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

package runtime

import (
	"encoding/json"
	"fmt"
	"math"
	"net/http"
	"regexp"
	"strings"
	"time"

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

// 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）。
// 回合循环交给 runtime.Run；这里负责模型路由、提示词、上下文组装和工具声明。
type LLMRunner struct {
	// Prompts 提示词库（nil = 回退画像 SystemPrompt 单段，兼容测试与骨架）。
	Prompts *prompts.Service
	// SkillsDir 内置技能目录（<data>/skills）。写入环境段，供模型定位 house-style。
	SkillsDir string
	// Stream 提供方边界注入点（nil = 按模型路由构造 agent-native 客户端）。
	Stream agentllm.StreamFunc
	// Now 注入时钟（nil = time.Now）。测试用它固定思考与工具耗时。
	Now func() time.Time
	// VerifyStream 是终轮验证器的注入点（nil = 按轻量模型构造客户端）。
	VerifyStream agentllm.StreamFunc
}

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

func (l *LLMRunner) now() time.Time {
	if l.Now != nil {
		return l.Now()
	}
	return time.Now()
}

// durationMs 把 [start, end] 换成毫秒。时钟未打点时返回 0。
func durationMs(start, end time.Time) float64 {
	if start.IsZero() || end.Before(start) {
		return 0
	}
	return float64(end.Sub(start).Microseconds()) / 1000
}

const (
	toolDelegate    = "delegate_task"
	toolReview      = "review_findings"
	toolPlan        = "update_plan"
	toolTodos       = "update_todos"
	toolSquad       = "squad_orchestrate"
	toolRunWorkflow = "run_workflow"
	toolCard        = "submit_card" // 产物卡提交（ui.cardflow 开启时声明；D39 显式路由）
	// 流式产卡三段式（同门控声明；D44/D45）：begin 建卡、append 逐段追加、
	// commit 收尾——长正文优先走此路径，submit_card 留给一次性短产物。
	toolCardBegin  = "card_begin"
	toolCardAppend = "card_append"
	toolCardCommit = "card_commit"
)

// resolveSystem 经提示词库组装 system 段（基座+环境+命中技能），并产出可观测条目。
func (l *LLMRunner) resolveSystem(rc RunCtx, model string) ([]agentllm.Message, []string) {
	userText := currentUserText(rc)
	atts := rc.Attachments()
	names := make([]string, len(atts))
	for i, a := range atts {
		names[i] = a.Meta.Filename
	}
	now := time.Now()
	env := prompts.EnvData{
		Date:             now.Format("2006-01-02"),
		Weekday:          weekdayCN(now),
		Model:            model,
		DeliverableStyle: deliverableStyleLine(rc.Settings()),
		SkillsDir:        l.SkillsDir,
	}
	if len(names) > 0 {
		env.Attachments = strings.Join(names, "、")
	}
	if listed, ok := rc.(interface{ WorkspaceInputs() []string }); ok {
		if paths := listed.WorkspaceInputs(); len(paths) > 0 {
			env.Inputs = strings.Join(paths, "、")
		}
	}
	prof := rc.Profile()
	blocks, hits, err := l.Prompts.ResolveSystem(prof, prompts.TriggerCtx{
		Text: userText, HasAttachments: len(atts) > 0,
	}, env, prof.Owner)
	if err != nil {
		return nil, []string{"prompt(库解析失败：" + err.Error() + ")"}
	}
	// 卡流注入（D39/D42）：ui.cardflow 开启时附加 style.cardflow 模块正文。
	// 不走 PromptModules（向空列表追加会翻转 pick 的"显式引用"语义），独立成块。
	if rc.Settings().GetBool("ui.cardflow", false) {
		if body, ok, err := l.Prompts.ModuleBody(prof.Owner, "style.cardflow"); err == nil && ok {
			blocks = append(blocks, core.PromptBlock{ModuleID: "style.cardflow", Layer: core.LayerStyle, Body: body})
			hits = append(hits, core.PromptHitRecord{ModuleID: "style.cardflow", Injected: true, Reason: "setting(ui.cardflow)"})
		}
	}
	msgs := make([]agentllm.Message, 0, len(blocks))
	var obs []string
	for _, b := range blocks {
		msgs = append(msgs, agentllm.Message{Role: agentllm.RoleSystem, Content: b.Body})
	}
	for _, h := range hits {
		if h.Injected {
			obs = append(obs, fmt.Sprintf("prompt(%s v%d · %s)", h.ModuleID, versionOf(blocks, h.ModuleID), h.Reason))
		} else {
			obs = append(obs, fmt.Sprintf("skill-yield(%s · %s)", h.ModuleID, h.Reason))
		}
	}
	return msgs, obs
}

