Sakasegawa's blog. Most posts on this blog are written with the help of coding agents.

Hi there! This is Sakasegawa-chan (@gyakuse)!

Today I want to write about HarnessBench, a benchmark I built for comparing coding agent harnesses.

Hi there! This is Sakasegawa-chan (@gyakuse)!

Today I want to walk through a mojibake (character corruption) issue that shows up when writing Japanese on recent versions of Claude Code, along with a simple workaround using hooks.

Update 2026-04-08: The Changelog says this issue was fixed in Claude Code v2.1.94, but there are reports that it still isn't fully resolved. I recommend keeping the hook in place for now. I also switched the hook back from PreToolUse to PostToolUse, because PostToolUse only has to repair the corrupted spots after the fact, which costs fewer tokens to fix.

Introduction: The Problem at Hand

In the practical domain of software development, a certain quasi-ritual practice has been observed sporadically. Some users of Claude Code, Anthropic's AI coding agent, place a file titled .claudeignore in the root directory of their projects.

Hi there! This is Sakasegawa-chan (@gyakuse)!

I've been doing so much solo dev work that my bank account has reached zero, so I'm opening up for business!

Hi there! This is Sakasegawa (@gyakuse)!

I was about to pick Cursor back up after a while, but then I saw that Composer 2 is based on Kimi, and I thought, "wait, is Cursor really ZDR (Zero Data Retention)?" One thing led to another and I ended up investigating all the other Coding Agents too (?). If you're motivated to contribute your data to training, you might be surprised to find out that most of them don't actually train on it. And if you'd rather not contribute, this should help you eliminate the risk of your code being used. Kimi Code stood out: it trains on your data enthusiastically unless you email them. If you want to help push the frontier, that's a great option. For what it's worth, even the API usage of Kimi (Moonshot AI) gets used for model training. Pretty bold.

Update 2026-03-27: GitHub Copilot announced that starting April 24, data from Free/Pro/Pro+ users will be used for AI model training by default. See the GitHub Copilot section below. I also updated the rest of the tool information to reflect the latest terms.