Karpathy LLM & Agent Knowledge Wiki Plugin

The karpathy-llm plugin packages Andrej Karpathy's agent behavioral principles, multi-repo context engineering standards, and the autonomous Karpathy LLM Wiki knowledge architecture for AI coding assistants.

It requires zero external vector databases, embedding pipelines, or heavy MCP infrastructure — knowledge is represented as interlinked, version-controlled markdown pages with typed YAML frontmatter that AI agents navigate natively.

---

What It Provides

1. The 4 Cardinal Principles for Coding Agents (rules/rnex-behavioral.md)

• Think Before Coding: State explicit assumptions, formulate the problem clearly, and evaluate trade-offs before touching code.
• Simplicity First: Minimal abstractions, readable implementations, and zero boilerplate bloat. Avoid premature optimization or speculative generalization.
• Surgical Changes: Minimize blast radius. Modify only what is strictly necessary, preserve existing comments/docstrings, and produce clean, reviewable diffs.
• Goal-Driven Execution: Define upfront verification criteria, write/run automated tests, and review diffs rigorously before completing tasks.

2. Multi-Repo Context Engineering & Autonomy (rules/rnex-behavioral.md)

• Guidelines tailored specifically for multi-repo virtual workspaces:
  • Always read workspace configuration (.local.rnex.yaml and rnex.yaml) first.
  • Respect member repository autonomy (repos/<name>/).
  • Read wiki/hot.md and wiki/index.md once per session for orientation.
  • Perform Git commits directly inside individual member repositories.

3. Autonomous LLM Wiki Architecture

A self-healing, agent-maintained knowledge graph structured as:

• /raw/ Intake Directory:
  • Append-only directory where unmodified source material (articles, transcripts, docs, API references, meeting notes) is deposited.
  • Raw sources are immutable source-of-truth references for citations.
• /wiki/ Curated Knowledge Base:
  • Interlinked atomic markdown pages, each representing a single concept, architectural pattern, or domain decision.
  • Each page contains structured YAML frontmatter defining typed relations:

    ---
    title: Workspace Configuration Protocol
    sources: [raw/rnex-design-spec.md]
    related: [wiki/member-repos.md, wiki/ai-routing.md]
    extends: [wiki/multi-repo-architecture.md]
    contradicts: []
    mentioned_in: []
    ---

• /wiki/index.md Navigation Index:
  • Master categorized catalog organizing all atomic concepts, architecture decisions, and cross-project knowledge.
• /wiki/hot.md Rolling Context Cache:
  • High-density ~500-word orientation summary for rapid agent onboarding without full wiki scans. Agents read this once on startup.
• /wiki/_log.md Audit Trail:
  • Append-only log recording every ingest and lint operation with timestamped details and source attribution.

4. Standardized Agent Workflows

• wiki-ingest.md (8-Step Ingestion Protocol):
  • Decomposes raw source documents from /raw/ into 5–25 atomic wiki pages.
  • Formulates bi-directional frontmatter relations (sources, related, extends, contradicts).
  • Appends operations to wiki/_log.md and rebuilds wiki/index.md.
• wiki-lint.md (10-Step Integrity Audit Protocol):
  • Validates inter-page link paths and file existence.
  • Recomputes reverse backlinks (mentioned_in).
  • Identifies and eliminates orphan pages.
  • Merges duplicate or overlapping pages.
  • Audits logical inconsistencies and flags outdated information.
  • Updates wiki/hot.md and rebuilds wiki/index.md.

5. Specialized Agent (agents/wiki-curator.agent.md)

• A dedicated autonomous curator agent configured to maintain the LLM Wiki graph, execute ingestion protocols, and perform periodic graph integrity audits.

