superlocalmemory 3.4.10 → 3.4.11

package/src/superlocalmemory/evolution/triggers.py

@@ -0,0 +1,367 @@
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under AGPL-3.0-or-later - see LICENSE file
+# Part of SuperLocalMemory V3 | https://qualixar.com | https://varunpratap.com
+
+"""Evolution Triggers — detect when skills need evolution.
+
+3-trigger system (adopted from OpenSpace):
+  1. Post-session analysis: session ends, scan for skill failures
+  2. Skill degradation: behavioral assertion confidence drops
+  3. Periodic health check: consolidation cycle scans all skills
+
+All triggers are zero-LLM, zero-embedding. Pure SQLite queries.
+They produce EvolutionCandidates — the evolver decides what to do.
+
+Part of Qualixar | Author: Varun Pratap Bhardwaj
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import logging
+import re
+import sqlite3
+from datetime import datetime, timezone
+from pathlib import Path
+
+from superlocalmemory.evolution.types import (
+    EvolutionCandidate,
+    EvolutionType,
+    TriggerType,
+)
+
+logger = logging.getLogger(__name__)
+
+# Thresholds — conservative, matching C1 design doc
+DEGRADATION_THRESHOLD = 0.4      # Effective score below this → FIX candidate
+RECOVERY_THRESHOLD = 0.6         # Above this → skill recovered, prune from addressed
+DERIVED_THRESHOLD = 0.55         # Moderate effectiveness → DERIVED candidate
+MIN_INVOCATIONS = 5              # Don't trigger on insufficient data
+NEGATIVE_SIGNALS_THRESHOLD = 2   # Min negative signals in one session for post-session trigger
+MAX_RSS_MB = 1024                # Skip evolution if process exceeds 1GB RSS
+
+
+def _check_memory_pressure() -> bool:
+    """Return True if process RSS exceeds threshold. Skip evolution if so."""
+    try:
+        import psutil
+        rss_mb = psutil.Process().memory_info().rss / (1024 * 1024)
+        if rss_mb > MAX_RSS_MB:
+            logger.warning(
+                "Memory pressure: %dMB RSS > %dMB limit, skipping evolution",
+                int(rss_mb), MAX_RSS_MB,
+            )
+            return True
+    except ImportError:
+        pass
+    return False
+
+
+class PostSessionTrigger:
+    """Trigger 1: Analyze a completed session for skill failures.
+
+    Scans tool_events for the given session_id. If any Skill call
+    had 2+ negative signals (retries, errors), creates a FIX candidate.
+
+    Zero-LLM. Runs in ~10ms on indexed table.
+    """
+
+    def __init__(self, db_path: str | Path):
+        self._db_path = str(db_path)
+
+    def scan(self, session_id: str, profile_id: str = "default") -> list[EvolutionCandidate]:
+        if _check_memory_pressure():
+            return []
+
+        candidates = []
+        conn = sqlite3.connect(self._db_path, timeout=10)
+        conn.row_factory = sqlite3.Row
+
+        try:
+            # Get all Skill events for this session
+            rows = conn.execute(
+                "SELECT id, tool_name, input_summary, output_summary, created_at "
+                "FROM tool_events "
+                "WHERE session_id = ? AND profile_id = ? AND tool_name = 'Skill' "
+                "ORDER BY id ASC",
+                (session_id, profile_id),
+            ).fetchall()
+
+            if not rows:
+                return []
+
+            # Parse skill names and count per-skill signals
+            skill_signals: dict[str, dict] = {}
+            for row in rows:
+                d = dict(row)
+                skill_name = self._extract_skill_name(d)
+                if not skill_name:
+                    continue
+
+                if skill_name not in skill_signals:
+                    skill_signals[skill_name] = {
+                        "positive": 0, "negative": 0, "event_ids": [],
+                    }
+                skill_signals[skill_name]["event_ids"].append(d["id"])
+
+            # Check execution traces for each skill
+            for skill_name, signals in skill_signals.items():
+                for event_id in signals["event_ids"]:
+                    trace = conn.execute(
+                        "SELECT tool_name, output_summary FROM tool_events "
+                        "WHERE profile_id = ? AND id > ? ORDER BY id ASC LIMIT 5",
+                        (profile_id, event_id),
+                    ).fetchall()
+
+                    has_error = any(
+                        "error" in (dict(t).get("output_summary", "") or "").lower()
+                        for t in trace[:3]
+                        if dict(t)["tool_name"] == "Bash"
+                    )
+                    has_retry = any(
+                        dict(t)["tool_name"] == "Skill"
+                        for t in trace[:3]
+                    )
+
+                    if has_error or has_retry:
