I build stuff; Software Engineer @nutanix; @BITSPilaniGoa alum
Bengaluru, India
- Tweets 175
- Following 674
- Followers 321
- Likes 597
Pinned Tweet
git blame says I wrote this code.
I didn't. Two different AI agents did, in two different tools... and the conversations explaining why are gone.
So I built TracesHub: line-level AI attribution for your codebase. Which model wrote each line, and the prompt behind it. 🧵
3
1
8
2,819
I think Jev and this wave of System One models brings best of both worlds. The value of doing meaningful work by conveying non-contractual information and reliability of deterministic (sort of) output. This being cheap, fast, and "high quality" will unlock amazing new software
It feels there is a lot of room to reimagine the nature of most software, the UX, the form factor of how most technology deliver value now. Even without any AGI/ASI coming, meaningful improvements and bringing reliability in current models would let us do all that.
1
1
324
Ujjal Sharma retweeted
Asked Claude Code to 'clean up the magic numbers' in a retry loop. It ran 'agent-trace context', found the rate-limit ticket behind them, and refused.
Your agent, reading the why before it edits.
traceshub.com
1
227
Everything here is local. 'agent-trace viewer' shows full blame UI in your browser, no account, hooks never touch the network.
traceshub.com
335
So it has been 3.5 years since Messi won the world cup
and you are telling me there is chance I won't get to say 4 years for another 4 years
420
This repo is 43.7% AI-written across 2 tools and 3 models. Do you know your number?
traceshub.com
155
TracesHub: Line-level AI attribution for your codebase producthunt.com/products/tra… via @producthunt
170
See the model and prompt/context behind each line of code
youtube.com/watch?v=J4LPhV9w…
325
Ujjal Sharma retweeted
git blame says I wrote this code.
I didn't. Two different AI agents did, in two different tools... and the conversations explaining why are gone.
So I built TracesHub: line-level AI attribution for your codebase. Which model wrote each line, and the prompt behind it. 🧵
3
1
8
2,819
Ujjal Sharma retweeted
The whole flow is like git:
pip install agent-trace-cli
(or curl)
agent-trace init
(…work normally, commit… )
agent-trace push
Local-first: hooks never touch the network. Nothing leaves your machine until you push. There's even a full local viewer. No account needed.
1
1
148
Ujjal Sharma retweeted
My favorite part: your agents read it too.
'agent-trace context' gives an agent the intent behind existing code.
Asked ours to "clean up magic numbers" in a retry loop ... it looked up the context, found the rate-limit ticket, and kept them.
1
1
123
git blame says I wrote this code.
I didn't. Two different AI agents did, in two different tools... and the conversations explaining why are gone.
So I built TracesHub: line-level AI attribution for your codebase. Which model wrote each line, and the prompt behind it. 🧵
3
1
8
2,819
The CLI is open source, the service is self-hostable, and
TracesHub (the hosted hub - think git -> GitHub) is free to try.
CLI: github.com/ujjalsharma100/ag…
Hub: traceshub.com?utm_source=x&u…
Demo video: youtube.com/watch?v=J4LPhV9w…
1
141
Built solo...feedback very welcome, especially the brutal kind.
traceshub.com
112