nexo-brain 1.2.3 → 1.4.0

package/src/scripts/nexo-sleep.py

@@ -1,31 +1,21 @@
 #!/usr/bin/env python3
 """
-NEXO Sleep System — Daily memory cleanup and pruning.
+NEXO Sleep System v2 — The brain dreams.
 
-Triggered hourly via LaunchAgent. Runs ONCE per day, first time the Mac is awake.
+Before: 834 lines with word-overlap "intelligence" for learning consolidation.
+Now: Stage A (mechanical cleanup) stays pure Python. Stage B (dreaming) uses
+Claude CLI (opus) to understand, deduplicate, and prune with real intelligence.
+
+Triggered hourly via LaunchAgent. Runs ONCE per day, first time Mac is awake.
 If interrupted (power loss, crash), resumes on next trigger.
 
-Stage A — Mechanical cleanup (Python pure, always runs):
-  A1: Delete daily_summaries >90 days
-  A2: Delete session_archive >30 days
-  A3: Rotate coordination stdout logs >5MB
-  A4: Delete compressed_memories/week_*.md >180 days
-  A5: Trim heartbeat-log.json to 200 entries
-  A6: Trim reflection-log.json to 60 entries
-  A7: Delete daemon/logs/ dirs >14 days
-
-Stage C — Learning Consolidation (Python pure, always runs):
-  C1: Duplicate detection (>80% word overlap in titles)
-  C2: Age distribution of learnings
-  C3: Category health (counts, hottest last 7d, categories >20)
-  C4: Contradiction detection (NUNCA pairs in same category)
-
-Stage B — Intelligent pruning (Claude CLI, conditional):
-  Only activates if MEMORY.md >170 lines, nexo.db preferences table has >5 rows,
-  or claude-mem.db has >500 observations >60 days.
-  Uses Claude CLI (sonnet) to compress and prune.
-
-Zero external dependencies beyond stdlib + sqlite3. Claude CLI for Stage B only.
+Stage A — Housekeeping (Python pure):
+  Delete old logs, rotate files, trim JSON. No intelligence needed.
+
+Stage B — Dreaming (Claude CLI opus):
+  Review learnings for duplicates and contradictions with UNDERSTANDING.
+  Prune MEMORY.md if over limit. Clean preferences. Compress old observations.
+  One CLI call that does what 500 lines of word-overlap couldn't.
 """
 
 import fcntl
@@ -40,7 +30,7 @@ from datetime import datetime, date, timedelta
 from pathlib import Path
 
 # ─── Paths ────────────────────────────────────────────────────────────────────
-CLAUDE_DIR = Path.home() / "claude"
+CLAUDE_DIR = Path.home() / ".nexo"
 BRAIN_DIR = CLAUDE_DIR / "brain"
 COORD_DIR = CLAUDE_DIR / "coordination"
 MEMORY_DIR = CLAUDE_DIR / "memory"
@@ -54,9 +44,8 @@ HEARTBEAT_LOG = COORD_DIR / "heartbeat-log.json"
 REFLECTION_LOG = COORD_DIR / "reflection-log.json"
 SLEEP_LOG = COORD_DIR / "sleep-log.json"
 
-NEXO_HOME = os.environ.get("NEXO_HOME", str(Path.home() / ".nexo"))
-MEMORY_MD = Path.home() / ".claude" / "projects" / f"-Users-{os.environ.get('USER', 'user')}" / "memory" / "MEMORY.md"
-NEXO_DB = Path(NEXO_HOME) / "nexo.db"
+MEMORY_MD = Path.home() / ".nexo" / "memory" / "MEMORY.md"
+NEXO_DB = Path.home() / ".nexo" / "nexo.db"
 CLAUDE_MEM_DB = Path.home() / ".claude-mem" / "claude-mem.db"
 CLAUDE_CLI = Path.home() / ".local" / "bin" / "claude"
 
@@ -69,30 +58,25 @@ NOW = datetime.now()
 TIMESTAMP = NOW.strftime("%Y-%m-%d %H:%M")
 
 
-# ─── Run-once & resume logic ────────────────────────────────────────────────
+# ─── Run-once & resume logic (unchanged from v1) ──────────────────────────────
 
 def already_ran_today() -> bool:
-    """Check if sleep already completed today."""
     if not LAST_RUN_FILE.exists():
         return False
     try:
-        last_date = LAST_RUN_FILE.read_text().strip()
-        return last_date == str(TODAY)
+        return LAST_RUN_FILE.read_text().strip() == str(TODAY)
     except Exception:
         return False
 
 
 def was_interrupted() -> bool:
-    """Check if a previous run was interrupted (lock file exists with dead PID)."""
     if not LOCK_FILE.exists():
         return False
     try:
         lock_data = json.loads(LOCK_FILE.read_text())
-        lock_date = lock_data.get("date", "")
-        if lock_date != str(TODAY):
+        if lock_data.get("date") != str(TODAY):
             LOCK_FILE.unlink()
             return False
-
         lock_pid = lock_data.get("pid")
         if lock_pid:
             try:
@@ -100,39 +84,29 @@ def was_interrupted() -> bool:
                 log(f"Another instance running (PID {lock_pid}). Exiting.")
                 return False
             except ProcessLookupError:
-                log(f"Interrupted run detected (phase: {lock_data.get('phase', '?')}, dead PID {lock_pid}). Resuming.")
+                log(f"Interrupted run (phase: {lock_data.get('phase', '?')}). Resuming.")
                 return True
             except PermissionError:
                 return False
-        else:
-            LOCK_FILE.unlink()
-            return False
+        LOCK_FILE.unlink()
+        return False
     except Exception:
         LOCK_FILE.unlink(missing_ok=True)
         return False
 
