Show HN: LLM control of deterministic coder – Sif 1.0 – LLMs as vibe coders
github.comWas looking to speed up coding and reduce token usage. Theory is an LLM can provide a plan to a deterministic local coder faster and with less token use than writing the code itself, and so far that's panned out. Revisions and teaching the coder new local skills can take time and tokens in the short run, but Sif remembers the new skills and doesn't have to be taught those skills again. Name: Sif 1.0 License: Apache 2.0 Developed and tested on Windows. Most of my testing involved converting Python to C++, though I have experimented with general coding tasks. Python to C++ conversion has worked well so far. Takes about 250-300 tokens for a frontier model, 400 to 500 tokens for a flash open source model to produce the plan. If Sif has experienced all requirements before, can produce one shot conversion with no repairs needed. Several test/benchmark reports can be found in the repository highlighting successes and failures along the way if interested. For those who are interested, I would greatly appreciate any feedback offered.
This is the only way I can think llm could be used meaningfully in an enterprise environment, sooner or later this is going to be a category, separation of plan and execution, exactly like terraform, I’m exploring this for a while now, saw many interesting ideas in your repo,especially ‚learn’ , would love to connect, mine is rigorix written in rust.
Thanks for the comment! I agree, this seems like the inevitable future. I would argue that plan and execution is already separated, but between human and AI. This concept pushes that down a level. As a coding tool it is a good demonstration of the larger concept of probabilistic control of deterministic infrastructure. I too think it will become a category. I found Rigorix-oss on GitHub. Is it mainly to be used by human coders? Can an LLM use it? Mine started out as an LLM skill and developed into a MCP hybrid.
yes, llm can use it via mcp, claude plans and writes the template , then handover to rigorix , it executes the DAG, without llm , completely deteministic
The idea looks good, however I am of the opinion that LLMs cant be deterministic because of its inherent nature based on probablisitic mathematics, that is why we are building something to make AI agents auditable, if you want i can share it here its opensource
I totally agree, you are correct, LLMs are inherently probabilistic. Sorry if I was confusing. Yes, please share.
i do agree, llm can not be deterministic, and that's why i sepereated the execution from planning in rigorix, execution phase is a deteministic DAG.would love to take a look at what you are doing, could you please share