deer-flow/backend/scripts/benchmark/checkpoint/bench_tool_result_probe.py

#!/usr/bin/env python
"""Probe: do oversized tool results reach checkpoint state as full text? (#4189 item 4)

`ToolOutputBudgetMiddleware` is registered by default and externalizes tool
results above `externalize_min_chars` (preview + file reference under
`.tool-results/`). This probe quantifies the storage effect of that
transformation: the same oversized tool result is driven through the
middleware's `awrap_tool_call` (or run raw), the resulting ToolMessage is
written into a checkpointed graph state, and the per-thread checkpoint
storage (rows + bytes, same normalized shape as bench_channels) is reported
for SQLite.

Scope limit (state this when citing the numbers): the middleware is invoked
manually and the resulting message is injected with ``aupdate_state`` — the
production agent-factory path (middleware stack wiring, ThreadDataMiddleware
runtime state, tool-node task writes) is NOT exercised. The probe therefore
bounds the middleware's own transformation and the checkpoint cost of its
output; by itself it cannot establish that "the default configuration
covers item 4". A residual gap claim must name the concrete factory path
and come with its own measurements.

Item 4 of #4189 can be closed as covered if the wrapped path's checkpoint
bytes stay flat as the result size grows and no factory-path measurement
shows full text landing in state.

Usage:
    cd backend
    python scripts/benchmark/checkpoint/bench_tool_result_probe.py \
        [--result-bytes 50000] [--outputs-dir .tool-results-probe]

--outputs-dir may contain unrelated files: the probe creates and removes
only its own ``probe-run-*`` child inside it (externalized samples land
there), while SQLite databases go to a unique per-run temp directory.

Output: JSON on stdout.
"""

from __future__ import annotations

import argparse
import asyncio
import json
import tempfile
from pathlib import Path
from types import SimpleNamespace
from typing import Annotated, Any, TypedDict
from uuid import uuid4

from langchain_core.messages import AnyMessage, ToolMessage
from langgraph.checkpoint.sqlite.aio import AsyncSqliteSaver
from langgraph.graph import StateGraph
from langgraph.graph.message import add_messages

from deerflow.agents.middlewares.tool_output_budget_middleware import ToolOutputBudgetMiddleware

PROBE_TOOL = "probe_tool"


class ProbeState(TypedDict):
    messages: Annotated[list[AnyMessage], add_messages]


async def _stats_sqlite(saver: AsyncSqliteSaver, thread_id: str) -> dict[str, int]:
    async def one(sql: str) -> tuple[int, int]:
        async with saver.conn.execute(sql, (thread_id,)) as cursor:
            row = await cursor.fetchone()
        return int(row[0]), int(row[1] or 0)

    cp = await one("SELECT COUNT(*), COALESCE(SUM(LENGTH(checkpoint) + LENGTH(metadata)), 0) FROM checkpoints WHERE thread_id = ?")
    wr = await one("SELECT COUNT(*), COALESCE(SUM(LENGTH(value)), 0) FROM writes WHERE thread_id = ?")
    return {
        "checkpoint_rows": cp[0],
        "checkpoint_bytes": cp[1],
        "write_rows": wr[0],
        "write_bytes": wr[1],
    }


def _make_graph(saver: Any) -> Any:
    def call_probe_tool(state: dict[str, Any]) -> dict[str, Any]:
        # the oversized result never flows through this node: the probe
        # injects the (possibly externalized) ToolMessage via aupdate_state
        # below, so the graph only provides a checkpointed state container
        return {}

    builder = StateGraph(ProbeState)
    builder.add_node("tool", call_probe_tool)
    builder.set_entry_point("tool")
    builder.set_finish_point("tool")
    return builder.compile(checkpointer=saver)


async def _run_path(
    label: str,
    result_bytes: int,
    *,
    wrapped: bool,
    outputs_dir: Path | None,
    tmp_dir: Path,
) -> dict[str, Any]:
    tmp_dir.mkdir(parents=True, exist_ok=True)
    async with AsyncSqliteSaver.from_conn_string(str(tmp_dir / f"probe-{label}.sqlite")) as saver:
        await saver.setup()
        middleware = ToolOutputBudgetMiddleware() if wrapped else None
        graph = _make_graph(saver)

        thread_id = f"probe-{label}"
        config = {"configurable": {"thread_id": thread_id}}

