17 Stages
全链路治理管线
10 阶段 ControlPlanePipeline + 14 个 WorkerLoop Mixin 协同运作
追踪: ... · 运行: ... · 耗时: 0ms (模拟)
速度:
执行中
已完成
等待中
Mixin
管线阶段
基础设施
NL 输入接收HTTP API
endpointPOST /intent
text"帮我创建一个TODO应用HTML页面"
sessiontest-agentic-final
RPC 提交 → 内核守护RPC
pathPrimary (RPC Daemon)
methodsubmit_run()
port9071
WorkerLoop.tick_once()Worker Loop
mixins14 active
modelocal_dev
BrainLearningCommunicationEnterpriseGovGoalSynthesisMemoryKbLlmRoutingModeDispatchResourceOrchPerceptionSafetyMonitorMetacognitionToolReasoningWebIntelHelpers
记忆上下文加载MemoryKb
eventMEMORY_READ
portSqliteMemoryPort.query()
records5 recalled
ControlPlanePipeline (10 阶段)10-Stage
typeSequential governance pipeline
1.NormalizeInput2.IntentRouting3.RiskAssess4.Governance5.ModelSelect6.CostGov7.CapSchedule8.ExecuteCap9.EvidenceCommit10.FeedbackDistill
意图路由 (LLM)Stage 2
eventRIS_ROUTING_RECOMMENDATION
routeagentic_task
layerllm (keyword fallback failed)
ms3106ms
风险评估 + 治理Stage 3-4
eventPOLICY_DECISION
riskacceptable
govgovernance_required: false
目标合成 + 计划生成GoalSynthesis
eventPLAN_GENERATED
depts2 departments
tasks2 tasks
能力调度 + 发现Stage 7
handleragentic_handler
learnedlearned_capabilities_found: 0
sourceSkill Discovery scan
模式分发ModeDispatch
eventMODE_DISPATCH_DECIDED
modeexecute_enqueued_intent
安全 + 元认知门控Safety
corrigibilityPASS
confidencePASS
perceptionmultimodal scan OK
执行能力 (Agent Executor)Stage 8
executoragent_execute()
modelclaude-sonnet-4-6 (Anthropic)
max_tok16384
Step 1: write_fileStep 2: read_fileStep 3: done
LLM 工具循环 (3 步)Anthropic API
step 1write_file → /tmp/todo.html (7678 bytes)
step 2read_file → 验证内容
step 3done → 任务完成
记忆写入 (执行摘要)MemoryKb
methodMemoryPort.append(MemoryPoint)
idagentic-7e33b87f413a
resultcompleted
证据提交 + 反馈蒸馏Stage 9-10
events18 events committed
digeste58b8d58a604...
sinkfeedback_sink → belief update
资源编排 (计划构建)ResourceOrch
eventPLAN_BUILD_STARTED → PLAN_BUILD_FINISHED
KBRefbase.default (KnowledgePort stub)
RUN_FINISHED → HTTP 响应Output
statuscompleted
sourceprimary (RPC path)
total45930ms (45.9s)
file/tmp/todo.html (7678 bytes)