Agent-System-Backend/internal/runtime/llm.go

package runtime

import (
	"encoding/json"
	"fmt"
	"strings"

	agentllm "agent-native/runtime/llm"
	core "agent-system-core"
	"agent-system-backend/internal/ctxasm"
	"agent-system-backend/internal/squadadapter"
)

// LLMRunner 主对话执行器：复用 agent-native 的 OpenAI 兼容流式客户端（不自研接入层）。
// 多轮循环：流式 → 工具调用（画像策略）→ 回灌结果 → 直至无工具调用或预算耗尽。
// 内置三个编排型工具（不经注册表）：delegate_task（子代理）、update_plan、update_todos。
type LLMRunner struct{}

func (l *LLMRunner) Name() string { return "llm" }

const (
	toolDelegate  = "delegate_task"
	toolPlan      = "update_plan"
	toolTodos     = "update_todos"
	toolSquad     = "squad_orchestrate"
)

func (l *LLMRunner) Execute(rc RunCtx) error {
	eff := rc.Settings()
	prof := rc.Profile()
	route := ResolveChatModel(prof, eff)
	if route.Model == "" || route.Base == "" {
		return fmt.Errorf("llm: 主模型未配置（当前解析：model=%q base=%q；检查 settings.models.chat 与画像 ModelRoute；演示模式可开 runtime.mock）", route.Model, route.Base)
	}

	client := &agentllm.Client{BaseURL: route.Base, APIKey: route.Key}

	// 知识注入：当前消息包含 @knowledge 引用时检索注入（最小实现：显式引用）。
	var knowledge []core.RetrievalHit
	if strings.Contains(currentUserText(rc), "@knowledge") {
		q := strings.TrimSpace(strings.ReplaceAll(currentUserText(rc), "@knowledge", ""))
		if hits, err := rc.Retrieve(core.RetrievalQuery{Text: q}); err == nil {
			knowledge = hits
		}
	}

	// 组装初始上下文。
	hist := rc.History()
	var currentUser core.Message
	for i := len(hist) - 1; i >= 0; i-- {
		if hist[i].Role == core.RoleUser {
			currentUser = hist[i]
			break
		}
	}
	msgs, asmReport := ctxasm.AssembleInput(prof, eff, hist, currentUser, rc.Attachments(), knowledge)
	if asmReport != nil {
		rc.SetContextObservation(core.ContextObservation{
			Items: asmReport.Items, Truncations: asmReport.Truncations, EstChars: asmReport.EstChars,
		})
	}

	// 声明工具：编排型内置 + 注册表中画像允许的工具。
	decls := []agentllm.ToolDecl{
		{Name: toolSquad, Description: "用小队编排引擎执行多步骤目标：规则规划器拆解为 梳理→专项→成文→终审 步骤并逐步执行（每步一个子代理），挂起决策会以提问卡片征求用户", Parameters: json.RawMessage(`{"type":"object","properties":{"goal":{"type":"string","description":"要完成的整体目标描述"}},"required":["goal"]}`)},
		{Name: toolDelegate, Description: "派生一个子代理执行独立子任务（如检索、撰写），返回其结果摘要", Parameters: json.RawMessage(`{"type":"object","properties":{"name":{"type":"string","description":"子代理角色名"},"task":{"type":"string","description":"子任务描述"},"hint":{"type":"string","description":"执行提示"}},"required":["name","task"]}`)},
		{Name: toolPlan, Description: "创建或更新当前会话的计划（步骤列表）", Parameters: json.RawMessage(`{"type":"object","properties":{"title":{"type":"string"},"steps":{"type":"array","items":{"type":"object","properties":{"id":{"type":"string"},"text":{"type":"string"},"status":{"type":"string","enum":["pending","in_progress","done","skipped"]}},"required":["id","text","status"]}}},"required":["title","steps"]}`)},
		{Name: toolTodos, Description: "更新当前会话的待办清单", Parameters: json.RawMessage(`{"type":"object","properties":{"items":{"type":"array","items":{"type":"object","properties":{"id":{"type":"string"},"text":{"type":"string"},"status":{"type":"string","enum":["pending","in_progress","done"]}},"required":["id","text","status"]}}},"required":["items"]}`)},
	}
	var toolDescs []core.ToolDesc
	for _, t := range toolList(rc) {
		if prof.Tools.ToolAllowed(t.Name) {
			toolDescs = append(toolDescs, t)
			decls = append(decls, agentllm.ToolDecl{Name: t.Name, Description: t.Description, Parameters: t.Parameters})
		}
	}

