Files
Cyrene_v2/internal/runtime/loop/request.go
T
Aska 33eeadfc38
CI / Backend (build / vet / test / cross-compile) (push) Has been cancelled
CI / Frontend (tsc + vite build) (push) Has been cancelled
fix(backthink): 后台思考静默广播——不再把内部推理泄到聊天思维链 + XML 动作/旁白解析
【静默广播修复(bug:思维链露出 backthink 推理)】
- 根因:LoopThinker.Think 复用同一 loop.Loop 且 loop 全局带 OnStream,backthink 后台思考时
  把 [background autonomous reflection] 的 thinking/流式事件广播到前端,被当聊天思维链显示。
- loop/mute_ctx.go:WithMutedBroadcast(ctx) + isMutedBroadcast(ctx)——context 上静默标志。
- loop 四处广播抑制(stream.go 流式 / step.go 静默走非流式 agentStep / request.go capacity 事件 /
  tools.go tool_execution),静默时跳过 UI 广播,推理照常返回结论。
- thinker.go LoopThinker.Think 用 loop.WithMutedBroadcast(ctx),backthink 不再泄 thinking。

【XML 动作/旁白结构化解析】
- cyrene.yaml systemPrompt 加「表达方式:动作与旁白」——动作 <action>…</action>、
  旁白 <narration>…</narration>;克制≤1-2个/段、贴合唤醒态、重要信息不加、朗读正文忽略标签。
- 前端 parseContent.ts:ContentPart[](text/action/narration),非贪婪+反向引用、跨行、非法/
  未闭合兜底纯文本不崩;summarizeContent 供思维导图/状态。
- 前端 renderContent.tsx:action/narration → 灰斜体(mu-text-lo italic)句中原位渲染。
- 应用到 chat.tsx UserBubble/AssistantBubble + messages 各 mode body。

go build/vet/test(loop/backthink)绿 + 前端 tsc/vite 绿。
2026-09-06 22:09:51 +08:00

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 && !isMutedBroadcast(ctx) {
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
}