foxclaw 0.1.0 → 0.1.2

package/python/vendor/hmr/engines/thought_chain.py

@@ -0,0 +1,635 @@
+"""
+ThoughtChain Engine - 显式推理链引擎
+HMR v2.0 新增
+
+解决的问题：
+  v1.x 只存"发生了什么"（execution/decision），但不知道"为什么这样想"
+  ThoughtChain 保存完整的思维过程，让 AI 能：
+    1. 回溯推理路径（我当时为什么这样决策？）
+    2. 反思错误（预期 vs 实际，找出错误思维节点）
+    3. 传递认知链（接手任务时直接继承完整推理过程）
+
+思维节点类型：
+  observation  → 观察到的事实
+  hypothesis   → 提出的假设
+  decision     → 做出的决策
+  action       → 执行的行动
+  outcome      → 行动的结果
+  reflection   → 事后反思
+  insight      → 提炼的洞察
+
+链的生命周期：
+  open → active → reflecting → closed
+"""
+
+import json
+import uuid
+import threading
+from pathlib import Path
+from typing import List, Dict, Any, Optional, Tuple
+from dataclasses import dataclass, field
+from datetime import datetime
+from enum import Enum
+
+
+# ============================================================================
+# 数据结构
+# ============================================================================
+
+class ThoughtType(Enum):
+    OBSERVATION = "observation"    # 观察：发现了什么事实
+    HYPOTHESIS  = "hypothesis"     # 假设：可能的原因或方案
+    DECISION    = "decision"       # 决策：选择了哪个方案
+    ACTION      = "action"         # 行动：具体执行了什么
+    OUTCOME     = "outcome"        # 结果：行动产生了什么结果
+    REFLECTION  = "reflection"     # 反思：事后回顾和评价
+    INSIGHT     = "insight"        # 洞察：提炼的普适知识
+
+
+class ChainStatus(Enum):
+    OPEN       = "open"       # 刚创建，正在思考中
+    ACTIVE     = "active"     # 已执行行动，等待结果
+    REFLECTING = "reflecting" # 收到结果，正在反思
+    CLOSED     = "closed"     # 已完成，产生洞察
+
+
+@dataclass
+class Thought:
+    """单个思维节点"""
+    thought_id: str
+    chain_id: str
+    thought_type: ThoughtType
+    content: str
+
+    # 关联
+    parent_id: Optional[str] = None      # 基于哪个思维得出
+    memory_ids: List[str] = field(default_factory=list)  # 依据的记忆
+
+    # 质量指标
+    confidence: float = 0.7              # 确信度 0-1
+    was_correct: Optional[bool] = None   # 事后评估：是否正确（None=未评估）
+    correction: Optional[str] = None     # 如果错了，正确答案是什么
+
+    timestamp: datetime = field(default_factory=datetime.utcnow)
+
+    def to_dict(self) -> dict:
+        return {
+            "thought_id": self.thought_id,
+            "chain_id": self.chain_id,
+            "thought_type": self.thought_type.value,
+            "content": self.content,
+            "parent_id": self.parent_id,
+            "memory_ids": self.memory_ids,
+            "confidence": self.confidence,
+            "was_correct": self.was_correct,
+            "correction": self.correction,
+            "timestamp": self.timestamp.isoformat(),
+        }
+
+    @classmethod
+    def from_dict(cls, d: dict) -> "Thought":
+        d = d.copy()
+        d["thought_type"] = ThoughtType(d["thought_type"])
+        d["timestamp"] = datetime.fromisoformat(d["timestamp"])
+        return cls(**d)
+
+
+@dataclass
+class ReflectionResult:
+    """反思结果：评估整条链的质量"""
+    chain_id: str
+    accuracy: float                      # 整体准确率（正确思维/总思维）
+    correct_nodes: List[str]             # 正确的思维 ID
+    wrong_nodes: List[str]               # 错误的思维 ID
+    key_mistakes: List[str]              # 主要错误描述
+    insights: List[str]                  # 提炼的洞察
+    suggested_memory: Optional[str]      # 建议存储的关键学习
+    timestamp: datetime = field(default_factory=datetime.utcnow)
+
+
+@dataclass
+class ThoughtChain:
+    """完整的思维链"""
+    chain_id: str
+    goal: str
+    status: ChainStatus = ChainStatus.OPEN
+
+    thoughts: List[Thought] = field(default_factory=list)
+
+    # 结果
+    expected_outcome: Optional[str] = None
+    actual_outcome: Optional[str] = None
+    reflection: Optional[ReflectionResult] = None
