refactor(local-llm): выбранность модели из настроек (не из рантайма); back-fill llmLocalModelId для старых настроек (порт 8011/18200); token_to_piece_bytes вместо устаревшего API; +юнит-тесты

src-tauri/src/bin/talkis-llm.rs

@@ -14,8 +14,9 @@ use llama_cpp_2::context::params::LlamaContextParams;
 use llama_cpp_2::llama_backend::LlamaBackend;
 use llama_cpp_2::llama_batch::LlamaBatch;
 use llama_cpp_2::model::params::LlamaModelParams;
-use llama_cpp_2::model::{AddBos, LlamaModel, Special};
+use llama_cpp_2::model::{AddBos, LlamaModel};
 use llama_cpp_2::sampling::LlamaSampler;
+use llama_cpp_2::TokenToStringError;
 
 const SERVER_NAME: &str = "talkis-llm";
 const ENGINE_NAME: &str = "llama.cpp";
@@ -350,7 +351,18 @@ fn generate(
         // Accumulate raw token bytes and emit only COMPLETE UTF-8 sequences. A
         // single token can carry a partial multibyte char (Cyrillic spans token
         // boundaries), so decoding tokens one-by-one yielded garbled glyphs.
-        if let Ok(bytes) = model.token_to_bytes(token, Special::Tokenize) {
+        let piece_bytes = match model.token_to_piece_bytes(token, 8, true, None) {
+            Err(TokenToStringError::InsufficientBufferSpace(size)) => {
+                let needed = size
+                    .checked_neg()
+                    .and_then(|value| usize::try_from(value).ok())
+                    .unwrap_or(32);
+                model.token_to_piece_bytes(token, needed, true, None)
+            }
+            result => result,
+        };
+
+        if let Ok(bytes) = piece_bytes {
             pending.extend_from_slice(&bytes);
             loop {
                 match std::str::from_utf8(&pending) {

src/components/LocalLlmModels.tsx

@@ -5,6 +5,11 @@ import { AlertCircle, Check, Download, HardDrive, Loader2, MemoryStick, Trash2,
 
 import { AppSettings } from "../lib/store";
 import { useI18n } from "../lib/i18n";
+import {
+  isLocalLlmEndpoint,
+  localLlmDeleteSettingsPatch,
+  selectedLocalLlmModelId,
+} from "../lib/localLlmSelection";
 import qwenAvatar from "../assets/adapters/qwen.png";
 
 interface LocalLlmModel {
@@ -17,12 +22,6 @@ interface LocalLlmModel {
   downloaded: boolean;
 }
 
-interface LocalLlmStatus {
-  running: boolean;
-  model_id: string | null;
-  base_url: string | null;
-}
-
 interface DownloadProgress {
   model_id: string;
   status: string;
@@ -56,7 +55,6 @@ export function LocalLlmModels({
 }) {
   const { t } = useI18n();
   const [models, setModels] = useState<LocalLlmModel[]>([]);
-  const [status, setStatus] = useState<LocalLlmStatus | null>(null);
   const [progress, setProgress] = useState<Record<string, DownloadProgress>>({});
   const [busy, setBusy] = useState<string | null>(null);
   const [deleting, setDeleting] = useState<string | null>(null);
@@ -65,12 +63,8 @@ export function LocalLlmModels({
 
   const refresh = async (): Promise<void> => {
     try {
-      const [list, runtime] = await Promise.all([
-        invoke<LocalLlmModel[]>("list_local_llm_models"),
-        invoke<LocalLlmStatus>("get_local_llm_status"),
-      ]);
+      const list = await invoke<LocalLlmModel[]>("list_local_llm_models");
       setModels(list);
-      setStatus(runtime);
     } catch {
       /* keep last known state */
     }
@@ -152,6 +146,10 @@ export function LocalLlmModels({
     setError(null);
     try {
       await invoke("delete_local_llm_model", { modelId: model.id });
+      const cleanupPatch = localLlmDeleteSettingsPatch(settings, model.id);
+      if (cleanupPatch) {
+        update(cleanupPatch);
+      }
       setProgress((prev) => {
         const next = { ...prev };
         delete next[model.id];
@@ -182,10 +180,8 @@ export function LocalLlmModels({
     }
   };
 
