模型保持占用(held)机制:回合结束不再立即释放模型,进入 held 租约(config.model_hold 可配:idle_release_ms 默认3分钟——页面无人观看即释放;active_max_ms 默认30分钟——有人观看的保持上限;sweep_ms 巡检间隔);held 期间追问直接复用已绑模型免排队(delivered=live);sessions 表迁移新增 held_at 列;sse 增加 clientCount 判定观看者;前端新增保持中状态标签/徽章色,终止入口扩展至 held(主动放弃占用);smoke/README/示例配置同步更新
This commit is contained in:
@@ -19,3 +19,9 @@ agent:
|
||||
temperature: 1.0
|
||||
max_tool_iterations: 20
|
||||
max_retries: 1
|
||||
|
||||
# 冒烟测试用短租约,验证 held → 自动释放
|
||||
model_hold:
|
||||
idle_release_ms: 1500
|
||||
active_max_ms: 8000
|
||||
sweep_ms: 300
|
||||
|
||||
Reference in New Issue
Block a user