+
+    # 元数据
+    agent_id: Optional[str] = None
+    runtime_id: Optional[str] = None
+    created_at: datetime = field(default_factory=datetime.utcnow)
+    closed_at: Optional[datetime] = None
+
+    # ── 便利方法 ──────────────────────────────────────────────────────────
+
+    def add_thought(self, thought: Thought):
+        self.thoughts.append(thought)
+
+    def get_latest(self, n: int = 1) -> List[Thought]:
+        return self.thoughts[-n:]
+
+    def get_by_type(self, ttype: ThoughtType) -> List[Thought]:
+        return [t for t in self.thoughts if t.thought_type == ttype]
+
+    def get_decision_chain(self) -> List[Thought]:
+        """返回从第一个 observation 到最后一个 decision 的主干路径"""
+        key_types = {ThoughtType.OBSERVATION, ThoughtType.HYPOTHESIS,
+                     ThoughtType.DECISION, ThoughtType.ACTION}
+        return [t for t in self.thoughts if t.thought_type in key_types]
+
+    def to_summary(self) -> str:
+        """生成人类可读的链摘要"""
+        lines = [f"目标: {self.goal}", f"状态: {self.status.value}", ""]
+        for t in self.thoughts:
+            icon = {
+                "observation": "👁", "hypothesis": "💭", "decision": "✅",
+                "action": "⚡", "outcome": "📊", "reflection": "🔍", "insight": "💡"
+            }.get(t.thought_type.value, "·")
+            correct_marker = (
+                " ✓" if t.was_correct is True else
+                " ✗" if t.was_correct is False else ""
+            )
+            lines.append(f"  {icon} [{t.thought_type.value}]{correct_marker} {t.content[:60]}")
+        if self.reflection:
+            lines.append(f"\n反思准确率: {self.reflection.accuracy:.0%}")
+            for ins in self.reflection.insights:
+                lines.append(f"  💡 {ins}")
+        return "\n".join(lines)
+
+    def to_dict(self) -> dict:
+        return {
+            "chain_id": self.chain_id,
+            "goal": self.goal,
+            "status": self.status.value,
+            "thoughts": [t.to_dict() for t in self.thoughts],
+            "expected_outcome": self.expected_outcome,
+            "actual_outcome": self.actual_outcome,
+            "reflection": self._reflection_to_dict(),
+            "agent_id": self.agent_id,
+            "runtime_id": self.runtime_id,
+            "created_at": self.created_at.isoformat(),
+            "closed_at": self.closed_at.isoformat() if self.closed_at else None,
+        }
+
+    def _reflection_to_dict(self) -> Optional[dict]:
+        if not self.reflection:
+            return None
+        r = self.reflection
+        return {
+            "chain_id": r.chain_id,
+            "accuracy": r.accuracy,
+            "correct_nodes": r.correct_nodes,
+            "wrong_nodes": r.wrong_nodes,
+            "key_mistakes": r.key_mistakes,
+            "insights": r.insights,
+            "suggested_memory": r.suggested_memory,
+            "timestamp": r.timestamp.isoformat(),
+        }
+
+    @classmethod
+    def from_dict(cls, d: dict) -> "ThoughtChain":
+        d = d.copy()
+        d["status"] = ChainStatus(d["status"])
+        d["thoughts"] = [Thought.from_dict(t) for t in d.get("thoughts", [])]
+        d["created_at"] = datetime.fromisoformat(d["created_at"])
+        if d.get("closed_at"):
+            d["closed_at"] = datetime.fromisoformat(d["closed_at"])
+
+        raw_ref = d.pop("reflection", None)
+        chain = cls(**d)
+        if raw_ref:
+            raw_ref["timestamp"] = datetime.fromisoformat(raw_ref["timestamp"])
+            chain.reflection = ReflectionResult(**raw_ref)
+        return chain
+
+
