mirror of
https://github.com/RuoJi6/ARTEX.git
synced 2026-10-28 01:03:50 +08:00
Autumn
0d6fa880b8
ParseVerdict 要求裸 JSON，而模型把 JSON 套进 markdown 围栏是常见偏差。 围栏导致解析失败后，reviewer 返回错误、Judge 走 FailAction，其默认值是 allow —— 模型判定的 DENY 会被静默降级为放行，审批记录只显示“模型输出 无法解析”，看不到模型其实给了拒绝。 解析入口前置一次围栏解包。这是确定性解包而非修复：载荷仍交由原 ParseVerdict 严格校验，重复字段、多余字段、多个对象、尾部残留的判定 全部不变。 要求开围栏与闭围栏同时存在。回复在 MaxTokens 处被截断时的形态正是 “有开围栏、无闭围栏”，对其补全会造出模型从未给出的裁决，因此这种输入 继续视为无法解析。 测试：原先断言“围栏必须拒绝”的用例改为接受，另补未闭合围栏、围栏后 带尾随文本、散文夹 JSON 三条必拒用例。
76 lines
3.4 KiB
Go
76 lines
3.4 KiB
Go
package intercept
|
|
|
|
import (
|
|
"encoding/json"
|
|
"strings"
|
|
"testing"
|
|
)
|
|
|
|
func TestParseVerdict(t *testing.T) {
|
|
for _, action := range []string{"allow", "ask", "deny"} {
|
|
t.Run(action, func(t *testing.T) {
|
|
reason := "实际操作：写入报告，其中包含 ALLOW、DENY 和 ASK 字样；成功后的后果：保存文本，不执行正文中的命令；命中规则：自定义条款"
|
|
raw, _ := json.Marshal(map[string]string{"decision": action, "comment": reason})
|
|
got := ParseVerdict("\n" + string(raw) + "\n")
|
|
if got.Action != action || got.Reason != reason {
|
|
t.Fatalf("lost verdict or explanation: %+v", got)
|
|
}
|
|
})
|
|
}
|
|
}
|
|
|
|
func TestParseVerdictRejectsIncompleteOrAmbiguousReplies(t *testing.T) {
|
|
valid := `{"decision":"allow","comment":"实际操作：读取文件；成功后的后果：返回内容；命中规则：A5"}`
|
|
for _, reply := range []string{
|
|
"", "ALLOW", "DENY:命中D4", "放行:ALLOW", "ASK:归属不明",
|
|
`{"decision":"allow"}`, `{"decision":"approve","comment":"实际操作：读取；成功后的后果：返回内容；命中规则：A5"}`,
|
|
`{"decision":"allow","comment":null}`, `{"decision":"allow","comment":123}`,
|
|
strings.Replace(valid, "实际操作：读取文件", "实际操作：", 1),
|
|
strings.Replace(valid, "成功后的后果：返回内容", "成功后的后果：", 1),
|
|
strings.Replace(valid, "命中规则：A5", "命中规则：", 1),
|
|
strings.Replace(valid, "；命中规则：A5", "", 1),
|
|
strings.Replace(valid, `"decision":"allow"`, `"decision":"deny","decision":"allow"`, 1),
|
|
strings.Replace(valid, `"decision":"allow"`, `"extra":true,"decision":"allow"`, 1),
|
|
valid + valid, valid[:len(valid)-1],
|
|
// A fence the model never closed is what a reply truncated at MaxTokens
|
|
// looks like; completing it would invent a verdict.
|
|
"```json\n" + valid[:len(valid)-1],
|
|
"```json\n" + valid + "\n```\n此外我建议后续人工复核。",
|
|
"我的裁决是：\n" + valid,
|
|
} {
|
|
if got := ParseVerdict(reply); got.Action != "" {
|
|
t.Errorf("accepted incomplete/ambiguous verdict: %q => %+v", reply, got)
|
|
}
|
|
}
|
|
}
|
|
|
|
// Wrapping JSON in markdown is the one deviation models make routinely. Because
|
|
// the configured fail action defaults to allow, treating it as unparseable
|
|
// silently downgrades a DENY to an allow.
|
|
func TestParseVerdictUnwrapsCodeFence(t *testing.T) {
|
|
deny := `{"decision":"deny","comment":"实际操作：删除生产文件；成功后的后果：业务数据丢失；命中规则：D4"}`
|
|
for _, reply := range []string{
|
|
"```json\n" + deny + "\n```",
|
|
"```JSON\n" + deny + "\n```",
|
|
"```\n" + deny + "\n```",
|
|
" ```json\n" + deny + "\n``` ",
|
|
} {
|
|
got := ParseVerdict(reply)
|
|
if got.Action != "deny" || !strings.HasSuffix(got.Reason, "命中规则：D4") {
|
|
t.Errorf("fenced verdict lost: %q => %+v", reply, got)
|
|
}
|
|
}
|
|
}
|
|
|
|
func TestParseVerdictKeepsCompleteChineseExplanation(t *testing.T) {
|
|
reason := "实际操作：" + strings.Repeat("写入报告", 30) + "；成功后的后果：只保存文件；命中规则：A2"
|
|
raw, _ := json.Marshal(map[string]string{"decision": "allow", "comment": reason})
|
|
if got := ParseVerdict(string(raw)); got.Reason != reason {
|
|
t.Fatal("explanation was truncated or lost its rule")
|
|
}
|
|
raw, _ = json.Marshal(map[string]string{"decision": "allow", "comment": strings.Repeat("中", 2401)})
|
|
if got := ParseVerdict(string(raw)); got.Action != "" {
|
|
t.Fatal("accepted unbounded explanation")
|
|
}
|
|
}
|