mirror of
https://gitcode.com/JianFeeeee/HomeAgent.git
synced 2026-09-22 18:08:04 +00:00
fix(obs): 抢占日志改在判决点打(修自伤)+ scheduler 事件接进 SSE
## 修我上一版的自伤
上一版把 preempt 日志打在 `executeNewTask`,但那时 `nextRef` 已经把
`s.running` 换成了抢占者自己,于是输出成了
preempt start: task#2 ... -> victim task#2 (cli)
victim 打印的是入侵者本人。判据必须落在 `registerInterrupt`——那一刻
running 还是真正的受害者。改为在抢占判决点打:
[agent] preempt: task#2 class=interrupt level=3 from cli (L3) preempts task#1 class=queued level=0 (qq)
## 补上「入队而非抢占」的日志
中断到了却没生效,此前完全不可解释。现在两种成因分开写:
[agent] interrupt queued: ... vs ... (qq) — running in critical section; queue=N
[agent] interrupt queued: ... vs ... (qq) — preempt cooldown; queue=N
[agent] interrupt queued: ... vs ... (qq) — level insufficient; queue=N
没有这条,`interrupt from X` 打过之后任务为什么没让位就只能猜。
## scheduler 事件接进 SSE
`EventScheduler` 此前既不在 `handler_chat.go` 的 subTypes、也没有任何订阅者
(全仓 grep 零命中)——内核里 suspend/resume 只 publishEvent,于是事件发出来
就掉地上,对内对外都不可见。加进 subTypes 后前端/客户端能看到抢占链。
## 验证
- preempt_logging_test.go 增一条:victim 与入侵者必须是不同来源(qq vs cli),
且排队输入对排队任务 `canPreempt` 必为假。
- 实测输出含 `cli (L3) preempts task#1 class=queued level=0 (qq)`。
- 全量 `go test ./internal/... ./cmd/...` 与 `go vet ./internal/...` 全绿。
This commit is contained in:
@ -62,3 +62,28 @@ func TestSourceOf_SelfAndNilAreSafe(t *testing.T) {
|
||||
t.Fatalf("self task -> %q", got)
|
||||
}
|
||||
}
|
||||
|
||||
// 抢占日志必须把「入侵者」与「受害者」分开写,且受害者是**真正在跑的那个**。
|
||||
//
|
||||
// 镇的是一个自伤:第一版把日志打在 executeNewTask 里,而那时 nextRef 已经把
|
||||
// s.running 换成了抢占者自己,于是日志写成 "victim = 入侵者"(实测输出过
|
||||
// `preempt start: task#2 ... -> victim task#2 (cli)`)。判据必须落在
|
||||
// registerInterrupt —— 那一刻 running 还没被换。
|
||||
func TestSourceOf_DistinctVictimAndIntruder(t *testing.T) {
|
||||
if sourceOf(newInputTask(&agentIO.InputEvent{Source: "qq"}), nil) != "qq" {
|
||||
t.Fatal("queued qq 任务的来源应为 qq")
|
||||
}
|
||||
it := newInterruptTask(&agentIO.InputEvent{Source: "homeagent-mail-bridge"}, LevelMessage)
|
||||
if got := sourceOf(it, nil); got != "homeagent-mail-bridge" {
|
||||
t.Fatalf("中断来源应为 homeagent-mail-bridge,实际 %q", got)
|
||||
}
|
||||
if it.Class != TaskInterrupt || it.Level != LevelMessage {
|
||||
t.Fatalf("中断任务应 class=interrupt level=2,实际 %v/%v", it.Class, it.Level)
|
||||
}
|
||||
// 邮件若走排队注入则会变成 TaskQueued —— 那种情况下 canPreempt 必为假。
|
||||
q := &Task{Class: TaskQueued, Kind: TaskKindInput,
|
||||
Event: &agentIO.InputEvent{Source: "homeagent-mail-bridge"}}
|
||||
if canPreempt(q, newInputTask(&agentIO.InputEvent{Source: "qq"})) {
|
||||
t.Fatal("排队输入永远不得抢占另一个排队任务")
|
||||
}
|
||||
}
|
||||
|
||||
@ -706,11 +706,29 @@ func (s *scheduler) registerInterrupt(t *Task) bool {
|
||||
s.preemptArmed = true
|
||||
s.preemptLevel = t.Level
|
||||
