tzt
6247e053c1
- list_dir/read_file/write_file 三工具，OpenAI tools 声明与 tool_calls 解析 - 路径 join 后 resolve 必须仍位于工作区根内（防 ../ 与绝对路径逃逸） - ToolLoop：轮数上限 + token 熔断双护栏、事件回调逐条产出、触顶强制总结 - 测试 11 项：关押/往返/截断/循环编排/熔断/错误折叠（假 chat_fn，零外部依赖）
325 lines
14 KiB
Python
325 lines
14 KiB
Python
"""工具调用内核 —— 让 LLM 以 OpenAI function-calling 协议操作工作区文件。
|
||
|
||
组成（对齐《实现方案_v4_模型池与工具智能体.md》D4）：
|
||
- TOOLS_SPEC：list_dir / read_file / write_file 三个工具的 OpenAI tools 声明
|
||
- WorkspaceTools：被"关押"在根目录内的文件工具（路径越界一律拒绝，Windows pathlib）
|
||
- parse_tool_calls：解析 OpenAI 响应里的 tool_calls（arguments 容错为 {}）
|
||
- ToolLoop：通用智能体循环。chat_fn 注入（网关传 OpenAI 兼容客户端，测试传假实现），
|
||
本模块只负责循环编排：调用 -> 执行工具 -> 回喂结果 -> 直到模型给出最终答复。
|
||
|
||
工程约束：
|
||
- 纯标准库（router_system 零第三方依赖不变）
|
||
- 工具结果回喂前截断（防止上下文爆炸），轮数与 token 双上限（金额护栏）
|
||
- 事件回调 on_event 逐条产出过程事件（供 SSE 透出"智能体在做什么"）
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import json
|
||
from pathlib import Path
|
||
from typing import Any, Awaitable, Callable, Dict, List, Optional
|
||
|
||
# 回喂给模型的工具结果/读取内容上限（字符）
|
||
MAX_READ_CHARS = 8000
|
||
MAX_LIST_ENTRIES = 200
|
||
MAX_RESULT_CHARS = 8000
|
||
MAX_WRITE_CHARS = 200_000
|
||
|
||
# 默认循环上限
|
||
DEFAULT_MAX_ROUNDS = 8
|
||
|
||
# OpenAI tools 声明（chat/completions 请求的 tools 参数）
|
||
TOOLS_SPEC: List[Dict[str, Any]] = [
|
||
{
|
||
"type": "function",
|
||
"function": {
|
||
"name": "list_dir",
|
||
"description": "列出工作区内目录的内容（文件与子目录，含大小）。",
|
||
"parameters": {
|
||
"type": "object",
|
||
"properties": {
|
||
"path": {"type": "string", "description": "工作区内的相对路径，默认根目录"}
|
||
},
|
||
},
|
||
},
|
||
},
|
||
{
|
||
"type": "function",
|
||
"function": {
|
||
"name": "read_file",
|
||
"description": "读取工作区内一个文本文件的内容（过长自动截断）。",
|
||
"parameters": {
|
||
"type": "object",
|
||
"properties": {
|
||
"path": {"type": "string", "description": "工作区内的相对路径"}
|
||
},
|
||
"required": ["path"],
|
||
},
|
||
},
|
||
},
|
||
{
|
||
"type": "function",
|
||
"function": {
|
||
"name": "write_file",
|
||
"description": "把文本内容写入（或创建/覆盖）工作区内的一个文件。",
|
||
"parameters": {
|
||
"type": "object",
|
||
"properties": {
|
||
"path": {"type": "string", "description": "工作区内的相对路径"},
|
||
"content": {"type": "string", "description": "要写入的全文"},
|
||
},
|
||
"required": ["path", "content"],
|
||
},
|
||
},
|
||
},
|
||
]
|
||
|
||
TOOL_NAMES = {t["function"]["name"] for t in TOOLS_SPEC}
|
||
|
||
|
||
class ToolError(Exception):
|
||
"""工具执行失败（路径越界/不存在/参数非法）。"""
|
||
|
||
|
||
class WorkspaceTools:
|
||
"""被限制在根目录内的文件工具（智能体的"手"）。
|
||
|
||
安全：所有路径先 join 再 resolve，解析结果必须仍位于根目录内
|
||
（根目录自身允许），否则抛 ToolError——防 ../ 越界与绝对路径逃逸。
|
||
"""
|
||
|
||
def __init__(self, root: str | Path):
|
||
self.root = Path(root).resolve()
|
||
self.root.mkdir(parents=True, exist_ok=True)
|
||
|
||
# ---------- 路径关押 ----------
|
||
def resolve(self, rel_path: str) -> Path:
|
||
rel = (rel_path or "").strip().replace("\\", "/").lstrip("/")
