.end()
+      }
+    }
+  })
+}
+
+export async function requestText(urlString, options = {}) {
+  let url = new URL(urlString)
+  let method = options.method || 'GET', body = options.body
+  let headers = { 'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/126.0.0.0 Safari/537.36',
+    Accept: 'application/json,text/plain,*/*', ...options.headers }
+  const deadline = Date.now() + (options.timeoutMs || 25000)
+  for (let redirects = 0; redirects <= 5; redirects++) {
+    if (!['http:', 'https:'].includes(url.protocol)) throw new Error('Unsupported request protocol: ' + url.protocol)
+    const remaining = deadline - Date.now()
+    if (remaining <= 0) throw new Error('request timeout')
+    // Re-evaluate loopback on every redirect so a local model never uses a proxy.
+    const response = await requestOnce(url, { ...options, method, body, headers,
+      timeoutMs: remaining, proxy: isLoopback(url.href) ? null : options.proxy })
+    if ([301, 302, 303, 307, 308].includes(response.status) && response.headers.location) {
+      if (redirects === 5) throw new Error('Too many redirects')
+      const next = new URL(response.headers.location, url)
+      if (next.origin !== url.origin) {
+        headers = Object.fromEntries(Object.entries(headers).filter(([key]) =>
+          !['authorization', 'cookie', 'x-api-key', 'api-key', 'host'].includes(key.toLowerCase())))
+      }
+      if ((response.status === 303 && method !== 'HEAD') || ([301, 302].includes(response.status) && method === 'POST')) {
+        method = 'GET'; body = undefined
+        headers = Object.fromEntries(Object.entries(headers).filter(([key]) =>
+          !['content-type', 'content-length', 'transfer-encoding'].includes(key.toLowerCase())))
+      }
+      url = next
+      continue
+    }
+    if (response.status < 200 || response.status >= 300) throw new Error('HTTP ' + response.status)
+    return response.text
+  }
+}

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "dsh-think-translate",
   "description": "Display-layer translation for the DSH web app: thinking chain, task cards, and optional answer text, in 8 target languages (local Ollama / any OpenAI-compatible or Anthropic endpoint / Google / Bing, fail-open)",
-  "version": "1.2.5",
+  "version": "1.2.8",
   "type": "module",
   "main": "lib/index.js",
   "exports": {
@@ -12,6 +12,8 @@
   "files": [
     "lib/index.js",
     "lib/client.js",
+    "lib/ollama.js",
+    "lib/transport.js",
     "cordis.patch.yml",
     "demo/demo.gif",
     "demo/demo2.gif"