cortico 0.1.1 → 0.1.3

package/src/providers/llamacpp/console/models-panel.ts

@@ -1,10 +1,13 @@
 /**
- * Model section of one endpoint: what llama-server knows, and pulling more.
- * The endpoint comes from `ctx.scope.instance`.
+ * Model section of one endpoint: what llama-server knows, a HuggingFace search that feeds the
+ * pull field, and `use` to put a served model into the model section. The endpoint comes from
+ * `ctx.scope.instance`.
  */
 import type { ConsolePanel, ConsolePanelContext } from '../../../web/shared/client-panel.ts';
+import type { HfFile, HfRepo } from '../huggingface.ts';
 import type { ModelsState } from './server.ts';
 import { panel } from '../strings.ts';
+import { HITS_PER_PAGE } from '../../../web/client/features/extensions/index.ts';
 
 const POLL_MS = 2_000;
 
@@ -22,8 +25,12 @@ export const modelsPanel: ConsolePanel = {
     const card = ui.sheet({ title: S.modelsTitle });
     const message = ui.msgline();
     root.append(card.el, message);
-    /** The pull input survives a re-render: the operator types while the poll redraws the card. */
+    /** Typed text and search results survive a re-render: the poll redraws the card around them. */
     let draft = '';
+    let query = '';
+    let repos: HfRepo[] | null = null;
+    let searchNote = '';
+    let chosen: { repo: string; files: HfFile[] } | null = null;
     let lastSnapshot = '';
 
     async function act(method: string, extra: Record<string, unknown> = {}): Promise<void> {
@@ -36,6 +43,76 @@ export const modelsPanel: ConsolePanel = {
       await load(true);
     }
 
+    async function search(text: string): Promise<void> {
+      const wanted = text.trim();
+      if (!wanted) return;
+      query = wanted;
+      chosen = null;
+      try {
+        // One page of results, the console's page size for search hits.
+        const result = await ctx.invoke<{ repos: HfRepo[] }>('search', [{ name, query: wanted, limit: HITS_PER_PAGE }]);
+        repos = result.repos;
+        searchNote = repos.length ? '' : S.noRepos;
+      } catch (error) {
+        repos = [];
+        searchNote = S.searchFailed(String(error));
+      }
+      await load(true);
+    }
+
+    async function showFiles(repo: string): Promise<void> {
+      try {
+        chosen = { repo, files: (await ctx.invoke<{ files: HfFile[] }>('files', [{ name, repo }])).files };
+      } catch (error) {
+        message.textContent = String(error);
+        return;
+      }
+      await load(true);
+    }
+
+    function renderSearch(body: HTMLElement): void {
+      const bar = ui.rowbar();
+      const input = ui.input({
+        placeholder: S.searchPlaceholder,
+        value: query,
+        onCommit: (value) => void search(value),
+      });
+      input.setAttribute('aria-label', S.search);
+      bar.append(input, ui.button(S.search, { onClick: () => void search(input.value) }));
+      body.append(ui.field(S.search, bar));
+      if (repos === null) return;
+      if (searchNote) body.append(ui.msgline(searchNote, repos.length === 0 && searchNote !== S.noRepos));
+      if (repos.length) {
+        const table = ui.table({ head: [S.repo, S.downloads, S.likes, S.updated, ''] });
+        for (const repo of repos) {
+          const open = ui.button(S.showFiles, { size: 'sm', onClick: () => void showFiles(repo.id) });
+          table.addRow([
+            { text: repo.id, cls: 'mono' },
+            ui.fmt.count(repo.downloads),
+            ui.fmt.count(repo.likes),
+            repo.updatedAt ? repo.updatedAt.slice(0, 10) : '',
+            open,
+          ]);
+        }
+        body.append(table.el);
+      }
+      if (!chosen) return;
+      body.append(ui.section(S.filesOf(chosen.repo)));
+      const files = ui.table({ head: [S.file, S.size, S.tag, ''] });
+      if (chosen.files.length === 0) files.clear(S.noFiles);
+      for (const file of chosen.files) {
+        const pull = ui.button(S.pull, { size: 'sm', variant: 'primary', onClick: () => void act('pull', { model: file.pull }) });
+        pull.title = file.pull;
+        files.addRow([
+          { text: file.name, cls: 'mono' },
+          file.bytes === null ? '' : bytes(file.bytes),
+          { text: file.tag ?? '', cls: 'mono' },
+          pull,
+        ]);
+      }
+      body.append(files.el);
+    }
+
     function render(state: ModelsState, body: HTMLElement): void {
       body.append(ui.kv([
         { k: S.cacheDir, v: state.cacheDir },
@@ -79,13 +156,18 @@ export const modelsPanel: ConsolePanel = {
           });
           status.append(progress.el);
         }
-        const actions = ui.h('div');
+        const actions = ui.rowbar();
         if (model.status === 'downloading')
           actions.append(ui.button(S.cancel, { size: 'sm', onClick: () => void act('cancel', { model: model.id }) }));
         else if (model.status === 'loaded' || model.status === 'sleeping')
           actions.append(ui.button(S.unload, { size: 'sm', onClick: () => void act('unload', { model: model.id }) }));
         else if (model.status === 'unloaded' || model.status === 'failed')
           actions.append(ui.button(S.load, { size: 'sm', onClick: () => void act('load', { model: model.id }) }));
+        if (model.status !== 'downloading') {
+          const use = ui.button(S.use, { size: 'sm', onClick: () => void act('use', { model: model.id }) });
+          use.title = S.useTitle;
+          actions.append(use);
+        }
         table.addRow([
           { text: model.id, cls: 'mono' },
           status,
@@ -95,6 +177,7 @@ export const modelsPanel: ConsolePanel = {
         ]);
       }
       body.append(table.el);
+      renderSearch(body);
     }
 
