mirror of
https://github.com/bytedance/deer-flow.git
synced 2026-10-26 21:26:39 +00:00
35023bb908
* feat(scheduler): create conversation schedules and support agent self-stop * fix(scheduler): validate goals and trial requests before dispatch --------- Co-authored-by: Totoro-qaq <279883115+Totoro-qaq@users.noreply.github.com>
190 lines
11 KiB
Python
190 lines
11 KiB
Python
"""Offline protocol checks, independently graded artifacts and spend accounting."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import hashlib
|
|
import json
|
|
import math
|
|
from pathlib import Path
|
|
from typing import Any
|
|
from urllib.parse import urlsplit
|
|
|
|
import yaml
|
|
|
|
from app.gateway.routers.console import _ModelPricing, _run_cost
|
|
from deerflow.config.database_config import DEFAULT_CHECKPOINT_SNAPSHOT_FREQUENCY
|
|
|
|
ROOT = Path(__file__).resolve().parent
|
|
PROTOCOL = "scheduled-tasks-synthetic-v1"
|
|
|
|
|
|
def load_fixtures() -> dict[str, Any]:
|
|
value = json.loads((ROOT / "fixtures.json").read_text(encoding="utf-8"))
|
|
if value.get("protocol") != PROTOCOL or value.get("synthetic") is not True:
|
|
raise ValueError("Only the committed, identified synthetic fixture is supported")
|
|
return value
|
|
|
|
|
|
def fixture(case_id: str) -> dict[str, Any]:
|
|
for case in load_fixtures()["cases"]:
|
|
if case["id"] == case_id:
|
|
return case
|
|
raise ValueError("Unknown synthetic fixture case ID")
|
|
|
|
|
|
def task_prompt(case_id: str, filename: str) -> str:
|
|
case = fixture(case_id)
|
|
return (
|
|
f"This is an operator-authorized synthetic benchmark. Read read_scheduled_fixture(case_id='{case_id}'). "
|
|
f"{case['instruction']} Use write_file to create /mnt/user-data/outputs/{filename}, "
|
|
"then present_files so the user can open it. Do not use external network sources or subagents. "
|
|
"Report the computed values in your final answer."
|
|
)
|
|
|
|
|
|
def grade_artifact(text: str, expected: dict[str, Any]) -> dict[str, Any]:
|
|
"""Grade the opened file independently of agent/evaluator self-reports."""
|
|
try:
|
|
observed = json.loads(text)
|
|
except (ValueError, TypeError):
|
|
return {"passed": False, "reason": "artifact_not_json"}
|
|
passed = isinstance(observed, dict) and all(key in observed and observed[key] == value and type(observed[key]) is type(value) for key, value in expected.items())
|
|
return {"passed": passed, "reason": "expected_fields_match" if passed else "artifact_values_wrong", "artifact_sha256": hashlib.sha256(text.encode("utf-8")).hexdigest()}
|
|
|
|
|
|
def require_loopback(base_url: str) -> str:
|
|
parsed = urlsplit(base_url)
|
|
if parsed.scheme != "http" or parsed.hostname not in {"127.0.0.1", "localhost", "::1"} or parsed.username or parsed.password or parsed.query or parsed.fragment or parsed.path not in {"", "/"}:
|
|
raise ValueError("Use a disposable Gateway's explicit loopback HTTP origin")
|
|
return base_url.rstrip("/")
|
|
|
|
|
|
def load_pilot_config(path: Path, model_alias: str) -> dict[str, Any]:
|
|
"""Inspect named fields only; secrets never enter the returned provenance."""
