Built Deja for the @hydra_db hackathon: a graph that remembers which tools past agent tasks used. 67% vs 37% for tool search, plus the baseline that beat my first version. https://t.co/AUhv75daKa
@divyansh_1729 I have used it more than Opus 5.5 since its release and have also used it to land multiple PRs in T3code. Plus it's really good when working in a large codebase. 6.2 sol is better than 6.2 astra but openai in general is behind right now
I finally figured this out:
The best way to contribute to OSS is to actually use the software. When you rely on it daily, youβll naturally spot bugs and features worth improving.
I used to think I had to read the entire codebase first just to understand how everything connects. You don't need to, just start using it.
I want to contribute to @t3dotcodes but going through the repo alone to understand the codebase would take me at least a week or twoπ How do I get started does anyone have some advice?
The PR: https://t.co/JFjN4H8ud2
The fix was low severity, but if it had been a critical one, an agent ignoring a repo's rules would be a serious problem. Always read the diff.
Has your agent ever ignored your rules?
My AI agent read the rule in AGENTS.md and ignored it anyway.
I was setting up custom themes in @t3dotcodes and found a bug: my theme didn't show up in new chats. I blamed Gemini Neon at first, but it was a real bug. I fixed it with a 1-line change and I'm now a contributor π
But Gemini Neon's first attempt was 29 lines. 28 of them were UI tests, which T3 Code's AGENTS.md explicitly says not to include.
When I asked why, it said this π
@jullerino you need to chill you are too quick broπ. Was just gonna open a PR for this issue that I found and julius already opened up a PR and fixed it.
https://t.co/RPoWwaxV5K
That day has finally come I'm a gemini glazer nowπ
Jokes apart idk how they closed the gap from all the other gemini models to neon but neon is fucking incredible it's so good to work with too.
@DopestBigge@Super_Nius Claude is simply better than gemini I'm just stating facts π If some day gemini becomes better than claude then im gonna glaze gemini dw
@aakashH242 Not right away. For a new server, Deja finds tools from their descriptions about as well as normal tool search. It gets better as you use them. Each task that works teaches it what to load next time. It pays off when you have lots of tools and use them regularly
Built Deja for the @hydra_db hackathon: a graph that remembers which tools past agent tasks used. 67% vs 37% for tool search, plus the baseline that beat my first version. https://t.co/AUhv75daKa