     async function load(force = false): Promise<void> {

package/src/providers/llamacpp/console/runtime-panel.ts

@@ -3,11 +3,12 @@
  * The endpoint comes from `ctx.scope.instance`; edits are written on change.
  */
 import type { ConsolePanel, ConsolePanelContext } from '../../../web/shared/client-panel.ts';
-import type { LaunchOptions } from '../options.ts';
-import type { RuntimeState } from '../runtime.ts';
+import type { ConfigValues } from '../../../core/config-schema.ts';
+import { configField } from '../../../web/client/features/config/view.ts';
+import type { RuntimePanelState } from './server.ts';
 import { panel } from '../strings.ts';
 
-type Row = RuntimeState & { name: string };
+type Row = RuntimePanelState;
 
 const POLL_MS = 2_000;
 
@@ -42,18 +43,35 @@ export const runtimePanel: ConsolePanel = {
       await load(true);
     }
 
-    /** One launch field; empty input sends NaN so the endpoint reports the same error as a bad value. */
-    function launchNumber(row: Row, key: keyof LaunchOptions, label: string, min: number): HTMLElement {
-      const current = row.launch?.[key];
-      const input = ui.input({
-        type: 'number',
-        value: current === undefined ? '' : String(current),
-        onChange: (value) => void act('configure', { launch: { [key]: Number(value) } }),
-      });
-      input.min = String(min);
-      input.step = '1';
-      input.setAttribute('aria-label', label);
-      return ui.field(label, input);
+    async function saveConfig(groupId: string, values: ConfigValues): Promise<void> {
+      if (busy) return;
+      busy = true;
+      message.textContent = '';
+      try {
+        await ctx.setConfig(groupId, values);
+      } catch (error) {
+        message.textContent = String(error);
+      } finally {
+        busy = false;
+      }
+      await load(true);
+    }
+
+    function renderConfig(row: Row, index: number, body: HTMLElement): void {
+      const { group, values } = row.config[index];
+      body.append(ui.section(group.schema.title, group.schema.description));
+      for (const [path, property] of Object.entries(group.schema.properties)) {
+        const field = configField(ui, property, values[path], () => {
+          if (field.read) void saveConfig(group.id, { [path]: field.read() });
+        }, ctx.signal);
+        field.node.setAttribute('aria-label', property.title);
+        // 自备运行时目录时后端由目录里的二进制决定,后端那一格禁用。
+        if (path.endsWith('.runtime.backend') && row.own) {
+          (field.node as HTMLSelectElement).disabled = true;
+        }
+        body.append(ui.field(property.title, field.node));
+        if (property.description) body.append(ui.msgline(property.description));
+      }
     }
 
     function renderUnmanaged(row: Row, body: HTMLElement): void {
@@ -74,42 +92,8 @@ export const runtimePanel: ConsolePanel = {
         : install.phase === 'extracting' ? ui.pill(S.extracting, 'plain')
         : ui.pill(S.absent, 'off');
 
