OSS Radar

OpenScholar, STORM, TruLens, and real Codex/Claude runs show where AI citations improve, where they fail, and what makes engineering evidence useful.

interp-engine, Neuronpedia's new interpretability engine, names 34 hook points once and compares 27 across engines. I audited the code and ran it on a Mac.

Bun 1.4 moves fifteen package jobs into one Rust-based binary. The package graph shrinks, but compatibility and maintenance move into the runtime.

Herdr gives coding agents durable terminal sessions, agent state, and a control API. Screen-derived semantics remain its limit.

Buzz gives agents signed accounts and shared local compute. Delegation reliability and private code access remain unfinished.

How the Astral acquisition reveals the real war being fought in AI coding — and why everyone is talking about the wrong thing.

A new series: I find a source, go through it, distill what's worth your time, and tell you what I actually think. First up — Vercel's Open Source Program Winter 2026 cohort.