+# ============================================================================
+# Reflector（反思器）
+# ============================================================================
+
+class Reflector:
+    """
+    对一条思维链做事后反思：
+    - 对比预期结果 vs 实际结果
+    - 标记哪些思维节点是错误的
+    - 提炼洞察
+    - 建议存储哪些学习到长期记忆
+    """
+
+    def reflect(
+        self,
+        chain: ThoughtChain,
+        actual_outcome: str,
+        outcome_rating: float = None  # 0-1，人工评分（可选）
+    ) -> ReflectionResult:
+        """
+        执行反思。
+
+        Args:
+            chain: 要反思的思维链
+            actual_outcome: 实际发生了什么
+            outcome_rating: 人工评分（0=完全失败, 1=完全成功），不传则自动推断
+        """
+        chain.actual_outcome = actual_outcome
+        chain.status = ChainStatus.REFLECTING
+
+        # 推断评分
+        if outcome_rating is None:
+            outcome_rating = self._infer_rating(chain.expected_outcome, actual_outcome)
+
+        # 标记思维节点
+        correct_nodes, wrong_nodes = self._evaluate_thoughts(chain, outcome_rating)
+
+        # 找主要错误
+        key_mistakes = self._identify_mistakes(chain, wrong_nodes)
+
+        # 提炼洞察
+        insights = self._extract_insights(chain, actual_outcome, outcome_rating)
+
+        # 建议记忆
+        suggested_memory = self._suggest_memory(chain, insights, outcome_rating)
+
+        accuracy = len(correct_nodes) / max(len(chain.thoughts), 1)
+
+        result = ReflectionResult(
+            chain_id=chain.chain_id,
+            accuracy=accuracy,
+            correct_nodes=correct_nodes,
+            wrong_nodes=wrong_nodes,
+            key_mistakes=key_mistakes,
+            insights=insights,
+            suggested_memory=suggested_memory,
+        )
+
+        chain.reflection = result
+        chain.status = ChainStatus.CLOSED
+        chain.closed_at = datetime.utcnow()
+
+        return result
+
+    def _infer_rating(self, expected: Optional[str], actual: str) -> float:
+        """推断结果评分（简单词级重叠）"""
+        if not expected:
+            # 无预期，中性评分
+            neg_words = ["失败", "错误", "超时", "崩溃", "fail", "error", "crash", "timeout"]
+            pos_words = ["成功", "完成", "正常", "通过", "success", "done", "ok", "pass"]
+            actual_lower = actual.lower()
+            neg_score = sum(1 for w in neg_words if w in actual_lower)
+            pos_score = sum(1 for w in pos_words if w in actual_lower)
+            if neg_score > pos_score:
+                return 0.2
+            elif pos_score > neg_score:
+                return 0.9
+            return 0.5
+
+        # 有预期：计算词级重叠度
+        exp_words = set(expected.lower().split())
+        act_words = set(actual.lower().split())
+        if not exp_words:
+            return 0.5
+        overlap = len(exp_words & act_words) / len(exp_words)
+        return min(1.0, overlap * 1.5)
+
+    def _evaluate_thoughts(
+        self, chain: ThoughtChain, rating: float
+    ) -> Tuple[List[str], List[str]]:
+        """
+        根据最终评分，倒推哪些决策/假设节点是错误的。
+
+        规则：
+        - 结果好（rating > 0.7）→ decisions 和 hypotheses 都标为正确
+        - 结果差（rating < 0.4）→ 找最后的 decision，标为错误；hypothesis 有一个错误
+        - 中等 → 随机保留部分正确
+        """
+        correct_ids = []
+        wrong_ids = []
+
+        for t in chain.thoughts:
+            if t.thought_type in (ThoughtType.OBSERVATION, ThoughtType.OUTCOME):
+                # 观察和结果节点本身不评判对错
+                t.was_correct = None
+                continue
+
+            if rating >= 0.7:
+                t.was_correct = True
+                correct_ids.append(t.thought_id)
+            elif rating <= 0.3:
+                if t.thought_type in (ThoughtType.DECISION, ThoughtType.HYPOTHESIS):