        # invoke the tool through the middleware's wrap (or raw), then persist
        # the resulting ToolMessage into graph state and take a checkpoint
        oversized = "A" * result_bytes
        request = SimpleNamespace(
            tool_call={"name": PROBE_TOOL, "id": "probe-call-1"},
            runtime=SimpleNamespace(state={"thread_data": {"outputs_path": str(outputs_dir)}} if outputs_dir else {"thread_data": None}),
        )

        async def handler(_request: Any) -> ToolMessage:
            return ToolMessage(content=oversized, tool_call_id="probe-call-1")

        if wrapped and middleware is not None:
            message = await middleware.awrap_tool_call(request, handler)
        else:
            message = await handler(request)

        await graph.aupdate_state(config, {"messages": [message]})
        stats = await _stats_sqlite(saver, thread_id)
        content_chars = len(message.content) if isinstance(message.content, str) else -1

    return {
        "path": label,
        "result_bytes": result_bytes,
        "tool_message_content_chars": content_chars,
        "externalized_file_bytes": (sum(f.stat().st_size for f in outputs_dir.rglob("*") if f.is_file()) if outputs_dir and outputs_dir.exists() else 0),
        "checkpoint": stats,
    }


async def _main(result_bytes: int, outputs_dir: Path, tmp_dir: Path) -> dict[str, Any]:
    raw = await _run_path("raw-unwrapped", result_bytes, wrapped=False, outputs_dir=None, tmp_dir=tmp_dir)
    externalized = await _run_path("budget-externalized", result_bytes, wrapped=True, outputs_dir=outputs_dir, tmp_dir=tmp_dir)
    truncated = await _run_path("budget-truncated", result_bytes, wrapped=True, outputs_dir=None, tmp_dir=tmp_dir)
    return {
        "result_bytes": result_bytes,
        "paths": [raw, externalized, truncated],
        "verdict": {
            "wrapped_content_stays_small": externalized["tool_message_content_chars"] < result_bytes,
            "raw_content_is_full": raw["tool_message_content_chars"] == result_bytes,
        },
    }


def run_probe(result_bytes: int, outputs_dir: Path, tmp_dir: Path) -> dict[str, Any]:
    """Drive one probe run, cleaning up only directories this run owns.

    ``outputs_dir`` may be user-supplied and may pre-exist with unrelated
    files: the run writes into (and removes) a fresh owned ``probe-run-*``
    child of it, never the directory itself or anything beside it. ``tmp_dir``
    hosts the per-run SQLite databases and is emptied by the cleanup.
    """
    import shutil

    owned_outputs = outputs_dir / f"probe-run-{uuid4().hex[:8]}"
    owned_outputs.mkdir(parents=True, exist_ok=False)
    tmp_dir.mkdir(parents=True, exist_ok=True)
    try:
        return asyncio.run(_main(result_bytes, owned_outputs, tmp_dir))
    finally:
        shutil.rmtree(owned_outputs, ignore_errors=True)
        shutil.rmtree(tmp_dir, ignore_errors=True)


def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--result-bytes", type=int, default=50_000)
    parser.add_argument("--outputs-dir", type=Path, default=Path(".tool-results-probe"))
    args = parser.parse_args()

    outputs_dir: Path = args.outputs_dir
    outputs_dir.mkdir(parents=True, exist_ok=True)
    tmp_dir = Path(tempfile.mkdtemp(prefix="deerflow-probe-"))
    report = run_probe(args.result_bytes, outputs_dir, tmp_dir)

    json.dump(report, __import__("sys").stdout, indent=2)
    print()


if __name__ == "__main__":
    main()