MoFin/scripts/session_to_cron_bridge.py

#!/usr/bin/env python3
"""session_to_cron_bridge.py — 将Hermes session DB中的cron报告写到cron/output目录

Hermes cron jobs（如快速盯盘）将LLM输出存在 session DB (state.db) 中。
cron_to_xmpp.py 扫描 ~/.hermes/cron/output/ 目录的 .md 文件推送到XMPP。
这个脚本弥补这个缺口：从state.db读取最新的cron输出，生成.md文件。

工作方式：
1. 查询 state.db 中最近的 cron 会话（source='cron'）
2. 提取 assistant 的最后一条非空消息
3. 与 relay journal 对比去重
4. 新消息写入 cron/output/<job_id>/ 目录
5. cron_to_xmpp.py 自然捡起并推送
"""

import json
import sqlite3
import subprocess
import re
import sys
from datetime import datetime
from pathlib import Path

REAL_HOME = Path("/home/hmo")
PROFILE = "position-analyst"

# 要中继的 cron job ID 列表（需要推送到 XMPP 的）
RELAY_JOBS = {
    "62a2ba59f7ff": "快速盯盘-15分钟",
    "e27e2e92ed80": "知识萃取-盘后",
    "9d1236d8a07f": "策略评估-每日",
    "5dde4e1a42ce": "分析师-持仓复查",
}

# 输出目录（与 cron_to_xmpp.py 一致）
# 注意：~/.hermes 是 symlink 到 /home/hmo/.hermes/profiles/position-analyst/home/.hermes
# cron_to_xmpp.py 使用绝对路径 REAL_HOME / ".hermes" / "cron" / "output"
# 所以这里必须用绝对路径，不要相信 ~/.hermes 的解析
OUTPUT_DIRS = [
    REAL_HOME / ".hermes" / "cron" / "output",
    REAL_HOME / ".hermes" / "profiles" / PROFILE / "cron" / "output",
]

JOURNAL = REAL_HOME / ".hermes" / "cron" / ".relay_journal.json"
STATE_DB = REAL_HOME / ".hermes" / "profiles" / PROFILE / "state.db"

MAX_AGE_MINUTES = 70  # 只处理最近70分钟内的报告
TRACK_FILE = REAL_HOME / ".hermes" / "cron" / ".bridge_track.json"  # 追踪已桥接的session


def load_track():
    try:
        return set(json.loads(TRACK_FILE.read_text()))
    except:
        return set()


def save_track(entries):
    TRACK_FILE.write_text(json.dumps(sorted(entries)))


def load_journal():
    try:
        return set(json.loads(JOURNAL.read_text()))
    except:
        return set()


def save_journal(entries):
    JOURNAL.write_text(json.dumps(sorted(entries)))


def ensure_output_dirs():
    for d in OUTPUT_DIRS:
        d.mkdir(parents=True, exist_ok=True)
        for job_id in RELAY_JOBS:
            (d / job_id).mkdir(exist_ok=True)


def extract_report_content(content):
    """从assistant消息中提取报告正文"""
    if not content or content.strip() in ("", " ", "\n", "\n\n"):
        return None
    
    text = content.strip()
    
    # 跳过太短的消息
    if len(text) < 20:
        return None
    
    # 跳过 [SILENT]
    if "[SILENT]" in text:
        return None
    
    # 跳过思考过程（只留下实际报告内容）
    # 如果消息以"Now let me"/"Let me"/"I need"等开头，尝试找后面的报告正文
    lines = text.split('\n')
    report_lines = []
    in_report = False
    for line in lines:
        if not in_report:
            # 报告特征：以【开头 或 包含📊 或 包含【知微】
            if any(x in line for x in ["【", "📊", "【知微", "【⚡", "## "]):
                in_report = True
                report_lines.append(line)
        else:
            report_lines.append(line)
    
    if report_lines:
        text = '\n'.join(report_lines)
    
    if len(text) < 20:
        return None
    
    return text


def scan():
    processed = load_journal()
    tracked = load_track()
    new = set()
    n_written = 0
    
    if not STATE_DB.exists():
        print(f"state.db not found: {STATE_DB}", file=sys.stderr)
        return
    
    conn = sqlite3.connect(str(STATE_DB))
    conn.row_factory = sqlite3.Row
    cur = conn.cursor()
    
    now = datetime.now()
    
    for job_id, job_name in RELAY_JOBS.items():
        # Find recent sessions for this job
        cur.execute('''
            SELECT id, started_at, message_count, source
            FROM sessions 
            WHERE id LIKE ?
            ORDER BY started_at DESC 
            LIMIT 10
        ''', (f'cron_{job_id}_%',))
        
        sessions = cur.fetchall()
        
        for s in sessions:
            session_id = s['id']
            
            # Skip already bridged sessions
            if session_id in tracked:
                continue
            
            started_at = datetime.fromtimestamp(s['started_at']) if s['started_at'] else now
            
            # Skip too old sessions
            age_minutes = (now - started_at).total_seconds() / 60
            if age_minutes > MAX_AGE_MINUTES:
                continue
            
            # Find the last assistant message
            cur.execute('''
                SELECT content, timestamp
                FROM messages 
                WHERE session_id = ? AND role = 'assistant'
                AND content NOT IN ('', ' ', '\n', '\n\n', '\n\n\n')
                ORDER BY timestamp DESC
                LIMIT 1
            ''', (session_id,))
            
            msg = cur.fetchone()
            if not msg:
                tracked.add(session_id)
                continue
            
            content = msg['content'].strip()
            report = extract_report_content(content)
            if not report:
                tracked.add(session_id)
                continue
            
            # Mark as tracked even before writing
            tracked.add(session_id)
            
            # Generate a unique key for this report
            ts = datetime.fromtimestamp(msg['timestamp']).strftime('%Y%m%d_%H%M%S') if msg['timestamp'] else started_at.strftime('%Y%m%d_%H%M%S')
            filename = f"{job_name}_{ts}.md"
            
            for out_dir in OUTPUT_DIRS:
                out_path = out_dir / job_id / filename
                key = str(out_path.resolve())
                
                if key in processed or key in new:
                    continue
                
                # Write the report as an .md file (matching cron_to_xmpp.py format)
                md_content = f"# Cron Job: {job_name} ({session_id})\n\n## Response\n\n{report}\n"
                out_path.write_text(md_content, encoding='utf-8')
                new.add(key)
                n_written += 1
                print(f"  Written: {out_path.relative_to(REAL_HOME)}", file=sys.stderr)
    
    conn.close()
    
    if tracked:
        save_track(tracked)
    
    print(f"桥接完成：写入{n_written}份新报告", file=sys.stderr)
    # 桥接脚本只负责写入 .md 文件，不做去重追踪
    # 这样可以避免重复推送的复杂问题
    # 可能每次运行会写重复的文件，但cron_to_xmpp.py会用journal去重
    
    print(f"桥接完成：写入{n_written}份新报告", file=sys.stderr)


if __name__ == "__main__":
    scan()