Thanks for this fix.

The stdio reconfigure portion of this PR overlaps with #1259, which addresses the same Windows encoding problem using sys.stdin.reconfigure(encoding="utf-8"). Both approaches work. The reconfigure() idiom is what the other Windows stdio PRs in flight (#1280, #1282) are using, so we are likely to take #1259 to keep the codebase consistent.

That said, the ensure_ascii=False additions on the two json.dumps calls in this PR are independently valuable and not present in #1259. Without them, non-ASCII content in tool results is emitted as \uXXXX escapes, which is verbose and can trip downstream MCP clients that do not unescape correctly.

Would you be willing to open a separate PR with just the ensure_ascii=False change, plus a small test that round-trips a Chinese string through handle_request so the fix has a regression guard? That change can land on its own merits and stay decoupled from the stdio-reconfigure work.

Thanks for this fix.

The stdio reconfigure portion of this PR overlaps with #1259, which addresses the same Windows encoding problem using sys.stdin.reconfigure(encoding="utf-8"). Both approaches work. The reconfigure() idiom is what the other Windows stdio PRs in flight (#1280, #1282) are using, so we are likely to take #1259 to keep the codebase consistent.

That said, the ensure_ascii=False additions on the two json.dumps calls in this PR are independently valuable and not present in #1259. Without them, non-ASCII content in tool results is emitted as \uXXXX escapes, which is verbose and can trip downstream MCP clients that do not unescape correctly.

Would you be willing to open a separate PR with just the ensure_ascii=False change, plus a small test that round-trips a Chinese string through handle_request so the fix has a regression guard? That change can land on its own merits and stay decoupled from the stdio-reconfigure work.

Thanks for the clear and helpful feedback!

I've split the changes as you suggested into one separate PR:

• PR fix: MCP server JSON output ensure_ascii=False for non-ASCII support #1293: ensure_ascii=False for all platforms

I'll close this original PR in favor of the split one, to keep things clean and decoupled as you recommended.