feat: add harvester.py — session knowledge extractor (#8)

scripts/harvester.py

@@ -0,0 +1,447 @@
+#!/usr/bin/env python3
+"""
+harvester.py — Extract durable knowledge from Hermes session transcripts.
+
+Combines session_reader + extraction prompt + LLM inference to pull
+facts, pitfalls, patterns, and tool quirks from finished sessions.
+
+Usage:
+    python3 harvester.py --session ~/.hermes/sessions/session_xxx.jsonl --output knowledge/
+    python3 harvester.py --batch --since 2026-04-01 --limit 100
+    python3 harvester.py --session session.jsonl --dry-run  # Preview without writing
+"""
+
+import argparse
+import json
+import os
+import sys
+import time
+import hashlib
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Optional
+
+# Add scripts dir to path for sibling imports
+SCRIPT_DIR = Path(__file__).parent.absolute()
+sys.path.insert(0, str(SCRIPT_DIR))
+
+from session_reader import read_session, extract_conversation, truncate_for_context, messages_to_text
+
+# --- Configuration ---
+
+DEFAULT_API_BASE = os.environ.get("HARVESTER_API_BASE", "https://api.nousresearch.com/v1")
+DEFAULT_API_KEY = os.environ.get("HARVESTER_API_KEY", "")
+DEFAULT_MODEL = os.environ.get("HARVESTER_MODEL", "xiaomi/mimo-v2-pro")
+KNOWLEDGE_DIR = os.environ.get("HARVESTER_KNOWLEDGE_DIR", "knowledge")
+PROMPT_PATH = os.environ.get("HARVESTER_PROMPT_PATH", str(SCRIPT_DIR.parent / "templates" / "harvest-prompt.md"))
+
+# Where to look for API keys if not set via env
+API_KEY_PATHS = [
+    os.path.expanduser("~/.config/nous/key"),
+    os.path.expanduser("~/.hermes/keymaxxing/active/minimax.key"),
+    os.path.expanduser("~/.config/openrouter/key"),
+]
+
+
+def find_api_key() -> str:
+    """Find API key from common locations."""
+    for path in API_KEY_PATHS:
+        if os.path.exists(path):
+            with open(path) as f:
+                key = f.read().strip()
+                if key:
+                    return key
+    return ""
+
+
+def load_extraction_prompt() -> str:
+    """Load the extraction prompt template."""
+    path = Path(PROMPT_PATH)
+    if not path.exists():
+        print(f"ERROR: Extraction prompt not found at {path}", file=sys.stderr)
+        print("Expected templates/harvest-prompt.md from issue #7", file=sys.stderr)
+        sys.exit(1)
+    return path.read_text(encoding='utf-8')
+
+
+def call_llm(prompt: str, transcript: str, api_base: str, api_key: str, model: str) -> Optional[list[dict]]:
+    """Call the LLM API to extract knowledge from a transcript."""
+    import urllib.request
+    
+    messages = [
+        {"role": "system", "content": prompt},
+        {"role": "user", "content": f"Extract knowledge from this session transcript:\n\n{transcript}"}
+    ]
+    
+    payload = json.dumps({
+        "model": model,
+        "messages": messages,
+        "temperature": 0.1,  # Low temp for consistent extraction
+        "max_tokens": 4096
+    }).encode('utf-8')
+    
+    req = urllib.request.Request(
+        f"{api_base}/chat/completions",
+        data=payload,
+        headers={
+            "Authorization": f"Bearer {api_key}",
+            "Content-Type": "application/json"
+        },
+        method="POST"
+    )
+    
+    try:
+        with urllib.request.urlopen(req, timeout=60) as resp:
+            result = json.loads(resp.read().decode('utf-8'))
+            content = result["choices"][0]["message"]["content"]
+            return parse_extraction_response(content)
+    except Exception as e:
+        print(f"ERROR: LLM API call failed: {e}", file=sys.stderr)
+        return None
+
+
+def parse_extraction_response(content: str) -> Optional[list[dict]]:
+    """Parse the LLM response to extract knowledge items.
+    
+    Handles various response formats: raw JSON, markdown-wrapped JSON, etc.
+    """
+    # Try direct JSON parse first
+    try:
+        data = json.loads(content)
+        if isinstance(data, dict) and 'knowledge' in data:
+            return data['knowledge']
+        if isinstance(data, list):
+            return data
+    except json.JSONDecodeError:
+        pass
+    
+    # Try extracting JSON from markdown code blocks
+    import re
+    json_match = re.search(r'```(?:json)?\s*({.*?})\s*```', content, re.DOTALL)
+    if json_match:
+        try:
+            data = json.loads(json_match.group(1))
+            if isinstance(data, dict) and 'knowledge' in data:
+                return data['knowledge']
+            if isinstance(data, list):
+                return data
+        except json.JSONDecodeError:
+            pass
+    
+    # Try finding any JSON object with knowledge array
+    json_match = re.search(r'({[^{}]*"knowledge"[^{}]*[[sS]*?][^{}]*})', content)
+    if json_match:
+        try:
+            data = json.loads(json_match.group(1))
+            return data.get('knowledge', [])
+        except json.JSONDecodeError:
+            pass
+    
+    print(f"WARNING: Could not parse LLM response as JSON", file=sys.stderr)
+    print(f"Response preview: {content[:500]}", file=sys.stderr)
+    return None
+
+
+def load_existing_knowledge(knowledge_dir: str) -> dict:
+    """Load the existing knowledge index."""
+    index_path = Path(knowledge_dir) / "index.json"
+    if not index_path.exists():
+        return {"version": 1, "last_updated": "", "total_facts": 0, "facts": []}
+    
+    try:
+        with open(index_path, 'r', encoding='utf-8') as f:
+            return json.load(f)
+    except (json.JSONDecodeError, IOError) as e:
+        print(f"WARNING: Could not load knowledge index: {e}", file=sys.stderr)
+        return {"version": 1, "last_updated": "", "total_facts": 0, "facts": []}
+
+
+def fact_fingerprint(fact: dict) -> str:
+    """Generate a deduplication fingerprint for a fact.
+    
+    Uses the fact text normalized (lowercase, stripped) as the key.
+    Similar facts will have similar fingerprints.
+    """
+    text = fact.get('fact', '').lower().strip()
+    # Normalize whitespace
+    text = ' '.join(text.split())
+    return hashlib.md5(text.encode('utf-8')).hexdigest()
+
+
+def deduplicate(new_facts: list[dict], existing: list[dict], similarity_threshold: float = 0.8) -> list[dict]:
+    """Remove duplicate facts from new_facts that already exist in the knowledge store.
+    
+    Uses fingerprint matching for exact dedup and simple overlap check for near-dupes.
+    """
+    existing_fingerprints = set()
+    existing_texts = []
+    for f in existing:
+        fp = fact_fingerprint(f)
+        existing_fingerprints.add(fp)
+        existing_texts.append(f.get('fact', '').lower().strip())
+    
+    unique = []
+    for fact in new_facts:
+        fp = fact_fingerprint(fact)
+        if fp in existing_fingerprints:
+            continue
+        
+        # Check for near-duplicates using simple word overlap
+        fact_words = set(fact.get('fact', '').lower().split())
+        is_dup = False
+        for existing_text in existing_texts:
+            existing_words = set(existing_text.split())
+            if not fact_words or not existing_words:
+                continue
+            overlap = len(fact_words & existing_words) / max(len(fact_words | existing_words), 1)
+            if overlap >= similarity_threshold:
+                is_dup = True
+                break
+        
+        if not is_dup:
+            unique.append(fact)
+            existing_fingerprints.add(fp)
+            existing_texts.append(fact.get('fact', '').lower().strip())
+    
+    return unique
+
+
+def validate_fact(fact: dict) -> bool:
+    """Validate a single knowledge item has required fields."""
+    required = ['fact', 'category', 'repo', 'confidence']
+    for field in required:
+        if field not in fact:
+            return False
+    
+    if not isinstance(fact['fact'], str) or not fact['fact'].strip():
+        return False
+    
+    valid_categories = ['fact', 'pitfall', 'pattern', 'tool-quirk', 'question']
+    if fact['category'] not in valid_categories:
+        return False
+    
+    if not isinstance(fact.get('confidence', 0), (int, float)):
+        return False
+    
+    if not (0.0 <= fact['confidence'] <= 1.0):
+        return False
+    
+    return True
+
+
+def write_knowledge(index: dict, new_facts: list[dict], knowledge_dir: str, source_session: str = ""):
+    """Write new facts to the knowledge store."""
