VS Code extension: Bash tool returns exit 0 but stdout is always empty (WSL2)

Summary

In the VS Code extension (moonshot-ai.kimi-code 0.6.6, running on the WSL remote server), the Bash tool executes commands and reports exit 0, but the tool result handed back to the model contains no stdout — only the literal string Command executed successfully.

The same CLI binary (0.31.0) captures stdout correctly when used as a terminal CLI in the same working directory.

Environment

• VS Code extension: moonshot-ai.kimi-code 0.6.6 (linux-x64), activated on the WSL remote (remote: wsl)
• CLI: kimi-code 0.31.0 (~/.kimi-code/bin/kimi)
• Host: Windows 11 + WSL2, kernel 6.18.33.2-microsoft-standard-WSL2, Ubuntu userland
• Extension host env includes SHELL=/usr/bin/zsh, TERM_PROGRAM=vscode
• Permission mode: manual (auto mode disabled); commands were approved via the permission prompt

Evidence

From the extension session's wire file (~/.kimi-code/sessions/<wd>/<session>/agents/main/wire.jsonl), an approved ls . call records:

{"type": "tool.call", "name": "Bash", "args": {"command": "ls ."}}
{"type": "tool.result", "toolCallId": "tool_NcuQOd43Ul7hozch48nRHGdG",
 "result": {"output": "Command executed successfully."}}

Same for a second approved command (pwd && ls -1 | head -5) — exit 0, stdout missing. The directory is not empty; a CLI session in the same directory returns full listings.

What I ruled out

• SHELL value: reproducing with SHELL=/usr/bin/zsh kimi -p "run echo MARKER..." in the terminal captures stdout fine.
• CLI itself: both uiMode=shell (interactive CLI) and uiMode=print (kimi -p) return stdout correctly on the same machine, same project, same binary.
• VS Code user settings: no kimi/terminal settings that affect bash.

So the stdout loss happens only on the extension ↔ backend path (bridge mode), not in the CLI's bash execution itself.

Repro

1. Open VS Code connected to a WSL2 workspace.
2. In the Kimi Code panel, ask the agent to run a bash command, e.g. ls ..
3. Approve the permission prompt.
4. The agent reports the command ran with exit 0 but sees no output.

Start with the extension↔backend bridge path for the Bash tool and compare it with the CLI execution path. Use the wire file at ~/.kimi-code/sessions///agents/main/wire.jsonl and reproduce with ls . or pwd && ls -1 | head -5; done means the tool result handed to the model contains the command's stdout instead of only the success message.