fix(engine): 种子意图任务误触发首轮 planner(Frontier 只数 open 与 worker claim 竞态)
Run 里「跳过首轮」依赖种子意图让 frontier 非空,但 Frontier 只数 state='open', 而 Run 先起 worker、后查 frontier——worker 抢先 claim 种子意图(open→running)后, Frontier(open) 变空 → 误发 t.Notify() → 首轮 planner 立刻触发(日志 "edge 触发")。 改为 HasActiveIntent(open 或 running):种子无论 open 还是已被 claim 成 running 都算 「有活干」,一律跳过首轮。首轮延后到种子 worker 收工(NotifyDone)/出 finding/心跳。 顺带修好重启恢复时只剩 running 意图也会误 kick 的情况。
This commit is contained in:
1 parent
47ba58bda6
commit
4835cb57bd
2 files changed
+18
-4
No files matched your search
@@ -381,6 +381,18 @@ ORDER BY priority DESC, id ASC LIMIT $2`, s.expID, limit)
|
||||
return scanNodes(rows)
|
||||
}
|
||||
|
||||
// HasActiveIntent reports whether this exploration has any intent still in play
|
||||
// (state open OR running). Used to decide whether to kick the FIRST planner round:
|
||||
// a seeded task's intent may already have been claimed (open→running) by a worker
|
||||
// before the check, so counting only 'open' (Frontier) races with worker claim and
|
||||
// wrongly kicks a redundant first round. Counting open+running is race-proof.
|
||||
func (s *ExplorationStore) HasActiveIntent() (bool, error) {
|
||||
var exists bool
|
||||
err := s.db.QueryRow(`SELECT EXISTS(SELECT 1 FROM exploration_nodes
|
||||
WHERE exploration_id=$1 AND kind='intent' AND state IN ('open','running'))`, s.expID).Scan(&exists)
|
||||
return exists, err
|
||||
}
|
||||
|
||||
// ClaimIntent atomically moves an open intent to running. Returns true if claimed.
|
||||
func (s *ExplorationStore) ClaimIntent(id int64, owner string) (bool, error) {
|
||||
res, err := s.db.Exec(`UPDATE exploration_nodes SET state='running', owner=$1
|
||||
|
||||
+6
-4
@@ -433,10 +433,12 @@ func (e *Engine) Run(ctx context.Context, t *Task) {
|
||||
go e.workerLoop(ctx, t, fmt.Sprintf("work#%d", i+1))
|
||||
}
|
||||
e.startDeadlineCoordinator(ctx, t) // 任务级超时定时器(仅 timeout>0;去重)
|
||||
// 仅在 frontier 为空时才 kick 首轮规划:普通任务(空 frontier)照常触发;带种子意图的
|
||||
// 任务 frontier 已非空 → 跳过首轮 planner,worker 直接领种子意图开跑,跑完由 NotifyDone
|
||||
// 唤醒 planner。重启自动恢复时 frontier 通常也非空 → 顺带避免白烧一轮。
|
||||
if fr, _ := t.Store.Frontier(1); len(fr) == 0 {
|
||||
// 仅在「完全没有活动意图(open+running)」时才 kick 首轮规划。带种子意图的任务:种子已
|
||||
// 是 open,或已被上面刚起的 worker 抢先 claim 成 running——两种都算「有活干」,一律跳过
|
||||
// 首轮 planner,worker 直接领种子意图开跑,跑完由 NotifyDone/心跳唤醒 planner。
|
||||
// ⚠️ 不能用 Frontier(只数 open):worker 领取(open→running)与本检查存在竞态,会误 kick。
|
||||
// 重启自动恢复时也可能只剩 running 意图,同样应跳过。
|
||||
if has, _ := t.Store.HasActiveIntent(); !has {
|
||||
t.Notify() // kick the first planning round (acted on once LLM is ready)
|
||||
}
|
||||
}
|
||||
|
||||
Reference in new issue
Block a user