tracking: Windows support

[stdrc]

• Directory listing feat: add Windows support for directory listing #72
• Keyboard input fix(ui-termios): add Windows keyboard input support #59
• Grep binary downloading feat(windows): grep tool support and CI workflow #194
• CI workflow feat(windows): grep tool support and CI workflow #194
• CMD/PowerShell tool feat: use CMD tool on Windows #229
• Release Windows binaries ci: release binary for Windows #231

[2308087369]

CY

[255doesnotexist]

almost works with cygwin/msys2, but kimi-cli keeps trying use "bash" on Windows it is a bit weird.

[Axonsin]

Does prompt in kimi cli wont guide AI to use other commands to read files? When grep not available on Windows, it will still repeatedly try to search files using grep. Once I ask a job, it make over ten incorrect grep calls.
But on ClaudeCode, once grep fails, AI will not continue to attempt to use grep.
I know can download grep, just an advice that can upgrade prompts

[stdrc]

Does prompt in kimi cli wont guide AI to use other commands to read files? When grep not available on Windows, it will still repeatedly try to search files using grep. Once I ask a job, it make over ten incorrect grep calls.

Have you tried latest version (0.51)? It should auto download ripgrep now.

[Axonsin]

Does prompt in kimi cli wont guide AI to use other commands to read files? When grep not available on Windows, it will still repeatedly try to search files using grep. Once I ask a job, it make over ten incorrect grep calls.

Have you tried latest version (0.51)? It should auto download ripgrep now.

Oh my bad. My version is 0.50. Thanks for answer :)

[leslieo2]

Huge thanks for driving the Windows support roadmap—the steady progress has been inspiring to watch. One key piece we’ve found is teaching the agent that it’s on Windows by injecting the host OS details (shell/version/etc.) directly into the system prompt; without that context it keeps defaulting to POSIX assumptions. I just opened PR (#230) to wire that in—would love your eyes on it so we can build the rest of the Windows behaviors on top.

[stdrc]

Huge thanks for driving the Windows support roadmap—the steady progress has been inspiring to watch. One key piece we’ve found is teaching the agent that it’s on Windows by injecting the host OS details (shell/version/etc.) directly into the system prompt; without that context it keeps defaulting to POSIX assumptions. I just opened PR (#230) to wire that in—would love your eyes on it so we can build the rest of the Windows behaviors on top.

Seems to be useful but I think env context is not necessary related to Windows support. In my mind the only existing thing that gives wrong assumption is the Bash tool, so we only need to update the prompt and tool name for Bash. And I did it in #229.

Don't get me wrong. We do need a summary of the OS environment in the system prompt. I'll review your PR later. Thank you!

[leslieo2]

Huge thanks for driving the Windows support roadmap—the steady progress has been inspiring to watch. One key piece we’ve found is teaching the agent that it’s on Windows by injecting the host OS details (shell/version/etc.) directly into the system prompt; without that context it keeps defaulting to POSIX assumptions. I just opened PR (#230) to wire that in—would love your eyes on it so we can build the rest of the Windows behaviors on top.

Seems to be useful but I think env context is not necessary related to Windows support. In my mind the only existing thing that gives wrong assumption is the Bash tool, so we only need to update the prompt and tool name for Bash. And I did it in #229.

Don't get me wrong. We do need a summary of the OS environment in the system prompt. I'll review your PR later. Thank you!

The reason I pushed for a broader env summary is because of what I learned digging into Gemini CLI’s architecture: they keep a six-layer context stack (env → memory → tooling → IDE → conversation → subagents). Capturing the host OS is just the foundation for that “Layer 1,” so tools, prompts, and subagents can all make consistent decisions long-term—especially once we start mixing IDE signals or multiple runtimes.
If you’re curious, I wrote up the breakdown in https://leslieo2.github.io/posts/context-engineering/, it walks through how Gemini wires those layers together. Would love your take after you skim it!