ultra-memory 3.0.5 → 3.2.0

package/scripts/evolve_profile.py

@@ -0,0 +1,414 @@
+#!/usr/bin/env python3
+"""
+ultra-memory: 用户画像进化引擎 (Evolution Engine Phase 4)
+基于贝叶斯置信度更新实现 user_profile.json 的动态演化。
+
+贝叶斯更新公式（Beta 分布）：
+  alpha_post = max(0.5, old_conf × old_count) + new_evidence_conf
+  beta_post  = max(0.5, (1-old_conf) × old_count) + (1 - new_evidence_conf)
+  new_conf   = alpha_post / (alpha_post + beta_post)
+
+当新事实与画像矛盾时：
+  1. 追加 correction_history
+  2. 新值胜出，旧值降权记录
+  3. 如果旧置信度 > 0.85 → 标记为 pending manual resolution
+
+被 extract_facts.py 或 SKILL.md Step 7B 触发。
+"""
+
+import os
+import sys
+import json
+import argparse
+import math
+from datetime import datetime, timezone
+from pathlib import Path
+
+if sys.stdout.encoding != "utf-8":
+    sys.stdout.reconfigure(encoding="utf-8")
+if sys.stderr.encoding != "utf-8":
+    sys.stderr.reconfigure(encoding="utf-8")
+
+ULTRA_MEMORY_HOME = Path(os.environ.get("ULTRA_MEMORY_HOME", Path.home() / ".ultra-memory"))
+
+# ── 贝叶斯更新 ─────────────────────────────────────────────────────────────
+
+
+def _now_iso() -> str:
+    return datetime.now(timezone.utc).isoformat().replace("+00:00", "Z")
+
+
+def _parse_ts(ts_str: str) -> datetime:
+    if not ts_str:
+        return datetime.now(timezone.utc)
+    try:
+        ts_str = ts_str.replace("Z", "+00:00")
+        return datetime.fromisoformat(ts_str)
+    except ValueError:
+        return datetime.now(timezone.utc)
+
+
+def bayesian_update(
+    existing_confidence: float,
+    new_evidence_confidence: float,
+    existing_count: int,
+) -> tuple[float, int]:
+    """
+    Beta 分布贝叶斯更新。
+
+    Jeffreys prior: Beta(0.5, 0.5)
+    posterior alpha' = prior_alpha + new_weight
+    posterior beta'  = prior_beta  + (1 - new_weight)
+    new_confidence    = alpha' / (alpha' + beta')
+
+    Returns: (new_confidence, new_count)
+    """
+    # Jeffreys prior 保护
+    alpha_prior = max(0.5, existing_confidence * existing_count)
+    beta_prior = max(0.5, (1 - existing_confidence) * existing_count)
+
+    alpha_post = alpha_prior + new_evidence_confidence
+    beta_post = beta_prior + (1 - new_evidence_confidence)
+
+    new_confidence = alpha_post / (alpha_post + beta_post)
+    new_count = existing_count + 1
+
+    return round(new_confidence, 3), new_count
+
+
+# ── 画像加载/保存 ─────────────────────────────────────────────────────────
+
+
+def _load_profile() -> dict:
+    """加载 user_profile.json，兼容 v1 和 v2 格式"""
+    profile_file = ULTRA_MEMORY_HOME / "semantic" / "user_profile.json"
+    if not profile_file.exists():
+        return {"version": 2, "fields": {}, "last_reflection": None, "last_distillation": None}
+
+    try:
+        with open(profile_file, encoding="utf-8") as f:
+            profile = json.load(f)
+    except (json.JSONDecodeError, IOError):
+        return {"version": 2, "fields": {}, "last_reflection": None, "last_distillation": None}
+
+    # v1 → v2 迁移
+    if profile.get("version") != 2:
+        migrated = {"version": 2, "fields": {}, "last_reflection": None, "last_distillation": None}
+
+        # 迁移 tech_stack
+        if "tech_stack" in profile:
+            migrated["fields"]["tech_stack"] = {
+                "value": profile["tech_stack"] if isinstance(profile["tech_stack"], list) else [profile["tech_stack"]],
+                "confidence": 0.6,
+                "evidence_count": 1,
+                "last_updated": profile.get("last_updated", _now_iso()),
+                "sources": [],
+                "corrected_at": None,
+            }
+
+        # 迁移 language
+        if "language" in profile:
+            migrated["fields"]["language"] = {
+                "value": profile["language"],
+                "confidence": 0.7,
+                "evidence_count": 1,
+                "last_updated": profile.get("last_updated", _now_iso()),
+                "sources": [],
+                "corrected_at": None,
+            }
+
+        # 迁移 work_style
+        if "work_style" in profile:
+            for k, v in profile["work_style"].items():
+                migrated["fields"][f"work_style.{k}"] = {
+                    "value": v,
+                    "confidence": 0.5,
+                    "evidence_count": 1,
+                    "last_updated": profile.get("last_updated", _now_iso()),
+                    "sources": [],
+                    "corrected_at": None,
+                }
+
+        # 迁移 observed_patterns
+        if "observed_patterns" in profile:
+            migrated["fields"]["observed_patterns"] = {
+                "value": profile["observed_patterns"],
+                "confidence": 0.5,
+                "evidence_count": 1,
+                "last_updated": profile.get("last_updated", _now_iso()),
+                "sources": [],