|
||
p = (self.root / rel).resolve()
|
||
if p != self.root and self.root not in p.parents:
|
||
raise ToolError(f"路径越界（不允许访问工作区之外）: {rel_path}")
|
||
return p
|
||
|
||
# ---------- 三个工具 ----------
|
||
def list_dir(self, rel_path: str = "") -> Dict[str, Any]:
|
||
d = self.resolve(rel_path)
|
||
if not d.exists():
|
||
return {"ok": False, "error": f"目录不存在: {rel_path}"}
|
||
if not d.is_dir():
|
||
return {"ok": False, "error": f"不是目录: {rel_path}"}
|
||
entries = []
|
||
for child in sorted(d.iterdir(), key=lambda c: (c.is_file(), c.name.lower())):
|
||
if child.is_dir():
|
||
entries.append({"name": child.name + "/", "type": "dir"})
|
||
else:
|
||
entries.append({
|
||
"name": child.name, "type": "file",
|
||
"size": child.stat().st_size,
|
||
})
|
||
if len(entries) >= MAX_LIST_ENTRIES:
|
||
entries.append({"name": f"…（超过 {MAX_LIST_ENTRIES} 项已截断）", "type": "notice"})
|
||
break
|
||
return {"ok": True, "path": rel_path or ".", "entries": entries}
|
||
|
||
def read_file(self, rel_path: str) -> Dict[str, Any]:
|
||
p = self.resolve(rel_path)
|
||
if not p.exists():
|
||
return {"ok": False, "error": f"文件不存在: {rel_path}"}
|
||
if not p.is_file():
|
||
return {"ok": False, "error": f"不是文件: {rel_path}"}
|
||
try:
|
||
text = p.read_text(encoding="utf-8")
|
||
except UnicodeDecodeError:
|
||
return {"ok": False, "error": f"非文本文件（UTF-8 解码失败）: {rel_path}"}
|
||
truncated = len(text) > MAX_READ_CHARS
|
||
return {
|
||
"ok": True,
|
||
"path": rel_path,
|
||
"content": text[:MAX_READ_CHARS],
|
||
"truncated": truncated,
|
||
"total_chars": len(text),
|
||
}
|
||
|
||
def write_file(self, rel_path: str, content: str) -> Dict[str, Any]:
|
||
if len(content) > MAX_WRITE_CHARS:
|
||
return {"ok": False, "error": f"内容过长（>{MAX_WRITE_CHARS} 字符），拒绝写入"}
|
||
p = self.resolve(rel_path)
|
||
p.parent.mkdir(parents=True, exist_ok=True)
|
||
p.write_text(content, encoding="utf-8")
|
||
return {"ok": True, "path": rel_path, "bytes_written": len(content.encode("utf-8"))}
|
||
|
||
# ---------- 统一执行入口 ----------
|
||
def execute(self, name: str, arguments: Dict[str, Any]) -> Dict[str, Any]:
|
||
"""按名字执行工具；任何异常折叠为 {"ok": False, "error": ...}。"""
|
||
try:
|
||
if name == "list_dir":
|
||
return self.list_dir(str(arguments.get("path", "")))
|
||
if name == "read_file":
|
||
return self.read_file(str(arguments.get("path", "")))
|
||
if name == "write_file":
|
||
return self.write_file(
|
||
str(arguments.get("path", "")), str(arguments.get("content", "")))
|
||
return {"ok": False, "error": f"未知工具: {name}"}
|
||
except ToolError as e:
|
||
return {"ok": False, "error": str(e)}
|
||
except OSError as e:
|
||
return {"ok": False, "error": f"文件系统错误: {type(e).__name__}: {e}"}
|
||
|
||
|
||
def parse_tool_calls(message: Dict[str, Any]) -> List[Dict[str, Any]]:
|
||
"""从 OpenAI 响应的 message 解析 tool_calls。
|
||
|
||
