weichen
b6ff0ab55f
Details (original message): Add catalog-driven orchestration: intent catalog injection, deterministic plan validation, and system-function steps.
299 lines
11 KiB
Go
299 lines
11 KiB
Go
package squad
|
||
|
||
import (
|
||
"context"
|
||
"encoding/json"
|
||
"fmt"
|
||
"strings"
|
||
|
||
"agent-native/runtime/llm"
|
||
)
|
||
|
||
// LLMPlanner LLM 规划器：用模型理解目标 → 拆解任务 → 选/派角色 → 生成步骤。
|
||
// 模型客户端与默认模型在构造时注入（角色级模型在派生角色时由模型自行指定）。
|
||
// 输出为严格 JSON；解析失败时重试一次，仍失败则 fail-closed 返回错误。
|
||
type LLMPlanner struct {
|
||
Client *llm.Client
|
||
Model string
|
||
APIBase string
|
||
APIKey string
|
||
MaxTokens int
|
||
}
|
||
|
||
// NewLLMPlanner 创建 LLM 规划器。
|
||
func NewLLMPlanner(client *llm.Client, model string) *LLMPlanner {
|
||
return &LLMPlanner{Client: client, Model: model}
|
||
}
|
||
|
||
func (p *LLMPlanner) Name() string { return "llm" }
|
||
|
||
// llmPlanSchema 期望模型输出的 JSON 结构（宽松解析：仅取必要字段）。
|
||
type llmPlanSchema struct {
|
||
Clarify []llmQuestion `json:"clarify,omitempty"`
|
||
Confirm bool `json:"confirm,omitempty"`
|
||
Reason string `json:"confirm_reason,omitempty"`
|
||
Agents []llmAgent `json:"agents"`
|
||
Steps []llmStep `json:"steps"`
|
||
Template string `json:"template,omitempty"`
|
||
}
|
||
|
||
type llmQuestion struct {
|
||
ID string `json:"id,omitempty"`
|
||
Text string `json:"text"`
|
||
Options []HumanOption `json:"options,omitempty"`
|
||
}
|
||
|
||
type llmAgent struct {
|
||
Action string `json:"action"` // use | derive
|
||
Role string `json:"role"`
|
||
Name string `json:"name,omitempty"`
|
||
BaseRole string `json:"base_role,omitempty"`
|
||
Model string `json:"model,omitempty"`
|
||
Backend string `json:"backend,omitempty"`
|
||
DomainTopic string `json:"domain_topic,omitempty"`
|
||
Responsibilities []string `json:"responsibilities,omitempty"`
|
||
}
|
||
|
||
type llmStep struct {
|
||
ID string `json:"id"`
|
||
Role string `json:"role"`
|
||
Title string `json:"title"`
|
||
Parallel bool `json:"parallel,omitempty"`
|
||
Targets []string `json:"targets,omitempty"`
|
||
RequiresApproval bool `json:"requires_approval,omitempty"`
|
||
IntentID string `json:"intent_id,omitempty"`
|
||
IntentSlots map[string]string `json:"intent_slots,omitempty"`
|
||
}
|
||
|
||
const llmPlannerSystem = `你是智能体小队的规划器。给定用户目标、已注册角色清单和可选的任务模板，输出一个可执行的小队计划。
|
||
规则：
|
||
1. 只输出一个 JSON 对象，不要输出任何解释文字或代码围栏。
|
||
2. 目标信息不足时必须输出 {"clarify":[{"id":"q1","text":"问题","options":[{"value":"continue","label":"…"}]}]} 并停止，不要替用户假设。判定标准（任一命中即 clarify）：目标少于 15 个字；没有明确的对象/主题（如"帮我写个东西"）；没有可辨识的交付物；关键约束互相矛盾。
|
||
3. 正常情况输出 {"agents":[...],"steps":[...],"confirm":bool,"confirm_reason":"...","template":"markdown 骨架"}。
|
||
4. agents: 优先复用已注册角色（{"action":"use","role":"已注册key"}）；确无合适角色才派生（{"action":"derive","role":"新key","base_role":"基线key","name":"中文名","domain_topic":"领域","responsibilities":[...]}）。派生数量尽量少。
|
||
5. steps: 每步 {"id":"step_01_x","role":"角色key","title":"步骤职责"}；可并行的相邻步骤 parallel=true；高风险/破坏性/对外发布类步骤 requires_approval=true；步骤通常 2~6 个，含梳理、专业执行、汇总成文、终审把关。
|
||
6. 涉及写入、删除、对外提交或超出常规预算的计划 confirm=true。
|
||
7. 通用守则：缺失的信息若会实质改变结果、执行路径、权限或安全，必须 clarify 问人；否则自行决定并在计划中说明你决定了什么。严禁假设交付完成：无法覆盖的目标范围要如实体现在步骤或假设中，不许静默缩水。
|
||
8. 目标下方若给出"系统能力目录"：子任务命中目录能力时，该步输出 {"role":"intent","intent_id":"目录id","intent_slots":{...},"title":"..."}，槽位必须按目录填齐；必填槽位信息不足时禁止编造，改为 clarify 并使用目录的提问建议。未命中目录的子任务照常派给角色。没有目录时不得输出 intent_id 步骤。`
|
||
|
||
