File size: 16,233 Bytes
9abace2 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 | """
Supervisor Agent with Tool-Lookup Orchestration
Part of SOVEREIGN PYTHON LLM ENGINE
Orchestrates sub-agents by:
1. Breaking a high-level task into SubTasks via model planning.
2. Assigning the best tool to each SubTask via ToolLookupRegistry.
3. Executing each subtask (delegating to a callable or simulating).
4. Collecting and returning results.
"""
from __future__ import annotations
import asyncio
import uuid
from dataclasses import dataclass, field
from typing import Any, Awaitable, Callable
from .lookup import ToolCheckout, ToolLookupRegistry, ToolQuery
from .registry import RiskClass
# ==========================================
# SubTask
# ==========================================
@dataclass
class SubTask:
"""
A single unit of work within a larger plan.
Attributes:
id: Unique task identifier (auto-generated if not provided).
description: Human-readable description of what must be done.
required_capability: Natural-language description of the tool
capability needed to complete this subtask.
assigned_tool: tool_id of the checked-out tool (None until assigned).
status: Lifecycle state:
"pending" — not yet assigned.
"assigned" — tool checked out, ready to execute.
"done" — executed successfully.
"failed" — execution raised an exception.
result: Output produced by execution; None until done.
"""
description: str
required_capability: str
id: str = field(default_factory=lambda: str(uuid.uuid4())[:8])
assigned_tool: str | None = None
status: str = "pending"
result: dict[str, Any] | None = None
# ==========================================
# Model provider protocol (minimal duck-typing)
# ==========================================
class ModelProvider:
"""
Minimal interface for a text-generation model.
Implementations must provide an async ``complete`` method.
The default stub returns empty completions so that
SupervisorAgent can be instantiated without a real model.
"""
async def complete(self, prompt: str) -> str: # pragma: no cover
"""
Generate a text completion for the given prompt.
Args:
prompt: Instruction / context string.
Returns:
Model-generated text.
"""
return ""
# ==========================================
# SupervisorAgent
# ==========================================
class SupervisorAgent:
"""
Orchestrator that decomposes tasks and routes them to specialised tools.
The agent follows a plan → assign → execute loop:
1. ``plan(task)`` — ask model to decompose task into SubTasks.
2. ``assign(subtask)`` — find the best tool via ToolLookupRegistry.
3. ``execute_plan`` — run the full loop and collect results.
A ``delegate`` helper lets the supervisor hand off individual subtasks
to sub-agent callables while still managing tool checkout lifecycle.
"""
# Prompt templates
_PLAN_PROMPT = (
"You are a planning assistant. Break the following task into a numbered "
"list of discrete subtasks. For each subtask provide:\n"
" - A short description (one sentence).\n"
" - The tool capability required (what kind of tool is needed).\n\n"
"Output format (repeat for each subtask):\n"
"SUBTASK: <description>\n"
"CAPABILITY: <required tool capability>\n\n"
"Task: {task}\n"
)
def __init__(
self,
lookup: ToolLookupRegistry,
model_provider: ModelProvider | None = None,
agent_id: str | None = None,
risk_max: RiskClass = RiskClass.REVERSIBLE_REMOTE_WRITE,
max_parallel_tasks: int = 4,
) -> None:
"""
Initialise the supervisor.
Args:
lookup: ToolLookupRegistry used for tool discovery and checkout.
model_provider: Language model used for planning. A no-op stub
is used when None.
agent_id: Stable identifier for this supervisor instance.
Auto-generated if not provided.
risk_max: Hard ceiling on the risk class of tools this agent
may check out.
max_parallel_tasks: Maximum subtasks to execute concurrently.
"""
self._lookup = lookup
self._model = model_provider or ModelProvider()
self.agent_id = agent_id or f"supervisor-{uuid.uuid4().hex[:8]}"
self._risk_max = risk_max
self._max_parallel = max_parallel_tasks
# ------------------------------------------------------------------
# Planning
# ------------------------------------------------------------------
async def plan(self, task: str) -> list[SubTask]:
"""
Decompose a high-level task into SubTasks using the model.
