fix(core): 规则结论改在脱敏前算，会话路上 fabric.loader_0_12 能命中（M241）
会话 JSONL 是写前脱敏的，而脱敏的「≥64 无点号不透明串」是有损的： fabric.loader_0_12 唯一的证据行（Sponge mixin 类路径 66 字符）被抹成 <redacted:opaque>，规则在会话路上永不命中。 修法不动脱敏策略、也不往盘上多写一份原文： 1. SessionWriter 在内存里留一个 2000 行的原始尾窗（不落盘）； 2. 收尾时把这些**原始**事实交给新的 diagnose::session::conclude 跑规则， 产出契约形状的 diagnosis payload —— 它落盘前仍过 Redactor::scrub_json； 3. 日志页优先读这条落盘结论，老会话才回落到「在脱敏文本上重跑规则」。 为什么不是「原文与脱敏文本并存」：会话 JSONL 里存原文会踩 core/tests/launch/pipeline_cases.rs 早已钉死的「事件流不得含凭据明文」。 顺带把 EventKind 纯搬移到 launch/session/event_kind.rs 守 400 行上限。
This commit is contained in:
AskaEth committed
1 parent
74be510114
commit
edd6402fd5
10 files changed
+633
-65
No files matched your search
@@ -28,6 +28,8 @@ use crate::dto::{
|
||||
|
||||
use super::env::HostEnv;
|
||||
|
||||
#[path = "logs/diagnosis.rs"]
|
||||
mod diagnosis;
|
||||
#[path = "logs/location.rs"]
|
||||
mod location;
|
||||
|
||||
@@ -141,30 +143,43 @@ pub fn line_of(event: &Value) -> Option<LogLine> {
|
||||
})
|
||||
}
|
||||
|
||||
/// 整个事件流 → 日志行（解析失败的行跳过，容忍被 kill 时截断的尾行）。
|
||||
pub fn read_lines(path: &Path) -> Vec<LogLine> {
|
||||
/// 整个事件流的原始 JSON 行（解析失败的行跳过，容忍被 kill 时截断的尾行）。
|
||||
///
|
||||
/// 一次读完所有事件：日志行与诊断结论都从这同一份读里取，不重复读盘。
|
||||
pub fn read_events(path: &Path) -> Vec<Value> {
|
||||
let Ok(text) = std::fs::read_to_string(path) else {
|
||||
return Vec::new();
|
||||
};
|
||||
text.lines()
|
||||
.filter(|line| !line.trim().is_empty())
|
||||
.filter_map(|line| serde_json::from_str::<Value>(line).ok())
|
||||
.filter_map(|event| line_of(&event))
|
||||
.collect()
|
||||
}
|
||||
|
||||
/// 整个事件流 → 日志行（**展示面**：`payload.line`，即写盘前已脱敏的文本）。
|
||||
pub fn read_lines(path: &Path) -> Vec<LogLine> {
|
||||
read_events(path).iter().filter_map(line_of).collect()
|
||||
}
|
||||
|
||||
/// 打开一个会话：日志行 + 规则引擎结论。
|
||||
///
|
||||
/// 结论优先取会话里**落盘的那份**（写会话时用原始行算过，见 [`diagnosis`] 模块文档）；
|
||||
/// 老会话没有它，才回落到「在脱敏文本上重跑规则」。
|
||||
pub fn open(env: &HostEnv, key: &str) -> (Vec<LogLine>, Option<DiagnosisReport>) {
|
||||
let Some(entry) = find(env, key) else {
|
||||
return (Vec::new(), None);
|
||||
};
|
||||
let path = event_path(env, &entry);
|
||||
let lines = read_lines(&path);
|
||||
let diagnosis = diagnose_lines(&entry, &lines);
|
||||
let events = read_events(&event_path(env, &entry));
|
||||
let lines: Vec<LogLine> = events.iter().filter_map(line_of).collect();
|
||||
let diagnosis = diagnosis::stored(&entry, &events).or_else(|| diagnose_lines(&entry, &lines));
|
||||
(lines, diagnosis)
|
||||
}
|
||||
|
||||
/// 规则引擎跑一遍（**规则先跑，模型补位**：这里只出规则结论）。
|
||||
/// 在**脱敏文本**上重跑规则引擎（老会话的兜底路径）。
|
||||
///
|
||||
/// ⚠️ 这条路的证据是「写盘前脱敏过的」行：`≥64` 无点号长串（Sponge 类路径、长 hash）会被
|
||||
/// 抹成 `<redacted:opaque>`，依赖那些字面量的规则（`fabric.loader_0_12`）在这里**不可能命中**。
|
||||
/// 新会话通常不必走这条路——写会话时用原始行算过结论，见 [`diagnosis::stored`]。
|
||||
pub fn diagnose_lines(entry: &SessionIndexEntry, lines: &[LogLine]) -> Option<DiagnosisReport> {
|
||||
let stderr_text: String = lines
|
||||
.iter()
|
||||
@@ -355,6 +370,9 @@ pub fn session_finished(env: &HostEnv, session: &str) -> bool {
|
||||
})
|
||||
}
|
||||
|
||||
#[cfg(test)]
|
||||
#[path = "logs/m241_tests.rs"]
|
||||
mod m241_tests;
|
||||
#[cfg(test)]
|
||||
#[path = "logs_tests.rs"]
|
||||
mod tests;
|
||||
@@ -0,0 +1,82 @@
|
||||
//! 会话里落盘的诊断结论（M241）：读回时优先用它，不要拿脱敏文本重跑规则。
|
||||
//!
