bc7600625d
接了 checkpoint 后端时,审批节点从「阻塞 goroutine 等 5min」改为持久化中断:
首次执行发待审 + 置 waiting + compose.Interrupt → compose 把整图状态(含 board)
落进 checkpoint store 并返回中断错误 → Handle 释放 goroutine、任务停在 waiting,
不收尾不评测不判 done。抗 dispatcher 重启。
- orchestrator: 新增 errInterrupted 哨兵 + checkpoints 字段 + SetCheckpoints
setter(沿用 guard/usageSink 的 setter 注入,不动构造签名);Handle 识别
errInterrupted → 释放 goroutine、保留 waiting、SSE 流不关。
- compose_compiler: 编译挂 WithCheckPointStore + WithGraphName("root"),Invoke
带 WithCheckPointID(task_id);审批节点接 checkpoint 时改走专用
approvalInterruptLambda(须把中断错误作节点返回值上抛,泛型 lambda 会吞掉);
ExtractInterruptInfo 识别中断 → 上抛 errInterrupted。
- graph.go: 抽出 approvalSummary / applyApprovalDecision,阻塞式与中断式审批共用,
杜绝两路文案/语义漂移。
双路径并存:未接 checkpoint 后端(store=nil)维持阻塞模型,行为不变;main 暂不
接,生产保持阻塞,待增量3 resume 闭环补齐再打开(建在 flag 后)。
测试:中断半边端到端——errInterrupted + 置 waiting + checkpoint 落 KV(key=task_id)
+ 下游 agent 不执行。既有阻塞式审批/等价测试全绿。
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
172 lines
6.8 KiB
Go
172 lines
6.8 KiB
Go
package eino
|
||
|
||
import (
|
||
"context"
|
||
"encoding/json"
|
||
"errors"
|
||
"strings"
|
||
"testing"
|
||
"time"
|
||
|
||
"github.com/sundynix/sundynix-dispatcher/internal/harness"
|
||
"github.com/sundynix/sundynix-dispatcher/internal/llm"
|
||
"github.com/sundynix/sundynix-shared/contract"
|
||
)
|
||
|
||
// echoLLM 回显最后一条 user 消息内容(确定性),便于两条执行路径逐字对比。
|
||
func echoLLM() *fakeLLM {
|
||
return &fakeLLM{
|
||
ready: true,
|
||
stream: func(m []llm.ChatMessage) string {
|
||
for i := len(m) - 1; i >= 0; i-- {
|
||
if m[i].Role == "user" {
|
||
return "ANS:" + m[i].Content
|
||
}
|
||
}
|
||
return "ANS:"
|
||
},
|
||
}
|
||
}
|
||
|
||
func runBoth(t *testing.T, graph string) (interp, comp string) {
|
||
t.Helper()
|
||
task := &contract.Task{ID: "t_eq", Graph: []byte(graph)}
|
||
|
||
o1 := &Orchestrator{pool: echoLLM(), breaker: harness.NewCircuitBreaker(), sink: &fakeSink{}}
|
||
a1, _, err := o1.runGraph(context.Background(), task, &execTracer{})
|
||
if err != nil {
|
||
t.Fatalf("runGraph: %v", err)
|
||
}
|
||
o2 := &Orchestrator{pool: echoLLM(), breaker: harness.NewCircuitBreaker(), sink: &fakeSink{}}
|
||
a2, _, err := o2.runComposeGraph(context.Background(), task, &execTracer{})
|
||
if err != nil {
|
||
t.Fatalf("runComposeGraph: %v", err)
|
||
}
|
||
return a1, a2
|
||
}
|
||
|
||
// rejectWaiter 是恒拒绝的 HITL 审批替身(fail-safe 路径回归用)。
|
||
type rejectWaiter struct{ note string }
|
||
|
||
func (r *rejectWaiter) WaitApproval(context.Context, string, time.Duration) (*contract.ApprovalDecision, error) {
|
||
