Files
Cyrene_v2/internal/runtime/loop/spanlog.go
T
AskaEth de1fa83b79 feat(orchestration,loop,flow): 接线子代理编排 — 派发入口 + 分支归父会话 + 节点日志(含脱敏)
修的是「文档说 ✅、生产不可达」:此前 Scheduler.Spawn 只有测试调用方、无 bus 入口、无派发工具;Span.Record 无生产调用方 → 泳道画不出、节点日志恒空。
- 入口:subagent_spawn 工具(internal/runtime/orchestration/spawntool.go,同包便于用真 Scheduler/loop/Tracer/Flow 做集成测试;注册在 cmd/cyrene/subagent_tool.go 照 terminal_tool 范式)。默认 owner-only,不设 AllowInGroup/AllowInStranger;安全仍由 restrictSchemas + applyToolSecurityGate 双层保证,未绕代审
- 父会话归属:AgentSpec.ParentSession(请求语义)→ Scheduler 转运行期 ctx 标记 loop.WithFlowBranch → 分支 span 以父 session 发布;**allow 谓词一行未改**;单图语义靠「分支 Run 不调 StartRound/EndRound」+ FlowPublisher 不新建图;未归属 spawn 一张图都不发(有测试锁)
- 前台最多等 2s(对齐 task.autoBackgroundThreshold)→ 合并结果作工具结果;转后台则广播 bus subagent.result(design/05 §9 回填 hook);新增 Scheduler.Wait(ids) 按批等待(Join 会串并发批次,故不重用)
- 子代理跑在 WithoutCancel(ctx):保留 span 归属、不被工具超时/回合结束顺手杀掉;生命周期由 PerAgentTimeout + CancelAll 约束;内部 WithMutedBroadcast 静音(不进前台 stream)
- 节点日志:loop/spanlog.go 新增 llm.done / tool.done / subagent.done;工具参数**只记键名不记值**;仅步骤真正结束才追加(跳过/被拒路径不发,不伪造)
- 脱敏(红线):新增 trace/redact.go(形状规则 sk-/Bearer/JWT/key:value + 键名规则 api_key/token/authorization/password…,递归嵌套、不改调用方事件),graph.go 发布前必经;**顺带修掉一个既有泄漏**:result_summary prop 原先未脱敏(工具结果里的密钥会进图),由图级变异测试抓出
- 测试:新增 22 个(工具级 / 图级 / 日志级),含 4 次变异验证(去掉 WithFlowBranch、去掉 record*、还原未脱敏、去掉静音 各自 FAIL);-race 过 orchestration+trace+loop
- 已知限制:真实模型端到端未验(无凭据/不起服务);工具白名单是 design gap 未做;subagent.result 暂无消费者(design/15 未实现);后台子代理跨轮次边界迟到 span 会丢泳道根;PerAgentTimeout=10min 待确认
2026-10-03 19:39:38 +08:00

119 lines
3.8 KiB
Go

package loop
import (
"sort"
"time"
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/agent"
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/trace"
)
// Log event types recorded on a span so the flow graph's node log[] shows what
// happened at each step (design/19 §1 节点实时日志). An event is appended only when
// the step really finished, so the log never announces work that did not run.
const (
// EventLLMDone is recorded on an llm_call span once the model call returned.
EventLLMDone = "llm.done"
// EventToolDone is recorded on a tool_call span once the tool call settled.
EventToolDone = "tool.done"
)
// maxLogSummaryRunes bounds a log value before publication. The graph redacts and
// truncates again on the way out; this only keeps an individual span small.
const maxLogSummaryRunes = 240
// recordLLMDone appends a desensitized summary of a finished model call to its
// span: the configured model, both stop reasons, how many tool calls were
// requested and how long the call took. No prompt or completion text is recorded,
// so an llm node's log can never carry conversation content.
func (l *Loop) recordLLMDone(span *trace.Span, out *stepOutcome) {
if span == nil || out == nil {
return
}
props := map[string]any{
"stop_reason": out.stopReason,
"tool_calls": len(out.toolCalls),
"elapsed_ms": elapsedMS(span),
}
if m := l.modelName(out); m != "" {
props["model"] = m
}
if out.modelStopReason != "" {
props["model_stop_reason"] = out.modelStopReason
}
_ = span.Record(trace.Event{Type: EventLLMDone, Props: props})
}
// recordToolDone appends a desensitized summary of one tool call to its span: the
// tool name, whether it succeeded, the argument NAMES (never their values), the
// error kind on failure, and a truncated result summary. Argument values are
// deliberately omitted — a tool argument is exactly where a credential, a message
// body or a file path with a secret would appear — and everything recorded here is
// scrubbed again by the graph before publication.
func (l *Loop) recordToolDone(span *trace.Span, tc agent.ToolCall, out ToolOutcome) {
if span == nil {
return
}
props := map[string]any{
"tool": tc.Name,
"ok": out.Err == nil,
"args_keys": argKeys(tc.Args),
"elapsed_ms": elapsedMS(span),
}
if out.Err != nil {
if out.ErrKind != "" {
props["error_kind"] = out.ErrKind
}
props["error"] = trimSummary(out.Err.Error())
} else {
props["result_summary"] = trimSummary(formatToolResult(out.Result))
}
_ = span.Record(trace.Event{Type: EventToolDone, Props: props})
}
// modelName reports the model that served a step: the routed request model when
// the streaming path knows it, else the loop's configured model. It is empty when
// neither is available.
func (l *Loop) modelName(out *stepOutcome) string {
if out.model != "" {
return out.model
}
if l.agent == nil {
return ""
}
return l.agent.Model()
}
// elapsedMS reports how long a span has been open so far. A span's Duration is
// only stamped by Finish, so a log event recorded before Finish measures it here.
func elapsedMS(span *trace.Span) int64 {
if span == nil || span.StartedAt.IsZero() {
return 0
}
return time.Since(span.StartedAt).Milliseconds()
}
// argKeys returns a tool call's argument names in sorted order. Only the names are
// logged; the values stay out of the trace (see recordToolDone).
func argKeys(args map[string]any) []string {
if len(args) == 0 {
return nil
}
keys := make([]string, 0, len(args))
for k := range args {
keys = append(keys, k)
}
sort.Strings(keys)
return keys
}
// trimSummary shortens a log value to maxLogSummaryRunes runes without splitting a
// multi-byte rune.
func trimSummary(s string) string {
r := []rune(s)
if len(r) <= maxLogSummaryRunes {
return s
}
return string(r[:maxLogSummaryRunes]) + "…"
}