@objectivex666/dsh-settings-search 1.10.0 → 1.11.0

package/README.en.md

@@ -22,7 +22,7 @@
 - 📂 **In-page option search (v1.2.0)** — search beyond pages and reach the specific options inside them: Plugins-page tabs (`settings.plugins.tab`), Web UI plugin cards (`web-ui.plugin.item`), general rows, and more, each shown with a "page › option" breadcrumb.
 - 🔤 **Pinyin association search (v1.5.0)** — type initials (`sz`) or full pinyin (`shezhi`) to surface Chinese settings items; navigate with `↑`/`↓`, confirm with `Enter`, dismiss with `Esc`, and matched text is highlighted.
 - 🧠 **Intent search (v1.6.0)** — describe the goal in Chinese (e.g. 「太亮了」 "too bright", 「字太小」 "text too small") to surface matching settings without knowing their names, shown in a "You may want" group.
-- 🤖 **AI-assisted search (v1.7.0)** — when nothing matches locally, your own model (OpenAI-compatible or Anthropic: Base URL / model / API key) understands your search intent and suggests settings. Off by default; configure it on the plugin's settings page.
+- 🤖 **AI-assisted search** — when nothing matches locally, your OpenAI-compatible, Anthropic, or local Ollama model understands your search intent and suggests settings. Off by default; configure it on the plugin's settings page. Ollama needs no API key.
 - 🧾 **Logging & export (v1.7.0)** — records key plugin actions, searches, and AI requests; export, copy, or clear them from the plugin's settings page for troubleshooting. API keys are never logged.
 - 🧭 **Click-to-jump** — selecting a result clicks the matching left-nav entry to open its section; for tab options it also opens the tab and flashes the target row. A manual-path hint appears if automatic navigation is not possible.
 - 🌱 **Auto-index on startup (v1.8.0)** — the first time the settings panel opens, the plugin automatically visits every settings section, silently reads in-page options (e.g. "Enable Workshop card") into the search index, then restores your current section, so all section options are searchable without having to open each page manually. A "Importing section options…" hint shows under the search box and clears when done.
@@ -84,6 +84,22 @@ The package ships its own `cordis.patch.yml` (that exact content), which `dsh pl
 
 ---
 
+## Local Ollama (v1.11.0)
+
+Enable "AI-assisted search" on the Settings Search page and select "Ollama (local)" as the API type:
+
+- **Base URL**: defaults to `http://127.0.0.1:11434`; custom ports are supported. Accepts the root URL, `/api`, `/api/chat`, or `/v1`.
+- **Model**: enter the full name, including its tag, of an installed model. Use `ollama list` to see installed models.
+- **API key**: not required. Ollama requests never send an existing cloud API key.
+
+Save the configuration and use "Test API" to check connectivity. Ollama requests allow 120 seconds for cold starts; cancelling a search aborts the request.
+
+The plugin does not install or start Ollama, or download models. Prepare a running Ollama service and model yourself. Requests originate in the browser, so "local" means the browser's device, not a remote DSH server.
+
+For connection failures, check the service URL. For cross-origin failures, add the exact DSH page origin (for example `http://127.0.0.1:3080`) to Ollama's `OLLAMA_ORIGINS` and restart Ollama. Avoid unnecessary wildcard origins. Non-local URLs still require HTTPS.
+
+---
+
 ## Package layout
 
 ```
@@ -91,12 +107,13 @@ lib/index.js        Host half (no-op — makes the package a valid Cordis plugin
 lib/client.js       Browser half (search UI, __ModuleLoader__ format)
 cordis.patch.yml    Composition patch: inserts the settings-search plugin row
 scripts/check.mjs   Structural self-checks (npm test)
+scripts/ai.test.mjs Mock AI regression tests (no model execution)
 ```
 
 ## Development
 
 ```bash
-npm test   # validates the host half, the dsh.client manifest, and the client bundle handoff
+npm test   # structural checks and mocked Ollama / OpenAI / Anthropic regression tests
 ```
 
 > Historical note: up to v1.1 the repo shipped `dsh-settings-search.js` (an

package/README.md

@@ -22,7 +22,7 @@
 - 📂 **页内选项搜索（v1.2.0）**：不只搜页面，还能直达每个设置页内部的具体选项——插件页的标签页（`settings.plugins.tab`）、Web UI 插件的卡片（`web-ui.plugin.item`）、通用设置行等，结果以「页面 › 选项」面包屑展示。
 - 🔤 **拼音联想搜索（v1.5.0）**：输入拼音首字母（如 `sz`）或全拼（如 `shezhi`）即可联想中文设置项；支持 `↑`/`↓` 选择、`Enter` 跳转、`Esc` 关闭，命中的文字自动高亮。
 - 🧠 **意图联想搜索（v1.6.0）**：直接描述想做的事（如「太亮了」「字太小」「忘记密码」），无需知道选项名即可联想真实设置项，结果在「猜你想找」分组展示。
-- 🤖 **AI 联想搜索（v1.7.0）**：没有本地匹配时，用你配置的模型（OpenAI 兼容或 Anthropic：Base URL / 模型名 / API Key）理解搜索意图并推荐设置项。默认关闭，可在「设置搜索」页开启并配置。
+- 🤖 **AI 联想搜索**：没有本地匹配时，用你配置的 OpenAI 兼容、Anthropic 或本机 Ollama 模型理解搜索意图并推荐设置项。默认关闭，可在「设置搜索」页开启并配置；Ollama 无需 API Key。