// deliverableStyleLine 把设置里的默认主题、强调色和模板压成环境段的一句话。
func deliverableStyleLine(eff core.EffectiveSettings) string {
	style := strings.TrimSpace(eff.GetString("deliverables.style", "modern"))
	if style == "" {
		style = "modern"
	}
	line := style
	if accent := strings.TrimSpace(eff.GetString("deliverables.accentColor", "")); accent != "" {
		line += "，强调色 " + accent
	}
	if strings.TrimSpace(eff.GetString("deliverables.templateArtifactId", "")) != "" {
		line += "；已配置默认模板（见 inputs 中的 _default-template）"
	}
	return line
}

// toolDesc 工具描述：库覆盖（tool.<name>）优先，回落注册表文案。
func (l *LLMRunner) toolDesc(owner, name, fallback string) string {
	if l.Prompts != nil {
		if body, ok, err := l.Prompts.ToolPrompt(owner, name); err == nil && ok && strings.TrimSpace(body) != "" {
			return body
		}
	}
	return fallback
}

// artifactFacadeCard 构造工具大结果的门面卡（D47）：kind=result、status=done，
// body = 摘要 markdown + 来源 kv + doc-ref（正文权威在工件，卡只做门面与锚）。
func artifactFacadeCard(tool string, art core.Artifact, summary string) core.ActivityCard {
	md := strings.TrimSpace(summary)
	if md == "" {
		md = "工具结果超限，完整内容已存为工件"
	}
	return core.ActivityCard{
		Kind: core.CardKindResult, Title: "工具大结果 · " + art.Name,
		Status: core.CardDone, Tone: "info",
		Body: []core.UINode{
			{Type: "markdown", Markdown: md},
			{Type: "kv", KV: [][2]string{{"来源工具", tool}, {"工件", art.Name}}},
			{Type: "doc-ref", DocID: core.ID(art.ID), Title: art.Name},
		},
	}
}

func versionOf(blocks []core.PromptBlock, id string) int {
	for _, b := range blocks {
		if b.ModuleID == id {
			return b.Version
		}
	}
	return 0
}

func weekdayCN(t time.Time) string {
	switch t.Weekday() {
	case time.Sunday:
		return "周日"
	case time.Monday:
		return "周一"
	case time.Tuesday:
		return "周二"
	case time.Wednesday:
		return "周三"
	case time.Thursday:
		return "周四"
	case time.Friday:
		return "周五"
	default:
		return "周六"
	}
}

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)
	}

	stream := l.Stream
	if stream == nil {
		client := openLLMClient(route.Base, route.Key)
		stream = client.Stream
	}

