AskaEth
de1fa83b79
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
80 lines
2.9 KiB
Go
80 lines
2.9 KiB
Go
package loop
|
|
|
|
import (
|
|
"context"
|
|
"fmt"
|
|
|
|
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/agent"
|
|
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
|
|
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/trace"
|
|
)
|
|
|
|
// stepOutcome is the loop's internal view of one model step, normalized from the
|
|
// agent's *agent.StepResult (non-streaming) or the equivalent built by the
|
|
// streaming path.
|
|
type stepOutcome struct {
|
|
final bool
|
|
text string
|
|
toolCalls []agent.ToolCall
|
|
stopReason string
|
|
// modelStopReason is the provider's raw stop/finish reason for the model
|
|
// call (Anthropic max_tokens / end_turn / tool_use), recorded on the llm_call
|
|
// span so a truncation is visible even when the step itself ended normally.
|
|
modelStopReason string
|
|
// model is the routed model that served the call, known on the streaming path
|
|
// (which builds the request itself) and empty on the agent path. It is
|
|
// recorded on the llm_call span's log event.
|
|
model string
|
|
}
|
|
|
|
// step runs a single model call. When an OnStream callback is configured it
|
|
// drives the adapter's ChatStream directly, threading deltas to OnStream and
|
|
// applying the loop's guardrails; otherwise it delegates to the agent's RunStep,
|
|
// which runs the agent's own guardrails and records its own CallRecord. The whole
|
|
// call is wrapped in an llm_call span so the M9 panel sees each LLM invocation.
|
|
func (l *Loop) step(ctx context.Context, sessionID string, messages []model.Message) (*stepOutcome, error) {
|
|
_, llmSpan := l.span(ctx, "llm.call", trace.SpanKindLLM)
|
|
|
|
var (
|
|
out *stepOutcome
|
|
err error
|
|
)
|
|
// Muted (internal autonomous pass) bypasses streaming: it neither streams to
|
|
// the UI nor pays the streaming path's extra overhead. Normal chat streams.
|
|
if l.onStream != nil && !isMutedBroadcast(ctx) {
|
|
out, err = l.streamStep(ctx, messages)
|
|
} else {
|
|
out, err = l.agentStep(ctx, sessionID, messages)
|
|
}
|
|
if err != nil {
|
|
l.spanProp(llmSpan, "error", err.Error())
|
|
l.finishSpan(llmSpan, trace.SpanStatusError)
|
|
return nil, fmt.Errorf("loop: model step: %w", err)
|
|
}
|
|
|
|
l.spanProp(llmSpan, "stop_reason", out.stopReason)
|
|
if out.modelStopReason != "" {
|
|
l.spanProp(llmSpan, "model_stop_reason", out.modelStopReason)
|
|
}
|
|
l.spanProp(llmSpan, "tool_calls", len(out.toolCalls))
|
|
l.recordLLMDone(llmSpan, out)
|
|
l.finishSpan(llmSpan, trace.SpanStatusOK)
|
|
return out, nil
|
|
}
|
|
|
|
// agentStep delegates a non-streaming step to the agent, which runs input and
|
|
// output guardrails, calls the model and decodes any tool calls.
|
|
func (l *Loop) agentStep(ctx context.Context, sessionID string, messages []model.Message) (*stepOutcome, error) {
|
|
sr, err := l.agent.RunStep(ctx, sessionID, messages)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("loop: agent step: %w", err)
|
|
}
|
|
return &stepOutcome{
|
|
final: sr.Final,
|
|
text: sr.Text,
|
|
toolCalls: sr.ToolCalls,
|
|
stopReason: sr.StopReason,
|
|
modelStopReason: sr.ModelStopReason,
|
|
}, nil
|
|
}
|