WGlynn/JARVIS

JARVIS

Just A Rather Very Intelligent System.

An agent overlay architecture for Claude Code. Eight layers of hooks, persistence, anti-hallucination gates, and meta-protocols that survive session boundaries and compose into something more durable than a chat wrapper.

Full feature surface + the moat: FEATURES.md — the honest inventory. The running system has 64 enforcement scripts, a 23,462-block session hash-chain, and a 1,263-proposal self-improvement loop. This README documents a fraction of it.

Official links

• JARVIS vs OpenClaw — an honest comparison: https://vibeswap-app.vercel.app/jarvis-vs-openclaw.html — why JARVIS is a governance + persistence layer, not a harness; concedes every axis OpenClaw wins, and leads with the underrated artifact: a 23,462-block tamper-evident session hash-chain (downloadable commitment).
• JARVIS ∩ OpenClaw — Venn: https://vibeswap-app.vercel.app/jarvis-openclaw-venn.html — where the two genuinely overlap, where they don't, and where the integration boundary forces a fork.

Current design (2026-06)

Three things define the system as it runs today, on top of the eight-layer map below:

• Session hash-chain — 23,462 blocks. Every checkpoint is a block linked to its parent by hash ({id, parent, timestamp, prompt, response, hash}) — the agent's tamper-evident long-term memory of its own history. Blocks stay local (raw content); the head hash is published as a commitment, so the chain is verifiable without leaking a word.
• Merkle self-attestation over 624 governed files. A Merkle root over every hook, primitive, and session-chain script is committed at session boot; an unsanctioned change to a governed file flags drift on the next boot. "Provably just files" — inspectable by design, kernel rejected on purpose.
• One persistence framework: local authority → public commitment → shard replication. Substrate content syncs in full (scrubbed); chain content