+                    t.was_correct = False
+                    wrong_ids.append(t.thought_id)
+                else:
+                    t.was_correct = True
+                    correct_ids.append(t.thought_id)
+            else:
+                # 中等结果：行动正确但决策可能有问题
+                if t.thought_type == ThoughtType.DECISION:
+                    t.was_correct = (rating > 0.5)
+                    (correct_ids if t.was_correct else wrong_ids).append(t.thought_id)
+                else:
+                    t.was_correct = True
+                    correct_ids.append(t.thought_id)
+
+        return correct_ids, wrong_ids
+
+    def _identify_mistakes(
+        self, chain: ThoughtChain, wrong_ids: List[str]
+    ) -> List[str]:
+        mistakes = []
+        for t in chain.thoughts:
+            if t.thought_id in wrong_ids:
+                mistakes.append(
+                    f"[{t.thought_type.value}] 思维有误: {t.content[:50]}"
+                )
+        return mistakes
+
+    def _extract_insights(
+        self, chain: ThoughtChain, actual_outcome: str, rating: float
+    ) -> List[str]:
+        insights = []
+
+        if rating >= 0.8:
+            decisions = chain.get_by_type(ThoughtType.DECISION)
+            if decisions:
+                insights.append(
+                    f"有效决策路径：{decisions[-1].content[:40]}（评分={rating:.0%}）"
+                )
+        elif rating <= 0.3:
+            hypotheses = chain.get_by_type(ThoughtType.HYPOTHESIS)
+            if hypotheses:
+                insights.append(
+                    f"假设方向有误：{hypotheses[0].content[:40]} → 实际:{actual_outcome[:30]}"
+                )
+            insights.append("下次应先验证假设，再做决策")
+
+        # 通用洞察：基于目标和结果
+        insights.append(
+            f"目标「{chain.goal[:30]}」: "
+            f"{'达成' if rating > 0.6 else '未达成'}（评分={rating:.0%}）"
+        )
+
+        return insights
+
+    def _suggest_memory(
+        self, chain: ThoughtChain, insights: List[str], rating: float
+    ) -> Optional[str]:
+        if not insights:
+            return None
+
+        # 只在有明确学习价值时建议存储
+        if rating > 0.85 or rating < 0.25:
+            decisions = chain.get_by_type(ThoughtType.DECISION)
+            decision_str = decisions[-1].content[:80] if decisions else ""
+
+            return (
+                f"[思维链洞察] 目标:{chain.goal[:30]} | "
+                f"结果:{'成功' if rating > 0.6 else '失败'} | "
+                f"关键决策:{decision_str} | "
+                f"洞察:{' / '.join(insights[:2])}"
+            )
+        return None
+
+
+# ============================================================================
+# ThoughtChain Engine（主类）
+# ============================================================================
+
+class ThoughtChainEngine:
+    """
+    思维链引擎
+
+    核心功能：
+    1. 创建思维链（一次推理任务的开始）
+    2. 追加思维节点（推理过程记录）
+    3. 触发反思（执行完毕后评估推理质量）
+    4. 查询思维链（找到类似问题的历史推理）
+    5. 持久化（重启后恢复推理历史）
+
+    使用示例：
+        engine = ThoughtChainEngine("./hmr_data/chains")
+
+        # 开始推理任务
+        chain = engine.create_chain("为什么调度器超时", agent_id="agent_01")
+
+        # 记录推理过程
+        engine.add_thought(chain.chain_id, "IPC 队列深度超过 1000", ThoughtType.OBSERVATION)
+        engine.add_thought(chain.chain_id, "可能是生产者速度超过消费者", ThoughtType.HYPOTHESIS)
+        engine.add_thought(chain.chain_id, "决定增加消费者线程数", ThoughtType.DECISION)
+        engine.add_thought(chain.chain_id, "已部署消费者 x3", ThoughtType.ACTION)
+
+        # 执行后触发反思
+        result = engine.reflect(chain.chain_id, actual_outcome="延迟降低 80%", rating=0.9)
