Add files via upload

2026-08-09 20:54:31 +02:00 · 2026-07-11 10:47:04 +08:00
parent bec296ae3f
commit 2fcda3a57d
87 changed files with 11962 additions and 1 deletions
@@ -0,0 +1,683 @@
+package multiagent
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"cyberstrike-ai/internal/agent"
+	"cyberstrike-ai/internal/config"
+	"cyberstrike-ai/internal/database"
+	copenai "cyberstrike-ai/internal/openai"
+	"cyberstrike-ai/internal/project"
+
+	"github.com/bytedance/sonic"
+	einoopenai "github.com/cloudwego/eino-ext/components/model/openai"
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/adk/middlewares/summarization"
+	"github.com/cloudwego/eino/components/model"
+	"github.com/cloudwego/eino/schema"
+	"go.uber.org/zap"
+)
+
+// einoSummarizeUserInstruction：压缩历史时保留渗透测试与用户约束关键信息。
+// 结构对齐 Eino 最佳实践（禁止工具、<analysis>+<summary>、<all_user_messages>），章节为安全测试领域化。
+const einoSummarizeUserInstruction = `关键：仅以纯文本响应。禁止调用任何工具（read_file、exec、grep、glob、write、edit 等）。
+上述对话中已包含全部待压缩上下文；不要要求用户粘贴历史，不要输出「请提供待压缩的对话历史」等占位/meta 回复。
+工具调用将被拒绝并浪费唯一一次摘要机会。
+
+你的任务：在保持所有关键安全测试信息完整的前提下压缩对话历史，使后续代理能无缝继续同一授权测试任务。
+
+压缩原则：
+- 必须保留：已确认漏洞与攻击路径、工具输出核心发现、凭证与认证细节、架构与薄弱点、当前进度、失败尝试与死路、策略决策
+- 保留精确技术细节（URL、路径、参数、Payload、版本号；报错原文可摘要但要点不丢）
+- 冗长扫描输出概括为结论；重复发现合并表述
+- 已枚举资产须保留可继承摘要：主域、关键子域/主机短表（或数量+代表样例）、高价值目标、已识别服务/端口要点
+
+输出格式（严格遵循，仅一轮回复）：
+1. 先输出 <analysis> 块：按时间顺序梳理对话，检查是否涵盖下方各章节要点；analysis 仅供自检，保持简洁（建议 ≤400 字）
+2. 再输出 <summary> 块：按以下章节写入可继承的压缩报告（无信息处写「无」，禁止留空模板占位符）
+
+<summary>
+## 1. 授权范围与约束
+- 目标/范围/禁止项（域名、路径、IP、环境）
+- 凭证/认证信息（账号、Token、Cookie；敏感值原文保留）
+- 用户指定的方法、工具、优先级与待办
+- 否定约束（不测什么、不用什么手法）
+
+## 2. 资产与服务枚举摘要
+- 主域/核心资产、关键子域或主机短表（或数量+代表样例）
+- 高价值目标、已识别服务/端口要点
+- 资产状态（存活/可攻/已排除/待验证）
+
+## 3. 架构与已知薄弱点
+- 技术栈/部署拓扑/信任边界
+- 已识别薄弱点列表
+
+## 4. 已确认漏洞与攻击路径
+- 漏洞名/CVE、URL/路径、参数/Payload、PoC 要点、影响等级
+- 攻击链/利用路径（步骤化）
+
+## 5. 工具核心发现与扫描结论
+- 各工具结论（概括核心输出，非冗长日志）
+- 重复发现合并表述
+
+## 6. 所有用户消息
+<all_user_messages>
+- [逐条列出非 tool 结果的用户消息要点；敏感约束与原文措辞尽量保留]
+</all_user_messages>
+
+## 7. 当前进度、策略决策与下一步
+- 当前位置（已完成/进行中/卡点）
+- 失败尝试与死路（方法、现象/报错摘要、结论）
+- 策略决策与下一步具体操作（须与最近用户请求及未完成任务一致）
+</summary>
+
+提醒：不要调用任何工具；必须基于上文已有对话直接输出 <analysis> 与 <summary>，勿输出 analysis 以外的正文。`
+
+// newEinoSummarizationMiddleware 使用 Eino ADK Summarization 中间件（见 https://www.cloudwego.io/zh/docs/eino/core_modules/eino_adk/eino_adk_chatmodelagentmiddleware/middleware_summarization/）。
+// 触发阈值：估算 token 超过 openai.max_total_tokens * summarization_trigger_ratio（默认 0.8）时摘要。
+func newEinoSummarizationMiddleware(
+	ctx context.Context,
+	summaryModel model.BaseChatModel,
+	appCfg *config.Config,
+	mwCfg *config.MultiAgentEinoMiddlewareConfig,
+	conversationID string,
+	db *database.DB,
+	projectID string,
+	logger *zap.Logger,