-      body.append(ui.section(S.runtimeSection, S.runtimeSectionDesc));
-      const release = ui.input({
-        value: row.release ?? '',
-        cls: 'mono',
-        onChange: (value) => void act('configure', { release: value }),
-      });
-      release.setAttribute('aria-label', S.release);
-      const backend = ui.select({
-        value: row.backend ?? '',
-        options: row.backendChoices,
-        onChange: (value) => void act('configure', { backend: value }),
-      });
-      backend.setAttribute('aria-label', S.backend);
-      backend.disabled = row.own;
-      const dir = ui.input({
-        value: row.own ? row.runtimeDir ?? '' : '',
-        cls: 'mono',
-        onChange: (value) => void act('configure', { runtimeDir: value }),
-      });
-      dir.setAttribute('aria-label', S.runtimeDir);
-      const dirRow = ui.rowbar();
-      dirRow.append(dir, ui.button(S.browse, {
-        size: 'sm',
-        onClick: async () => {
-          const picked = await ctx.pickPath({ kind: 'directory', title: S.runtimeDir, currentPath: dir.value });
-          if (picked) await act('configure', { runtimeDir: picked });
-        },
-      }));
-      body.append(
-        ui.field(S.release, release),
-        ui.msgline(S.releaseHint),
-        ui.field(S.backend, backend),
-        ui.field(S.runtimeDir, dirRow),
-        ui.msgline(S.runtimeDirHint),
-        ui.kv([{ k: S.installStatus, v: installPill }]),
-      );
+      renderConfig(row, 0, body);
+      body.append(ui.kv([{ k: S.installStatus, v: installPill }]));
       if (row.smartAppControl === 1) body.append(ui.msgline(S.sacWarning, true));
       if (install.phase === 'downloading' || install.phase === 'extracting') {
         const progress = ui.progress({
@@ -133,28 +117,7 @@ export const runtimePanel: ConsolePanel = {
       installBar.append(ui.h('span', 'grow'), ui.button(S.disable, { onClick: () => void act('disable') }));
       body.append(installBar);
 
-      body.append(ui.section(S.launchSection, S.launchSectionDesc));
-      const launchRow = ui.rowbar();
-      launchRow.append(
-        launchNumber(row, 'contextSize', S.contextSize, 1),
-        launchNumber(row, 'nGpuLayers', S.nGpuLayers, 0),
-        launchNumber(row, 'parallel', S.parallel, 1),
-      );
-      const extraArgs = ui.input({
-        value: row.launch?.extraArgs ?? '',
-        cls: 'mono',
-        onChange: (value) => void act('configure', { launch: { extraArgs: value } }),
-      });
-      extraArgs.setAttribute('aria-label', S.extraArgs);
-      body.append(
-        launchRow,
-        ui.field(S.extraArgs, extraArgs),
-        ui.msgline(S.extraArgsHint),
-        ui.checkbox(S.autoStart, {
-          checked: row.autoStart,
-          onChange: (checked) => void act('configure', { autoStart: checked }),
-        }).el,
-      );
+      renderConfig(row, 1, body);
 
       body.append(ui.section(S.serverSection, S.serverSectionDesc));
       const server = row.server;

package/src/providers/llamacpp/console/server.ts

@@ -5,9 +5,14 @@
  */
 import type { ConsolePageContribution } from '../../../web/shared/console-protocol.ts';
 import type { ProviderConsoleHost } from '../../console/types.ts';
+import { connectionBlocks } from '../../console/config.ts';
 import type { RouterCatalog, RouterModel } from '../catalog.ts';
 import { LAUNCH_DEFAULTS, PINNED_RELEASE, backendChoices, defaultBackend, llamacppOptions, type LaunchOptions } from '../options.ts';
 import type { LlamaRuntime } from '../runtime.ts';
+import type { RuntimeState } from '../runtime.ts';
+import { runtimeConfig } from '../config.ts';
+import { listGguf, searchGguf } from '../huggingface.ts';
+import { getByPath, type ConfigGroup, type ConfigValues } from '../../../core/config-schema.ts';
 import { text } from '../strings.ts';
 
 interface Control {
@@ -23,6 +28,11 @@ export interface ModelsState {
   models: RouterModel[];
 }
 