+        print(result.insights)
+    """
+
+    def __init__(self, storage_path: Optional[str] = None):
+        self.storage_path = Path(storage_path) if storage_path else None
+        self._lock = threading.Lock()
+        self.chains: Dict[str, ThoughtChain] = {}
+        self.reflector = Reflector()
+        self._ingest_fn = None  # 注入 HMR.ingest（存储洞察到长期记忆）
+
+        if self.storage_path:
+            self.storage_path.mkdir(parents=True, exist_ok=True)
+            self._load_all()
+
+    def register_ingest_fn(self, fn):
+        """注入 HMR.ingest 函数，用于把洞察存入长期记忆"""
+        self._ingest_fn = fn
+
+    # ── 创建和管理 ─────────────────────────────────────────────────────────
+
+    def create_chain(
+        self,
+        goal: str,
+        agent_id: Optional[str] = None,
+        runtime_id: Optional[str] = None,
+        expected_outcome: Optional[str] = None,
+    ) -> ThoughtChain:
+        chain = ThoughtChain(
+            chain_id=f"tc_{uuid.uuid4().hex[:8]}",
+            goal=goal,
+            agent_id=agent_id,
+            runtime_id=runtime_id,
+            expected_outcome=expected_outcome,
+        )
+        with self._lock:
+            self.chains[chain.chain_id] = chain
+        self._save(chain)
+        return chain
+
+    def add_thought(
+        self,
+        chain_id: str,
+        content: str,
+        thought_type: ThoughtType,
+        confidence: float = 0.7,
+        memory_ids: Optional[List[str]] = None,
+        parent_id: Optional[str] = None,
+    ) -> Optional[Thought]:
+        chain = self.chains.get(chain_id)
+        if not chain or chain.status == ChainStatus.CLOSED:
+            return None
+
+        # 自动推断 parent_id（默认接在最后一个思维节点后）
+        if parent_id is None and chain.thoughts:
+            parent_id = chain.thoughts[-1].thought_id
+
+        thought = Thought(
+            thought_id=f"th_{uuid.uuid4().hex[:8]}",
+            chain_id=chain_id,
+            thought_type=thought_type,
+            content=content,
+            confidence=confidence,
+            memory_ids=memory_ids or [],
+            parent_id=parent_id,
+        )
+
+        with self._lock:
+            chain.add_thought(thought)
+            if thought_type == ThoughtType.ACTION:
+                chain.status = ChainStatus.ACTIVE
+
+        self._save(chain)
+        return thought
+
+    def reflect(
+        self,
+        chain_id: str,
+        actual_outcome: str,
+        outcome_rating: Optional[float] = None,
+    ) -> Optional[ReflectionResult]:
+        chain = self.chains.get(chain_id)
+        if not chain:
+            return None
+
+        result = self.reflector.reflect(chain, actual_outcome, outcome_rating)
+        self._save(chain)
+
+        # 把高价值洞察存入长期记忆
+        if result.suggested_memory and self._ingest_fn:
+            try:
+                self._ingest_fn(
+                    content=result.suggested_memory,
+                    memory_type="reflection",
+                    title=f"[链洞察] {chain.goal[:30]}",
+                    metadata={
+                        "tags": ["thought_chain", "reflection", "insight"],
+                        "confidence": result.accuracy,
+                        "runtime_dependencies": [chain.goal]
+                    }
+                )
+            except Exception as e:
+                print(f"[ThoughtChain] 洞察存储失败: {e}")
+
+        return result
+
+    # ── 查询 ───────────────────────────────────────────────────────────────
+
+    def get_chain(self, chain_id: str) -> Optional[ThoughtChain]:
+        return self.chains.get(chain_id)
+
+    def find_similar_chains(