return &contract.ApprovalDecision{Approved: false, Note: r.note}, nil
|
||
}
|
||
|
||
// TestComposeApprovalRejectStopsDownstream 钉死 compose 路径的 HITL 审批:
|
||
// 审批拒绝须置 errRejected、产出拒绝语,且下游 agent 绝不执行(否则等于审批形同虚设)。
|
||
// 这是翻默认前最高危的回归点——execDSLNode 一度没有 approval 分支会把审批节点当未识别跳过。
|
||
func TestComposeApprovalRejectStopsDownstream(t *testing.T) {
|
||
graph := `{"version":"1","nodes":[
|
||
{"id":"in","kind":"input","config":{"text":"hi"}},
|
||
{"id":"ap","kind":"approval","config":{"title":"上线审批"}},
|
||
{"id":"a","kind":"agent","config":{"system":"机密操作"}}
|
||
],"edges":[
|
||
{"source":"in","target":"ap"},{"source":"ap","target":"a"}
|
||
]}`
|
||
o := &Orchestrator{pool: echoLLM(), breaker: harness.NewCircuitBreaker(),
|
||
sink: &fakeSink{}, approval: &rejectWaiter{note: "不批"}}
|
||
ans, refs, err := o.runComposeGraph(context.Background(),
|
||
&contract.Task{ID: "t_ap", Graph: []byte(graph)}, &execTracer{})
|
||
if !errors.Is(err, errRejected) {
|
||
t.Fatalf("审批拒绝应返回 errRejected,got %v", err)
|
||
}
|
||
if !strings.Contains(ans, "已被拒绝") {
|
||
t.Fatalf("应产出拒绝语,got %q", ans)
|
||
}
|
||
if strings.Contains(ans, "ANS:") {
|
||
t.Fatalf("审批拒绝后下游 agent 不应执行,但成稿含 agent 产出: %q", ans)
|
||
}
|
||
if refs != nil {
|
||
t.Fatalf("拒绝终态 refs 应为 nil,got %v", refs)
|
||
}
|
||
}
|
||
|
||
// TestComposeApprovalInterruptCheckpoints 钉死中断式 HITL(接了 checkpoint 后端):
|
||
// 审批节点应 compose.Interrupt —— 返回 errInterrupted、置 waiting、checkpoint 落盘、下游 agent 不执行。
|
||
// 这是「释放 goroutine + 抗重启」的中断半边;恢复半边(resume)由增量3 覆盖。
|
||
func TestComposeApprovalInterruptCheckpoints(t *testing.T) {
|
||
graph := `{"version":"1","nodes":[
|
||
{"id":"in","kind":"input","config":{"text":"hi"}},
|
||
{"id":"ap","kind":"approval","config":{"title":"上线审批"}},
|
||
{"id":"a","kind":"agent","config":{"system":"机密操作"}}
|
||
],"edges":[
|
||
{"source":"in","target":"ap"},{"source":"ap","target":"a"}
|
||
]}`
|
||
kv := newMemKV()
|
||
st := &fakeStatus{}
|
||
fs := &fakeSink{}
|
||
o := &Orchestrator{pool: echoLLM(), breaker: harness.NewCircuitBreaker(), sink: fs, status: st}
|
||
o.SetCheckpoints(kv) // 接 checkpoint 后端 → 审批走中断模型
|
||
|
||
task := &contract.Task{ID: "t_intr", Graph: []byte(graph)}
|
||
_, _, err := o.runComposeGraph(context.Background(), task, &execTracer{})
|
||
|
||
if !errors.Is(err, errInterrupted) {
|
||
t.Fatalf("审批应中断并返回 errInterrupted,got %v", err)
|
||
}
|
||
if st.last() != contract.TaskWaiting {
|
||
t.Fatalf("中断应置 waiting,got %q", st.last())
|
||
}
|
||
if _, ok, _ := kv.Get(context.Background(), task.ID); !ok {
|
||
t.Fatalf("compose 应把图状态落进 checkpoint(key=task_id),但 KV 未命中")
|