+export type RuntimePanelState = RuntimeState & {
+  name: string;
+  config: Array<{ group: ConfigGroup; values: ConfigValues }>;
+};
+
 /** The launch fields the panel edits, each optional and applied over the stored values. */
 type LaunchPatch = Partial<Record<keyof LaunchOptions, unknown>>;
 
@@ -41,16 +51,32 @@ export function llamacppConsole(host: ProviderConsoleHost): Partial<ConsolePageC
     if (!found) throw new Error(S.instanceNameRequired);
     return found.entry;
   };
+  const blocks = connectionBlocks(host.language);
   return {
+    config: [],
     panels: [
-      { id: 'runtime', title: S.runtimePanel, description: S.runtimePanelDescription, slot: 'instance' },
-      { id: 'models', title: S.modelsPanel, description: S.modelsPanelDescription, slot: 'instance' },
+      { ...blocks.endpoint, title: S.endpointPanel, description: S.endpointPanelDescription },
+      { id: 'runtime', title: S.runtimePanel, description: S.runtimePanelDescription },
+      { id: 'models', title: S.modelsPanel, description: S.modelsPanelDescription },
+      blocks.model,
+      blocks.pricing,
+      blocks.protocol,
     ],
     invoke: async (panel, method, args) => {
       if (panel === 'runtime') {
         const value = body(args);
         const name = value.name as string;
-        if (method === 'state') return { name, ...(await control(name).runtime.state(host.language)) };
+        if (method === 'state') {
+          const entry = entryOf(name);
+          const prefix = `providers.${name}.`;
+          const config = runtimeConfig(name, entry, host.language).map((group) => ({
+            group,
+            values: Object.fromEntries(Object.keys(group.schema.properties).map((path) => [
+              path, getByPath(entry as unknown as Record<string, unknown>, path.slice(prefix.length)) ?? '',
+            ])) as ConfigValues,
+          }));
+          return { name, ...(await control(name).runtime.state(host.language)), config } satisfies RuntimePanelState;
+        }
         if (method === 'enable') {
           const entry = entryOf(name);
           const options = llamacppOptions(entry);
@@ -66,7 +92,7 @@ export function llamacppConsole(host: ProviderConsoleHost): Partial<ConsolePageC
           return { ok: true };
         }
         if (method === 'disable') {
-          await control(name).runtime.stop(host.language);
+          if (!host.editing) await control(name).runtime.stop(host.language);
           const entry = entryOf(name);
           const { runtime: _runtime, launch: _launch, ...rest } = entry.options ?? {};
           host.save(name, { ...entry, options: rest });
@@ -96,6 +122,7 @@ export function llamacppConsole(host: ProviderConsoleHost): Partial<ConsolePageC
           });
           return { ok: true };
         }
+        if (host.editing) throw new Error(host.language === 'zh' ? '请先保存配置，再执行运行时操作。' : 'Save configuration before runtime operations.');
         if (method === 'install') {
           await control(name).runtime.install(host.language);
           return { ok: true };
@@ -120,6 +147,22 @@ export function llamacppConsole(host: ProviderConsoleHost): Partial<ConsolePageC
           await catalog.list(true);
           return { ok: true };
         }
+        if (method === 'search') {
+          if (typeof value.query !== 'string' || !value.query.trim()) throw new Error(S.queryRequired);
+          if (!Number.isInteger(value.limit) || (value.limit as number) < 1) throw new Error(S.limitRequired);
+          return { repos: await searchGguf(value.query.trim(), value.limit as number) };
+        }
+        if (method === 'files') {
+          if (typeof value.repo !== 'string' || !value.repo.trim()) throw new Error(S.repoRequired);
+          return { files: await listGguf(value.repo.trim()) };
+        }
+        if (method === 'use') {
+          if (typeof value.model !== 'string' || !value.model.trim()) throw new Error(S.modelIdRequired);
+          const entry = entryOf(name);
+          host.save(name, { ...entry, spec: { thinking: false, ...entry.spec, model: value.model.trim() } });
+          return { ok: true };
+        }
+        if (host.editing) throw new Error(host.language === 'zh' ? '请先保存配置，再操作模型。' : 'Save configuration before model operations.');
         if (typeof value.model !== 'string' || !value.model.trim()) throw new Error(S.modelIdRequired);
         const model = value.model.trim();
         if (method === 'pull') await catalog.download(model);

