修复 AI 回复为空与开头数字被吞

日志里一批「成功但回复为空」的调用，completion tokens 恰好等于配置的
上限 256：推理模型把预算全花在思维链上，正文没写出来就被截断。

- 解析 OpenAI 兼容响应时读取 reasoning_content，并据此与 finish_reason
  判断「被截断」。
- 被截断时自动加大预算重试一次（4 倍，下限 1024、上限 4096），两次调用
  都记入日志以便核算成本；重试后仍为空才报错，并说明是推理占满预算，
  提示调高 tokens 或改用非推理模型。
- sanitizeAIReply 原本用 TrimLeft 去掉列表符号，连同任何开头的数字一起
  吃掉了——「30岁啦」发出去变成「岁啦」。改为只匹配真正的列表前缀
  （- * • 或 "1. " "2、"），号码、年龄、时长都不再受损。

线上两台已将模型 max_tokens 由 256 调整为 2048。

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

im/backend/internal/app/ai_agent.go

@@ -8,6 +8,7 @@ import (
 	"fmt"
 	"log"
 	"math/rand"
+	"regexp"
 	"strings"
 	"sync"
 	"time"
@@ -258,14 +259,31 @@ func (a *App) processAIReplyJob(ctx context.Context, job aiReplyJob) error {
 	result, callErr := provider.Chat(callCtx, request)
 	latency := int(time.Since(started).Milliseconds())
 	a.logAICall(ctx, model.ID, "chat", job.AgentUserID, job.ConversationID, latency, result, callErr)
+	// Reasoning models spend the budget thinking before they answer. With a
+	// small ceiling the call "succeeds" with nothing to say, so give it room and
+	// ask once more rather than logging a silent empty reply.
+	if callErr == nil && result.truncated() {
+		retry := request
+		retry.MaxTokens = aiRetryMaxTokens(aiRequestMaxTokens(request, model))
+		retryCtx, cancelRetry := context.WithTimeout(ctx, time.Duration(model.TimeoutMS+2000)*time.Millisecond)
+		started = time.Now()
+		result, callErr = provider.Chat(retryCtx, retry)
+		cancelRetry()
+		a.logAICall(ctx, model.ID, "chat", job.AgentUserID, job.ConversationID, int(time.Since(started).Milliseconds()), result, callErr)
+	}
 	if callErr != nil {
 		if fallback := strings.TrimSpace(a.configPlain(ctx, "ai.fallback_text", "")); fallback != "" {
 			_ = a.sendAIReply(ctx, job, fallback)
 		}
 		return callErr
 	}
+	// A short answer that survived the retry is still worth sending; only an
+	// empty one is a failure.
 	text := sanitizeAIReply(result.Text)
 	if text == "" {
+		if result.truncated() {
+			return fmt.Errorf("模型把 %d tokens 全部用于推理，未产出正文；请在模型配置中调高最大 tokens，或改用非推理模型", result.CompletionTokens)
+		}
 		return fmt.Errorf("模型返回内容为空")
 	}
 	if first && a.configBool(ctx, "ai.disclose_in_chat", true) {
@@ -433,6 +451,11 @@ func aiSystemPrompt(binding aiAgentBinding, disclose bool) string {
 }
 
 // Models drift into assistant formatting; the chat bubble wants one plain line.
+// Only an actual list marker is removed. Trimming the digit characters
+// themselves ate the number out of every reply that opened with one — "30岁啦"
+// arrived as "岁啦".
+var aiListMarker = regexp.MustCompile(`^(?:[-*•]+\s*|\d{1,2}(?:[.)]\s+|、\s*))`)
+
 func sanitizeAIReply(text string) string {
 	text = strings.TrimSpace(text)
 	if text == "" {
@@ -445,7 +468,7 @@ func sanitizeAIReply(text string) string {
 	text = replacer.Replace(text)
 	lines := []string{}
 	for _, line := range strings.Split(text, "\n") {
-		line = strings.TrimSpace(strings.TrimLeft(strings.TrimSpace(line), "-*•0123456789. "))
+		line = strings.TrimSpace(aiListMarker.ReplaceAllString(strings.TrimSpace(line), ""))
 		if line != "" {
 			lines = append(lines, line)
 		}

im/backend/internal/app/ai_enqueue_integration_test.go

@@ -4,6 +4,10 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
 	"testing"
 )
 
