feat(harness): momentum.py — async 慢更新动量生成

app/harness/momentum.py

@@ -0,0 +1,178 @@
+"""慢更新动量生成 — epoch 末为单个 skill 产出新的动量指导。
+
+对标 SkillOpt 的 slow_update 机制：拿上一 epoch 末与当前 epoch 末两版 skill，
+在固定样本上各跑一遍得到纵向对比（comparison_pairs），反思上一轮动量指导是否奏效、
+本轮正文改动是改善还是漂移，据此重写动量指导。新指导经 patch 引擎的 replace_momentum
+写回 skill 的 momentum 受保护区，作为下一轮进化的方向锚。
+
+从 TRM4 core/harness/momentum.py（156 行）迁移 + async 化。
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any
+
+from loguru import logger
+
+from core.evolution.diagnose import extract_json_from_response
+
+if TYPE_CHECKING:
+    from pathlib import Path
+
+    from core.protocols import LLMProvider
+
+
+# =========================================================================
+# 四类纵向对比类别名（单一真源）
+# =========================================================================
+
+IMPROVED = "improved"  # 错→对
+REGRESSED = "regressed"  # 对→错
+PERSISTENT_FAIL = "persistent_fail"  # 错→错
+STABLE_SUCCESS = "stable_success"  # 对→对
+
+# 类别名 → 展示标题，列表顺序即展示顺序。
+# 回退（REGRESSED）刻意排在改善（IMPROVED）之前——它是最该警惕的伤害信号。
+_CATEGORY_LABELS: tuple[tuple[str, str], ...] = (
+    (REGRESSED, "从对变错（回退，最高优先级）"),
+    (PERSISTENT_FAIL, "始终答错（持续失败）"),
+    (IMPROVED, "从错变对（改善）"),
+    (STABLE_SUCCESS, "始终答对（稳定成功）"),
+)
+
+
+# =========================================================================
+# 辅助函数
+# =========================================================================
+
+
+def _categorize_pair(pair: dict[str, Any]) -> str:
+    """按两版正误派生纵向对比类别。
+
+    用键值的真值（bool(...)）表示该题在两版上各自的正误：缺 correct_prev/
+    correct_curr 键时直接抛 KeyError 向上传播——这是上游数据损坏（不是裁判语义
+    歧义），静默当 False 会伪造 persistent_fail 证据、污染动量指导，故不掩盖。
+
+    参数:
+        pair: 单个纵向对比对，须含 correct_prev/correct_curr 两键。
+
+    返回:
+        四个类别命名常量之一：IMPROVED（错→对）/REGRESSED（对→错）/
+        PERSISTENT_FAIL（错→错）/STABLE_SUCCESS（对→对）。
+
+    异常:
+        KeyError: 缺 correct_prev 或 correct_curr 键时。
+    """
+    correct_prev = bool(pair["correct_prev"])
+    correct_curr = bool(pair["correct_curr"])
+    if not correct_prev and correct_curr:
+        return IMPROVED
+    if correct_prev and not correct_curr:
+        return REGRESSED
+    if not correct_prev and not correct_curr:
+        return PERSISTENT_FAIL
+    return STABLE_SUCCESS
+
+
+def _format_comparison_pairs(comparison_pairs: list[dict[str, Any]]) -> str:
+    """将纵向对比对格式化为裁判可读文本，按 _CATEGORY_LABELS 分组与排序。
+
+    参数:
+        comparison_pairs: 每个 dict 含 question/prev_prediction/curr_prediction/
+            correct_prev/correct_curr 字段，描述一道固定样本上两版的成对结果。
+
+    返回:
+        可读的纵向对比文本；空列表返回占位说明。
+
+    异常:
+        KeyError: 任一 pair 缺 correct_prev/correct_curr 键时；不掩盖的理由见
+            _categorize_pair docstring。
+    """
+    if not comparison_pairs:
+        return "（本轮无可用纵向对比样本）"
+
+    grouped: dict[str, list[dict[str, Any]]] = {key: [] for key, _ in _CATEGORY_LABELS}
+    for pair in comparison_pairs:
+        grouped[_categorize_pair(pair)].append(pair)
+
+    lines: list[str] = [f"固定样本总数：{len(comparison_pairs)}"]
+    for key, label in _CATEGORY_LABELS:
+        entries = grouped[key]
+        lines.append(f"\n### {label}（{len(entries)} 题）")
+        if not entries:
+            lines.append("（无）")
+            continue
+        for pair in entries:
+            lines.append(
+                f"- 题目：{pair.get('question', '')}\n"
+                f"  上版预测：{pair.get('prev_prediction', '')} | "
+                f"当前版预测：{pair.get('curr_prediction', '')}"
+            )
+    return "\n".join(lines)
+
+
+# =========================================================================
+# 入口
+# =========================================================================
+
+
+async def run_slow_momentum(