The model is prompted to produce a structured list of subtasks.
The response is parsed line by line; any subtask whose description
or capability cannot be extracted is silently skipped.
Falls back to a single pass-through SubTask when the model
returns an empty or unparseable response.
Args:
task: Natural-language task description.
Returns:
Ordered list of SubTask objects with status "pending".
"""
prompt = self._PLAN_PROMPT.format(task=task)
raw_response = await self._model.complete(prompt)
subtasks = self._parse_plan(raw_response, task)
return subtasks
def _parse_plan(self, response: str, original_task: str) -> list[SubTask]:
"""
Parse model response into SubTask objects.
Expects lines of the form:
SUBTASK: <description>
CAPABILITY: <capability>
Args:
response: Raw model text.
original_task: Used as fallback when parsing yields nothing.
Returns:
List of SubTask objects.
"""
subtasks: list[SubTask] = []
pending_desc: str | None = None
pending_cap: str | None = None
for line in response.splitlines():
line = line.strip()
if line.upper().startswith("SUBTASK:"):
# Flush any complete pending pair
if pending_desc and pending_cap:
subtasks.append(
SubTask(
description=pending_desc,
required_capability=pending_cap,
)
)
pending_desc = line[len("SUBTASK:"):].strip()
pending_cap = None
elif line.upper().startswith("CAPABILITY:"):
pending_cap = line[len("CAPABILITY:"):].strip()
# Flush final pair
if pending_desc and pending_cap:
subtasks.append(
SubTask(
description=pending_desc,
required_capability=pending_cap,
)
)
# Fallback: treat the whole task as one subtask
if not subtasks:
subtasks.append(
SubTask(
description=original_task,
required_capability=original_task,
)
)
return subtasks
# ------------------------------------------------------------------
# Assignment
# ------------------------------------------------------------------
async def assign(self, subtask: SubTask) -> ToolCheckout:
"""
Find and check out the best tool for a subtask.
Uses ``ToolLookupRegistry.search`` to rank tools by keyword
relevance, respecting the supervisor's risk ceiling. The
highest-scored tool is checked out and the subtask is marked
"assigned".
Args:
subtask: The subtask to assign a tool to.
Returns:
ToolCheckout for the assigned tool.
Raises:
RuntimeError: If no matching tool is found.
"""
query = ToolQuery(
query=subtask.required_capability,
agent_id=self.agent_id,
risk_max=self._risk_max,
top_k=1,
)
result = self._lookup.search(query)
if not result.tools:
subtask.status = "failed"
subtask.result = {
"error": f"No tool found for capability: {subtask.required_capability!r}"
}
raise RuntimeError(
f"No tool found for subtask {subtask.id!r}: "
f"{subtask.required_capability!r}"
)
best_tool = result.tools[0]
checkout = self._lookup.checkout(self.agent_id, best_tool.tool_id)
subtask.assigned_tool = best_tool.tool_id
subtask.status = "assigned"
return checkout
# ------------------------------------------------------------------
# Execution
# ------------------------------------------------------------------
async def _execute_subtask(
self,
subtask: SubTask,
inputs: dict[str, Any] | None = None,
) -> dict[str, Any]:
"""
Execute a single assigned subtask using its checked-out tool.
Calls the tool handler with ``inputs`` (defaults to empty dict).
Records the tool use and marks the subtask done or failed.
Args:
subtask: The assigned subtask to execute.
inputs: Input parameters forwarded to the tool handler.
Returns:
Tool execution result dict.
Raises:
RuntimeError: If the subtask is not in "assigned" state.
"""
if subtask.status != "assigned":
raise RuntimeError(
f"Cannot execute subtask {subtask.id!r} in state {subtask.status!r}. "
"Call assign() first."
