跳到主要内容

代码执行(编程式工具调用)

execute_code 工具允许代理编写能够以编程方式调用 Hermes 工具的 Python 脚本,从而将多步骤工作流压缩为一次 LLM 轮次。脚本在代理主机上的沙盒子进程中运行,通过 Unix 域套接字 RPC 进行通信。

工作原理

  1. 代理使用 from hermes_tools import ... 编写 Python 脚本
  2. Hermes 生成一个包含 RPC 函数的 hermes_tools.py 存根模块
  3. Hermes 打开一个 Unix 域套接字并启动一个 RPC 监听线程
  4. 脚本在子进程中运行 —— 工具调用通过套接字传回 Hermes
  5. 只有脚本的 print() 输出会返回给 LLM;中间工具结果永远不会进入上下文窗口
# 代理可以编写如下脚本:
from hermes_tools import web_search, web_extract

results = web_search("Python 3.13 新特性", limit=5)
for r in results["data"]["web"]:
content = web_extract([r["url"]])
# ... 过滤和处理 ...
print(summary)

沙盒中可用的工具: web_search, web_extract, read_file, write_file, search_files, patch, terminal(仅前台模式)。

代理何时使用此功能

当出现以下情况时,代理会使用 execute_code

  • 3 次以上工具调用,且调用之间存在处理逻辑
  • 批量数据过滤或条件分支
  • 对结果进行循环处理

关键优势:中间工具结果永远不会进入上下文窗口 —— 只有最终的 print() 输出会返回,从而显著减少令牌使用量。

实际示例

数据处理流水线

from hermes_tools import search_files, read_file
import json

# 查找所有配置文件并提取数据库设置
matches = search_files("database", path=".", file_glob="*.yaml", limit=20)
configs = []
for match in matches.get("matches", []):
content = read_file(match["path"])
configs.append({"file": match["path"], "preview": content["content"][:200]})

print(json.dumps(configs, indent=2))

多步骤网络研究

from hermes_tools import web_search, web_extract
import json

# 在一次轮次中完成搜索、提取和总结
results = web_search("2025年 Rust 异步运行时比较", limit=5)
summaries = []
for r in results["data"]["web"]:
page = web_extract([r["url"]])
for p in page.get("results", []):
if p.get("content"):
summaries.append({
"title": r["title"],
"url": r["url"],
"excerpt": p["content"][:500]
})

print(json.dumps(summaries, indent=2))

批量文件重构

from hermes_tools import search_files, read_file, patch

# 查找所有使用已弃用 API 的 Python 文件并修复它们
matches = search_files("old_api_call", path="src/", file_glob="*.py")
fixed = 0
for match in matches.get("matches", []):
result = patch(
path=match["path"],
old_string="old_api_call(",
new_string="new_api_call(",
replace_all=True
)
if "error" not in str(result):
fixed += 1

print(f"修复了 {fixed} 个文件,共匹配到 {len(matches.get('matches', []))} 个")

构建和测试流水线

from hermes_tools import terminal, read_file
import json

# 运行测试,解析结果并报告
result = terminal("cd /project && python -m pytest --tb=short -q 2>&1", timeout=120)
output = result.get("output", "")

# 解析测试输出
passed = output.count(" passed")
failed = output.count(" failed")
errors = output.count(" error")

report = {
"passed": passed,
"failed": failed,
"errors": errors,
"exit_code": result.get("exit_code", -1),
"summary": output[-500:] if len(output) > 500 else output
}

print(json.dumps(report, indent=2))

资源限制

资源限制说明
超时5 分钟 (300秒)脚本先收到 SIGTERM,5秒宽限期后收到 SIGKILL
标准输出50 KB输出会被截断,并添加 [output truncated at 50KB] 提示
标准错误10 KB在非零退出时包含在输出中,用于调试
工具调用每次执行 50 次达到限制时返回错误

所有限制都可以通过 config.yaml 配置:

# 在 ~/.hermes/config.yaml 中
code_execution:
timeout: 300 # 每个脚本的最大秒数(默认:300)
max_tool_calls: 50 # 每次执行的最大工具调用次数(默认:50)

脚本内部工具调用如何工作

当你的脚本调用像 web_search("query") 这样的函数时:

  1. 调用被序列化为 JSON 并通过 Unix 域套接字发送到父进程
  2. 父进程通过标准的 handle_function_call 处理程序进行分发
  3. 结果通过套接字发送回来
  4. 函数返回解析后的结果

这意味着脚本内部的工具调用行为与普通工具调用完全相同 —— 相同的速率限制、相同的错误处理、相同的能力。唯一的限制是 terminal() 仅限前台模式(没有 backgroundptycheck_interval 参数)。

错误处理

当脚本失败时,代理会收到结构化的错误信息:

  • 非零退出码:标准错误包含在输出中,因此代理可以看到完整的回溯信息
  • 超时:脚本被终止,代理看到 "Script timed out after 300s and was killed."
  • 中断:如果用户在执行期间发送了新消息,脚本会被终止,代理看到 [execution interrupted — user sent a new message]
  • 工具调用限制:当达到 50 次调用限制时,后续的工具调用会返回错误消息

响应始终包含 status(success/error/timeout/interrupted)、outputtool_calls_madeduration_seconds

安全性

安全模型

子进程在最小化环境中运行。默认情况下,API 密钥、令牌和凭据会被剥离。脚本只能通过 RPC 通道访问工具 —— 除非明确允许,否则它无法从环境变量中读取密钥。

名称中包含 KEYTOKENSECRETPASSWORDCREDENTIALPASSWDAUTH 的环境变量会被排除。只有安全的系统变量(PATHHOMELANGSHELLPYTHONPATHVIRTUAL_ENV 等)会被传递。

技能环境变量透传

当技能在其 frontmatter 中声明了 required_environment_variables 时,这些变量在技能加载后会自动透传execute_codeterminal 沙盒中。这使得技能可以使用其声明的 API 密钥,而不会削弱任意代码的安全性。

对于非技能用例,你可以在 config.yaml 中明确地将变量加入允许列表:

terminal:
env_passthrough:
- MY_CUSTOM_KEY
- ANOTHER_TOKEN

完整细节请参阅安全指南

脚本在临时目录中运行,执行后会被清理。子进程在其自己的进程组中运行,以便在超时或中断时能够被干净地终止。

execute_code 与 terminal 对比

用例execute_codeterminal
在工具调用之间进行多步骤工作流
简单的 shell 命令
过滤/处理大量工具输出
运行构建或测试套件
循环处理搜索结果
交互式/后台进程
需要在环境中使用 API 密钥⚠️ 仅通过透传✅(大多数会透传)

经验法则: 当你需要以编程方式调用 Hermes 工具,并在调用之间加入逻辑时,使用 execute_code。对于运行 shell 命令、构建和进程,使用 terminal

平台支持

代码执行需要 Unix 域套接字,仅在 Linux 和 macOS 上可用。在 Windows 上会自动禁用 —— 代理会回退到常规的顺序工具调用。