+    llm: LLMProvider,
+    diagnose_prompts_dir: Path,
+    skill_content: str,
+    prev_skill: str,
+    prev_guidance: str,
+    comparison_pairs: list[dict[str, Any]],
+) -> str:
+    """为单个 skill 生成新的慢更新动量指导。
+
+    参数:
+        llm: LLM 端口（async chat）。
+        diagnose_prompts_dir: 诊断 prompt 目录（根 prompts/，slow_momentum.md 在此）。
+        skill_content: 当前版 skill 正文。
+        prev_skill: 上一版 skill 正文。
+        prev_guidance: 上一轮写下的动量指导。
+        comparison_pairs: 固定样本上两版 rollout 的成对结果（含 question/
+            prev_prediction/curr_prediction/correct_prev/correct_curr）。
+
+    返回:
+        新的动量指导文本；解析失败时保留 prev_guidance。
+
+    关键实现细节:
+        - _format_comparison_pairs 刻意置于 try 块之外（prompt 构造阶段）：它对每个
+          pair 取 correct_prev/correct_curr，缺键抛 KeyError 直接向上传播，不被下方
+          针对裁判语义歧义的 except ValueError 吞掉。
+        - 解析失败保留上轮指导：extract_json_from_response 抛 ValueError、缺
+          slow_update_content 字段、或该字段非 str，均视为语义解析失败，返回
+          prev_guidance（判不准时保守保留上轮指导，对标 diagnose 的保护性 fallback）。
+        - P5 边界：仅捕 ValueError 这一语义歧义；llm.chat 的基础设施失败
+          （网络/API 异常）刻意不捕，向上传播，绝不用默认值掩盖。
+    """
+    system_prompt = (diagnose_prompts_dir / "slow_momentum.md").read_text(encoding="utf-8")
+    # _format_comparison_pairs 刻意置于下方 try 块之外（prompt 构造阶段）：它对每个
+    # pair 取 correct_prev/correct_curr，缺键抛 KeyError 直接向上传播，不被下方针对
+    # 裁判语义歧义的 except ValueError 吞掉。异常类型选 KeyError（非 ValueError），
+    # 即便位置疏忽落入 try 也不会被误吞。
+    user_prompt = (
+        f"## 上一版 skill 正文\n{prev_skill}\n\n"
+        f"## 当前版 skill 正文\n{skill_content}\n\n"
+        f"## 上一轮的动量指导\n{prev_guidance}\n\n"
+        f"## 固定样本纵向对比（上版 vs 当前版）\n"
+        f"{_format_comparison_pairs(comparison_pairs)}"
+    )
+    response = await llm.chat(
+        [
+            {"role": "system", "content": system_prompt},
+            {"role": "user", "content": user_prompt},
+        ]
+    )
+    raw = response.content
+    try:
+        parsed = extract_json_from_response(raw)
+        new_guidance = parsed.get("slow_update_content")
+        if not isinstance(new_guidance, str):
+            raise ValueError("slow_update_content 字段缺失或非字符串")
+    except ValueError:
+        logger.warning("慢更新动量解析失败，保留上轮动量指导")
+        return prev_guidance
+    return new_guidance

tests/unit/test_harness_momentum.py

@@ -0,0 +1,261 @@
+"""慢更新动量生成（app/harness/momentum.py）单元测试。
+
+覆盖:
+- _categorize_pair 四分类 + 缺键 KeyError
+- _format_comparison_pairs 分组排序 + 空列表
+- run_slow_momentum 正常解析 + 解析失败保留 prev_guidance
+"""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any
+from unittest.mock import AsyncMock
+
+import pytest
+
+from app.harness.momentum import (
+    IMPROVED,
+    PERSISTENT_FAIL,
+    REGRESSED,
+    STABLE_SUCCESS,
+    _categorize_pair,
+    _format_comparison_pairs,
+    run_slow_momentum,
+)
+
+if TYPE_CHECKING:
+    from pathlib import Path
+
+
+# =========================================================================
+# _categorize_pair
+# =========================================================================
+
+
+class TestCategorizePair:
+    """_categorize_pair 四分类测试。"""
+
+    def test_categorize_pair_all_four(self) -> None:
+        """四种 correct_prev/correct_curr 组合应返回对应类别常量。"""
+        assert _categorize_pair({"correct_prev": False, "correct_curr": True}) == IMPROVED
+        assert _categorize_pair({"correct_prev": True, "correct_curr": False}) == REGRESSED
+        assert _categorize_pair({"correct_prev": False, "correct_curr": False}) == PERSISTENT_FAIL
+        assert _categorize_pair({"correct_prev": True, "correct_curr": True}) == STABLE_SUCCESS
+
+    def test_categorize_pair_truthy_values(self) -> None:
+        """非布尔真值（int / str）也能正确分类。"""