package/src/providers/llamacpp/huggingface.ts

@@ -0,0 +1,93 @@
+/**
+ * HuggingFace lookups behind the pull field: repositories tagged `gguf`, and the GGUF files of
+ * one repository as `repo:tag` pull targets. The host is the one llama-server itself pulls from:
+ * `MODEL_ENDPOINT`, else `HF_ENDPOINT`, else huggingface.co.
+ */
+
+export interface HfRepo {
+  id: string;
+  downloads: number;
+  likes: number;
+  /** ISO timestamp of the last commit; null when the listing omits it. */
+  updatedAt: string | null;
+}
+
+export interface HfFile {
+  name: string;
+  /** Sum over shards; null when the listing carries no sizes. */
+  bytes: number | null;
+  /** Quantization tag read from the file name; null when the name carries none. */
+  tag: string | null;
+  /** What `POST /models` takes to pull this file: `repo:tag`, or the repository alone. */
+  pull: string;
+}
+
+export interface HfOptions {
+  fetchImpl?: typeof fetch;
+  env?: NodeJS.ProcessEnv;
+}
+
+const DEFAULT_ENDPOINT = 'https://huggingface.co';
+const SHARD_RE = /-(\d{5})-of-\d{5}\.gguf$/i;
+const QUANT_RE = /^(?:i?q\d[a-z0-9_]*|tq\d[a-z0-9_]*|bf16|f16|f32|f64|mxfp4[a-z0-9_]*)$/i;
+
+export function hfEndpoint(env: NodeJS.ProcessEnv = process.env): string {
+  return (env.MODEL_ENDPOINT || env.HF_ENDPOINT || DEFAULT_ENDPOINT).replace(/\/+$/, '');
+}
+
+async function getJson(url: string, fetchImpl: typeof fetch): Promise<unknown> {
+  const res = await fetchImpl(url, { headers: { Accept: 'application/json' } });
+  if (!res.ok) throw new Error(`GET ${url} ${res.status}`);
+  return res.json();
+}
+
+/** The `limit` most-downloaded matches; without one the listing returns up to a thousand rows. */
+export async function searchGguf(query: string, limit: number, options: HfOptions = {}): Promise<HfRepo[]> {
+  const params = new URLSearchParams({ search: query, filter: 'gguf', sort: 'downloads', direction: '-1', limit: String(limit) });
+  // The plain listing omits lastModified; `expand` names every field wanted back.
+  for (const field of ['downloads', 'likes', 'lastModified']) params.append('expand[]', field);
+  const rows = await getJson(`${hfEndpoint(options.env)}/api/models?${params}`, options.fetchImpl ?? fetch);
+  if (!Array.isArray(rows)) throw new Error('HuggingFace search returned no list');
+  const repos: HfRepo[] = [];
+  for (const row of rows as Array<Record<string, unknown>>) {
+    const id = typeof row.id === 'string' ? row.id : typeof row.modelId === 'string' ? row.modelId : null;
+    if (!id) continue;
+    repos.push({
+      id,
+      downloads: typeof row.downloads === 'number' ? row.downloads : 0,
+      likes: typeof row.likes === 'number' ? row.likes : 0,
+      updatedAt: typeof row.lastModified === 'string' ? row.lastModified : null,
+    });
+  }
+  return repos;
+}
+
+/** The quantization tag in a GGUF file name: the last `-` or `.` separated token that names one, with an `UD-` prefix kept. */
+export function quantTag(filename: string): string | null {
+  const stem = filename.replace(SHARD_RE, '.gguf').replace(/\.gguf$/i, '');
+  const tokens = stem.split(/[-.]/);
+  for (let i = tokens.length - 1; i >= 0; i--) {
+    if (!QUANT_RE.test(tokens[i])) continue;
+    return (tokens[i - 1]?.toUpperCase() === 'UD' ? 'UD-' : '') + tokens[i];
+  }
+  return null;
+}
+
+/** GGUF files of a repository, shards folded into one entry, projector files (`mmproj-*`) left out. */
+export async function listGguf(repo: string, options: HfOptions = {}): Promise<HfFile[]> {
+  const info = await getJson(`${hfEndpoint(options.env)}/api/models/${repo.split('/').map(encodeURIComponent).join('/')}?blobs=true`, options.fetchImpl ?? fetch) as { siblings?: unknown };
+  if (!Array.isArray(info.siblings)) throw new Error('HuggingFace repository listing carries no files');
+  const files = new Map<string, HfFile>();
+  for (const sibling of info.siblings as Array<Record<string, unknown>>) {
+    const path = typeof sibling.rfilename === 'string' ? sibling.rfilename : '';
+    if (!/\.gguf$/i.test(path) || /(^|\/)mmproj/i.test(path)) continue;
+    const name = path.replace(SHARD_RE, '.gguf');
+    const lfs = sibling.lfs as { size?: unknown } | undefined;
+    const size = typeof sibling.size === 'number' ? sibling.size : typeof lfs?.size === 'number' ? lfs.size : null;
+    const entry = files.get(name);
+    if (entry) { if (size !== null && entry.bytes !== null) entry.bytes += size; continue; }
+    const tag = quantTag(name);
+    files.set(name, { name, bytes: size, tag, pull: tag ? `${repo}:${tag}` : repo });
+  }
+  return [...files.values()];
+}

