Building a coding stack around GPT-6 Astra with 8 open source repos
OpenHands, Serena, Context7, Browser Use, E2B, Ruff, pytest and Langfuse each handle one part of the loop, leaving Astra to understand the problem and write the change.
GPT 6 Astra + GitHub is getting f...cking ridiculous I’d build the entire coding loop around 8 open source repos 01 OpenHands ▸ https://github.com/All-Hands-AI/OpenHands → takes the issue and runs the engineering task 02 Serena ▸ https://github.com/oraios/serena → finds symbols, references and the exact code Astra needs to touch 03 Context7 ▸ https://github.com/upstash/context7 → pulls current docs and examples before Astra writes against an outdated API 04 Browser Use ▸ https://github.com/browser-use/browser-use → researches anything missing from the repo or docs 05 E2B ▸ https://github.com/e2b-dev/E2B → runs the generated code inside an isolated sandbox 06 Ruff ▸ https://github.com/astral-sh/ruff → catches formatting, lint and common Python problems before testing 07 pytest ▸ https://github.com/pytest-dev/pytest → checks whether the patch actually works 08 Langfuse ▸ https://github.com/langfuse/langfuse → keeps the trace when the agent screws something up Give each of those jobs to a dedicated layer and the model gets one job Understand the problem and write the right change That is the stack I’d clone before trying another model
This post quotes another post: x.com/promixyz/status/2102424036416348494