I like to build stuff - father of two 👦👦 - Teacher, Tinkerer - Senior Staff at @PayFitEng - TanStack router maintainer - Nx Champion 🏆 - @traefik Ambassador
Tried @OpenAIDevs Codex code mode on a real task today
Same outcome, but context usage looked meaningfully better:
- with code mode: ~88k input tokens in the final request
- without it: ~118k
So the result is there!
The annoying bit: it’s not surfaced properly in the Codex
@thsottiaux there is one big feature missing for large codebases in codex, that is present in opencode is the auto load of the AGENTS.md found in sub folders. In opencode, when the agent reads a file, the read tools auto loads sub agent md found, that's the main feature I'm
I wrote about how I use AI with Obsidian without letting it take over my notes
The core idea: every AI run should leave a visible diff, an audit trail, and a clear point where I decide whether I trust the result
If you publish private packages and use Pnpm 11, you may need to check if your packages are still private if you use publishConfig access to configure this 👀
@pnpmjs