	// 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()
	currentUser, _ := lastAddressed(hist)
	var sys []agentllm.Message
	var promptObs []string
	if l.Prompts != nil {
		sys, promptObs = l.resolveSystem(rc, route.Model)
	}
	// 工件正文回填（D46）：artifact_ref 与 card 部件同样回填，下一轮不失忆。
	resolveArtifact := func(id core.ID) (core.Artifact, []byte, bool) { return rc.Artifact(id) }
	if ms, ok := rc.(interface{ Memories() []core.MemoryEntry }); ok {
		sys = ctxasm.InjectMemory(sys, ms.Memories())
	}
	msgs, asmReport := ctxasm.AssembleInput(prof, eff, hist, currentUser, rc.Attachments(), knowledge, sys, promptObs, rc.Card, resolveArtifact)
	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；描述文案库覆盖优先。
	// runtime.entry=squad 时对话本身就是小队入口，不再提供 squad_orchestrate。
	decls := []agentllm.ToolDecl{}
	if squadToolEnabled(eff) {
		decls = append(decls, agentllm.ToolDecl{Name: toolSquad, Description: l.toolDesc(prof.Owner, toolSquad, "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"]}`)})
	}
	decls = append(decls,
		agentllm.ToolDecl{Name: toolDelegate, Description: l.toolDesc(prof.Owner, toolDelegate, "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。模型和工具由子代理 profile 决定，调用方不能指定模型。background 为真时立即返回，子代理完成后系统会自动发起一轮汇报"), Parameters: json.RawMessage(`{"type":"object","properties":{"agent":{"type":"string","enum":["general","explore","research","plan","verify","writer"],"description":"子代理 profile，缺省 general"},"name":{"type":"string","description":"显示名"},"task":{"type":"string","description":"子任务描述"},"hint":{"type":"string","description":"执行提示"},"background":{"type":"boolean","description":"为真时在后台运行，当前回合不等待结果"}},"required":["name","task"]}`)},
		agentllm.ToolDecl{Name: toolReview, Description: l.toolDesc(prof.Owner, toolReview, "对一段材料做评审：多个审查子代理并行找问题，再逐条独立验证，只保留验证通过的问题"), Parameters: json.RawMessage(`{"type":"object","properties":{"target":{"type":"string","description":"要评审的文本或工件"},"focus":{"type":"string","description":"评审重点"},"reviewers":{"type":"integer","minimum":1,"maximum":3,"description":"并行审查人数，1 到 3，缺省 1"}},"required":["target"]}`)},
		agentllm.ToolDecl{Name: toolPlan, Description: l.toolDesc(prof.Owner, toolPlan, "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"]}`)},
		agentllm.ToolDecl{Name: toolTodos, Description: l.toolDesc(prof.Owner, toolTodos, "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"]}`)},
	)
	if eff.GetString("runtime.entry", "direct") != "squad" {
		decls = append(decls, agentllm.ToolDecl{
			Name:        toolRunWorkflow,
			Description: l.toolDesc(prof.Owner, toolRunWorkflow, "运行一个工作流。workflow 运行已注册的定义并带上 inputs；goal 让模型先生成步骤再请用户确认；script 在开关打开后运行一段脚本。三者只能填一个"),
			Parameters:  json.RawMessage(`{"type":"object","properties":{"workflow":{"type":"string","description":"已注册工作流 id"},"inputs":{"type":"object","additionalProperties":{"type":"string"},"description":"槽位初值"},"goal":{"type":"string","description":"用模型生成步骤"},"script":{"type":"string","description":"要运行的脚本"}},"additionalProperties":false}`),
		})
	}
	cardflow := eff.GetBool("ui.cardflow", false)
	if cardflow {
		decls = append(decls, agentllm.ToolDecl{
			Name:        toolCard,
			Description: l.toolDesc(prof.Owner, toolCard, "提交产物型长内容（报告/纪要/方案/对比/文档草稿）为右侧收件箱卡片：卡片承载全文，对话中只保留结论摘要。解释型回答不要用此工具"),
			Parameters:  json.RawMessage(`{"type":"object","properties":{"title":{"type":"string","description":"产物标题"},"summary":{"type":"string","description":"一两句结论摘要（显示在卡片与对话锚点上）"},"body":{"type":"string","description":"完整 Markdown 正文"},"scope":{"type":"string","enum":["conversation","project"],"description":"卡片归属：conversation=当前对话（默认）、project=项目收件箱（跨会话可见）"}},"required":["title","summary","body"]}`),
		})
		decls = append(decls, agentllm.ToolDecl{
			Name:        toolCardBegin,
			Description: l.toolDesc(prof.Owner, toolCardBegin, "开始流式生成右侧收件箱产物卡（长正文推荐）：先给出标题与可选摘要，随后用 card_append 逐段追加正文，写完用 card_commit 收尾"),
			Parameters:  json.RawMessage(`{"type":"object","properties":{"title":{"type":"string","description":"产物标题"},"summary":{"type":"string","description":"一两句结论摘要（可选）"},"scope":{"type":"string","enum":["conversation","project"],"description":"卡片归属：conversation=当前对话（默认）、project=项目收件箱（跨会话可见）"}},"required":["title"]}`),
		})
		decls = append(decls, agentllm.ToolDecl{
			Name:        toolCardAppend,
			Description: l.toolDesc(prof.Owner, toolCardAppend, "向当前流式产物卡追加一段 Markdown 正文（只追加，不改写已提交内容）"),
			Parameters:  json.RawMessage(`{"type":"object","properties":{"text":{"type":"string","description":"本段 Markdown 文本"}},"required":["text"]}`),
		})
		decls = append(decls, agentllm.ToolDecl{
			Name:        toolCardCommit,
			Description: l.toolDesc(prof.Owner, toolCardCommit, "收尾当前流式产物卡：卡片转为完成态并锚定到本条对话"),
			Parameters:  json.RawMessage(`{"type":"object","properties":{}}`),
		})
	}
	for _, t := range toolList(rc) {
		if prof.Tools.ToolAllowed(t.Name) {
			td := t
			td.Description = l.toolDesc(prof.Owner, td.Name, td.Description)
			decls = append(decls, agentllm.ToolDecl{Name: td.Name, Description: td.Description, Parameters: td.Parameters})
		}
	}

