Building a full coding system around GPT-6 Astra with 8 open-source repos
The author lists Serena, Context7, LangGraph, SWE-agent, E2B, pytest, DeepEval and Langfuse as a closed loop from reading a repo to producing a PR.
GPT 6 Astra + GitHub can already be turned into a full coding system I’d use 8 open source repos around the model, with each one solving a different part of the loop 01 Serena ▸ https://github.com/oraios/serena → reads the codebase at symbol level, finds references and gives Astra a real map of the repo. 02 Context7 ▸ https://github.com/upstash/context7 → pulls current docs and code examples into context, so the agent is not guessing old APIs. 03 LangGraph ▸ https://github.com/langchain-ai/langgraph → controls the loop and keeps state while the agent plans, edits, checks results and tries again 04 SWE agent ▸ https://github.com/SWE-agent/SWE-agent → takes the GitHub issue and runs the actual repair loop against the repository 05 E2B ▸ https://github.com/e2b-dev/E2B → gives the agent an isolated machine for commands, installs and code execution 06 pytest ▸ https://github.com/pytest-dev/pytest → runs the project tests after the code changes 07 DeepEval ▸ https://github.com/confident-ai/deepeval → checks the agent itself, including complete trajectories, tool calls and individual steps 08 Langfuse ▸ https://github.com/langfuse/langfuse → records the full run so you can inspect model calls, retrieval, tools and failures instead of guessing what broke the build: → Serena understands the repo → Context7 gets the docs → LangGraph controls the loop → SWE agent edits the code → E2B runs it → pytest tests it → DeepEval checks the agent → Langfuse shows what happened → PR The useful part is that every failure now has somewhere to go Bad repo context goes back to Serena Wrong API knowledge goes back to Context7 Broken code gets caught by pytest Bad agent behavior gets caught by DeepEval And if the whole run turns into a mess, Langfuse gives you the trace instead of another hour of guessing That is the kind of stack I’d build around GPT 6 Astra before spending time swapping one frontier model for another
This post quotes another post: x.com/lummox_eth/status/2100288151897981255