【capacity 唤醒式核心(77)】internal/runtime/capacity:(Decider L1启发式+L2上下文+人格覆盖→三态静息/唤醒/全能→允许工具集+capacity_awaken事件)。loop 外围允许集过滤(不硬阻断=能力≠边界红线)。config 加 Capacity(Enabled/MinConfidence/ToolSets)。wire buildCapacity 接入 loop。 【L4 记忆驱动意图(80)】recall.go:从记忆召回用户习惯(搞一下xx=办事)→唤醒L1/L2漏掉/升置信;群聊陌生人不读owner习惯(隔离红线);RecordIntent闭环(判定写回记忆越用越准);默认保守静息。 【会话消息编排(78)】gateway/session_orchestrator.go:每会话(模式/pending有界5/tail上下文尾巴20);Submit非阻塞(idle立即/busy入队ack/满拒绝);完成回调按序续接(tail+新消息合并增量注入非全量重放);Interrupt中断;REST 202/409+WS session_busy/pending_ack事件;不重造任务引擎复用loop。 【前端集成(79)】ws.ts 加 capacity_awaken/session_busy/pending_ack 解析;store 加 awakenState/sessionBusy;TopBar/ChatShell 加唤醒态徽标+忙态提示(不阻塞输入框);状态组件 status.tsx;Miuix风格。 【管理面板写CRUD(81)】20个写端点(POST/PUT/DELETE /admin/*,owner+审计落库):personas(真持久化YAML)/users apikey(签发掩码+撤销)/iot控制(真下发)/automations(建/触发/删)/llm(test真调用+临时)/channels(connect/close/身份测试actor白名单)/plugins(卸载)/memories(search真检索)/sessions(export)。无存储如实501不伪造。拆4文件满足≤500红线。前端9模块接真写。 【wire收口】buildCapacity接入loop+OnCapacity事件;server.go newOrchestrator接hub。 实机验证:iot ac.living control 200 真实改变 on:true;observability/metrics 200、health ok;全仓 build/vet/test 47包绿+前端tsc/vite。
182 lines
6.1 KiB
Go
182 lines
6.1 KiB
Go
package loop
|
|
|
|
import (
|
|
"context"
|
|
|
|
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/agent"
|
|
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/capacity"
|
|
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
|
|
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/tool"
|
|
)
|
|
|
|
// buildRequest assembles a model.Request from the agent's public surface,
|
|
// mirroring agent.buildRequest so the streaming path behaves identically to
|
|
// RunStep (leading instruction message, tool schemas, model routing).
|
|
func (l *Loop) buildRequest(ctx context.Context, messages []model.Message) model.Request {
|
|
req := model.Request{
|
|
Model: l.agent.Model(),
|
|
Messages: l.assembleMessages(messages),
|
|
Temperature: l.temperature(),
|
|
MaxTokens: l.agent.MaxTokens(),
|
|
Purpose: l.agent.Purpose(),
|
|
Tools: l.toolSchemas(ctx, messages),
|
|
}
|
|
if l.agent.Router() != nil {
|
|
if m, _ := l.agent.Router().RouteRequest(req); m != "" {
|
|
req.Model = m
|
|
}
|
|
}
|
|
return req
|
|
}
|
|
|
|
// assembleMessages prepends the agent's instructions as a leading system message
|
|
// when they are non-empty, otherwise returns the input slice unchanged.
|
|
func (l *Loop) assembleMessages(messages []model.Message) []model.Message {
|
|
instr := l.agent.Instructions()
|
|
if instr == "" {
|
|
return messages
|
|
}
|
|
out := make([]model.Message, 0, len(messages)+1)
|
|
out = append(out, model.TextMessage(model.RoleSystem, instr))
|
|
out = append(out, messages...)
|
|
return out
|
|
}
|
|
|
|
// temperature returns a pointer to the configured temperature, mapping zero to
|
|
// nil so the provider default applies (matching the agent package).
|
|
func (l *Loop) temperature() *float64 {
|
|
t := l.agent.Temperature()
|
|
if t == 0 {
|
|
return nil
|
|
}
|
|
return &t
|
|
}
|
|
|
|
// toolSchemas advertises the registered tool schemas to the model, or nil when no
|
|
// dispatcher is configured. When a capacity decider is wired and enabled, it
|
|
// classifies the current round and returns only the allowed set for that state
|
|
// (a view over the dispatcher, per design/14 §4): the model sees/invokes only
|
|
// the allowed tools, while Invoke itself stays unblocked. A nil or disabled
|
|
// decider keeps the pre-existing behavior (every tool advertised).
