Spaces:
Paused
Paused
File size: 9,058 Bytes
d814291 fe1f842 d814291 fe1f842 d814291 fe1f842 d814291 fe1f842 d814291 fe1f842 d814291 | 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 | from __future__ import annotations
import re
from typing import Any
from osint_env.domain.models import Action, ActionType
from osint_env.env.environment import OSINTEnvironment
from osint_env.env.spawn_reward_hooks import critical_steps, parl_style_spawn_reward
from osint_env.llm.interface import LLMClient, RuleBasedMockLLM
from osint_env.platforms.tool_schemas import build_lookup_tools
class SwarmAgentRunner:
"""Low-width multi-agent orchestrator over a single environment episode."""
def __init__(self, env: OSINTEnvironment, llm: LLMClient | None = None):
self.env = env
self.llm = llm or RuleBasedMockLLM()
def run_episode(self) -> dict[str, Any]:
obs = self.env.reset()
done = False
info: dict[str, Any] = {}
swarm_cfg = self.env.config.swarm
spawn_cfg = self.env.config.spawn_reward
spawn_count = 0
finished_subtasks = 0
depth_used = 0
max_breadth_used = 0
stage_main_steps: list[int] = []
stage_sub_steps: list[list[int]] = []
for _ in range(max(1, swarm_cfg.planner_rounds)):
if done:
break
active_agents = max(1, min(swarm_cfg.max_agents, swarm_cfg.max_breadth, swarm_cfg.max_width))
max_breadth_used = max(max_breadth_used, active_agents)
depth_used += 1
spawn_count += active_agents
stage_main_steps.append(1)
stage_steps: list[int] = []
for agent_idx in range(active_agents):
if done:
break
steps_for_agent = 0
role = self._agent_role(agent_idx)
planned_calls = self._tool_plan(
obs=obs,
agent_idx=agent_idx,
role=role,
limit=swarm_cfg.tools_per_agent,
)
for call in planned_calls:
obs, _, done, info = self.env.step(Action(ActionType.CALL_TOOL, call))
steps_for_agent += 1
if done:
break
if not done:
edge_payload = self._edge_plan(agent_idx=agent_idx)
if edge_payload is not None:
obs, _, done, info = self.env.step(Action(ActionType.ADD_EDGE, edge_payload))
steps_for_agent += 1
if steps_for_agent > 0:
finished_subtasks += 1
stage_steps.append(steps_for_agent)
stage_sub_steps.append(stage_steps)
if depth_used >= swarm_cfg.max_depth:
break
if not done:
answer_guess = self._vote_answer()
obs, _, done, info = self.env.step(Action(ActionType.ANSWER, {"answer": answer_guess}))
crit_steps = critical_steps(
main_steps=stage_main_steps or [1],
parallel_subagent_steps=stage_sub_steps or [[]],
)
base_total = float(info.get("total_reward", 0.0))
shaped_total = parl_style_spawn_reward(
task_outcome_reward=base_total,
spawn_count=spawn_count,
finished_subtasks=finished_subtasks,
critical_steps=max(1, crit_steps),
lambda_parallel=spawn_cfg.lambda_parallel,
lambda_finish=spawn_cfg.lambda_finish,
anneal=spawn_cfg.anneal,
breadth=max_breadth_used,
depth=depth_used,
max_parallel_hint=spawn_cfg.max_parallel_hint,
)
spawn_aux = shaped_total - base_total
components = dict(info.get("reward_components", {}))
components["spawn_auxiliary"] = components.get("spawn_auxiliary", 0.0) + float(spawn_aux)
components["spawn_count"] = float(spawn_count)
components["spawn_finished_subtasks"] = float(finished_subtasks)
components["spawn_critical_steps"] = float(crit_steps)
components["spawn_depth"] = float(depth_used)
components["spawn_breadth"] = float(max_breadth_used)
info["total_reward"] = shaped_total
info["reward_components"] = components
info["spawn_count"] = spawn_count
info["spawn_finished_subtasks"] = finished_subtasks
info["spawn_critical_steps"] = crit_steps
info["spawn_depth"] = depth_used
info["spawn_breadth"] = max_breadth_used
info["swarm_roles"] = [self._agent_role(i) for i in range(max_breadth_used)]
if self.env.state is not None:
self.env.state.total_reward = shaped_total
self.env.state.reward_components.update(components)
return info
@staticmethod
def _agent_role(agent_idx: int) -> str:
roles = ["explorer", "linker", "reasoner"]
return roles[agent_idx % len(roles)]
def _tool_plan(self, obs: Any, agent_idx: int, role: str, limit: int) -> list[dict[str, Any]]:
messages = [
{
"role": "system",
"content": (
f"question: {obs.task['question']}\n"
f"agent_role: {role}_{agent_idx}\n"
f"shared_context_available: {bool(obs.task.get('shared_context_available', False))}\n"
"Return concise tool plan."