+    kdir = Path(knowledge_dir)
+    kdir.mkdir(parents=True, exist_ok=True)
+    
+    # Add source tracking to each fact
+    for fact in new_facts:
+        fact['source_session'] = source_session
+        fact['harvested_at'] = datetime.now(timezone.utc).isoformat()
+    
+    # Update index
+    index['facts'].extend(new_facts)
+    index['total_facts'] = len(index['facts'])
+    index['last_updated'] = datetime.now(timezone.utc).isoformat()
+    
+    # Write index
+    index_path = kdir / "index.json"
+    with open(index_path, 'w', encoding='utf-8') as f:
+        json.dump(index, f, indent=2, ensure_ascii=False)
+    
+    # Also write per-repo markdown files for human reading
+    repos = {}
+    for fact in new_facts:
+        repo = fact.get('repo', 'global')
+        repos.setdefault(repo, []).append(fact)
+    
+    for repo, facts in repos.items():
+        if repo == 'global':
+            md_path = kdir / "global" / "harvested.md"
+        else:
+            md_path = kdir / "repos" / f"{repo}.md"
+        
+        md_path.parent.mkdir(parents=True, exist_ok=True)
+        
+        # Append to existing or create new
+        mode = 'a' if md_path.exists() else 'w'
+        with open(md_path, mode, encoding='utf-8') as f:
+            if mode == 'w':
+                f.write(f"# Knowledge: {repo}\n\n")
+            f.write(f"## Harvested {datetime.now(timezone.utc).strftime('%Y-%m-%d %H:%M')}\n\n")
+            for fact in facts:
+                icon = {'fact': '📋', 'pitfall': '⚠️', 'pattern': '🔄', 'tool-quirk': '🔧', 'question': '❓'}.get(fact['category'], '•')
+                f.write(f"- {icon} **{fact['category']}** (conf: {fact['confidence']:.1f}): {fact['fact']}\n")
+            f.write("\n")
+
+
+def harvest_session(session_path: str, knowledge_dir: str, api_base: str, api_key: str, 
+                    model: str, dry_run: bool = False, min_confidence: float = 0.3) -> dict:
+    """Harvest knowledge from a single session.
+    
+    Returns: dict with stats (facts_found, facts_new, facts_dup, elapsed_seconds, error)
+    """
+    start_time = time.time()
+    stats = {
+        'session': session_path,
+        'facts_found': 0,
+        'facts_new': 0,
+        'facts_dup': 0,
+        'elapsed_seconds': 0,
+        'error': None
+    }
+    
+    try:
+        # 1. Read session
+        messages = read_session(session_path)
+        if not messages:
+            stats['error'] = "Empty session file"
+            return stats
+        
+        # 2. Extract conversation
+        conv = extract_conversation(messages)
+        if not conv:
+            stats['error'] = "No conversation turns found"
+            return stats
+        
+        # 3. Truncate for context window
+        truncated = truncate_for_context(conv, head=50, tail=50)
+        transcript = messages_to_text(truncated)
+        
+        # 4. Load extraction prompt
+        prompt = load_extraction_prompt()
+        
+        # 5. Call LLM
+        raw_facts = call_llm(prompt, transcript, api_base, api_key, model)
+        if raw_facts is None:
+            stats['error'] = "LLM extraction failed"
+            return stats
+        
+        # 6. Validate
+        valid_facts = [f for f in raw_facts if validate_fact(f) and f.get('confidence', 0) >= min_confidence]
+        stats['facts_found'] = len(valid_facts)
+        
+        # 7. Deduplicate
+        existing_index = load_existing_knowledge(knowledge_dir)
+        existing_facts = existing_index.get('facts', [])
+        new_facts = deduplicate(valid_facts, existing_facts)
+        stats['facts_new'] = len(new_facts)
+        stats['facts_dup'] = len(valid_facts) - len(new_facts)
+        
+        # 8. Write (unless dry run)
+        if new_facts and not dry_run:
+            write_knowledge(existing_index, new_facts, knowledge_dir, source_session=session_path)
+        
+        stats['elapsed_seconds'] = round(time.time() - start_time, 2)
+        return stats
+        
+    except Exception as e:
+        stats['error'] = str(e)
+        stats['elapsed_seconds'] = round(time.time() - start_time, 2)
+        return stats
+
+
+def batch_harvest(sessions_dir: str, knowledge_dir: str, api_base: str, api_key: str,
+                  model: str, since: str = "", limit: int = 0, dry_run: bool = False) -> list[dict]:
+    """Harvest knowledge from multiple sessions in batch."""
