feat(bench): live competitor lane - black-box MCP stdio adapters, pinned checkouts, published results
- adapters/mcp_stdio.py: drives competitors over MCP stdio like a real host; payload-level error detection, fastmcp output-wrapper normalization, cold-start warmup, artifact-based verification (harness re-opens every exported DXF with ezdxf) - competitors_env.py: reproducible pinned-SHA checkout + isolated venv per competitor - adapters for puran-water/autocad-mcp (95476a3) and beiming183-cloud/AutoCAD-MCP (11f7c47) mapped to their own documented tool contracts; unmapped tasks report 'unsupported' with reasons - run_competitors.py: registry-driven --server from competitors.yaml - published reports (pro 100.0 / beiming 50.0 / puran 45.0) + render_live_chart.py -> docs/assets/autocad-mcp-livebench.svg - benchmark.yml: weekly + manual dispatch matrix Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Umutcan Edizsalan
co-authored by
Cursor
Cursor
parent
2f56a73093
commit
6d7c39e012
@@ -0,0 +1,304 @@
|
||||
"""Generic MCP-stdio benchmark adapter.
|
||||
|
||||
Drives a *competitor* MCP server as a black box over the standard stdio
|
||||
transport (fastmcp ``Client``), exactly the way an agent host would. No
|
||||
competitor code is imported; verification never trusts the competitor's
|
||||
response payload — every task that claims to have produced geometry must
|
||||
save a DXF which the harness re-opens with ezdxf and inspects itself.
|
||||
|
||||
Subclasses provide the launch runtime (via ``benchmarks.competitors_env``)
|
||||
and per-task playbooks named ``_task_<task_id>``; anything unmapped is
|
||||
reported as ``unsupported`` (coverage-aware scoring already accounts for it).
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import dataclasses
|
||||
import json
|
||||
import re
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
import ezdxf
|
||||
from fastmcp import Client
|
||||
from fastmcp.client.transports import StdioTransport
|
||||
|
||||
from benchmarks.adapters.base import BenchmarkAdapter, TaskResult
|
||||
from benchmarks.competitors_env import CompetitorRuntime, CompetitorSpec, ensure_competitor
|
||||
from benchmarks.tasks_v2 import TaskSpec
|
||||
|
||||
CALL_TIMEOUT_SECONDS = 15.0
|
||||
WARMUP_TIMEOUT_SECONDS = 90.0
|
||||
|
||||
|
||||
class ToolCallFailed(RuntimeError):
|
||||
"""A competitor tool call returned an error result."""
|
||||
|
||||
|
||||
def read_dxf(path: Path):
|
||||
return ezdxf.readfile(str(path))
|
||||
|
||||
|
||||
def modelspace_types(doc) -> list[str]:
|
||||
return [entity.dxftype() for entity in doc.modelspace()]
|
||||
|
||||
|
||||
def find_dxf(artifact_dir: Path, preferred: Path | None = None) -> Path | None:
|
||||
"""Locate the DXF a competitor produced.
|
||||
|
||||
Servers with managed output roots may relocate writes, so fall back to
|
||||
the newest ``*.dxf`` anywhere under the artifact directory.
|
||||
"""
|
||||
if preferred is not None and preferred.is_file():
|
||||
return preferred
|
||||
candidates = sorted(
|
||||
artifact_dir.rglob("*.dxf"), key=lambda item: item.stat().st_mtime, reverse=True
|
||||
)
|
||||
return candidates[0] if candidates else None
|
||||
|
||||
|
||||
class McpStdioAdapter(BenchmarkAdapter):
|
||||
"""Black-box adapter base for competitor MCP servers."""
|
||||
|
||||
name = "mcp-stdio"
|
||||
spec: CompetitorSpec # set by subclasses
|
||||
|
||||
def __init__(self, *, backend: str = "ezdxf"):
|
||||
if backend != "ezdxf":
|
||||
raise ValueError(
|
||||
f"{type(self).__name__} only benchmarks the headless ezdxf lane; "
|
||||
f"got backend={backend!r}. COM lanes run locally, not in CI."
|
||||
)
|
||||
self.backend_name = backend
|
||||
self.artifact_dir = Path()
|
||||
self.runtime: CompetitorRuntime | None = None
|
||||
self._client: Client | None = None
|
||||
self.tool_names: set[str] = set()
|
||||
|
||||
# -- lifecycle ----------------------------------------------------------
|
||||
|
||||
def extra_launch_env(self) -> dict[str, str]:
|
||||
"""Environment computed at setup time (``self.artifact_dir`` is set)."""
