153 lines
5.0 KiB
Python
153 lines
5.0 KiB
Python
"""主 agent loop: ReAct 风格,LLM ↔ Tool 反复直到无 tool_call。
|
|
|
|
loop 不直接 print —— 进度通过 sink.emit(event) 上抛。Sink 决定怎么呈现
|
|
(本地 console / SSE / 日志)。事件类型见 core/sinks.py 头部说明。
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import time
|
|
from typing import Any, Dict, Optional, Tuple
|
|
|
|
from .capabilities import ModelCapabilities
|
|
from .llm import LLM
|
|
from .session import Session
|
|
|
|
|
|
def _extract_usage(usage: Any) -> Tuple[int, int]:
|
|
"""从 litellm response.usage 提 (prompt_tokens, completion_tokens)。"""
|
|
if not usage:
|
|
return 0, 0
|
|
if hasattr(usage, "model_dump"):
|
|
usage = usage.model_dump()
|
|
elif hasattr(usage, "dict"):
|
|
usage = usage.dict()
|
|
if isinstance(usage, dict):
|
|
return int(usage.get("prompt_tokens") or 0), int(usage.get("completion_tokens") or 0)
|
|
return 0, 0
|
|
|
|
|
|
class AgentLoop:
|
|
def __init__(
|
|
self,
|
|
llm: LLM,
|
|
tools: Dict[str, Any],
|
|
session: Session,
|
|
capabilities: ModelCapabilities,
|
|
sink: Optional[Any] = None,
|
|
max_iterations: Optional[int] = None,
|
|
) -> None:
|
|
self.llm = llm
|
|
self.tools = tools
|
|
self.session = session
|
|
self.caps = capabilities
|
|
self.max_iterations = max_iterations or capabilities.max_iterations
|
|
self.sink = sink
|
|
|
|
def _emit(self, event: dict) -> None:
|
|
if self.sink is not None:
|
|
self.sink.emit(event)
|
|
|
|
def run(self, user_message: str) -> str:
|
|
self.session.append({"role": "user", "content": user_message})
|
|
|
|
for _ in range(self.max_iterations):
|
|
self._emit({"type": "llm_start"})
|
|
start = time.monotonic()
|
|
response = self.llm.chat(
|
|
messages=self.session.messages,
|
|
tools=[t.schema for t in self.tools.values()],
|
|
reasoning_effort=self.caps.default_reasoning_effort or None,
|
|
)
|
|
elapsed = time.monotonic() - start
|
|
msg = response.choices[0].message
|
|
self.session.append(msg)
|
|
|
|
pt, ct = _extract_usage(getattr(response, "usage", None))
|
|
self._emit({
|
|
"type": "llm_end",
|
|
"prompt_tokens": pt,
|
|
"completion_tokens": ct,
|
|
"elapsed": elapsed,
|
|
})
|
|
|
|
tool_calls = getattr(msg, "tool_calls", None) or []
|
|
content = getattr(msg, "content", None)
|
|
if content:
|
|
self._emit({"type": "text", "content": content})
|
|
|
|
if not tool_calls:
|
|
self._emit({"type": "done"})
|
|
return content or ""
|
|
|
|
for tc in tool_calls:
|
|
result = self._execute_tool_call(tc)
|
|
self.session.append(
|
|
{
|
|
"role": "tool",
|
|
"tool_call_id": tc.id,
|
|
"content": result,
|
|
}
|
|
)
|
|
|
|
self._emit({"type": "done"})
|
|
return "[reached max iterations]"
|
|
|
|
def _execute_tool_call(self, tc: Any) -> str:
|
|
name = tc.function.name
|
|
raw_args = tc.function.arguments or "{}"
|
|
try:
|
|
args = json.loads(raw_args)
|
|
except json.JSONDecodeError as e:
|
|
return f"[Error] invalid JSON arguments for {name}: {e}"
|
|
|
|
args_preview = json.dumps(args, ensure_ascii=False)
|
|
if len(args_preview) > 200:
|
|
args_preview = args_preview[:200] + "..."
|
|
self._emit({
|
|
"type": "tool_call",
|
|
"name": name,
|
|
"args": args,
|
|
"args_preview": args_preview,
|
|
})
|
|
|
|
tool = self.tools.get(name)
|
|
if tool is None:
|
|
err = f"[Error] unknown tool: {name}"
|
|
self._emit({"type": "tool_result", "name": name, "result": err,
|
|
"preview": err, "truncated": False})
|
|
return err
|
|
|
|
try:
|
|
result = tool.execute(**args)
|
|
except TypeError as e:
|
|
err = f"[Error] bad arguments to {name}: {e}"
|
|
self._emit({"type": "tool_result", "name": name, "result": err,
|
|
"preview": err, "truncated": False})
|
|
return err
|
|
except Exception as e:
|
|
err = f"[Error executing {name}] {type(e).__name__}: {e}"
|
|
self._emit({"type": "tool_result", "name": name, "result": err,
|
|
"preview": err, "truncated": False})
|
|
return err
|
|
|
|
if not isinstance(result, str):
|
|
result = str(result)
|
|
|
|
# 控制返回给模型的 tool 结果体量,避免炸 context
|
|
MAX_LEN = 16_000
|
|
truncated = False
|
|
if len(result) > MAX_LEN:
|
|
result = result[:MAX_LEN] + f"\n[... truncated, {len(result) - MAX_LEN} chars ...]"
|
|
truncated = True
|
|
|
|
preview = result if len(result) < 400 else result[:400] + "..."
|
|
self._emit({
|
|
"type": "tool_result",
|
|
"name": name,
|
|
"result": result,
|
|
"preview": preview,
|
|
"truncated": truncated,
|
|
})
|
|
return result
|