+                        signals["negative"] += 1
+                    else:
+                        signals["positive"] += 1
+
+                total_invocations = len(signals["event_ids"])
+                if signals["negative"] >= NEGATIVE_SIGNALS_THRESHOLD and total_invocations >= MIN_INVOCATIONS:
+                    evidence = (
+                        f"{signals['negative']} negative signals in session {session_id}",
+                        f"{signals['positive']} positive signals",
+                    )
+                    candidates.append(EvolutionCandidate(
+                        skill_name=skill_name,
+                        evolution_type=EvolutionType.FIX,
+                        trigger=TriggerType.POST_SESSION,
+                        evidence=evidence,
+                        effective_score=signals["positive"] / max(1, signals["positive"] + signals["negative"]),
+                        invocation_count=len(signals["event_ids"]),
+                        session_id=session_id,
+                    ))
+
+        except Exception as exc:
+            logger.debug("Post-session trigger scan failed: %s", exc)
+        finally:
+            conn.close()
+
+        return candidates
+
+    def _extract_skill_name(self, event: dict) -> str:
+        for field_key in ("input_summary", "output_summary"):
+            raw = event.get(field_key, "") or ""
+            if not raw:
+                continue
+            try:
+                parsed = json.loads(raw)
+                name = parsed.get("skill") or parsed.get("commandName", "")
+                if name:
+                    return name
+            except (json.JSONDecodeError, TypeError):
+                # M-JSONPARSE: Fallback for non-JSON plain text
+                m = re.search(r"skill[:\s]+['\"]?(\S+)", raw, re.IGNORECASE)
+                if m:
+                    return m.group(1).strip("'\"")
+        return ""
+
+
+class DegradationTrigger:
+    """Trigger 2: Detect skills with declining performance assertions.
+
+    Scans behavioral_assertions for skill_performance category.
+    If a skill's confidence dropped below threshold → FIX candidate.
+    If moderate with specific failure pattern → DERIVED candidate.
+
+    Zero-LLM. Runs in ~5ms.
+    """
+
+    def __init__(self, db_path: str | Path):
+        self._db_path = str(db_path)
+
+    def scan(self, profile_id: str = "default") -> list[EvolutionCandidate]:
+        if _check_memory_pressure():
+            return []
+
+        candidates = []
+        conn = sqlite3.connect(self._db_path, timeout=10)
+        conn.row_factory = sqlite3.Row
+
+        try:
+            rows = conn.execute(
+                "SELECT trigger_condition, action, confidence, evidence_count "
+                "FROM behavioral_assertions "
+                "WHERE profile_id = ? AND category = 'skill_performance' "
+                "AND confidence > 0",
+                (profile_id,),
+            ).fetchall()
+
+            for row in rows:
+                d = dict(row)
+                # H-TRIGGERPARSE: Extract skill name robustly.
+                # Primary: regex on trigger_condition. Fallback: old replace approach.
+                skill_name = ""
+                tc = d["trigger_condition"]
+                m = re.search(r"skill\s+(\S+)", tc)
+                if m:
+                    skill_name = m.group(1)
+                else:
+                    skill_name = tc.replace("when considering skill ", "")
+                confidence = d["confidence"]
+                evidence = d["evidence_count"]
+
+                if evidence < MIN_INVOCATIONS:
+                    continue
+
+                # Parse effective score from action text
+                effective_score = self._parse_effective_score(d["action"])
+
+                if effective_score < DEGRADATION_THRESHOLD:
+                    candidates.append(EvolutionCandidate(
+                        skill_name=skill_name,
+                        evolution_type=EvolutionType.FIX,
+                        trigger=TriggerType.DEGRADATION,
+                        evidence=(d["action"],),
+                        effective_score=effective_score,
+                        invocation_count=evidence,
+                    ))
+                elif effective_score < DERIVED_THRESHOLD:
+                    candidates.append(EvolutionCandidate(
+                        skill_name=skill_name,
+                        evolution_type=EvolutionType.DERIVED,
+                        trigger=TriggerType.DEGRADATION,
+                        evidence=(d["action"],),
+                        effective_score=effective_score,
+                        invocation_count=evidence,
