A portable, version-controlled configuration for
Claude Code: a crew of specialised agents, a
memory that grows from finished work, and a learning loop. install.py
symlinks agents/, skills/, memory/, commands/ and hooks/ into
~/.claude, then asks how to handle settings.json and CLAUDE.md.
Two things it does that most agent setups do not:
Principled refusal.“I don't know / no competent path” is a
first-class output. The orchestrator will not relay a sub-answer it trusts
less than its own read.
A built-in devil's advocate.critic red-teams a plan before it is
built, in one bounded pass, and is free to conclude that the plan holds.
Ten agents, each with a narrow remit. Delegation goes one way — down — and
every result comes back through a routing quality gate rather than straight to
you. The full roster with descriptions is on Agents.
A Stop hook records what actually happened in a session, locally and
uncurated. /harvest-memory is what turns that trail into durable notes under
memory/ — a step you run deliberately, not a background process. Nothing
reaches memory without you.