---

Enabling the Plugin

Via CLI

# Enable across the workspace in rnex.yaml
rnex plugin enable karpathy-llm

# Enable locally in .local.rnex.yaml for your workstation only
rnex plugin enable --local karpathy-llm

---

Configuration Options

Configure options under plugins.karpathy-llm in rnex.yaml or .local.rnex.yaml:

plugins:
  karpathy-llm:
    # Directory paths relative to workspace root (defaults shown)
    wiki_dir: wiki              # Directory for curated atomic markdown pages
    raw_dir: raw                # Append-only directory for unmodified intake documents

    # Automated Git hook trigger for wiki verification
    # Options: git-post-merge, git-post-commit, git-pre-commit, git-pre-push, manual
    lint_trigger: git-post-merge

    # Notify developer when changes are committed to raw/ or wiki/
    notify_on_changes: true

    # Enable or disable
    enabled: true

Option	Type	Default	Description
wiki_dir	string	wiki	Directory containing curated atomic markdown pages.
raw_dir	string	raw	Append-only directory for unmodified source intake documents.
lint_trigger	string	git-post-merge	Git hook event that triggers wiki change detection (git-post-merge, git-post-commit, git-pre-commit, git-pre-push, manual).
notify_on_changes	boolean	true	Show notification banner when changes are committed to raw/ or wiki/.
enabled	boolean	true	Enable or disable the plugin. Overridable in .local.rnex.yaml.

---

Automated Git Hooks Integration

When workspace Git hooks are installed (rnex hooks install), the karpathy-llm plugin links directly into Git events:

• post-merge: Detects if incoming commits touched raw/ or wiki/, warning the developer to review or re-lint.
• post-commit: Reminds the developer to run wiki-lint or wiki-ingest if wiki files were committed.
• pre-commit & pre-push: Emits trigger notices if unvalidated wiki files are staged or about to be pushed.

Inspect active trigger status anytime:

rnex hooks status

---

Generated Directory Structure

When karpathy-llm is enabled, it scaffolds:

my-workspace/
├── raw/
│   └── .gitkeep                     # Deposit unmodified source files here
├── wiki/
│   ├── index.md                     # Master navigation catalog
│   ├── hot.md                       # High-density ~500-word orientation summary
│   └── _log.md                      # Append-only audit history
├── .rnex/
│   └── plugins/
│       └── karpathy-llm/
│           ├── rules/
│           │   └── rnex-behavioral.md    # 4 Cardinal principles & context engineering
│           ├── instructions/
│           │   └── rnex-wiki-architecture.md # Schema, relations, and conventions
│           ├── workflows/
│           │   ├── rnex-wiki-ingest.md   # Ingestion checklist
│           │   └── rnex-wiki-lint.md     # Lint & integrity checklist
│           ├── prompts/
│           │   ├── rnex-wiki-ingest.md
│           │   └── rnex-wiki-lint.md
│           ├── skills/
│           │   ├── rnex-wiki-ingest/
│           │   │   └── SKILL.md
│           │   └── rnex-wiki-lint/
│           │       └── SKILL.md
│           ├── agents/
│           │   └── rnex-wiki-curator.agent.md
│           └── templates/
│               ├── rnex-wiki-page.template.md
│               └── rnex-llm-wiki.sample.md
└── rnex.yaml

---

Cross-Plugin Provider Mirroring

Because Repo Nexus features decoupled cross-plugin asset discovery:

• If copilot is enabled, rnex-wiki-ingest and rnex-wiki-lint prompts mirror to .github/prompts/, skills mirror to .github/skills/, and rnex-wiki-curator.agent.md mirrors to .github/agents/.
• If claude is enabled, rnex-wiki-ingest and rnex-wiki-lint mirror to .claude/commands/, .claude/prompts/, and .claude/skills/.
• If gemini, cursor, or windsurf is enabled, prompts, skills, and agents mirror into .gemini/, .cursor/, and .windsurf/.

---

Disabling the Plugin

rnex plugin disable karpathy-llm

When disabled:

• Scoped assets under .rnex/plugins/karpathy-llm/ and member repository mirrors are cleanly removed.
• Your data in raw/ and wiki/ is preserved 100% intact. Repo Nexus never deletes user-generated knowledge bases upon plugin disablement.