返回 [{"id","name","arguments"(dict)}]；arguments 非法 JSON 时容错为 {}。
|
||
"""
|
||
out: List[Dict[str, Any]] = []
|
||
for tc in message.get("tool_calls") or []:
|
||
fn = tc.get("function") or {}
|
||
raw = fn.get("arguments")
|
||
if isinstance(raw, dict):
|
||
args = raw
|
||
elif isinstance(raw, str) and raw.strip():
|
||
try:
|
||
parsed = json.loads(raw)
|
||
args = parsed if isinstance(parsed, dict) else {}
|
||
except json.JSONDecodeError:
|
||
args = {}
|
||
else:
|
||
args = {}
|
||
out.append({
|
||
"id": tc.get("id") or f"call_{len(out)}",
|
||
"name": fn.get("name") or "",
|
||
"arguments": args,
|
||
})
|
||
return out
|
||
|
||
|
||
class ToolLoop:
|
||
"""通用智能体工具循环（zcode 式：思考 -> 调工具 -> 看结果 -> 再思考）。
|
||
|
||
chat_fn(messages, tools_spec) -> {"content": str|None,
|
||
"tool_calls": [ {id,name,arguments}, ... ],
|
||
"usage": {"prompt_tokens", "completion_tokens"}}
|
||
由网关注入真实 OpenAI 兼容客户端；测试注入脚本化假实现。
|
||
|
||
on_event(ev) 为可选同步回调，逐条收到过程事件：
|
||
{"type":"round","round":n}
|
||
{"type":"tool_call","round":n,"name":...,"arguments":...}
|
||
{"type":"tool_result","round":n,"name":...,"ok":...,"preview":...}
|
||
{"type":"usage","prompt_tokens":...,"completion_tokens":...}
|
||
{"type":"final","round":n,"reason":"answer"|"max_rounds"|"token_cap"|"error"}
|
||
"""
|
||
|
||
def __init__(
|
||
self,
|
||
tools: WorkspaceTools,
|
||
chat_fn: Callable[[List[Dict[str, Any]], List[Dict[str, Any]]], Awaitable[Dict[str, Any]]],
|
||
max_rounds: int = DEFAULT_MAX_ROUNDS,
|
||
token_cap: int = 0,
|
||
on_event: Optional[Callable[[Dict[str, Any]], None]] = None,
|
||
result_preview_chars: int = MAX_RESULT_CHARS,
|
||
):
|
||
self.tools = tools
|
||
self.chat_fn = chat_fn
|
||
self.max_rounds = max(1, int(max_rounds))
|
||
self.token_cap = int(token_cap) # 0 = 不限
|
||
self.on_event = on_event
|
||
self.result_preview_chars = result_preview_chars
|
||
|
||
def _emit(self, ev: Dict[str, Any]) -> None:
|
||
if self.on_event is not None:
|
||
try:
|
||
self.on_event(ev)
|
||
except Exception:
|
||
pass # 事件回调不允许打断主循环
|
||
|
||
def _total_tokens(self, usage: Dict[str, int]) -> int:
|
||
return int(usage.get("prompt_tokens", 0)) + int(usage.get("completion_tokens", 0))
|
||
|
||
async def run(self, task: str, system: str = "") -> Dict[str, Any]:
|
||
"""执行任务直到模型给出最终答复或触顶。返回最终结果与账目。"""
|
||
messages: List[Dict[str, Any]] = []
|
||
if system:
|
||
messages.append({"role": "system", "content": system})
|
||
messages.append({"role": "user", "content": task})
|
||
|
||
total_in = 0
|
||
total_out = 0
|
||
last_content = ""
|
||
|
||
for round_no in range(1, self.max_rounds + 1):
|
||
self._emit({"type": "round", "round": round_no})
|
||
try:
|
||
resp = await self.chat_fn(messages, TOOLS_SPEC)
|
||
except Exception as e:
|
||
self._emit({"type": "final", "round": round_no, "reason": "error",
|
||
"error": f"{type(e).__name__}: {e}"})
|
||