+                    ))
+
+        except Exception as exc:
+            logger.debug("Degradation trigger scan failed: %s", exc)
+        finally:
+            conn.close()
+
+        return candidates
+
+    def get_active_degraded(self, profile_id: str = "default") -> set[str]:
+        """Return names of currently degraded skills (for anti-loop pruning)."""
+        conn = sqlite3.connect(self._db_path, timeout=10)
+        conn.row_factory = sqlite3.Row
+        degraded = set()
+        try:
+            rows = conn.execute(
+                "SELECT trigger_condition, action FROM behavioral_assertions "
+                "WHERE profile_id = ? AND category = 'skill_performance'",
+                (profile_id,),
+            ).fetchall()
+            for row in rows:
+                d = dict(row)
+                score = self._parse_effective_score(d["action"])
+                if score < RECOVERY_THRESHOLD:
+                    tc = d["trigger_condition"]
+                    m = re.search(r"skill\s+(\S+)", tc)
+                    name = m.group(1) if m else tc.replace("when considering skill ", "")
+                    degraded.add(name)
+        except Exception:
+            pass
+        finally:
+            conn.close()
+        return degraded
+
+    def _parse_effective_score(self, action_text: str) -> float:
+        """Extract effective score from assertion action text."""
+        import re
+        match = re.search(r"effective score:\s*([-\d.]+)%", action_text)
+        if match:
+            return float(match.group(1)) / 100.0
+        return 0.5
+
+
+class HealthCheckTrigger:
+    """Trigger 3: Periodic scan of all skill entities.
+
+    Runs every N consolidation cycles. Checks Entity Explorer for
+    skill entities with low performance.
+
+    Cycle count is persisted to the evolution_cycle_state table so it
+    survives process restarts (H-CYCLECNT fix).
+
+    Zero-LLM. Runs in ~20ms.
+    """
+
+    _STATE_KEY = "health_check_cycle_count"
+
+    def __init__(self, db_path: str | Path):
+        self._db_path = str(db_path)
+        self._check_every_n = 3  # Every 3rd consolidation (~18h)
+
+    def _read_cycle_count(self) -> int:
+        """Read persisted cycle count from DB."""
+        conn = sqlite3.connect(self._db_path, timeout=10)
+        try:
+            row = conn.execute(
+                "SELECT value FROM evolution_cycle_state WHERE key = ?",
+                (self._STATE_KEY,),
+            ).fetchone()
+            return int(row[0]) if row else 0
+        except sqlite3.OperationalError:
+            # Table may not exist yet
+            return 0
+        finally:
+            conn.close()
+
+    def _write_cycle_count(self, count: int) -> None:
+        """Persist cycle count to DB.
+
+        Relies on evolution_cycle_state table created by EvolutionStore.
+        Falls back to CREATE IF NOT EXISTS for standalone use.
+        """
+        conn = sqlite3.connect(self._db_path, timeout=10)
+        try:
+            # Table schema matches evolution_store.py: value INTEGER
+            conn.execute(
+                "CREATE TABLE IF NOT EXISTS evolution_cycle_state "
+                "(key TEXT PRIMARY KEY, value INTEGER DEFAULT 0, updated_at TEXT)",
+            )
+            now = datetime.now(timezone.utc).isoformat()
+            conn.execute(
+                "INSERT OR REPLACE INTO evolution_cycle_state (key, value, updated_at) "
+                "VALUES (?, ?, ?)",
+                (self._STATE_KEY, count, now),
+            )
+            conn.commit()
+        except sqlite3.OperationalError as exc:
+            logger.warning("Failed to persist cycle count: %s", exc)
+        finally:
+            conn.close()
+
+    def should_run(self) -> bool:
+        cycle_count = self._read_cycle_count() + 1
+        self._write_cycle_count(cycle_count)
+        return cycle_count % self._check_every_n == 0
+
+    def scan(self, profile_id: str = "default") -> list[EvolutionCandidate]:
+        if _check_memory_pressure():
+            return []
+        if not self.should_run():
+            return []
+
+        # Delegate to DegradationTrigger — same logic, different trigger label
+        deg = DegradationTrigger(self._db_path)
+        deg_candidates = deg.scan(profile_id)
+
+        # Re-label as health_check trigger
+        return [
+            EvolutionCandidate(
+                skill_name=c.skill_name,
+                evolution_type=c.evolution_type,
+                trigger=TriggerType.HEALTH_CHECK,
+                evidence=c.evidence,
+                effective_score=c.effective_score,
+                invocation_count=c.invocation_count,
+            )
+            for c in deg_candidates
+        ]

