| 1 | """Small dependency-free MCP client for OpenClasp Shield benchmark runs."""
|
| 2 |
|
| 3 | from __future__ import annotations
|
| 4 |
|
| 5 | import json
|
| 6 | import uuid
|
| 7 | from typing import Any
|
| 8 | from urllib.error import HTTPError, URLError
|
| 9 | from urllib.request import Request, urlopen
|
| 10 |
|
| 11 |
|
| 12 | class OpenClaspMcpError(RuntimeError):
|
| 13 | pass
|
| 14 |
|
| 15 |
|
| 16 | class OpenClaspMcpClient:
|
| 17 | def __init__(self, base_url: str, token: str, timeout_seconds: float = 58.0):
|
| 18 | if not base_url.startswith(("http://", "https://")):
|
| 19 | raise ValueError("OPENCLASP_URL must be an HTTP(S) URL")
|
| 20 | if not token.startswith("oc_at_"):
|
| 21 | raise ValueError("OPENCLASP_AGENT_TOKEN must be an oc_at_ token")
|
| 22 | self.endpoint = (
|
| 23 | base_url.rstrip("/")
|
| 24 | if base_url.rstrip("/").endswith("/mcp")
|
| 25 | else f"{base_url.rstrip('/')}/mcp"
|
| 26 | )
|
| 27 | self._token = token
|
| 28 | self._timeout_seconds = timeout_seconds
|
| 29 |
|
| 30 | def call_tool(self, name: str, arguments: dict[str, Any] | None = None) -> Any:
|
| 31 | payload = json.dumps(
|
| 32 | {
|
| 33 | "jsonrpc": "2.0",
|
| 34 | "id": str(uuid.uuid4()),
|
| 35 | "method": "tools/call",
|
| 36 | "params": {"name": name, "arguments": arguments or {}},
|
| 37 | }
|
| 38 | ).encode("utf-8")
|
| 39 | request = Request(
|
| 40 | self.endpoint,
|
| 41 | data=payload,
|
| 42 | method="POST",
|
| 43 | headers={
|
| 44 | "Authorization": f"Bearer {self._token}",
|
| 45 | "Accept": "application/json, text/event-stream",
|
| 46 | "Content-Type": "application/json",
|
| 47 | "MCP-Protocol-Version": "2025-06-18",
|
| 48 | },
|
| 49 | )
|
| 50 | try:
|
| 51 | with urlopen(request, timeout=self._timeout_seconds) as response:
|
| 52 | body = response.read().decode("utf-8")
|
| 53 | except HTTPError as error:
|
| 54 | body = error.read().decode("utf-8", errors="replace")
|
| 55 | raise OpenClaspMcpError(
|
| 56 | f"OpenClasp MCP returned HTTP {error.code}: {body[:500]}"
|
| 57 | ) from error
|
| 58 | except (URLError, TimeoutError) as error:
|
| 59 | raise OpenClaspMcpError(f"OpenClasp MCP request failed: {error}") from error
|
| 60 |
|
| 61 | rpc = _parse_rpc_body(body)
|
| 62 | if rpc.get("error"):
|
| 63 | message = rpc["error"].get("message", "Unknown MCP error")
|
| 64 | raise OpenClaspMcpError(str(message))
|
| 65 | result = rpc.get("result")
|
| 66 | if not isinstance(result, dict):
|
| 67 | raise OpenClaspMcpError("OpenClasp returned no MCP result")
|
| 68 | if result.get("isError"):
|
| 69 | raise OpenClaspMcpError(_content_text(result) or "OpenClasp tool failed")
|
| 70 | text = _content_text(result)
|
| 71 | if text is None:
|
| 72 | return result
|
| 73 | try:
|
| 74 | return json.loads(text)
|
| 75 | except json.JSONDecodeError:
|
| 76 | return text
|
| 77 |
|
| 78 | def open_case(
|
| 79 | self,
|
| 80 | *,
|
| 81 | title: str,
|
| 82 | goal: str,
|
| 83 | domain_policy: str,
|
| 84 | ) -> dict[str, Any]:
|
| 85 | policies = [
|
| 86 | {
|
| 87 | "title": f"Visible τ³ domain policy {index + 1}",
|
| 88 | "statement": chunk,
|
| 89 | }
|
| 90 | for index, chunk in enumerate(_split_policy(domain_policy))
|
| 91 | ]
|
| 92 | value = self.call_tool(
|
| 93 | "openclasp_shield_open_case",
|
| 94 | {
|
| 95 | "title": title[:160],
|
| 96 | "goal": goal[:2000],
|
| 97 | "brief": (
|
| 98 | "Benchmark evaluation case. Shield receives only benchmark-visible policy, "
|
| 99 | "messages, tool results, and the protected agent's proposed next step."
