Spaces:
Running
Running
File size: 10,469 Bytes
df568bc 5627f55 11a4e6e 5627f55 df568bc 5627f55 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 | """LLM-driven agent loop for the AWM web UI."""
from __future__ import annotations
import json
import re
from dataclasses import dataclass, field
from typing import Any, AsyncIterator
from openai import AsyncOpenAI
from .prompts import DEFAULT_SYSTEM_PROMPT
def parse_tool_call(content: str) -> dict | None:
m = re.search(r"<tool_call>\s*(.*?)\s*</tool_call>", content, re.DOTALL)
if not m:
return None
try:
data = json.loads(m.group(1).strip())
except json.JSONDecodeError:
return None
if isinstance(data, list):
data = data[0] if data else None
if not isinstance(data, dict) or "name" not in data:
return None
return data
def format_tools(tools: list[dict]) -> str:
lines = [f"Available MCP Tools ({len(tools)} tools):", "=" * 60]
for i, t in enumerate(tools, 1):
name = t.get("name") or t.get("tool_name", "")
desc = t.get("description", "")
schema = t.get("input_schema") or t.get("inputSchema") or {}
lines.append(f"{i}. {name}")
lines.append(f" Description: {desc}")
props = schema.get("properties", {})
required = set(schema.get("required", []))
if props:
lines.append(" Parameters:")
for pname, pinfo in props.items():
req = " (required)" if pname in required else ""
lines.append(
f" - {pname}: {pinfo.get('type', 'any')}{req} — "
f"{pinfo.get('description', '')}"
)
else:
lines.append(" Parameters: None")
lines.append("")
return "\n".join(lines)
def _truncate(s: str, limit: int = 2000) -> str:
if len(s) <= limit:
return s
return s[:limit] + f"\n... (truncated, full length {len(s)} chars)"
@dataclass
class AgentEvent:
kind: str # "info" | "llm_response" | "tool_call" | "tool_result" | "verify" | "done" | "error"
text: str = ""
payload: dict = field(default_factory=dict)
class AwmAgent:
def __init__(
self,
web_manager: Any,
llm_base_url: str,
llm_api_key: str,
llm_model: str,
system_prompt: str = DEFAULT_SYSTEM_PROMPT,
max_iterations: int = 10,
temperature: float = 1.0,
max_tokens: int = 2048,
):
self._web = web_manager
self._client = AsyncOpenAI(base_url=llm_base_url, api_key=llm_api_key)
self._model = llm_model
self._system_prompt = system_prompt
self._max_iterations = max_iterations
self._temperature = temperature
self._max_tokens = max_tokens
self._stop_requested = False
def request_stop(self) -> None:
self._stop_requested = True
async def _list_tools_dict(self) -> list[dict]:
result = await self._web.step_environment({"type": "list_tools"})
obs = result.get("observation", {}) or {}
tools = obs.get("tools", []) or []
out = []
for t in tools:
if isinstance(t, dict):
out.append(t)
else:
out.append(
{
"name": getattr(t, "name", ""),
"description": getattr(t, "description", ""),
"input_schema": getattr(t, "input_schema", {}),
}
)
return out
async def _call_tool(self, tool_name: str, args: dict) -> dict:
return await self._web.step_environment(
{"type": "call_tool", "tool_name": tool_name, "arguments": args}
)
async def run(
self,
task: str,
verifier_mode: str | None = None,
final_answer_fallback: str = "",
auto_verify: bool = True,
auto_done: bool = True,
) -> AsyncIterator[AgentEvent]:
"""Run the agent on the already-reset env, yielding events as it goes."""
try:
tools = await self._list_tools_dict()
except Exception as e:
yield AgentEvent(kind="error", text=f"list_tools failed: {e}")
return
yield AgentEvent(
kind="info",
text=f"Discovered {len(tools)} tools.",
payload={"tool_names": [t.get("name") for t in tools]},
)
tools_text = format_tools(tools)
messages: list[dict] = [
{"role": "system", "content": self._system_prompt},
{"role": "user", "content": task},
{
"role": "user",
"content": f"Available tools:\n{tools_text}",
},
]
last_assistant_content = ""
for step in range(1, self._max_iterations + 1):
if self._stop_requested:
yield AgentEvent(kind="info", text="Stopped by user.")