	maxTurns, maxTools := loopCaps(prof.Budget)
	streamOpts := agentllm.StreamOptions{MaxTokens: outputTokenCap(prof.Budget, eff, "models.chat.maxOutputTokens")}
	return l.runDirect(rc, stream, route.Model, msgs, decls, streamOpts, cardflow, eff, maxTurns, maxTools)
}

// loopCaps 把画像预算换成内核参数。轮次或工具数为 0 / 未设置表示不限制；
// 内核对 MaxTurns<=0 会回落到 12，所以不限制时传 MaxInt32。MaxToolCalls 的 0 才是不限制。
func loopCaps(budget core.RunBudget) (maxTurns, maxTools int) {
	maxTurns = math.MaxInt32
	if budget.MaxTurns != nil && *budget.MaxTurns > 0 {
		maxTurns = *budget.MaxTurns
	}
	if budget.MaxToolCalls != nil && *budget.MaxToolCalls > 0 {
		maxTools = *budget.MaxToolCalls
	}
	return maxTurns, maxTools
}

func toolBudget(maxTools int) *nruntime.Budget {
	if maxTools <= 0 {
		return nil
	}
	return &nruntime.Budget{MaxToolCalls: maxTools}
}

// outputTokenCap 优先用画像 budget.maxTokens，否则用设置项。两者都是 0 表示不下发，由服务端决定。
func outputTokenCap(budget core.RunBudget, eff core.EffectiveSettings, key string) int {
	if budget.MaxTokens != nil && *budget.MaxTokens > 0 {
		return *budget.MaxTokens
	}
	n := int(eff.GetNumber(key, 0))
	if n > 0 {
		return n
	}
	return 0
}

// openLLMClient 不设 HTTP 客户端超时。单次生成可能很长，停止按钮通过 run 的 context 取消。
func openLLMClient(base, key string) *agentllm.Client {
	return &agentllm.Client{BaseURL: base, APIKey: key, HTTPClient: &http.Client{}}
}

// SubLLMRunner 子代理执行器：按内置 profile 跑一轮带工具的内核循环。
type SubLLMRunner struct {
	// Prompts 提示词库（nil = 使用内联回退文案与只读回退 spec）。
	Prompts *prompts.Service
	// Stream 提供方边界注入点（nil = 按 profile 的模型档构造客户端）。
	Stream agentllm.StreamFunc
}

func currentUserText(rc RunCtx) string {
	msg, ok := lastAddressed(rc.History())
	if !ok {
		return ""
	}
	for _, p := range msg.Parts {
		if p.Type == core.PartText {
			return p.Text
		}
	}
	return ""
}

// lastAddressed 取最后一条用户消息或系统通知，避免通知回合重答更早的问题。
func lastAddressed(hist []core.Message) (core.Message, bool) {
	for i := len(hist) - 1; i >= 0; i-- {
		if hist[i].Role == core.RoleUser || hist[i].Role == core.RoleSystem {
			return hist[i], true
		}
	}
	return core.Message{}, false
}