package/src/superlocalmemory/evolution/types.py

@@ -0,0 +1,92 @@
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under AGPL-3.0-or-later - see LICENSE file
+# Part of SuperLocalMemory V3 | https://qualixar.com | https://varunpratap.com
+
+"""Skill Evolution type definitions.
+
+Immutable data classes for evolution candidates, records, and lineage.
+All types are frozen dataclasses — no mutation after creation.
+
+Part of Qualixar | Author: Varun Pratap Bhardwaj
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from enum import Enum
+from typing import Optional
+
+
+class EvolutionType(str, Enum):
+    """How the skill is being evolved."""
+    FIX = "fix"           # Repair broken skill in-place
+    DERIVED = "derived"   # Create specialized variant
+    CAPTURED = "captured" # Extract new skill from patterns
+
+
+class TriggerType(str, Enum):
+    """What triggered the evolution."""
+    POST_SESSION = "post_session"   # Session Stop hook analysis
+    DEGRADATION = "degradation"     # Behavioral assertion confidence drop
+    HEALTH_CHECK = "health_check"   # Periodic consolidation scan
+
+
+class EvolutionStatus(str, Enum):
+    """Pipeline status."""
+    CANDIDATE = "candidate"     # Detected, not yet confirmed
+    CONFIRMED = "confirmed"     # LLM gate passed
+    MUTATED = "mutated"         # New SKILL.md generated
+    VERIFIED = "verified"       # Blind verification passed
+    PROMOTED = "promoted"       # Live — evolved skill active
+    REJECTED = "rejected"       # Failed verification or gate
+    FAILED = "failed"           # Error during evolution
+
+
+@dataclass(frozen=True)
+class EvolutionCandidate:
+    """A skill flagged for potential evolution."""
+    skill_name: str
+    evolution_type: EvolutionType
+    trigger: TriggerType
+    evidence: tuple[str, ...] = ()
+    effective_score: float = 0.0
+    invocation_count: int = 0
+    session_id: str = ""
+    project_path: str = ""
+
+
+@dataclass(frozen=True)
+class EvolutionRecord:
+    """Persisted record of an evolution attempt."""
+    id: str
+    skill_name: str
+    parent_skill_id: Optional[str]
+    evolution_type: EvolutionType
+    trigger: TriggerType
+    generation: int = 0
+    status: EvolutionStatus = EvolutionStatus.CANDIDATE
+    mutation_summary: str = ""
+    evidence: tuple[str, ...] = ()
+    original_content: str = ""
+    evolved_content: str = ""
+    content_diff: str = ""
+    blind_verified: bool = False
+    rejection_reason: str = ""
+    created_at: str = ""
+    completed_at: str = ""
+
+
+@dataclass(frozen=True)
+class SkillLineage:
+    """Lineage metadata for an evolved skill."""
+    skill_id: str
+    parent_skill_id: Optional[str]
+    evolution_type: EvolutionType
+    generation: int
+    trigger: TriggerType
+    mutation_summary: str = ""
+    created_at: str = ""
+
+    @property
+    def is_root(self) -> bool:
+        return self.parent_skill_id is None

package/src/superlocalmemory/hooks/hook_handlers.py

@@ -301,6 +301,19 @@ def _hook_stop() -> None:
         except Exception:
             pass
 
+    # --- Post-session skill evolution trigger (best-effort) ---
+    try:
+        session_id = os.environ.get("CLAUDE_SESSION_ID", "")
+        if session_id:
+            subprocess.Popen(
+                ["slm", "evolve", "--session", session_id],
+                stdout=subprocess.DEVNULL,
+                stderr=subprocess.DEVNULL,
+                start_new_session=True,
+            )
+    except Exception:
+        pass
+
     # --- Auto-consolidation (if >24h since last run) ---
     _maybe_consolidate()
 

package/src/superlocalmemory/learning/skill_performance_miner.py

@@ -171,6 +171,8 @@ class SkillPerformanceMiner:
         - Signal 2 (NEGATIVE): Same Skill re-invoked within 5 min
         - Signal 3 (NEGATIVE): Bash errors in next 3 events
         - Signal 4 (WEAK POSITIVE): Session continues 10+ events
+
+        H-N1QUERY: Batch-loads all trace events in one query instead of N+1.
         """
         metrics: dict[str, dict] = defaultdict(lambda: {
             "total_invocations": 0,
@@ -180,6 +182,29 @@ class SkillPerformanceMiner:
             "projects": set(),
         })
 
