Running GPT-6 Astra autonomously on a repo for 30 minutes to make 38 code changes
The author built a spec, routing, patching, review, test and release pipeline with OpenSpec, Kestra, Dagger and Reviewdog, letting Astra autonomously complete 38 changes.
AI coding with GPT-6 Astra is getting f...cking dangerous i gave Astra one GitHub repo and a 30 minute autonomous run and watched it route 38 changes through planning, patching, review and tests without letting a failed check reach release [the stack was basically this 👇] 1. OpenSpec turns the request into an explicit spec before Astra starts touching code, so the agent has something concrete to build and verify against ▸ https://github.com/Fission-AI/OpenSpec 2. Kestra handles the execution flow and keeps each coding job moving through separate stages instead of letting one agent freestyle the entire run ▸ https://github.com/kestra-io/kestra 3. Dagger gives the patches a programmable environment for builds and tests, so the same verification can run locally and inside CI ▸ https://github.com/dagger/dagger 4. Reviewdog takes the outputs from linters and analyzers and puts the failures directly back into the code review loop ▸ https://github.com/reviewdog/reviewdog the useful part is the order spec → route → patch → review → test → release Astra can keep moving fast GitHub decides what survives
This post quotes another post: x.com/lummox_eth/status/2100288151897981255