|
|
raw_bytes = path.read_bytes()
|
|
raw = yaml.safe_load(raw_bytes)
|
|
models = raw.get("models") or []
|
|
if not models or models[0].get("name") != model_alias:
|
|
raise ValueError("The pilot model must be the Gateway's configured default model")
|
|
model = models[0]
|
|
if not isinstance(model.get("model"), str) or not model["model"]:
|
|
raise ValueError("The configured provider model ID is required")
|
|
pricing = model.get("pricing") or {}
|
|
if pricing.get("currency") != "USD":
|
|
raise ValueError("Explicit USD pricing is required for the pilot spend guard")
|
|
for field in ("input_per_million", "output_per_million"):
|
|
if not isinstance(pricing.get(field), (int, float)) or not math.isfinite(pricing[field]) or pricing[field] < 0:
|
|
raise ValueError("Explicit finite input/output model prices are required")
|
|
if not raw.get("scheduler", {}).get("enabled") or not raw["scheduler"].get("tool_enabled"):
|
|
raise ValueError("Both scheduler flags must be enabled in the disposable Gateway")
|
|
if raw.get("database", {}).get("backend") != "sqlite":
|
|
raise ValueError("The pilot requires the disposable Gateway's local SQLite database")
|
|
budget = raw.get("token_budget") or {}
|
|
if not budget.get("enabled") or not isinstance(budget.get("max_tokens"), int):
|
|
raise ValueError("An explicit enabled per-run token budget is required")
|
|
if model.get("max_retries") != 0:
|
|
raise ValueError("Set the pilot provider model's max_retries to zero")
|
|
tools = {entry.get("use") for entry in raw.get("tools", [])}
|
|
required = {"scripts.benchmark.scheduled_tasks.fixture_tools:read_scheduled_fixture", "deerflow.sandbox.tools:write_file_tool"}
|
|
if not required <= tools:
|
|
raise ValueError("Configure the synthetic fixture reader and production write_file")
|
|
if raw.get("channels", {}).get("enabled"):
|
|
raise ValueError("Disable IM channels for the disposable pilot")
|
|
if raw.get("plugins") or (raw.get("extensions") or {}).get("middlewares"):
|
|
raise ValueError("The native budget reader is scoped to the default lead schema without plugin/custom middleware")
|
|
return {
|
|
"protocol": PROTOCOL,
|
|
"config_sha256": hashlib.sha256(raw_bytes).hexdigest(),
|
|
"fixture_sha256": hashlib.sha256((ROOT / "fixtures.json").read_bytes()).hexdigest(),
|
|
"model_alias": model_alias,
|
|
"provider_model": model["model"],
|
|
"temperature": model.get("temperature"),
|
|
"model_max_tokens": model.get("max_tokens"),
|
|
"max_retries": model["max_retries"],
|
|
"thinking_type": ((model.get("extra_body") or {}).get("thinking") or {}).get("type"),
|
|
"request_timeout_seconds": model.get("timeout", model.get("request_timeout")),
|
|
"recursion_limit": raw.get("recursion_limit"),
|
|
"case_order": ["inventory", "corrected-ledger", "missing-reviewer"],
|
|
"provider_seed": None,
|
|
"token_budget": budget,
|
|
"normal_pilot_budget_reference": 60000,
|
|
"completion_rule": "correct_artifact_source_host_delivery_successful_run",
|
|
"checkpoint_channel_mode": raw["database"].get("checkpoint_channel_mode", "full"),
|
|
"checkpoint_snapshot_frequency": (raw["database"].get("checkpoint_delta") or {}).get("snapshot_frequency", DEFAULT_CHECKPOINT_SNAPSHOT_FREQUENCY),
|
|
"pricing": {field: pricing[field] for field in ("currency", "input_per_million", "output_per_million")},
|
|
"sqlite_dir": raw["database"]["sqlite_dir"],
|
|
"scheduler_min_delay": raw["scheduler"].get("min_once_delay_seconds", 60),
|
|
"conversation_reader_enabled": "deerflow.tools.conversation:read_conversation" in tools,
|
|
}
|
|
|
|
|
|
def run_cost(row: dict[str, Any], plan: dict[str, Any]) -> float | None:
|
|
accepted = {plan["model_alias"], plan["provider_model"]}
|
|
if row.get("model_name") and row["model_name"] not in accepted:
|
|
raise ValueError("Run used an unexpected model; stop the pilot")
|
|
usage = row.get("token_usage_by_model") or {}
|
|
if set(usage) - accepted:
|
|
raise ValueError("Unpriced or unexpected model bucket; stop the pilot")
|
|
prices = plan["pricing"]
|
|