 
 def get_interrupted_phase() -> str:
-    """Get which phase was interrupted."""
     try:
-        lock_data = json.loads(LOCK_FILE.read_text())
-        return lock_data.get("phase", "stage_a")
+        return json.loads(LOCK_FILE.read_text()).get("phase", "stage_a")
     except Exception:
         return "stage_a"
 
 
 def set_lock(phase: str):
-    """Set lock file indicating current phase with PID for race detection."""
-    save_json(LOCK_FILE, {
-        "date": str(TODAY),
-        "phase": phase,
-        "started": TIMESTAMP,
-        "pid": os.getpid()
-    })
+    save_json(LOCK_FILE, {"date": str(TODAY), "phase": phase, "started": TIMESTAMP, "pid": os.getpid()})
 
 
 def mark_complete():
-    """Mark today's run as complete."""
     LAST_RUN_FILE.write_text(str(TODAY))
     LOCK_FILE.unlink(missing_ok=True)
 
@@ -140,7 +114,8 @@ def mark_complete():
 # ─── Helpers ──────────────────────────────────────────────────────────────────
 
 def log(msg: str):
-    print(f"[{TIMESTAMP}] {msg}")
+    ts = datetime.now().strftime("%Y-%m-%d %H:%M")
+    print(f"[{ts}] {msg}")
 
 
 def load_json(path: Path, default=None):
@@ -148,8 +123,7 @@ def load_json(path: Path, default=None):
         return default if default is not None else {}
     try:
         return json.loads(path.read_text())
-    except Exception as e:
-        log(f"WARN: Failed to load {path}: {e}")
+    except Exception:
         return default if default is not None else {}
 
 
@@ -158,8 +132,7 @@ def save_json(path: Path, data):
     path.write_text(json.dumps(data, indent=2, ensure_ascii=False))
 
 
-def parse_date_from_stem(stem: str) -> date | None:
-    """Extract YYYY-MM-DD date from a filename stem."""
+def parse_date_from_stem(stem: str):
     m = re.search(r'(\d{4}-\d{2}-\d{2})', stem)
     if m:
         try:
@@ -170,24 +143,19 @@ def parse_date_from_stem(stem: str) -> date | None:
 
 
 def append_sleep_log(entry: dict):
-    """Append entry to sleep-log.json, keeping last 90 entries."""
     entries = load_json(SLEEP_LOG, [])
     if not isinstance(entries, list):
         entries = []
     entries.append(entry)
-    # Keep last 90
     if len(entries) > 90:
         entries = entries[-90:]
     save_json(SLEEP_LOG, entries)
 
 
-# ─── Stage A: Mechanical cleanup ─────────────────────────────────────────────
+# ─── Stage A: Mechanical cleanup (UNCHANGED from v1) ─────────────────────────
 
 def stage_a_cleanup() -> dict:
-    """
-    Pure Python cleanup. No LLM calls.
-    Returns stats dict with counts per sub-task.
-    """
+    """Pure Python cleanup. No LLM calls."""
     stats = {
         "a1_daily_summaries_deleted": 0,
         "a2_session_archives_deleted": 0,
@@ -207,9 +175,8 @@ def stage_a_cleanup() -> dict:
                 try:
                     f.unlink()
                     stats["a1_daily_summaries_deleted"] += 1
-                    log(f"A1: Deleted {f.name} (>{90}d)")
-                except Exception as e:
-                    log(f"A1: WARN: Could not delete {f.name}: {e}")
+                except Exception:
+                    pass
 
     # A2: Delete session_archive/*.jsonl >30 days
     cutoff_30 = TODAY - timedelta(days=30)
@@ -220,22 +187,20 @@ def stage_a_cleanup() -> dict:
                 try:
                     f.unlink()
                     stats["a2_session_archives_deleted"] += 1
-                    log(f"A2: Deleted {f.name} (>{30}d)")
-                except Exception as e:
-                    log(f"A2: WARN: Could not delete {f.name}: {e}")
+                except Exception:
+                    pass
 
-    # A3: Rotate coordination/*-stdout.log if >5MB (keep last 500 lines)
+    # A3: Rotate coordination/*-stdout.log if >5MB
     if COORD_DIR.exists():
         for f in COORD_DIR.glob("*-stdout.log"):
             try:
-                if f.stat().st_size > 5 * 1024 * 1024:  # >5MB
+                if f.stat().st_size > 5 * 1024 * 1024:
                     lines = f.read_text().splitlines()
-                    keep = lines[-500:] if len(lines) > 500 else lines
+                    keep = lines[-500:]
                     f.write_text("\n".join(keep) + "\n")
                     stats["a3_logs_rotated"] += 1
-                    log(f"A3: Rotated {f.name} ({len(lines)}→{len(keep)} lines)")
-            except Exception as e:
-                log(f"A3: WARN: Could not rotate {f.name}: {e}")
+            except Exception:
+                pass
 
     # A4: Delete compressed_memories/week_*.md >180 days
     cutoff_180 = TODAY - timedelta(days=180)
@@ -246,37 +211,30 @@ def stage_a_cleanup() -> dict:
                 try:
                     f.unlink()
                     stats["a4_compressed_memories_deleted"] += 1
-                    log(f"A4: Deleted {f.name} (>{180}d)")
-                except Exception as e:
-                    log(f"A4: WARN: Could not delete {f.name}: {e}")
+                except Exception:
+                    pass
 