+        assert _categorize_pair({"correct_prev": 0, "correct_curr": 1}) == IMPROVED
+        assert _categorize_pair({"correct_prev": "yes", "correct_curr": ""}) == REGRESSED
+
+    def test_categorize_pair_missing_key(self) -> None:
+        """缺 correct_prev 或 correct_curr 键时抛 KeyError。"""
+        with pytest.raises(KeyError):
+            _categorize_pair({"correct_prev": True})
+        with pytest.raises(KeyError):
+            _categorize_pair({"correct_curr": False})
+        with pytest.raises(KeyError):
+            _categorize_pair({})
+
+
+# =========================================================================
+# _format_comparison_pairs
+# =========================================================================
+
+
+class TestFormatComparisonPairs:
+    """_format_comparison_pairs 分组排序测试。"""
+
+    def test_format_comparison_pairs_empty(self) -> None:
+        """空列表返回占位说明。"""
+        result = _format_comparison_pairs([])
+        assert "无可用" in result
+
+    def test_format_comparison_pairs_order(self) -> None:
+        """REGRESSED 标题应出现在 IMPROVED 标题之前（伤害信号优先）。"""
+        pairs = [
+            {
+                "question": "Q1",
+                "prev_prediction": "A",
+                "curr_prediction": "B",
+                "correct_prev": False,
+                "correct_curr": True,
+            },
+            {
+                "question": "Q2",
+                "prev_prediction": "C",
+                "curr_prediction": "D",
+                "correct_prev": True,
+                "correct_curr": False,
+            },
+        ]
+        result = _format_comparison_pairs(pairs)
+        regressed_pos = result.index("回退")
+        improved_pos = result.index("改善")
+        assert regressed_pos < improved_pos, "REGRESSED 应排在 IMPROVED 之前"
+
+    def test_format_comparison_pairs_all_categories(self) -> None:
+        """四种类别的 pair 都能被正确分组。"""
+        pairs = [
+            {"question": "Q1", "correct_prev": False, "correct_curr": True},
+            {"question": "Q2", "correct_prev": True, "correct_curr": False},
+            {"question": "Q3", "correct_prev": False, "correct_curr": False},
+            {"question": "Q4", "correct_prev": True, "correct_curr": True},
+        ]
+        result = _format_comparison_pairs(pairs)
+        assert "固定样本总数：4" in result
+        # 每个类别都应标注 1 题
+        assert "1 题" in result
+
+    def test_format_comparison_pairs_missing_key(self) -> None:
+        """pair 缺键时 KeyError 不被吞。"""
+        with pytest.raises(KeyError):
+            _format_comparison_pairs([{"question": "Q1"}])
+
+
+# =========================================================================
+# run_slow_momentum
+# =========================================================================
+
+
+def _make_mock_llm(response_content: str) -> Any:
+    """构造一个返回指定 content 的 mock LLMProvider。"""
+
+    @dataclass(frozen=True)
+    class _FakeResponse:
+        content: str
+        thinking: str = ""
+        model: str = "mock"
+        provider: str = "mock"
+        prompt_tokens: int = 0
+        completion_tokens: int = 0
+        latency_ms: int = 0
+        ttft_ms: float | None = None
+        max_inter_token_ms: float | None = None
+        cache_hit: bool = False
+        call_id: str = "test-call-id"
+
+    mock_llm = AsyncMock()
+    mock_llm.chat.return_value = _FakeResponse(content=response_content)
+    return mock_llm
+
+
+@pytest.mark.asyncio
+async def test_run_slow_momentum_basic(tmp_path: Path) -> None:
+    """正常解析时返回新的动量指导文本。"""
+    # 准备 prompt 文件
+    prompt_file = tmp_path / "slow_momentum.md"
+    prompt_file.write_text("你是一个慢更新动量裁判。", encoding="utf-8")
+