s.stats.bumpInterruptLevel(&s.stats.PreemptsByLevel, t.Level)
|
||||
// 抢占判决点:此刻 running 还是真正的受害者,记下来才能回答
|
||||
// 「我的任务被谁打断了」。放在 executeNewTask 里打是错的——
|
||||
// 那时 nextRef 已把 running 换成抢占者自己,日志会写成
|
||||
// "victim = 入侵者自己"(实测过这个错误输出)。
|
||||
log.Printf("[agent] preempt: %s from %s (L%d) preempts %s (%s)",
|
||||
describeTask(t), sourceOf(t, nil), int(t.Level),
|
||||
describeTask(running), sourceOf(running, nil))
|
||||
}
|
||||
}
|
||||
}
|
||||
if !arm {
|
||||
s.enqueueInterruptLocked(t)
|
||||
// 入队而不是抢占:这也是要能看见的(否则「中断明明到了却没生效」无从解释)。
|
||||
// 两种成因分开写:临界区 vs 级别不够/冷却期。
|
||||
reason := "level insufficient"
|
||||
if critical {
|
||||
reason = "running in critical section"
|
||||
} else if canPreempt(t, running) && running != nil && !running.LastPreemptAt.IsZero() && time.Since(running.LastPreemptAt) < preemptCooldown {
|
||||
reason = "preempt cooldown"
|
||||
}
|
||||
log.Printf("[agent] interrupt queued: %s from %s (L%d) vs %s (%s) — %s; queue=%d",
|
||||
describeTask(t), sourceOf(t, nil), int(t.Level),
|
||||
describeTask(running), sourceOf(running, nil), reason, s.interruptCountLocked())
|
||||
}
|
||||
s.mu.Unlock()
|
||||
|
||||
@ -1164,17 +1182,6 @@ func (a *Agent) executeNewTask(t *Task) {
|
||||
var f *TaskFrame
|
||||
var out stepOutcome = outcomeDone
|
||||
|
||||
if t.Class == TaskInterrupt {
|
||||
// 让位之前先记一笔「谁将要被打断」——这是排查「任务被莫名打断」的锚点。
|
||||
// victim 从调度器取:此刻 s.running 还是被抢占者本身(suspend 里才清)。
|
||||
a.sched.mu.Lock()
|
||||
victim := describeTask(a.sched.running)
|
||||
victimSrc := sourceOf(a.sched.running, nil)
|
||||
a.sched.mu.Unlock()
|
||||
log.Printf("[agent] preempt start: %s from %s (level=%d) -> victim %s (%s)",
|
||||
describeTask(t), sourceOf(t, nil), int(t.Level), victim, victimSrc)
|
||||
}
|
||||
|
||||
func() {
|
||||
defer func() {
|
||||
if r := recover(); r != nil {
|
||||
|
||||
@ -728,7 +728,13 @@ func (h *Handler) handleChatEvents(w http.ResponseWriter, r *http.Request) {
|
||||
}
|
||||
}
|
||||
|
||||
subTypes := []string{"agent_output", "reasoning", "agent_error", "tool_call", "stage", "agent_llm_chain", "terminal_output"}
|
||||
// scheduler:输入调度器的状态变更(抢占 / 挂起 / 恢复)。
|
||||
//
|
||||
// 为何必须转发:内核里 suspend/resume 只发事件(见 core/scheduler.go 的
|
||||
// executeNewTask/resumeTask),而 `scheduler` 曾既不在本列表里、也没有任何
|
||||
// 订阅方—— 事件发出来就掉地上。后果是「我的任务被谁打断了」对内对外都
|
||||
// 不可见,排查时只能按时序猜。这里补上对外那一半。
|
||||
subTypes := []string{"agent_output", "reasoning", "agent_error", "tool_call", "stage", "agent_llm_chain", "terminal_output", "scheduler"}
|
||||
// token 级流式增量事件:实时转发给浏览器做逐 token 渲染。
|
||||
// 不进 sseEventRing —— 断线重连只重放聚合事件(最终真相),
|
||||
// 避免重放 delta 与聚合内容重复追加。
|
||||
|
||||
Reference in New Issue
Block a user