Merge pull request #73 from Autumn-27/fix/llm-test-verify-reply

fix(llm): 测试连接校验模型确有回复，并按配置的收发模式测

agent/provider.go

@@ -357,11 +357,12 @@ func quotaAwareHTTPClient(proxy string) (*http.Client, error) {
 }
 
 // TestConnection makes a minimal real completion to verify the provider/model/
-// endpoint/key actually work. Returns the round-trip latency.
-func TestConnection(ctx context.Context, c Config) (time.Duration, error) {
+// endpoint/key actually work. Returns the round-trip latency and the model's
+// reply text.
+func TestConnection(ctx context.Context, c Config) (time.Duration, string, error) {
 	prov, err := c.NewProvider()
 	if err != nil {
-		return 0, err
+		return 0, "", err
 	}
 	ctx, cancel := context.WithTimeout(ctx, 30*time.Second)
 	defer cancel()
@@ -371,7 +372,7 @@ func TestConnection(ctx context.Context, c Config) (time.Duration, error) {
 	// 就撞到输出上限(finish=length)、被截断,连接测试虽仍算通(err=nil)但显示成
 	// "已中断/length/resume" 一团糟。给足预算让它把 OK 干净吐完(finish=stop)。
 	// EscalateMaxTokens 保持 false:不因截断而抬额重试,避免 resume 循环空烧。
-	_, err = agentcore.Run(ctx, agentcore.Options{
+	reply, err := agentcore.Run(ctx, agentcore.Options{
 		Provider:       prov,
 		SystemPrompt:   []string{"你是连接测试。直接输出两个字符 OK 即可，不要思考、不要解释、不要别的。"},
 		PermissionMode: acperm.ModeBypass,
@@ -379,5 +380,16 @@ func TestConnection(ctx context.Context, c Config) (time.Duration, error) {
 		MaxTokens:      8192,
 		NonStreaming:   !c.Stream, // 用该 profile 的真实收发模式做连接测试
 	}, "ping")
-	return time.Since(start), err
+	lat := time.Since(start)
+	if err != nil {
+		return lat, "", err
+	}
+	// err==nil 还不够：请求通了但模型一个字都不吐的情况真实存在(思考把预算烧光、
+	// 正文被安全策略吞掉、兼容层把 content 丢了)。这种配置在会话里就是"不回话",
+	// 测试却报成功——正是本项要消除的落差。没有可见正文一律判失败。
+	reply = strings.TrimSpace(reply)
+	if reply == "" {
+		return lat, "", fmt.Errorf("模型无回复内容（请求已通，但未返回任何文本）")
+	}
+	return lat, reply, nil
 }

server/server.go

@@ -1337,6 +1337,7 @@ func (s *Server) testLLM(w http.ResponseWriter, r *http.Request) {
 		ThinkingType    string `json:"thinking_type"`
 		ReasoningEffort string `json:"reasoning_effort"`
 		ProfileID       *int64 `json:"profile_id"` // 测已存 profile 时传入：api_key 为空则用它存的 key
+		Streaming       *bool  `json:"streaming"`  // 省略=流式，与保存 profile 时同一套默认
 	}
 	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
 		writeErr(w, 400, err.Error())
@@ -1347,6 +1348,11 @@ func (s *Server) testLLM(w http.ResponseWriter, r *http.Request) {
 	// reasoning_effort/thinking field fails the test too (no false "test ok, run 400").
 	cfg.ThinkingType = req.ThinkingType
 	cfg.ReasoningEffort = req.ReasoningEffort
+	// 同理，收发模式也照该 profile 的选择来：只支持其中一种通道的端点必须在这里就
+	// 暴露，而不是等会话里才发现"测试通过的配置根本跑不动"。
+	if req.Streaming != nil {
+		cfg.Stream = *req.Streaming
+	}
 	// API Key 解析优先级：表单输入 > 指定 profile 存的 key > 全局配置的 key。
 	// 已存 profile 的 key 不回传浏览器，所以测试已存配置时表单为空，需从 DB 取。
 	if cfg.APIKey == "" && req.ProfileID != nil {
@@ -1363,12 +1369,27 @@ func (s *Server) testLLM(w http.ResponseWriter, r *http.Request) {
 		writeJSON(w, 200, map[string]any{"ok": false, "error": "未提供 API Key"})
 		return
 	}
-	lat, err := agent.TestConnection(r.Context(), cfg)
+	lat, reply, err := agent.TestConnection(r.Context(), cfg)
 	if err != nil {
 		writeJSON(w, 200, map[string]any{"ok": false, "error": err.Error()})
 		return
 	}
-	writeJSON(w, 200, map[string]any{"ok": true, "latency_ms": lat.Milliseconds(), "model": cfg.Model})
+	// 回传模型实际回复，让"测试通过"有据可查：看得见它确实说了话，而不只是 HTTP 200。
+	writeJSON(w, 200, map[string]any{
+		"ok": true, "latency_ms": lat.Milliseconds(), "model": cfg.Model, "reply": truncateReply(reply),
+	})
+}
+
+// truncateReply clips a connection-test reply for display. A model told to answer
+// "OK" can still ramble (or think out loud); the UI only needs enough to show it
+// really said something. Rune-based so multibyte text never splits mid-character.
+func truncateReply(s string) string {
+	const max = 200
+	r := []rune(s)
+	if len(r) <= max {
+		return s
+	}
+	return string(r[:max]) + "…"
 }
 
