modelcontextprotocol / modelcontextprotocol/python-sdk

Duplicate `initialize` with changed parameters can overwrite `ServerSession.client_params`

Abierto
#2,605 4 comentarios 0 reacciones 0 asignados Ver en GitHub

Nadie ha tomado este issue todavía.

bug fix proposed P2 ready for work
Lenguaje dominante
Python
Estrellas
24.3k
Forks
4k
Merge medio
1 d 1 h
PR fusionados (30 d)
31

Descripción

Initial Checks
Description

After a normal MCP initialization flow, a Python SDK server accepts another initialize request on the same live connection/session and returns a normal initialize result.

I agree that the spec does not explicitly say every repeated initialize must be rejected. My concern is narrower: after the session has already moved into normal operation, a later same-session initialize with different parameters can update the server's stored client initialization parameters.

The MCP lifecycle documentation describes initialization as the first client/server interaction and says the operation phase uses the negotiated protocol version and capabilities. In that model, silently changing the stored client identity/capabilities/protocol version after operation has begun seems surprising unless the SDK intentionally supports same-session renegotiation.

  • src/mcp/server/session.py stores the client initialization request in ServerSession._client_params.
  • ServerSession.client_params exposes that value.
  • In both the stable release and the main snapshot I tested, the InitializeRequest handler assigns self._client_params = params.
  • ServerSession.check_client_capability() later reads self._client_params.capabilities, so a duplicate initialize with different capabilities can affect capability-gated server logic.
What I observed

Reproduced on stable v1.27.1 (77431ebe7dda9ed0c61451b22d3e7f8d981bc092):

  • stdio: a second initialize was accepted after notifications/initialized; a later ping still succeeded.
  • stateful Streamable HTTP: a second initialize was accepted after notifications/initialized; a later ping still succeeded.
  • stateless Streamable HTTP: a second initialize was accepted after notifications/initialized; a later ping still succeeded.

For stateful Streamable HTTP on v1.27.1, I added a local diagnostic tool that reads ctx.session.client_params before and after the duplicate initialize on the same session:

Duplicate initialize parameters Observed result
Same parameters No state change
Changed clientInfo and capabilities client_params.clientInfo and client_params.capabilities changed
Older protocolVersion (2024-11-05) client_params.protocolVersion and client_params.clientInfo changed

On the frozen main snapshot (161834d4aee2633c42d3976c8f8751b6c4d947d5), the same duplicate-initialize acceptance still reproduced. My diagnostic snapshot directly observed client_params.capabilities changing after a duplicate initialize with changed capabilities on both stdio and stateful Streamable HTTP.

Expected behavior

If same-session duplicate initialization is not intended, I would expect the server to reject the later initialize after normal operation has begun, for example with a JSON-RPC -32600 Invalid Request error.

If this tolerant behavior is intentional, I think it should be either:

  • a no-op once the session has already initialized, especially for changed parameters, or
  • documented as same-session renegotiation, including whether ServerSession.client_params, client capabilities, client identity, and protocol version are allowed to change.
Example Code
import json

try:
    # Current main branch naming.
    from mcp.server.mcpserver import Context, MCPServer
except ImportError:
    # Latest stable release naming.
    from mcp.server.fastmcp import Context, FastMCP as MCPServer


mcp = MCPServer("dup-init-repro")


def field(obj, *names):
    if obj is None:
        return None
    for name in names:
        if hasattr(obj, name):
            return getattr(obj, name)
    return None


def dump(value):
    if value is None:
        return None
    if hasattr(value, "model_dump"):
        return value.model_dump(mode="json", by_alias=True)
    return value


@mcp.tool()
async def session_snapshot(ctx: Context) -> str:
    params = ctx.session.client_params
    return json.dumps(
        {
            "protocolVersion": field(params, "protocolVersion", "protocol_version"),
            "clientInfo": dump(field(params, "clientInfo", "client_info")),
            "clientCapabilities": dump(field(params, "capabilities")),
        },
        separators=(",", ":"),
        default=str,
    )


if __name__ == "__main__":
    mcp.run(transport="streamable-http")

'''
Request sequence:

1. Start the server.
2. Send a normal `initialize`.
3. Send `notifications/initialized`.
4. Call the `session_snapshot` tool and record the returned JSON.
5. Send another `initialize` on the same live session, but change `clientInfo` and/or `capabilities`.
6. Call `session_snapshot` again on the same session.

On `v1.27.1`, the second snapshot reflects the changed `clientInfo` / `capabilities`. On the `main` snapshot I tested, the second snapshot still reflects changed `capabilities`.
'''
Python & MCP Python SDK
- Python: `3.12.3`
- MCP Python SDK: `mcp==1.27.1` (`v1.27.1`, `77431ebe7dda9ed0c61451b22d3e7f8d981bc092`)
- Also reproduced on `main` snapshot `161834d4aee2633c42d3976c8f8751b6c4d947d5` (`v1.25.0-159-g161834d4`, pinned 2026-05-11)

Guía de contribución

Abrir la guía de contribución

Primeros pasos

  1. Lee el issue completo y luego la guía de contribución del proyecto.
  2. Comenta en el issue que vas a ocuparte — evita que dos personas hagan lo mismo.
  3. Haz un fork del repositorio y trabaja en una rama.
  4. Abre un pull request que haga referencia al número del issue.

Línea de trabajo

Comienza con src/mcp/server/session.py, especialmente con el controlador de InitializeRequest y ServerSession.client_params. Reproduce la secuencia de inicialización duplicada descrita en el issue tanto en stdio como en Streamable HTTP con estado o sin estado, y después inspecciona cómo se leen los parámetros del cliente. Se considera terminado cuando el comportamiento elegido para la inicialización duplicada es coherente y no se sobrescriben parámetros del cliente de forma no intencionada.

Escrito por el modelo de indexación a partir del texto del issue.

Evaluación

Stack tecnológico
python
Área
api, backend
Tipo de issue
Error
Dificultad
4/5
Tiempo estimado
3-5 días
Estado de actividad
Activo
Claridad
Bastante claro
Aptitud para principiantes
48/100

Recibe los nuevos issues en tu correo

Un resumen breve de issues de GitHub para principiantes.