|
||||
return {}
|
||||
|
||||
async def setup(self, artifact_dir: Path) -> None:
|
||||
self.artifact_dir = Path(artifact_dir)
|
||||
self.runtime = ensure_competitor(self.spec)
|
||||
env = {**self.runtime.env, **self.extra_launch_env()}
|
||||
env.setdefault("PYTHONIOENCODING", "utf-8")
|
||||
transport = StdioTransport(
|
||||
command=str(self.runtime.python_exe),
|
||||
args=list(self.runtime.entry_args),
|
||||
env=env,
|
||||
cwd=str(self.runtime.repo_dir),
|
||||
)
|
||||
self._client = Client(transport)
|
||||
await self._client.__aenter__()
|
||||
self.tool_names = {tool.name for tool in await self._client.list_tools()}
|
||||
await self._warmup()
|
||||
|
||||
def warmup_calls(self) -> tuple[tuple[str, dict[str, Any]], ...]:
|
||||
"""Calls issued once after connect to absorb the competitor's lazy
|
||||
backend initialization (which can exceed the per-call timeout)."""
|
||||
return ()
|
||||
|
||||
async def _warmup(self) -> None:
|
||||
for tool, arguments in self.warmup_calls():
|
||||
if tool not in self.tool_names:
|
||||
continue
|
||||
try:
|
||||
await self._client.call_tool(
|
||||
tool, arguments, timeout=WARMUP_TIMEOUT_SECONDS, raise_on_error=False
|
||||
)
|
||||
except Exception: # cold-start absorption is best effort
|
||||
continue
|
||||
|
||||
async def cleanup(self) -> None:
|
||||
if self._client is not None:
|
||||
await self._client.__aexit__(None, None, None)
|
||||
self._client = None
|
||||
|
||||
def metadata(self) -> dict:
|
||||
return {
|
||||
"backend": self.backend_name,
|
||||
"kind": "mcp-stdio-blackbox",
|
||||
"repo": self.spec.repo_url,
|
||||
"pinned_sha": self.spec.pinned_sha,
|
||||
"tools_discovered": sorted(self.tool_names),
|
||||
}
|
||||
|
||||
# -- call helpers ---------------------------------------------------------
|
||||
|
||||
async def call(self, tool: str, arguments: dict[str, Any]) -> Any:
|
||||
"""Call one tool; raise ``ToolCallFailed`` on an error result.
|
||||
|
||||
Competitor servers in this family report failures *inside* the payload
|
||||
(``CommandResult(ok=False, error=...)`` serialized as text) rather than
|
||||
via the MCP ``isError`` flag, so the payload is inspected as well.
|
||||
"""
|
||||
if tool not in self.tool_names:
|
||||
raise ToolCallFailed(f"tool {tool!r} is not exposed by {self.name}")
|
||||
result = await self._client.call_tool(
|
||||
tool, arguments, timeout=CALL_TIMEOUT_SECONDS, raise_on_error=False
|
||||
)
|
||||
if getattr(result, "is_error", False):
|
||||
raise ToolCallFailed(f"{tool}({arguments}) -> {_result_text(result)}")
|
||||
data = getattr(result, "data", None)
|
||||
payload = data if data is not None else _result_text(result)
|
||||
payload = _normalize_payload(payload)
|
||||
error = _payload_error(payload)
|
||||
if error:
|
||||
raise ToolCallFailed(f"{tool}({arguments}) -> {error}")
|
||||
return payload
|
||||
|
||||
async def call_op(
|
||||
self,
|
||||
tool: str,
|
||||
operation: str,
|
||||
data: dict[str, Any] | None = None,
|
||||
*,
|
||||
shapes: tuple[dict[str, Any], ...] | None = None,
|
||||
) -> Any:
|
||||
"""Call a consolidated ``{tool}(operation=..., ...)`` style tool.
|
||||
|
||||
Competitors disagree on whether payload fields live under ``data`` or
|
||||
at the top level, and on field spellings. ``shapes`` supplies explicit
|
||||
argument dicts to try in order; otherwise both common layouts of
|
||||
``data`` are attempted. The first non-error result wins.
|
||||
"""
|
||||
attempts: list[dict[str, Any]] = []
|
||||
if shapes:
|
||||
attempts.extend({"operation": operation, **shape} for shape in shapes)
|
||||
else:
|
||||
payload = data or {}
|
||||
attempts.append({"operation": operation, "data": payload})
|
||||
if payload:
|
||||
attempts.append({"operation": operation, **payload})
|
||||
last_error: Exception | None = None
|
||||
for arguments in attempts:
|
||||
try:
|
||||
return await self.call(tool, arguments)
|
||||
except ToolCallFailed as exc: # try the next argument shape
|
||||
last_error = exc
|
||||
raise ToolCallFailed(str(last_error))
|
||||
|
||||
# -- task dispatch --------------------------------------------------------
|
||||
|
||||
async def run_task(self, task: TaskSpec) -> TaskResult:
|
||||
handler = getattr(self, f"_task_{task.task_id}", None)
|
||||
if handler is None:
|
||||
return TaskResult(
|
||||
task.task_id,
|
||||
"unsupported",
|
||||
0.0,
|
||||
f"{self.name} has no documented equivalent for this task",
|
||||
)
|
||||
try:
|
||||
return await handler()
|
||||
except ToolCallFailed as exc:
|
||||
return TaskResult(task.task_id, "fail", 0.0, str(exc)[:500])
|
||||
|
||||
# -- verification ---------------------------------------------------------
|
||||
|
||||
def save_target(self, filename: str) -> Path:
|
||||
return self.artifact_dir / filename
|
||||
|
||||
def verify_dxf(self, preferred: Path | None = None):
|
||||
located = find_dxf(self.artifact_dir, preferred)
|
||||
if located is None:
|
||||
raise ToolCallFailed("no DXF artifact was produced")
|
||||
return located, read_dxf(located)
|
||||
|
||||
|
||||
def _result_text(result: Any) -> str:
|
||||
parts: list[str] = []
|
||||
for item in getattr(result, "content", None) or []:
|
||||
text = getattr(item, "text", None)
|
||||
if text:
|
||||
parts.append(str(text))
|
||||
return " | ".join(parts)[:1000] if parts else "<no text content>"
|
||||
|
||||
|
||||
def _normalize_payload(payload: Any) -> Any:
|
||||
"""Unwrap fastmcp's dynamic output wrappers and JSON-in-text payloads.