-  const activeModelId = status?.running ? status.model_id : null;
-  const usingLocalEndpoint = Boolean(
-    settings.llmEndpoint?.trim().includes("127.0.0.1"),
-  );
+  const usingLocalEndpoint = isLocalLlmEndpoint(settings.llmEndpoint);
+  const selectedModelId = selectedLocalLlmModelId(settings);
 
   return (
     <div style={{ display: "flex", flexDirection: "column", gap: 12 }}>
@@ -241,7 +237,7 @@ export function LocalLlmModels({
             prog?.status === "downloading" || prog?.status === "starting";
           const isBusy = busy === model.id;
           const isDeleting = deleting === model.id;
-          const isActive = activeModelId === model.id && usingLocalEndpoint;
+          const isSelected = selectedModelId === model.id && usingLocalEndpoint;
           // Collapsed by default like the recognition cards; force-open while a
           // download runs so its progress bar stays visible.
           const open = expandedId === model.id || isDownloading;
@@ -250,12 +246,12 @@ export function LocalLlmModels({
             ? t("localLlm.status.deleting")
             : isDownloading
               ? t("localLlm.status.downloading")
-              : isActive
+              : isSelected
                 ? t("localLlm.status.selected")
                 : model.downloaded
                   ? t("localLlm.status.ready")
                   : t("localLlm.status.notDownloaded");
-          const statusColor = isActive
+          const statusColor = isSelected
             ? "var(--success-bright)"
             : model.downloaded
               ? "var(--success-bright)"
@@ -464,7 +460,7 @@ export function LocalLlmModels({
                       )
                     ) : (
                       <>
-                        {!isActive && (
+                        {!isSelected && (
                           <button
                             onClick={() => void useForSummary(model)}
                             disabled={isBusy}

src/lib/localLlmSelection.test.ts

@@ -0,0 +1,59 @@
+import { describe, expect, it } from "bun:test";
+
+import {
+  isLocalLlmEndpoint,
+  localLlmDeleteSettingsPatch,
+  selectedLocalLlmModelId,
+} from "./localLlmSelection";
+
+describe("local LLM selection helpers", () => {
+  it("recognizes localhost and 127.0.0.1 endpoints", () => {
+    expect(isLocalLlmEndpoint("http://127.0.0.1:8011/v1")).toBe(true);
+    expect(isLocalLlmEndpoint("http://localhost:11434/v1")).toBe(true);
+    expect(isLocalLlmEndpoint("https://api.openai.com/v1")).toBe(false);
+  });
+
+  it("prefers the bundled runtime marker over the legacy model field", () => {
+    expect(
+      selectedLocalLlmModelId({
+        llmEndpoint: "http://127.0.0.1:8011/v1",
+        llmModel: "legacy-model",
+        llmLocalModelId: "qwen2.5-3b-instruct-q4",
+      }),
+    ).toBe("qwen2.5-3b-instruct-q4");
+  });
+
+  it("falls back to llmModel for pre-marker local selections", () => {
+    expect(
+      selectedLocalLlmModelId({
+        llmEndpoint: "http://127.0.0.1:8011/v1",
+        llmModel: "qwen2.5-3b-instruct-q4",
+        llmLocalModelId: "",
+      }),
+    ).toBe("qwen2.5-3b-instruct-q4");
+  });
+
+  it("clears persisted text-model fields only when deleting the selected model", () => {
+    expect(
+      localLlmDeleteSettingsPatch(
+        {
+          llmEndpoint: "http://127.0.0.1:8011/v1",
+          llmModel: "qwen2.5-3b-instruct-q4",
+          llmLocalModelId: "qwen2.5-3b-instruct-q4",
+        },
+        "qwen2.5-3b-instruct-q4",
+      ),
+    ).toEqual({ llmEndpoint: "", llmModel: "none", llmLocalModelId: "" });
+
+    expect(
+      localLlmDeleteSettingsPatch(
+        {
+          llmEndpoint: "http://127.0.0.1:8011/v1",
+          llmModel: "qwen2.5-7b-instruct-q4",
+          llmLocalModelId: "qwen2.5-7b-instruct-q4",
+        },
+        "qwen2.5-3b-instruct-q4",
+      ),
+    ).toBeNull();
+  });
+});