 - 🧾 **日志与导出（v1.7.0）**：记录插件运行、搜索与 AI 请求的关键动作，可在「设置搜索」页一键导出、复制或清空，便于排查问题。不会记录 API Key。
 - 🧭 **点击即达**：选中结果后自动点击对应的左侧导航项打开所在分区；若是页内标签页会继续点开对应标签，并高亮闪烁目标选项行。无法自动跳转时显示手动路径提示。
 - 🌱 **启动自动索引（v1.8.0）**：首次打开设置面板时自动遍历各设置分区，静默读取分区内选项（如「启用创意工坊卡片」）并导入搜索索引；此后分区选项无需逐页手动打开即可直接搜索，读取完成后自动恢复当前所在分区，导入过程搜索框下方会显示提示，完成后自动消失。
@@ -84,6 +84,21 @@ dsh plugin --profile web add /path/to/dsh-settings-search
 
 ---
 
+## 本机 Ollama（v1.11.0）
+
+在「设置搜索」页开启「AI 联想搜索」，将「接口类型」选为「Ollama（本机）」：
+
+- **Base URL**：默认 `http://127.0.0.1:11434`，也可填写自定义端口；支持根地址、`/api`、`/api/chat` 和 `/v1`。
+- **模型名**：填写本机已安装的完整模型名（包括标签），可用 `ollama list` 查看。
+- **API Key**：无需填写，Ollama 请求不会携带原有云端密钥。
+
+保存后可用「测试接口」检查连接。本地模型首次加载可能较慢，Ollama 请求的超时时间为 120 秒；取消搜索会中止请求。
+
+插件不会安装或启动 Ollama，也不会下载模型。使用时需自行准备已运行的 Ollama 服务和模型。请求由浏览器发出，因此「本机」指浏览器所在设备，而非远程 DSH 服务器。
+
+连接被拒绝时检查服务地址；跨域访问失败时，将当前 DSH 页面的确切来源（如 `http://127.0.0.1:3080`）加入 Ollama 的 `OLLAMA_ORIGINS`，并重启 Ollama。不要使用不必要的通配符来源。非本机地址仍要求 HTTPS。
+
+---
 
 
 ## 包结构
@@ -93,12 +108,13 @@ lib/index.js        Host half（空实现，保证作为合法 Cordis 插件行
 lib/client.js       Browser half（搜索框 UI，__ModuleLoader__ 格式）
 cordis.patch.yml    组合补丁：插入 settings-search 插件行
 scripts/check.mjs   结构自检（npm test）
+scripts/ai.test.mjs AI 接口回归测试（模拟请求，不运行模型）
 ```
 
 ## 开发
 
 ```bash
-npm test   # 校验 host 半、dsh.client 清单、client bundle handoff
+npm test   # 结构检查及 Ollama / OpenAI / Anthropic 模拟接口回归测试
 ```
 
 > 历史说明：v1.1 及之前仓库曾附带 `dsh-settings-search.js`（会话内

package/lib/client.js

@@ -30,7 +30,7 @@ window.__ModuleLoader__.load({
 
     const NS = 'settings-search-v5'
     /** Single source of truth is package.json — scripts/check.mjs asserts this matches. */
-    const PKG_VERSION = '1.10.0'
+    const PKG_VERSION = '1.11.0'
     const NPM_NAME = '@objectivex666/dsh-settings-search'
     const zh = {
       title: '设置',
@@ -72,16 +72,17 @@ window.__ModuleLoader__.load({
       updFetchingNotes: '正在获取…',
       updRestart: '更新完成后重启 DSH，新版本才会生效。',
       aiSearch: 'AI 联想搜索',
-      aiDesc: '没有直接匹配时，用你配置的模型（OpenAI 兼容或 Anthropic）理解目标并推荐设置项。',
+      aiDesc: '没有直接匹配时，用你配置的模型（OpenAI 兼容、Anthropic 或本机 Ollama）理解目标并推荐设置项。',
       aiPrivacy: 'AI 请求会发送到你配置的模型服务；API Key 仅保存在本机浏览器。',
       aiOn: '已开启',
       aiOff: '已关闭',
       aiThinking: '正在用 AI 理解搜索意图…',
       aiUnavailable: 'AI 联想暂不可用',
-      aiNeedConfig: '请先在「设置搜索」页配置 Base URL、模型名与 API Key。',
+      aiNeedConfig: '请先在「设置搜索」页配置 Base URL、模型名与 API Key（Ollama 无需密钥）。',
       aiProvider: '接口类型',
       aiOpenAI: 'OpenAI 兼容',
       aiAnthropic: 'Anthropic',
+      aiOllama: 'Ollama（本机）',
       aiBaseUrl: 'Base URL',
       aiBaseUrlHint: '仅支持 http(s) 地址；非本机地址必须使用 https，以免泄露 API Key。',
       aiModel: '模型名',
@@ -112,6 +113,9 @@ window.__ModuleLoader__.load({
       aiTestErrBadResponse: '响应格式异常，请确认接口类型与服务兼容',
       aiTestErrHttp: '请求失败（HTTP 状态码错误），请检查地址与配置',
       aiTestErrUnsafeUrl: '接口地址不安全：仅支持 http(s)，非本机必须为 https，且不可包含账号密码',
+      aiOllamaErrNetwork: '无法连接 Ollama，请确认本机服务已启动、地址可达，且 OLLAMA_ORIGINS 允许当前 DSH 页面来源',
+      aiOllamaErrForbidden: 'Ollama 拒绝访问，请检查 OLLAMA_ORIGINS 是否允许当前 DSH 页面来源',
+      aiOllamaErrTimeout: 'Ollama 响应超时，请检查本机服务及模型加载状态',
       logTitle: '日志',
       logDesc: '记录插件运行、搜索与 AI 请求的关键动作，便于排查问题。',
       logEmpty: '暂无日志',
@@ -183,16 +187,17 @@ window.__ModuleLoader__.load({
       updFetchingNotes: 'Fetching…',
       updRestart: 'Restart DSH after updating for the new version to take effect.',
       aiSearch: 'AI-assisted search',
-      aiDesc: 'When no direct match exists, your own model (OpenAI-compatible or Anthropic) understands your goal and suggests settings.',
+      aiDesc: 'When no direct match exists, your own model (OpenAI-compatible, Anthropic, or local Ollama) understands your goal and suggests settings.',
       aiPrivacy: 'AI requests are sent to the model service you configure; the API key stays only in your browser.',
       aiOn: 'On',
       aiOff: 'Off',
       aiThinking: 'Understanding your intent with AI…',
       aiUnavailable: 'AI suggestions are unavailable',
-      aiNeedConfig: 'Configure Base URL, model, and API key in the Settings Search panel first.',
+      aiNeedConfig: 'Configure Base URL, model, and API key in the Settings Search panel first (Ollama needs no key).',
       aiProvider: 'API type',
       aiOpenAI: 'OpenAI-compatible',
       aiAnthropic: 'Anthropic',
+      aiOllama: 'Ollama (local)',
       aiBaseUrl: 'Base URL',
       aiBaseUrlHint: 'Only http(s) URLs are allowed; remote hosts must use https to avoid exposing your API key.',
       aiModel: 'Model',
@@ -223,6 +228,9 @@ window.__ModuleLoader__.load({
       aiTestErrBadResponse: 'Unexpected response — confirm the API type matches the service',
