Spaces:
Sleeping
Sleeping
File size: 10,101 Bytes
6bedc74 2411e3d 6bedc74 2411e3d 6bedc74 2411e3d 6bedc74 2411e3d 6bedc74 2411e3d 6bedc74 2411e3d 6bedc74 2411e3d 6bedc74 2411e3d 6bedc74 2411e3d 6bedc74 2411e3d 6bedc74 2411e3d 6bedc74 2411e3d 6bedc74 2411e3d 6bedc74 2411e3d 6bedc74 2411e3d 6bedc74 2411e3d 6bedc74 2411e3d | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 | import json
import os
from collections.abc import Iterator
from typing import Any, Dict, List
from dotenv import load_dotenv
from openai import OpenAI
from src.audit import write_audit_record
from src.executor import authorize_tool_call, execute_authorized_tool
from src.schemas import AgentRunEvent, AgentRunResult, ToolCall
from src.tool_registry import get_tool_definitions
load_dotenv(dotenv_path=".env")
HF_BASE_URL = "https://router.huggingface.co/v1"
DEFAULT_MODEL = "openai/gpt-oss-120b:cerebras"
SYSTEM_PROMPT = """
You are a controlled tool-using operations agent.
You have access only to the tools explicitly provided to you.
Rules:
1. Use tools when they are needed to answer the user's request accurately.
2. Do not claim that a tool was used unless you actually called it.
3. Never invent tool results.
4. Do not invent tools that are not available.
5. Use the calculator for arithmetic rather than calculating mentally.
6. Use search_inventory for questions about the inventory database.
7. Use lookup_country for structured country information.
8. You may call multiple tools when the user's request requires them.
9. After receiving tool results, use those results to construct your answer.
10. If a tool is blocked or returns an error, do not hide the failure.
11. Do not expose internal chain-of-thought. Provide concise conclusions only.
"""
def get_client() -> OpenAI:
"""Create the Hugging Face OpenAI-compatible client."""
token = os.getenv("HF_TOKEN")
if not token:
raise RuntimeError(
"HF_TOKEN is not configured. "
"Set it as an environment variable before running the agent."
)
return OpenAI(
base_url=HF_BASE_URL,
api_key=token,
)
def _parse_tool_arguments(raw_arguments: str) -> Dict[str, Any]:
"""Decode model-provided tool arguments into an object for validation."""
try:
arguments = json.loads(raw_arguments)
if not isinstance(arguments, dict):
raise ValueError(
"Tool arguments must decode to an object."
)
return arguments
except (json.JSONDecodeError, ValueError):
return {
"__invalid_arguments__": raw_arguments,
}
def run_agent_iter(
user_request: str,
model: str | None = None,
max_tool_rounds: int = 5,
) -> Iterator[AgentRunEvent]:
"""Run the agent while emitting real control-boundary events."""
if not isinstance(user_request, str):
raise TypeError("user_request must be a string.")
user_request = user_request.strip()
if not user_request:
raise ValueError("user_request cannot be empty.")
if max_tool_rounds < 1:
raise ValueError(
"max_tool_rounds must be at least 1."
)
client = get_client()
selected_model = (
model
or os.getenv("MODEL_ID")
or DEFAULT_MODEL
)
tools = get_tool_definitions()
messages: List[Dict[str, Any]] = [
{
"role": "system",
"content": SYSTEM_PROMPT,
},
{
"role": "user",
"content": user_request,
},
]
executed_calls: List[ToolCall] = []
executed_results = []
yield AgentRunEvent(
event="run_started",
message=(
"Request accepted. The model may propose only tools exposed "
"through the approved registry."
),
)
for round_index in range(1, max_tool_rounds + 1):
yield AgentRunEvent(
event="model_thinking",
round_index=round_index,
message="The model is deciding whether an approved tool is needed.",
)
response = client.chat.completions.create(
model=selected_model,
messages=messages,
tools=tools,
tool_choice="auto",
reasoning_effort="low",
)
message = response.choices[0].message
assistant_message: Dict[str, Any] = {
"role": "assistant",
"content": message.content,
}
if message.tool_calls:
assistant_message["tool_calls"] = [
tool_call.model_dump()
for tool_call in message.tool_calls
]
messages.append(assistant_message)
if not message.tool_calls:
final_answer = (
message.content
or "The agent returned no final answer."
)
final_result = AgentRunResult(
user_request=user_request,
final_answer=final_answer,
tool_calls=executed_calls,
tool_results=executed_results,
)
yield AgentRunEvent(
event="final_answer",
round_index=round_index,
message="The model produced the final answer from the available results.",
final_result=final_result,
)
return
for model_tool_call in message.tool_calls:
call = ToolCall(
call_id=model_tool_call.id,
tool_name=model_tool_call.function.name,
arguments=_parse_tool_arguments(
model_tool_call.function.arguments
),
)
yield AgentRunEvent(
event="tool_requested",
round_index=round_index,
message=f"The model requested the '{call.tool_name}' tool.",
call=call,
)
authorization = authorize_tool_call(call)
if authorization.status == "approved":
yield AgentRunEvent(
event="tool_approved",
round_index=round_index,
message=(
f"Application approved '{call.tool_name}': "
f"{authorization.reason}"
),
call=call,
authorization=authorization,
)
else:
result = execute_authorized_tool(
call=call,
authorization=authorization,
)
write_audit_record(call=call, result=result)
executed_calls.append(call)
executed_results.append(result)
yield AgentRunEvent(
event="tool_blocked",
round_index=round_index,
message=(
f"Application blocked '{call.tool_name}' before execution: "
f"{authorization.reason}"
),
call=call,
authorization=authorization,
result=result,
)
messages.append(
{
"role": "tool",
"tool_call_id": model_tool_call.id,
"content": json.dumps(
{
"status": result.status,
"output": result.output,
"error": result.error,
},
default=str,
),
}
)
continue
result = execute_authorized_tool(
call=call,
authorization=authorization,
)
write_audit_record(
call=call,
result=result,
)
executed_calls.append(call)
executed_results.append(result)
event_name = (
"tool_succeeded"
if result.status == "success"
else "tool_failed"
)
event_message = (
f"Approved tool '{call.tool_name}' executed successfully."
if result.status == "success"
else (
f"Approved tool '{call.tool_name}' failed during execution: "
f"{result.error}"
)
)
yield AgentRunEvent(
event=event_name,
round_index=round_index,
message=event_message,
call=call,
authorization=authorization,
result=result,
)
messages.append(
{
"role": "tool",
"tool_call_id": model_tool_call.id,
"content": json.dumps(
{
"status": result.status,
"output": result.output,
"error": result.error,
},
default=str,
),
}
)
final_result = AgentRunResult(
user_request=user_request,
final_answer=(
"The agent reached the maximum number "
"of tool-execution rounds before producing "
"a final answer."
),
tool_calls=executed_calls,
tool_results=executed_results,
)
yield AgentRunEvent(
event="max_rounds_reached",
round_index=max_tool_rounds,
message="The configured tool-round limit was reached.",
final_result=final_result,
)
def run_agent(
user_request: str,
model: str | None = None,
max_tool_rounds: int = 5,
) -> AgentRunResult:
"""Run the tool-using agent and return the final structured result."""
final_result: AgentRunResult | None = None
for event in run_agent_iter(
user_request=user_request,
model=model,
max_tool_rounds=max_tool_rounds,
):
if event.final_result is not None:
final_result = event.final_result
if final_result is None:
raise RuntimeError(
"Agent run ended without a final result."
)
return final_result
|