src/lib/localLlmSelection.ts

@@ -0,0 +1,39 @@
+import type { AppSettings } from "./store";
+
+export interface LocalLlmSelectionSettings {
+  llmEndpoint?: string;
+  llmModel?: string;
+  llmLocalModelId?: string;
+}
+
+export function isLocalLlmEndpoint(endpoint?: string): boolean {
+  return /127\.0\.0\.1|localhost/i.test(endpoint || "");
+}
+
+export function selectedLocalLlmModelId(
+  settings: LocalLlmSelectionSettings,
+): string {
+  if (!isLocalLlmEndpoint(settings.llmEndpoint)) {
+    return "";
+  }
+
+  return settings.llmLocalModelId?.trim() || settings.llmModel?.trim() || "";
+}
+
+export function isSelectedLocalLlmModel(
+  settings: LocalLlmSelectionSettings,
+  modelId: string,
+): boolean {
+  return selectedLocalLlmModelId(settings) === modelId;
+}
+
+export function localLlmDeleteSettingsPatch(
+  settings: LocalLlmSelectionSettings,
+  modelId: string,
+): Partial<AppSettings> | null {
+  if (!isSelectedLocalLlmModel(settings, modelId)) {
+    return null;
+  }
+
+  return { llmEndpoint: "", llmModel: "none", llmLocalModelId: "" };
+}

src/lib/store.ts

@@ -246,6 +246,11 @@ const MAIN_KEY_ALIASES: Record<string, string> = {
 const FUNCTION_KEY_PATTERN = /^F(?:[1-9]|1[0-2])$/;
 const DEFAULT_MAC_HOTKEY = "Command+Shift+Space";
 const DEFAULT_DESKTOP_HOTKEY = "Control+Alt+Space";
+const BUNDLED_LOCAL_LLM_PORTS = new Set([8011]);
+const BUNDLED_LOCAL_LLM_MODEL_IDS = new Set([
+  "qwen2.5-3b-instruct-q4",
+  "qwen2.5-7b-instruct-q4",
+]);
 
 export function isMacPlatform(): boolean {
   if (typeof navigator === "undefined") return false;
@@ -597,6 +602,20 @@ function parseProvider(value: unknown): ApiProvider | undefined {
   return undefined;
 }
 
+function isBundledLocalLlmEndpoint(value: unknown): boolean {
+  if (typeof value !== "string") return false;
+  const match = value.match(/:(\d{4,5})(?:\/|$)/);
+  if (!match) return false;
+  const port = Number(match[1]);
+  return BUNDLED_LOCAL_LLM_PORTS.has(port) || (port >= 18200 && port <= 18249);
+}
+
+function parseBundledLocalLlmModelId(value: unknown): string | undefined {
+  if (typeof value !== "string") return undefined;
+  const modelId = value.trim();
+  return BUNDLED_LOCAL_LLM_MODEL_IDS.has(modelId) ? modelId : undefined;
+}
+
 function parseRealtimeInterpreterSettings(
   value: unknown,
 ): RealtimeInterpreterSettings | undefined {
@@ -630,7 +649,7 @@ function parseRealtimeInterpreterSettings(
   };
 }
 
-function normalizeSavedSettings(saved: unknown): Partial<AppSettings> {
+export function normalizeSavedSettings(saved: unknown): Partial<AppSettings> {
   if (!saved || typeof saved !== "object") {
     return {};
   }
@@ -772,7 +791,11 @@ function normalizeSavedSettings(saved: unknown): Partial<AppSettings> {
     llmEndpoint:
       typeof raw.llmEndpoint === "string" ? raw.llmEndpoint : undefined,
     llmLocalModelId:
-      typeof raw.llmLocalModelId === "string" ? raw.llmLocalModelId : undefined,
+      typeof raw.llmLocalModelId === "string"
+        ? raw.llmLocalModelId
+        : isBundledLocalLlmEndpoint(raw.llmEndpoint)
+          ? parseBundledLocalLlmModelId(raw.llmModel)
+          : undefined,
     useOwnKey: typeof raw.useOwnKey === "boolean" ? raw.useOwnKey : undefined,
     deviceToken:
       typeof raw.deviceToken === "string" ? raw.deviceToken : undefined,