+                "corrected_at": None,
+            }
+
+        migrated["last_reflection"] = profile.get("last_reflection")
+        migrated["last_distillation"] = profile.get("last_distillation")
+
+        return migrated
+
+    return profile
+
+
+def _save_profile(profile: dict):
+    """原子写入 user_profile.json"""
+    semantic_dir = ULTRA_MEMORY_HOME / "semantic"
+    semantic_dir.mkdir(parents=True, exist_ok=True)
+    profile_file = semantic_dir / "user_profile.json"
+    profile["version"] = 2
+
+    tmp_file = profile_file.with_suffix(".tmp")
+    with open(tmp_file, "w", encoding="utf-8") as f:
+        json.dump(profile, f, ensure_ascii=False, indent=2)
+    tmp_file.replace(profile_file)
+
+
+# ── 字段更新 ───────────────────────────────────────────────────────────────
+
+
+# 偏好类谓词 → 画像字段映射
+PREFERENCE_MAPPINGS = {
+    "user_prefers": "preferences",
+    "user_avoids": "preferences",
+    "adopted": "preferences",
+    "chose": "preferences",
+}
+
+
+def update_profile_from_fact(fact: dict, session_id: str):
+    """
+    基于提取到的事实更新用户画像。
+    判断事实是否涉及用户偏好/行为，并更新对应字段。
+    """
+    predicate = fact.get("predicate", "")
+    subject = fact.get("subject", "")
+    obj = fact.get("object", "")
+    confidence = fact.get("confidence", 0.7)
+
+    profile = _load_profile()
+    fields = profile.setdefault("fields", {})
+
+    changed = False
+
+    # 1. 处理用户偏好类谓词
+    if predicate in PREFERENCE_MAPPINGS:
+        pref_key = f"preferences.{subject}"
+
+        existing = fields.get(pref_key)
+        if existing:
+            new_conf, new_count = bayesian_update(
+                existing.get("confidence", 0.5),
+                confidence,
+                existing.get("evidence_count", 0),
+            )
+            existing["confidence"] = new_conf
+            existing["evidence_count"] = new_count
+            existing["last_updated"] = _now_iso()
+            existing["sources"] = existing.get("sources", []) + [session_id]
+            existing["sources"] = existing["sources"][-10:]  # 最多保留10个来源
+
+            # 如果新值与旧值不同 → 追加修正历史
+            if existing.get("value") != obj:
+                existing.setdefault("correction_history", []).append({
+                    "corrected_at": _now_iso(),
+                    "old_value": existing.get("value"),
+                    "new_value": obj,
+                    "confidence_delta": round(new_conf - existing.get("confidence", 0.5), 3),
+                    "source": "auto",
+                })
+                existing["value"] = obj
+        else:
+            fields[pref_key] = {
+                "value": obj,
+                "confidence": confidence,
+                "evidence_count": 1,
+                "last_updated": _now_iso(),
+                "sources": [session_id],
+                "corrected_at": None,
+                "correction_history": [],
+            }
+        changed = True
+
+    # 2. 处理技术栈推断（从 depends_on / requires / uses 谓词推断）
+    elif predicate in ("depends_on", "requires", "uses", "installed_as"):
+        tech_key = "tech_stack"
+        existing = fields.get(tech_key)
+
+        if existing:
+            new_conf, new_count = bayesian_update(
+                existing.get("confidence", 0.5),
+                confidence,
+                existing.get("evidence_count", 0),
+            )
+            existing["confidence"] = new_conf
+            existing["evidence_count"] = new_count
+            existing["last_updated"] = _now_iso()
+
+            # 如果技术不在列表中，追加
+            if isinstance(existing["value"], list) and obj not in existing["value"]:
+                existing["value"] = existing["value"] + [obj]
+        else:
+            fields[tech_key] = {
+                "value": [obj],
+                "confidence": confidence,
+                "evidence_count": 1,
+                "last_updated": _now_iso(),
+                "sources": [session_id],
+                "corrected_at": None,
+                "correction_history": [],
+            }
+        changed = True
+
+    # 3. 处理行为模式（从 skip / fail_on / blocks 谓词推断）
+    elif predicate in ("skips", "fails_on", "blocks"):
+        pattern_key = f"work_style.behavior.{subject}"
+        existing = fields.get(pattern_key)
+
+        if existing:
+            new_conf, new_count = bayesian_update(
+                existing.get("confidence", 0.5),
+                confidence,
+                existing.get("evidence_count", 0),
+            )
+            existing["confidence"] = new_conf
+            existing["evidence_count"] = new_count
+            existing["last_updated"] = _now_iso()
+            if existing.get("value") != obj:
+                existing.setdefault("correction_history", []).append({
+                    "corrected_at": _now_iso(),