     # A5: Trim heartbeat-log.json to 200 entries
     if HEARTBEAT_LOG.exists():
         try:
             data = load_json(HEARTBEAT_LOG, [])
             if isinstance(data, list) and len(data) > 200:
-                before = len(data)
-                data = data[-200:]
-                save_json(HEARTBEAT_LOG, data)
+                save_json(HEARTBEAT_LOG, data[-200:])
                 stats["a5_heartbeat_trimmed"] = True
-                log(f"A5: Trimmed heartbeat-log.json {before}→200 entries")
-        except Exception as e:
-            log(f"A5: WARN: {e}")
+        except Exception:
+            pass
 
     # A6: Trim reflection-log.json to 60 entries
     if REFLECTION_LOG.exists():
         try:
             data = load_json(REFLECTION_LOG, [])
             if isinstance(data, list) and len(data) > 60:
-                before = len(data)
-                data = data[-60:]
-                save_json(REFLECTION_LOG, data)
+                save_json(REFLECTION_LOG, data[-60:])
                 stats["a6_reflection_trimmed"] = True
-                log(f"A6: Trimmed reflection-log.json {before}→60 entries")
-        except Exception as e:
-            log(f"A6: WARN: {e}")
+        except Exception:
+            pass
 
-    # A7: Delete daemon/logs/ dirs >14 days (subdirs named YYYY-MM-DD)
+    # A7: Delete daemon/logs/ dirs >14 days
     cutoff_14 = TODAY - timedelta(days=14)
     if DAEMON_LOGS_DIR.exists():
         for d_path in sorted(DAEMON_LOGS_DIR.iterdir()):
@@ -287,513 +245,335 @@ def stage_a_cleanup() -> dict:
                 try:
                     shutil.rmtree(d_path)
                     stats["a7_daemon_logs_deleted"] += 1
-                    log(f"A7: Deleted daemon/logs/{d_path.name}/ (>{14}d)")
-                except Exception as e:
-                    log(f"A7: WARN: Could not delete {d_path.name}: {e}")
+                except Exception:
+                    pass
 
-    # A8: Delete cortex/logs/*.log >7 days, truncate launchd logs >5MB
+    # A8: Delete cortex/logs/*.log >7 days, truncate launchd >5MB
     cutoff_7 = TODAY - timedelta(days=7)
-    cortex_logs = Path.home() / "claude" / "cortex" / "logs"
+    cortex_logs = Path.home() / ".nexo" / "cortex" / "logs"
     if cortex_logs.exists():
         for f in cortex_logs.glob("*.log"):
             if f.name.startswith("launchd-"):
                 try:
                     if f.stat().st_size > 5 * 1024 * 1024:
                         lines = f.read_text().splitlines()
-                        keep = lines[-500:] if len(lines) > 500 else lines
-                        f.write_text("\n".join(keep) + "\n")
+                        f.write_text("\n".join(lines[-500:]) + "\n")
                         stats["a3_logs_rotated"] += 1
-                        log(f"A8: Truncated cortex {f.name}")
-                except Exception as e:
-                    log(f"A8: WARN: {e}")
+                except Exception:
+                    pass
                 continue
             d = parse_date_from_stem(f.stem)
             if d and d < cutoff_7:
                 try:
                     f.unlink()
-                    log(f"A8: Deleted cortex log {f.name} (>7d)")
-                except Exception as e:
-                    log(f"A8: WARN: Could not delete {f.name}: {e}")
+                except Exception:
+                    pass
 
     return stats
 
 
-# ─── Stage C: Learning Consolidation ─────────────────────────────────────────
+# ─── Stage B: Dreaming (Claude CLI) ─────────────────────────────────────────
 
-STOPWORDS = {
-    "el", "la", "los", "las", "un", "una", "unos", "unas",
-    "de", "del", "al", "en", "y", "o", "a", "con", "por", "para",
-    "que", "es", "se", "no", "si", "lo", "le", "su", "sus",
-    "the", "a", "an", "of", "in", "and", "or", "to", "for", "is",
-    "it", "on", "at", "by", "from", "with", "not", "be", "as",
-    "this", "that", "are", "was", "were",
-}
+def collect_brain_state() -> dict:
+    """Collect all data the CLI needs to dream."""
+    state = {"learnings": [], "preferences": [], "memory_md_lines": 0,
+             "claude_mem_old": 0, "feedback_count": 0}
 
+    if NEXO_DB.exists():
+        try:
+            conn = sqlite3.connect(str(NEXO_DB))
+            conn.row_factory = sqlite3.Row
 
-def _title_words(title: str) -> set:
-    """Lowercase, tokenize, remove stopwords from a title."""
-    words = re.findall(r'[a-záéíóúüñA-ZÁÉÍÓÚÜÑ\w]+', title.lower())
-    return {w for w in words if w not in STOPWORDS and len(w) > 2}
+            # Learnings
+            rows = conn.execute(
+                "SELECT id, title, content, category, created_at FROM learnings "
+                "WHERE status='active' ORDER BY id"
+            ).fetchall()
+            state["learnings"] = [dict(r) for r in rows]
 
+            # Preferences
+            rows = conn.execute("SELECT key, value, category, updated_at FROM preferences").fetchall()
+            state["preferences"] = [dict(r) for r in rows]
 
-def _word_overlap(words_a: set, words_b: set) -> float:
-    """Jaccard-like overlap: intersection / union."""
-    if not words_a or not words_b:
-        return 0.0
-    return len(words_a & words_b) / len(words_a | words_b)
+            conn.close()
+        except Exception as e:
+            log(f"DB error: {e}")
 
+    # MEMORY.md
+    if MEMORY_MD.exists():
+        state["memory_md_lines"] = len(MEMORY_MD.read_text().splitlines())
 