+        self, goal: str, top_k: int = 3
+    ) -> List[ThoughtChain]:
+        """找与当前目标相似的历史思维链（关键词匹配）"""
+        goal_words = set(goal.lower().split())
+        scored = []
+        for chain in self.chains.values():
+            if chain.status != ChainStatus.CLOSED:
+                continue
+            chain_words = set(chain.goal.lower().split())
+            overlap = len(goal_words & chain_words) / max(len(goal_words), 1)
+            if overlap > 0.3:
+                scored.append((chain, overlap))
+
+        scored.sort(key=lambda x: x[1], reverse=True)
+        return [c for c, _ in scored[:top_k]]
+
+    def get_active_chains(
+        self, agent_id: Optional[str] = None
+    ) -> List[ThoughtChain]:
+        return [
+            c for c in self.chains.values()
+            if c.status != ChainStatus.CLOSED
+            and (agent_id is None or c.agent_id == agent_id)
+        ]
+
+    def get_best_decision_for(self, goal: str) -> Optional[str]:
+        """从历史成功链中找出对同类问题最好的决策"""
+        similar = self.find_similar_chains(goal)
+        best_chain = None
+        best_accuracy = 0.0
+
+        for chain in similar:
+            if chain.reflection and chain.reflection.accuracy > best_accuracy:
+                best_accuracy = chain.reflection.accuracy
+                best_chain = chain
+
+        if not best_chain or best_accuracy < 0.7:
+            return None
+
+        decisions = best_chain.get_by_type(ThoughtType.DECISION)
+        correct_decisions = [d for d in decisions if d.was_correct]
+        if correct_decisions:
+            return (
+                f"[历史参考 准确率={best_accuracy:.0%}] "
+                f"{correct_decisions[-1].content}"
+            )
+        return None
+
+    def get_stats(self) -> Dict[str, Any]:
+        total = len(self.chains)
+        by_status = {}
+        total_accuracy = 0.0
+        reflected_count = 0
+
+        for chain in self.chains.values():
+            s = chain.status.value
+            by_status[s] = by_status.get(s, 0) + 1
+            if chain.reflection:
+                total_accuracy += chain.reflection.accuracy
+                reflected_count += 1
+
+        avg_accuracy = total_accuracy / reflected_count if reflected_count > 0 else 0.0
+
+        return {
+            "total_chains": total,
+            "by_status": by_status,
+            "reflected_chains": reflected_count,
+            "avg_reflection_accuracy": round(avg_accuracy, 3),
+            "active_chains": by_status.get("open", 0) + by_status.get("active", 0),
+        }
+
+    # ── 持久化 ─────────────────────────────────────────────────────────────
+
+    def _save(self, chain: ThoughtChain):
+        if not self.storage_path:
+            return
+        path = self.storage_path / f"{chain.chain_id}.json"
+        tmp = path.with_suffix(".tmp")
+        with open(tmp, "w", encoding="utf-8") as f:
+            json.dump(chain.to_dict(), f, ensure_ascii=False, indent=2, default=str)
+        tmp.replace(path)
+
+    def _load_all(self):
+        if not self.storage_path:
+            return
+        count = 0
+        for path in self.storage_path.glob("tc_*.json"):
+            try:
+                with open(path, "r", encoding="utf-8") as f:
+                    data = json.load(f)
+                chain = ThoughtChain.from_dict(data)
+                self.chains[chain.chain_id] = chain
+                count += 1
+            except Exception as e:
+                print(f"[ThoughtChain] 加载失败 {path.name}: {e}")
+        if count:
+            print(f"[ThoughtChain] 加载 {count} 条思维链")