+    sessions_path = Path(sessions_dir)
+    if not sessions_path.is_dir():
+        print(f"ERROR: Sessions directory not found: {sessions_dir}", file=sys.stderr)
+        return []
+    
+    # Find session files
+    session_files = sorted(sessions_path.glob("*.jsonl"), reverse=True)  # Newest first
+    
+    # Filter by date if --since provided
+    if since:
+        since_dt = datetime.fromisoformat(since.replace('Z', '+00:00'))
+        filtered = []
+        for sf in session_files:
+            # Try to parse timestamp from filename (common format: session_YYYYMMDD_HHMMSS_hash.jsonl)
+            try:
+                parts = sf.stem.split('_')
+                if len(parts) >= 3:
+                    date_str = parts[1]
+                    file_dt = datetime.strptime(date_str, '%Y%m%d').replace(tzinfo=timezone.utc)
+                    if file_dt >= since_dt:
+                        filtered.append(sf)
+            except (ValueError, IndexError):
+                # If we can't parse the date, include the file (be permissive)
+                filtered.append(sf)
+        session_files = filtered
+    
+    # Apply limit
+    if limit > 0:
+        session_files = session_files[:limit]
+    
+    print(f"Harvesting {len(session_files)} sessions...")
+    
+    results = []
+    for i, sf in enumerate(session_files, 1):
+        print(f"[{i}/{len(session_files)}] {sf.name}...", end=" ", flush=True)
+        stats = harvest_session(str(sf), knowledge_dir, api_base, api_key, model, dry_run)
+        if stats['error']:
+            print(f"ERROR: {stats['error']}")
+        else:
+            print(f"{stats['facts_new']} new, {stats['facts_dup']} dup ({stats['elapsed_seconds']}s)")
+        results.append(stats)
+    
+    return results
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Harvest knowledge from session transcripts")
+    parser.add_argument('--session', help='Path to a single session JSONL file')
+    parser.add_argument('--batch', action='store_true', help='Batch mode: process multiple sessions')
+    parser.add_argument('--sessions-dir', default=os.path.expanduser('~/.hermes/sessions'), 
+                        help='Directory containing session files (default: ~/.hermes/sessions)')
+    parser.add_argument('--output', default='knowledge', help='Output directory for knowledge store')
+    parser.add_argument('--since', default='', help='Only process sessions after this date (YYYY-MM-DD)')
+    parser.add_argument('--limit', type=int, default=0, help='Max sessions to process (0=unlimited)')
+    parser.add_argument('--api-base', default=DEFAULT_API_BASE, help='LLM API base URL')
+    parser.add_argument('--api-key', default='', help='LLM API key (or set HARVESTER_API_KEY)')
+    parser.add_argument('--model', default=DEFAULT_MODEL, help='Model to use for extraction')
+    parser.add_argument('--dry-run', action='store_true', help='Preview without writing to knowledge store')
+    parser.add_argument('--min-confidence', type=float, default=0.3, help='Minimum confidence threshold')
+    
+    args = parser.parse_args()
+    
+    # Resolve API key
+    api_key = args.api_key or DEFAULT_API_KEY or find_api_key()
+    if not api_key:
+        print("ERROR: No API key found. Set HARVESTER_API_KEY or store in one of:", file=sys.stderr)
+        for p in API_KEY_PATHS:
+            print(f"  {p}", file=sys.stderr)
+        sys.exit(1)
+    
+    # Resolve knowledge directory
+    knowledge_dir = args.output
+    if not os.path.isabs(knowledge_dir):
+        knowledge_dir = os.path.join(SCRIPT_DIR.parent, knowledge_dir)
+    
+    if args.session:
+        # Single session mode
+        stats = harvest_session(
+            args.session, knowledge_dir, args.api_base, api_key, args.model,
+            dry_run=args.dry_run, min_confidence=args.min_confidence
+        )
+        print(json.dumps(stats, indent=2))
+        if stats['error']:
+            sys.exit(1)
+    elif args.batch:
+        # Batch mode
+        results = batch_harvest(
+            args.sessions_dir, knowledge_dir, args.api_base, api_key, args.model,
+            since=args.since, limit=args.limit, dry_run=args.dry_run
+        )
+        total_new = sum(r['facts_new'] for r in results)
+        total_dup = sum(r['facts_dup'] for r in results)
+        errors = sum(1 for r in results if r['error'])
+        print(f"\nDone: {total_new} new facts, {total_dup} duplicates, {errors} errors")
+    else:
+        parser.print_help()
+        sys.exit(1)
+
+
+if __name__ == '__main__':
+    main()

scripts/session_reader.py

@@ -0,0 +1,142 @@
+#!/usr/bin/env python3
+"""
+session_reader.py — Parse Hermes session JSONL transcripts.