+) (adk.ChatModelAgentMiddleware, error) {
+	if summaryModel == nil || appCfg == nil {
+		return nil, fmt.Errorf("multiagent: summarization 需要 model 与配置")
+	}
+	maxTotal := appCfg.OpenAI.MaxTotalTokens
+	if maxTotal <= 0 {
+		maxTotal = 120000
+	}
+	triggerRatio := 0.8
+	emitInternalEvents := true
+	outputReserve := config.DefaultSummarizationOutputReserveTokens
+	userLedgerMaxRunes := config.DefaultSummarizationUserIntentLedgerMaxRunes
+	userLedgerEntryMaxRunes := config.DefaultSummarizationUserIntentLedgerEntryMaxRunes
+	toolMaxBytes := config.MultiAgentEinoMiddlewareConfig{}.ReductionMaxLengthForTruncEffective()
+	if mwCfg != nil {
+		triggerRatio = mwCfg.SummarizationTriggerRatioEffective()
+		emitInternalEvents = mwCfg.SummarizationEmitInternalEventsEffective()
+		outputReserve = mwCfg.SummarizationOutputReserveTokensEffective()
+		userLedgerMaxRunes = mwCfg.SummarizationUserIntentLedgerMaxRunesEffective()
+		userLedgerEntryMaxRunes = mwCfg.SummarizationUserIntentLedgerEntryMaxRunesEffective()
+		toolMaxBytes = mwCfg.ReductionMaxLengthForTruncEffective()
+	}
+	// The ledger is merged into the leading system message and cannot be removed as
+	// an ordinary conversation round. Bound it relative to the configured window so
+	// it cannot crowd out the summary/latest turn.
+	ledgerWindowCap := modelFacingRuneBudget(maxTotal, 0.20)
+	userLedgerMaxRunes = minPositiveInt(userLedgerMaxRunes, ledgerWindowCap)
+	userLedgerEntryMaxRunes = minPositiveInt(userLedgerEntryMaxRunes, userLedgerMaxRunes)
+	// Keep enough safety margin for tokenizer/model-side accounting mismatch.
+	trigger := int(float64(maxTotal) * triggerRatio)
+	if trigger < 4096 {
+		trigger = maxTotal
+		if trigger < 4096 {
+			trigger = 4096
+		}
+	}
+	preserveMax := trigger / 3
+	if preserveMax < 2048 {
+		preserveMax = 2048
+	}
+
+	modelName := strings.TrimSpace(appCfg.OpenAI.Model)
+	if modelName == "" {
+		modelName = "gpt-4o"
+	}
+	tokenCounter := einoSummarizationTokenCounter(modelName)
+	recentTrailMax := trigger / 4
+	if recentTrailMax < 2048 {
+		recentTrailMax = 2048
+	}
+	if recentTrailMax > trigger/2 {
+		recentTrailMax = trigger / 2
+	}
+	// Summarization input aligns with the trigger threshold, minus explicit output reserve.
+	summaryInputMax := trigger - outputReserve
+	if summaryInputMax < 4096 {
+		summaryInputMax = trigger * 80 / 100
+	}
+	if summaryInputMax < 4096 {
+		summaryInputMax = 4096
+	}
+	transcriptPath := ""
+	if conv := strings.TrimSpace(conversationID); conv != "" {
+		baseRoot := filepath.Join(os.TempDir(), "cyberstrike-summarization")
+		if dbPath := strings.TrimSpace(appCfg.Database.Path); dbPath != "" {
+			// Persist with the same lifecycle as local conversation storage.