 type createTaskReq struct {

web/src/app/(main)/system/llm/page.tsx

@@ -369,8 +369,22 @@ function ProfileSheet({
     try {
       // 用配置实际会跑的思考参数来测，这样不支持该字段的模型在这里就失败，
       // 而不是等到跑任务时才炸。传 profile id：Key 输入框留空时用已存的 Key。
-      const r = await api.testLLM(format, model, baseUrl, apiKey, proxy, toStore(thinkingType), toStore(effort), profileId);
-      if (r.ok) toast.success(`连接成功 · ${r.latency_ms ?? "?"}ms · ${r.model ?? model}`);
+      const r = await api.testLLM(
+        format,
+        model,
+        baseUrl,
+        apiKey,
+        proxy,
+        toStore(thinkingType),
+        toStore(effort),
+        profileId,
+        streaming,
+      );
+      // 回复内容一并展示：看得见模型确实说了话，才算和会话里跑通是一回事。
+      if (r.ok)
+        toast.success(`连接成功 · ${r.latency_ms ?? "?"}ms · ${r.model ?? model}`, {
+          description: r.reply ? `回复：${r.reply}` : undefined,
+        });
       else toast.error(`连接失败：${r.error ?? "未知"}`);
     } catch (e) {
       toast.error(`测试出错：${(e as Error).message}`);

web/src/lib/api.ts

@@ -646,8 +646,10 @@ export const api = {
     thinking_type = "",
     reasoning_effort = "",
     profile_id?: number,
+    streaming = true, // 用该配置真实的收发模式来测，别让"流式能通、非流式不通"漏到会话里
   ) =>
-    post<{ ok: boolean; error?: string; latency_ms?: number; model?: string }>("/llm/test", {
+    // reply = 模型实际回复(已截断);一个字都不回的配置后端直接判失败
+    post<{ ok: boolean; error?: string; latency_ms?: number; model?: string; reply?: string }>("/llm/test", {
       provider,
       model,
       base_url,
@@ -656,6 +658,7 @@ export const api = {
       thinking_type,
       reasoning_effort,
       profile_id,
+      streaming,
     }),
   llmProfiles: () => get<{ profiles: LLMProfile[] }>("/llm/profiles").then((r) => arr(r.profiles)),
   saveLLMProfile: (p: {

web/src/lib/mock/handler.ts

@@ -1155,7 +1155,8 @@ function route(m: string, path: string, seg: string[], q: URLSearchParams, b: Re
   }
   if (path === "/llm" && m === "GET") return D.llmConfig;
   if (path === "/llm" && m === "POST") return { ok: true };
-  if (path === "/llm/test") return { ok: true, latency_ms: 128, model: String(b.model ?? "claude-opus-4-8") };
+  if (path === "/llm/test")
+    return { ok: true, latency_ms: 128, model: String(b.model ?? "claude-opus-4-8"), reply: "OK" };
   if (path === "/llm/profiles" && m === "GET") return { profiles: D.llmProfiles };
   if (path === "/llm/profiles" && m === "POST") return { id: Number(b.id) || 3 };
   if (path === "/llm/profiles/active") return { ok: true };