@@ -168,3 +172,77 @@ func TestDefaultAIModelAcceptsNameOrIDMySQL(t *testing.T) {
 		t.Fatalf("留空时使用标记为默认的一条: id = %d", model.ID)
 	}
 }
+
+// End to end over a real database and a real HTTP model: the first call comes
+// back blank because the reasoning ate the budget, and the reply still has to
+// arrive.
+func TestAIReplyRetriesAfterReasoningExhaustsBudgetMySQL(t *testing.T) {
+	db := isolatedIMDatabase(t)
+	a := &App{db: db, hub: NewHub(), config: Config{Environment: "development", JWTSecret: "isolated-im-regression-secret-only"}}
+	budgets := []int{}
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		var body map[string]any
+		raw, _ := io.ReadAll(io.LimitReader(r.Body, 1<<20))
+		_ = json.Unmarshal(raw, &body)
+		budgets = append(budgets, int(body["max_tokens"].(float64)))
+		w.Header().Set("Content-Type", "application/json")
+		if len(budgets) == 1 {
+			_, _ = w.Write([]byte(`{"id":"a","choices":[{"message":{"content":"","reasoning_content":"想一想"},"finish_reason":"length"}],"usage":{"prompt_tokens":240,"completion_tokens":256}}`))
+			return
+		}
+		_, _ = w.Write([]byte(`{"id":"b","choices":[{"message":{"content":"在的，刚忙完","reasoning_content":"想好了"},"finish_reason":"stop"}],"usage":{"prompt_tokens":240,"completion_tokens":40}}`))
+	}))
+	defer server.Close()
+	for _, statement := range []string{
+		`INSERT INTO system_configs(config_key,config_value,value_type,description) VALUES('ai.enabled','true','boolean',''),('ai.disclose_in_chat','false','boolean','') ON DUPLICATE KEY UPDATE config_value=VALUES(config_value)`,
+		`UPDATE users SET is_test=1,test_batch='regression' WHERE id=2`,
+		`INSERT INTO ai_agents(user_id,model_id,enabled) VALUES(2,1,1)`,
+	} {
+		if _, err := db.Exec(statement); err != nil {
+			t.Fatal(err)
+		}
+	}
+	cipher, err := a.encryptSecret("test-key")
+	if err != nil {
+		t.Fatal(err)
+	}
+	// max_tokens 256 is the setting that produced blank replies in production.
+	if _, err := db.Exec(`INSERT INTO ai_models(id,name,protocol,base_url,model_name,api_key_cipher,temperature,max_tokens,timeout_ms,status,is_default) VALUES(1,'reasoning','openai',?,'reasoning-model',?,0.80,256,8000,1,1)`, server.URL, cipher); err != nil {
+		t.Fatal(err)
+	}
+	var created struct {
+		ID int64 `json:"id"`
+	}
+	if err := json.Unmarshal(imTestCall(t, a.directConversation, 1, "POST", "/api/v1/im/conversations", `{"userId":2}`, 200), &created); err != nil {
+		t.Fatal(err)
+	}
+	if _, _, err := a.persistMessageContext(context.Background(), created.ID, 1, "trigger", 1, map[string]any{"text": "在吗"}); err != nil {
+		t.Fatal(err)
+	}
+	var job aiReplyJob
+	if err := db.QueryRow(`SELECT id,conversation_id,agent_user_id,attempts FROM ai_reply_jobs WHERE conversation_id=?`, created.ID).
+		Scan(&job.ID, &job.ConversationID, &job.AgentUserID, &job.Attempts); err != nil {
+		t.Fatal(err)
+	}
+	if err := a.processAIReplyJob(context.Background(), job); err != nil {
+		t.Fatalf("推理耗尽预算后应当重试成功: %v", err)
+	}
+	if len(budgets) != 2 || budgets[0] != 256 || budgets[1] != 1024 {
+		t.Fatalf("budgets = %v, want [256 1024]", budgets)
+	}
+	var text string
+	if err := db.QueryRow(`SELECT CAST(body AS CHAR) FROM im_messages WHERE conversation_id=? AND sender_id=2 ORDER BY id DESC LIMIT 1`, created.ID).Scan(&text); err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(text, "刚忙完") {
+		t.Fatalf("回复没有落库: %s", text)
+	}
+	// Both attempts are on the record, so the cost of the retry is visible.
+	var calls int
+	if err := db.QueryRow(`SELECT COUNT(*) FROM ai_call_logs WHERE conversation_id=?`, created.ID).Scan(&calls); err != nil {
+		t.Fatal(err)
+	}
+	if calls != 2 {
+		t.Fatalf("calls = %d, want 2", calls)
+	}
+}