+			baseRoot = filepath.Join(filepath.Dir(dbPath), "conversation_artifacts", sanitizeEinoPathSegment(conv), "summarization")
+		}
+		base := baseRoot
+		if mkErr := os.MkdirAll(base, 0o755); mkErr == nil {
+			transcriptPath = filepath.Join(base, "transcript.txt")
+		}
+	}
+
+	retryPolicy := einoTransientRunRetryPolicyFromMW(mwCfg)
+	retryMax := retryPolicy.maxAttempts
+	var summaryOverflowRetries int
+
+	// ModelOptions apply only to summarization Generate (same ChatModel instance as the agent).
+	// Strip thinking/reasoning on this call path; mark requests for empty-choices diagnostics.
+	summaryModelOpts := []model.Option{
+		einoopenai.WithExtraHeader(map[string]string{
+			copenai.SummarizationRequestHeader: "1",
+		}),
+		einoopenai.WithRequestPayloadModifier(func(_ context.Context, in []*schema.Message, rawBody []byte) ([]byte, error) {
+			if logger != nil {
+				logger.Info("eino summarization generate request",
+					zap.Int("input_messages", len(in)),
+					zap.Int("payload_bytes", len(rawBody)),
+					zap.String("model", modelName),
+				)
+			}
+			return stripReasoningFromSummarizationPayload(rawBody)
+		}),
+	}
+
+	mw, err := summarization.New(ctx, &summarization.Config{
+		Model:        summaryModel,
+		ModelOptions: summaryModelOpts,
+		GenModelInput: func(ctx context.Context, sysInstruction, userInstruction adk.Message, originalMsgs []adk.Message) ([]adk.Message, error) {
+			if transcriptPath != "" && len(originalMsgs) > 0 {
+				if werr := writeSummarizationTranscript(transcriptPath, originalMsgs); werr != nil && logger != nil {
+					logger.Warn("eino summarization transcript preflight 写入失败",
+						zap.String("path", transcriptPath), zap.Error(werr))
+				}
+			}
+			budget := summaryInputMax
+			aggressive := summaryOverflowRetries > 0
+			if aggressive {
+				budget = summaryInputMax * 70 / 100
+				if budget < 4096 {
+					budget = 4096
+				}
+			}
+			input, dropped, berr := buildBudgetedSummarizationModelInput(
+				ctx, sysInstruction, userInstruction, originalMsgs, tokenCounter, budget,
+				summarizationInputBudgetOpts{
+					toolMaxBytes: toolMaxBytes,
+					spillRef:     transcriptPath,
+					aggressive:   aggressive,
+				},
+			)
+			if logger != nil && (berr != nil || dropped > 0 || aggressive) {
+				fields := []zap.Field{
+					zap.Int("max_input_tokens", budget),
+					zap.Int("trigger_context_tokens", trigger),
+					zap.Int("output_reserve_tokens", outputReserve),
+					zap.Int("dropped_rounds", dropped),
+					zap.Bool("aggressive", aggressive),
+				}
+				if berr != nil {
+					fields = append(fields, zap.Error(berr))
+					logger.Warn("eino summarization input budget failed", fields...)
+				} else {
+					logger.Info("eino summarization input bounded", fields...)
