Description
Git and Visual Studio Code ("VSCode") are essential tools for software development. The built-in VSCode Git tools leave a lot to be desired, so VSCode extensions have been developed to support additional Git features. I used the GitLens extension for years, and even contributed a feature to it. The extension was later acquired by GitKraken. Over time, it became bloated with enormous numbers of settings, overcomplicated panels with graphs and dashboards, obtrusive paid features, and AI integrations. It also began regularly invalidating extension versions, which would disable the extension and display obnoxious notifications forcing users to update.
I eventually got fed up with GitLens and decided to write my own extension from scratch. Introducing GitLess, a VSCode extension that offers the essential features of GitLens with less bloat. It is available on the Open VSX registry and Microsoft marketplace.
What GitLess does not do:
- GitLess does not collect telemetry.
- GitLess does not expire extension versions or disable functionality until you update.
- GitLess does not require a separate account, subscription, or cloud service.
- GitLess does not turn your Source Control panel into a full workflow dashboard.
- GitLess does not run heavyweight background services just to show Git history.
Development
I built the initial version 0.1.0 of this extension with the Kilo Code CLI (based on OpenCode) and their Auto Frontier models, which were mostly Claude Opus 4.6 at the time. The initial turn took about 15 minutes in the (now deprecated) "orchestrator" mode and costed $3.15 in credits. As usual for code generated by Claude Opus models, the code was incoherent and nonfunctional. The extension did not start, "settings" did nothing, and "tests" were little more than placeholders. Anecdotally, I've found that Claude models often ignore their instructions, so I had to interrupt turns and steer the model when it started doing things I didn't want it to do. It was also relatively expensive because inference with Claude Opus 4.6 is twice as expensive per input token as inference with GPT-5.4. My experience with Claude models in Claude Code has been largely the same, so I think this is more of a Claude model problem than a Kilo Code problem. The rest of the work on 0.1.0 took about 15 hours and costed about $30 in credits. So the initial turn only accomplished a small percentage of the work, and then it still needed a lot more work before it was functional. This is a typical experience in prompt engineering. Prompting is an iterative process.
After burning through some more Kilo Code credits (about $100 in total) fixing bugs with Kilo Code and Claude Opus 4.6, I switched to Codex with GPT-5.4. Development went more smoothly -- bugs were fixed, features actually worked, and tests evaluated real-world extension behavior. Codex paid attention to my instructions, and I appreciated its relatively concise communication style. It was also cheaper. I was able to get a lot of work done on the $20/month Plus plan, especially if I paced myself to stay under the usage limits.
See my dotfiles repo for more details on the configurations I use for Codex, Kilo Code, and other AI coding tools.