Show HN: Open-source coding assistant with terminal access
3 by martinye022 | 0 comments on Hacker News.
Sweep is an AI coding assistant that reads your files and runs code in your terminal. It takes a list of files and instructions, and performs edits using GPT4. For example, you can tell Sweep to reference a.tsx to localize b.tsx. We previously worked on Sweep as a bot that generated pull requests from GitHub issues. We learned that AI needs validation (like formatters and unit tests) to code properly. We tried setting up a code sandbox for each user but this didn't work well. We’re now relying on the prebuilt dev environment on their machine to avoid having to install dependencies. LLM based pipelines are also error-prone. Our previous project was a pipeline that turns the issue into relevant files as context, relevant files to a generated plan of files to change, and then the generated plan to a series of code changes. It’s frustrating when one/all of these steps fail and you’re forced to start over. You can now intervene at each step, making the entire process more reliable. Sweep is open source (licensed under AGPL) and runs entirely on your laptop. The only remote calls will be to OpenAI. We'd love to help you perform software chores like version upgrades, where you describe a set of changes to have Sweep apply over multiple files. You can install Sweep here with a one-line command: https://ift.tt/bcgB3Gv . We'd love to hear any feedback you have!
3 by martinye022 | 0 comments on Hacker News.
Sweep is an AI coding assistant that reads your files and runs code in your terminal. It takes a list of files and instructions, and performs edits using GPT4. For example, you can tell Sweep to reference a.tsx to localize b.tsx. We previously worked on Sweep as a bot that generated pull requests from GitHub issues. We learned that AI needs validation (like formatters and unit tests) to code properly. We tried setting up a code sandbox for each user but this didn't work well. We’re now relying on the prebuilt dev environment on their machine to avoid having to install dependencies. LLM based pipelines are also error-prone. Our previous project was a pipeline that turns the issue into relevant files as context, relevant files to a generated plan of files to change, and then the generated plan to a series of code changes. It’s frustrating when one/all of these steps fail and you’re forced to start over. You can now intervene at each step, making the entire process more reliable. Sweep is open source (licensed under AGPL) and runs entirely on your laptop. The only remote calls will be to OpenAI. We'd love to help you perform software chores like version upgrades, where you describe a set of changes to have Sweep apply over multiple files. You can install Sweep here with a one-line command: https://ift.tt/bcgB3Gv . We'd love to hear any feedback you have!