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>
This commit is contained in:
Your Name
2026-09-04 10:50:25 +08:00
co-authored by Claude Opus 5
parent acd8933dbb
commit 812d8032e1
4 changed files with 235 additions and 1 deletions
+24 -1
View File
@@ -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)
}