	var finalParts []core.Part
	usageTotal := core.Usage{}
	maxTurns := 16
	if prof.Budget.MaxTurns != nil {
		maxTurns = *prof.Budget.MaxTurns
	}
	maxTools := 40
	if prof.Budget.MaxToolCalls != nil {
		maxTools = *prof.Budget.MaxToolCalls
	}
	// 输出预算：画像 budget.maxTokens 配置时下发（0 = 走服务端默认，不限制输出长度）。
	streamOpts := agentllm.StreamOptions{}
	if prof.Budget.MaxTokens != nil && *prof.Budget.MaxTokens > 0 {
		streamOpts.MaxTokens = *prof.Budget.MaxTokens
	}
	toolCallsMade := 0
	thinkingIdx, textIdx := 0, 0

	for turn := 0; turn < maxTurns; turn++ {
		select {
		case <-rc.Context().Done():
			return rc.Context().Err()
		default:
		}

		var turnThinking, turnText strings.Builder
		var calls []agentllm.ToolCall
		var usage agentllm.Usage
		ev := client.Stream(rc.Context(), route.Model, msgs, decls, streamOpts)
		phaseEmitted := false
		for e := range ev {
			switch e.Kind {
			case agentllm.KindTextDelta:
				if !phaseEmitted {
					rc.Emit(core.EvRunPhase, core.RunPhasePayload{Phase: core.PhaseResponding})
					phaseEmitted = true
				}
				turnText.WriteString(e.Text)
				rc.Emit(core.EvMessageDelta, core.MessageDeltaPayload{Index: textIdx, Text: e.Text})
			case agentllm.KindReasoningDelta:
				turnThinking.WriteString(e.ReasoningDelta)
				rc.Emit(core.EvThinkingDelta, core.ThinkingDeltaPayload{Index: thinkingIdx, Text: e.ReasoningDelta})
			case agentllm.KindToolCallDelta:
				// 累积由 message_end 的 ToolCalls 提供；此处忽略增量。
			case agentllm.KindMessageEnd:
				calls = e.Message.ToolCalls
				usage = e.Usage
			case agentllm.KindError:
				return fmt.Errorf("llm stream: %s (retryable=%v)", e.ErrorMessage, e.Retryable)
			}
		}
		usageTotal.PromptTokens += usage.PromptTokens
		usageTotal.CompletionTokens += usage.CompletionTokens
		usageTotal.TotalTokens += usage.TotalTokens

		// 记录本轮部件。
		if turnThinking.Len() > 0 {
			finalParts = append(finalParts, core.Part{Type: core.PartThinking, Text: turnThinking.String(), Process: true})
			thinkingIdx++
		}
		if turnText.Len() > 0 {
			finalParts = append(finalParts, core.Part{Type: core.PartText, Text: turnText.String()})
			textIdx++
		}

		if len(calls) == 0 {
			// 终轮：无工具调用。
			if len(finalParts) == 0 {
				finalParts = append(finalParts, core.Part{Type: core.PartText, Text: "（模型未返回内容）"})
			}
			rc.SetAssistantParts(finalParts)
			return nil
		}

		// 工具轮。
		rc.Emit(core.EvRunPhase, core.RunPhasePayload{Phase: core.PhaseActing})
		msgs = append(msgs, agentllm.Message{Role: agentllm.RoleAssistant, Content: turnText.String(), ToolCalls: calls, StopReason: agentllm.StopToolCalls})

		for _, call := range calls {
			toolCallsMade++
			if toolCallsMade > maxTools {
				return fmt.Errorf("budget: 工具调用超过上限 %d（profile.budget.maxToolCalls）", maxTools)
			}
			rc.Emit(core.EvToolCall, core.ToolCallPayload{CallID: call.ID, Name: call.Name, Arguments: call.Arguments})

			var out string
			var err error
			var artifactRef *core.ArtifactRef
			switch call.Name {
			case toolSquad:
				if !eff.GetBool("runtime.squad", true) {
					out = "小队编排已被设置禁用（runtime.squad）"
					break
				}
				var a struct{ Goal string }
				json.Unmarshal([]byte(call.Arguments), &a)
				dataDir := eff.GetString("storage.dataDir", "./data")
				summary, serr := squadadapter.Orchestrate(rc, dataDir, a.Goal)
				if serr != nil {
					out = "小队编排失败：" + serr.Error()
				} else {
					out = summary
				}
			case toolDelegate:
				var a struct{ Name, Task, Hint string }
				json.Unmarshal([]byte(call.Arguments), &a)
				res, serr := rc.SpawnSubagent(orDefault(a.Name, "子代理"), a.Task, a.Hint)
				if serr != nil {
					out = "子代理失败：" + serr.Error()
				} else {
					out = orDefault(res.Summary, "子代理完成")
				}
			case toolPlan:
				var p core.PlanSnapshot
				if err := json.Unmarshal([]byte(call.Arguments), &p); err == nil {
					rc.UpdatePlan(p)
					out = "计划已更新"
				} else {
					out = "计划参数错误: " + err.Error()
				}
			case toolTodos:
				var t core.TodoList
				if err := json.Unmarshal([]byte(call.Arguments), &t); err == nil {
					rc.UpdateTodos(t)
					out = "待办已更新"
				} else {
					out = "待办参数错误: " + err.Error()
				}
			default:
				res := rc.ExecuteTool(core.ToolCallRequest{CallID: call.ID, Name: call.Name, Arguments: json.RawMessage(call.Arguments)})
				if res.Status != "ok" {
					err = fmt.Errorf("%s", res.Output)
				} else {
					out = res.Output
				}
				if res.Artifact != nil {
					artifactRef = &core.ArtifactRef{ArtifactID: res.Artifact.ID, Name: res.Artifact.Name, Kind: res.Artifact.Kind}
				}
				rc.Emit(core.EvToolResult, core.ToolResultPayload{CallID: call.ID, Status: res.Status, Summary: summarize(out, err), Truncated: res.Truncated})
			}
			if call.Name == toolDelegate || call.Name == toolPlan || call.Name == toolTodos || call.Name == toolSquad {
				rc.Emit(core.EvToolResult, core.ToolResultPayload{CallID: call.ID, Status: "ok", Summary: summarize(out, err)})
			}