+				}
+			}
+			return input, berr
+		},
+		Trigger: &summarization.TriggerCondition{
+			ContextTokens: trigger,
+		},
+		TokenCounter:       tokenCounter,
+		UserInstruction:    einoSummarizeUserInstruction,
+		EmitInternalEvents: emitInternalEvents,
+		TranscriptFilePath: transcriptPath,
+		PreserveUserMessages: &summarization.PreserveUserMessages{
+			Enabled:   true,
+			MaxTokens: preserveMax,
+		},
+		Retry: &summarization.RetryConfig{
+			MaxRetries: &retryMax,
+			ShouldRetry: func(_ context.Context, _ adk.Message, err error) bool {
+				if isEinoContextOverflowError(err) && summaryOverflowRetries < 1 {
+					summaryOverflowRetries++
+					if logger != nil {
+						logger.Warn("eino summarization context overflow, retrying with aggressive compaction",
+							zap.Error(err),
+						)
+					}
+					return true
+				}
+				retry := isEinoTransientRunError(err)
+				if retry && logger != nil {
+					logger.Warn("eino summarization generate transient error, will retry if attempts remain",
+						zap.Error(err),
+						zap.Int("max_retries", retryMax),
+					)
+				}
+				return retry
+			},
+		},
+		Finalize: func(ctx context.Context, originalMessages []adk.Message, summary adk.Message) ([]adk.Message, error) {
+			summary = stripAnalysisFromSummarizationMessage(summary)
+			userLedger := buildOriginalUserIntentLedgerMessage(originalMessages, userLedgerMaxRunes, userLedgerEntryMaxRunes)
+			compactionMessages := stripOriginalUserIntentLedgerFromMessages(originalMessages)
+			out, ferr := summarizeFinalizeWithRecentAssistantToolTrail(ctx, compactionMessages, summary, tokenCounter, recentTrailMax)
+			if ferr != nil {
+				return nil, ferr
+			}
+			out = mergeMessageIntoLeadingSystem(out, userLedger)
+			if appCfg != nil {
+				out = refreshFactIndexInMessages(out, db, projectID, appCfg.Project, logger)
+			}
+			return out, nil
+		},
+		Callback: func(ctx context.Context, before, after adk.ChatModelAgentState) error {
+			if transcriptPath != "" && len(before.Messages) > 0 {
+				if werr := writeSummarizationTranscript(transcriptPath, before.Messages); werr != nil && logger != nil {
+					logger.Warn("eino summarization transcript 写入失败",
+						zap.String("path", transcriptPath),
+						zap.Error(werr),
+					)
+				}
+			}
+			if logger != nil {
+				beforeTokens, _ := tokenCounter(ctx, &summarization.TokenCounterInput{Messages: before.Messages})
+				afterTokens, _ := tokenCounter(ctx, &summarization.TokenCounterInput{Messages: after.Messages})
+				logger.Info("eino summarization 已压缩上下文",
+					zap.Int("messages_before", len(before.Messages)),
+					zap.Int("messages_after", len(after.Messages)),
+					zap.Int("tokens_before_estimated", beforeTokens),
+					zap.Int("tokens_after_estimated", afterTokens),
+					zap.Int("max_total_tokens", maxTotal),
+					zap.Int("trigger_context_tokens", trigger),
+					zap.String("transcript_file", transcriptPath),
+				)
+			}
+			return nil
+		},
+	})
+	if err != nil {
+		return nil, fmt.Errorf("summarization.New: %w", err)
+	}
+	return mw, nil
+}
+
+// summarizationInputBudgetOpts controls spill/truncation behavior when a round alone exceeds budget.
+type summarizationInputBudgetOpts struct {
+	toolMaxBytes int
+	spillRef     string
+	aggressive   bool
+}
+
+// buildBudgetedSummarizationModelInput builds the exact payload sent to the summary model.
+// It retains the newest complete conversation rounds within budget and emits an explicit
+// marker when older rounds are omitted. The full pre-compaction transcript is persisted
+// separately; omitted raw messages never become model-facing history again.
+func buildBudgetedSummarizationModelInput(
+	ctx context.Context,
+	sysInstruction adk.Message,
+	userInstruction adk.Message,
+	originalMsgs []adk.Message,
+	tokenCounter summarization.TokenCounterFunc,
+	maxTokens int,
+	opts summarizationInputBudgetOpts,
+) ([]adk.Message, int, error) {
+	base := []adk.Message{sysInstruction, userInstruction}
+	baseTokens, err := tokenCounter(ctx, &summarization.TokenCounterInput{Messages: base})
+	if err != nil {
+		return nil, 0, err
+	}
+	remaining := maxTokens - baseTokens
+	markerTemplate := schema.UserMessage("[Context budget guard omitted older conversation rounds; summarize the retained recent rounds and preserve the omission marker.]")