       aiTestErrHttp: 'Request failed (HTTP status error) — check the URL and config',
       aiTestErrUnsafeUrl: 'Unsafe URL — only http(s) is allowed, remote hosts must use https, and embedded credentials are not permitted',
+      aiOllamaErrNetwork: 'Cannot connect to Ollama. Check that the local service is running, the URL is reachable, and OLLAMA_ORIGINS allows this DSH page origin',
+      aiOllamaErrForbidden: 'Ollama denied access. Check that OLLAMA_ORIGINS allows this DSH page origin',
+      aiOllamaErrTimeout: 'Ollama timed out. Check the local service and model loading status',
       logTitle: 'Logs',
       logDesc: 'Logs key plugin actions, searches, and AI requests to help troubleshoot.',
       logEmpty: 'No logs yet',
@@ -463,6 +471,20 @@ window.__ModuleLoader__.load({
     const AI_BASE_URL_KEY = 'dsh-settings-search:ai-base-url'
     const AI_MODEL_KEY = 'dsh-settings-search:ai-model'
     const AI_API_KEY_KEY = 'dsh-settings-search:ai-api-key'
+    const OLLAMA_BASE_URL = 'http://127.0.0.1:11434'
+    const normalizeAiConfig = (cfg) => {
+      const provider = ['anthropic', 'ollama'].includes(cfg.provider) ? cfg.provider : 'openai'
+      return {
+        provider,
+        baseUrl: String(cfg.baseUrl ?? '').trim() || (provider === 'ollama' ? OLLAMA_BASE_URL : ''),
+        model: String(cfg.model ?? '').trim(),
+        apiKey: String(cfg.apiKey ?? '').trim(),
+      }
+    }
+    const aiConfigReady = (cfg) => {
+      const { provider, baseUrl, model, apiKey } = normalizeAiConfig(cfg)
+      return Boolean(baseUrl && model && (provider === 'ollama' || apiKey))
+    }
     /** Session-only API key storage avoids durable browser storage exposure. */
     const readApiKey = () => {
       const legacy = typeof localStorage !== 'undefined' ? (localStorage.getItem(AI_API_KEY_KEY) ?? '').trim() : ''
@@ -491,24 +513,23 @@ window.__ModuleLoader__.load({
       window.dispatchEvent(new CustomEvent(AI_TOGGLE_EVENT, { detail: enabled }))
     }
     const aiConfig = () => {
-      if (typeof localStorage === 'undefined' && typeof sessionStorage === 'undefined') {
-        return { provider: 'openai', baseUrl: '', model: '', apiKey: '' }
-      }
-      const providerRaw = (localStorage?.getItem(AI_PROVIDER_KEY) ?? 'openai').trim()
-      return {
-        provider: providerRaw === 'anthropic' ? 'anthropic' : 'openai',
-        baseUrl: (localStorage?.getItem(AI_BASE_URL_KEY) ?? '').trim(),
-        model: (localStorage?.getItem(AI_MODEL_KEY) ?? '').trim(),
+      const storage = typeof localStorage === 'undefined' ? null : localStorage
+      return normalizeAiConfig({
+        provider: (storage?.getItem(AI_PROVIDER_KEY) ?? 'openai').trim(),
+        baseUrl: storage?.getItem(AI_BASE_URL_KEY),
+        model: storage?.getItem(AI_MODEL_KEY),
         apiKey: readApiKey(),
-      }
+      })
     }
     const setAiConfig = (cfg) => {
+      AI_CACHE.clear()
       if (typeof localStorage === 'undefined' && typeof sessionStorage === 'undefined') return
-      writeApiKey(cfg.apiKey)
+      const normalized = normalizeAiConfig(cfg)
+      writeApiKey(normalized.apiKey)
       if (typeof localStorage === 'undefined') return
-      localStorage.setItem(AI_PROVIDER_KEY, cfg.provider === 'anthropic' ? 'anthropic' : 'openai')
-      localStorage.setItem(AI_BASE_URL_KEY, cfg.baseUrl ?? '')
-      localStorage.setItem(AI_MODEL_KEY, cfg.model ?? '')
+      localStorage.setItem(AI_PROVIDER_KEY, normalized.provider)
+      localStorage.setItem(AI_BASE_URL_KEY, normalized.baseUrl)
+      localStorage.setItem(AI_MODEL_KEY, normalized.model)
       window.dispatchEvent(new CustomEvent(AI_CONFIG_EVENT))
     }
     /** First-run onboarding: remembered via localStorage so it shows once. */
@@ -599,62 +620,109 @@ window.__ModuleLoader__.load({
       if (/\/v1\/messages$/i.test(safe)) return safe
       return /\/v1$/i.test(safe) ? `${safe}/messages` : `${safe}/v1/messages`
     }
-    const requestUserAi = async (query, labels, signal) => {
-      const { provider, baseUrl, model, apiKey } = aiConfig()
-      if (!baseUrl || !model || !apiKey) return []
-      const isAnthropic = provider === 'anthropic'
-      const url = isAnthropic ? anthropicMessagesUrl(baseUrl) : chatCompletionsUrl(baseUrl)
-      if (!url) return []
-      const startedAt = Date.now()
-      log('info', 'AI 请求', { provider, model, query: String(query).slice(0, 50) })
+    const ollamaChatUrl = (baseUrl) => {
+      const safe = safeBaseUrl(baseUrl)
+      if (!safe) return ''
+      if (/\/api\/chat$/i.test(safe)) return safe
+      if (/\/api$/i.test(safe)) return `${safe}/chat`
+      // Accept the base URL used by Ollama's OpenAI-compatible endpoint too.