return {"response": "", "rounds": round_no, "reason": "error",
|
||
"error": f"{type(e).__name__}: {e}",
|
||
"prompt_tokens": total_in, "completion_tokens": total_out}
|
||
|
||
usage = resp.get("usage") or {}
|
||
total_in += int(usage.get("prompt_tokens", 0))
|
||
total_out += int(usage.get("completion_tokens", 0))
|
||
self._emit({"type": "usage", "prompt_tokens": total_in,
|
||
"completion_tokens": total_out})
|
||
|
||
# 金额护栏（D6 同源）：token 触顶立即停
|
||
if self.token_cap and total_in + total_out > self.token_cap:
|
||
self._emit({"type": "final", "round": round_no, "reason": "token_cap"})
|
||
return {"response": last_content, "rounds": round_no, "reason": "token_cap",
|
||
"error": f"token 熔断（{total_in + total_out}/{self.token_cap}）",
|
||
"prompt_tokens": total_in, "completion_tokens": total_out}
|
||
|
||
calls = resp.get("tool_calls") or []
|
||
if not calls:
|
||
self._emit({"type": "final", "round": round_no, "reason": "answer"})
|
||
return {"response": resp.get("content") or "", "rounds": round_no,
|
||
"reason": "answer", "error": None,
|
||
"prompt_tokens": total_in, "completion_tokens": total_out}
|
||
|
||
# 有工具调用：回填 assistant 消息（带原始 tool_calls 结构）+ 逐个执行
|
||
messages.append({
|
||
"role": "assistant",
|
||
"content": resp.get("content") or None,
|
||
"tool_calls": [
|
||
{
|
||
"id": c["id"],
|
||
"type": "function",
|
||
"function": {"name": c["name"],
|
||
"arguments": json.dumps(c["arguments"], ensure_ascii=False)},
|
||
}
|
||
for c in calls
|
||
],
|
||
})
|
||
for c in calls:
|
||
self._emit({"type": "tool_call", "round": round_no,
|
||
"name": c["name"], "arguments": c["arguments"]})
|
||
result = self.tools.execute(c["name"], c["arguments"])
|
||
preview = json.dumps(result, ensure_ascii=False)
|
||
if len(preview) > self.result_preview_chars:
|
||
preview = preview[:self.result_preview_chars] + "…(截断)"
|
||
self._emit({"type": "tool_result", "round": round_no,
|
||
"name": c["name"], "ok": bool(result.get("ok")),
|
||
"preview": preview})
|
||
messages.append({
|
||
"role": "tool",
|
||
"tool_call_id": c["id"],
|
||
"content": preview,
|
||
})
|
||
last_content = resp.get("content") or last_content
|
||
|
||
# 轮次耗尽：不再给工具，让模型立即总结（无 tools 的最后一次调用）
|
||
self._emit({"type": "final", "round": self.max_rounds, "reason": "max_rounds"})
|
||
try:
|
||
messages.append({"role": "user",
|
||
"content": "工具轮次已达上限。请基于以上信息立即给出最终答复，不要再调用工具。"})
|
||
resp = await self.chat_fn(messages, [])
|
||
usage = resp.get("usage") or {}
|
||
total_in += int(usage.get("prompt_tokens", 0))
|
||
total_out += int(usage.get("completion_tokens", 0))
|
||
final_text = resp.get("content") or last_content
|
||
except Exception:
|
||
final_text = last_content
|
||
return {"response": final_text, "rounds": self.max_rounds, "reason": "max_rounds",
|
||
"error": "工具轮次达上限，已强制总结",
|
||
"prompt_tokens": total_in, "completion_tokens": total_out}
|