i kept repeating the same workflows and hitting similar walls. ai coding tools like Claude Code, OpenCode, and Codex are amazing around a single prompt, but they fall short when it comes to end-to-end software engineering without constantly being steered and corrected, esp. around a real flow. one where you need to do deep requirements gathering, even by reaching out to people, research anything and everything you can get your hands on, and then plan it with all the nitty gritty in mind, to finally produce something that works. just to start refactoring it endlessly until it becomes something you're proud of.
that's why i built this. so it can leverage these incredible tools and their never-ending new features and instead of trying to do the same thing, use them to drive the flow at a meta layer instead. this way, we get all the cool new features they add while still having a reliable orchestrator that does what we need without us having to repeat ourselves and steer it.
once it starts, it begins polling for new tasks from Github issues that are assigned to it, then drives whichever CLI coding tool you use through the whole flow until it comes back with a pull request. it only stops to get your feedback at two main checkpoints. when it needs requirements (and you haven't provided enough), and when it's done and the work needs to be reviewed. and it only merges the work after your approval.
the thing that's most useful about it is that it isn't tied to any one tool. you can swap Claude Code for OpenCode, or Github for Gitlab, all without modifying the Core.
it's still a work in progress though. works like a charm but needs to be enhanced to really get to the perfect output we expect in one flow run and without needing comments in the PR to refine its changes. oh did i mention, you can leave comments in the PR and it just starts the flow again, applying every single piece of feedback, just like a real software engineer!
other constraints and feature ideas are documented in the docs as well and are a work in progress. one of the first i'd love to do is sandboxing each session. stay tuned!
any feedback is greatly appreciated. thanks for reading!
hey, creator of The Engineer here.
i kept repeating the same workflows and hitting similar walls. ai coding tools like Claude Code, OpenCode, and Codex are amazing around a single prompt, but they fall short when it comes to end-to-end software engineering without constantly being steered and corrected, esp. around a real flow. one where you need to do deep requirements gathering, even by reaching out to people, research anything and everything you can get your hands on, and then plan it with all the nitty gritty in mind, to finally produce something that works. just to start refactoring it endlessly until it becomes something you're proud of.
that's why i built this. so it can leverage these incredible tools and their never-ending new features and instead of trying to do the same thing, use them to drive the flow at a meta layer instead. this way, we get all the cool new features they add while still having a reliable orchestrator that does what we need without us having to repeat ourselves and steer it.
once it starts, it begins polling for new tasks from Github issues that are assigned to it, then drives whichever CLI coding tool you use through the whole flow until it comes back with a pull request. it only stops to get your feedback at two main checkpoints. when it needs requirements (and you haven't provided enough), and when it's done and the work needs to be reviewed. and it only merges the work after your approval.
the thing that's most useful about it is that it isn't tied to any one tool. you can swap Claude Code for OpenCode, or Github for Gitlab, all without modifying the Core.
it's still a work in progress though. works like a charm but needs to be enhanced to really get to the perfect output we expect in one flow run and without needing comments in the PR to refine its changes. oh did i mention, you can leave comments in the PR and it just starts the flow again, applying every single piece of feedback, just like a real software engineer!
other constraints and feature ideas are documented in the docs as well and are a work in progress. one of the first i'd love to do is sandboxing each session. stay tuned!
any feedback is greatly appreciated. thanks for reading!
oh and here's a full end-to-end demo of it: https://youtu.be/48t6g_co9j4