Soften short-intent gates and add tool-result refs plus MCP wire invalidate.

Prefer report tools without hard-hiding CLI, keep soft CLI budgets, store compact+result_ref for the model, and clear frozen wire after MCP sync.

Co-authored-by: Cursor <cursoragent@cursor.com>

runtime/tools/mcp/sync_tools.py

@@ -85,6 +85,12 @@ def sync_mcp_server_tools(store: Any, row: dict[str, Any]) -> dict[str, Any] | N
         store.replace_mcp_server_tools(server_id=sid, tools=tools)
         detail = {"synced_tools": len(tools), "compat_mode": "bailian_webparser"}
         store.set_mcp_server_health(server_id=sid, status="ok", detail=detail)
+        try:
+            from runtime.direct_loop import invalidate_tool_wire_cache
+
+            invalidate_tool_wire_cache(reason=f"mcp_sync:{sid}")
+        except Exception:
+            pass
         return {"server_id": sid, "ok": True, "health": detail, "tools_synced": len(tools)}
 
     rt = mcp_runtime_for_row(row, store=store)
@@ -128,6 +134,12 @@ def sync_mcp_server_tools(store: Any, row: dict[str, Any]) -> dict[str, Any] | N
                 "health": health,
             },
         )
+        try:
+            from runtime.direct_loop import invalidate_tool_wire_cache
+
+            invalidate_tool_wire_cache(reason=f"mcp_sync:{sid}")
+        except Exception:
+            pass
         return {
             "server_id": sid,
             "ok": True,

runtime/tools/public/fetch_tool_result_tool.py

@@ -0,0 +1,56 @@
+from __future__ import annotations
+
+from typing import Any
+
+from runtime.chat.tool_invocation_context import current_tool_lane_sessions
+from runtime.chat.tool_result_store import load_tool_result_blob
+from runtime.tools.base import ToolSpec
+
+
+def fetch_tool_result_tool() -> ToolSpec:
+    def handler(args: dict[str, Any]) -> dict[str, Any]:
+        ref = str(args.get("result_ref") or "").strip()
+        if not ref:
+            return {"ok": False, "error_code": "result_ref_required", "error": "result_ref_required"}
+        owner, sid = current_tool_lane_sessions()
+        session_id = str(sid or owner or "").strip()
+        max_chars = args.get("max_chars")
+        try:
+            max_i = int(max_chars) if max_chars is not None else None
+        except Exception:
+            max_i = None
+        return load_tool_result_blob(ref, session_id=session_id, max_chars=max_i)
+
+    return ToolSpec(
+        name="fetch_tool_result",
+        description=(
+            "Fetch a previously truncated/guarded tool result by result_ref. "
+            "Use when a tool payload includes result_ref / _truncated_for_llm / _tool_result_guarded "
+            "and you need more detail than the compact preview."
+        ),
+        parameters={
+            "type": "object",
+            "properties": {
+                "result_ref": {
+                    "type": "string",
+                    "description": "Opaque ref from a prior tool message (e.g. tr:…).",
+                },
+                "max_chars": {
+                    "type": "integer",
+                    "minimum": 4000,
+                    "maximum": 500000,
+                    "description": "Optional cap for returned JSON size (default ~120k).",
+                },
+            },
+            "required": ["result_ref"],
+            "additionalProperties": False,
+        },
+        handler=handler,
+        tags=frozenset({"system", "read", "tool_result"}),
+        read_only=True,
+        risk_level="low",
+        timeout_s=8.0,
+    )
+
+
+__all__ = ["fetch_tool_result_tool"]