feat(scheduler): M7 可观测性 + 压力测试 + 端到端测试

设计依据 docs/zh/input-scheduler-design.md §11.5(O1/O2)、§11.6(E1/E2)。

- 可观测性:KernelStatus 新增 Scheduler 段(running/三集合深度/计数/
  深度上限),由 GetKernelStatus 从 DumpScheduler 原子快照填充;
  新增 events.EventScheduler,挂起/恢复各发一条(action/task/level)
- TaskKind.String() 便于日志与状态输出
- 新增 scheduler_e2e_test.go 3 项:
  · 压力:200 排队输入 + 50 中断全部经真实 loop 执行,结束时三集合排空、
    LLM 调用数精确等于输入数、无 Rejected
  · 可观测性:挂起/恢复事件齐备,状态快照计数一致
  · 端到端:完整启动 schedulerLoop+interceptLoop,经真实 channel 投递
    L1 任务与 L4 中断,验证「LLM 流式中断 → 挂起 → 中断先完成 → 原任务恢复」
    整条链路(LLM 调用数 = 丢弃1+中断1+恢复1+常规1)
- 验收:agent 全量 + -race;全仓 build/vet 通过
This commit is contained in:
JianFeeeee
2026-09-13 00:45:28 +08:00
parent 4e4e0ad656
commit f11de37bf2
5 changed files with 311 additions and 10 deletions

View File

@ -25,6 +25,8 @@ import (
agentAPI "gitcode.com/JianFeeeee/HomeAgent/internal/agent/api"
agentIO "gitcode.com/JianFeeeee/HomeAgent/internal/agent/io"
"gitcode.com/JianFeeeee/HomeAgent/internal/events"
sdk "gitcode.com/JianFeeeee/HomeAgent/internal/sdk"
)
// Level 是任务优先级,由内核预定义四级(设计文档 §3.1)。
@ -73,6 +75,17 @@ const (
TaskKindSelf
)
func (k TaskKind) String() string {
switch k {
case TaskKindInput:
return "input"
case TaskKindSelf:
return "self"
default:
return "unknown"
}
}
// Task 是调度器的最小单位。
//
// M2 只承载"一份待处理的输入";M3 起把 TaskFrame(现场)挂上来,
@ -124,9 +137,11 @@ func effectiveLevel(t *Task) Level {
type SchedulerStats struct {
Enqueued uint64
Executed uint64
// Rejected 是因队列满而未入队的次数。M2 由泵入侧节流,正常为 0;
// 出现非 0 说明消费端长期慢于生产端。
// Rejected 是因队列满(或深度超限)而未被接纳的次数。
Rejected uint64
// Suspended / Resumed 是挂起与恢复的次数。
Suspended uint64
Resumed uint64
}
// SchedulerSnapshot 是调度器的原子快照。
@ -136,6 +151,33 @@ type SchedulerSnapshot struct {
PendingInterrupts []*Task
SuspendPool []*suspendedTask
Stats SchedulerStats
MaxSuspendDepth int
}
// schedulerStatus 把快照转成对外的状态 DTO(不暴露帧内容)。
func (a *Agent) schedulerStatus() sdk.SchedulerStatus {
if a.sched == nil {
return sdk.SchedulerStatus{}
}
snap := a.DumpScheduler()
out := sdk.SchedulerStatus{
ReadyQueueDepth: len(snap.Queue),
PendingInterrupts: len(snap.PendingInterrupts),
SuspendPool: len(snap.SuspendPool),
MaxSuspendDepth: snap.MaxSuspendDepth,
Enqueued: snap.Stats.Enqueued,
Executed: snap.Stats.Executed,
Rejected: snap.Stats.Rejected,
Suspended: snap.Stats.Suspended,
Resumed: snap.Stats.Resumed,
Preempted: snap.Stats.Suspended,
}
if snap.Running != nil {
out.Running = &sdk.SchedulerTask{
ID: snap.Running.ID, Level: int(snap.Running.Level), Kind: snap.Running.Kind.String(),
}
}
return out
}
type scheduler struct {
@ -342,6 +384,7 @@ func (s *scheduler) suspend(t *Task, f *TaskFrame) {
s.stats.Rejected++
}
s.suspendPool = append(s.suspendPool, &suspendedTask{Task: t, Frame: f})
s.stats.Suspended++
// 饥饿防护:抢占计数 +1(提升有效级)并记录冷却起点。
t.PreemptCount++
t.LastPreemptAt = time.Now()
@ -487,6 +530,7 @@ func (a *Agent) DumpScheduler() SchedulerSnapshot {
snap.Queue = append(snap.Queue, a.sched.queue...)
snap.PendingInterrupts = append(snap.PendingInterrupts, a.sched.pendingInterrupts...)
snap.SuspendPool = append(snap.SuspendPool, a.sched.suspendPool...)
snap.MaxSuspendDepth = a.sched.maxSuspendDepth
return snap
}
@ -576,6 +620,9 @@ func (a *Agent) executeNewTask(t *Task) {
if out == outcomeSuspended && f != nil {
a.sched.suspend(t, f)
a.publishEvent(events.EventScheduler, map[string]interface{}{
"action": "suspend", "task": t.ID, "level": int(t.Level),
})
return
}
a.sched.done(t)
@ -585,6 +632,12 @@ func (a *Agent) executeNewTask(t *Task) {
//
// 关键:不重建帧、不重跑 prepare 段——否则会重复提交上下文与事件。
func (a *Agent) resumeTask(t *Task, f *TaskFrame) {
a.sched.mu.Lock()
a.sched.stats.Resumed++
a.sched.mu.Unlock()
a.publishEvent(events.EventScheduler, map[string]interface{}{
"action": "resume", "task": t.ID, "level": int(t.Level),
})
defer func() {
if r := recover(); r != nil {
log.Printf("[agent] resume task#%d panic recovered: %v\n%s",