src/lib/summarize.test.ts

@@ -0,0 +1,84 @@
+import { describe, expect, it } from "bun:test";
+
+import { normalizeSavedSettings, type AppSettings } from "./store";
+import { isSummaryAvailable, resolveSummaryBackend } from "./summarize";
+
+function settings(overrides: Partial<AppSettings>): AppSettings {
+  return {
+    apiKey: "",
+    llmApiKey: "",
+    llmEndpoint: "",
+    llmModel: "gpt-4o-mini",
+    llmLocalModelId: "",
+    useOwnKey: true,
+    deviceToken: "",
+    ...overrides,
+  } as AppSettings;
+}
+
+describe("resolveSummaryBackend local LLM behavior", () => {
+  it("treats bundled runtime port 8011 without a selected model as unavailable", () => {
+    const backend = resolveSummaryBackend(
+      settings({ llmEndpoint: "http://127.0.0.1:8011/v1" }),
+    );
+    expect(backend).toBeNull();
+    expect(
+      isSummaryAvailable(settings({ llmEndpoint: "http://127.0.0.1:8011/v1" })),
+    ).toBe(false);
+  });
+
+  it("treats bundled fallback ports without a selected model as unavailable", () => {
+    expect(
+      resolveSummaryBackend(
+        settings({ llmEndpoint: "http://127.0.0.1:18210/v1" }),
+      ),
+    ).toBeNull();
+  });
+
+  it("resolves bundled local runtime when llmLocalModelId is set", () => {
+    const backend = resolveSummaryBackend(
+      settings({
+        llmEndpoint: "http://127.0.0.1:8011/v1",
+        llmModel: "qwen2.5-3b-instruct-q4",
+        llmLocalModelId: "qwen2.5-3b-instruct-q4",
+      }),
+    );
+    expect(backend?.kind).toBe("local");
+  });
+
+  it("does not block a user's own localhost server outside bundled ports", () => {
+    const backend = resolveSummaryBackend(
+      settings({
+        llmEndpoint: "http://localhost:11434/v1",
+        llmModel: "qwen2.5:7b",
+      }),
+    );
+    expect(backend?.kind).toBe("local");
+  });
+});
+
+describe("stored settings migration for bundled local LLM", () => {
+  it("migrates old 8011 bundled runtime settings into llmLocalModelId", () => {
+    const normalized = normalizeSavedSettings({
+      llmEndpoint: "http://127.0.0.1:8011/v1",
+      llmModel: "qwen2.5-3b-instruct-q4",
+    });
+    expect(normalized.llmLocalModelId).toBe("qwen2.5-3b-instruct-q4");
+  });
+
+  it("migrates fallback bundled runtime ports into llmLocalModelId", () => {
+    const normalized = normalizeSavedSettings({
+      llmEndpoint: "http://127.0.0.1:18208/v1",
+      llmModel: "qwen2.5-7b-instruct-q4",
+    });
+    expect(normalized.llmLocalModelId).toBe("qwen2.5-7b-instruct-q4");
+  });
+
+  it("does not mark a user-managed localhost endpoint as bundled runtime", () => {
+    const normalized = normalizeSavedSettings({
+      llmEndpoint: "http://localhost:11434/v1",
+      llmModel: "qwen2.5:7b",
+    });
+    expect(normalized.llmLocalModelId).toBeUndefined();
+  });
+});

src/lib/summarize.ts

@@ -168,15 +168,15 @@ function localLlmRunner(
   };
 }
 
-// The bundled local text runtime listens on 127.0.0.1:18200–18249 (see
-// src-tauri/src/llm_runtime.rs). It can only auto-start when a model is selected
-// (llmLocalModelId); an endpoint in that range with NO selected model is a dead
-// port, so we must not treat it as a usable backend.
+// The bundled local text runtime prefers 127.0.0.1:8011, then falls back to
+// 18200–18249 (see src-tauri/src/llm_runtime.rs). It can only auto-start when a
+// model is selected (llmLocalModelId); an endpoint in that range with NO selected
+// model is a dead port, so we must not treat it as a usable backend.
 function isBundledLocalRuntime(endpoint: string): boolean {
   const match = endpoint.match(/:(\d{4,5})(?:\/|$)/);
   if (!match) return false;
   const port = Number(match[1]);
-  return port >= 18200 && port <= 18249;
+  return port === 8011 || (port >= 18200 && port <= 18249);
 }
 
 /**