-def stage_c_learning_consolidation() -> dict:
-    """
-    Pure Python analysis of the learnings table in nexo.db.
-    Reads only — no deletions.
-    Returns stats dict stored under run_log['stage_c'].
-    """
-    stats = {
-        "total_learnings": 0,
-        "potential_duplicates": [],       # max 10
-        "age_distribution": {"<7d": 0, "7-30d": 0, "30-90d": 0, ">90d": 0},
-        "category_counts": {},
-        "hottest_category_7d": None,
-        "categories_over_20": [],
-        "potential_contradictions": [],   # max 5
-    }
+    # claude-mem.db old observations
+    if CLAUDE_MEM_DB.exists():
+        try:
+            cutoff = int((datetime.now() - timedelta(days=60)).timestamp() * 1000)
+            conn = sqlite3.connect(str(CLAUDE_MEM_DB))
+            state["claude_mem_old"] = conn.execute(
+                "SELECT COUNT(*) FROM observations WHERE created_at_epoch < ?", (cutoff,)
+            ).fetchone()[0]
+            conn.close()
+        except Exception:
+            pass
 
-    if not NEXO_DB.exists():
-        log("Stage C: nexo.db not found, skipping.")
-        return stats
+    # Feedback count
+    state["feedback_count"] = len(list(MEMORY_MD.parent.glob("feedback_*.md")))
 
-    try:
-        conn = sqlite3.connect(str(NEXO_DB))
-        conn.row_factory = sqlite3.Row
-        cursor = conn.cursor()
+    return state
 
-        # Check table exists
-        cursor.execute(
-            "SELECT name FROM sqlite_master WHERE type='table' AND name='learnings'"
-        )
-        if not cursor.fetchone():
-            log("Stage C: learnings table not found, skipping.")
-            conn.close()
-            return stats
 
-        cursor.execute(
-            "SELECT id, title, content, category, created_at FROM learnings ORDER BY id"
-        )
-        rows = cursor.fetchall()
-        conn.close()
-    except Exception as e:
-        log(f"Stage C: DB error: {e}")
-        return stats
-
-    if not rows:
-        log("Stage C: No learnings found.")
-        return stats
-
-    stats["total_learnings"] = len(rows)
-    now_dt = datetime.now()
-    cutoff_7 = now_dt - timedelta(days=7)
-    cutoff_30 = now_dt - timedelta(days=30)
-    cutoff_90 = now_dt - timedelta(days=90)
-
-    # Pre-compute per-row data
-    parsed = []
-    category_7d_counts: dict[str, int] = {}
-
-    for row in rows:
-        # Parse created_at (stored as epoch float or ISO string)
-        created_dt = None
-        raw_ts = row["created_at"]
-        if raw_ts:
-            # Try epoch first (nexo.db uses epoch floats)
-            try:
-                ts_float = float(raw_ts)
-                if ts_float > 1_000_000_000:  # reasonable epoch
-                    created_dt = datetime.fromtimestamp(ts_float)
-            except (ValueError, TypeError, OSError):
-                pass
-            # Fallback to ISO string formats
-            if created_dt is None:
-                for fmt in ("%Y-%m-%d %H:%M:%S", "%Y-%m-%dT%H:%M:%S", "%Y-%m-%d"):
-                    try:
-                        created_dt = datetime.strptime(str(raw_ts)[:19], fmt)
-                        break
-                    except ValueError:
-                        continue
-
-        words = _title_words(row["title"] or "")
-        cat = (row["category"] or "uncategorized").strip()
-
-        parsed.append({
-            "id": row["id"],
-            "title": row["title"] or "",
-            "words": words,
-            "category": cat,
-            "created_dt": created_dt,
-        })
-
-        # C2: age distribution
-        if created_dt:
-            if created_dt >= cutoff_7:
-                stats["age_distribution"]["<7d"] += 1
-                category_7d_counts[cat] = category_7d_counts.get(cat, 0) + 1
-            elif created_dt >= cutoff_30:
-                stats["age_distribution"]["7-30d"] += 1
-            elif created_dt >= cutoff_90:
-                stats["age_distribution"]["30-90d"] += 1
-            else:
-                stats["age_distribution"][">90d"] += 1
-        else:
-            # Unknown age → bucket as >90d
-            stats["age_distribution"][">90d"] += 1
-
-        # C3: category counts
-        stats["category_counts"][cat] = stats["category_counts"].get(cat, 0) + 1
-
-    # C3: hottest category last 7d + categories over 20
-    if category_7d_counts:
-        stats["hottest_category_7d"] = max(category_7d_counts, key=lambda k: category_7d_counts[k])
-    stats["categories_over_20"] = [
-        cat for cat, cnt in stats["category_counts"].items() if cnt > 20
-    ]
-
-    # C1: Duplicate detection — O(n²) but learnings table is small
-    duplicates = []
-    for i in range(len(parsed)):
-        if len(duplicates) >= 10:
-            break
-        for j in range(i + 1, len(parsed)):
-            if len(duplicates) >= 10:
-                break
-            overlap = _word_overlap(parsed[i]["words"], parsed[j]["words"])
-            if overlap >= 0.80:
-                duplicates.append({
-                    "id1": parsed[i]["id"],
-                    "id2": parsed[j]["id"],
-                    "title1": parsed[i]["title"],
-                    "title2": parsed[j]["title"],
-                    "overlap": round(overlap, 2),
-                })
-    stats["potential_duplicates"] = duplicates
-
-    # C4: Contradiction detection — NUNCA pairs in same category
-    nunca_entries = [p for p in parsed if "nunca" in p["title"].lower()]
-    contradictions = []
-    for nunca in nunca_entries:
-        if len(contradictions) >= 5:
-            break
-        # Look for same-category entries that don't contain NUNCA
-        # and whose remaining words overlap significantly (same subject, opposite stance)
-        nunca_words_no_nunca = nunca["words"] - {"nunca"}
-        for other in parsed:
-            if len(contradictions) >= 5:
-                break
-            if other["id"] == nunca["id"]:
-                continue
-            if other["category"] != nunca["category"]:
-                continue
-            if "nunca" in other["title"].lower():
-                continue
-            # Check if they share meaningful subject words
-            overlap = _word_overlap(nunca_words_no_nunca, other["words"])
-            if overlap >= 0.50:
-                contradictions.append({
-                    "id1": nunca["id"],
-                    "id2": other["id"],
-                    "title1": nunca["title"],
-                    "title2": other["title"],
-                })
-    stats["potential_contradictions"] = contradictions
-
-    log(f"Stage C: {stats['total_learnings']} learnings analyzed. "
-        f"Potential duplicates: {len(duplicates)}. "
-        f"Categories over 20: {len(stats['categories_over_20'])}. "
-        f"Potential contradictions: {len(contradictions)}.")
-    if stats["hottest_category_7d"]:
-        log(f"Stage C: Hottest category last 7d: {stats['hottest_category_7d']} "
-            f"({category_7d_counts.get(stats['hottest_category_7d'], 0)} new).")
+def should_dream(state: dict) -> bool:
+    """Check if there's enough to justify a CLI call."""
+    return (
+        len(state["learnings"]) > 10
+        or state["memory_md_lines"] > 170
+        or len(state["preferences"]) > 5
+        or state["claude_mem_old"] > 500
+    )
 
