LLM-agnostic BRAIN_HARNESS.md — decouple from Claude-specific files

[0SxD]

Context

OpenBrainLM is designed to be LLM-agnostic (bridge.py = pluggable backends).
But the harness files are Claude-specific (CLAUDE.md, hookify rules).
Need a neutral BRAIN_HARNESS.md that any LLM reads, with LLM-specific files inheriting from it.

Mapping

• BRAIN_HARNESS.md → universal (rules, memory protocol, research protocol)
• CLAUDE.md → Claude-specific (hooks, tools, plugins)
• GEMINI.md → Gemini-specific (hooks, tools, extensions)

Why now

Multiple LLM backends being used in parallel.
Both need the same brain harness.

[0SxD]

Expanded scope — every LLM gets its own adapter file:

LLM	Adapter File	Reads on boot
Claude Code	CLAUDE.md	Auto-loaded
Gemini CLI	GEMINI.md	Auto-loaded
Codex (OpenAI)	CODEX.md	Auto-loaded
Any future LLM	[LLM_NAME].md	Must read BRAIN_HARNESS.md first

Architecture:

BRAIN_HARNESS.md          ← universal (rules, memory protocol, research protocol, agent patterns)
├── CLAUDE.md             ← Claude-specific (hooks, hookify, MCP tools, plugins)
├── GEMINI.md             ← Gemini-specific (extensions, tools, Antigravity integration)
├── CODEX.md              ← Codex-specific (sandbox, tools, exec model)
└── [FUTURE_LLM].md      ← template provided

BRAIN_HARNESS.md must include:

• "Any LLM wearing this harness MUST create its own [LLM_NAME].md adapter file"
• "The adapter file documents: tool names, hook mechanisms, permission models, and any LLM-specific behaviors"
• "The adapter inherits ALL rules from BRAIN_HARNESS.md — it only ADDS LLM-specific details, never overrides"

This makes OpenBrainLM truly LLM-agnostic. The brain is the harness. The LLM is the muscle.