MCP stdio server crashes on Windows with UnicodeEncodeError: 'charmap' codec can't encode character '\u2713' #886

Activity

1. svetkis

   added

   bugSomething isn't workingSomething isn't working

   on Jun 18, 2026
2. maxim-ultra commented on Jun 18, 2026

   

   maxim-ultra

   on Jun 18, 2026

   More actions

   From the public issue context, a small diagnostic structure that may help keep this actionable:

   • exact MCP client and version
   • exact server launch command or config entry
   • whether transport is stdio, SSE, or streamable HTTP
   • first stderr/stdout line before the failure
   • whether the same command works outside the MCP client
   • expected tool/list response vs actual error

   For a durable fix, I would also add a tiny regression fixture around the failing config shape so this does not reappear after a release. That seems especially relevant for this issue: MCP stdio server crashes on Windows with UnicodeEncodeError: 'charmap' codec can't encode character '\u2713'
3. svetkis commented on Jun 18, 2026

   

   svetkis

   on Jun 18, 2026

   Author

   More actions

   Updated the issue with requested details:

   • Kimi Code version: 0.5.10
   • Platform: Kimi Code VS Code extension, Kimi Allegretto subscription
   • Model: kimi-k2.7
   • Added Context7 3.2.1 as second affected server
   • Added "first log line before crash" and "works outside Kimi Code" sections

   Bug confirmed in Kimi Code's stdio UTF-8 handling on Windows. Two independent servers (Python and Node.js) fail with identical error. Same servers work in Codex and manually from console.
4. maxim-ultra commented on Jun 18, 2026

   

   maxim-ultra

   on Jun 18, 2026

   More actions

   From the public issue context and the reporter's public update, thanks for adding the details. Since both Serena and Context7 work outside Kimi Code, I would narrow the next public-only check to Kimi's Windows stdio launcher boundary rather than either MCP server.

   A useful regression shape would be:

   • fake stdio MCP server emits a Unicode checkmark in a tools/list response
   • launch it under a Windows non-UTF-8 locale/codepage fixture
   • assert Kimi reads protocol bytes as UTF-8 at the JSON-RPC boundary instead of using the process locale text codec
   • capture the effective child-process env for PYTHONUTF8, PYTHONIOENCODING, and any shell/codepage setup
   • make the failure report the launcher encoding decision, not just the server stderr line

   No private logs, credentials, or account data are needed. That should separate server emitted invalid output from client launcher decoded valid UTF-8 through cp1252/charmap.
5. svetkis commented on Jun 18, 2026

   

   svetkis

   on Jun 18, 2026

   Last edited by svetkis

   Author

   More actions

   Following up on the regression shape suggested above.

   Narrowed boundary: The crash is in Kimi Code's Windows stdio launcher, not in Serena or Context7. When Kimi spawns a Python MCP stdio server on Windows, the child inherits the console codepage (cp1252 / "charmap") for stdout/stderr. Emitting U+2713 (✓) triggers:

   UnicodeEncodeError: 'charmap' codec can't encode character '\u2713' in position 0: character maps to <undefined>
   

   Regression reproduction:

   I created a minimal Python fake MCP stdio server that:

   • emits a U+2713 checkmark in its tools/list response;
   • logs the effective stdout.encoding, stderr.encoding, PYTHONUTF8, and PYTHONIOENCODING of the child process.

   Launching it without Kimi's UTF-8 setup on a Windows cp1252 console gives:

   stdout.encoding=cp1252, stderr.encoding=cp1252, PYTHONUTF8=<unset>
   \u2713 Fake Unicode MCP server running on stdio
   

   With PYTHONUTF8=1 and PYTHONIOENCODING=utf-8 injected by the launcher:

   stdout.encoding=utf-8, stderr.encoding=utf-8, PYTHONUTF8=1
   ✓ Fake Unicode MCP server running on stdio
   

   This separates "server emitted invalid output" from "client launcher decoded valid UTF-8 through cp1252/charmap" — the failure now reports the launcher encoding decision (cp1252) explicitly.

   Fix: PR #899 defaults the child env to PYTHONUTF8=1 and PYTHONIOENCODING=utf-8 in mergeStdioEnv(). It also adds:

   • a unicode-stdio-server.mjs fixture that emits U+2713 in tools/list and tool results;
   • integration tests asserting UTF-8 round-trip at the JSON-RPC boundary and env injection;
   • unit tests for mergeStdioEnv defaults/override semantics.

   pnpm exec vitest run test/mcp/client-stdio.test.ts → 18/18 passed.

   I also verified the fix end-to-end with Serena: after the patch Serena activates projects successfully instead of crashing on the checkmark character.