Getting Started

You need Claude Code installed and running.

# Install Claude Code (if not already)
curl -fsSL https://claude.ai/install.sh | bash
claude --version

macOS / Linux / WSL. For Windows: irm https://claude.ai/install.ps1 | iex (PowerShell). Documentation →

git clone https://github.com/Arcanada-one/datarim.git
cd datarim
./install.sh --with-claude              # Claude Code (symlink, default)
./install.sh --with-codex               # Codex CLI (multi-runtime, v2.0+)
./install.sh --with-claude --with-codex # both runtimes at once

The installer creates symlinks ~/.claude/{skills,agents,commands,templates} → code/datarim/<scope>/. Runtime IS the repo: editing any file under ~/.claude/ shows up immediately as git diff — no drift, no manual sync.

Existing copy-mode install? The installer detects it and offers conversion (3 options: convert / keep / abort) with auto-backup.

Personal additions / overrides go to gitignored overlay ~/.claude/local/{skills,agents,commands,templates}/. Local wins on name conflict.

Already have Datarim installed? Update to the latest version:

cd /path/to/datarim
./update.sh                      # topology-aware: symlink → git pull, copy → install --copy --force --yes

Navigate to any project and start Claude Code:

cd /path/to/your/project
claude

# Inside Claude Code:
/dr-init Add dark mode toggle to settings page

Datarim analyzes your request, determines complexity (L1-L4), and creates task tracking in datarim/.

Since v2.7.0: Step 2.5b adds a soft advisory when your description overlaps in topic with an item already pending in the backlog, so you do not spawn a second ID for the same work. Non-blocking — you decide duplicate / refine-scope / orthogonal before continuing.

Not every task needs all 8 stages. Datarim routes by complexity:

After /dr-init, your project gets:

your-project/
├── datarim/                 # workflow state (add to .gitignore)
│   ├── tasks.md             # active tasks
│   ├── backlog.md           # pending items
│   ├── activeContext.md     # current focus
│   ├── progress.md          # status overview
│   ├── insights/            # research findings
│   └── ...                  # prd/, creative/, reflection/
└── documentation/
    └── archive/             # completed tasks (commit this)
        ├── web/
        ├── infrastructure/
        └── ...

datarim/ is your local workflow — add to .gitignore. documentation/archive/ is project history — commit to git.

Init-task file and Q&A log (v2.9.0): /dr-init stores your original prompt in datarim/tasks/{ID}-init-task.md verbatim. Every later clarification — the agent's question and your answer — is auto-appended to the file's Append-log as a structured «Q&A by /dr-<stage>» block. If you do not answer, the agent decides autonomously by best practices and records a rationale of at least 50 characters; the review stage verifies every autonomous decision against the implementation.

Extend Datarim with opt-in plugins via /dr-plugin. Each plugin is a directory with plugin.yaml + {skills,agents,commands,templates}/ subdirs. Enabling creates symlinks under ~/.claude/<cat>/<plugin-id>/.

/dr-plugin list                              # active set
/dr-plugin enable /path/to/my-plugin     # enable a local plugin
/dr-plugin disable my-plugin              # disable
/dr-plugin sync                              # reconcile runtime ↔ manifest
/dr-plugin doctor [--fix]                    # 9 health checks

Active-set manifest: datarim/enabled-plugins.md. Pre-mutation snapshot/rollback on every enable. Full reference — command page →

Secure-by-default contract for network bind targets: docker-compose ports, redis.conf bind, postgresql.conf listen_addresses, systemd .socket. Public-by-default = breach-by-default; the new network-exposure-baseline.md skill formalises the tier model with TTL-bounded waivers.

Local pre-commit check:

dev-tools/network-exposure-check.sh \
  --compose docker-compose.yml \
  --redis-conf redis.conf \
  --postgres-conf postgresql.conf
# exit 0 — clean; exit 1 — violation

Reusable CI workflow in .github/workflows/ci.yml:

jobs:
  network-exposure:
    uses: Arcanada-one/datarim/.github/workflows/network-exposure-lint.yml@main
    with:
      compose-paths: 'docker-compose.yml,docker-compose.prod.yml'

Pipeline commands (/dr-prd, /dr-plan, /dr-do, /dr-archive) auto-load the skill on network-surface changes. Tiered gate: P0 / security / infrastructure / framework-hardening tasks → hard_block; otherwise → advisory_warn with operator override.

When operator requirements are incomplete, complex, or ambiguous, the agent must not stall and escalate unnecessarily. Eight canonical rules (FB-1..FB-8) define autonomous behaviour under Supreme Directive Laws 1-5. The mandate lives in the consumer's ecosystem CLAUDE.md (canonical text + audit tagging); Datarim ships only the machine-readable contract block.

Machine-readable policy block:

plugins/dr-orchestrate/rules/fb-rules.yaml
# rules: 8 × {rule_id, enforcement_layer, tier, default_action,
#               reversibility_required, audit_required, conflicts_with_law}
# hard_gated_actions: [prod_deploy, secret_rotation, irreversible_db_op,
#                      public_communication, finance_action, legal_action,
#                      force_push_main, git_history_delete, multi_user_visible_action]

Hierarchy: Supreme Directive (Laws 1-5) > Autonomous Agent Operating Rules > AAL Mandate > project-specific mandates. hard_gated_actions NEVER auto-execute regardless of FB-5. Loader: scripts/rules_loader.sh § load_fb_policy() / load_fb_hard_gates().