fix(cli): redirect stderr before MCP loading to prevent terminal pollution

CHANGELOG.md

@@ -11,6 +11,7 @@ Only write entries that are worth mentioning to users.
 
 ## Unreleased
 
+- Core: Fix MCP server stderr pollution — stderr redirection is now installed before MCP servers start, so subprocess logs (e.g., OAuth debug output from `mcp-remote`) are captured into the log file instead of being printed to the terminal
 - Shell: Fix subprocess hang on interactive prompts — the `Shell` tool now closes stdin immediately and sets `GIT_TERMINAL_PROMPT=0` so commands that require credentials (e.g. `git push` over HTTPS) fail fast instead of blocking until timeout
 - Core: Fix JSON parsing error when LLM tool call arguments contain unescaped control characters — use `json.loads(strict=False)` across all LLM output parsing paths to prevent tool execution failure and session corruption
 - Shell: Auto-trigger agent when background tasks complete while idle — the shell now detects when a background bash command or agent task finishes and automatically starts a new agent turn to process the results, instead of waiting for the user to type something

docs/en/release-notes/changelog.md

@@ -4,6 +4,7 @@ This page documents the changes in each Kimi Code CLI release.
 
 ## Unreleased
 
+- Core: Fix MCP server stderr pollution — stderr redirection is now installed before MCP servers start, so subprocess logs (e.g., OAuth debug output from `mcp-remote`) are captured into the log file instead of being printed to the terminal
 - Shell: Fix subprocess hang on interactive prompts — the `Shell` tool now closes stdin immediately and sets `GIT_TERMINAL_PROMPT=0` so commands that require credentials (e.g. `git push` over HTTPS) fail fast instead of blocking until timeout
 - Core: Fix JSON parsing error when LLM tool call arguments contain unescaped control characters — use `json.loads(strict=False)` across all LLM output parsing paths to prevent tool execution failure and session corruption
 - Shell: Auto-trigger agent when background tasks complete while idle — the shell now detects when a background bash command or agent task finishes and automatically starts a new agent turn to process the results, instead of waiting for the user to type something

docs/zh/release-notes/changelog.md

@@ -4,6 +4,7 @@
 
 ## 未发布
 
+- Core：修复 MCP 服务器 stderr 污染问题——stderr 重定向现在在 MCP 服务器启动前安装，子进程日志（如 `mcp-remote` 的 OAuth 调试输出）将被捕获到日志文件，而非输出到终端
 - Shell：修复子进程遇到交互式提示时挂起的问题——`Shell` 工具现在会立即关闭 stdin 并设置 `GIT_TERMINAL_PROMPT=0`，使需要凭证的命令（如通过 HTTPS 执行 `git push`）快速失败，而非阻塞至超时
 - Core：修复 LLM 工具调用参数包含未转义控制字符时 JSON 解析失败的问题——在所有 LLM 输出解析路径使用 `json.loads(strict=False)`，防止工具执行失败和会话永久损坏
 - Shell：空闲时自动响应后台任务完成——Shell 现在会检测后台 Bash 命令或 Agent 任务的完成，并自动发起新的 Agent 轮次处理结果，无需等待用户输入

src/kimi_cli/cli/__init__.py

@@ -337,9 +337,10 @@ def kimi(
 
     from .mcp import get_global_mcp_config_file
 
-    # Don't redirect stderr yet. Our stderr redirector replaces fd=2 with a pipe, which
-    # would swallow Click/Typer startup errors (e.g. config parsing / BadParameter).
-    # We re-enable stderr redirection after KimiCLI.create() succeeds.
+    # Don't redirect stderr during argument parsing. Our stderr redirector
+    # replaces fd=2 with a pipe, which would swallow Click/Typer startup errors.
+    # Redirection is installed later, right before KimiCLI.create(), so that
+    # MCP server stderr noise is captured into logs from the start.
     enable_logging(debug, redirect_stderr=False)
 
     def _emit_fatal_error(message: str) -> None:
@@ -526,6 +527,15 @@ async def _run(session_id: str | None) -> tuple[Session, bool]:
                 if changed:
                     session.save_state()
 
+            # Redirect stderr *before* KimiCLI.create() so that MCP server
+            # subprocesses (e.g. mcp-remote OAuth debug logs) write to the log
+            # file instead of polluting the user's terminal.  CLI argument
+            # parsing has already succeeded at this point, so Typer/Click
+            # startup errors are no longer a concern.  Fatal errors from
+            # create() are still visible because _emit_fatal_error() writes to
+            # the saved original stderr fd.
+            redirect_stderr_to_logger()
+
             instance = await KimiCLI.create(
                 session,
                 config=config,
@@ -542,10 +552,6 @@ async def _run(session_id: str | None) -> tuple[Session, bool]:
                 defer_mcp_loading=ui == "shell" and prompt is None,
             )
             startup_progress.stop()
-
-            # Install stderr redirection only after initialization succeeded, so runtime
-            # stderr noise is captured into logs without hiding startup failures.
-            redirect_stderr_to_logger()
             preserve_background_tasks = False
             try:
                 match ui: