[v2] JSON-RPC error responses leave client OpenTelemetry spans UNSET

[HarperZ9]

Initial Checks

• I confirm that I'm using the latest version of MCP Python SDK
• I confirm that I searched for my issue in https://github.com/modelcontextprotocol/python-sdk/issues before opening this issue

Description

At current main (11934c90aeff5e1e68aee223edd00c5d1fce1d5c), JSON-RPC error responses handled by JSONRPCDispatcher leave the corresponding client OpenTelemetry span looking like a normal exit:

• span status: UNSET
• error.type: absent
• rpc.response.status_code: absent
• exception events: none

The request still fails correctly with MCPError; this report is limited to client-side telemetry correctness.

The ordering appears to explain the result: send_raw_request() receives outcome inside the client span, exits the span, and only afterward converts ErrorData into MCPError at lines 432–433. The span context manager therefore observes no exception.

This differs from the current server OTel middleware, which records the error attributes and status, and from the design assumption documented in #2381 that a propagated MCPError would be recorded by the client span.

I ran the same in-memory success/error pair five times. Every run produced:

METHOD=resources/list CONTROL=UNSET RED_CODE=-32602 RED_MESSAGE='forced failure' RED_STATUS=UNSET ERROR_TYPE=None RPC_STATUS=None EVENTS=0

Expected: the failed client span should be distinguishable from the successful control and carry the JSON-RPC error code/status consistently with the SDK's server-side instrumentation.

Scope and claim boundary: this concerns JSON-RPC/stream-backed dispatcher calls, not direct in-process dispatch generally. The reproducer is entirely in-memory. I have not tested or claimed anything about a deployed transport, collector, or production frequency.

Closest adjacent work found: #2132 targets the removed BaseSession architecture and predates the current dispatcher; #2854 adds GenAI attributes but does not handle JSON-RPC errors. I did not find an issue or current PR covering this ordering on the present dispatcher.

Disclosure: I used AI assistance to help reproduce and analyze this issue. I reviewed the source, reran the evidence independently, and take responsibility for the report.

Example Code

from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import SimpleSpanProcessor
from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter
from opentelemetry.trace import SpanKind

exporter = InMemorySpanExporter()
provider = TracerProvider()
provider.add_span_processor(SimpleSpanProcessor(exporter))
trace.set_tracer_provider(provider)

import anyio
from mcp.client.client import Client
from mcp.server.lowlevel.server import Server
from mcp.shared.exceptions import MCPError
from mcp_types import INVALID_PARAMS, ListResourcesResult

state = {"fail": False}

async def list_resources(ctx, params):
    if state["fail"]:
        raise MCPError(INVALID_PARAMS, "forced failure")
    return ListResourcesResult(resources=[])

async def main():
    server = Server(name="otel-audit", version="0.0.0", on_list_resources=list_resources)
    async with Client(server, mode="legacy", cache=None) as client:
        state["fail"] = False
        exporter.clear()
        await client.session.list_resources()
        [control] = [s for s in exporter.get_finished_spans() if s.kind == SpanKind.CLIENT]

        state["fail"] = True
        exporter.clear()
        try:
            await client.session.list_resources()
        except MCPError as exc:
            assert exc.error.code == INVALID_PARAMS
        else:
            raise AssertionError("expected MCPError")

        [failed] = [s for s in exporter.get_finished_spans() if s.kind == SpanKind.CLIENT]
        print(
            "control=", control.status.status_code.name,
            "failed=", failed.status.status_code.name,
            "error.type=", failed.attributes.get("error.type"),
            "rpc.response.status_code=", failed.attributes.get("rpc.response.status_code"),
            "events=", len(failed.events),
        )
        assert control.name == failed.name == "MCP send resources/list"
        assert failed.status.status_code.name == "UNSET"
        assert failed.attributes.get("error.type") is None
        assert failed.attributes.get("rpc.response.status_code") is None
        assert len(failed.events) == 0

anyio.run(main)

Python & MCP Python SDK

Python: 3.12.10
mcp: 2.0.0b2.dev33+11934c90
opentelemetry-sdk: 1.39.1
OS: Windows 11

Clean checkout command:
uv run --python 3.12 --frozen --no-default-groups --with opentelemetry-sdk==1.39.1 python repro.py

[Ranga-Prasath-22]

Hi, I'd like to work on this fix.

Root cause: In send_raw_request(), the otel_span context manager exits before ErrorData is converted to MCPError, so the span never observes an error — it stays UNSET with no error attributes.

Proposed fix: Inside the span, after receiving outcome, check for ErrorData and set error attributes (error.type,
pc.response.status_code, StatusCode.ERROR) on the span before exiting. This mirrors the server-side pattern in _otel.py:35-60.

The change is ~5 lines in send_raw_request() plus a test using InMemorySpanExporter (adapting the reproducer from the issue).

Happy to take this on if no one else is assigned. Thanks!