func (p *LLMPlanner) Plan(ctx context.Context, input PlanInput) (*PlanOutput, error) {
|
||
if p.Client == nil || p.Model == "" {
|
||
return nil, fmt.Errorf("llm planner requires client and model")
|
||
}
|
||
userMsg := p.buildPrompt(input)
|
||
|
||
var parsed *llmPlanSchema
|
||
var lastErr error
|
||
for attempt := 0; attempt < 2; attempt++ {
|
||
sys := llmPlannerSystem
|
||
if attempt > 0 {
|
||
sys += "\n注意：上一次输出无法解析为 JSON。这次只输出 JSON 对象本身。"
|
||
}
|
||
raw, err := p.complete(ctx, sys, userMsg)
|
||
if err != nil {
|
||
return nil, fmt.Errorf("llm planner call failed: %w", err)
|
||
}
|
||
schema, err := parseLLMPlan(raw)
|
||
if err != nil {
|
||
lastErr = err
|
||
continue
|
||
}
|
||
parsed = schema
|
||
break
|
||
}
|
||
if parsed == nil {
|
||
return nil, fmt.Errorf("llm planner output unparsable after retry: %w", lastErr)
|
||
}
|
||
|
||
if len(parsed.Clarify) > 0 && len(parsed.Agents) == 0 {
|
||
cl := &Clarification{Reason: "模型判定目标信息不足"}
|
||
for _, q := range parsed.Clarify {
|
||
cl.Questions = append(cl.Questions, Question{ID: q.ID, Text: q.Text, Options: q.Options})
|
||
}
|
||
return &PlanOutput{Clarification: cl}, nil
|
||
}
|
||
if len(parsed.Steps) == 0 {
|
||
return nil, fmt.Errorf("llm planner produced no steps")
|
||
}
|
||
return p.materialize(input, parsed)
|
||
}
|
||
|
||
// materialize 把模型输出落成 RoleSpec/StepSpec：use 校验存在，derive 走注册表派生。
|
||
func (p *LLMPlanner) materialize(input PlanInput, parsed *llmPlanSchema) (*PlanOutput, error) {
|
||
var agents []RoleSpec
|
||
addAgent := func(spec RoleSpec) {
|
||
for _, a := range agents {
|
||
if a.Role == spec.Role {
|
||
return
|
||
}
|
||
}
|
||
agents = append(agents, spec)
|
||
}
|
||
for _, la := range parsed.Agents {
|
||
switch strings.ToLower(la.Action) {
|
||
case "derive":
|
||
derived, err := input.Registry.Derive(DeriveRequest{
|
||
RoleKey: la.Role,
|
||
Name: la.Name,
|
||
BaseRole: la.BaseRole,
|
||
Backend: la.Backend,
|
||
Model: la.Model,
|
||
DomainTopic: la.DomainTopic,
|
||
Responsibilities: la.Responsibilities,
|
||
})
|
||
if err != nil {
|
||
return nil, fmt.Errorf("derive role %s: %w", la.Role, err)
|
||
}
|
||
addAgent(derived)
|
||
default: // use
|
||
spec, ok := input.Registry.Get(la.Role)
|
||
if !ok {
|
||
return nil, fmt.Errorf("plan references unregistered role %q (use derive action for new roles)", la.Role)
|
||
}
|
||
addAgent(spec)
|
||
}
|
||
}
|
||
|
||
steps := make([]StepSpec, 0, len(parsed.Steps))
|
||
for i, ls := range parsed.Steps {
|
||
if ls.Role == "" {
|
||
return nil, fmt.Errorf("step %d missing role", i)
|
||
}
|
||
if _, ok := input.Registry.Get(ls.Role); !ok {
|
||
return nil, fmt.Errorf("step %d references unregistered role %q", i, ls.Role)
|
||
}
|
||
if _, inSquad := findAgent(agents, ls.Role); !inSquad {
|
||
if spec, ok := input.Registry.Get(ls.Role); ok {
|
||
addAgent(spec)
|
||
}
|
||
}
|
||
id := ls.ID
|
||
if id == "" {
|
||
id = fmt.Sprintf("step_%02d", i+1)
|
||
}
|
||
targets := ls.Targets
|
||
if len(targets) == 0 {
|
||
targets = []string{"output.md"}
|
||
}
|
||
steps = append(steps, StepSpec{
|
||
ID: id, Role: ls.Role, Title: ls.Title, Parallel: ls.Parallel,
|
||
Targets: targets, RequiresApproval: ls.RequiresApproval, Status: "pending",
|
||
IntentID: ls.IntentID, IntentSlots: ls.IntentSlots,
|
||
})
|
||
}
|
||
|
||
plan := &SquadPlan{
|
||
Squad: SquadConfig{SquadName: truncate(input.Goal.Text, 40), Version: "1", Agents: agents},
|
||
Workflow: WorkflowPlan{
|
||