|
||||
//! ## 为什么不能重跑
|
||||
//!
|
||||
//! 会话 JSONL 是**写前脱敏**的，脱敏的两条形状规则有损：`≥64` 无点号不透明串会把证据行
|
||||
//! 一起抹掉（`fabric.loader_0_12` 的 `org/spongepowered/…/FabricMixinTransformerProxy`
|
||||
//! 66 字符 → `<redacted:opaque>`）。在脱敏文本上重跑规则必然空手——M239 实测那条 critical
|
||||
//! 规则在会话路上「永不命中」。
|
||||
//!
|
||||
//! 修法是**换证据**而不是换判据：写会话时原始行还在内存里，那时已经算过一份结论并落进
|
||||
//! `diagnosis` 事件（`launch::session::SessionWriter::finish` → `diagnose::session::conclude`），
|
||||
//! 那份结论在盘上仍然是脱敏过的（`Redactor::scrub_json`）。展示侧只读它。
|
||||
//!
|
||||
//! 老会话（本次改动之前写的 JSONL）没有 `diagnosis` 事件 → 调用方回落到
|
||||
//! [`super::diagnose_lines`]（在脱敏文本上重跑），行为与今天一致，不会凭空少一张卡片。
|
||||
|
||||
use serde_json::Value;
|
||||
use unlucky_core::launch::session_index::SessionIndexEntry;
|
||||
|
||||
use crate::dto::{Diagnosis, DiagnosisReport};
|
||||
|
||||
use super::confidence_percent;
|
||||
|
||||
/// 事件流里最后一条 `diagnosis` 事件 → 日志页的诊断卡片 DTO。
|
||||
///
|
||||
/// 正常只会有一条（挂钩写的，或收尾补的，二选一——见 `SessionWriter::conclude_from_raw` 的门槛）。
|
||||
/// 真出现多条时取**最后一条**：收尾补的那份证据最新鲜。
|
||||
pub(super) fn stored(entry: &SessionIndexEntry, events: &[Value]) -> Option<DiagnosisReport> {
|
||||
let payload = events
|
||||
.iter()
|
||||
.rev()
|
||||
.find(|event| event.get("kind").and_then(Value::as_str) == Some("diagnosis"))?