-    return stats
 
+def dream(state: dict) -> dict:
+    """The brain dreams — CLI does the intelligent work."""
 
-# ─── Stage B: Intelligent pruning (Claude CLI) ──────────────────────────────
-
-def check_stage_b_conditions() -> dict:
-    """
-    Check if Stage B should activate.
-    Returns dict with condition results and whether to trigger.
-    """
-    conditions = {
-        "memory_md_lines": 0,
-        "memory_md_over_limit": False,
-        "preferences_auto_sections": 0,
-        "preferences_over_limit": False,
-        "claude_mem_old_observations": 0,
-        "claude_mem_over_limit": False,
-        "should_trigger": False,
-    }
+    # Truncate learnings JSON if too large
+    learnings_json = json.dumps(state["learnings"], ensure_ascii=False, indent=1)
+    if len(learnings_json) > 15000:
+        learnings_json = learnings_json[:15000] + "\n... (truncated)"
 
-    # Check MEMORY.md line count
-    if MEMORY_MD.exists():
-        try:
-            lines = MEMORY_MD.read_text().splitlines()
-            conditions["memory_md_lines"] = len(lines)
-            conditions["memory_md_over_limit"] = len(lines) > 170
-        except Exception as e:
-            log(f"Stage B check: WARN reading MEMORY.md: {e}")
+    tasks = []
 
-    # Check preferences count in SQLite
-    if NEXO_DB.exists():
-        try:
-            conn = sqlite3.connect(str(NEXO_DB))
-            cursor = conn.cursor()
-            cursor.execute("SELECT COUNT(*) FROM preferences")
-            count = cursor.fetchone()[0]
-            conn.close()
-            conditions["preferences_auto_sections"] = count
-            conditions["preferences_over_limit"] = count > 5
-        except Exception as e:
-            log(f"Stage B check: WARN reading nexo.db preferences: {e}")
+    tasks.append(f"""TASK 1: LEARNING CONSOLIDATION ({len(state['learnings'])} active)
+Review these learnings and identify:
+a) DUPLICATES: learnings that say the same thing differently.
+b) CONTRADICTIONS: learnings that contradict each other.
+c) STALE: learnings about bugs/issues fixed >60 days ago that are never referenced.
 
-    # Check claude-mem.db observations >60 days
-    if CLAUDE_MEM_DB.exists():
-        try:
-            cutoff_epoch = int((datetime.now() - timedelta(days=60)).timestamp() * 1000)
-            conn = sqlite3.connect(str(CLAUDE_MEM_DB))
-            cursor = conn.cursor()
-            cursor.execute(
-                "SELECT COUNT(*) FROM observations WHERE created_at_epoch < ?",
-                (cutoff_epoch,)
-            )
-            count = cursor.fetchone()[0]
-            conn.close()
-            conditions["claude_mem_old_observations"] = count
-            conditions["claude_mem_over_limit"] = count > 500
-        except Exception as e:
-            log(f"Stage B check: WARN reading claude-mem.db: {e}")
+Write your findings to {COORD_DIR}/sleep-report.md with sections:
+- "## Duplicates to archive" — list learning IDs to archive and why
+- "## Contradictions" — pairs of conflicting learnings
+- "## Stale candidates" — IDs of learnings that may be obsolete
 
-    conditions["should_trigger"] = (
-        conditions["memory_md_over_limit"]
-        or conditions["preferences_over_limit"]
-        or conditions["claude_mem_over_limit"]
-    )
+Also write a machine-readable file {COORD_DIR}/sleep-actions.json:
+{{"archive_ids": [1, 2, 3], "contradiction_pairs": [[4, 5]], "stale_ids": [6, 7]}}
 
-    return conditions
+The wrapper will execute the actual DB operations based on this JSON.
 
