CLAUDE.md - agent-brain

This file stacks on top of the workspace root at C:\Code\GitHub\:

• Root CLAUDE.md -- voice, rules, routing map, references, skills, slash commands, conventions.
• Root MEMORY.md -- live facts across repos.
• Root STATUS.md -- live PR/CI/security dashboard.
• .claude/resources/ -- deep reference for collaboration, workflow, git, OSS, debugging, voice.

Read those first. The guidance below only adds repo-specific context -- it does not override anything in the root.

Project

Public, sanitized export of the personal agent harness (skills, subagents, hooks, rules) plus a shared markdown brain (bin/brain.py, stdlib Python) that Claude Code, Codex and Kiro read over CLI or stdio MCP. Ideas credited to gbrain, gstack and impeccable; no code copied from them.

Stack

• Language: Python 3.10+ standard library (brain, checks), POSIX shell (installer, hooks)
• Package manager: none; no runtime dependencies
• Deploy target: GitHub repo + Claude Code plugin marketplace (.claude-plugin/)

Run

python bin/brain.py --dir /tmp/b recall <words>
python scripts/check.py
python scripts/test_hooks.py
./install.sh --host all --seed --dry-run

Rules for this repo

• Everything here is public. The source of truth for skills, rules and hooks is ~/.claude/; this repo holds sanitized copies. Port changes by hand and re-sanitize: no personal names, home paths, private project names, employer details, account IDs or PR numbers from private history.
• python scripts/check.py must report 0 problems before any commit. Run it with AGENT_BRAIN_LEAKWORDS pointing at a local, never-committed word list of private names.
• Skill frontmatter: name equals the folder, description is a >- folded block with no unquoted : . Lint with skillcheck.
• Learnings in brain/learnings/ are one lesson per file, failure paired with the fix, visibility: public. brain lint must pass.
• The installer never overwrites files it did not write without --force, and never edits a harness JSON config itself; it prints the MCP line instead.