|
||||
.get("payload")?;
|
||||
let candidate = payload
|
||||
.get("candidates")
|
||||
.and_then(Value::as_array)
|
||||
.and_then(|items| items.first())
|
||||
.or_else(|| {
|
||||
payload
|
||||
.get("secondary")
|
||||
.and_then(Value::as_array)
|
||||
.and_then(|items| items.first())
|
||||
})?;
|
||||
let rule = candidate.get("id").and_then(Value::as_str)?;
|
||||
let title = candidate.get("message").and_then(Value::as_str)?;
|
||||
if title.is_empty() {
|
||||
return None;
|
||||
}
|
||||
let confidence = candidate
|
||||
.get("confidence")
|
||||
.and_then(Value::as_f64)
|
||||
.unwrap_or_default();
|
||||
let hint = candidate
|
||||
.get("detail")
|
||||
.and_then(Value::as_str)
|
||||
.map(str::to_string)
|
||||
.or_else(|| {
|
||||
payload
|
||||
.get("exit")
|
||||
.and_then(|exit| exit.get("exit_type"))
|
||||
.and_then(Value::as_str)
|
||||
.map(str::to_string)
|
||||
})
|
||||
.unwrap_or_else(|| "未知退出".to_string());
|
||||
let fix = candidate
|
||||
.get("fix_actions")
|
||||
.and_then(Value::as_array)
|
||||
.and_then(|items| items.first())
|
||||
.and_then(|action| action.get("tool"))
|
||||
.and_then(Value::as_str)
|
||||
.map(str::to_string);
|
||||
Some(DiagnosisReport {
|
||||
session: entry.session_id.clone(),
|
||||
diagnosis: Diagnosis {
|
||||
rule: rule.to_string(),
|
||||
title: title.to_string(),
|
||||
confidence: confidence_percent(confidence as f32),
|
||||
hint,
|
||||
fix,
|
||||
},
|
||||
})
|
||||
}
|
||||
@@ -0,0 +1,136 @@
|
||||
//! M241：诊断吃**原始**文本、展示/导出只吃脱敏文本（拆文件守 400 行上限）。
|
||||
|
||||
use super::*;
|
||||
use unlucky_core::launch::exit_analysis::{self, ExitSignals};
|
||||
use unlucky_core::launch::pump::{LogLine as CoreLine, Stream};
|
||||
use unlucky_core::launch::redact::Redactor;
|
||||
use unlucky_core::launch::session::{EndReason, SessionMeta, SessionWriter};
|
||||
|
||||
/// 一个临时数据根（`<root>/.ULL` + `<root>/.minecraft`）。
|
||||
fn temp_root(tag: &str) -> PathBuf {
|
||||
let root = std::env::temp_dir().join(format!("ull-m241-{tag}-{}", std::process::id()));
|
||||
let _ = std::fs::remove_dir_all(&root);
|
||||
root
|
||||
}
|
||||
|
||||
/// `fabric.loader_0_12`（critical / priority 80）在 `rules.d/fabric.yaml:114` 的唯一证据行。
|
||||
///
|
||||
/// 类路径 66 字符、无点号 → 撞上 `redact.rs` 的「≥64 不透明串」规则 → 脱敏后被抹成
|
||||
/// `<redacted:opaque>`。这正是 M239 报的那个「会话路上永不命中」。
|
||||
const EVIDENCE: &str = "java.lang.NoClassDefFoundError: \
|
||||
org/spongepowered/asm/mixin/transformer/FabricMixinTransformerProxy";
|
||||
/// 真凭据（写盘前必须被抹掉的形状）。
|
||||
const TOKEN: &str = "SECRET-TOKEN-123456";
|
||||
|
||||
fn meta() -> SessionMeta {
|
||||
SessionMeta::new("inst", "1.20.1", "1.20.1")
|
||||
}
|
||||
|
||||
/// 判据（M241）：① 规则引擎对含长类名的行**能命中**；② 用户可见/导出的文本**仍是被脱敏的**。
|
||||
///
|
||||
/// 端到端走真链路：`SessionWriter` 写盘（写前脱敏）→ 真 JSONL → [`open`] 读回。
|
||||
///
|
||||
/// 变异证明：
|
||||
/// - 把 `SessionWriter::finish` 里补结论的输入换成脱敏后的行 ⇒ ① 红（`fabric.loader_0_12`
|
||||
/// 命不中，`diagnosis_top` 不是它）；
|
||||
/// - 把 `SessionWriter::line` 里的 `scrub` 去掉（导出面改成原始文本）⇒ ② 红（类路径与凭据
|
||||
/// 原样出现在展示行与盘上）。
|
||||
#[test]
|
||||
fn raw_evidence_hits_the_rule_while_the_export_surface_stays_redacted() {
|
||||
let root = temp_root("diagraw");
|
||||
let env = HostEnv::at_root(&root);
|
||||
let redactor = Redactor::new().with_secret(TOKEN);
|
||||
let mut writer =
|
||||
SessionWriter::create_with_id(&env.sessions_root(), "diagraw-1", meta(), redactor)
|
||||