+LEARNINGS:
+{learnings_json}""")
 
-def build_stage_b_prompt(conditions: dict) -> str:
-    """Build the prompt for Claude CLI based on which conditions triggered."""
-    tasks = []
+    if state["memory_md_lines"] > 170:
+        tasks.append(f"""TASK 2: MEMORY.MD COMPRESSION ({state['memory_md_lines']} lines, limit 200)
+File: {MEMORY_MD}
+Read it, compress resolved incidents >21 days, merge duplicates.
+NEVER delete: credentials, legal entity info, CRITICAL rules, infrastructure.
+Target: <180 lines.""")
+
+    if len(state["preferences"]) > 5:
+        tasks.append(f"""TASK 3: PREFERENCES CLEANUP ({len(state['preferences'])} entries)
+Review the preferences and identify duplicate keys.
+Add to sleep-actions.json: "duplicate_preference_keys": ["key1", "key2", ...]
+The wrapper will handle the actual DB cleanup safely.""")
 
-    if conditions["memory_md_over_limit"]:
-        tasks.append(f"""TAREA 1: MEMORY.md ({conditions['memory_md_lines']} lineas, limite 200)
-Archivo: {MEMORY_MD}
-Lee con Read tool, comprime incidentes resueltos >21 dias, fusiona duplicados, mantener <180 lineas.
-PRESERVA toda la estructura de secciones existente. No elimines secciones enteras.""")
-
-    if conditions["preferences_over_limit"]:
-        tasks.append(f"""TAREA 2: preferences en SQLite ({conditions['preferences_auto_sections']} registros)
-DB: {NEXO_DB}, tabla: preferences (columnas: key, value, category, updated_at)
-Conecta con sqlite3. Elimina preferencias duplicadas (mismo key) manteniendo la mas reciente.
-Elimina preferencias con updated_at mas antiguo de 30 dias si hay un duplicado mas reciente.
-Reporta cuantos registros eliminaste.""")
-
-    if conditions["claude_mem_over_limit"]:
-        tasks.append(f"""TAREA 3: claude-mem observations ({conditions['claude_mem_old_observations']} registros >60d)
-DB: {CLAUDE_MEM_DB}
-Conecta con sqlite3. Ejecuta:
-  DELETE FROM observations WHERE created_at_epoch < {int((datetime.now() - timedelta(days=60)).timestamp() * 1000)}
-    AND discovery_tokens < 300
-    AND id NOT IN (SELECT id FROM observations WHERE
-        title LIKE '%CRITICO%' OR title LIKE '%MAXIMA%'
-        OR title LIKE '%credential%' OR title LIKE '%token%' OR title LIKE '%API%'
-        OR narrative LIKE '%CRITICO%' OR narrative LIKE '%MAXIMA%')
-  LIMIT 200;
-Luego: DELETE FROM observations_fts WHERE rowid NOT IN (SELECT id FROM observations);
-Luego: VACUUM;
-Reporta cuantos registros eliminaste.""")
-
-    if not tasks:
-        return ""
+    if state["claude_mem_old"] > 500:
+        tasks.append(f"""TASK 4: OLD OBSERVATIONS ({state['claude_mem_old']} entries >60d)
+Note in sleep-report.md that old observations should be cleaned.
+Add to sleep-actions.json: "clean_old_observations": true
+The wrapper will handle the actual DB cleanup safely.""")
 
     tasks_str = "\n\n".join(tasks)
 
-    return f"""You are NEXO Sleep System. Your job is to PRUNE memory.
-You are NOT interactive. Do NOT wait for input. Execute the following tasks and exit.
+    prompt = f"""You are NEXO Sleep — the nightly brain maintenance process.
+Like a human brain during sleep: consolidate important memories, discard noise,
+detect conflicts, prepare state for tomorrow.
 
-ABSOLUTE RULES:
-- NEVER delete credentials, tokens, account IDs, API endpoints, keys, secrets.
-- NEVER delete operational rules marked as "CRITICAL" or "HIGHEST PRIORITY".
-- NEVER delete information about infrastructure (servers, repos, deploys).
-- You CAN merge redundant sections.
-- You CAN remove obsolete technical information (fixed >30 days ago and never referenced since).
-- You CAN compress long paragraphs into concise bullets.
-- Every line you remove must have a clear reason. When in doubt, do NOT delete.
+BRAIN STATE:
+- {len(state['learnings'])} active learnings
+- {state['memory_md_lines']} lines in MEMORY.md (limit: 200)
+- {len(state['preferences'])} preferences
+- {state['feedback_count']} feedback files
+- {state['claude_mem_old']} old observations (>60d)
 
 {tasks_str}
 
-Al terminar, imprime un resumen JSON con las acciones realizadas."""
-
+ABSOLUTE RULES:
+- NEVER delete legal entity info (LLC, SLU, EIN, NIF, project)
+- NEVER delete credentials, tokens, API keys, secrets
+- NEVER delete rules marked CRITICAL or MAX PRIORITY
+- NEVER delete infrastructure info (servers, repos, deploys)
+- When in doubt, DON'T delete
 
-def run_stage_b(conditions: dict) -> dict:
-    """Run Stage B using Claude CLI."""
-    prompt = build_stage_b_prompt(conditions)
-    if not prompt:
-        return {"skipped": True, "reason": "No tasks to run"}
+Write a summary to {COORD_DIR}/sleep-report.md when done.
+Execute without asking."""
 
-    if not CLAUDE_CLI.exists():
-        return {"error": f"Claude CLI not found at {CLAUDE_CLI}"}
+    log("Stage B: Invoking Claude CLI (opus) — dreaming...")
 