			summary := ""
			if err != nil {
				summary = "error: " + err.Error()
				out = summary
			}
			finalParts = append(finalParts,
				core.Part{Type: core.PartToolCall, Process: true, ToolCall: &core.ToolCallPart{ID: call.ID, Name: call.Name, Arguments: call.Arguments}},
				core.Part{Type: core.PartToolResult, Process: true, ToolResult: &core.ToolResultPart{CallID: call.ID, Status: statusOf(err), Summary: summary}},
			)
			if artifactRef != nil {
				finalParts = append(finalParts, core.Part{Type: core.PartArtifactRef, Artifact: artifactRef})
			}
			// 回灌工具结果（截断保护）。
			if len(out) > 16000 {
				out = out[:16000] + "…（超长截断）"
			}
			msgs = append(msgs, agentllm.Message{Role: agentllm.RoleTool, ToolCallID: call.ID, Name: call.Name, Content: out, IsError: err != nil})
		}
	}
	return fmt.Errorf("budget: 轮次超过上限 %d（profile.budget.maxTurns）", maxTurns)
}

// SubLLMRunner 子代理执行器：轻量模型单轮完成任务并返回摘要。
type SubLLMRunner struct{}

func (l *SubLLMRunner) Name() string { return "llm-subagent" }

func (l *SubLLMRunner) Execute(rc RunCtx) error {
	route := ResolveLightweightModel(rc.Profile(), rc.Settings())
	if route.Model == "" || route.Base == "" {
		return fmt.Errorf("llm-subagent: 模型未配置（当前解析：model=%q base=%q）", route.Model, route.Base)
	}
	rc.Emit(core.EvAgentProgress, core.AgentProgressPayload{AgentRun: rc.Run().ID, Note: "执行子任务……", Phase: core.PhaseActing})
	client := &agentllm.Client{BaseURL: route.Base, APIKey: route.Key}
	msgs := []agentllm.Message{
		{Role: agentllm.RoleSystem, Content: "你是被派生的子代理。独立完成给定子任务，输出完整、结构化的结果；需要分点时分条列出，不要为了简短而省略关键内容。"},
		{Role: agentllm.RoleUser, Content: rc.Subtask()},
	}
	var sb strings.Builder
	for e := range client.Stream(rc.Context(), route.Model, msgs, nil, agentllm.StreamOptions{}) {
		if e.Kind == agentllm.KindTextDelta {
			sb.WriteString(e.Text)
			rc.Emit(core.EvAgentProgress, core.AgentProgressPayload{AgentRun: rc.Run().ID, Note: truncateRun(sb.String(), 80)})
		}
		if e.Kind == agentllm.KindError {
			return fmt.Errorf("subagent stream: %s", e.ErrorMessage)
		}
	}
	rc.SetAssistantParts([]core.Part{{Type: core.PartText, Text: sb.String()}})
	return nil
}

func currentUserText(rc RunCtx) string {
	hist := rc.History()
	for i := len(hist) - 1; i >= 0; i-- {
		if hist[i].Role == core.RoleUser {
			for _, p := range hist[i].Parts {
				if p.Type == core.PartText {
					return p.Text
				}
			}
		}
	}
	return ""
}

func toolList(rc RunCtx) []core.ToolDesc { return rc.Tools() }

func orDefault(s, def string) string {
	if strings.TrimSpace(s) == "" {
		return def
	}
	return s
}

func summarize(out string, err error) string {
	if err != nil {
		return truncateRun(err.Error(), 160)
	}
	return truncateRun(out, 160)
}

func statusOf(err error) string {
	if err != nil {
		return "error"
	}
	return "ok"
}

func truncateRun(s string, n int) string {
	r := []rune(s)
	if len(r) <= n {
		return s
	}
	return string(r[:n]) + "…"
}