+	markerTokens, err := tokenCounter(ctx, &summarization.TokenCounterInput{Messages: []adk.Message{markerTemplate}})
+	if err != nil {
+		return nil, 0, err
+	}
+	remaining -= markerTokens
+	if remaining < 0 {
+		remaining = 0
+	}
+
+	contextMsgs := make([]adk.Message, 0, len(originalMsgs))
+	for _, msg := range originalMsgs {
+		if msg != nil && msg.Role != schema.System {
+			contextMsgs = append(contextMsgs, msg)
+		}
+	}
+	rounds := splitMessagesIntoRounds(contextMsgs)
+	selectedReverse := make([]messageRound, 0, len(rounds))
+	used := 0
+	toolMaxBytes := opts.toolMaxBytes
+	if toolMaxBytes <= 0 {
+		toolMaxBytes = 12000
+	}
+	if opts.aggressive {
+		toolMaxBytes /= aggressiveToolTruncDivisor
+		if toolMaxBytes < 2048 {
+			toolMaxBytes = 2048
+		}
+	}
+	for i := len(rounds) - 1; i >= 0; i-- {
+		n, countErr := tokenCounter(ctx, &summarization.TokenCounterInput{Messages: rounds[i].messages})
+		if countErr != nil {
+			return nil, 0, countErr
+		}
+		if used+n > remaining {
+			if len(selectedReverse) == 0 {
+				slot := remaining - used
+				if slot > 0 {
+					truncated, truncErr := truncateRoundMessagesToTokenBudget(
+						ctx, rounds[i], slot, tokenCounter, toolMaxBytes, opts.spillRef,
+					)
+					if truncErr != nil {
+						return nil, 0, truncErr
+					}
+					if len(truncated) > 0 {
+						selectedReverse = append(selectedReverse, messageRound{messages: truncated})
+					}
+				}
+			}
+			break
+		}
+		used += n
+		selectedReverse = append(selectedReverse, rounds[i])
+	}
+
+	dropped := len(rounds) - len(selectedReverse)
+	input := make([]adk.Message, 0, 3+len(contextMsgs))
+	input = append(input, sysInstruction)
+	if dropped > 0 {
+		input = append(input, schema.UserMessage(fmt.Sprintf(
+			"[Context budget guard omitted %d older conversation round(s); summarize the retained recent rounds and preserve the omission marker.]",
+			dropped,
+		)))
+	}
+	for i := len(selectedReverse) - 1; i >= 0; i-- {
+		input = append(input, selectedReverse[i].messages...)
+	}
+	input = append(input, userInstruction)
+	return input, dropped, nil
+}
+
+// refreshFactIndexInMessages 在 summarization 压缩后，用 DB 最新索引替换 system 中已有的项目黑板索引段。
+func refreshFactIndexInMessages(msgs []adk.Message, db *database.DB, projectID string, cfg config.ProjectConfig, logger *zap.Logger) []adk.Message {
+	if db == nil || !cfg.Enabled {
+		return msgs
+	}
+	projectID = strings.TrimSpace(projectID)
+	if projectID == "" {
+		return msgs
+	}
+	freshIndex, err := project.BuildFactIndexBlock(db, projectID, cfg)
+	if err != nil {
+		if logger != nil {
+			logger.Warn("summarization: 刷新项目黑板索引失败", zap.String("projectId", projectID), zap.Error(err))
+		}
+		return msgs
+	}
+	freshIndex = strings.TrimSpace(freshIndex)
+	if freshIndex == "" {
+		return msgs
+	}
+
+	changed := false
+	out := make([]adk.Message, len(msgs))
+	for i, msg := range msgs {
+		if msg == nil || msg.Role != schema.System {
+			out[i] = msg
+			continue
+		}
+		newContent, ok := project.ReplaceFactIndexSection(msg.Content, freshIndex)
+		if !ok {
+			out[i] = msg
+			continue
+		}
+		cloned := *msg
+		cloned.Content = newContent
+		out[i] = &cloned
+		changed = true
+	}
+	if changed && logger != nil {
+		logger.Info("summarization: 已刷新项目黑板索引", zap.String("projectId", projectID))
+	}
+	return out
+}
+
+// summarizeFinalizeWithRecentAssistantToolTrail 在摘要消息后保留最近 assistant/tool 轨迹，避免压缩后执行链断裂。
+//
+// 关键不变量：tool_call ↔ tool_result 的 pair 必须整体保留或整体丢弃。
+// 把消息切成 round（回合）为原子单位：
+//   - user(...) 单条为一个 round；
+//   - assistant(tool_calls=[...]) 及其后连续的 role=tool 消息合成一个 round；
+//   - 其它 assistant(reply, 无 tool_calls) 单条为一个 round。
+//
+// 倒序挑 round（预算不够即放弃该 round），保证 tool 消息不会跨 round 被孤立。
+func summarizeFinalizeWithRecentAssistantToolTrail(
+	ctx context.Context,
+	originalMessages []adk.Message,
+	summary adk.Message,
+	tokenCounter summarization.TokenCounterFunc,
+	recentTrailTokenBudget int,
+) ([]adk.Message, error) {
+	systemMsgs := make([]adk.Message, 0, len(originalMessages))
+	nonSystem := make([]adk.Message, 0, len(originalMessages))
+	for _, msg := range originalMessages {
+		if msg == nil {
+			continue
+		}
+		if msg.Role == schema.System {
+			systemMsgs = append(systemMsgs, msg)
+			continue
+		}
+		nonSystem = append(nonSystem, msg)
+	}
+
+	mergedSystem := mergeCollectedSystemMessages(systemMsgs)
+
+	if recentTrailTokenBudget <= 0 || len(nonSystem) == 0 {
+		out := make([]adk.Message, 0, len(mergedSystem)+1)
+		out = append(out, mergedSystem...)