-    log("Stage B: Invoking Claude CLI (sonnet)...")
+    env = os.environ.copy()
+    env.pop("CLAUDECODE", None)
+    env.pop("CLAUDE_CODE", None)
 
     try:
-        env = os.environ.copy()
-        # Remove env vars that would cause Claude CLI to think it's inside Claude Code
-        env.pop("CLAUDECODE", None)
-        env.pop("CLAUDE_CODE", None)
-
         result = subprocess.run(
-            [str(CLAUDE_CLI), "-p", prompt, "--model", "sonnet"],
-            capture_output=True,
-            text=True,
-            timeout=600,
-            env=env
+            [str(CLAUDE_CLI), "-p", prompt, "--model", "opus",
+             "--allowedTools", "Read,Write,Edit,Glob,Grep"],
+            capture_output=True, text=True, timeout=600, env=env
         )
 
-        stdout = result.stdout.strip() if result.stdout else ""
-        stderr = result.stderr.strip() if result.stderr else ""
-
         if result.returncode != 0:
-            log(f"Stage B: Claude CLI returned code {result.returncode}")
-            if stderr:
-                log(f"Stage B: stderr: {stderr[:500]}")
-            return {
-                "returncode": result.returncode,
-                "stderr": stderr[:500],
-                "stdout": stdout[:500],
-            }
-
-        log(f"Stage B: Completed. Output length: {len(stdout)} chars")
-        return {
-            "returncode": 0,
-            "output_length": len(stdout),
-            "output_preview": stdout[:800],
-        }
+            log(f"Stage B: CLI error ({result.returncode}): {(result.stderr or '')[:300]}")
+            return {"error": result.returncode}
+
+        log(f"Stage B: Dreaming complete. Output: {len(result.stdout or '')} chars")
+        return {"ok": True, "output_len": len(result.stdout or "")}
 
     except subprocess.TimeoutExpired:
-        log("Stage B: Claude CLI timed out (600s)")
+        log("Stage B: CLI timed out (600s)")
         return {"error": "timeout"}
     except Exception as e:
         log(f"Stage B: Exception: {e}")
         return {"error": str(e)}
 
 
-# ─── Main ─────────────────────────────────────────────────────────────────────
+def execute_dream_actions(actions: dict, state: dict):
+    """Execute the DB actions decided by CLI, safely in Python."""
+    log("Stage B2: Executing dream actions...")
+
+    # Archive duplicate/stale learnings
+    archive_ids = actions.get("archive_ids", []) + actions.get("stale_ids", [])
+    if archive_ids and NEXO_DB.exists():
+        try:
+            conn = sqlite3.connect(str(NEXO_DB))
+            for lid in archive_ids:
+                if isinstance(lid, int):
+                    conn.execute(
+                        "UPDATE learnings SET status='archived' WHERE id=? AND status='active'",
+                        (lid,)
+                    )
+            conn.commit()
+            conn.close()
+            log(f"  Archived {len(archive_ids)} learnings: {archive_ids}")
+        except Exception as e:
+            log(f"  Error archiving learnings: {e}")
+
+    # Clean duplicate preferences
+    dup_keys = actions.get("duplicate_preference_keys", [])
+    if dup_keys and NEXO_DB.exists():
+        try:
+            conn = sqlite3.connect(str(NEXO_DB))
+            for key in dup_keys:
+                if isinstance(key, str):
+                    # Keep newest, delete older duplicates
+                    conn.execute(
+                        "DELETE FROM preferences WHERE key = ? AND rowid NOT IN "
+                        "(SELECT rowid FROM preferences WHERE key = ? ORDER BY updated_at DESC LIMIT 1)",
+                        (key, key)
+                    )
+            conn.commit()
+            conn.close()
+            log(f"  Cleaned {len(dup_keys)} duplicate preference keys")
+        except Exception as e:
+            log(f"  Error cleaning preferences: {e}")
+
+    # Clean old observations
+    if actions.get("clean_old_observations") and CLAUDE_MEM_DB.exists():
+        try:
+            cutoff_ms = int((datetime.now() - timedelta(days=60)).timestamp() * 1000)
+            conn = sqlite3.connect(str(CLAUDE_MEM_DB))
+            deleted = conn.execute(
+                "DELETE FROM observations WHERE created_at_epoch < ? "
+                "AND discovery_tokens < 300 "
+                "AND id NOT IN (SELECT id FROM observations WHERE "
+                "title LIKE '%CRITICO%' OR title LIKE '%credential%' "
+                "OR title LIKE '%token%' OR title LIKE '%API%' "
+                "OR title LIKE '%LLC%' OR title LIKE '%SLU%') "
+                "LIMIT 200",
+                (cutoff_ms,)
+            ).rowcount
+            conn.execute(
+                "DELETE FROM observations_fts WHERE rowid NOT IN "
+                "(SELECT id FROM observations)"
+            )
+            conn.execute("VACUUM")
+            conn.commit()
+            conn.close()
+            log(f"  Cleaned {deleted} old observations")
+        except Exception as e:
+            log(f"  Error cleaning observations: {e}")
+
+    log("Stage B2: Actions complete.")
+
+
+# ─── Main ────────────────────────────────────────────────────────────────────
 
 def main():
     log("=" * 60)
-    log("NEXO Sleep System starting")
+    log("NEXO Sleep System v2 starting")
 
-    # Process lock via fcntl to prevent concurrent instances
+    # Process lock
     try:
         lock_fd = open(PROCESS_LOCK, "w")
         fcntl.flock(lock_fd, fcntl.LOCK_EX | fcntl.LOCK_NB)
         lock_fd.write(str(os.getpid()))
         lock_fd.flush()
     except (IOError, OSError):
-        log("Another sleep instance is already running. Exiting.")
+        log("Another sleep instance running. Exiting.")
         sys.exit(0)
 
     try:
-        # Check if already completed today
         if already_ran_today():
             log("Already ran today. Exiting.")
             sys.exit(0)
 
