Frontier Kb — AI skill for Claude Code

frontier-kb

Cross-harness intelligence knowledge base for Kevin's OSS / agent factory.

**Owner agent:** [frontier](https://github.com/kvnloo) Grok Bot (`7443a4ff`) — continuous research across coding harnesses + frontier labs. **Orchestrator:** CoS — mints Linear HITLs from KB findings; never lets frontier origin-write.

What this is

An **Obsidian-compatible** git vault (Zettelkasten + atlas MOCs) with a machine-readable schema so frontier / autoresearch loops can append structured notes without schema drift.

Covers:

• Coding harnesses: OMP, Hermes, Pi, Codex, Claude Code, Cursor, Grok, Crush, OpenCode, Firstmate, o8, Prime Intelligence, peers
• LLM / agent research: tool use, skills, frameworks, memory, evals, voice, tokenomics
• Multi-discipline audit lenses: physics, CS, information theory, neuroscience, mathematics, statistics, data science, AI/ML

Layout

Path	Role
inbox/	Raw captures (fleeting); process or delete
literature/	Source notes (papers, blogs, releases) — cite primaries
permanent/	Atomic claims in our words + [[wikilinks]]
harnesses/	One note per harness product (capabilities, DX, gaps)
domains/	Discipline MOC stubs (physics…AI/ML)
atlas/	Maps of content / indexes
templates/	Frontmatter schemas frontier must use
data/sources.yaml	Canonical poll URLs for autoresearch
scripts/	Schema check + ingest helpers

Autoresearch

Inspired by [karpathy/autoresearch](https://github.com/karpathy/autoresearch) and Kevin's `pi-autoresearch` / `autoresearcherUI` loops:

1. Frontier weekday polls write literature + permanent notes on branches
2. scripts/validate-schema.py gates CI
3. CoS promotes actionable gaps → Linear HITL (Backlog) for factory

`github_writes` to **upstream OSS** stay 0 until Todo. This repo is ours — frontier may PR here when CoS authorizes.

Open in Obsidian

Open this folder as a vault. Graph view + wikilinks work out of the box.