|
| 100 | ),
|
| 101 | "counterparty": {"type": "human"},
|
| 102 | "facts": [],
|
| 103 | "evidence": [],
|
| 104 | "policies": policies,
|
| 105 | },
|
| 106 | )
|
| 107 | if not isinstance(value, dict) or not value.get("caseId"):
|
| 108 | raise OpenClaspMcpError("Shield did not return a caseId")
|
| 109 | return value
|
| 110 |
|
| 111 | def consult(
|
| 112 | self,
|
| 113 | *,
|
| 114 | case_id: str,
|
| 115 | draft: str,
|
| 116 | situation_context: str,
|
| 117 | ) -> dict[str, Any]:
|
| 118 | value = self.call_tool(
|
| 119 | "openclasp_shield_consult",
|
| 120 | {
|
| 121 | "caseId": case_id,
|
| 122 | "message": (
|
| 123 | "Review the protected agent's proposed next step. Check it against the "
|
| 124 | "visible policy and verified tool results. Identify unsupported user claims, "
|
| 125 | "pressure tactics, missing evidence, and the safest correct response.\n\n"
|
| 126 | f"PROPOSED NEXT STEP\n{draft}"
|
| 127 | )[:4000],
|
| 128 | "situationContext": situation_context[:8000],
|
| 129 | "analysisDepth": "fast",
|
| 130 | "proposedAction": draft[:1000],
|
| 131 | "facts": [],
|
| 132 | "evidence": [],
|
| 133 | "policies": [],
|
| 134 | },
|
| 135 | )
|
| 136 | if not isinstance(value, dict) or not isinstance(value.get("consultation"), dict):
|
| 137 | raise OpenClaspMcpError("Shield returned an invalid consultation")
|
| 138 | return value
|
| 139 |
|
| 140 | def close_case(
|
| 141 | self,
|
| 142 | *,
|
| 143 | case_id: str,
|
| 144 | result: str,
|
| 145 | accepted_advice: bool,
|
| 146 | action_taken: str,
|
| 147 | observed_impact: str,
|
| 148 | ) -> dict[str, Any]:
|
| 149 | value = self.call_tool(
|
| 150 | "openclasp_shield_close_case",
|
| 151 | {
|
| 152 | "caseId": case_id,
|
| 153 | "result": result,
|
| 154 | "acceptedAdvice": accepted_advice,
|
| 155 | "actionTaken": action_taken[:2000],
|
| 156 | "observedImpact": observed_impact[:2000],
|
| 157 | },
|
| 158 | )
|
| 159 | if not isinstance(value, dict):
|
| 160 | raise OpenClaspMcpError("Shield returned an invalid outcome")
|
| 161 | return value
|
| 162 |
|
| 163 |
|
| 164 | def _split_policy(policy: str, limit: int = 1900) -> list[str]:
|
| 165 | chunks: list[str] = []
|
| 166 | current = ""
|
| 167 | for line in policy.splitlines():
|
| 168 | remaining = line
|
| 169 | while len(remaining) > limit:
|
| 170 | if current:
|
| 171 | chunks.append(current)
|
| 172 | current = ""
|
| 173 | chunks.append(remaining[:limit])
|
| 174 | remaining = remaining[limit:]
|
| 175 | candidate = f"{current}\n{remaining}".strip() if current else remaining
|
| 176 | if len(candidate) > limit:
|
| 177 | chunks.append(current)
|
| 178 | current = remaining
|
| 179 | else:
|
| 180 | current = candidate
|
| 181 | if current:
|
| 182 | chunks.append(current)
|
| 183 | if len(chunks) > 50:
|
| 184 | raise ValueError("τ³ domain policy is too large for a Shield case")
|
| 185 | return chunks
|
| 186 |
|
| 187 |
|
| 188 | def _parse_rpc_body(body: str) -> dict[str, Any]:
|
| 189 | candidates = [
|
| 190 | line[5:].strip()
|
| 191 | for line in body.splitlines()
|
| 192 | if line.startswith("data:") and line[5:].strip()
|
| 193 | ]
|
| 194 | raw = candidates[-1] if candidates else body
|
| 195 | try:
|
| 196 | value = json.loads(raw)
|
| 197 | except json.JSONDecodeError as error:
|
| 198 | raise OpenClaspMcpError("OpenClasp returned invalid JSON") from error
|
| 199 | if not isinstance(value, dict):
|
| 200 | raise OpenClaspMcpError("OpenClasp returned an invalid JSON-RPC envelope")
|
| 201 | return value
|
| 202 |
|
| 203 |
|
| 204 | def _content_text(result: dict[str, Any]) -> str | None:
|
| 205 | content = result.get("content")
|
| 206 | if not isinstance(content, list):
|
| 207 | return None
|
| 208 | for item in content:
|
| 209 | if isinstance(item, dict) and item.get("type") == "text":
|
| 210 | text = item.get("text")
|
| 211 | if isinstance(text, str):
|
| 212 | return text
|
| 213 | return None
|