devstack

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

devstack merges four popular Claude Code add-ons into one plugin and switches off everything they duplicate:

• Hindsight: long-term agent memory
• Context+: semantic code intelligence
• graph-engineering: knowledge-graph and task-graph methodology
• oh-my-openagent: the ulw / mass ulw workflow

Installed side by side, these tools keep three kinds of memory and two ways to undo, and each ships instructions telling the agent to prefer itself. devstack picks one owner for each job and removes the rest.

Documentation · Installation · How it works · Workflow · What's inside · Philosophy

---

How it works

Capability	Owner	Switched off
Memory across sessions and branches	Hindsight (one bank per repo, local daemon by default)	Context+ memory-graph tools · OmO's Kibitzer
Live code structure, semantic search, blast radius	Context+ (6 tools)	Hindsight's codebase survey and "Component map" page
Writing code and undoing it	Edit / Write + git	Context+ propose_commit, list_restore_points, undo_change
Lint, types, tests	your repo's own scripts	Context+ run_static_analysis
Feature-to-code maps	Hindsight knowledge pages	Context+ get_feature_hub
One disciplined change (ulw)	ultrawork skill	OmO's harness
A large job across parallel agents (mass ulw)	ultrawork-graph skill + scout / worker / verifier agents	OmO's harness
Building a knowledge graph as a product	graph-engineering skill + nine /devstack:kg-* commands	—

Four mechanisms enforce the table:

1. Deny rules in ~/.claude/settings.json remove the 11 overlapping Context+ tools from Claude's context.
2. Hindsight config turns off the features Context+ already covers.
3. Context+'s bundled server instructions are deleted at install. They order the agent to use the very tools devstack switches off.
4. Two hooks: one injects the routing table at session start, and one maps the ulw / mass ulw keywords to skills.

Installation

Requirements: Claude Code, Node.js 18+, git, uv (for Hindsight's local daemon), Ollama (for Context+ embeddings).

git clone https://github.com/codicovibes/devstack.git
cd devstack
./install.sh

Run it from your own terminal, not from inside a Claude Code session. It's idempotent, so re-run it to update. To keep memory on a server instead of this machine:

HINDSIGHT_SERVER=self-hosted HINDSIGHT_API_URL=http://localhost:8888 ./install.sh
HINDSIGHT_SERVER=cloud HINDSIGHT_API_TOKEN=<token> ./install.sh

Verify: start a new session and run /mcp. You should see plugin:devstack:contextplus with 6 tools and hindsight connected.

Uninstall: ./uninstall.sh (add --purge-hindsight to also remove Hindsight's hooks).

The basic workflow

1. Work normally. Hindsight injects what the repo remembers (decisions, conventions, earlier sessions) on your first prompt, and writes each session back when it ends.
2. ulw <task> for a change that must be right. The agent reads before it edits, reproduces the bug or states the acceptance check, checks the blast radius, makes the smallest change, and proves it with commands run in the session.
3. mass ulw <job> for work that splits. The agent first checks that the work really is parallel; if it isn't, it falls back to ulw. Then it draws a task graph, runs Haiku scouts and Sonnet workers (each in its own git worktree), has a separate verifier check every node, and merges once.
4. /devstack:kg-tutor to learn knowledge-graph engineering. For a real build, run /devstack:kg-scope → kg-schema → kg-extract → … → kg-rag.

What's inside

Skills

• ultrawork: one change, evidence at every step. It defers to your own implement-change skill if one is installed.
• ultrawork-graph: parallel fan-out with the stop rule, separate verifiers and a single merge.
• graph-engineering: the 9-stage knowledge-graph pipeline, plus the task-graph rules.

Commands: kg-tutor, kg-scope, kg-schema, kg-extract, kg-relations, kg-events, kg-fuse, kg-eval, kg-rag.

Agents

• scout (Haiku, read-only)
• worker (Sonnet, git worktree)
• verifier (session model, read-only)

MCP server: Context+, pinned and launched from ~/.local/share/devstack/contextplus.

Hooks: SessionStart injects the routing table. UserPromptSubmit maps ulw and mass ulw to their skills.

Philosophy

• One owner per capability. Two tools holding the same kind of state drift apart. Choose one and remove the other.
• The shape of the work decides the team. Parallel agents win on work that splits and lose on sequential work.
• Evidence over claims. A worker's report of a passing test is a claim. The verifier reruns it.
• Memory is a past record. When memory and code disagree, the code wins.

Documentation

Full docs are at codicovibes.github.io/devstack: installation, how the ownership map is enforced, usage, configuration and troubleshooting.

Credits

• graph-engineering (MIT) is vendored in plugins/devstack/skills/graph-engineering with its licence. Only its description is changed. It distils Southeast University's Knowledge Graph course.
• The ulw / mass ulw keywords and the fan-out idea come from oh-my-openagent. devstack reimplements them and copies none of its code or prompts.
• Context+ (MIT) and Hindsight are installed from npm, not vendored.

License

MIT © codicoscepticos