Fetching from the wire…
01
02
03
04
05
06
07
08
Vibe Coding2026-09-22 · source-backed
MIT, Python, appeared September 20, converting yes/no, pick-one and rubric-score evaluation into calibrated probabilities in a single forward pass. Claimed: a few thousandths of a cent per trace, p50 244ms and p95 371ms against several seconds for the same eight metrics through Ragas. It runs on Jev via TypeSafe or Vercel, on Kev locally, or on Laya at about 10ms on Apple silicon. First artifact of this wave aimed at production agent loops rather than at reverse-engineering the model. (GitHub)
Each link below shares sources, entities, or timing with this story.
Compaction is where long sessions go to die. The model summarizes what happened, the summary drops the exact string you needed three hours later, and you don't find out until the agent confidently references a file path that never existed. It's the biggest source of silent con...
kruzovic7/ai-data-extractor pulls chat histories from Claude Code, Cursor, Windsurf, Aider and Cline/Roo Code, in Python under MIT (GitHub). The fork ratio of about one in six is unusually high and suggests people are running and modifying it rather than bookmarking. No push s...
A single Rust binary running MCP and HTTP servers over a local data directory, at 7,975 stars with about 217 added September 22. Agents write observations through lifecycle hooks that consolidate into ordinary markdown you can grep, edit in Obsidian and diff as commits, with r...
Every browser-automation setup I've built in the past year has the same shape. A Playwright process. An MCP wrapper around it. A headless Chromium that drifts from the browser I actually use, with a different cookie jar, a different rendering path, and a login session I have t...
tirth8205/code-review-graph topped GitHub daily trending at 22.5k total stars, MIT licensed, building a persistent Tree-sitter map so coding agents read only relevant context. Claimed ~82x median per-question token reduction across six repos (range 38x to 528x) at 0.714 averag...
openai/PrimeGaps186, created September 2 under Apache-2.0, at 117 stars, formalizes lim inf(p_{n+1} - p_n) ≤ 186 via the Dickman-Hardy-Littlewood conjecture applied to a 40-element admissible tuple. It proves three theorems but remains conditional on a Kloosterman3 bound from...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.