im/backend/internal/app/ai_providers.go

@@ -36,12 +36,21 @@ type aiChatRequest struct {
 
 type aiChatResult struct {
 	Text             string
+	Reasoning        string
 	PromptTokens     int
 	CompletionTokens int
 	FinishReason     string
 	RequestID        string
 }
 
+// The model ran out of budget mid-thought. For a reasoning model that usually
+// means the whole ceiling went on the chain of thought and Text is empty; for
+// any model it can also mean the answer itself was cut off. Both deserve one
+// retry with room to finish rather than a blank or half-finished message.
+func (r aiChatResult) truncated() bool {
+	return r.FinishReason == "length" || (r.Text == "" && r.Reasoning != "")
+}
+
 type aiProvider interface {
 	Chat(ctx context.Context, req aiChatRequest) (aiChatResult, error)
 }
@@ -160,6 +169,20 @@ func aiRequestMaxTokens(req aiChatRequest, model aiModel) int {
 	return 256
 }
 
+// One retry, with enough room for the thinking plus a short answer, and a cap
+// so a misconfigured model cannot bill an unbounded completion.
+func aiRetryMaxTokens(current int) int {
+	const floor, ceiling = 1024, 4096
+	raised := current * 4
+	if raised < floor {
+		raised = floor
+	}
+	if raised > ceiling {
+		raised = ceiling
+	}
+	return raised
+}
+
 func aiPostJSON(ctx context.Context, client *http.Client, endpoint string, headers map[string]string, payload any) ([]byte, error) {
 	body, err := json.Marshal(payload)
 	if err != nil {
@@ -234,6 +257,11 @@ func (p openAIProvider) Chat(ctx context.Context, req aiChatRequest) (aiChatResu
 		Choices []struct {
 			Message struct {
 				Content string `json:"content"`
+				// Reasoning models put their chain of thought here and the answer
+				// in Content. Spending the whole budget on the former leaves the
+				// latter empty, which is indistinguishable from a broken model
+				// unless this field is read.
+				ReasoningContent string `json:"reasoning_content"`
 			} `json:"message"`
 			FinishReason string `json:"finish_reason"`
 		} `json:"choices"`
@@ -256,6 +284,7 @@ func (p openAIProvider) Chat(ctx context.Context, req aiChatRequest) (aiChatResu
 	}
 	return aiChatResult{
 		Text:             strings.TrimSpace(payload.Choices[0].Message.Content),
+		Reasoning:        strings.TrimSpace(payload.Choices[0].Message.ReasoningContent),
 		PromptTokens:     payload.Usage.PromptTokens,
 		CompletionTokens: payload.Usage.CompletionTokens,
 		FinishReason:     payload.Choices[0].FinishReason,

