Version: 0.9.20.dev.260415
后端: 1. 修复 query_available_slots section_from/section_to 错误覆盖 duration 并使用精确匹配而非范围包含 - 更新backend/newAgent/tools/schedule/read_filter_tools.go:移除 span = exactTo - exactFrom + 1 对 duration 的覆盖;matchSectionRange 从精确匹配改为范围包含语义(slotStart < exactFrom || slotEnd > exactTo) 2. Execute 上下文窗口从硬编码裁剪改造为 80k token 动态预算 + LLM滚动压缩 - 基础设施层:AgentChat 新增 compaction 三个持久化字段,dao 新增 CRUD,Redis 新增缓存;pkg 新增 ExecuteTokenBudget常量、ExecuteTokenBreakdown 结构体、CheckExecuteTokenBudget 预算检查函数 - prompt 层:新建 compact_msg1.go / compact_msg2.go 分别实现msg1(历史对话)和 msg2(ReAct Loop)的 LLM 压缩;execute_context.go 移除 msg1 的 1400 字符/30 轮/120 字符三重裁剪和 msg2 的 8 条窗口限制,改为全量加载 - node 层:新建 execute_compact.go(compactExecuteMessagesIfNeeded:预算检查 → msg1 优先压缩 → msg2 兜底 → SSE 通知 → token 分布持久化);execute.go ReAct 循环插入 compact 调用 - 服务/API 层:AgentGraphDeps / AgentService 新增 CompactionStore 注入链路;新增 GET /api/v1/agent/context-stats 查询接口 - 启动层:cmd/start.go 注入 agentRepo 为 CompactionStore 3. 新增 Execute Context Compaction 决策报告 - 新建docs/功能决策记录/Execute_Context_Compaction_决策记录.md 前端:无 仓库:无
This commit is contained in:
@@ -21,6 +21,12 @@ const (
|
||||
SessionWindowMin = 32
|
||||
SessionWindowMax = 4096
|
||||
SessionWindowBuffer = 2
|
||||
|
||||
// ---- Execute Context Compaction 预算 ----
|
||||
// Execute 阶段 prompt 总 token 上限
|
||||
ExecuteTokenBudget = 80000
|
||||
// msg0 + msg3 固定开销 + 安全余量
|
||||
ExecuteReserveTokens = 8000
|
||||
)
|
||||
|
||||
// MaxContextTokensByModel 返回指定模型的最大上下文 token。
|
||||
@@ -144,3 +150,42 @@ func CalcSessionWindowSize(trimmedHistoryLen int) int {
|
||||
func isCJK(r rune) bool {
|
||||
return unicode.Is(unicode.Han, r) || unicode.Is(unicode.Hiragana, r) || unicode.Is(unicode.Katakana, r) || unicode.Is(unicode.Hangul, r)
|
||||
}
|
||||
|
||||
// ExecuteTokenBreakdown 是 Execute 阶段四条消息的 token 分布。
|
||||
type ExecuteTokenBreakdown struct {
|
||||
Msg0 int `json:"msg0"`
|
||||
Msg1 int `json:"msg1"`
|
||||
Msg2 int `json:"msg2"`
|
||||
Msg3 int `json:"msg3"`
|
||||
Total int `json:"total"`
|
||||
Budget int `json:"budget"`
|
||||
}
|
||||
|
||||
// EstimateExecuteMessagesTokens 估算 Execute 四条消息的 token 分布。
|
||||
func EstimateExecuteMessagesTokens(msg0, msg1, msg2, msg3 string) ExecuteTokenBreakdown {
|
||||
b := ExecuteTokenBreakdown{
|
||||
Msg0: EstimateTextTokens(msg0),
|
||||
Msg1: EstimateTextTokens(msg1),
|
||||
Msg2: EstimateTextTokens(msg2),
|
||||
Msg3: EstimateTextTokens(msg3),
|
||||
Budget: ExecuteTokenBudget,
|
||||
}
|
||||
b.Total = b.Msg0 + b.Msg1 + b.Msg2 + b.Msg3
|
||||
return b
|
||||
}
|
||||
|
||||
// CheckExecuteTokenBudget 检查是否超出 token 预算。
|
||||
// 返回 breakdown、是否超限、是否需要压缩 msg1、是否需要压缩 msg2。
|
||||
func CheckExecuteTokenBudget(msg0, msg1, msg2, msg3 string) (breakdown ExecuteTokenBreakdown, overBudget bool, needCompactMsg1 bool, needCompactMsg2 bool) {
|
||||
breakdown = EstimateExecuteMessagesTokens(msg0, msg1, msg2, msg3)
|
||||
overBudget = breakdown.Total > ExecuteTokenBudget
|
||||
if !overBudget {
|
||||
return
|
||||
}
|
||||
// msg1 超过可用预算的一半时需要压缩
|
||||
available := ExecuteTokenBudget - ExecuteReserveTokens
|
||||
needCompactMsg1 = breakdown.Msg1 > available/2
|
||||
// 压缩 msg1 后仍超限,则压缩 msg2
|
||||
needCompactMsg2 = (breakdown.Total - breakdown.Msg1 + available/4) > ExecuteTokenBudget
|
||||
return
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user