feat(llm): LLM 配置轮询(故障转移),某个配置没额度/失效时自动切下一个
默认关闭。开启后「未指定模型」的 agent 在当前配置不可用(402 余额不足 /
401 key 失效 / 429 限流 / 5xx)时自动切到下一个配置;agent 绑定或任务 pin
了配置的默认不轮询(绑定即独占,可由 llm_pool_bind_fallback 打开兜底)。
链序 = 激活配置 + 其余按 priority DESC,排除 pool_exclude 的。只给
llm_profiles 加两列,不引入关联表,所以 agents/tasks/conversations 对
llm_profile_id 的四处引用一行未改。
核心约束:只在本次调用尚未 yield 出任何事件时切换成员。已经吐出
text_delta / tool_use_start 后中途断流一律原样上抛(交给 agentcore 的
resume),否则重发同一 request 会重复输出、污染会话历史。SDK 的 doStream
保证 4xx/5xx 在读 body 前返回,所以目标场景都落在这个安全窗口内。
context.Canceled / DeadlineExceeded 绝不触发切换——那是用户取消或任务
超时,误判会白烧备用 key 并污染 a6bd180 的终止原因诊断。400 也不切
(请求本身有问题,换模型结果一样)。
熔断:402/401/403/404 一次即熔断,429/5xx/网络错误连续 3 次;冷却
60s→5min→30min 退避,到期半开试探,成功清零。Registry 是进程级的,独立
于 provider 缓存,重建链路不丢状态,且全进程共享(A 任务撞见的 402,B 任务
立刻知道)。状态旁路落库,重启只恢复尚未到期的冷却窗口。
上下文窗口装不下当前请求的成员直接跳过(必然 400),不把主力模型的窗口
压到全链最小值。全员熔断时仍探链首,不让引擎停摆。
Pool 包在 llmrec 外面、每个成员自带 recorder,失败调用会以 status=error
+ 正确的 profile 名进 llm_records;切换/熔断走 log.Printf,自动出现在
系统日志页。
顺带修两个既有问题:
- DecomposeGoals 原本自己 NewProvider(),绕过限速器/录制/轮询,且每个
任务新建一个实例——改为由 server 传入共享 provider
- saveLLMConfig 会把 priority/pool_exclude 重置为零值
前端:LLM 页新增轮询卡片(总开关/兜底开关/实际链路顺序/健康徽章/冷却
倒计时/立即恢复),配置编辑区加优先级与「不参与轮询」。
只有一个配置时链长度为 1,poolChain 返回 nil 走裸 provider,行为与开关
打开前完全一致。设计见 docs/LLM轮询设计.md。
This commit is contained in:
1 parent
89c5e9b26f
commit
0c653a2135
19 files changed
+1750
-50
No files matched your search
+8
-3
@@ -7,6 +7,7 @@ import (
|
||||
|
||||
"github.com/Autumn-27/artex/db"
|
||||
"github.com/Autumn-27/norma/agentcore"
|
||||
"github.com/Autumn-27/norma/llm"
|
||||
acperm "github.com/Autumn-27/norma/permission"
|
||||
actool "github.com/Autumn-27/norma/tool"
|
||||
)
|
||||
@@ -63,6 +64,11 @@ type GoalSpec struct {
|
||||
// no provider is configured or the call yields nothing — the caller then falls
|
||||
// back to a rule-based split so goal nodes always exist.
|
||||
//
|
||||
// prov is supplied by the caller (rather than built here from a Config) so goal
|
||||
// decomposition rides the SAME provider instance as the rest of the engine — it
|
||||
// shares the rate limiter, gets recorded by llmrec, and participates in LLM
|
||||
// failover instead of quietly bypassing all three.
|
||||
//
|
||||
// desc is the task's free-text description (背景:靶标范围/flag 数量/交战说明等).
|
||||
// It is fed alongside the goal so the decomposer no longer splits blind — the
|
||||
// prompt still forbids inventing anything the two texts don't state.
|
||||
@@ -77,9 +83,8 @@ type GoalSpec struct {
|
||||
// straight into it (the same managed tool the main agent uses to add goals at
|
||||
// runtime). The returned specs are read back from the store so callers can emit
|
||||
// per-goal activity and detect the "LLM produced nothing" case for their fallback.
|
||||
func DecomposeGoals(ctx context.Context, c Config, dataDir, goalText, desc string, as *db.AssetStore, ts *db.ExplorationStore, taskID int64, emit func(db.Activity)) []GoalSpec {
|
||||
prov, err := c.NewProvider()
|
||||
if err != nil {
|
||||
func DecomposeGoals(ctx context.Context, prov llm.Provider, dataDir, goalText, desc string, as *db.AssetStore, ts *db.ExplorationStore, taskID int64, emit func(db.Activity)) []GoalSpec {
|
||||
if prov == nil {
|
||||
return nil
|
||||
}
|
||||
// worker="goals" tags the goal nodes' provenance; ts/taskID let set_goals link
|
||||
|
||||
Reference in new issue
Block a user