This repo adds context on "how things should be done", much like a CONTRIBUTING.md or CONTEXT.md
What it provides beyond that is:
* profiles, which can also be implemented as a directive in any prompt, eg "use CONTEXT-lite.md" as opposed to a different agent reasoning level. You'll just have to memorize them?
* opinionated choices about host implementation (for Codex use CONTEXT.md and Claude use CLAUDE.md, etc)
* cpp/Python/Universal rulesets
There's some nice invariant rules and structure[1], to be sure, but that doesn't justify the install for me.
[1] Everything in the packs are machine-readable rule atoms that are similar to SARIF and RFC 9457. There is also an avante-guarde routing system for selecting to add additional packs to context.
I think we are converging on similar approaches to how deterministic quality verification with coding agents should be performed, but from different angles.
I’ve been tinkering on a package to capture linking, testing and gating quality output in a framework unified way across coding agents.
I would love to have a chat with you about how you see this space developing and how you look at it.
Problem I kept hitting: ask the agent for a 15-line fix, get a 500-line renovation that still compiles and passes tests.
Boffin is a small control layer for AI coding agents. Before an edit, it feeds the agent only the architectural constraints for the file it is about to touch, then makes it verify the result. Not another static AGENTS.md for the whole repo.
On DuckDB, a guided refactor landed at +17 / -17 lines with 2,104 assertions passing. Same shape of cases also on FastAPI and LangChain (in the repo examples/).
Try it:
npx boffinit cursor
Also works with Claude Code, Codex, and OpenCode.
Happy to answer questions — especially if you have a place where the agent keeps "improving" load-bearing code.
To be blunt the degree to which I take these projects seriously is basically the same as how extensive their eval suite is. A single guided refactor is better than nothing, but one success could just be a fluke. I want to see hundreds, if not thousands, of evals to convince me you’re a project worth looking into.
The AI generated text is also not doing you any favors.
This repo adds context on "how things should be done", much like a CONTRIBUTING.md or CONTEXT.md
What it provides beyond that is:
* profiles, which can also be implemented as a directive in any prompt, eg "use CONTEXT-lite.md" as opposed to a different agent reasoning level. You'll just have to memorize them?
* opinionated choices about host implementation (for Codex use CONTEXT.md and Claude use CLAUDE.md, etc)
* cpp/Python/Universal rulesets
There's some nice invariant rules and structure[1], to be sure, but that doesn't justify the install for me.
[1] Everything in the packs are machine-readable rule atoms that are similar to SARIF and RFC 9457. There is also an avante-guarde routing system for selecting to add additional packs to context.
[flagged]
I think we are converging on similar approaches to how deterministic quality verification with coding agents should be performed, but from different angles.
I’ve been tinkering on a package to capture linking, testing and gating quality output in a framework unified way across coding agents.
I would love to have a chat with you about how you see this space developing and how you look at it.
https://github.com/mateffy/gesetz
[flagged]
Hi HN - author here.
Problem I kept hitting: ask the agent for a 15-line fix, get a 500-line renovation that still compiles and passes tests.
Boffin is a small control layer for AI coding agents. Before an edit, it feeds the agent only the architectural constraints for the file it is about to touch, then makes it verify the result. Not another static AGENTS.md for the whole repo.
On DuckDB, a guided refactor landed at +17 / -17 lines with 2,104 assertions passing. Same shape of cases also on FastAPI and LangChain (in the repo examples/).
Try it: npx boffinit cursor
Also works with Claude Code, Codex, and OpenCode.
Happy to answer questions — especially if you have a place where the agent keeps "improving" load-bearing code.
To be blunt the degree to which I take these projects seriously is basically the same as how extensive their eval suite is. A single guided refactor is better than nothing, but one success could just be a fluke. I want to see hundreds, if not thousands, of evals to convince me you’re a project worth looking into.
The AI generated text is also not doing you any favors.
[flagged]
So your headliner case is that it managed to edit 17 lines in DuckDB? Why is the line count relevant here? What is the 17-line change even about?
[flagged]