-        # Determine start phase (for resume after interruption)
         start_phase = "stage_a"
         if was_interrupted():
             start_phase = get_interrupted_phase()
-            log(f"Resuming from phase: {start_phase}")
-
-        run_log = {
-            "date": str(TODAY),
-            "started": TIMESTAMP,
-            "stage_a": None,
-            "stage_c": None,
-            "stage_b_conditions": None,
-            "stage_b": None,
-            "completed": None,
-        }
-
-        # Stage A: Mechanical cleanup
-        if start_phase in ("stage_a",):
+
+        run_log = {"date": str(TODAY), "started": TIMESTAMP,
+                   "stage_a": None, "stage_b": None, "completed": None}
+
+        # Stage A: Housekeeping (mechanical)
+        if start_phase == "stage_a":
             set_lock("stage_a")
-            log("─── Stage A: Mechanical cleanup ───")
-            stage_a_stats = stage_a_cleanup()
-            run_log["stage_a"] = stage_a_stats
-
-            total_cleaned = (
-                stage_a_stats["a1_daily_summaries_deleted"]
-                + stage_a_stats["a2_session_archives_deleted"]
-                + stage_a_stats["a3_logs_rotated"]
-                + stage_a_stats["a4_compressed_memories_deleted"]
-                + stage_a_stats["a7_daemon_logs_deleted"]
-            )
-            log(f"Stage A complete: {total_cleaned} items cleaned, "
-                f"heartbeat trimmed={stage_a_stats['a5_heartbeat_trimmed']}, "
-                f"reflection trimmed={stage_a_stats['a6_reflection_trimmed']}")
-
-        # Stage C: Learning Consolidation (always runs, pure Python)
-        if start_phase in ("stage_a", "stage_c", "stage_b"):
-            set_lock("stage_c")
-            log("─── Stage C: Learning Consolidation ───")
-            stage_c_stats = stage_c_learning_consolidation()
-            run_log["stage_c"] = stage_c_stats
-
-        # Stage B: Intelligent pruning (conditional)
-        if start_phase in ("stage_a", "stage_c", "stage_b"):
-            set_lock("stage_b")
-            log("─── Stage B: Checking conditions ───")
-            conditions = check_stage_b_conditions()
-            run_log["stage_b_conditions"] = conditions
-
-            log(f"  MEMORY.md: {conditions['memory_md_lines']} lines "
-                f"(trigger={conditions['memory_md_over_limit']})")
-            log(f"  nexo.db preferences: {conditions['preferences_auto_sections']} rows "
-                f"(trigger={conditions['preferences_over_limit']})")
-            log(f"  claude-mem old observations: {conditions['claude_mem_old_observations']} "
-                f"(trigger={conditions['claude_mem_over_limit']})")
-
-            if conditions["should_trigger"]:
-                log("Stage B: Conditions met, running intelligent pruning...")
-                stage_b_result = run_stage_b(conditions)
-                run_log["stage_b"] = stage_b_result
-            else:
-                log("Stage B: No conditions met, skipping.")
-                run_log["stage_b"] = {"skipped": True, "reason": "No conditions met"}
-
-        # Mark complete
+            log("─── Stage A: Housekeeping ───")
+            run_log["stage_a"] = stage_a_cleanup()
+
+        # Stage B: Dreaming (intelligent)
+        set_lock("stage_b")
+        log("─── Stage B: Dreaming ───")
+        state = collect_brain_state()
+
+        if should_dream(state):
+            log(f"Brain state: {len(state['learnings'])} learnings, "
+                f"{state['memory_md_lines']} MEMORY lines, "
+                f"{state['claude_mem_old']} old observations")
+            run_log["stage_b"] = dream(state)
+
+            # Stage B2: Execute actions from CLI output
+            actions_file = COORD_DIR / "sleep-actions.json"
+            if actions_file.exists():
+                try:
+                    actions = json.loads(actions_file.read_text())
+                    execute_dream_actions(actions, state)
+                except Exception as e:
+                    log(f"Stage B2: Error executing actions: {e}")
+        else:
+            log("Brain is clean — no dreaming needed.")
+            run_log["stage_b"] = {"skipped": True}
+
+        # Done
         run_log["completed"] = datetime.now().strftime("%Y-%m-%d %H:%M")
         mark_complete()
         append_sleep_log(run_log)
+        log(f"NEXO Sleep v2 complete at {run_log['completed']}")
 
-        log(f"NEXO Sleep complete at {run_log['completed']}")
-
-        # Register successful run for catch-up
+        # Register for catch-up
         try:
-            import json as _json
-            _state_file = Path.home() / "claude" / "operations" / ".catchup-state.json"
-            _state = _json.loads(_state_file.read_text()) if _state_file.exists() else {}
-            _state["sleep"] = datetime.now().isoformat()
-            _state_file.write_text(_json.dumps(_state, indent=2))
+            state_file = Path.home() / ".nexo" / "operations" / ".catchup-state.json"
+            st = json.loads(state_file.read_text()) if state_file.exists() else {}
+            st["sleep"] = datetime.now().isoformat()
+            state_file.write_text(json.dumps(st, indent=2))
         except Exception:
             pass
 
-        log("=" * 60)
-
     finally:
-        # Release process lock
         try:
             fcntl.flock(lock_fd, fcntl.LOCK_UN)
             lock_fd.close()