WorkflowID: NewPlanID(),
|
||
Title: truncate(input.Goal.Text, 80),
|
||
PlanningMode: PlanModeAutonomous,
|
||
Status: "pending",
|
||
Steps: steps,
|
||
},
|
||
Template: parsed.Template,
|
||
}
|
||
return &PlanOutput{Plan: plan, NeedsConfirmation: parsed.Confirm, ConfirmReason: parsed.Reason}, nil
|
||
}
|
||
|
||
func (p *LLMPlanner) buildPrompt(input PlanInput) string {
|
||
var b strings.Builder
|
||
b.WriteString("## 用户目标\n")
|
||
b.WriteString(input.Goal.Text + "\n")
|
||
if len(input.Goal.Constraints) > 0 {
|
||
b.WriteString("\n## 约束\n")
|
||
for k, v := range input.Goal.Constraints {
|
||
fmt.Fprintf(&b, "- %s: %s\n", k, v)
|
||
}
|
||
}
|
||
if len(input.Goal.Materials) > 0 {
|
||
b.WriteString("\n## 输入材料（文件名）\n")
|
||
for _, m := range input.Goal.Materials {
|
||
fmt.Fprintf(&b, "- %s\n", m.Filename)
|
||
}
|
||
}
|
||
b.WriteString("\n## 已注册角色\n")
|
||
for _, r := range input.Registry.List() {
|
||
fmt.Fprintf(&b, "- key=%s name=%s backend=%s keywords=%v responsibilities=%v\n",
|
||
r.Role, r.Name, r.Backend, r.Keywords, r.Responsibilities)
|
||
}
|
||
if input.Blueprint != nil {
|
||
b.WriteString("\n## 命中任务模板（可作为基线微调）\n")
|
||
enc, _ := json.Marshal(input.Blueprint)
|
||
b.Write(enc)
|
||
b.WriteString("\n")
|
||
}
|
||
if input.IntentCatalog != nil && input.IntentCatalog.Markdown != "" {
|
||
b.WriteString("\n## 系统能力目录（可直接调用，不必派智能体）\n")
|
||
b.WriteString(input.IntentCatalog.Markdown)
|
||
b.WriteString("\n")
|
||
}
|
||
return b.String()
|
||
}
|
||
|
||
// complete 调一次模型并拼接完整文本。
|
||
func (p *LLMPlanner) complete(ctx context.Context, system, user string) (string, error) {
|
||
msgs := []llm.Message{
|
||
{Role: llm.RoleSystem, Content: system},
|
||
{Role: llm.RoleUser, Content: user},
|
||
}
|
||
opts := llm.StreamOptions{MaxTokens: p.MaxTokens, APIBase: p.APIBase, APIKey: p.APIKey}
|
||
var sb strings.Builder
|
||
for ev := range p.Client.Stream(ctx, p.Model, msgs, nil, opts) {
|
||
if ev.Kind == llm.KindError {
|
||
return "", fmt.Errorf("%s", ev.ErrorMessage)
|
||
}
|
||
if ev.Kind == llm.KindTextDelta {
|
||
sb.WriteString(ev.Text)
|
||
}
|
||
if ev.Kind == llm.KindMessageEnd {
|
||
if ev.Message.IsError {
|
||
return "", fmt.Errorf("assistant error: %s", ev.Message.Content)
|
||
}
|
||
if sb.Len() == 0 {
|
||
sb.WriteString(ev.Message.Content)
|
||
}
|
||
break
|
||
}
|
||
}
|
||
if ctx.Err() != nil {
|
||
return "", ctx.Err()
|
||
}
|
||
return sb.String(), nil
|
||
}
|
||
|
||
// parseLLMPlan 从模型输出稳健解析 JSON（剥离 think 块/代码围栏/前后杂文）。
|
||
func parseLLMPlan(raw string) (*llmPlanSchema, error) {
|
||
s := stripThinkBlocks(strings.TrimSpace(raw))
|
||
s = strings.TrimPrefix(s, "```json")
|
||
s = strings.TrimPrefix(s, "```")
|
||
s = strings.TrimSuffix(s, "```")
|
||
s = strings.TrimSpace(s)
|
||
// 截取首个 { 到最后一个 }，容忍模型前后夹杂说明文字。
|
||
if i := strings.IndexByte(s, '{'); i >= 0 {
|
||
if j := strings.LastIndexByte(s, '}'); j > i {
|
||
s = s[i : j+1]
|
||
}
|
||
}
|
||
var out llmPlanSchema
|
||
if err := json.Unmarshal([]byte(s), &out); err != nil {
|
||
return nil, err
|
||
}
|
||
return &out, nil
|
||
}
|
||
|
||
// stripThinkBlocks 移除 MiniMax 等模型内嵌在 content 中的 <think>…</think> 块。
|
||
func stripThinkBlocks(s string) string {
|
||
for {
|
||
i := strings.Index(s, "<think>")
|
||
if i < 0 {
|
||
return s
|
||
}
|
||
j := strings.Index(s[i:], "</think>")
|
||
if j < 0 {
|
||
return s[:i] // 未闭合：丢弃思考余段
|
||
}
|
||
s = s[:i] + s[i+j+len("</think>"):]
|
||
}
|
||
}
|