“Don’t ask again” shouldn’t mean “save my password.”
Researchers found 1,379 secrets in Claude Code approval files on public GitHub.
In Druks, AI/GitHub tokens stay outside the sandbox. Approval records stay outside the repo.
Source: https://t.co/oXrzmH4vWf
Before leaving Sherpany, I integrated Druks into their infra during the quarterly hackathon.
When the eMudhra news broke, security asked it in Slack: “Are we affected?”
It checked 118 repos and helped engineering prepare fixes that the team then merged 🤘
After eight years at Sherpany, I’ve taken the leap to build my own thing: Druks 🚀
Feels a bit crazy to leave a stable job, especially in this market. But I’m excited to see where this goes.
More soon.
https://t.co/LTfDaG4ITw
I built Druks so agents could ship my side projects. Now they review PRs, write test plans, and find sales leads too. Open source.
You write the workflow once, and agents run it every time.https://t.co/OwfkDyK8y5
@oldstackjournal Agreed, mostly. But "that smells wrong, try again" assumes you can smell it. Without having actually built things the hard way, most people can't spot subtly broken code.
@omkarships I've flipped on this. Used to think no. Now I think it depends, for most products, yes. For regulated stuff like banking, you'll still need someone technical to survive an audit.
Been playing with @Cloudflare's EmDash, the agent/MCP integration is pretty good. AI-native content ops baked in from the ground up. Genuinely impressed.
@rauchg Also the biggest patience test in history, negotiating daily with something that has zero accountability, no stakes, and doesn't run on any of the feedback loops we subconsciously use (guilt, embarrassment, 'I'll do better'). You just keep re-explaining until it lands.
@mitsuhiko Python at least made __name mangling ugly enough that nobody mistakes it for ergonomic. JavaScript’s # feels designed to maximize diff noise.