156 lines
4.9 KiB
Go
156 lines
4.9 KiB
Go
package approvalreview
|
||
|
||
import (
|
||
"context"
|
||
"errors"
|
||
"strings"
|
||
"sync"
|
||
"testing"
|
||
|
||
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
|
||
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/orchestration/audit"
|
||
)
|
||
|
||
// recordingLLM 是记录请求的模型适配器桩：返回既定内容或既定错误。
|
||
type recordingLLM struct {
|
||
content string
|
||
err error
|
||
|
||
mu sync.Mutex
|
||
reqs []model.Request
|
||
}
|
||
|
||
func (m *recordingLLM) Name() string { return "recording-llm" }
|
||
|
||
func (m *recordingLLM) Chat(_ context.Context, req model.Request, _ model.Options) (model.Response, error) {
|
||
m.mu.Lock()
|
||
m.reqs = append(m.reqs, req)
|
||
m.mu.Unlock()
|
||
if m.err != nil {
|
||
return model.Response{}, m.err
|
||
}
|
||
return model.Response{Content: m.content, Message: model.TextMessage(model.RoleAssistant, m.content)}, nil
|
||
}
|
||
|
||
func (m *recordingLLM) ChatStream(context.Context, model.Request, func(model.StreamEvent)) (model.Response, error) {
|
||
return model.Response{}, nil
|
||
}
|
||
|
||
func (m *recordingLLM) last() model.Request {
|
||
m.mu.Lock()
|
||
defer m.mu.Unlock()
|
||
return m.reqs[len(m.reqs)-1]
|
||
}
|
||
|
||
func (m *recordingLLM) count() int {
|
||
m.mu.Lock()
|
||
defer m.mu.Unlock()
|
||
return len(m.reqs)
|
||
}
|
||
|
||
// promptText 拼接请求里所有消息的文本，便于断言问句确实带上了任务与调用摘要。
|
||
func (m *recordingLLM) promptText() string {
|
||
req := m.last()
|
||
var b strings.Builder
|
||
for _, msg := range req.Messages {
|
||
for _, c := range msg.Content {
|
||
b.WriteString(c.Text)
|
||
b.WriteString("\n")
|
||
}
|
||
}
|
||
return b.String()
|
||
}
|
||
|
||
// TestNewLLMJudge_NilAdapter：无模型适配器时返回 nil ⇒ 门对子代理的敏感调用走
|
||
// fail-closed（绝不因为没有判定器就放行）。
|
||
func TestNewLLMJudge_NilAdapter(t *testing.T) {
|
||
t.Parallel()
|
||
|
||
if got := NewLLMJudge(nil, nil); got != nil {
|
||
t.Fatalf("NewLLMJudge(nil) = %v, want nil", got)
|
||
}
|
||
}
|
||
|
||
// TestJudgeSubagentCall_AsksAStructuredQuestion：单轮轻量问句里带上了被派任务与
|
||
// 本次调用摘要，且不塞正文（问句只由这两段结构化输入组成）。
|
||
func TestJudgeSubagentCall_AsksAStructuredQuestion(t *testing.T) {
|
||
t.Parallel()
|
||
|
||
llm := &recordingLLM{content: `{"allow": true, "reason": ""}`}
|
||
judge := NewLLMJudge(llm, nil)
|
||
verdict, err := judge.JudgeSubagentCall(context.Background(), audit.JudgeRequest{
|
||
Task: "整理 data/ 下的导出文件",
|
||
Actor: "昔涟",
|
||
Action: "执行命令：terminal rm -rf /",
|
||
})
|
||
if err != nil {
|
||
t.Fatalf("JudgeSubagentCall: %v", err)
|
||
}
|
||
if !verdict.Allowed {
|
||
t.Fatalf("verdict = %+v, want 放行", verdict)
|
||
}
|
||
prompt := llm.promptText()
|
||
for _, want := range []string{"整理 data/ 下的导出文件", "执行命令：terminal rm -rf /", "昔涟"} {
|
||
if !strings.Contains(prompt, want) {
|
||
t.Fatalf("问句里缺少 %q，got: %s", want, prompt)
|
||
}
|
||
}
|
||
if req := llm.last(); req.Purpose != judgePurpose || req.MaxTokens != judgeMaxTokens {
|
||
t.Fatalf("模型请求 = (purpose=%q, maxTokens=%d), want (%q, %d)",
|
||
req.Purpose, req.MaxTokens, judgePurpose, judgeMaxTokens)
|
||
}
|
||
if llm.count() != 1 {
|
||
t.Fatalf("模型调用次数 = %d, want 1（单轮轻量）", llm.count())
|
||
}
|
||
}
|
||
|
||
// TestJudgeSubagentCall_ParsesDenial：判「越界」时把理由交回调用方（它要进审计与
|
||
// 前端小字）。围栏包裹的 JSON 也能识别。
|
||
func TestJudgeSubagentCall_ParsesDenial(t *testing.T) {
|
||
t.Parallel()
|
||
|
||
llm := &recordingLLM{content: "```json\n{\"allow\": false, \"reason\": \"任务里没有执行命令这一项\"}\n```"}
|
||
verdict, err := NewLLMJudge(llm, nil).JudgeSubagentCall(context.Background(), audit.JudgeRequest{
|
||
Task: "整理导出文件", Action: "执行命令：terminal rm -rf /",
|
||
})
|
||
if err != nil {
|
||
t.Fatalf("JudgeSubagentCall: %v", err)
|
||
}
|
||
if verdict.Allowed {
|
||
t.Fatal("verdict.Allowed = true, want false")
|
||
}
|
||
if verdict.Reason != "任务里没有执行命令这一项" {
|
||
t.Fatalf("理由 = %q, want 判定器给出的理由", verdict.Reason)
|
||
}
|
||
}
|
||
|
||
// TestJudgeSubagentCall_FailClosedOnBadOutput：模型报错、输出不是 JSON、缺 allow
|
||
// 字段 —— 一律报错（由 Approver 按「判不出来就不放行」处理）。
|
||
func TestJudgeSubagentCall_FailClosedOnBadOutput(t *testing.T) {
|
||
t.Parallel()
|
||
|
||
tests := []struct {
|
||
name string
|
||
llm *recordingLLM
|
||
}{
|
||
{name: "model error", llm: &recordingLLM{err: errors.New("provider down")}},
|
||
{name: "not json", llm: &recordingLLM{content: "允许"}},
|
||
{name: "broken json", llm: &recordingLLM{content: `{"allow": `}},
|
||
{name: "missing allow field", llm: &recordingLLM{content: `{"reason": "看起来没问题"}`}},
|
||
{name: "empty output", llm: &recordingLLM{content: ""}},
|
||
}
|
||
|
||
for _, tt := range tests {
|
||
tt := tt
|
||
t.Run(tt.name, func(t *testing.T) {
|
||
t.Parallel()
|
||
_, err := NewLLMJudge(tt.llm, nil).JudgeSubagentCall(context.Background(), audit.JudgeRequest{
|
||
Task: "整理导出文件", Action: "执行命令：terminal rm -rf /",
|
||
})
|
||
if err == nil {
|
||
t.Fatal("err = nil, want 判不出来（fail-closed）")
|
||
}
|
||
})
|
||
}
|
||
}
|