+      return `${safe.replace(/\/v1(?:\/chat\/completions)?$/i, '')}/api/chat`
+    }
+    const aiRequest = (cfg, userContent, testing = false) => {
+      const { provider, baseUrl, model, apiKey } = normalizeAiConfig(cfg)
+      const messages = [{ role: 'user', content: userContent }]
+      const headers = { 'content-type': 'application/json' }
+      if (provider === 'anthropic') {
+        return {
+          url: anthropicMessagesUrl(baseUrl),
+          headers: { ...headers, 'x-api-key': apiKey, 'anthropic-version': '2023-06-01', 'anthropic-dangerous-direct-browser-access': 'true' },
+          body: { model, messages, max_tokens: testing ? 1 : 1024, ...(testing ? {} : { system: AI_PROMPT, temperature: 0 }) },
+        }
+      }
+      if (!testing) messages.unshift({ role: 'system', content: AI_PROMPT })
+      if (provider === 'ollama') {
+        return {
+          url: ollamaChatUrl(baseUrl),
+          headers,
+          body: {
+            model,
+            messages,
+            stream: false,
+            ...(testing ? {} : { format: 'json' }),
+            options: { temperature: 0, num_predict: testing ? 1 : 1024 },
+          },
+        }
+      }
+      return {
+        url: chatCompletionsUrl(baseUrl),
+        headers: { ...headers, authorization: `Bearer ${apiKey}` },
+        body: { model, messages, temperature: 0, ...(testing ? { max_tokens: 1 } : {}) },
+      }
+    }
+    const aiResponseContent = (provider, data) => {
+      if (provider === 'ollama') return data?.message?.content
+      if (provider === 'anthropic') {
+        if (!Array.isArray(data?.content) || data.content.length === 0) return undefined
+        return data.content.filter((block) => block?.type === 'text').map((block) => block.text).join('')
+      }
+      return data?.choices?.[0]?.message?.content
+    }
+    const fetchAiData = async (request, provider, signal) => {
+      const controller = new AbortController()
+      const abort = () => controller.abort()
+      if (signal?.aborted) abort()
+      else signal?.addEventListener('abort', abort, { once: true })
       let timer
       const deadline = new Promise((_, reject) => {
         timer = setTimeout(() => {
           const error = new Error('AI request timed out')
           error.name = 'AI_TIMEOUT'
           reject(error)
-        }, 12000)
+          controller.abort()
+        }, provider === 'ollama' ? 120000 : 12000)
       })
       const work = (async () => {
-        const userContent = `用户目标：${query}\n设置索引：\n${labels}`
-        const headers = {
-          'content-type': 'application/json',
-          ...(isAnthropic
-            ? { 'x-api-key': apiKey, 'anthropic-version': '2023-06-01', 'anthropic-dangerous-direct-browser-access': 'true' }
-            : { authorization: `Bearer ${apiKey}` }),
-        }
-        const body = isAnthropic
-          ? {
-              model,
-              max_tokens: 1024,
-              system: AI_PROMPT,
-              messages: [{ role: 'user', content: userContent }],
-              temperature: 0,
-            }
-          : {
-              model,
-              messages: [
-                { role: 'system', content: AI_PROMPT },
-                { role: 'user', content: userContent },
-              ],
-              temperature: 0,
-            }
-        const response = await fetch(url, {
+        const response = await fetch(request.url, {
           method: 'POST',
-          headers,
-          signal,
+          headers: request.headers,
+          signal: controller.signal,
           redirect: 'error',
-          body: JSON.stringify(body),
+          body: JSON.stringify(request.body),
+        })
+        if (!response.ok) {
+          const error = new Error(`HTTP ${response.status}`)
+          error.status = response.status
+          throw error
+        }
+        const data = await response.json().catch((error) => {
+          if (error.name === 'AbortError') throw error
+          return null
         })
-        if (!response.ok) throw new Error(`HTTP ${response.status}`)
-        const data = await response.json()
-        const content = isAnthropic
-          ? (data?.content ?? []).filter((block) => block?.type === 'text').map((block) => block.text).join('')
-          : data?.choices?.[0]?.message?.content ?? ''
-        return extractAiKeywords(content)
+        if (data?.error || typeof aiResponseContent(provider, data) !== 'string') {
+          const error = new Error('Unexpected AI response')
+          error.name = 'AI_BAD_RESPONSE'
+          throw error
+        }
+        return data
       })()
       try {
-        const keywords = await Promise.race([work, deadline])
+        return await Promise.race([work, deadline])
+      } finally {
+        clearTimeout(timer)
+        signal?.removeEventListener('abort', abort)
+      }
+    }
+    const requestUserAi = async (query, labels, signal) => {
+      const cfg = aiConfig()
+      if (!aiConfigReady(cfg)) return []
+      const { provider, model } = cfg
+      const request = aiRequest(cfg, `用户目标：${query}\n设置索引：\n${labels}`)
+      if (!request.url) return []
+      const startedAt = Date.now()
+      log('info', 'AI 请求', { provider, model, query: String(query).slice(0, 50) })
+      try {
+        const data = await fetchAiData(request, provider, signal)
+        const keywords = extractAiKeywords(aiResponseContent(provider, data))
         log('info', 'AI 结果', { count: keywords.length, duration: Date.now() - startedAt })
         return keywords
       } catch (error) {
@@ -664,15 +732,14 @@ window.__ModuleLoader__.load({
         }
         log('error', 'AI 请求失败', { message: String(error?.message ?? error).slice(0, 160), duration: Date.now() - startedAt })
         throw error
-      } finally {
-        clearTimeout(timer)
       }
     }
     const requestAiKeywords = async (query, labels, signal) => {
-      const cacheKey = query.toLowerCase()
+      const { provider, baseUrl, model } = aiConfig()
+      const cacheKey = JSON.stringify([provider, baseUrl, model, query.toLowerCase(), labels])
       if (AI_CACHE.has(cacheKey)) return AI_CACHE.get(cacheKey)
       const keywords = await requestUserAi(query, labels, signal)
-      if (keywords.length > 0) {
+      if (!signal?.aborted && keywords.length > 0) {
         // bound the cache so a long session cannot grow it without limit
         if (AI_CACHE.size >= 100) AI_CACHE.clear()
         AI_CACHE.set(cacheKey, keywords)
@@ -698,68 +765,31 @@ window.__ModuleLoader__.load({
       const code = Number(status)
       return HTTP_TEST_CODES[code] ?? (code >= 500 && code <= 599 ? 'server' : 'http')
     }
-    const testAiService = async (cfg) => {
-      const provider = cfg.provider === 'anthropic' ? 'anthropic' : 'openai'
-      const baseUrl = String(cfg.baseUrl ?? '').trim()
-      const model = String(cfg.model ?? '').trim()
-      const apiKey = String(cfg.apiKey ?? '').trim()