)
tool_id = subtask.assigned_tool
if tool_id is None:
raise RuntimeError(f"Subtask {subtask.id!r} has no assigned_tool.")
tool_def = self._lookup._registry.get(tool_id)
if tool_def is None:
subtask.status = "failed"
subtask.result = {"error": f"Tool {tool_id!r} disappeared from registry."}
return subtask.result
try:
output = await tool_def.handler(inputs or {})
self._lookup.record_use(self.agent_id, tool_id)
subtask.status = "done"
subtask.result = output
return output
except Exception as exc:
subtask.status = "failed"
subtask.result = {"error": str(exc), "tool_id": tool_id}
raise
async def execute_plan(
self,
task: str,
inputs_by_subtask: dict[str, dict[str, Any]] | None = None,
) -> dict[str, Any]:
"""
Full plan → assign → execute → collect loop.
Steps:
1. Call ``plan`` to decompose the task.
2. Assign a tool to each subtask (sequential, fast).
3. Execute all subtasks concurrently (up to ``max_parallel_tasks``).
4. Return a summary with per-subtask results.
Assignment failures do not abort the loop; the subtask is marked
"failed" and execution is skipped for it.
Args:
task: High-level task description.
inputs_by_subtask: Optional map of subtask_id -> input params.
Subtasks not in the map receive an empty input dict.
Returns:
Summary dict with keys:
"task", "subtasks" (list of serialised SubTask results),
"success_count", "failure_count".
"""
subtasks = await self.plan(task)
inputs_by_subtask = inputs_by_subtask or {}
# Assignment is fast (registry lookup) — do sequentially
for subtask in subtasks:
try:
await self.assign(subtask)
except RuntimeError:
# subtask already marked failed inside assign()
pass
# Execute assigned subtasks with bounded concurrency
semaphore = asyncio.Semaphore(self._max_parallel)
async def _bounded_execute(subtask: SubTask) -> None:
async with semaphore:
if subtask.status != "assigned":
return
inputs = inputs_by_subtask.get(subtask.id, {})
try:
await self._execute_subtask(subtask, inputs)
except Exception:
# Error already recorded on subtask.result
pass
await asyncio.gather(*[_bounded_execute(st) for st in subtasks])
success = sum(1 for st in subtasks if st.status == "done")
failure = sum(1 for st in subtasks if st.status == "failed")
return {
"task": task,
"subtasks": [
{
"id": st.id,
"description": st.description,
"required_capability": st.required_capability,
"assigned_tool": st.assigned_tool,
"status": st.status,
"result": st.result,
}
for st in subtasks
],
"success_count": success,
"failure_count": failure,
}
# ------------------------------------------------------------------
# Delegation
# ------------------------------------------------------------------
async def delegate(
self,
task: str,
agent_callable: Callable[[str, list[str]], Awaitable[dict[str, Any]]],
) -> dict[str, Any]:
"""
Delegate a task to a sub-agent callable with pre-checked-out tools.
The supervisor:
1. Plans the task to determine which tools are needed.
2. Checks out all required tools.
3. Passes the task and list of tool_ids to ``agent_callable``.
4. Checks in all tools after the callable completes (or fails).
The sub-agent callable signature:
async def agent(task: str, tool_ids: list[str]) -> dict
Args:
task: High-level task description.
agent_callable: Coroutine function that executes the task.
Returns:
Result dict from ``agent_callable``, augmented with
"delegated_tools" (list of tool_ids that were checked out).
"""
subtasks = await self.plan(task)
checked_out_ids: list[str] = []
for subtask in subtasks:
try:
checkout = await self.assign(subtask)
checked_out_ids.append(checkout.tool_id)
except RuntimeError:
pass # Best-effort; sub-agent must handle missing tools
try:
result = await agent_callable(task, checked_out_ids)
finally:
for tool_id in checked_out_ids:
self._lookup.checkin(self.agent_id, tool_id)
result["delegated_tools"] = checked_out_ids
return result
# ------------------------------------------------------------------
# Convenience
# ------------------------------------------------------------------
def get_active_tools(self) -> list[str]:
"""
Return the tool_ids currently checked out by this supervisor.
Returns:
List of tool_id strings.
"""
return [t.tool_id for t in self._lookup.get_agent_tools(self.agent_id)]
def release_all(self) -> None:
"""
Check in all tools and clear the session cache for this supervisor.
"""
self._lookup.release_agent(self.agent_id)
|