dsh-model-health 0.2.18 → 0.2.20

package/README.md

@@ -53,6 +53,7 @@ dsh --profile web --dump-config | grep dsh-model-health   # 配置层含本行
 **特性一览：**
 
 - 支持 `llm-pi-ai`（多协议自定义提供商）与 `llm-deepseek`（官方）两类配置来源
+- 内置模型显示与健康探测：未写入配置文件的宿主内置模型（如 DeepSeek 默认目录 deepseek-flash / deepseek-v4-pro）自动从宿主 `llm` 注册表读取并去重合并；探测经宿主运行时 `ctx.llm.stream` 发起最小请求（max_tokens=1），协议、端点与凭据全部复用宿主适配器，无需额外配置
 - 可用性测试支持 `openai-completions` 与 `deepseek` 协议，其余协议自动标记「跳过」；对 200 响应校验 body 确实是成功应答（部分网关 200 但 body 是错误对象）
 - API Key 通过 DSH credential service（`ctx.credentials.resolve`）解析，密钥不会输出到浏览器
 - 测试结果（状态/延迟/错误）持久化到 localStorage，刷新页面后仍可见

package/dist/index.js

@@ -307,7 +307,7 @@ async function probeModel(row, apiKey) {
 		clearTimeout(timer);
 	}
 }
-function createTestRouteHandler(resolveApiKey, profile) {
+function createTestRouteHandler(resolveApiKey, profile, probeExtra) {
 	return async function handler(req, res) {
 		if (!isLocalOrigin(req)) {
 			sendJson(res, 403, {
@@ -331,6 +331,17 @@ function createTestRouteHandler(resolveApiKey, profile) {
 		try {
 			const found = collectModels(readSettingsCached(profile)).find((r) => r.key === key);
 			if (!found) {
+				if (probeExtra) {
+					const probed = await probeExtra(key).catch(() => void 0);
+					if (probed) {
+						sendJson(res, 200, {
+							ok: true,
+							key,
+							...probed
+						});
+						return;
+					}
+				}
 				sendJson(res, 404, {
 					ok: false,
 					error: `未找到 key 为 ${key} 的模型（配置可能已变化，请刷新列表）`
@@ -374,6 +385,125 @@ function createTestRouteHandler(resolveApiKey, profile) {
 	};
 }
 //#endregion
+//#region src/host/registry.ts
+/** 从宿主 llm 注册表收集全部已注册模型（内置 + 配置驱动）。 */
+async function collectRegistryModels(llm) {
+	const providers = llm.listProviders();
+	return (await Promise.all(providers.map(async (p) => {
+		let models;
+		try {
+			models = await llm.listModels(p.id);
+		} catch {
+			return [];
+		}
+		return Promise.all(models.map(async (m) => {
+			let contextWindow = "-";
+			let maxTokens = "-";
+			if (typeof llm.resolveModelInfo === "function") try {
+				const info = await llm.resolveModelInfo(p.id, m.id);
+				if (info?.context?.contextWindow) contextWindow = info.context.contextWindow;
+				if (info?.defaultMaxTokens) maxTokens = info.defaultMaxTokens;
+			} catch {}
+			return {
+				key: `${p.id}/${m.id}`,
+				provider: p.id,
+				displayName: p.name || p.id,
+				modelId: m.id,
+				modelName: m.name || m.id,
+				contextWindow,
+				maxTokens,
+				input: m.inputModalities && m.inputModalities.length > 0 ? m.inputModalities.join("/") : "text",
+				api: "registry",
+				baseURL: "-",
+				apiKeyEnv: ""
+			};
+		}));
+	}))).flat();
+}
+/** 去掉注册表中与文件配置重复的模型（同一 provider 下的同一 modelId）。 */
+function dedupeRegistryRows(registryRows, fileRows) {
+	const seen = new Set(fileRows.map((r) => `${r.provider}\u0000${r.modelId}`));
+	return registryRows.filter((r) => !seen.has(`${r.provider}\u0000${r.modelId}`));
+}
+const RUNTIME_PROBE_TIMEOUT_MS = 3e4;
+/** 经宿主 llm 运行时对单个模型发起最小探测请求。 */
+async function probeViaRuntime(llm, provider, model) {
+	if (typeof llm.stream !== "function") return {
+		status: "fail",
+		latency: 0,
+		error: "宿主 llm 服务不支持 stream 调用"
+	};
+	const controller = new AbortController();
+	const timer = setTimeout(() => controller.abort(), RUNTIME_PROBE_TIMEOUT_MS);
+	const start = Date.now();
+	try {
+		const stream = llm.stream({
+			provider,
+			model,
+			messages: [{
+				role: "user",
+				content: [{
+					type: "text",
+					text: "hi"
+				}]
+			}],
+			maxTokens: 1,
+			signal: controller.signal
+		});
+		let firstChunkLatency = 0;