||
}
|
||
if strings.Contains(fs.text(), "ANS:") {
|
||
t.Fatalf("审批中断后下游 agent 不应执行,但 sink 含 agent 产出: %q", fs.text())
|
||
}
|
||
}
|
||
|
||
// TestComposeReturnsRefs 钉死 compose 路径回传检索来源——曾被硬写成 nil,导致忠实度评测静默失效。
|
||
// 与 runGraph 同图对照:两路径都应回流含检索片段的 refs(喂 grounded judge)。
|
||
func TestComposeReturnsRefs(t *testing.T) {
|
||
g := `{"nodes":[
|
||
{"id":"i","kind":"input","config":{"text":"介绍杭州西湖"}},
|
||
{"id":"r","kind":"retriever","config":{"kb":"travel"}},
|
||
{"id":"a","kind":"agent","config":{"system":"你是导游"}}
|
||
],"edges":[{"source":"i","target":"r"},{"source":"r","target":"a"}]}`
|
||
ll := &fakeLLM{ready: true, stream: func(m []llm.ChatMessage) string { return m[0].Content }}
|
||
o := newOrch(ll, kbTool(nil), &fakeSink{}, &fakeExec{})
|
||
tk := &contract.Task{ID: "t_refs", Graph: json.RawMessage(g), Meta: map[string]any{contract.MetaUserID: "u42"}}
|
||
|
||
_, refs, err := o.runComposeGraph(context.Background(), tk, &execTracer{})
|
||
if err != nil {
|
||
t.Fatalf("runComposeGraph: %v", err)
|
||
}
|
||
if len(refs) == 0 || !strings.Contains(strings.Join(refs, ""), ragSnippet) {
|
||
t.Fatalf("compose 路径应回流含检索片段的 refs,got %v", refs)
|
||
}
|
||
}
|
||
|
||
// TestComposeEquivalentLinear 多节点线性图:input→memory→agent→output,两路径成稿应逐字一致。
|
||
func TestComposeEquivalentLinear(t *testing.T) {
|
||
graph := `{"version":"1","nodes":[
|
||
{"id":"in","kind":"input","config":{"text":"什么是图编排"}},
|
||
{"id":"m","kind":"memory","config":{}},
|
||
{"id":"a","kind":"agent","config":{"system":"你是助手"}},
|
||
{"id":"out","kind":"output","config":{}}
|
||
],"edges":[
|
||
{"source":"in","target":"m"},{"source":"m","target":"a"},{"source":"a","target":"out"}
|
||
]}`
|
||
interp, comp := runBoth(t, graph)
|
||
if interp == "" || interp != comp {
|
||
t.Fatalf("线性图不等价: interp=%q compose=%q", interp, comp)
|
||
}
|
||
}
|
||
|
||
// TestComposeEquivalentBranch 分支图:input→branch→(真)A/(假)B,条件恒真应都走 A,两路径一致。
|
||
func TestComposeEquivalentBranch(t *testing.T) {
|
||
graph := `{"version":"1","nodes":[
|
||
{"id":"in","kind":"input","config":{"text":"hi"}},
|
||
{"id":"br","kind":"branch","config":{"condition":""}},
|
||
{"id":"a","kind":"agent","config":{"system":"A"}},
|
||
{"id":"b","kind":"agent","config":{"system":"B"}}
|
||
],"edges":[
|
||
{"source":"in","target":"br"},
|
||
{"source":"br","target":"a","sourceHandle":"true"},
|
||
{"source":"br","target":"b","sourceHandle":"false"}
|
||
]}`
|
||
interp, comp := runBoth(t, graph)
|
||
if interp == "" || interp != comp {
|
||
t.Fatalf("分支图不等价: interp=%q compose=%q", interp, comp)
|
||
}
|
||
}
|