MCPServer completion handler returning more than 100 values fails on 2026-07-28 sessions and violates the 100-item limit on 2025-11-25 sessions

從 mcp/server/mcpserver.py 中的 MCPServer.completion 包裝器（約 759-766 行）開始，它原封不動地傳遞處理器的 Completion；100 項的上限位於 mcp_types/_v2026_07_28，並在 runner.py:381 強制執行，因此包裝器必須在驗證前截斷為 100 個值並設定 total/has_more，同時保留處理器設定的值。與 TypeScript SDK 的 createCompletionResult 比較。完成的標準是：在 tests/server/mcpserver/ 底下使用記憶體 Client 的回歸測試，對 mode="auto"（2026-07-28）與 mode="legacy"（2025-11-25）皆通過，並回傳 100 個值與 has_more=True。

Initial Checks

• I confirm that I'm using the newest release of my line (the latest 2.x, or the latest 1.x if I'm still on v1)
• I confirm that I searched for my issue in https://github.com/modelcontextprotocol/python-sdk/issues before opening this issue

Release line

2.x (current stable)

Description

An @mcp.completion() handler that returns more than 100 values behaves differently depending on the negotiated protocol version, and both behaviours are wrong:

• 2026-07-28 session: the whole completion/complete request fails with -32603 "Handler returned an invalid result". The user gets no suggestions at all.
• 2025-11-25 session: all values go out on the wire. The spec says values "Must not exceed 100 items", and total / hasMore are left unset, so the client can't tell the list was too long.

It's easy to hit with the filter-by-prefix pattern from docs/servers/completions.md. When the user hasn't typed anything yet (argument.value == ""), every candidate matches, so any handler over a list of more than 100 options (countries, time zones, repos, file names, and so on) fails as soon as the field is focused.

The 100-item cap is only enforced by the 2026-07-28 wire model (Completion.values: Annotated[list[str], Field(max_length=100)] in mcp_types/_v2026_07_28). runner.py:381 checks results against that model and turns the ValidationError into a generic INTERNAL_ERROR. The 2025-11-25 wire model has no such constraint, so nothing stops the oversize list there. The MCPServer.completion wrapper (server.py:759-766) passes the handler's Completion through as-is.

For comparison, the TypeScript SDK's McpServer truncates to 100 and fills the pagination hints (createCompletionResult: values.slice(0, 100), total: suggestions.length, hasMore: suggestions.length > 100).

Expected: the same handler works on both protocol versions. The client receives the first 100 values with total / hasMore telling it more exist, as Completion's own docstring ("total … can exceed the number of values actually sent") and the docs' total= / has_more= note describe.

Server log on the 2026-07-28 session:

ERROR    handler for 'completion/complete' returned an invalid result
pydantic_core._pydantic_core.ValidationError: 1 validation error for CompleteResult
completion.values
  List should have at most 100 items after validation, not 150 [type=too_long, ...]

I reported this and would like to fix it. Proposed approach: in the MCPServer.completion wrapper, when the handler returns more than 100 values, send the first 100 and set total (to the full count) and has_more=True. Values the handler set explicitly win. Handlers returning ≤100 values are unchanged. The lowlevel Server stays as it is, since a lowlevel handler builds the CompleteResult itself. I'd add a regression test in tests/server/mcpserver/ using an in-memory Client, covering both mode="auto" (2026-07-28) and mode="legacy" (2025-11-25). It's roughly 10 lines of source. If you'd prefer a different behaviour (for example, raising a clear error at the wrapper instead of truncating), I'm happy to do that.

I used an AI assistant to help narrow this down; I ran the reproduction myself on 2.3.0 and on main.

Example Code

import anyio
from mcp_types import Completion, PromptReference
from mcp.client import Client
from mcp.server.mcpserver import MCPServer

COUNTRIES = [f"country-{i:03}" for i in range(150)]

mcp = MCPServer("demo")


@mcp.prompt()
def travel(country: str) -> str:
    return f"Plan a trip to {country}"


@mcp.completion()
async def complete(ref, argument, context):
    # Filter by the typed prefix, as in docs/servers/completions.md; an empty prefix matches everything.
    return Completion(values=[c for c in COUNTRIES if c.startswith(argument.value)])


async def main() -> None:
    for mode in ("auto", "legacy"):
        async with Client(mcp, mode=mode) as client:
            try:
                result = await client.complete(
                    ref=PromptReference(type="ref/prompt", name="travel"),
                    argument={"name": "country", "value": ""},
                )
                c = result.completion
                print(f"{client.protocol_version}: {len(c.values)} values, total={c.total}, has_more={c.has_more}")
            except Exception as e:
                print(f"{client.protocol_version}: {type(e).__name__}: {e}")


anyio.run(main)

Output:

2026-07-28: MCPError: Handler returned an invalid result
2025-11-25: 150 values, total=None, has_more=None

Python & MCP Python SDK

Python 3.12.12, mcp 2.3.0, pydantic 2.13.5 (PyPI release)
Python 3.14.0, mcp main @ 91941ed, pydantic 2.12.5
Windows 11