[bot] Anthropic: MCP connector tool calls (`mcp_tool_use`/`mcp_tool_result`) are silently dropped from tool spans

问题位于 py/src/braintrust/integrations/anthropic/tracing.py。首先阅读第 1455 行附近的 _log_server_tool_spans 函数以及 _SERVER_TOOL_USE_TYPE 和 _is_server_tool_result_type 的定义。修复方法是扩展调用端检查，使其也匹配 "mcp_tool_use"。确保 mcp_tool_result 块与其调用正确配对。运行 Anthropic 集成的现有测试以验证修复。

Summary

Anthropic's MCP connector feature (tools=[{"type": "mcp_toolset", ...}] on client.messages.create(), gated by the mcp-client-2025-11-20 beta header) returns mcp_tool_use and mcp_tool_result content blocks when Claude calls a remote MCP server's tools. The repo's Anthropic tracing does not recognize mcp_tool_use as a tool call at all, so:

• MCP tool calls are never captured into a span (no input, no tool name, no call metadata).
• MCP tool results are still detected (because the result-type check is suffix-based) but are logged as orphaned spans with no call context, since the corresponding call was never registered.

This is a correctness gap, not just missing-coverage: half of each MCP tool exchange is dropped and the other half is logged incompletely/incorrectly.

What is missing

In py/src/braintrust/integrations/anthropic/tracing.py:

_SERVER_TOOL_USE_TYPE = "server_tool_use"          # line 1337

def _is_server_tool_result_type(item_type: Any) -> bool:   # line 1340
    return isinstance(item_type, str) and item_type.endswith("_tool_result") and item_type != "tool_result"

_log_server_tool_spans (line 1455) pairs calls and results by walking response content:

item_type = item.get("type")
if item_type == _SERVER_TOOL_USE_TYPE:      # line 1470 — only matches "server_tool_use"
    ...
    continue

if not _is_server_tool_result_type(item_type):   # line 1481
    continue

• The call-side check only matches the literal string "server_tool_use" (used for built-in server tools like web search / code execution). It does not match "mcp_tool_use", so mcp_tool_use blocks fall through the loop entirely and are never added to calls_by_id.
• The result-side check (_is_server_tool_result_type) matches anything ending in _tool_result except the literal tool_result, so mcp_tool_result does pass this check — but since no matching call was ever registered, it's appended as (None, item), producing a tool span with output only and no input/tool name.

Separately, _MANAGED_AGENTS_CALL_TYPES (line 905) does include "agent.mcp_tool_use" — but that's the distinct, agent.-prefixed type used by the Managed Agents API (client.beta.agents/client.beta.sessions), not the plain mcp_tool_use/mcp_tool_result types returned by the standard Messages API's MCP connector.

Braintrust docs status: unclear / not_found

The Anthropic integration page mentions mcp_servers exactly once, as one of many request parameters captured in span metadata for the Go SDK's request-param list. It does not mention mcp_toolset, mcp_tool_use, or mcp_tool_result anywhere, and does not document any tool-span behavior specific to the MCP connector for any language. The only documented span-splitting for server-side tools is generic ("server-side tool calls ... appear as child tool spans"), described for the Java SDK, and is not confirmed to apply to MCP connector blocks specifically.

Upstream sources

• Anthropic MCP connector docs: https://platform.claude.com/docs/en/agents-and-tools/mcp-connector — confirms the mcp_toolset tool type, the mcp-client-2025-11-20 beta header, and the exact response content block types "mcp_tool_use" and "mcp_tool_result" (example response blocks shown verbatim in the "How MCP connector tool calls work" section).

Local repo files inspected

• py/src/braintrust/integrations/anthropic/tracing.py (full file, 1627 lines) — specifically _SERVER_TOOL_USE_TYPE (line 1337), _is_server_tool_result_type (line 1340), _log_server_tool_spans (line 1455), _MANAGED_AGENTS_CALL_TYPES (line 905)
• py/src/braintrust/integrations/anthropic/integration.py
• py/src/braintrust/integrations/anthropic/patchers.py