+
+Each line in a session file is a JSON object representing a message.
+Standard fields: role (user|assistant|system), content (str), timestamp (str).
+Tool calls and tool results are also captured.
+"""
+
+import json
+import sys
+from pathlib import Path
+from typing import Iterator, Optional
+
+
+def read_session(path: str) -> list[dict]:
+    """Read a session JSONL file and return all messages as a list."""
+    messages = []
+    with open(path, 'r', encoding='utf-8') as f:
+        for line_num, line in enumerate(f, 1):
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                msg = json.loads(line)
+                messages.append(msg)
+            except json.JSONDecodeError as e:
+                print(f"WARNING: Skipping malformed JSON at line {line_num}: {e}", file=sys.stderr)
+    return messages
+
+
+def read_session_iter(path: str) -> Iterator[dict]:
+    """Iterate over session messages without loading all into memory."""
+    with open(path, 'r', encoding='utf-8') as f:
+        for line_num, line in enumerate(f, 1):
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                yield json.loads(line)
+            except json.JSONDecodeError as e:
+                print(f"WARNING: Skipping malformed JSON at line {line_num}: {e}", file=sys.stderr)
+
+
+def extract_conversation(messages: list[dict]) -> list[dict]:
+    """Extract user/assistant conversation turns, skipping tool-only messages."""
+    conversation = []
+    for msg in messages:
+        role = msg.get('role', '')
+        content = msg.get('content', '')
+        
+        # Skip empty messages and pure tool calls
+        if role in ('user', 'assistant', 'system'):
+            if isinstance(content, str) and content.strip():
+                conversation.append({
+                    'role': role,
+                    'content': content.strip(),
+                    'timestamp': msg.get('timestamp', '')
+                })
+            elif isinstance(content, list):
+                # Multimodal content — extract text parts
+                text_parts = []
+                for part in content:
+                    if isinstance(part, dict) and part.get('type') == 'text':
+                        text_parts.append(part.get('text', ''))
+                if text_parts:
+                    conversation.append({
+                        'role': role,
+                        'content': '\n'.join(text_parts),
+                        'timestamp': msg.get('timestamp', '')
+                    })
+    return conversation
+
+
+def truncate_for_context(messages: list[dict], head: int = 50, tail: int = 50) -> list[dict]:
+    """Truncate long sessions: keep first N + last N messages.
+    
+    This preserves session start (initial context) and end (final results),
+    skipping the messy middle of long debugging sessions.
+    """
+    if len(messages) <= head + tail:
+        return messages
+    
+    truncated = messages[:head]
+    truncated.append({
+        'role': 'system',
+        'content': f'[{len(messages) - head - tail} messages truncated]',
+        'timestamp': ''
+    })
+    truncated.extend(messages[-tail:])
+    return truncated
+
+
+def messages_to_text(messages: list[dict]) -> str:
+    """Convert message list to plain text for LLM consumption."""
+    lines = []
+    for msg in messages:
+        role = msg.get('role', 'unknown').upper()
+        content = msg.get('content', '')
+        if msg.get('role') == 'system' and 'truncated' in content:
+            lines.append(f'--- {content} ---')
+        else:
+            lines.append(f'{role}: {content}')
+    return '\n\n'.join(lines)
+
+
+def get_session_metadata(path: str) -> dict:
+    """Extract metadata from a session file (first message often has config info)."""
+    messages = read_session(path)
+    if not messages:
+        return {'path': path, 'message_count': 0}
+    
+    first = messages[0]
+    last = messages[-1]
+    
+    return {
+        'path': path,
+        'message_count': len(messages),
+        'first_timestamp': first.get('timestamp', ''),
+        'last_timestamp': last.get('timestamp', ''),
+        'first_role': first.get('role', ''),
+        'has_tool_calls': any(m.get('tool_calls') for m in messages),
+    }
+
+
+if __name__ == '__main__':
+    if len(sys.argv) < 2:
+        print(f"Usage: {sys.argv[0]} <session.jsonl>")
+        sys.exit(1)
+    
+    path = sys.argv[1]
+    meta = get_session_metadata(path)
+    print(json.dumps(meta, indent=2))
+    
+    messages = read_session(path)
+    conv = extract_conversation(messages)
+    print(f"\nConversation: {len(conv)} turns")
+    
+    truncated = truncate_for_context(conv)
+    print(f"After truncation: {len(truncated)} turns")
+    print(f"\nPreview (first 500 chars):")
+    print(messages_to_text(truncated[:5])[:500])