ai-agent-book 精选快照（<2MB 代码与文档，来自 github.com/bojieli/ai-agent-book）

chapter9/ai-style-skill/skill_manager.py

@@ -0,0 +1,142 @@
+"""Skill 的增量维护：按开放式规则 id 合并、prune 与 SKILL.md 生成。
+
+防膨胀原则：提炼模型看到当前规则，语义相同时复用稳定 id；本模块按 id 合并来源，
+而不是再用预置 detector 指纹把新发现筛掉。
+长期未被新证据确认、或被评估证据推翻的规则归档到 skill/archive/，不再进入
+SKILL.md。所有合并/激活/归档决定都发生在这里，不交给生成候选的模型。
+"""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any, Dict, List, Set, Tuple
+
+ROOT = Path(__file__).resolve().parent
+SKILL_DIR = ROOT / "skill"
+
+def rule_signature(rule: Dict[str, Any]) -> str:
+    """模型在已有规则语义相同时复用 id；该稳定 id 就是合并键。"""
+    return str(rule.get("id", "")).strip().lower()
+
+
+def merge_rules(
+    existing: List[Dict[str, Any]], candidates: List[Dict[str, Any]]
+) -> Tuple[List[Dict[str, Any]], Dict[str, Any]]:
+    """把候选规则合并进现有规则集，返回 (规则集, 合并报告)。"""
+    rules = [dict(rule) for rule in existing]
+    report: Dict[str, Any] = {"added": [], "merged": [], "conflicts": []}
+    for cand in candidates:
+        sig = rule_signature(cand)
+        match = next(
+            (rule for rule in rules if rule_signature(rule) == sig), None
+        )
+        if match is None:
+            rules.append(cand)
+            report["added"].append(cand["id"])
+            continue
+        # 去重合并：并集来源与作用域，保留首次通过 schema 检查的定义与范例。
+        match["source_ids"] = sorted(set(match.get("source_ids", [])) | set(cand.get("source_ids", [])))
+        match["scope"] = sorted(set(match.get("scope", [])) | set(cand.get("scope", [])))
+        report["merged"].append(cand["id"])
+    return rules, report
+
+
+def prune_rules(
+    rules: List[Dict[str, Any]],
+    *,
+    current_batch: int,
+    idle_batches: int = 2,
+    contradicted_ids: Set[str] | None = None,
+) -> Tuple[List[Dict[str, Any]], List[Dict[str, Any]]]:
+    """把长期未被新证据确认、或被证据推翻的规则归档，返回 (存活, 归档)。"""
+    contradicted = contradicted_ids or set()
+    active, archived = [], []
+    for rule in rules:
+        last = rule.get("last_confirmed_batch", current_batch)
+        reason = None
+        if rule["id"] in contradicted:
+            reason = "被评估证据推翻（误伤率过高或与金标集冲突）"
+        elif current_batch - last > idle_batches:
+            reason = f"连续超过 {idle_batches} 批反馈未再被触发"
+        if reason:
+            archived.append({**rule, "status": "archived", "archive_reason": reason})
+        else:
+            active.append(rule)
+    return active, archived
+
+
+def _describe_detector(detector: Dict[str, Any]) -> str:
+    if detector.get("type") != "llm":
+        return "无效检测器（规则不会激活）"
+    return "LLM judge 语义判定（上线前须通过独立人工金标集校准）"
+
+
+def render_skill_md(rules: List[Dict[str, Any]]) -> str:
+    """按 house 风格渲染 SKILL.md：何时加载 + 每条规则的定义/坏例/好例/作用域。"""
+    lines = [
+        "---",
+        "name: ai-style",
+        "description: 中文文案去「AI 味」检查清单，由用户纠正反馈持续提炼而来",
+        "---",
+        "",
+        "# 去 AI 味写作 Skill",
+        "",
+        "## 何时加载",
+        "",
+        "当任务是用中文撰写或改写面向读者的文案（产品发布稿、公众号文章、邮件、README 等），",
+        "或用户反馈文字「AI 味太重」「不像人写的」时，加载本 Skill。",
+        "",
+        "## 使用方式",
+        "",
+        "起草或改写时逐条对照下面的规则自查。每条规则都给出定义、可检查的检测方法、",
+        "坏例与好例；规则只在声明的作用域内生效，作用域之外的文体不要套用。",
+        "",
+        f"## 规则清单（共 {len(rules)} 条）",
+        "",
+    ]
+    for i, rule in enumerate(rules, 1):
+        lines += [
+            f"### 规则 {i}：{rule['name']}（`{rule['id']}`）",
+            "",
+            f"- **定义**：{rule['definition']}",
+            f"- **检测方法**：{_describe_detector(rule['detector'])}",
+            f"- **坏例**：{rule.get('bad_example', '')}",
+            f"- **好例**：{rule.get('good_example', '')}",
+            f"- **改写建议**：{rule.get('rewrite_hint', '按定义改写。')}",
+            f"- **作用域**：{'、'.join(rule.get('scope', [])) or '通用'}",
+            f"- **来源反馈**：{', '.join(rule.get('source_ids', [])) or '无'}",
+            "",
+        ]
+    return "\n".join(lines)
+
+
+def write_skill(rules: List[Dict[str, Any]], skill_dir: Path | None = None) -> Path:
+    out_dir = skill_dir or SKILL_DIR
+    out_dir.mkdir(parents=True, exist_ok=True)
+    path = out_dir / "SKILL.md"
+    path.write_text(render_skill_md(rules), encoding="utf-8")
+    # 同步保存机器可读的规则集，供 evaluate/judge 直接加载。
+    (out_dir / "rules.json").write_text(
+        json.dumps(rules, ensure_ascii=False, indent=2), encoding="utf-8"
+    )
+    return path
+
+
+def write_archive(archived: List[Dict[str, Any]], skill_dir: Path | None = None) -> Path | None:
+    if not archived:
+        return None
+    out_dir = (skill_dir or SKILL_DIR) / "archive"
+    out_dir.mkdir(parents=True, exist_ok=True)
+    lines = ["# 已归档规则", ""]
+    for rule in archived:
+        lines += [
+            f"## {rule['name']}（`{rule['id']}`）",
+            f"- 归档原因：{rule.get('archive_reason', '未说明')}",
+            f"- 原定义：{rule['definition']}",
+            f"- 来源反馈：{', '.join(rule.get('source_ids', [])) or '无'}",
+            "",
+        ]
+    path = out_dir / "ARCHIVED.md"
+    path.write_text("\n".join(lines), encoding="utf-8")
+    return path