devstack

One Claude Code setup: memory, code intelligence, knowledge graphs and parallel agents, with one owner per capability.

Why devstack exists

Four popular Claude Code add-ons each solve a real problem:

Project	What it brings
Hindsight	Long-term memory built from git history and past sessions
Context+	Semantic search, AST structure and blast radius over a live codebase
graph-engineering	A 9-stage knowledge-graph pipeline and task-graph design rules
oh-my-openagent	The `ulw` / `mass ulw` “just finish it” workflow

Installed together, they overlap:

three kinds of memory: Hindsight banks, Context+’s memory graph and OmO’s Kibitzer
two ways to write and undo code
two descriptions of the codebase’s structure
server instructions in which each tool tells the agent to prefer itself

The agent ends up spreading the same facts across stores that drift apart.

devstack keeps each tool’s unique part, picks a single owner for everything they share, and switches the duplicates off. See How it works for the full ownership map.

At a glance

ulw fix the off-by-one in pagination       → ultrawork: one change, proven
mass ulw migrate every package to v2 API   → ultrawork-graph: scouts, workers, verifiers, one merge
/devstack:kg-tutor                         → learn knowledge-graph engineering, module by module

Hindsight works in the background: it injects what the repo remembers on your first prompt and writes each session back when it ends.