|
||||
|
||||
Legacy servers return plain text; the fastmcp client may deserialize it
|
||||
into a generated *dataclass* (e.g. ``entityOutput(result='<json>')``) or a
|
||||
pydantic model. Normalize down to dict / list / str so verification and
|
||||
handle extraction see real data.
|
||||
"""
|
||||
if dataclasses.is_dataclass(payload) and not isinstance(payload, type):
|
||||
payload = dataclasses.asdict(payload)
|
||||
elif hasattr(payload, "model_dump"):
|
||||
payload = payload.model_dump()
|
||||
elif not isinstance(payload, (dict, list, str)) and hasattr(payload, "result"):
|
||||
payload = payload.result
|
||||
if isinstance(payload, dict) and set(payload) == {"result"}:
|
||||
payload = payload["result"]
|
||||
if isinstance(payload, str):
|
||||
try:
|
||||
parsed = json.loads(payload)
|
||||
except ValueError:
|
||||
parsed = None
|
||||
if isinstance(parsed, (dict, list)):
|
||||
payload = parsed
|
||||
return payload
|
||||
|
||||
|
||||
_TEXT_ERROR_MARKERS = ('"ok": false', "'ok': false", '"ok":false', "ok=false")
|
||||
|
||||
|
||||
def _payload_error(payload: Any) -> str | None:
|
||||
"""Detect failure reported inside a tool payload instead of MCP isError."""
|
||||
if isinstance(payload, dict):
|
||||
if payload.get("ok") is False or payload.get("success") is False:
|
||||
return str(payload.get("error") or payload.get("message") or "ok=false")[:300]
|
||||
error = payload.get("error")
|
||||
if isinstance(error, str) and error.strip():
|
||||
return error[:300]
|
||||
return None
|
||||
if isinstance(payload, str):
|
||||
lowered = payload.lower()
|
||||
if lowered.startswith("error") or any(m in lowered for m in _TEXT_ERROR_MARKERS):
|
||||
return payload[:300]
|
||||
return None
|
||||
|
||||
|
||||
def wrapped_shapes(payload: dict[str, Any], **top_level: Any) -> tuple[dict[str, Any], ...]:
|
||||
"""Produce the two common consolidated-tool argument layouts for a payload:
|
||||
nested under ``data`` and flattened at the top level."""
|
||||
return (
|
||||
{**top_level, "data": payload},
|
||||
{**top_level, **payload},
|
||||
)
|
||||
|
||||
|
||||
_HANDLE_KEYS = ("handle", "entity_id", "entity_handle", "id", "object_id")
|
||||
_HANDLE_RE = re.compile(
|
||||
r'"(?:handle|entity_id|entity_handle|object_id)"\s*:\s*"?([0-9A-Fa-f*][\w*-]*)"?'
|
||||
)
|
||||
|
||||
|
||||
def extract_handle(payload: Any) -> str | None:
|
||||
"""Best-effort extraction of a created-entity handle from a black-box
|
||||
tool response (structured dict, list, or raw text)."""
|
||||
if isinstance(payload, dict):
|
||||
for key in _HANDLE_KEYS:
|
||||
value = payload.get(key)
|
||||
if isinstance(value, (str, int)) and str(value):
|
||||
return str(value)
|
||||
for value in payload.values():
|
||||
found = extract_handle(value)
|
||||
if found:
|
||||
return found
|
||||
return None
|
||||
if isinstance(payload, (list, tuple)):
|
||||
for item in payload:
|
||||
found = extract_handle(item)
|
||||
if found:
|
||||
return found
|
||||
return None
|
||||
if isinstance(payload, str):
|
||||
match = _HANDLE_RE.search(payload)
|
||||
return match.group(1) if match else None
|
||||
return None
|
||||
Reference in New Issue
Block a user