[Bug]: MCP SSE transport rejects concurrent requests on connect with "Received request before initialization was complete" (-32602)

crawl4ai version

0.9.2 (Docker server, unclecode/crawl4ai:latest)

Expected Behavior

When an MCP client (such as OpenCode, Claude Code, or multi-agent runtimes) connects over SSE (/mcp/sse) and issues tool calls, inbound requests arriving while initialization is completing should either be briefly buffered/queued or await initialization before processing, similar to how /mcp/ws gates incoming traffic.

Current Behavior

When a client establishes an SSE connection and dispatches tool calls concurrently (e.g. parallel fan-out upon connecting), the requests hit /mcp/messages/?session_id=... before the initialize -> notifications/initialized handshake has completed inside mcp.server.session.ServerSession.

This causes mcp/server/session.py to raise:

WARNING - Failed to validate request: Received request before initialization was complete

The client receives:

{"jsonrpc": "2.0", "id": 1, "error": {"code": -32602, "message": "Invalid request parameters"}}

The error code -32602 misleads clients into reporting that tool call parameters are malformed, when in reality the arguments were never validated because the transport rejected the uninitialized session state.

Context & Contrast with WebSocket

In /app/mcp_bridge.py, the WebSocket handler (/mcp/ws) explicitly mitigates this race condition by synchronizing on initialization:

# /app/mcp_bridge.py (WebSocket handler)
first = adapter.validate_python(await ws.receive_json())
await c2s_send.send(first)
await init_done.wait()          # <-- explicitly waits for server readiness
while True:
    data = await ws.receive_json()
    await c2s_send.send(adapter.validate_python(data))

However, the SSE transport (/mcp/sse + /mcp/messages) directly passes through Starlette to SseServerTransport:

sse = SseServerTransport(f"{base}/messages/")

Without any gating on incoming POST requests during the handshake window, any concurrent requests arriving in the first ~100–500ms get rejected.

Upstream References

This is a known race condition in the underlying Python SDK:

• modelcontextprotocol/python-sdk #2583: "Race condition between SSE session init and tools/call rejection (-32602)"
• modelcontextprotocol/python-sdk #423: "MCP SSE Server: Received request before initialization was complete"

Suggested Fix

A minimal fix in mcp_bridge.py for the SSE transport:

1. Wrap or intercept incoming POST messages to /mcp/messages to await session initialization (with a small timeout, e.g. 2.0s) before feeding them to the read stream writer, OR
2. Provide an init_done event per SSE session matching the WebSocket pattern.

Steps to Reproduce

1. Connect an MCP client over SSE (/mcp/sse).
2. Immediately upon receiving the endpoint SSE event, dispatch 3+ parallel tools/call requests in under 50ms alongside initialize.
3. Observe Received request before initialization was complete in Docker logs and -32602 returned to the client.

Environment

• Docker container unclecode/crawl4ai:latest (0.9.2)
• Python 3.12 (inside container)
• Client: OpenCode / Claude Code via SSE transport

Start with /app/mcp_bridge.py, comparing the /mcp/ws initialization gate with the SSE /mcp/sse and /mcp/messages flow through SseServerTransport. Reproduce with parallel tools/call requests during the initialize handshake and inspect the resulting session behavior. Done means concurrent SSE requests no longer fail with the pre-initialization -32602 response.