- tests/repair_experiment.rs:真实消费 validate → repair_targets → 模型行编辑 → repair_apply(越界必拒) → 复验 的完整闭环(≤2 轮;容错重试;cflmy @ llm.cflmy.cn,.env 读取,#[ignore] 不进常规回归) - 双臂消融:anchor(诊断带 doc_anchor/doc_quote/suggestion)vs plain(剥掉)—— anchor 一轮修复 20/20=100%;plain 17/20=85%(3 例失败均为网络空体抖动) - tests/common/mod.rs:golden 20 例与诊断/应用共享件抽出(repair_loop + experiment 同源) - parse_edits 容错(数组/单对象/字符串化元素/围栏);越界必拒/abstain 由 repair_loop 硬断言覆盖 - 报告:doc/roadmap/repair-experiment-2026-09-22.md;.gitignore 补 .env(密钥不入库) Co-authored-by: Cursor <cursoragent@cursor.com>
85 lines
4.2 KiB
Rust
85 lines
4.2 KiB
Rust
//! T3.3 golden:**20 例易锚定错误**的修复循环(护栏机械执行 + 金修复可复现)。
|
||
//!
|
||
//! 每例走完整循环:
|
||
//! 诊断(含 `span`/`doc_anchor`)→ 推导靶点范围 → 金修复(源 diff 自动推导行编辑,
|
||
//! 顺便机器验证「易锚定」= 金编辑全落范围内)→ `mlsp repair_apply` 应用 → 复验通过。
|
||
//!
|
||
//! 护栏负例内嵌每例:越界编辑(第 1 行 `# main`,永不在靶点内)必拒
|
||
//! (`mlsp.repair_out_of_scope` + `abstain`),源零字节不改。
|
||
//! 模型实验见 `tests/repair_experiment.rs`(同一批 20 例)。
|
||
|
||
mod common;
|
||
|
||
use common::*;
|
||
use marqdo::repair::{apply_edits, Edit};
|
||
use serde_json::json;
|
||
|
||
#[test]
|
||
fn golden_repair_loop_all_20() {
|
||
let cases = cases();
|
||
assert_eq!(cases.len(), 20, "golden 必须正好 20 例");
|
||
|
||
for case in &cases {
|
||
// 1. 诊断(broken 必须给出期望错误码)
|
||
let code = case.code();
|
||
let diags = diagnose(case.verify, &case.broken, &case.name);
|
||
assert!(
|
||
diags.iter().any(|d| d["code"] == code),
|
||
"[{}] 期望诊断 {code},实得 {diags:?}",
|
||
case.name
|
||
);
|
||
let ranges: Vec<(u32, u32)> = diags.iter().flat_map(ranges_of).collect();
|
||
assert!(!ranges.is_empty(), "[{}] 诊断必须带可锚定范围: {diags:?}", case.name);
|
||
|
||
// 2. 金修复 → 行编辑(diff 自动推导);「易锚定」= 全部编辑落在靶点内
|
||
let edits = derive_edits(&case.broken, &case.gold);
|
||
assert!(!edits.is_empty(), "[{}] 金修复必须产生编辑", case.name);
|
||
let applied = apply_edits(&case.broken, &ranges, &edits).unwrap_or_else(|oos| {
|
||
panic!("[{}] 金修复越界(不是易锚定错误): 越界 {:?},范围 {:?}", case.name, oos.lines, oos.ranges)
|
||
});
|
||
assert_eq!(applied, case.gold, "[{}] 行编辑结果应与金修复逐字节一致", case.name);
|
||
|
||
// 3. 经 MLSP 面应用(护栏所在)
|
||
let resp = apply_via_mlsp(&case.broken, &ranges, &edits);
|
||
assert_eq!(resp["ok"], true, "[{}] repair_apply 应接受金修复: {resp}", case.name);
|
||
assert_eq!(resp["result"]["source"], case.gold, "[{}]", case.name);
|
||
|
||
// 4. 复验(干净 = 空诊断)
|
||
let after = diagnose(case.verify, &case.gold, &case.name);
|
||
assert!(after.is_empty(), "[{}] 修复后必须零诊断: {after:?}", case.name);
|
||
|
||
// 5. 护栏负例:越界编辑(第 1 行 `# main`)必拒 + 源零字节不改
|
||
let bad = vec![Edit::Replace { line: 1, text: "# 被篡改".to_string() }];
|
||
let resp = apply_via_mlsp(&case.broken, &ranges, &bad);
|
||
assert_eq!(resp["ok"], false, "[{}] 越界编辑必须被拒: {resp}", case.name);
|
||
assert_eq!(resp["error"]["code"], "mlsp.repair_out_of_scope", "[{}]", case.name);
|
||
assert_eq!(resp["error"]["on_violation"], "abstain", "[{}]", case.name);
|
||
}
|
||
}
|
||
|
||
#[test]
|
||
fn repair_apply_refuses_any_out_of_scope_line() {
|
||
// 独立护栏例:部分越界 = 全部拒绝(全有或全无),源零字节不改。
|
||
let src = "# main\n\n| 字段 | 类型 | 可空 |\n|------|------|------|\n| id | number | |\n| name | text | 是 |\n\n*1*\n";
|
||
let ranges = [(3u32, 6u32)]; // 只有字段契约表
|
||
let edits = vec![
|
||
Edit::Replace { line: 5, text: "| id | text | |".to_string() }, // 表内 ✓
|
||
Edit::Delete { line: 1 }, // 越界 ✗
|
||
];
|
||
let resp = apply_via_mlsp(src, &ranges, &edits);
|
||
assert_eq!(resp["ok"], false);
|
||
assert_eq!(resp["error"]["code"], "mlsp.repair_out_of_scope");
|
||
assert!(resp["error"]["lines"].as_array().unwrap().contains(&json!(1)));
|
||
|
||
// 范围内编辑照常可用(含表尾追加)
|
||
let edits = vec![
|
||
Edit::Replace { line: 5, text: "| id | text | |".to_string() },
|
||
Edit::InsertBefore { line: 7, text: "| email | text | 是 |".to_string() }, // end+1 追加 ✓
|
||
];
|
||
let resp = apply_via_mlsp(src, &ranges, &edits);
|
||
assert_eq!(resp["ok"], true, "{resp}");
|
||
let out = resp["result"]["source"].as_str().unwrap();
|
||
assert!(out.contains("| id | text | |"));
|
||
assert!(out.contains("| email | text | 是 |"));
|
||
}
|