package/src/providers/llamacpp/index.ts

@@ -4,6 +4,7 @@ import type { Language } from '../../core/language.ts';
 import { isContextOverflow } from '../transport/errors.ts';
 import { modelsRoot, runtimesRoot } from '../../paths.ts';
 import { RouterCatalog } from './catalog.ts';
+import { runtimeConfig } from './config.ts';
 import { llamacppConsole } from './console/server.ts';
 import { LlamaCppProvider } from './native.ts';
 import { backendChoices, llamacppOptions, normalizeLlamaCpp } from './options.ts';
@@ -23,12 +24,14 @@ const reasoningTiers = (language: Language): ReasoningTier[] => {
 export default {
   id: 'llamacpp',
   title: 'llama.cpp',
+  description: 'Connect to or manage local llama.cpp models.',
   defaultBaseUrl: 'http://127.0.0.1:8090/v1',
   baseUrlSuggestions: ['http://127.0.0.1:8090/v1', 'http://127.0.0.1:8080/v1'],
   normalize: normalizeLlamaCpp,
   console: llamacppConsole,
+  config: runtimeConfig,
   reasoningTiers: reasoningTiers('zh'),
-  localize: (language) => ({ reasoningTiers: reasoningTiers(language) }),
+  localize: (language) => ({ description: language === 'zh' ? '连接或托管本地 llama.cpp 模型。' : 'Connect to or manage local llama.cpp models.', reasoningTiers: reasoningTiers(language) }),
   serviceTiers: [],
   validateEntry: (entry, language) => {
     const S = text(language);
@@ -87,6 +90,7 @@ export default {
         apiKey,
         log: host.log,
         media: { enabled: () => entry.multimodal === true, read: host.readBlob },
+        keepThinking: host.keepThinking,
       }),
     };
   },

package/src/providers/llamacpp/native.ts

@@ -2,23 +2,26 @@
  * Chat Completions against llama-server. Thinking is the template's `enable_thinking` switch
  * sent per request through `chat_template_kwargs`; templates without that variable ignore it.
  * The server returns the chain of thought as `reasoning_content` (`--reasoning-format deepseek`),
- * which the shared Chat assembly already maps; past thinking is not replayed.
+ * which the shared Chat assembly already maps; history sends it back on every request.
+ * `keepPastThinking` off blanks every turn but the synthetic opening; thinking off sends none.
  */
 import type { NativeChatMessage } from '../transport/native-types.ts';
 import type { ModelSpec, ToolSchema, Logger } from '../../core/types.ts';
 import { nullLogger } from '../../core/util.ts';
 import { OpenAIHttpClient } from '../transport/chat.ts';
-import { mapTools, renderMessagesWithMedia, type CompatMediaOptions } from '../transport/history.ts';
+import { dropPastThinking, mapTools, renderMessagesWithMedia, type CompatMediaOptions } from '../transport/history.ts';
 