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

// squadToolEnabled 在小队已经是对话入口时关掉 squad_orchestrate，避免再套一层编排。
func squadToolEnabled(eff core.EffectiveSettings) bool {
	if eff.GetString("runtime.entry", "direct") == "squad" {
		return false
	}
	return eff.GetBool("runtime.squad", true)
}

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

const maxDeliverableGuards = 2

const deliverableGuardNote = "你的上一段回复声称已经生成或登记了文件，但这一轮没有产生对应的工件。请二选一：用 run_python 写出文件并调用 register_artifact 登记，或如实说明这一轮没有生成文件。不要编造工件 ID，也不要在回复里写 art_ 开头的编号。正文里的「调用工具」「工具结果」「产出工件」字样不会执行任何东西，调用工具必须用函数调用。"

const discardedClaimNote = "（上一段回复声称已交付，但本轮没有任何工具调用，已作废）"

const deliverableFallbackNotice = "本轮未实际生成文件（模型声称已生成，但没有调用工具）。请重试，或换一种说法再试一次。"

var (
	artIDRe      = regexp.MustCompile(`art_[0-9a-f]{12}`)
	fileClaimRe  = regexp.MustCompile(`(?s)(?:已生成|已保存|已导出|已写入).{0,120}?\.(?:docx|xlsx|pptx|pdf|png|jpe?g|csv|md|txt|json)\b`)
	softClaimRe  = regexp.MustCompile(`(?s)(?:文件|文档|工件).{0,40}?已(?:正常)?(?:生成|登记|保存|导出|写入)`)
	forgedCallRe = regexp.MustCompile(`\[(?:调用工具|工具结果|产出工件) `)
)

// claimsFalseDeliverable 保留给小队入口：命中即判失败，不走直连路径的纠正轮次。
func claimsFalseDeliverable(body string, produced []*core.ArtifactRef, rc RunCtx) bool {
	block, _ := deliverableGate(body, produced, rc)
	return block
}

// deliverableGate 判断终轮正文是否在没有对应工件的情况下声称交付。
// 返回值的形状与 Stop 门禁一致，P1 平移进内核时不用再改调用方。
// 引用本轮已产出、或项目里确实存在的工件，不算谎称；正文里的伪造调用标记始终拦截。
func deliverableGate(body string, produced []*core.ArtifactRef, rc RunCtx) (bool, string) {
	if strings.TrimSpace(body) == "" {
		return false, ""
	}
	if forgedCallRe.MatchString(body) {
		return true, "回复正文里出现了工具调用标记，但本轮没有真实的工具调用"
	}
	producedIDs := map[string]bool{}
	for _, ref := range produced {
		if ref != nil && ref.ArtifactID != "" {
			producedIDs[ref.ArtifactID] = true
		}
	}
	knownExisting := false
	for _, id := range artIDRe.FindAllString(body, -1) {
		if producedIDs[id] {
			continue
		}
		if _, _, ok := rc.Artifact(id); ok {
			knownExisting = true
			continue
		}
		return true, "回复引用了不存在的工件"
	}
	if len(producedIDs) > 0 || knownExisting {
		return false, ""
	}
	if fileClaimRe.MatchString(body) || softClaimRe.MatchString(body) {
		return true, "回复声称已经生成或登记了文件，但这一轮没有产生对应的工件"
	}
	return false, ""
}

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]) + "…"
}

// splitThinkBlock 剥离正文里内联的 <think>…</think> 推理块（部分 OpenAI 兼容
// 模型把思考混进 content 流）：返回（干净正文, 思考内容）。只处理首部成对标签；
// 未闭合或非首部出现按原文返回（宁可多显示也不误删正文）。
func splitThinkBlock(s string) (body, think string) {
	trimmed := strings.TrimLeft(s, " \t\r\n")
	if !strings.HasPrefix(trimmed, "<think>") {
		return s, ""
	}
	rest := strings.TrimPrefix(trimmed, "<think>")
	i := strings.Index(rest, "</think>")
	if i < 0 {
		return s, ""
	}
	think = strings.TrimSpace(rest[:i])
	body = strings.TrimSpace(rest[i+len("</think>"):])
	if think == "" {
		return body, ""
	}
	return body, think
}