262 lines
11 KiB
Python
262 lines
11 KiB
Python
"""LLM task planner (Manager-first fallback).
|
|
|
|
Ported from agent_swarm_v4. This planner is used ONLY as a fallback to decompose a
|
|
swarm objective into specialist subtasks when the Manager's orchestration_plan does
|
|
not provide an explicit agent breakdown AND the operator opts in via
|
|
``ENABLE_PLANNER_FALLBACK``. It never overrides a Manager-supplied plan.
|
|
|
|
It degrades gracefully: with no API key or on any error it returns a static
|
|
implementation -> testing -> documentation plan, so the runtime never hard-depends on
|
|
the model being reachable.
|
|
"""
|
|
import json
|
|
import logging
|
|
import os
|
|
from typing import List, Dict
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
try:
|
|
from openai import AsyncOpenAI
|
|
except Exception: # pragma: no cover - only when openai is absent
|
|
AsyncOpenAI = None
|
|
|
|
|
|
def _planner_timeout() -> float:
|
|
try:
|
|
return float(os.getenv("PLANNER_TIMEOUT_SECONDS", "45") or 45)
|
|
except ValueError:
|
|
return 45.0
|
|
|
|
|
|
def _max_subtasks() -> int:
|
|
try:
|
|
return int(os.getenv("MAX_SUBTASKS", "6") or 6)
|
|
except ValueError:
|
|
return 6
|
|
|
|
|
|
class Planner:
|
|
"""Decomposes an objective into specialist subtask specs."""
|
|
|
|
def __init__(self):
|
|
api_key = os.getenv("OPENAI_API_KEY") or os.getenv("MODEL_API_KEY")
|
|
api_base = (
|
|
os.getenv("OPENAI_API_BASE")
|
|
or os.getenv("MODEL_API_BASE")
|
|
or "https://api.openai.com/v1"
|
|
)
|
|
model = (
|
|
os.getenv("OPENAI_MODEL")
|
|
or os.getenv("MODEL_NAME")
|
|
or os.getenv("MODEL_ID")
|
|
or "gpt-4o-mini"
|
|
)
|
|
self.model = os.getenv("MASTER_REVIEW_MODEL", model)
|
|
self.client = (
|
|
AsyncOpenAI(api_key=api_key, base_url=api_base)
|
|
if (api_key and AsyncOpenAI is not None)
|
|
else None
|
|
)
|
|
|
|
def _static_plan(self, run_id: str) -> List[Dict]:
|
|
return [
|
|
{
|
|
"subtask_id": f"{run_id}-implementation",
|
|
"description": "Implement the core code required by the user request.",
|
|
"required_capabilities": ["python", "code_generation"],
|
|
"role": "implementation",
|
|
"depends_on": [],
|
|
},
|
|
{
|
|
"subtask_id": f"{run_id}-testing",
|
|
"description": "Write tests for the implemented functionality.",
|
|
"required_capabilities": ["testing", "pytest"],
|
|
"role": "testing",
|
|
"depends_on": [f"{run_id}-implementation"],
|
|
},
|
|
{
|
|
"subtask_id": f"{run_id}-documentation",
|
|
"description": "Write concise documentation based on the implementation and tests.",
|
|
"required_capabilities": ["technical-writing", "general"],
|
|
"role": "documentation",
|
|
"depends_on": [f"{run_id}-implementation", f"{run_id}-testing"],
|
|
},
|
|
]
|
|
|
|
async def build_plan(self, run_id: str, objective: str) -> List[Dict]:
|
|
"""Return specialist subtask specs for an objective, or a static fallback."""
|
|
fallback = self._static_plan(run_id)
|
|
if not self.client:
|
|
return fallback
|
|
try:
|
|
response = await self.client.chat.completions.create(
|
|
model=self.model,
|
|
messages=[
|
|
{
|
|
"role": "system",
|
|
"content": (
|
|
"Break the request into at most 6 specialist subtasks. "
|
|
"Return JSON with key 'subtasks'; each subtask has subtask_id, "
|
|
"description, role, required_capabilities (list), depends_on (list of subtask_ids)."
|
|
),
|
|
},
|
|
{"role": "user", "content": f"Run id: {run_id}\nObjective: {objective}"},
|
|
],
|
|
response_format={"type": "json_object"},
|
|
timeout=_planner_timeout(),
|
|
)
|
|
content = response.choices[0].message.content or "{}"
|
|
subtasks = (json.loads(content) or {}).get("subtasks")
|
|
if not subtasks:
|
|
return fallback
|
|
return subtasks[: _max_subtasks()]
|
|
except Exception as e:
|
|
logger.warning(f"Planner LLM call failed ({e}); using static fallback plan")
|
|
return fallback
|
|
|
|
async def review(self, objective: str, tasks: List[Dict], results: Dict) -> Dict:
|
|
"""Judge whether the combined specialist results are good enough.
|
|
|
|
Returns {accepted: bool, summary: str, retry_tasks: [task_id, ...]}. Falls back to a
|
|
deterministic consistency heuristic when no model is available or the call fails, so the
|
|
review gate degrades safely instead of blocking the run.