price = _ModelPricing(input_per_million=prices["input_per_million"], output_per_million=prices["output_per_million"], currency="USD")
|
|
return _run_cost(
|
|
{model: price for model in accepted}, model_name=row.get("model_name") or plan["model_alias"], total_input_tokens=row.get("total_input_tokens"), total_output_tokens=row.get("total_output_tokens"), token_usage_by_model=usage
|
|
)
|
|
|
|
|
|
def require_spend_reserve(spent: float, reserve: float, ceiling: float) -> None:
|
|
if not all(math.isfinite(value) and value >= 0 for value in (spent, reserve, ceiling)) or reserve == 0 or not 0 < ceiling <= 1:
|
|
raise ValueError("Use finite positive per-run reserve and a pilot ceiling no greater than USD 1")
|
|
if spent + reserve > ceiling:
|
|
raise ValueError("Pilot spend guard refuses another paid run")
|
|
|
|
|
|
def assert_arm_installed(arm: str, row: dict[str, Any]) -> None:
|
|
if arm == "goal":
|
|
if not row.get("scheduled_goal_metadata_present") or not row.get("scheduled_goal_objective") or not isinstance(row.get("goal_verdict"), dict):
|
|
raise AssertionError("The goal arm lacks its installed goal metadata or durable verdict; the comparison is invalid")
|
|
elif row.get("scheduled_goal_metadata_present") or row.get("goal_verdict") is not None:
|
|
raise AssertionError("The plain arm unexpectedly carried a goal; the comparison is invalid")
|
|
|
|
|
|
def comparison_summary(results: list[dict[str, Any]]) -> dict[str, Any]:
|
|
arms = {}
|
|
for arm in ("plain", "goal"):
|
|
selected = [result for result in results if result.get("arm") == arm and "run" in result]
|
|
if selected:
|
|
arms[arm] = {
|
|
"occurrences": len(selected),
|
|
"independent_passes": sum(bool(result.get("independent_pass")) for result in selected),
|
|
"completion_rate": sum(bool(result.get("independent_pass")) for result in selected) / len(selected),
|
|
"mean_lead_model_turns": sum(result["evidence"]["lead_model_turns"] for result in selected) / len(selected),
|
|
"mean_llm_calls": sum(result["run"]["llm_call_count"] for result in selected) / len(selected),
|
|
"total_tokens": sum(result["run"]["total_tokens"] for result in selected),
|
|
"estimated_cost_usd": None if any(result.get("estimated_cost_usd") is None for result in selected) else sum(result["estimated_cost_usd"] for result in selected),
|
|
}
|
|
return {"arms": arms, "interpretation": "Small synthetic pilot; descriptive evidence only, without statistical or general completion-rate claims"}
|
|
|
|
|
|
def goal_history_evidence(entries: list[dict[str, Any]], objective: str) -> dict[str, Any]:
|
|
snapshots = [(entry.get("values") or {}).get("goal") for entry in entries]
|
|
goals = [goal for goal in snapshots if isinstance(goal, dict) and goal.get("objective") == objective]
|
|
return {
|
|
"goal_snapshots_observed": len(goals),
|
|
"maximum_observed_continuation_count": max((goal["continuation_count"] for goal in goals), default=None),
|
|
"continuation_limits_observed": sorted({goal["max_continuations"] for goal in goals}),
|
|
"within_observed_limits": bool(goals) and all(goal["continuation_count"] <= goal["max_continuations"] for goal in goals),
|
|
}
|
|
|
|
|
|
def host_delivery_evidence(events: list[dict[str, Any]], filepath: str) -> dict[str, Any]:
|
|
from deerflow.runtime.runs.worker import _presented_path_covers_output
|
|
|
|
receipts = [event["content"] for event in events if event.get("event_type") == "run.delivery" and isinstance(event.get("content"), dict)]
|
|
delivered = any(receipt.get("satisfied") is True and any(isinstance(path, str) and _presented_path_covers_output(path, filepath) for path in (receipt.get("by_tool") or {}).get("present_files", [])) for receipt in receipts)
|
|
return {"delivered": delivered, "host_delivery_receipts": receipts}
|
|
|
|
|
|
def independent_completion(*, artifact_passed: bool, fixture_read: bool, delivered: bool, run_status: str, stop_reason: str | None) -> bool:
|
|
# A cap is a separate process observation; it does not erase a verified
|
|
# artifact that was already delivered successfully.
|
|
return bool(artifact_passed and fixture_read and delivered and run_status == "success")
|