return
try:
resp = await self._client.chat.completions.create(
model=self._model,
messages=messages,
temperature=self._temperature,
max_completion_tokens=self._max_tokens,
)
except Exception as e:
yield AgentEvent(
kind="error", text=f"LLM call failed at step {step}: {e}"
)
return
content = resp.choices[0].message.content or ""
last_assistant_content = content
messages.append({"role": "assistant", "content": content})
yield AgentEvent(
kind="llm_response",
text=content,
payload={"step": step},
)
tc = parse_tool_call(content)
if tc is None:
yield AgentEvent(
kind="info",
text=f"No <tool_call> in step {step}; treating as final answer.",
)
break
name = tc.get("name", "")
arguments = tc.get("arguments") or {}
yield AgentEvent(
kind="tool_call",
text=f"{name} {json.dumps(arguments, ensure_ascii=False)[:300]}",
payload={"name": name, "arguments": arguments, "step": step},
)
tool_response = ""
try:
if name == "list_tools":
result = await self._web.step_environment({"type": "list_tools"})
obs = result.get("observation", {}) or {}
tools = await self._list_tools_dict()
tool_response = format_tools(tools)
elif name == "call_tool":
inner_name = arguments.get("tool_name", "")
inner_args = arguments.get("arguments", "{}")
if isinstance(inner_args, str):
try:
inner_args = json.loads(inner_args)
except json.JSONDecodeError:
inner_args = {}
if not isinstance(inner_args, dict):
inner_args = {}
result = await self._call_tool(inner_name, inner_args)
obs = result.get("observation", {}) or {}
if obs.get("tool_result") is not None:
tr = obs["tool_result"]
tool_response = (
json.dumps(tr, ensure_ascii=False)
if not isinstance(tr, str)
else tr
)
elif obs.get("error"):
tool_response = f"Error: {obs['error']}"
else:
tool_response = json.dumps(obs, ensure_ascii=False)
else:
tool_response = (
f"Error: Unknown function '{name}'. "
"Use 'list_tools' or 'call_tool'."
)
except Exception as e:
tool_response = f"Error during tool dispatch: {e}"
yield AgentEvent(
kind="tool_result",
text=_truncate(tool_response),
payload={"step": step},
)
messages.append(
{
"role": "user",
"content": f"Tool response:\n{_truncate(tool_response, 6000)}",
}
)
else:
yield AgentEvent(
kind="info",
text=f"Max iterations ({self._max_iterations}) reached.",
)
if auto_verify and verifier_mode:
verify_args: dict = {"verifier_mode": verifier_mode}
if last_assistant_content:
verify_args["final_answer"] = last_assistant_content
elif final_answer_fallback:
verify_args["final_answer"] = final_answer_fallback
try:
verify_result = await self._call_tool("verify", verify_args)
obs = verify_result.get("observation", {}) or {}
yield AgentEvent(
kind="verify",
text=(
f"reward_type={obs.get('reward_type')} "
f"reward={verify_result.get('reward')}"
),
payload={
"reward_type": obs.get("reward_type"),
"reward": verify_result.get("reward"),
"verify_result": obs.get("verify_result"),
},
)
except Exception as e:
yield AgentEvent(kind="error", text=f"verify failed: {e}")
if auto_done:
try:
done_result = await self._call_tool("done", {"keep_session": True})
obs = done_result.get("observation", {}) or {}
yield AgentEvent(
kind="done",
text=(
f"Episode done. trajectory_path="
f"{obs.get('trajectory_path') or '(none)'}"
),
payload={
"trajectory_path": obs.get("trajectory_path"),
"session_dir": obs.get("session_dir"),
},
)
except Exception as e:
yield AgentEvent(kind="error", text=f"done failed: {e}")
|