+                    "old_value": existing.get("value"),
+                    "new_value": obj,
+                    "confidence_delta": round(new_conf - existing.get("confidence", 0.5), 3),
+                    "source": "auto",
+                })
+                existing["value"] = obj
+        else:
+            fields[pattern_key] = {
+                "value": obj,
+                "confidence": confidence,
+                "evidence_count": 1,
+                "last_updated": _now_iso(),
+                "sources": [session_id],
+                "corrected_at": None,
+                "correction_history": [],
+            }
+        changed = True
+
+    if changed:
+        _save_profile(profile)
+        print(f"[ultra-memory] ✅ 画像更新: {predicate} / {subject} = {obj}")
+
+
+# ── 手动修正（SKILL.md Step 7B）────────────────────────────────────────────
+
+
+def correct_profile_field(
+    field_path: str,
+    new_value,
+    evidence_confidence: float = 1.0,
+    session_id: str | None = None,
+):
+    """
+    手动修正画像字段（Step 7B：错误修正）。
+    强制覆盖旧值，追加 correction_history。
+    """
+    profile = _load_profile()
+    fields = profile.setdefault("fields", {})
+
+    old_value = None
+    old_confidence = 0.5
+    old_count = 0
+
+    if field_path in fields:
+        old_value = fields[field_path].get("value")
+        old_confidence = fields[field_path].get("confidence", 0.5)
+        old_count = fields[field_path].get("evidence_count", 0)
+
+    # 强制更新
+    fields[field_path] = {
+        "value": new_value,
+        "confidence": min(1.0, evidence_confidence),
+        "evidence_count": old_count + 1,
+        "last_updated": _now_iso(),
+        "sources": [session_id] if session_id else [],
+        "corrected_at": _now_iso(),
+        "correction_history": fields.get(field_path, {}).get("correction_history", []) + [
+            {
+                "corrected_at": _now_iso(),
+                "old_value": old_value,
+                "new_value": new_value,
+                "old_confidence": old_confidence,
+                "new_confidence": evidence_confidence,
+                "source": "manual",
+            }
+        ],
+    }
+
+    _save_profile(profile)
+    print(f"[ultra-memory] ✅ 画像修正: {field_path}")
+    print(f"  旧值: {old_value} (conf={old_confidence:.2f})")
+    print(f"  新值: {new_value} (conf={evidence_confidence:.2f})")
+
+
+# ── 时间戳更新 ─────────────────────────────────────────────────────────────
+
+
+def update_reflection_timestamp():
+    """Step 7A/7C 完成后更新 last_reflection 时间戳"""
+    profile = _load_profile()
+    profile["last_reflection"] = _now_iso()
+    _save_profile(profile)
+
+
+def update_distillation_timestamp():
+    """Step 7C 完成后更新 last_distillation 时间戳"""
+    profile = _load_profile()
+    profile["last_distillation"] = _now_iso()
+    _save_profile(profile)
+
+
+# ── CLI ─────────────────────────────────────────────────────────────────────
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="更新用户画像")
+    parser.add_argument("--field", help="字段路径（dot-notation，如 tech_stack）")
+    parser.add_argument("--value", help="新值")
+    parser.add_argument("--evidence", type=float, default=0.8,
+                        help="证据置信度 (0.0-1.0)")
+    parser.add_argument("--source-session", default=None,
+                        help="来源 session ID")
+    parser.add_argument(
+        "--correct", action="store_true",
+        help="手动修正模式（强制覆盖旧值）"
+    )
+    parser.add_argument(
+        "--update-reflection", action="store_true",
+        help="仅更新时间戳"
+    )
+    args = parser.parse_args()
+
+    if args.update_reflection:
+        update_reflection_timestamp()
+        print("[ultra-memory] ✅ last_reflection 已更新")
+    elif args.correct and args.field and args.value is not None:
+        correct_profile_field(
+            args.field, args.value, args.evidence, args.source_session
+        )
+    elif args.field and args.value is not None:
+        # 模拟 fact 结构
+        fact = {
+            "predicate": args.field,
+            "subject": "",
+            "object": args.value,
+            "confidence": args.evidence,
+        }
+        update_profile_from_fact(fact, args.source_session or "cli")
+    else:
+        profile = _load_profile()
+        print(f"[ultra-memory] 当前画像版本: {profile.get('version', '?')}")
+        print(f"字段数: {len(profile.get('fields', {}))}")
+        print(f"last_reflection: {profile.get('last_reflection')}")
+        print(f"last_distillation: {profile.get('last_distillation')}")
+        for path, data in profile.get("fields", {}).items():
+            print(f"  {path}: {data.get('value')} "
+                  f"(conf={data.get('confidence', 0):.2f}, "
+                  f"n={data.get('evidence_count', 0)})")