@OverHashDev

software developer (he/him)!

 New Zealand
Joined August 2017
  • Tweets2.6K
  • Following456
  • Followers315
241Media
Been using Herdr lately to manage my agent sessions and really enjoying it. There's still more work to be done to have better integration with jj, however it has excellent hooks with all agent harnesses. Highly recommend! Next on the list is to try out Collie to prompt remotely.
2
4
476
One of my favorite features of Herdr is that it has great ways for agents to introspect it's state and update it. tmux also has this capability, but it feels so much more natural using Herdr. Some example prompts and outputs, this is really powerful to get Herdr spawning agents.
87
GPT 5.4 has a much better ability to delegate work to sub-agents compared to GPT 5.3. Even within OpenCode, which has not been tuned for subagents as much as Codex has (now), GPT 5.4 is able to delegate work even at the planning stage! This is sometihng GPT 5.3 would never do.
1
227
Trying out new code review tools lately! On my list is using `jj`, from ben.gesoff.uk/posts/reviewin…. I'm particularly interested in comparing to a native Git approach of walking commit-by-commit and taking notes, then using an LLM to post the notes correctly on GitHub.
1
2
181
more replies
Related, but I am excited for GitHub to have an early preview of a native Stacked PR flow. Until then, I've been following davepacheco.net/blog/2025/st… which appears to have been a pretty good approach to help coworkers review my code.
1
120
Overall, I'm still exploring new methods to review code! I think the promise of `jj` to have `jj edit` is much better than Git's experience of amending old commits, so I hope that GitHub supports jj better in the future.
73
A good release week for Luau! Both of these issues were opened by me on GitHub 🙂
1
5
361
Tried out Cursor and the updated Gemini 2.5 Pro model today. Overall so far Claude Code seems to work out-the-box better. Gemini 2.5 Pro experience with Cursor is not perfect -- using wrong file extensions and missing out on files included in the context.
2
1
367
more replies
I tried Cursor for another day (in agent mode), and overall it was a net negative. The agent mode fired off too fast and left me no time to evaluate each step. Although I spent substantial time creating a detailed plan of implementation and finding relevant context, the model
1
219
continues to go down a rabbit hole and provided me no way to 'step-in' before I felt like it had wasted a whole bunch of work. Ask mode is much more slow-paced and careful, so I need to try it out again (and this time, for a full week, to give it a fair try!)
181
oops accidentally deleted everything with git, thank you vscode for having a local history of everything
1
2
201
One of my first (proper) luau user-defined type functions -- a readonly<T> type which accepts a table/union/intersection type and spits back a readonly form. Useful now that a table.freeze call will return the same type (hopefully) as the output of this type
1
4
262
Just realized I had to amend the end of this to support indexers (for e.g., arrays)! This stuff is so new that there's currently no support for read/write indexers -- only general indexers. So it's not perfect today.
1
158
Been using Claude Code lately -- code output is quite good, but it's expensive (relative to other LLM tools). Would definitely exceed the cost of a Cursor subscription. My late night thought is that this is a great tool for finding dead code / duplicate code across a codebase.
1
181
more replies
Claude Code is decently good at finding relevant files / code to get enough context before it tries to code. Asking the model to look around more or at a specific place also tends to lead to a good result. Overall, been an interesting experience!
1
133
Still very far from replacing an engineer -- there is a lot of mental overhead in tracking what the LLM is doing and making sure it executes correctly. Feels akin to the work of doing a code review + writing the code yourself, but just much faster.
113
Load more