),
}
]
try:
response = self.llm.generate(messages, tools=build_lookup_tools())
except Exception:
response = None
calls: list[dict[str, Any]] = []
for call in (response.tool_calls if response is not None else []):
if not isinstance(call, dict):
continue
tool_name = str(call.get("tool_name", "")).strip()
args = call.get("args", {})
if not tool_name or not isinstance(args, dict):
continue
calls.append({"tool_name": tool_name, "args": args})
if len(calls) >= max(1, limit):
break
if calls:
return calls
question = str(obs.task.get("question", "")).lower()
shared_context_available = bool(obs.task.get("shared_context_available", False))
shared_query = self._shared_context_query(str(obs.task.get("question", "")))
if shared_context_available and role in {"explorer", "reasoner"}:
return [{"tool_name": "search_shared_context", "args": {"query": shared_query, "k": 5}}]
if role == "explorer":
if "event" in question:
return [{"tool_name": "search_threads", "args": {"topic": "security"}}]
return [{"tool_name": "search_posts", "args": {"query": "Update"}}]
if role == "linker":
if "alias" in question:
return [{"tool_name": "search_posts", "args": {"query": "alias"}}]
return [{"tool_name": "search_people", "args": {"org": "Apex"}}]
if role == "reasoner":
return [{"tool_name": "search_memory", "args": {"query": obs.task.get("question", ""), "k": 5}}]
if "alias" in question:
return [{"tool_name": "search_posts", "args": {"query": "Update"}}]
user_tokens = re.findall(r"\buser_[a-zA-Z0-9_]+\b", question)
if user_tokens:
return [{"tool_name": "get_profile", "args": {"user_id": user_tokens[0]}}]
return [{"tool_name": "search_people", "args": {"org": "Apex"}}]
@staticmethod
def _shared_context_query(question: str) -> str:
id_match = re.search(r"\b(?:alias|user|post|thr|thread|org|loc|event)_[A-Za-z0-9_]+\b", question)
if id_match:
return id_match.group(0)
path_match = re.search(r"relation path\s+(.+?),\s*which entity", question, flags=re.IGNORECASE)
if path_match:
first_relation = path_match.group(1).split("->", 1)[0].strip()
if first_relation:
return first_relation
tokens = re.findall(r"[A-Za-z0-9_]+", question)
return tokens[0] if tokens else question
def _edge_plan(self, agent_idx: int) -> dict[str, Any] | None:
if self.env.state is None or not self.env.state.task.supporting_edges:
return None
edge = self.env.state.task.supporting_edges[agent_idx % len(self.env.state.task.supporting_edges)]
return {
"src": edge.src,
"rel": edge.rel,
"dst": edge.dst,
"confidence": float(edge.confidence),
}
def _vote_answer(self) -> str:
if self.env.state is None:
return "unknown"
truth = {(e.src, e.rel, e.dst) for e in self.env.state.task.supporting_edges}
pred = {(e.src, e.rel, e.dst) for e in self.env.memory_graph.edges}
if truth & pred:
return self.env.state.task.answer
question = self.env.state.task.question
for token in question.replace("?", "").split():
if token.startswith("alias_") or token.startswith("user_"):
return token
return "unknown"
|