 export function buildLlamaCppRequestBody(
   spec: ModelSpec,
   messages: NativeChatMessage[],
   tools?: ToolSchema[],
   media?: CompatMediaOptions,
+  opts?: { keepThinking?: boolean },
 ): Record<string, unknown> {
+  const history = opts?.keepThinking === false ? dropPastThinking(messages) : messages;
   const body: Record<string, unknown> = {
     model: spec.model,
-    messages: renderMessagesWithMedia(messages, media),
+    messages: renderMessagesWithMedia(history, media, { keepReasoning: spec.thinking }),
     chat_template_kwargs: { enable_thinking: spec.thinking },
   };
   if (spec.thinking && spec.reasoningEffort) body.reasoning_effort = spec.reasoningEffort;
@@ -32,15 +35,17 @@ export function buildLlamaCppRequestBody(
 export class LlamaCppProvider extends OpenAIHttpClient {
   private readonly apiKey?: string;
   private readonly media?: CompatMediaOptions;
+  private readonly keepThinking?: () => boolean;
 
-  constructor(opts: { baseUrl: string; apiKey?: string; log?: Logger; media?: CompatMediaOptions }) {
+  constructor(opts: { baseUrl: string; apiKey?: string; log?: Logger; media?: CompatMediaOptions; keepThinking?: () => boolean }) {
     super(opts.baseUrl, opts.log ?? nullLogger());
     this.apiKey = opts.apiKey;
     this.media = opts.media;
+    this.keepThinking = opts.keepThinking;
   }
 
   protected buildBody(spec: ModelSpec, messages: NativeChatMessage[], tools?: ToolSchema[]): Record<string, unknown> {
-    return buildLlamaCppRequestBody(spec, messages, tools, this.media);
+    return buildLlamaCppRequestBody(spec, messages, tools, this.media, { keepThinking: this.keepThinking?.() });
   }
 
   protected headers(): Record<string, string> {

package/src/providers/llamacpp/options.ts

@@ -42,6 +42,8 @@ export function normalizeLlamaCpp(entry: LLMProviderEntry): LLMProviderEntry {
   if (options.runtime && typeof options.runtime === 'object') {
     const runtime = options.runtime as Partial<RuntimeOptions>;
     options.runtime = { ...runtime, release: runtime.release ?? PINNED_RELEASE, backend: runtime.backend ?? defaultBackend() };
+    if (typeof options.runtime.release === 'string') options.runtime.release = options.runtime.release.trim();
+    if (typeof options.runtime.runtimeDir === 'string') options.runtime.runtimeDir = options.runtime.runtimeDir.trim();
     if (options.runtime.runtimeDir === '') delete options.runtime.runtimeDir;
     options.launch = { ...LAUNCH_DEFAULTS, ...(options.launch ?? {}) };
   }