|
|
"""
|
|
artifact_summary = self._summarize_results(results)
|
|
if self.client:
|
|
try:
|
|
response = await self.client.chat.completions.create(
|
|
model=self.model,
|
|
messages=[
|
|
{
|
|
"role": "system",
|
|
"content": (
|
|
"Evaluate whether the specialist results jointly satisfy the objective and are "
|
|
"semantically aligned. Compare claimed error behavior, function/API names, and usage "
|
|
"examples across specialists. Reject when behavior claims conflict or the objective is "
|
|
"unmet. Return JSON with keys: accepted (bool), summary (str), retry_tasks (list of the "
|
|
"result keys that must be redone)."
|
|
),
|
|
},
|
|
{
|
|
"role": "user",
|
|
"content": json.dumps(
|
|
{
|
|
"objective": objective,
|
|
"tasks": tasks,
|
|
"results": artifact_summary,
|
|
"result_keys": list(results.keys()),
|
|
},
|
|
ensure_ascii=False,
|
|
),
|
|
},
|
|
],
|
|
response_format={"type": "json_object"},
|
|
timeout=_planner_timeout(),
|
|
)
|
|
result = json.loads(response.choices[0].message.content or "{}")
|
|
if result:
|
|
result.setdefault("accepted", True)
|
|
result.setdefault("summary", "accepted")
|
|
result.setdefault("retry_tasks", [])
|
|
# Only keep retry targets that are real result keys.
|
|
result["retry_tasks"] = [t for t in result["retry_tasks"] if t in results]
|
|
return result
|
|
except Exception as e:
|
|
logger.warning(f"Review LLM call failed ({e}); using heuristic consistency check")
|
|
|
|
consistency = self._heuristic_consistency_check(results)
|
|
if not consistency["accepted"]:
|
|
return consistency
|
|
return {"accepted": True, "summary": "fallback acceptance", "retry_tasks": []}
|
|
|
|
async def synthesize(self, objective: str, results: Dict) -> str:
|
|
"""Compose one coherent answer from the specialist results.
|
|
|
|
Uses the model when available; otherwise concatenates specialist summaries so a
|
|
unified response always exists.
|
|
"""
|
|
summary = self._summarize_results(results)
|
|
deterministic = " | ".join(
|
|
f"{key}: {value.get('summary') or value.get('changes') or 'no summary'}"
|
|
for key, value in summary.items()
|
|
) or objective
|
|
if not self.client:
|
|
return deterministic
|
|
try:
|
|
response = await self.client.chat.completions.create(
|
|
model=self.model,
|
|
messages=[
|
|
{
|
|
"role": "system",
|
|
"content": (
|
|
"Synthesize the specialist results into one concise, user-facing answer to the "
|
|
"objective. Resolve overlaps in favor of implementation semantics."
|
|
),
|
|
},
|
|
{
|
|
"role": "user",
|
|
"content": json.dumps(
|
|
{"objective": objective, "results": summary}, ensure_ascii=False
|
|
),
|
|
},
|
|
],
|
|
timeout=_planner_timeout(),
|
|
)
|
|
return (response.choices[0].message.content or "").strip() or deterministic
|
|
except Exception as e:
|
|
logger.warning(f"Synthesis LLM call failed ({e}); using concatenated summary")
|
|
return deterministic
|
|
|
|
def _summarize_results(self, results: Dict) -> Dict:
|
|
summary = {}
|
|
for task_id, payload in results.items():
|
|
result = (payload or {}).get("result", {}) or {}
|
|
subtasks = result.get("subtasks", []) or []
|
|
files, combined_changes, combined_summary = [], [], []
|
|
for item in subtasks:
|
|
files.extend(item.get("files_modified", []) or [])
|
|
if item.get("changes"):
|
|
combined_changes.append(item.get("changes"))
|
|
if item.get("summary"):
|
|
combined_summary.append(item.get("summary"))
|
|
if not combined_summary and result.get("summary"):
|
|
combined_summary.append(result.get("summary"))
|
|
summary[task_id] = {
|
|
"files_modified": files,
|
|
"summary": " ".join(combined_summary),
|
|
"changes": " ".join(combined_changes),
|
|
}
|
|
return summary
|
|
|
|
def _heuristic_consistency_check(self, results: Dict) -> Dict:
|
|
combined_text, task_ids = [], []
|
|
for task_id, payload in results.items():
|
|
task_ids.append(task_id)
|
|
result = (payload or {}).get("result", {}) or {}
|
|
for item in result.get("subtasks", []) or []:
|
|
combined_text.append(item.get("summary", "") or "")
|
|
combined_text.append(item.get("changes", "") or "")
|
|
for file_item in item.get("files", []) or []:
|
|
combined_text.append(file_item.get("content", "") or "")
|
|
|
|
joined = "\n".join(combined_text).lower()
|
|
if ("valueerror" in joined) and ("zerodivisionerror" in joined or "zero division" in joined):
|
|
retry = [t for t in task_ids if "documentation" in t or "testing" in t] or task_ids
|
|
return {
|
|
"accepted": False,
|
|
"summary": "conflicting error semantics detected between specialists",
|
|
"retry_tasks": retry,
|
|
}
|
|
if ("pytest" in joined) and ("unittest" in joined):
|
|
retry = [t for t in task_ids if "testing" in t or "documentation" in t] or task_ids
|
|
return {
|
|
"accepted": False,
|
|
"summary": "conflicting test framework expectations detected between specialists",
|
|
"retry_tasks": retry,
|
|
}
|
|
return {"accepted": True, "summary": "heuristic consistency acceptance", "retry_tasks": []}
|
|
|
|
|
|
planner = Planner()
|