-      if (!baseUrl || !model || !apiKey) return { ok: false, code: 'need-config', detail: '' }
-      const isAnthropic = provider === 'anthropic'
-      const requestUrl = isAnthropic ? anthropicMessagesUrl(baseUrl) : chatCompletionsUrl(baseUrl)
-      if (!requestUrl) return { ok: false, code: 'unsafe-url', detail: '' }
+    const testAiService = async (cfg, signal) => {
+      cfg = normalizeAiConfig(cfg)
+      const { provider, model } = cfg
+      if (!aiConfigReady(cfg)) return { ok: false, code: 'need-config', detail: '' }
+      const request = aiRequest(cfg, 'ping', true)
+      if (!request.url) return { ok: false, code: 'unsafe-url', detail: '' }
       const startedAt = Date.now()
       log('info', '测试 API 服务', { provider, model })
-      let timer
-      const deadline = new Promise((_, reject) => {
-        timer = setTimeout(() => {
-          const error = new Error('AI request timed out')
-          error.name = 'AI_TIMEOUT'
-          reject(error)
-        }, 12000)
-      })
-      const work = (async () => {
-        const headers = {
-          'content-type': 'application/json',
-          ...(isAnthropic
-            ? { 'x-api-key': apiKey, 'anthropic-version': '2023-06-01', 'anthropic-dangerous-direct-browser-access': 'true' }
-            : { authorization: `Bearer ${apiKey}` }),
-        }
-        const body = isAnthropic
-          ? { model, max_tokens: 1, messages: [{ role: 'user', content: 'ping' }] }
-          : { model, messages: [{ role: 'user', content: 'ping' }], max_tokens: 1, temperature: 0 }
-        const response = await fetch(requestUrl, {
-          method: 'POST',
-          headers,
-          redirect: 'error',
-          body: JSON.stringify(body),
-        })
-        if (!response.ok) return { ok: false, code: testHttpCode(response.status), detail: `HTTP ${response.status}` }
-        const data = await response.json().catch(() => ({}))
-        const valid = isAnthropic
-          ? (Array.isArray(data?.content) && data.content.length > 0)
-          : Boolean(data?.choices?.[0]?.message)
-        if (!valid) return { ok: false, code: 'bad-response', detail: '' }
-        return { ok: true }
-      })()
       try {
-        const result = await Promise.race([work, deadline])
-        if (result.ok) log('info', '测试 API 成功', { duration: Date.now() - startedAt })
-        else log('error', '测试 API 失败', { code: result.code ?? 'http', detail: result.detail ?? '', duration: Date.now() - startedAt })
-        return result
+        await fetchAiData(request, provider, signal)
+        log('info', '测试 API 成功', { duration: Date.now() - startedAt })
+        return { ok: true }
       } catch (error) {
         if (error.name === 'AbortError') {
           log('warn', '测试 API 取消', { duration: Date.now() - startedAt })
-          throw error
+          return { ok: false, code: 'cancelled', detail: '' }
         }
         if (error.name === 'AI_TIMEOUT') {
           log('error', '测试 API 超时', { duration: Date.now() - startedAt })
           return { ok: false, code: 'timeout', detail: '' }
         }
+        const code = error.status ? testHttpCode(error.status) : error.name === 'AI_BAD_RESPONSE' ? 'bad-response' : 'network'
         const reason = String(error?.message ?? error).slice(0, 160)
         log('error', '测试 API 失败', { detail: reason, duration: Date.now() - startedAt })
-        return { ok: false, code: 'network', detail: reason }
-      } finally {
-        clearTimeout(timer)
+        return { ok: false, code, detail: reason }
       }
     }
 
@@ -1407,7 +1437,9 @@ window.__ModuleLoader__.load({
         const baseUrl = String(providerProfile.baseURL ?? '').trim()
         if (!modelId || !baseUrl) return null
         return {
-          provider: protocol === 'anthropic-messages' ? 'anthropic' : 'openai',
+          provider: providerId.toLowerCase() === 'ollama' || protocol === 'ollama'
+            ? 'ollama'
+            : protocol === 'anthropic-messages' ? 'anthropic' : 'openai',
           baseUrl,
           model: modelId,
         }
@@ -1659,12 +1691,12 @@ window.__ModuleLoader__.load({
         const results = [...shownPages, ...shownOptions, ...related.map((r) => r.item)]
 
         const cfg = aiConfig()
-        const aiConfigured = Boolean(cfg.baseUrl && cfg.model && cfg.apiKey)
+        const aiConfigured = aiConfigReady(cfg)
         const aiReady = aiEnabled && aiConfigured && q.length >= 2 && noLocalMatch
         React.useEffect(() => {
-          if (!aiReady) return
           setAiKeywords([])
           setAiStatus('idle')
+          if (!aiReady) return
           let alive = true
           const controller = new AbortController()
           const timer = setTimeout(() => {
@@ -1692,7 +1724,7 @@ window.__ModuleLoader__.load({
             clearTimeout(timer)
             controller.abort()
           }
-        }, [aiReady, aiEnabled, q])
+        }, [aiReady, aiEnabled, q, cfg.provider, cfg.baseUrl, cfg.model, cfg.apiKey])
 
         const loggedQueryRef = React.useRef('')
         React.useEffect(() => {
@@ -2031,6 +2063,8 @@ window.__ModuleLoader__.load({
         const [baseUrl, setBaseUrl] = React.useState(cfg.baseUrl)
         const [model, setModel] = React.useState(cfg.model)
         const [apiKey, setApiKey] = React.useState(cfg.apiKey)
+        const providerDrafts = React.useRef({})
+        const testAbortRef = React.useRef(null)
         const [aiSaved, setAiSaved] = React.useState(false)
         const [testState, setTestState] = React.useState('idle') // idle | loading | ok | fail
         const [testMsg, setTestMsg] = React.useState('')
@@ -2065,6 +2099,13 @@ window.__ModuleLoader__.load({
         useAutoReset(copied, () => { setCopied(false); setCopyFailed(false) })
         useAutoReset(aiSaved, () => setAiSaved(false))
 
+        React.useEffect(() => {
+          setTestState('idle')
+          setTestMsg('')
+          setAiSaved(false)
+          return () => testAbortRef.current?.abort()
+        }, [provider, baseUrl, model, apiKey])
+
         React.useEffect(() => {
           const sync = () => setLogVersion((v) => v + 1)
           window.addEventListener(LOG_EVENT, sync)
@@ -2075,13 +2116,22 @@ window.__ModuleLoader__.load({
         useAutoReset(logCleared, () => setLogCleared(false))
         useAutoReset(logExportFail, () => setLogExportFail(false))
 
+        const changeProvider = (next) => {
+          providerDrafts.current[provider] = { baseUrl, model }
+          const draft = providerDrafts.current[next]
+          setProvider(next)
+          setBaseUrl(draft?.baseUrl ?? (next === 'ollama' ? OLLAMA_BASE_URL : ''))
+          setModel(draft?.model ?? '')
+        }
+
         const saveAi = () => {
-          setAiConfig({
+          const next = normalizeAiConfig({
             provider,
-            baseUrl: baseUrl.trim(),
-            model: model.trim(),
-            apiKey: apiKey.trim(),
+            baseUrl,
+            model,
+            apiKey,
           })
+          setAiConfig(next)