+    new_guidance_text = "新一轮的动量指导内容"
+    llm_response = json.dumps({"slow_update_content": new_guidance_text}, ensure_ascii=False)
+    mock_llm = _make_mock_llm(llm_response)
+
+    result = await run_slow_momentum(
+        llm=mock_llm,
+        diagnose_prompts_dir=tmp_path,
+        skill_content="当前 skill 正文",
+        prev_skill="上一版 skill 正文",
+        prev_guidance="旧的动量指导",
+        comparison_pairs=[
+            {
+                "question": "Q1",
+                "prev_prediction": "A",
+                "curr_prediction": "B",
+                "correct_prev": False,
+                "correct_curr": True,
+            }
+        ],
+    )
+    assert result == new_guidance_text
+    mock_llm.chat.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_run_slow_momentum_parse_failure(tmp_path: Path) -> None:
+    """LLM 返回无法解析的内容时保留 prev_guidance。"""
+    prompt_file = tmp_path / "slow_momentum.md"
+    prompt_file.write_text("你是一个慢更新动量裁判。", encoding="utf-8")
+
+    mock_llm = _make_mock_llm("这不是 JSON，无法解析")
+    prev_guidance = "应该被保留的旧动量指导"
+
+    result = await run_slow_momentum(
+        llm=mock_llm,
+        diagnose_prompts_dir=tmp_path,
+        skill_content="当前 skill",
+        prev_skill="上一版 skill",
+        prev_guidance=prev_guidance,
+        comparison_pairs=[
+            {
+                "question": "Q1",
+                "prev_prediction": "A",
+                "curr_prediction": "B",
+                "correct_prev": True,
+                "correct_curr": True,
+            }
+        ],
+    )
+    assert result == prev_guidance
+
+
+@pytest.mark.asyncio
+async def test_run_slow_momentum_missing_field(tmp_path: Path) -> None:
+    """LLM 返回合法 JSON 但缺少 slow_update_content 字段时保留 prev_guidance。"""
+    prompt_file = tmp_path / "slow_momentum.md"
+    prompt_file.write_text("你是一个慢更新动量裁判。", encoding="utf-8")
+
+    llm_response = json.dumps({"other_field": "无关内容"}, ensure_ascii=False)
+    mock_llm = _make_mock_llm(llm_response)
+    prev_guidance = "应该被保留的旧动量指导"
+
+    result = await run_slow_momentum(
+        llm=mock_llm,
+        diagnose_prompts_dir=tmp_path,
+        skill_content="当前 skill",
+        prev_skill="上一版 skill",
+        prev_guidance=prev_guidance,
+        comparison_pairs=[],
+    )
+    assert result == prev_guidance
+
+
+@pytest.mark.asyncio
+async def test_run_slow_momentum_keyerror_not_swallowed(tmp_path: Path) -> None:
+    """comparison_pairs 缺键时 KeyError 不被 ValueError 吞掉。"""
+    prompt_file = tmp_path / "slow_momentum.md"
+    prompt_file.write_text("你是一个慢更新动量裁判。", encoding="utf-8")
+
+    mock_llm = _make_mock_llm('{"slow_update_content": "ok"}')
+
+    with pytest.raises(KeyError):
+        await run_slow_momentum(
+            llm=mock_llm,
+            diagnose_prompts_dir=tmp_path,
+            skill_content="当前 skill",
+            prev_skill="上一版 skill",
+            prev_guidance="旧指导",
+            comparison_pairs=[{"question": "Q1"}],  # 缺 correct_prev/correct_curr
+        )
+
+
+@pytest.mark.asyncio
+async def test_run_slow_momentum_fenced_json(tmp_path: Path) -> None:
+    """LLM 返回 fenced code block 中的 JSON 也能正确解析。"""
+    prompt_file = tmp_path / "slow_momentum.md"
+    prompt_file.write_text("你是一个慢更新动量裁判。", encoding="utf-8")
+
+    new_guidance = "从 fenced block 中提取的指导"
+    llm_response = f'```json\n{{"slow_update_content": "{new_guidance}"}}\n```'
+    mock_llm = _make_mock_llm(llm_response)
+
+    result = await run_slow_momentum(
+        llm=mock_llm,
+        diagnose_prompts_dir=tmp_path,
+        skill_content="当前 skill",
+        prev_skill="上一版 skill",
+        prev_guidance="旧指导",
+        comparison_pairs=[],
+    )
+    assert result == new_guidance