flowchart TB
    A[Cron 定时触发] --> B[Agent 启动]
    B --> C[加载 Skill 上下文]
    C --> D[读取任务 Map]
    D --> E{工具调用循环}
    E --> F[web_search 搜资料]
    E --> G[write_file 写文章]
    E --> H[publish.py 发布]
    F --> I[提取页面内容]
    I --> E
    G --> J[自检清单]
    J --> K{通过?}
| K -->|是| L[飞书通知] |
| K -->|否| M[patch 修复] |
    M --> G
    H --> L
    L --> N[退出]

| E -->|工具返回异常| O[重试或降级] |
    O --> E

# 每次调 API 前检查上下文用量
import tiktoken

enc = tiktoken.get_encoding("cl100k_base")

def check_context_budget(messages, max_tokens=64000):
    total = sum(len(enc.encode(m["content"])) for m in messages if "content" in m)
    usage_pct = total / max_tokens * 100
    if usage_pct > 80:
        # 触发压缩：用 LLM 摘要替代原始搜索内容
        return "summarize"
    elif usage_pct > 90:
        # 直接裁剪最旧的非关键消息
        return "trim"
    return "ok"

def web_extract(url):
    try:
        resp = requests.get(url, timeout=30)
        return resp.text
    except Exception:
        return ""  # <-- 这里就是万恶之源

def web_extract(url):
    try:
        resp = requests.get(url, timeout=30)
        resp.raise_for_status()
        return resp.text
    except requests.Timeout:
        return "[ERROR: 请求超时，目标服务器无响应]"
    except requests.HTTPError as e:
        return f"[ERROR: HTTP {e.response.status_code}]"
    except Exception as e:
        return f"[ERROR: {str(e)}]"

# cron 脚本开头
WORKDIR="/home/hnzwx/.hermes/scripts/articles"
mkdir -p "$WORKDIR"/archive

# 把昨天的文件归档
find "$WORKDIR" -name "*.md" -mtime +0 -exec mv {} "$WORKDIR"/archive/ \;

import tempfile, os
workdir = tempfile.mkdtemp(prefix="hermes_cron_")
os.environ["WORKDIR"] = workdir
# ... run agent ...
shutil.rmtree(workdir)

except RateLimitError:
    return "[请求过于频繁，请稍后重试]"

from tenacity import retry, wait_exponential, stop_after_attempt

@retry(
    wait=wait_exponential(multiplier=1, min=2, max=60),
    stop=stop_after_attempt(5),
    retry=lambda e: isinstance(e, RateLimitError)
)
def call_llm(messages):
    return client.chat.completions.create(
        model="deepseek-chat",
        messages=messages
    )

from datetime import datetime, timezone, timedelta

beijing_tz = timezone(timedelta(hours=8))
now = datetime.now(beijing_tz)

time_context = f"""
当前时间: {now.strftime('%Y年%m月%d日 %H:%M')}（北京时间，UTC+8）
今天是星期{['一','二','三','四','五','六','日'][now.weekday()]}
"""
system_prompt = time_context + original_prompt

import re

def validate_format(content):
    issues = []

    # 检查标题层级
    if content.startswith("# "):
        issues.append("禁止正文开头写一级标题")

    # 检查代码块语言标记
    code_blocks = re.findall(r'```(\w*)', content)
    for lang in code_blocks:
        if lang == "":
            issues.append("代码块缺少语言标记")
        elif lang == "py":
            issues.append("请使用 'python' 而非 'py'")

    # 表格首尾竖线检查
    table_lines = [l for l in content.split('\n') if l.startswith('|')]
    for line in table_lines:
        if not line.endswith('|'):
            issues.append(f"表格行缺少结尾竖线: {line[:30]}...")

    return issues

# 在写文件前校验
issues = validate_format(article_content)
if issues:
    # 把问题反馈给 Agent，让它自己修
    fix_prompt = f"请修复以下格式问题：\n" + "\n".join(f"- {i}" for i in issues)
    article_content = call_agent_fix(fix_prompt, article_content)

# 1. 工具调用次数上限
MAX_TOOL_CALLS = 30  # 任何任务不能超过这个值

# 2. 去重检查
visited_urls = set()
def web_extract(url):
    if url in visited_urls:
        return "[跳过：此 URL 已访问过]"
    visited_urls.add(url)
    # ... fetch

# 3. 费用预估
def estimate_cost(token_count, model="deepseek-chat"):
    rates = {"deepseek-chat": 2/1_000_000}  # ¥2 per 1M tokens
    return token_count * rates[model]