           setAiSaved(true)
           log('info', '保存 AI 配置', { provider, model: model.trim() })
         }
@@ -2093,6 +2143,7 @@ window.__ModuleLoader__.load({
           try {
             const imported = await readDshModelConfig()
             if (!imported) throw new Error(t('aiDshImportEmpty'))
+            providerDrafts.current[provider] = { baseUrl, model }
             setAiConfig({ ...imported, apiKey })
             setProvider(imported.provider)
             setBaseUrl(imported.baseUrl)
@@ -2128,26 +2179,37 @@ window.__ModuleLoader__.load({
           setBaseUrl('')
           setModel('')
           setApiKey('')
+          providerDrafts.current = {}
           setAiSaved(false)
           log('info', '清除 API 配置', {})
         }
 
         const testService = async () => {
           if (testState === 'loading') return
-          if (!baseUrl.trim() || !model.trim() || !apiKey.trim()) {
+          if (!aiConfigReady({ provider, baseUrl, model, apiKey })) {
             setTestState('fail')
             setTestMsg(t('aiNeedConfig'))
             return
           }
           setTestState('loading')
           setTestMsg('')
-          const result = await testAiService({ provider, baseUrl, model, apiKey })
+          const controller = new AbortController()
+          testAbortRef.current?.abort()
+          testAbortRef.current = controller
+          const result = await testAiService({ provider, baseUrl, model, apiKey }, controller.signal)
+          if (controller.signal.aborted) return
+          if (testAbortRef.current === controller) testAbortRef.current = null
           if (result.ok) {
             setTestState('ok')
             setTestMsg(t('aiTestOk'))
           } else {
             setTestState('fail')
-            const hintKey = {
+            const ollamaHintKey = provider === 'ollama' ? {
+              network: 'aiOllamaErrNetwork',
+              forbidden: 'aiOllamaErrForbidden',
+              timeout: 'aiOllamaErrTimeout',
+            }[result.code] : null
+            const hintKey = ollamaHintKey ?? {
               auth: 'aiTestErrAuth',
               forbidden: 'aiTestErrForbidden',
               'not-found': 'aiTestErrNotFound',
@@ -2339,14 +2401,16 @@ window.__ModuleLoader__.load({
               React.createElement('select', {
                 className: 'sss-vr-input sss-vr-select',
                 name: 'dsh-settings-search-provider',
+                'aria-label': t('aiProvider'),
                 value: provider,
                 autoComplete: 'off',
                 'data-lpignore': 'true',
                 'data-form-type': 'other',
-                onChange: (event) => setProvider(event.target.value),
+                onChange: (event) => changeProvider(event.target.value),
               }, [
                 React.createElement('option', { value: 'openai' }, t('aiOpenAI')),
                 React.createElement('option', { value: 'anthropic' }, t('aiAnthropic')),
+                React.createElement('option', { value: 'ollama' }, t('aiOllama')),
               ]),
               React.createElement('div', { className: 'sss-vr-field-head' }, [
                 React.createElement('label', { className: 'sss-vr-label' }, t('aiBaseUrl')),
@@ -2355,8 +2419,10 @@ window.__ModuleLoader__.load({
                 className: 'sss-vr-input',
                 type: 'url',
                 name: 'dsh-settings-search-base-url',
+                'aria-label': t('aiBaseUrl'),
                 value: baseUrl,
-                placeholder: provider === 'anthropic' ? 'https://api.anthropic.com' : 'https://api.openai.com/v1',
+                placeholder: provider === 'ollama' ? OLLAMA_BASE_URL
+                  : provider === 'anthropic' ? 'https://api.anthropic.com' : 'https://api.openai.com/v1',
                 spellCheck: false,
                 autoComplete: 'off',
                 'data-lpignore': 'true',
@@ -2372,8 +2438,10 @@ window.__ModuleLoader__.load({
                 className: 'sss-vr-input',
                 type: 'text',
                 name: 'dsh-settings-search-model',
+                'aria-label': t('aiModel'),
                 value: model,
-                placeholder: provider === 'anthropic' ? 'claude-sonnet-4-20250514' : 'gpt-4o-mini',
+                placeholder: provider === 'ollama' ? 'qwen3:8b'
+                  : provider === 'anthropic' ? 'claude-sonnet-4-20250514' : 'gpt-4o-mini',
                 spellCheck: false,
                 autoComplete: 'off',
                 'data-lpignore': 'true',
@@ -2381,13 +2449,14 @@ window.__ModuleLoader__.load({
                 'data-form-type': 'other',
                 onChange: (event) => setModel(event.target.value),
               }),
-              React.createElement('div', { className: 'sss-vr-field-head' }, [
+              provider !== 'ollama' ? React.createElement('div', { className: 'sss-vr-field-head' }, [
                 React.createElement('label', { className: 'sss-vr-label' }, t('aiApiKey')),
-              ]),
-              React.createElement('input', {
+              ]) : null,
+              provider !== 'ollama' ? React.createElement('input', {
                 className: 'sss-vr-input sss-vr-mask',
                 type: 'text',
                 name: 'dsh-settings-search-api-key',
+                'aria-label': t('aiApiKey'),
                 value: apiKey,
                 placeholder: 'sk-…',
                 spellCheck: false,
@@ -2396,7 +2465,7 @@ window.__ModuleLoader__.load({
                 'data-1p-ignore': 'true',
                 'data-form-type': 'other',
                 onChange: (event) => setApiKey(event.target.value),
-              }),
+              }) : null,
               React.createElement('div', { className: 'sss-vr-actions' }, [
                 React.createElement('button', { className: 'sss-vr-btn sss-vr-primary', type: 'button', onClick: () => { void saveAi() } },
                   aiSaved ? t('aiConfigSaved') : t('aiConfigSave')),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@objectivex666/dsh-settings-search",
-  "version": "1.10.0",
+  "version": "1.11.0",
   "description": "DSH settings panel search plugin – instantly filter setting sections and general items with navigation guidance. Installable via `dsh plugin add`.",
   "keywords": [
     "dsh",
@@ -55,7 +55,7 @@
     "node": ">=18"
   },
   "scripts": {
-    "test": "node scripts/check.mjs"
+    "test": "node scripts/check.mjs && node --test scripts/ai.test.mjs"
   },
   "publishConfig": {
     "access": "public"

package/scripts/ai.test.mjs

@@ -0,0 +1,265 @@
+import assert from 'node:assert/strict'
+import { readFileSync } from 'node:fs'
+import test from 'node:test'
+import vm from 'node:vm'
+
+const source = readFileSync(new URL('../lib/client.js', import.meta.url), 'utf8')
+const hook = '    exports.apply = apply'
+assert.equal(source.split(hook).length, 2)
+// Expose closure helpers only in the VM; the shipped module has no test exports.