im/backend/internal/app/ai_providers_test.go

@@ -216,3 +216,107 @@ func TestDebugProviderIsDeterministicAndOffline(t *testing.T) {
 		t.Fatal("debug provider must answer even without history")
 	}
 }
+
+// A reasoning model answers in two parts. When the token ceiling is reached
+// before it stops thinking, the reply field is empty and the whole budget shows
+// up as completion tokens — exactly what the admin log showed as "成功" with a
+// blank reply.
+func TestReasoningModelExhaustingItsBudgetIsRecognised(t *testing.T) {
+	server, _, _ := captureAIRequest(t, `{"id":"call-1","choices":[{"message":{"content":"","reasoning_content":"先想想怎么回答"},"finish_reason":"length"}],"usage":{"prompt_tokens":240,"completion_tokens":256}}`, 200)
+	result, err := chatOnce(t, "openai", server.URL)
+	if err != nil {
+		t.Fatalf("空正文本身不是传输错误: %v", err)
+	}
+	if result.Text != "" || result.Reasoning != "先想想怎么回答" {
+		t.Fatalf("result = %+v", result)
+	}
+	if !result.truncated() {
+		t.Fatal("必须能识别出被截断")
+	}
+
+	// The ordinary case must not be mistaken for it.
+	plain, _, _ := captureAIRequest(t, `{"id":"call-2","choices":[{"message":{"content":"在的，刚吃过"},"finish_reason":"stop"}],"usage":{"prompt_tokens":240,"completion_tokens":62}}`, 200)
+	answered, err := chatOnce(t, "openai", plain.URL)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if answered.Text != "在的，刚吃过" || answered.truncated() {
+		t.Fatalf("result = %+v", answered)
+	}
+
+	// An empty reply with no reasoning and a normal stop is still just empty.
+	blank, _, _ := captureAIRequest(t, `{"id":"call-3","choices":[{"message":{"content":""},"finish_reason":"stop"}],"usage":{"prompt_tokens":10,"completion_tokens":0}}`, 200)
+	empty, err := chatOnce(t, "openai", blank.URL)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if empty.truncated() {
+		t.Fatal("普通空回复不该被当成截断")
+	}
+}
+
+func TestAIRetryMaxTokensStaysWithinBounds(t *testing.T) {
+	for _, test := range []struct{ current, want int }{
+		{256, 1024},  // the configured value that produced blank replies
+		{64, 1024},   // never retry with less than a usable budget
+		{512, 2048},  // four times when that is already sensible
+		{2048, 4096}, // capped, so a misconfiguration cannot bill without limit
+		{8192, 4096},
+	} {
+		if got := aiRetryMaxTokens(test.current); got != test.want {
+			t.Fatalf("aiRetryMaxTokens(%d) = %d, want %d", test.current, got, test.want)
+		}
+	}
+}
+
+// The request the retry sends has to actually carry the raised ceiling.
+func TestChatRequestMaxTokensOverridesTheModelConfiguration(t *testing.T) {
+	server, body, _ := captureAIRequest(t, `{"id":"c","choices":[{"message":{"content":"好"},"finish_reason":"stop"}],"usage":{}}`, 200)
+	provider, err := newAIProvider(testModel("openai", server.URL), false)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err = provider.Chat(context.Background(), aiChatRequest{
+		Messages:  []aiChatMessage{{Role: "user", Text: "在吗"}},
+		MaxTokens: aiRetryMaxTokens(128),
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if got := (*body)["max_tokens"]; got != float64(1024) {
+		t.Fatalf("max_tokens = %v, want 1024", got)
+	}
+}
+
+// A reply that opens with a number used to arrive with the number missing: the
+// list-marker trimming removed the digits themselves.
+func TestSanitizeAIReplyKeepsLeadingNumbers(t *testing.T) {
+	for _, test := range []struct{ raw, want string }{
+		{"30岁啦，你呢", "30岁啦，你呢"},
+		{"1. 先自我介绍", "先自我介绍"},
+		{"2、我在深圳", "我在深圳"},
+		{"- 喜欢跑步", "喜欢跑步"},
+		{"• 也喜欢音乐", "也喜欢音乐"},
+		{"**在的**", "在的"},
+		{"18612725546 是我的号码", "18612725546 是我的号码"},
+		{"3 个月前搬来的", "3 个月前搬来的"},
+	} {
+		if got := sanitizeAIReply(test.raw); got != test.want {
+			t.Fatalf("sanitizeAIReply(%q) = %q, want %q", test.raw, got, test.want)
+		}
+	}
+}
+
+// A cut-off answer is worth one more attempt too, not just an empty one.
+func TestTruncatedAnswerIsRetriedAndKeptIfItSurvives(t *testing.T) {
+	cut, _, _ := captureAIRequest(t, `{"id":"c","choices":[{"message":{"content":"30岁啦，"},"finish_reason":"length"}],"usage":{"prompt_tokens":240,"completion_tokens":1024}}`, 200)
+	result, err := chatOnce(t, "openai", cut.URL)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.truncated() {
+		t.Fatal("被截断的正文也要触发重试")
+	}
+	if sanitizeAIReply(result.Text) != "30岁啦，" {
+		t.Fatalf("text = %q", result.Text)
+	}
+}