Version: 0.9.53.dev.260429

后端：
1. 流式思考链路从 raw reasoning_content 切到 `thinking_summary` 摘要协议，补齐摘要 prompt、digestor 与 Lite 压缩链路，plan / execute / fallback 统一改为“只出摘要、不透原始推理”，正文开始后自动关停摘要流。
2. thinking_summary 打通 timeline / SSE / outbox 持久化闭环，只落 detail_summary 与必要 metadata，并补强 seq 自检、冲突幂等识别与补 seq 回填，提升重放恢复稳定性。
3. 会话历史口径继续收紧，assistant 正文与时间线不再回写 raw reasoning_content，仅保留正文与思考耗时，避免刷新恢复时再次暴露内部推理文本。

前端：
4. 助手页开始接入 thinking_summary 实时流与历史恢复，补齐短摘要状态、长摘要折叠区、正文开流后自动收口，并增加调试入口用于协议联调与验收。
5. 当前前端助手页仍是残次过渡态，本版先以 thinking_summary 协议接通和基础渲染为主，样式、交互与细节体验暂未收平，下一版集中修复。

仓库：
6. 补充 thinking_summary 对接说明，明确 SSE 协议、timeline 恢复口径与 short/detail summary 的使用边界。

backend/service/agentsvc/reasoning_summary.go

@@ -0,0 +1,112 @@
+package agentsvc
+
+import (
+	"context"
+	"errors"
+	"log"
+	"strings"
+
+	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
+	newagentprompt "github.com/LoveLosita/smartflow/backend/newAgent/prompt"
+	newagentstream "github.com/LoveLosita/smartflow/backend/newAgent/stream"
+)
+
+const reasoningSummaryMaxTokens = 700
+
+type reasoningSummaryLLMResponse struct {
+	ShortSummary  string `json:"short_summary"`
+	DetailSummary string `json:"detail_summary"`
+}
+
+// makeReasoningSummaryFunc 把便宜模型封装成 stream 层可注入的摘要函数。
+//
+// 职责边界：
+// 1. service 层负责选择模型与 prompt，stream 层只负责调度和闸门；
+// 2. 这里不持久化摘要，持久化统一走 ChunkEmitter 的 extra hook；
+// 3. 摘要失败时返回 error，由 ReasoningDigestor 吞掉并等待下一次水位线/Flush 兜底。
+func (s *AgentService) makeReasoningSummaryFunc(client *infrallm.Client) newagentstream.ReasoningSummaryFunc {
+	if client == nil {
+		return nil
+	}
+
+	return func(ctx context.Context, input newagentstream.ReasoningSummaryInput) (newagentstream.StreamThinkingSummaryExtra, error) {
+		previousSummary := ""
+		if input.PreviousSummary != nil {
+			previousSummary = input.PreviousSummary.DetailSummary
+			if strings.TrimSpace(previousSummary) == "" {
+				previousSummary = input.PreviousSummary.ShortSummary
+			}
+		}
+
+		messages := newagentprompt.BuildReasoningSummaryMessages(newagentprompt.ReasoningSummaryPromptInput{
+			FullReasoning:   input.FullReasoning,
+			DeltaReasoning:  input.DeltaReasoning,
+			PreviousSummary: previousSummary,
+			CandidateSeq:    input.CandidateSeq,
+			Final:           input.Final,
+			DurationSeconds: input.DurationSeconds,
+		})
+
+		resp, rawResult, err := infrallm.GenerateJSON[reasoningSummaryLLMResponse](
+			ctx,
+			client,
+			messages,
+			infrallm.GenerateOptions{
+				Temperature: 0.1,
+				MaxTokens:   reasoningSummaryMaxTokens,
+				Thinking:    infrallm.ThinkingModeDisabled,
+				Metadata: map[string]any{
+					"stage":         "reasoning_summary",
+					"candidate_seq": input.CandidateSeq,
+					"final":         input.Final,
+				},
+			},
+		)
+		if err != nil {
+			log.Printf("[WARN] reasoning 摘要模型调用失败 seq=%d final=%v err=%v raw=%s",
+				input.CandidateSeq,
+				input.Final,
+				err,
+				truncateReasoningSummaryRaw(rawResult),
+			)
+			return newagentstream.StreamThinkingSummaryExtra{}, err
+		}
+
+		summary := newagentstream.StreamThinkingSummaryExtra{
+			ShortSummary: strings.TrimSpace(resp.ShortSummary),
+			DetailSummary: limitReasoningDetailSummary(
+				resp.DetailSummary,
+				newagentprompt.ReasoningSummaryDetailRuneLimit(input.FullReasoning, input.DeltaReasoning),
+			),
+		}
+		if summary.ShortSummary == "" && summary.DetailSummary == "" {
+			return newagentstream.StreamThinkingSummaryExtra{}, errors.New("reasoning 摘要模型返回空摘要")
+		}
+		return summary, nil
+	}
+}
+
+func limitReasoningDetailSummary(text string, maxRunes int) string {
+	text = strings.TrimSpace(text)
+	if text == "" || maxRunes <= 0 {
+		return text
+	}
+
+	runes := []rune(text)
+	if len(runes) <= maxRunes {
+		return text
+	}
+	return string(runes[:maxRunes])
+}
+
+func truncateReasoningSummaryRaw(raw *infrallm.TextResult) string {
+	if raw == nil {
+		return ""
+	}
+	text := strings.TrimSpace(raw.Text)
+	runes := []rune(text)
+	if len(runes) <= 200 {
+		return text
+	}
+	return string(runes[:200]) + "..."
+}