.expect("建会话");
|
||||
writer.session_start().expect("session_start");
|
||||
let lines = vec![
|
||||
CoreLine::new(Stream::Stderr, "Failed to start Fabric loader"),
|
||||
CoreLine::new(Stream::Stderr, EVIDENCE),
|
||||
CoreLine::new(Stream::Stdout, format!("auth: token={TOKEN}")),
|
||||
];
|
||||
for line in &lines {
|
||||
writer.line(line).expect("写行");
|
||||
}
|
||||
let analysis = exit_analysis::analyze(&ExitSignals::new(1, 40_000, &lines));
|
||||
writer.exit(&analysis).expect("exit");
|
||||
let entry = writer.finish(EndReason::Exit).expect("finish");
|
||||
|
||||
// ① 规则对含长类名的行能命中（结论在原始行上算过，落进索引）。
|
||||
assert_eq!(
|
||||
entry.diagnosis_top.as_deref(),
|
||||
Some("fabric.loader_0_12"),
|
||||
"会话路上必须命中 fabric.loader_0_12"
|
||||
);
|
||||
let (opened, diagnosis) = open(&env, "diagraw-1");
|
||||
assert_eq!(
|
||||
diagnosis.map(|report| report.diagnosis.rule),
|
||||
Some("fabric.loader_0_12".to_string())
|
||||
);
|
||||
|
||||
// ② 展示面（日志行 DTO）仍是被脱敏的。
|
||||
assert!(
|
||||
opened
|
||||
.iter()
|
||||
.any(|line| line.text.contains("<redacted:opaque>")),
|
||||
"长类名必须被抹成占位符：{opened:?}"
|
||||
);
|
||||
assert!(
|
||||
!opened
|
||||
.iter()
|
||||
.any(|line| line.text.contains("FabricMixinTransformerProxy")),
|
||||
"展示行不许出现原始类路径：{opened:?}"
|
||||
);
|
||||
assert!(
|
||||
!opened.iter().any(|line| line.text.contains(TOKEN)),
|
||||
"展示行不许出现凭据：{opened:?}"
|
||||
);
|
||||
|
||||
// ②' 盘上那份（= 导出包收进去的同一份文件）也没有原文——修 bug 不等于把密钥写进盘。
|
||||
let disk = std::fs::read_to_string(event_path(&env, &entry)).expect("读会话文件");
|
||||
assert!(!disk.contains("FabricMixinTransformerProxy"), "{disk}");
|
||||
assert!(!disk.contains(TOKEN), "{disk}");
|
||||
std::fs::remove_dir_all(&root).ok();
|
||||
}
|
||||
|
||||
/// 老会话（本次改动之前写的 JSONL，没有 `diagnosis` 事件）走兜底：在脱敏文本上重跑规则。
|
||||
///
|
||||
/// 判据：短字面量的规则（`memory.out_of_memory`）在脱敏文本上照样命中 ⇒ 兜底没坏。
|
||||
#[test]
|
||||
fn sessions_without_a_stored_conclusion_still_get_a_card() {
|
||||
let root = temp_root("diagraw-legacy");
|
||||
let env = HostEnv::at_root(&root);
|
||||
let entry = SessionIndexEntry {
|
||||
session_id: "legacy".to_string(),
|
||||
instance_id: "inst".to_string(),
|
||||
started_at: "2026-11-01T10:00:00.000Z".to_string(),
|
||||
ended_at: "2026-11-01T10:01:00.000Z".to_string(),
|
||||
exit_code: Some(1),
|
||||
exit_type: Some("application_error".to_string()),
|
||||
log_path: "game/legacy.jsonl".to_string(),
|
||||
..SessionIndexEntry::default()
|
||||
};
|
||||
session_index::append(&session_index::index_path(&env.sessions_root()), &entry)
|
||||
.expect("写索引");
|
||||
let file = env.sessions_root().join("game/legacy.jsonl");
|
||||
std::fs::create_dir_all(file.parent().expect("父目录")).expect("建目录");
|
||||
std::fs::write(
|
||||
&file,
|
||||
concat!(
|
||||
"{\"kind\":\"stderr\",\"payload\":{\"line\":\"java.lang.OutOfMemoryError: Java heap space\",\"level\":\"error\"}}\n",
|
||||
"{\"kind\":\"session_end\",\"payload\":{\"reason\":\"exit\"}}\n"
|
||||
),
|
||||
)
|
||||
.expect("写会话");
|
||||
|
||||
let (lines, diagnosis) = open(&env, "legacy");
|
||||
assert!(lines
|
||||
.iter()
|
||||
.any(|line| line.text.contains("OutOfMemoryError")));
|
||||
assert_eq!(
|
||||
diagnosis.map(|report| report.diagnosis.rule),
|
||||
Some("memory.out_of_memory".to_string()),
|
||||
"没有落盘结论时应回落到规则重跑"
|
||||
);
|
||||
std::fs::remove_dir_all(&root).ok();
|
||||
}
|
||||
Reference in new issue
Block a user