+		for await (const chunk of stream) {
+			if (chunk.type === "finish") {
+				const kind = chunk.reason?.kind;
+				if (kind === "error" || kind === "aborted") {
+					const failure = chunk.reason?.failure;
+					return {
+						status: "fail",
+						latency: Date.now() - start,
+						error: failure?.message ? `${failure.code ?? kind}${failure.status != null ? ` (HTTP ${failure.status})` : ""}：${failure.message}` : `模型调用失败（${kind}）`
+					};
+				}
+				return {
+					status: "ok",
+					latency: firstChunkLatency || Date.now() - start
+				};
+			}
+			if (!firstChunkLatency && chunk.type !== "usage") firstChunkLatency = Date.now() - start;
+		}
+		return {
+			status: "fail",
+			latency: Date.now() - start,
+			error: "模型流提前结束（未收到 finish）"
+		};
+	} catch (e) {
+		return {
+			status: "fail",
+			latency: Date.now() - start,
+			error: e?.name === "AbortError" ? `超时（${RUNTIME_PROBE_TIMEOUT_MS / 1e3}s）` : e?.message || String(e)
+		};
+	} finally {
+		clearTimeout(timer);
+	}
+}
+/**
+* 按列表 key（provider/modelId）探测注册表模型；key 不属于注册表时返回
+* undefined（调用方继续走 404）。modelId 可含 '/'（如 pi-ai 的路由模型 id），
+* 因此 provider 段取第一个 '/' 之前。
+*/
+async function probeRegistryKey(llm, key) {
+	if (typeof llm.stream !== "function") return void 0;
+	const slash = key.indexOf("/");
+	if (slash <= 0 || slash === key.length - 1) return void 0;
+	const provider = key.slice(0, slash);
+	const model = key.slice(slash + 1);
+	try {
+		if (!llm.listProviders().some((p) => p.id === provider)) return void 0;
+		if (!(await llm.listModels(provider)).some((m) => m.id === model)) return void 0;
+	} catch {
+		return;
+	}
+	return probeViaRuntime(llm, provider, model);
+}
+//#endregion
 //#region src/index.ts
 const name = "dsh-model-health";
 const inject = [
@@ -383,6 +513,18 @@ const inject = [
 ];
 function apply(ctx) {
 	const profile = ctx.get("profileContext");
+	const llm = ctx.get("llm");
+	/** 配置文件模型 + 注册表补充模型（内置等文件里没有的条目）。 */
+	const collectAllRows = async () => {
+		const fileRows = collectModels(readSettingsCached(profile));
+		if (!llm) return fileRows;
+		try {
+			const extra = dedupeRegistryRows(await collectRegistryModels(llm), fileRows);
+			return [...fileRows, ...extra];
+		} catch {
+			return fileRows;
+		}
+	};
 	ctx.tools.register(defineTool({
 		name: "list_models",
 		description: "列出当前 DSH 配置（settings.yaml 或 profile patch）中已配置的所有模型，以 Markdown 表格展示。用于在对话中查看当前可用模型清单。",
@@ -396,15 +538,15 @@ function apply(ctx) {
 		},
 		async execute() {
 			const source = settingsPaths(profile).join(" + ");
-			return renderMarkdownTable(collectModels(readSettingsCached(profile)), source);
+			return renderMarkdownTable(await collectAllRows(), source);
 		}
 	}));
 	ctx.webServer.register({
 		kind: "exact",
 		path: "/api/model-health/json",
-		handler: (_req, res) => {
+		handler: async (_req, res) => {
 			try {
-				const rows = collectModels(readSettingsCached(profile));
+				const rows = await collectAllRows();
 				sendJson(res, 200, {
 					ok: true,
 					count: rows.length,
@@ -429,7 +571,7 @@ function apply(ctx) {
 			} catch {
 				return;
 			}
-		}, profile)
+		}, profile, llm ? (key) => probeRegistryKey(llm, key) : void 0)
 	});
 	console.log(`[dsh-model-health] ready — tool "list_models" + routes GET /api/model-health/json, POST /api/model-health/test`);
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-model-health",
-  "version": "0.2.18",
+  "version": "0.2.20",
   "description": "DSH plugin that lists all configured models with availability/latency testing, by reading $DSH_HOME/settings.yaml.",
   "icon": "./icon.svg",
   "license": "MIT",