+		out = append(out, summary)
+		return out, nil
+	}
+
+	rounds := splitMessagesIntoRounds(nonSystem)
+	if len(rounds) == 0 {
+		out := make([]adk.Message, 0, len(mergedSystem)+1)
+		out = append(out, mergedSystem...)
+		out = append(out, summary)
+		return out, nil
+	}
+
+	// 目标：至少保留 minRounds 个 round 的执行轨迹；在预算允许时尽量多保留。
+	// 优先确保最后一个 round（通常是最新的 tool 往返或 assistant 回复）存在。
+	const minRounds = 2
+
+	selectedRoundsReverse := make([]messageRound, 0, 8)
+	selectedCount := 0
+	totalTokens := 0
+
+	tokensOfRound := func(r messageRound) (int, error) {
+		if len(r.messages) == 0 {
+			return 0, nil
+		}
+		n, err := tokenCounter(ctx, &summarization.TokenCounterInput{Messages: r.messages})
+		if err != nil {
+			return 0, err
+		}
+		if n <= 0 {
+			n = len(r.messages)
+		}
+		return n, nil
+	}
+
+	for i := len(rounds) - 1; i >= 0; i-- {
+		r := rounds[i]
+		n, err := tokensOfRound(r)
+		if err != nil {
+			return nil, err
+		}
+		// 预算不够：已经保留了足够 round 则停，否则跳过该 round 继续往前找
+		// （避免一个超大 round 挤占全部预算，至少保证有轨迹）。
+		if totalTokens+n > recentTrailTokenBudget {
+			if selectedCount >= minRounds {
+				break
+			}
+			continue
+		}
+		totalTokens += n
+		selectedRoundsReverse = append(selectedRoundsReverse, r)
+		selectedCount++
+	}
+
+	// 还原时间顺序。round 内为原始 *schema.Message 指针，保留 ReasoningContent（DeepSeek 工具续跑所必需）。
+	selectedMsgs := make([]adk.Message, 0, 8)
+	for i := len(selectedRoundsReverse) - 1; i >= 0; i-- {
+		selectedMsgs = append(selectedMsgs, selectedRoundsReverse[i].messages...)
+	}
+
+	out := make([]adk.Message, 0, len(mergedSystem)+1+len(selectedMsgs))
+	out = append(out, mergedSystem...)
+	out = append(out, summary)
+	out = append(out, selectedMsgs...)