package/src/providers/llamacpp/strings.ts

@@ -14,6 +14,8 @@ const zh = {
   notManaged: '这条端点没有开启托管',
   runtimeNotInstalled: '运行时还没装好',
   runtimeUnsupported: '本机平台没有这个后端的官方构建,填自备运行时目录',
+  endpointPanel: '服务地址',
+  endpointPanelDescription: 'llama-server 监听的地址;托管的进程按它启动。',
   runtimePanel: '运行时',
   runtimePanelDescription: 'llama.cpp 官方二进制的下载、安装与 llama-server 进程的启停。',
   modelsPanel: '模型',
@@ -22,6 +24,9 @@ const zh = {
   unknownMethod: '未知操作',
   instanceNameRequired: '需要端点名',
   modelIdRequired: '需要模型 id',
+  queryRequired: '需要搜索词',
+  limitRequired: '需要结果条数',
+  repoRequired: '需要仓库名',
   bodyRequired: '需要请求体',
   serverUnreachable: "llama-server 不可达；请检查端点地址与服务状态。",
   externalServer: '端点已有服务在跑(外部启动),不接管',
@@ -51,6 +56,8 @@ const en: typeof zh = {
   notManaged: 'This endpoint is not managed',
   runtimeNotInstalled: 'The runtime is not installed yet',
   runtimeUnsupported: 'No official build for this backend on this platform; set an own runtime directory',
+  endpointPanel: 'Server address',
+  endpointPanelDescription: 'The address llama-server listens on; a hosted process starts on it.',
   runtimePanel: 'Runtime',
   runtimePanelDescription: 'Download and install the official llama.cpp binaries; start and stop the llama-server process.',
   modelsPanel: 'Models',
@@ -59,6 +66,9 @@ const en: typeof zh = {
   unknownMethod: 'Unknown action',
   instanceNameRequired: 'An endpoint name is required',
   modelIdRequired: 'A model id is required',
+  queryRequired: 'A search query is required',
+  limitRequired: 'A result count is required',
+  repoRequired: 'A repository is required',
   bodyRequired: 'A request body is required',
   serverUnreachable: "llama-server is unreachable; check the endpoint URL and server status.",
   externalServer: 'A server is already running at the endpoint (started externally); not taking over',
@@ -139,6 +149,22 @@ const panelZh = {
   reload: '重扫',
   pull: '拉取',
   pullPlaceholder: 'HuggingFace 仓库,如 ggml-org/Qwen2.5-Coder-7B-Instruct-GGUF',
+  use: '选用',
+  useTitle: '把这个模型填进「模型与生成」',
+  search: '搜索',
+  searchPlaceholder: '搜 HuggingFace 上的 GGUF 仓库,如 qwen3',
+  searchFailed: (reason: string) => `搜索失败: ${reason}`,
+  noRepos: '没有匹配的仓库。',
+  repo: '仓库',
+  downloads: '下载',
+  likes: '点赞',
+  updated: '更新',
+  showFiles: '看文件',
+  filesOf: (repo: string) => `${repo} 里的 GGUF 文件`,
+  noFiles: '这个仓库里没有 GGUF 文件。',
+  file: '文件',
+  size: '大小',
+  tag: '标签',
   localDir: '本机 GGUF 目录',
   cacheDir: '缓存目录',
   noModels: '暂无可用的 GGUF 模型：可通过上方输入拉取，或将 .gguf 文件放入本机目录后点击重扫。',
@@ -213,6 +239,22 @@ const panelEn: typeof panelZh = {
   reload: 'Rescan',
   pull: 'Pull',
   pullPlaceholder: 'HuggingFace repo, e.g. ggml-org/Qwen2.5-Coder-7B-Instruct-GGUF',
+  use: 'Use',
+  useTitle: 'Put this model into the model section',
+  search: 'Search',
+  searchPlaceholder: 'Search GGUF repositories on HuggingFace, e.g. qwen3',
+  searchFailed: (reason: string) => `Search failed: ${reason}`,
+  noRepos: 'No matching repositories.',
+  repo: 'Repository',
+  downloads: 'Downloads',
+  likes: 'Likes',
+  updated: 'Updated',
+  showFiles: 'Files',
+  filesOf: (repo: string) => `GGUF files in ${repo}`,
+  noFiles: 'This repository has no GGUF files.',
+  file: 'File',
+  size: 'Size',
+  tag: 'Tag',
   localDir: 'Local GGUF directory',
   cacheDir: 'Cache directory',
   noModels: 'No models yet: pull one above, or drop GGUF files into the local directory and rescan.',

package/src/providers/name.ts

@@ -0,0 +1,13 @@
+/** New connection names are portable directory names; unchanged historical keys are exempt. */
+export function validateProviderName(name: string): string | null {
+  if (!/^[A-Za-z0-9][A-Za-z0-9_-]*$/.test(name))
+    return 'Use English letters, digits, - or _; start with a letter or digit. Spaces are not allowed.';
+  if (/^(con|prn|aux|nul|com[0-9]|lpt[0-9])$/i.test(name))
+    return 'This name is reserved by Windows.';
+  return null;
+}
+
+/** 密钥变量名缺省时的默认名:按端点名派生,`-` 折成 `_`,与 validateEntry 的环境变量名格式一致。 */
+export function defaultSecretName(name: string): string {
+  return `CORTICO_KEY_${name.toUpperCase().replace(/-/g, '_')}`;
+}