Replicating the orchestrata-ultimate environment

Everything lives in the repo now (SPEC-108). Two installers, both tested end-to-end on clean Ubuntu 24.04 (2026-07-21):

Quick path

git clone git@gitlab.com:MSamy/orchestrata-ultimate-architecture.git && cd orchestrata-ultimate-architecture
bash claude-env/install.sh              # ~/.claude: session CLI, hooks, commands, skills, settings
bash claude-env/install-infra.sh        # detects incus AND docker; you choose the backend
uv sync --dev                           # Python 3.14t + patched janus-swi + FT grpcio
bash scripts/hooks/install-git-hooks.sh
cp .env.example .env                    # or use the generated .env.docker
git clone git@gitlab.com:MSamy/knowledge-wiki.git ~/workspace/knowledge-wiki
python scripts/bootstrap_graphs.py --all   # seed canonical constraints (incl. per-worker DBs), AST graph, Memgraph content
python scripts/run_tests.py             # kernel-verified full suite

Install flow

What each installer does

• claude-env/install.sh — replicates the global ~/.claude layer (session CLI + package, autoformat/session/remember hooks, 14 slash commands, session skill, settings). Idempotent; timestamped backups; never clobbers a differing settings.json (writes settings.json.new); installs jinja2 pydantic neo4j PEP 668-aware. Fails fast (before any mutation) on missing rsync cmp date git python3.

• claude-env/install-infra.sh — reports BOTH backends (present/absent), asks which to use (--backend ... --yes for automation), installs Terraform either way, and finishes by installing the canonical .env pointed at the machine's own fleet:

  • incus (raw machine): 8-node public tier via infra/rawmachine terraform + infra/cloud-init provisioning on images:debian/12/cloud — redis, neo4j primary + 3 worker replicas, 3 prolog/app nodes, all systemd-supervised and restart-durable, pinned IPs on the machine's own bridge (default 10.95.0.x). Verified end-to-end in an ISOLATED VM (host fleet ACL-blocked). Overrides: ORCHESTRATA_SUBNET, ORCHESTRATA_CLOUD_IMAGE, ORCHESTRATA_READY_TIMEOUT. The orchgraph tier (PRIVATE Memgraph Enterprise fork) is added later from the private GitLab release + enroll_orchgraph.py.
  • docker: infra/docker/infra-compose.yml — redis, neo4j primary + 3 workers, Memgraph community 3.11.0 stand-in (no HA); localhost port map (6379, 7687-7690, 7691).

  To run the suite on the fresh fleet afterwards: python scripts/bootstrap_graphs.py --all (the three hard seed pre-conditions — canonical constraints incl. per-worker DBs, AST code-intelligence graph, Memgraph content); everything else self-seeds.

Key artifacts

Gotchas discovered during verification

• Clean Ubuntu ships without rsync, git, and pip — the preflights exist because all three bit during testing.
• The global tools carried 17 PEP 758 except A, B: sites — Python 3.14-only syntax; now parenthesized (works 3.10+).
• session_cli.py init used to require an existing project root (circular); now bootstraps from cwd.
• Docker Hub has no memgraph/memgraph:3.5 — the compose pins 3.11.0 (the orchgraph fork's base version).
• The original cloud-init files were unexercisable: terraform used images:debian/12, which has NO cloud-init — user-data was silently ignored. The /cloud image variant is mandatory.
• A 10GiB VM root exhausts mid-provision (installer now preflights 12GiB); four parallel neo4j tarball pulls need the 1800s readiness gate.
• On constrained networks the images: remote fetch stalls — import the image locally and set ORCHESTRATA_CLOUD_IMAGE to the alias.
• Mix-and-match backends (per-service) is the declared follow-up (REQ-108-013).