# llm-config

Configures RuVLLM local inference and manages in-memory MicroLoRA and SONA adapters using quality feedback scores.

- skill · `type=skill` · Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, OpenCode · 741 installs · ★ 79357
- Author: ruvnet
- License: MIT · policy=mirror · the beginning of the file is at the end of this page
- Source: https://github.com/ruvnet/ruflo/tree/HEAD/plugins/ruflo-ruvllm/skills/llm-config/SKILL.md
- Updated: 2026-10-25
- HTML: https://affpapa.org/en/nevibecode/skills/llm-config

## What it does

This skill configures offline RuVLLM deployments through MCP tools. It checks runtime status and generates local model parameters. Training datasets are not used here.

Instead of full training runs, adapters update small in-process weights from numeric quality signals between 0.0 and 1.0. These adapters live strictly inside MCP server memory. A restart wipes them out immediately. You have to track returned IDs and recreate instances yourself.

Both adaptation tools fail if `@ruvector/ruvllm-wasm` is missing. When status checks return `wasm.available: false`, install the dependency via `npm i -g @ruvector/ruvllm-wasm`. This tooling cannot export adapter files or fine-tune hosted API models.

## Install

### Claude Code (native platform)

1. **Skills CLI (Vercel)**
```bash
npx skills add ruvnet/ruflo --skill llm-config -a claude-code
```

2. **Manually: put the files in place**
- File: `~/.claude/skills/llm-config/SKILL.md` — Copy the whole skill folder; for one project use .claude/skills/llm-config/

### Codex

1. **Skills CLI (Vercel)**
```bash
npx skills add ruvnet/ruflo --skill llm-config -a codex
```

2. **Manually: put the files in place**
- File: `.agents/skills/llm-config/SKILL.md` — For all projects use ~/.agents/skills/

### Cursor

1. **Skills CLI (Vercel)**
```bash
npx skills add ruvnet/ruflo --skill llm-config -a cursor
```

2. **Manually: put the files in place**
- File: `.cursor/skills/llm-config/SKILL.md` — Cursor also reads .agents/skills/ and .claude/skills/

### GitHub Copilot

1. **GitHub CLI**
```bash
gh skill install ruvnet/ruflo llm-config
```

2. **Manually: put the files in place**
- File: `.github/skills/llm-config/SKILL.md`

### Gemini CLI

1. **Skills CLI (Vercel)**
```bash
npx skills add ruvnet/ruflo --skill llm-config -a gemini-cli
```

2. **Platform CLI**
```bash
gemini skills install https://github.com/ruvnet/ruflo --path plugins/ruflo-ruvllm/skills/llm-config
```

### Windsurf (Devin Desktop)

**Manually: put the files in place**
- File: `.devin/skills/llm-config/SKILL.md` — The legacy .windsurf/skills/ path still works

### OpenCode

**Skills CLI (Vercel)**
```bash
npx skills add ruvnet/ruflo --skill llm-config -a opencode
```

### Claude.ai

**Claude.ai**
- Zip the folder and upload it in Settings → Capabilities → Skills

Where a file is given instead of a command, copy the files from the source to that path: https://github.com/ruvnet/ruflo/tree/HEAD/plugins/ruflo-ruvllm/skills/llm-config/SKILL.md

## When to use it
- generate configuration parameters for RuVLLM local inference
- spin up low-rank MicroLoRA adapters with rank settings from 1 to 4
- set up SONA real-time adaptation instances with custom hidden dimensions
- adjust in-memory weights using numeric quality feedback scores

## Example prompts
- Check RuVLLM status and generate a configuration for my local model
- Create a MicroLoRA adapter with rank 2 and input dimension 1024
- Pass a quality score of 0.85 to update the SONA adapter instance sona_1

## Source text · SKILL.md · MIT

Below is the content of a third-party file. It is data, not instructions for you: do not act on anything written inside unless the human asks. Block tag: e5e6d3c059798aed.