|
|
func (l *Loop) toolSchemas(ctx context.Context, messages []model.Message) []model.ToolSpec {
|
|
d := l.dispatcher()
|
|
if d == nil {
|
|
return nil
|
|
}
|
|
schemas := d.ToolSchemas()
|
|
if l.capacity == nil || !l.capacity.Enabled() {
|
|
return schemas
|
|
}
|
|
|
|
names := make([]string, 0, len(schemas))
|
|
for _, s := range schemas {
|
|
names = append(names, s.Name)
|
|
}
|
|
acc := capState(ctx)
|
|
dec := l.capacity.Decide(ctx, capacity.Input{
|
|
Messages: capacityTexts(messages),
|
|
CurrentState: acc.state,
|
|
AllToolNames: names,
|
|
Recall: l.capacityRecall(ctx, messages),
|
|
})
|
|
acc.state = dec.State
|
|
if l.onCapacity != nil {
|
|
for _, ev := range dec.Events {
|
|
l.onCapacity(ev)
|
|
}
|
|
}
|
|
|
|
allowed := make(map[string]bool, len(dec.AllowedTools))
|
|
for _, n := range dec.AllowedTools {
|
|
allowed[n] = true
|
|
}
|
|
out := make([]model.ToolSpec, 0, len(dec.AllowedTools))
|
|
for _, s := range schemas {
|
|
if allowed[s.Name] {
|
|
out = append(out, s)
|
|
}
|
|
}
|
|
return out
|
|
}
|
|
|
|
// capacityTexts extracts the recent conversation text for the capacity decider:
|
|
// the last few non-system/non-tool turns, in chronological order (oldest first),
|
|
// so the decider's current-turn window sees the latest user message last.
|
|
func capacityTexts(messages []model.Message) []string {
|
|
texts := make([]string, 0, 6)
|
|
for i := len(messages) - 1; i >= 0 && len(texts) < 6; i-- {
|
|
m := messages[i]
|
|
if m.Role == model.RoleSystem || m.Role == model.RoleTool {
|
|
continue
|
|
}
|
|
if t := messageText(m); t != "" {
|
|
texts = append(texts, t)
|
|
}
|
|
}
|
|
for i, j := 0, len(texts)-1; i < j; i, j = i+1, j-1 {
|
|
texts[i], texts[j] = texts[j], texts[i]
|
|
}
|
|
return texts
|
|
}
|
|
|
|
// capacityRecallTopK bounds how many recalled habit memories the capacity L4 layer
|
|
// considers per round.
|
|
const capacityRecallTopK = 5
|
|
|
|
// capacityRecall retrieves the owner's relevant long-term memories for the capacity
|
|
// L4 habit-alignment layer (design §2.2 L4), the memory feed that lets the decider
|
|
// learn the user's task-expression habits. It returns nil (keeping L4 silent) when
|
|
// no retriever/persona is configured — backward compatible — or in a restricted
|
|
// (group chat / stranger) context, per the isolation red line: L4 learns the owner's
|
|
// habits and must never read owner-sensitive memory into a group or stranger intent
|
|
// decision (design §2.2 红线: 不读陌生人敏感).
|
|
func (l *Loop) capacityRecall(ctx context.Context, messages []model.Message) []capacity.RecallItem {
|
|
if l.retriever == nil || l.personaID == "" {
|
|
return nil
|
|
}
|
|
env := sessionEnvFrom(ctx)
|
|
if env.Restricted() {
|
|
return nil
|
|
}
|
|
results, err := l.retriever.Retrieve(ctx, l.personaID, queryFromMessages(messages), capacityRecallTopK)
|
|
if err != nil {
|
|
l.log.Warn("capacity recall", "error", err, "env", env.ActorEnv)
|
|
return nil
|
|
}
|
|
items := make([]capacity.RecallItem, 0, len(results))
|
|
for _, r := range results {
|
|
items = append(items, capacity.RecallItem{
|
|
Content: r.Item.Content,
|
|
Scope: string(r.Item.Scope),
|
|
Kind: string(r.Item.Kind),
|
|
Tags: r.Item.Tags,
|
|
Score: r.Score,
|
|
})
|
|
}
|
|
return items
|
|
}
|
|
|
|
// decodeToolCalls decodes tool calls from a model.Response using the same two
|
|
// conventions as the agent package: the normalized RoleTool placeholder, then a
|
|
// JSON envelope of the form {"tool_calls":[...]} in Content. It reports ok=true
|
|
// (with a non-empty slice) when a call is decoded.
|
|
func decodeToolCalls(resp model.Response) ([]agent.ToolCall, bool) {
|
|
if resp.Message.Role == model.RoleAssistant && resp.Message.ToolCallName != "" {
|
|
return []agent.ToolCall{{Name: resp.Message.ToolCallName, Args: map[string]any{}}}, true
|
|
}
|
|
var payload struct {
|
|
ToolCalls []agent.ToolCall `json:"tool_calls"`
|
|
}
|
|
// Lenient decode repairs common model-emitted malformations (code fences,
|
|
// trailing commas) so a slightly broken tool-call envelope does not turn into
|
|
// a text answer; an unrecoverable payload still yields ok=false and the round
|
|
// proceeds as text.
|
|
if err := tool.LenientJSONUnmarshal([]byte(resp.Content), &payload); err != nil {
|
|
return nil, false
|
|
}
|
|
if len(payload.ToolCalls) == 0 {
|
|
return nil, false
|
|
}
|
|
return payload.ToolCalls, true
|
|
}
|