+const instrumented = source.replace(hook, `
+    Object.assign(exports, {
+      aiConfig, setAiConfig, aiConfigReady, ollamaChatUrl,
+      requestUserAi, requestAiKeywords, testAiService, LOGS,
+    })
+${hook}`)
+const storage = () => {
+  const values = new Map()
+  return {
+    getItem: (key) => values.get(key) ?? null,
+    setItem: (key, value) => values.set(key, String(value)),
+    removeItem: (key) => values.delete(key),
+  }
+}
+const json = (data, status = 200) => ({
+  ok: status >= 200 && status < 300,
+  status,
+  json: async () => data,
+})
+const plain = (value) => JSON.parse(JSON.stringify(value))
+const ollama = { provider: 'ollama', baseUrl: '', model: 'local-model:7b', apiKey: '' }
+const openai = { provider: 'openai', baseUrl: 'https://example.test/v1', model: 'cloud-model', apiKey: 'test-secret' }
+const anthropic = { ...openai, provider: 'anthropic', baseUrl: 'https://example.test' }
+const reply = (provider, content = '{"keywords":["Language"]}') => provider === 'ollama'
+  ? { message: { role: 'assistant', content }, done: true }
+  : provider === 'anthropic'
+    ? { content: [{ type: 'text', text: content }] }
+    : { choices: [{ message: { role: 'assistant', content } }] }
+
+function load(fetchImpl = async () => json(reply('ollama'))) {
+  let handoff
+  const calls = []
+  const sandbox = {
+    window: { __ModuleLoader__: { load: (value) => { handoff = value } }, dispatchEvent() {} },
+    localStorage: storage(),
+    sessionStorage: storage(),
+    CustomEvent: class { constructor(type) { this.type = type } },
+    URL,
+    AbortController,
+    setTimeout,
+    clearTimeout,
+    fetch: async (url, options) => {
+      const call = { url, ...options, body: JSON.parse(options.body) }
+      calls.push(call)
+      return fetchImpl(call)
+    },
+  }
+  vm.runInNewContext(instrumented, sandbox, { filename: 'lib/client.js' })
+  const api = handoff.factory(() => ({}))
+  return { api, calls, sandbox }
+}
+
+test('Ollama configuration persists and requires a model but no key', () => {
+  const { api, sandbox } = load()
+  api.setAiConfig({ ...ollama, model: ' local-model:7b ' })
+  assert.deepEqual(plain(api.aiConfig()), { ...ollama, baseUrl: 'http://127.0.0.1:11434' })
+  assert.equal(api.aiConfigReady(api.aiConfig()), true)
+  assert.equal(api.aiConfigReady({ ...ollama, model: ' ' }), false)
+  assert.equal(sandbox.localStorage.getItem('dsh-settings-search:ai-provider'), 'ollama')
+  assert.equal(sandbox.sessionStorage.getItem('dsh-settings-search:ai-api-key'), null)
+})
+
+test('legacy providers still require keys and unknown provider falls back to OpenAI', () => {
+  const { api } = load()
+  for (const cfg of [openai, anthropic]) {
+    assert.equal(api.aiConfigReady(cfg), true)
+    assert.equal(api.aiConfigReady({ ...cfg, apiKey: ' ' }), false)
+  }
+  api.setAiConfig({ ...openai, provider: 'unknown' })
+  assert.equal(api.aiConfig().provider, 'openai')
+})
+
+test('Ollama accepts root, native, compatibility, proxy and IPv6 base URLs', () => {
+  const { api } = load()
+  for (const suffix of ['', '/', '/api', '/api/', '/api/chat', '/api/chat/', '/v1', '/v1/chat/completions']) {
+    assert.equal(api.ollamaChatUrl(`http://localhost:11434${suffix}`), 'http://localhost:11434/api/chat')
+  }
+  assert.equal(api.ollamaChatUrl('http://[::1]:11434/'), 'http://[::1]:11434/api/chat')
+  assert.equal(api.ollamaChatUrl('https://example.test/ollama/v1/'), 'https://example.test/ollama/api/chat')
+})
+
+test('Ollama search uses native non-streaming JSON and never sends a stored cloud key', async () => {
+  const { api, calls, sandbox } = load(async () => json(reply('ollama', '{"keywords":["Language","Language"," Theme ",42]}')))
+  api.setAiConfig({ ...ollama, apiKey: 'test-secret' })
+  const keywords = await api.requestUserAi('unmatched intent', 'Language\nTheme')
+  assert.deepEqual(plain(keywords), ['Language', 'Theme'])
+  const [call] = calls
+  assert.equal(call.url, 'http://127.0.0.1:11434/api/chat')
+  assert.equal(call.method, 'POST')
+  assert.equal(call.redirect, 'error')
+  assert.deepEqual(plain(call.headers), { 'content-type': 'application/json' })
+  assert.equal(call.body.model, ollama.model)
+  assert.equal(call.body.stream, false)
+  assert.equal(call.body.format, 'json')
+  assert.equal(call.body.options.temperature, 0)
+  assert.equal(call.body.options.num_predict, 1024)
+  assert.equal(call.body.messages[0].role, 'system')
+  assert.match(call.body.messages[1].content, /unmatched intent/)
+  assert.match(call.body.messages[1].content, /Language\nTheme/)
+  assert.equal(JSON.stringify(call.body).includes('test-secret'), false)
+  assert.equal(JSON.stringify(api.LOGS).includes('test-secret'), false)
+  assert.equal(sandbox.localStorage.getItem('dsh-settings-search:ai-api-key'), null)
+})
+
+test('Ollama connectivity test works without an API key and limits generation', async () => {
+  const { api, calls } = load()
+  assert.deepEqual(plain(await api.testAiService(ollama)), { ok: true })
+  assert.deepEqual(plain(calls[0].body), {
+    model: ollama.model,
+    messages: [{ role: 'user', content: 'ping' }],
+    stream: false,
+    options: { temperature: 0, num_predict: 1 },
+  })
+  assert.equal(calls[0].headers.authorization, undefined)
+})
+
+for (const cfg of [openai, anthropic]) {
+  test(`${cfg.provider} search and test preserve protocol, authentication and redirect protection`, async () => {
+    const { api, calls } = load(async () => json(reply(cfg.provider)))
+    api.setAiConfig(cfg)
+    assert.deepEqual(plain(await api.requestUserAi('intent', 'Language')), ['Language'])
+    assert.deepEqual(plain(await api.testAiService(cfg)), { ok: true })
+    for (const call of calls) {
+      assert.equal(call.redirect, 'error')
+      assert.equal(call.body.model, cfg.model)
+      assert.equal(call.body.stream, undefined)