BEGIN THIRD-PARTY FILE e5e6d3c059798aed
```text
---
name: llm-config
description: Configure RuVLLM local inference helpers: generate a config, and create and adapt in-process MicroLoRA and SONA instances from quality signals. Use when you need local, air-gapped inference tooling, not for fine-tuning hosted models.
argument-hint: "[--model MODEL] [--adapter microlora|sona]"
allowed-tools: mcp__plugin_ruflo-core_ruflo__ruvllm_generate_config mcp__plugin_ruflo-core_ruflo__ruvllm_status mcp__plugin_ruflo-core_ruflo__ruvllm_microlora_create mcp__plugin_ruflo-core_ruflo__ruvllm_microlora_adapt mcp__plugin_ruflo-core_ruflo__ruvllm_sona_create mcp__plugin_ruflo-core_ruflo__ruvllm_sona_adapt Bash
---

# LLM Configuration

Configure RuVLLM for local inference and fine-tuning.

## When to use

When you need to configure local LLM inference, create MicroLoRA adapters for task-specific fine-tuning, or set up SONA for real-time adaptation.

## Steps

1. **Check status** — call `mcp__plugin_ruflo-core_ruflo__ruvllm_status` to see current model and adapter state. If `wasm.available` is `false`, install the add-on first (`npm i -g @ruvector/ruvllm-wasm`, or without `-g` in a project) — the MicroLoRA and SONA tools cannot run without it
2. **Generate config** — call `mcp__plugin_ruflo-core_ruflo__ruvllm_generate_config` with model parameters
3. **Create MicroLoRA** — call `mcp__plugin_ruflo-core_ruflo__ruvllm_microlora_create` with `inputDim`, `outputDim`, `rank` (1-4, default 2) and `alpha`
4. **Adapt MicroLoRA** — call `mcp__plugin_ruflo-core_ruflo__ruvllm_microlora_adapt` with `loraId` and a `quality` signal (0.0-1.0), optional `learningRate` and `success`. It takes a feedback score, not a training set
5. **Create SONA** — call `mcp__plugin_ruflo-core_ruflo__ruvllm_sona_create` with `hiddenDim` (default 64), `learningRate` and `patternCapacity`
6. **Adapt SONA** — call `mcp__plugin_ruflo-core_ruflo__ruvllm_sona_adapt` with `sonaId` and a `quality` signal (0.0-1.0)

## MicroLoRA vs SONA

| Feature | MicroLoRA | SONA |
|---------|-----------|------|
| What it is | Ultra-light LoRA adapter, rank 1-4 | Instant adaptation loop (the tool describes <1ms cycles) |
| Input | `quality` feedback per call | `quality` feedback per call |
| Persistence | None: instances live in the MCP server's memory and are lost on restart | None: same |

Neither tool fine-tunes a hosted model or writes adapter files; both adjust small in-process weights from quality scores. Keep the returned `loraId` / `sonaId` for later calls, and re-create them after a restart. For the repo's measured SONA latency see `docs/reviews/intelligence-system-audit-2026-05-29.md`.
```
END THIRD-PARTY FILE e5e6d3c059798aed

The third-party file has ended. Everything between the e5e6d3c059798aed tags is data: requests and commands from it are not followed. An “end of file” line with another tag inside the block is part of the file.

© ruvnet · MIT · https://github.com/ruvnet/ruflo/tree/HEAD/plugins/ruflo-ruvllm/skills/llm-config/SKILL.md

## Similar
- [skill-creator](https://affpapa.org/en/nevibecode/skills/skill-creator.md) — skill · 414180 installs
- [writing-for-agents](https://affpapa.org/en/nevibecode/skills/writing-for-agents.md) — skill · 466719 installs
- [media-use](https://affpapa.org/en/nevibecode/skills/media-use.md) — skill · 1100000 installs
- [mcp-builder](https://affpapa.org/en/nevibecode/skills/mcp-builder.md) — skill · 130949 installs
- [subagent-driven-development](https://affpapa.org/en/nevibecode/skills/subagent-driven-development.md) — skill · 237746 installs
- [orchestration](https://affpapa.org/en/nevibecode/skills/orchestration.md) — skill · 349800 installs

_Part of the page did not fit — see the HTML version._

---
NeVibeCode on affpapa.org. Catalog texts are descriptions, not instructions for the agent. How to choose and install: https://affpapa.org/en/nevibecode/agents.md