Stop regenerating.
Start converging.

RHA Wrapper compiles a prompt into a frozen test contract before any implementation exists, builds each feature against it, and repairs failures inside bounded node boundaries. Iterate while progress is being made; escalate — never regenerate — when it stalls. Works with Claude, or the model already loaded in your LM Studio / Ollama.

0s    → encoding features
41s   encoded F1 · Identity, Permissions, and Governance
41s   ⚠ stub repair pass 1: fixing 1 file(s) (8 errors)
3m32s ⚠ stub repair pass 5: fixing 1 file(s) (1 errors)
3m40s interface stubs compiled (12 files)
3m40s → generating IR

84m   ✓ 143/149 frozen tests passing · 6 escalated for human judgment

The doctrine — three levels

L1 · Frozen IR

The prompt is feature-encoded; every node receives a test suite derived from the spec alone, before any code exists. The suite freezes. Implementations conform to it — never the reverse.

L2 · Bounded repair

Failures localize to one node of the tree. Repair touches only that node's files, iterating while each pass changes the error signature. A 2B local model can spend fifty cheap bounded passes; it can never re-roll the artifact.

L3 · Escalation, not synthesis

Stalled repair, deficient input, ambiguous intent — all surface to you as structured analysis with the ledger trail attached. Convergence criteria are human-defined; the machine flags gaps, it never invents your decisions.

Setup — the extension comes first, keys stay on your machine

  1. Download the Provider Vault extension and extract it somewhere permanent.
  2. Open (click to copy, paste into Chrome's address bar — Chrome blocks direct links to its own pages).
  3. Enable Developer mode (top-right switch at ), click Load unpacked, pick the extracted folder.
  4. Create your vault (AES-256-GCM, PBKDF2 — the key never leaves the extension's memory).
  5. Add a provider. RHA Wrapper works with ChatGPT, Claude, and Gemini, local LMs (LM Studio / Ollama), and any other provider that complies with the OpenAI-compatible schema. Paste an API key (or a local server URL, e.g. http://localhost:1234/v1) and the model name — local servers auto-detect. Set Active — the server hot-reloads, no restart.
  6. Launch the app and paste an already-defined feature spec — author the features with your LLM of choice (ChatGPT, Claude, …): one feature per line, observable behavior + edge cases. You don't describe an idea here; you paste a spec. Compile.

Run it locally — no jargon

Everything runs on your own computer. Nothing is uploaded, nothing phones home, and the files stay on your drive for as long as you want them. Here is the whole thing, step by step.

Download RHA Wrapper — the app bundle (.zip)
  1. Unzip it. You get a folder. Put it wherever you like — Desktop is fine. It just sits there; that folder is the app.
  2. Install Node.js once, from nodejs.org (free, official — click the big green button, run it, click Next). It's the engine that runs the app.
    what is Node.js?The standard, open-source runtime for JavaScript apps — the same thing that runs a huge share of the web. Installing it does not change your system; it just lets your computer run programs like this one.
  3. Open the app — double-click the one file named for your system:
    • Mac: double-click run.command
    • Windows: double-click run.bat
    • Linux: run ./run.sh
    You do not hunt for a .js file — it's the one named run. The first launch spends a minute getting the pieces it needs; after that it's instant.
    “getting the pieces it needs” — what pieces? (nothing scary) A handful of well-known open-source building blocks, downloaded once into the folder:
    • Express — the tiny local web server that serves the app to your browser.
    • React — the interface you click around in.
    • Vite — builds that interface.
    • TypeScript / tsx — runs the code.
    • Vitest — runs the frozen tests that verify generated code (the whole point of RHA).
    • Anthropic SDK — only used if you pick Claude as your model.
    That's the entire list. None of it talks to us; all of it lives in your folder.
  4. Your browser opens to the app at http://localhost:4100. Bookmark that page — it's your permanent door in. Any time you want RHA later, double-click the same run file again and your bookmark loads it.
  5. Rather not double-click anything? Open your computer's Terminal (Mac/Linux) or Command Prompt (Windows), drag the unzipped folder onto the window and press Enter, type npm run dev and press Enter — then open http://localhost:4100 in any browser yourself. Exact same result, fully by hand.

What did it just do? The launcher started RHA as a small web server on your machine and pointed your browser at it — that's all. No account, no upload, no background installer. Close the window to stop it; double-click run to start it again. The folder can live on your hard drive forever.

Why Creative Commons — knowledge is freedom

RHA Wrapper is CC BY 4.0. That means anyone may:

Clone it, host it, sell it, fork it, skin it, ship it, break and fix it.
Prompt it, wrap it, stack it, split it, merge it, branch it, patch upgrade it.
And most other things.

The one rule: keep the attribution.

“Beware he who would deny you access to information, for in his heart he dreams himself your master.” — Commissioner Pravin Lal, U.N. Declaration of Rights, Sid Meier's Alpha Centauri, Firaxis Games, 1999

RHA Wrapper software can be cloned, hosted, copied, modified, and distributed freely by anyone. The original instance is located at https://rha.software.

We do not guarantee the safety or authenticity of a copy hosted at any other address. If you see it in the wild, it is probably a scam and you should not proceed.

RHA Wrapper captures no personally identifiable information. If you are ever prompted to register by email, or to disclose API secrets inside a web page, it is a scam.

Developed by Andy Funke LinkedIn