+	return out, nil
+}
+
+// messageRound 表示一个"不可分割"的消息回合。
+//   - 对 assistant(tool_calls) + 随后若干 tool 消息的组合，round 内全部 call_id 成对完整；
+//   - 对独立的 user / assistant(reply) 消息，round 仅包含该条消息。
+type messageRound struct {
+	messages []adk.Message
+}
+
+// splitMessagesIntoRounds 将非 system 消息切分为若干 round，保证：
+//   - 每个 assistant(tool_calls) 与其对应的 role=tool 响应消息在同一个 round；
+//   - 孤立（无对应 assistant(tool_calls)）的 role=tool 消息不会单独成为 round，
+//     而是被丢弃（这些消息在 pair 完整性层面已属孤儿，保留反而会触发 LLM 400）。
+func splitMessagesIntoRounds(msgs []adk.Message) []messageRound {
+	if len(msgs) == 0 {
+		return nil
+	}
+	rounds := make([]messageRound, 0, len(msgs))
+	i := 0
+	for i < len(msgs) {
+		msg := msgs[i]
+		if msg == nil {
+			i++
+			continue
+		}
+		switch {
+		case msg.Role == schema.Assistant && len(msg.ToolCalls) > 0:
+			// 收集该 assistant 提供的 call_id 集合。
+			provided := make(map[string]struct{}, len(msg.ToolCalls))
+			for _, tc := range msg.ToolCalls {
+				if tc.ID != "" {
+					provided[tc.ID] = struct{}{}
+				}
+			}
+			round := messageRound{messages: []adk.Message{msg}}
+			j := i + 1
+			for j < len(msgs) {
+				next := msgs[j]
+				if next == nil {
+					j++
+					continue
+				}
+				if next.Role != schema.Tool {
+					break
+				}
+				if next.ToolCallID != "" {
+					if _, ok := provided[next.ToolCallID]; !ok {
+						// 下一条 tool 不属于当前 assistant，认为当前 round 结束。
+						break
+					}
+				}
+				round.messages = append(round.messages, next)
+				j++
+			}
+			rounds = append(rounds, round)
+			i = j
+		case msg.Role == schema.Tool:
+			// 孤儿 tool 消息：既不跟随在一个 assistant(tool_calls) 后，
+			// 说明它对应的 assistant 已被上游裁剪；直接丢弃，下一步到 orphan pruner
+			// 兜底也不会出错，但在 round 切分这里就剔除更干净。
+			i++
+		default:
+			// user / assistant(reply) / 其它：单条成 round。
+			rounds = append(rounds, messageRound{messages: []adk.Message{msg}})
+			i++
+		}
+	}
+	return rounds
+}
+
+// writeSummarizationTranscript persists pre-compaction history for read_file after summarization.
+// Eino TranscriptFilePath only embeds the path in summary text; the file must be written by the host app.
+func writeSummarizationTranscript(path string, msgs []adk.Message) error {
+	path = strings.TrimSpace(path)
+	if path == "" {
+		return nil
+	}
+	body := formatSummarizationTranscript(msgs)
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		return fmt.Errorf("mkdir transcript dir: %w", err)
+	}
+	if err := os.WriteFile(path, []byte(body), 0o600); err != nil {
+		return fmt.Errorf("write transcript: %w", err)
+	}
+	return nil
+}
+
+func einoSummarizationTokenCounter(openAIModel string) summarization.TokenCounterFunc {
+	tc := agent.NewTikTokenCounter()
+	return func(ctx context.Context, input *summarization.TokenCounterInput) (int, error) {
+		var sb strings.Builder
+		for _, msg := range input.Messages {
+			if msg == nil {
+				continue
+			}
+			sb.WriteString(string(msg.Role))
+			sb.WriteByte('\n')
+			if msg.Content != "" {
+				sb.WriteString(msg.Content)
+				sb.WriteByte('\n')
+			}
+			if msg.ReasoningContent != "" {
+				sb.WriteString(msg.ReasoningContent)
+				sb.WriteByte('\n')
+			}
+			if len(msg.ToolCalls) > 0 {
+				if b, err := sonic.Marshal(msg.ToolCalls); err == nil {
+					sb.Write(b)
+					sb.WriteByte('\n')
+				}
+			}
+			for _, part := range msg.UserInputMultiContent {
+				if part.Type == schema.ChatMessagePartTypeText && part.Text != "" {
+					sb.WriteString(part.Text)
+					sb.WriteByte('\n')
+				}
+			}
+		}
+		for _, tl := range input.Tools {
+			if tl == nil {
+				continue
+			}
+			cp := *tl
+			cp.Extra = nil
+			if text, err := sonic.MarshalString(cp); err == nil {
+				sb.WriteString(text)
+				sb.WriteByte('\n')
+			}
+		}
+		text := sb.String()
+		n, err := tc.Count(openAIModel, text)
+		if err != nil {
+			return (len(text) + 3) / 4, nil
+		}
+		return n, nil
+	}
+}