dsh-model-health 0.2.19 → 0.2.20

package/README.md

@@ -53,7 +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：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；此类模型协议与凭据由宿主适配器管理，暂不支持在线探测（标记「跳过」）
+- 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, findExtraRow) {
+function createTestRouteHandler(resolveApiKey, profile, probeExtra) {
 	return async function handler(req, res) {
 		if (!isLocalOrigin(req)) {
 			sendJson(res, 403, {
@@ -331,14 +331,16 @@ function createTestRouteHandler(resolveApiKey, profile, findExtraRow) {
 		try {
 			const found = collectModels(readSettingsCached(profile)).find((r) => r.key === key);
 			if (!found) {
-				if (findExtraRow ? await findExtraRow(key).catch(() => void 0) : void 0) {
-					sendJson(res, 200, {
-						ok: true,
-						key,
-						status: "skip",
-						error: "宿主内置模型：协议与凭据由宿主适配器管理，暂不支持在线探测"
-					});
-					return;
+				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,
@@ -423,6 +425,84 @@ 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";
@@ -491,9 +571,7 @@ function apply(ctx) {
 			} catch {
 				return;
 			}
-		}, profile, llm ? async (key) => {
-			return (await collectRegistryModels(llm)).find((r) => r.key === key);
-		} : void 0)
+		}, 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": "0.2.19",
+  "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",