+      assert.equal(call.body.options, undefined)
+      if (cfg.provider === 'anthropic') {
+        assert.equal(call.url, 'https://example.test/v1/messages')
+        assert.equal(call.headers['x-api-key'], cfg.apiKey)
+        assert.equal(call.headers['anthropic-version'], '2023-06-01')
+        assert.equal(call.headers['anthropic-dangerous-direct-browser-access'], 'true')
+        assert.equal(call.headers.authorization, undefined)
+      } else {
+        assert.equal(call.url, 'https://example.test/v1/chat/completions')
+        assert.equal(call.headers.authorization, `Bearer ${cfg.apiKey}`)
+      }
+    }
+    assert.equal(calls[1].body.max_tokens, 1)
+    assert.equal(calls[1].body.messages.length, 1)
+    if (cfg.provider === 'anthropic') assert.equal(typeof calls[0].body.system, 'string')
+    else assert.equal(calls[0].body.messages[0].role, 'system')
+  })
+}
+
+test('incomplete configuration never sends requests', async () => {
+  const { api, calls } = load()
+  for (const cfg of [{ ...ollama, model: '' }, { ...openai, apiKey: '' }, { ...anthropic, apiKey: '' }]) {
+    api.setAiConfig(cfg)
+    assert.deepEqual(plain(await api.requestUserAi('intent', 'Language')), [])
+    assert.equal((await api.testAiService(cfg)).code, 'need-config')
+  }
+  assert.equal(calls.length, 0)
+})
+
+test('unsafe URLs are rejected before search and connectivity requests', async () => {
+  const { api, calls } = load()
+  for (const baseUrl of ['javascript:alert(1)', 'file:///tmp/model', 'http://example.test', 'https://user:pass@example.test']) {
+    const cfg = { ...ollama, baseUrl }
+    api.setAiConfig(cfg)
+    assert.deepEqual(plain(await api.requestUserAi('intent', 'Language')), [])
+    assert.equal((await api.testAiService(cfg)).code, 'unsafe-url')
+  }
+  assert.equal(calls.length, 0)
+})
+
+for (const [status, code] of [[400, 'bad-request'], [401, 'auth'], [403, 'forbidden'], [404, 'not-found'], [429, 'rate-limit'], [500, 'server']]) {
+  test(`Ollama HTTP ${status} produces ${code}`, async () => {
+    const { api } = load(async () => json({ error: 'test error' }, status))
+    assert.equal((await api.testAiService(ollama)).code, code)
+    api.setAiConfig(ollama)
+    await assert.rejects(api.requestUserAi('intent', 'Language'), { message: `HTTP ${status}` })
+  })
+}
+
+test('malformed responses and native error payloads cannot pass connectivity checks', async () => {
+  for (const data of [null, {}, { message: {} }, { message: { content: 123 } }, { error: 'bad model', ...reply('ollama') }]) {
+    const { api } = load(async () => json(data))
+    assert.equal((await api.testAiService(ollama)).code, 'bad-response')
+  }
+  const { api } = load(async () => ({ ok: true, json: async () => { throw new SyntaxError('invalid JSON') } }))
+  assert.equal((await api.testAiService(ollama)).code, 'bad-response')
+})
+
+test('network failure is reported without claiming connection success', async () => {
+  const { api } = load(async () => { throw new TypeError('Failed to fetch') })
+  assert.equal((await api.testAiService(ollama)).code, 'network')
+})
+
+const abortedFetch = ({ signal }) => new Promise((_, reject) => {
+  const abort = () => reject(Object.assign(new Error('aborted'), { name: 'AbortError' }))
+  if (signal.aborted) abort()
+  else signal.addEventListener('abort', abort, { once: true })
+})
+
+for (const cfg of [ollama, openai, anthropic]) {
+  test(`${cfg.provider} timeout aborts the request and uses its provider deadline`, async () => {
+    const { api, sandbox, calls } = load(abortedFetch)
+    let expire
+    let cleared = false
+    sandbox.setTimeout = (callback, ms) => {
+      assert.equal(ms, cfg.provider === 'ollama' ? 120000 : 12000)
+      expire = callback
+      return 1
+    }
+    sandbox.clearTimeout = () => { cleared = true }
+    const result = api.testAiService(cfg)
+    expire()
+    assert.equal((await result).code, 'timeout')
+    assert.equal(calls[0].signal.aborted, true)
+    assert.equal(cleared, true)
+  })
+}
+
+test('search and connectivity propagate cancellation, including pre-aborted signals', async () => {
+  const { api, calls } = load(abortedFetch)
+  api.setAiConfig(ollama)
+  const controller = new AbortController()
+  const search = api.requestUserAi('intent', 'Language', controller.signal)
+  controller.abort()
+  await assert.rejects(search, { name: 'AbortError' })
+  assert.equal(calls[0].signal.aborted, true)
+  assert.equal((await api.testAiService(ollama, controller.signal)).code, 'cancelled')
+})
+
+test('a successful request clears its timer and detaches the caller abort listener', async () => {
+  const { api, calls, sandbox } = load()
+  const cleared = []
+  sandbox.setTimeout = () => 7
+  sandbox.clearTimeout = (timer) => cleared.push(timer)
+  const controller = new AbortController()
+  api.setAiConfig(ollama)
+  await api.requestUserAi('intent', 'Language', controller.signal)
+  controller.abort()
+  assert.equal(calls[0].signal.aborted, false)
+  assert.deepEqual(cleared, [7])
+})
+
+test('cache respects settings labels and is invalidated when provider configuration changes', async () => {
+  let provider = 'ollama'
+  const { api, calls } = load(async () => json(reply(provider)))
+  api.setAiConfig(ollama)
+  await api.requestAiKeywords('intent', 'Language')
+  await api.requestAiKeywords('intent', 'Language')
+  assert.equal(calls.length, 1)
+  await api.requestAiKeywords('intent', 'Theme')
+  assert.equal(calls.length, 2)
+  api.setAiConfig({ ...ollama, model: 'another-model' })
+  await api.requestAiKeywords('intent', 'Language')
+  assert.equal(calls.length, 3)
+  provider = 'openai'
+  api.setAiConfig(openai)
+  await api.requestAiKeywords('intent', 'Language')
+  assert.equal(calls.length, 4)
+  assert.equal(calls[3].headers.authorization, `Bearer ${openai.apiKey}`)
+})