+        if not invocations:
+            return {}
+
+        # H-N1QUERY: Batch-load all potential trace events in one query.
+        # Find the min event_id across all invocations so we can fetch
+        # all subsequent events in a single SELECT.
+        min_event_id = min(inv["event_id"] for inv in invocations)
+        all_trace_rows = conn.execute(
+            "SELECT id, tool_name, event_type, output_summary, created_at "
+            "FROM tool_events "
+            "WHERE profile_id = ? AND id > ? "
+            "ORDER BY id ASC",
+            (profile_id, min_event_id),
+        ).fetchall()
+        all_trace = [dict(r) for r in all_trace_rows]
+
+        # Build an index: for each event_id, find its position in all_trace
+        # so we can slice TRACE_WINDOW events after it in O(1).
+        trace_id_to_idx: dict[int, int] = {}
+        for idx, t in enumerate(all_trace):
+            if t["id"] not in trace_id_to_idx:
+                trace_id_to_idx[t["id"]] = idx
+
         for inv in invocations:
             skill = inv["skill_name"]
             m = metrics[skill]
@@ -188,16 +213,24 @@ class SkillPerformanceMiner:
             if inv["project_path"]:
                 m["projects"].add(inv["project_path"])
 
-            # Get surrounding tool events for execution trace
-            trace = conn.execute(
-                "SELECT tool_name, event_type, output_summary, created_at "
-                "FROM tool_events "
-                "WHERE profile_id = ? AND id > ? "
-                "ORDER BY id ASC LIMIT ?",
-                (profile_id, inv["event_id"], TRACE_WINDOW),
-            ).fetchall()
+            # Find trace window for this invocation from pre-loaded data
+            # Events with id > inv["event_id"], take first TRACE_WINDOW
+            start_idx = 0
+            eid = inv["event_id"]
+            # The first entry in all_trace with id > eid
+            # Since all_trace starts at min_event_id+1 and is sorted, we
+            # can bisect or scan. Use the index if the next id is present.
+            # Simple approach: events after eid start at the position of eid+1
+            # or the first id > eid in the sorted list.
+            for candidate_id in range(eid + 1, eid + TRACE_WINDOW + 2):
+                if candidate_id in trace_id_to_idx:
+                    start_idx = trace_id_to_idx[candidate_id]
+                    break
+            else:
+                # No trace events found after this invocation
+                start_idx = len(all_trace)
 
-            trace_list = [dict(r) for r in trace]
+            trace_list = all_trace[start_idx:start_idx + TRACE_WINDOW]
             outcome = self._evaluate_trace(skill, inv, trace_list, invocations)
 
             if outcome > 0:
@@ -318,7 +351,7 @@ class SkillPerformanceMiner:
 
         assertion_id = hashlib.sha256(
             f"{profile_id}:skill_perf:{skill_name}".encode(),
-        ).hexdigest()[:16]
+        ).hexdigest()[:32]
 
         existing = conn.execute(
             "SELECT id, confidence FROM behavioral_assertions WHERE id = ?",
@@ -363,7 +396,7 @@ class SkillPerformanceMiner:
 
         assertion_id = hashlib.sha256(
             f"{profile_id}:skill_corr:{pair[0]}:{pair[1]}".encode(),
-        ).hexdigest()[:16]
+        ).hexdigest()[:32]
 
         existing = conn.execute(
             "SELECT id FROM behavioral_assertions WHERE id = ?",

package/src/superlocalmemory/mcp/server.py

@@ -79,6 +79,8 @@ _ESSENTIAL_TOOLS: set[str] = {
     # v3.4.7: Two-way learning (4)
     "log_tool_event", "get_assertions",
     "reinforce_assertion", "contradict_assertion",
+    # v3.4.11: Skill evolution (3)
+    "evolve_skill", "skill_health", "skill_lineage",
 }
 
 # v3.4.4: Mesh tools — enabled if mesh_enabled in config or SLM_MCP_MESH_TOOLS=1
@@ -138,6 +140,7 @@ from superlocalmemory.mcp.resources import register_resources
 from superlocalmemory.mcp.tools_code_graph import register_code_graph_tools
 from superlocalmemory.mcp.tools_mesh import register_mesh_tools
 from superlocalmemory.mcp.tools_learning import register_learning_tools
+from superlocalmemory.mcp.tools_evolution import register_evolution_tools
 
 register_core_tools(_target, get_engine)
 register_v28_tools(_target, get_engine)
@@ -148,6 +151,7 @@ register_resources(server, get_engine)  # Resources always registered (not tools
 register_code_graph_tools(_target, get_engine)  # CodeGraph: filtered like other tools (SLM_MCP_ALL_TOOLS=1 to show all)
 register_mesh_tools(_target, get_engine)  # v3.4.4: Mesh P2P tools — ships with SLM, no separate slm-mesh needed
 register_learning_tools(_target, get_engine)  # v3.4.7: Two-way learning tools
+register_evolution_tools(_target, get_engine)  # v3.4.11: Skill evolution tools
 
 
 # V3.3.21: Eager engine warmup — start initializing BEFORE first tool call.