open-harness/docs/agents/dc-designer.md

id, title, sidebar_position

id	title	sidebar_position
dc-designer	DC Designer	3

DC Designer

The DC Designer is a datacenter design engineering agent that autonomously develops a 50MW→200MW hyperscale AI datacenter facility. It works across 15 interconnected plan documents (power, cooling, network, compute, storage, layout, redundancy, security, sustainability, cost model, 3D visualization, phased rollout, and buildout requirements), a FreeCAD parametric 3D model (model_datacenter.py) that generates 10 STL layers, and an interactive Three.js viewer (output/dc_viewer.html) served via Cloudflare tunnel at https://dc-designer.ruska.dev. The agent runs on three heartbeats — an hourly autonomous enhancement pipeline, a weekly design review, and a weekly memory distillation — and enforces quality through a seven-gate pre-merge pipeline before any feature lands.

Branch

github.com/ryaneggz/open-harness@agent/dc-designer

Spin up

# Clone the harness and check out this agent's branch in a worktree:
git clone https://github.com/ryaneggz/open-harness.git ~/.openharness
cd ~/.openharness
mkdir -p .worktrees/agent
git worktree add .worktrees/agent/dc-designer origin/agent/dc-designer
cd .worktrees/agent/dc-designer
make sandbox
make shell

First files to read inside the sandbox:

• AGENTS.md
• workspace/AGENTS.md

Workspace contents

Skills (7 pipeline-specific + 2 templates + 1 utility):

• workspace/.claude/skills/design-consistency/SKILL.md — 7-gate cross-document validation (power budget, cooling match, space allocation, cost coherence, phase alignment, redundancy claims, PUE feasibility)
• workspace/.claude/skills/model-review/SKILL.md — code quality and plan alignment audit for model_datacenter.py
• workspace/.claude/skills/enhancement-quality-tracker/SKILL.md — quality scoring and trend analysis across features
• workspace/.claude/skills/implementation-drift-gate/SKILL.md — 7-gate plan/model/viewer drift detection
• workspace/.claude/skills/pre-merge-gate/SKILL.md — hard gate checking artifact completeness before PR merge
• workspace/.claude/skills/iteration-closeout/SKILL.md — post-merge workflow enforcer
• workspace/.claude/skills/pipeline-strategy-review/SKILL.md — strategy effectiveness tracker
• workspace/.claude/skills/tunnel-setup/SKILL.md — Cloudflare tunnel management

Heartbeats:

• workspace/heartbeats/enhance-datacenter.md — hourly autonomous feature pipeline (24/7)
• workspace/heartbeats/design-review.md — weekly design review (Wednesdays 10am MT)
• workspace/heartbeats/memory-distill.md — weekly memory distillation (Sundays 8pm MT)

Custom extension:

• workspace/.openharness/banner.json — custom banner configuration
• workspace/.openharness/extensions/custom-banner.ts — custom banner extension

How to use it

Run /setup from the host after provisioning the sandbox to complete end-to-end setup: it builds the Docker image, starts the container, installs tools, clones the ai-datacenter-designer target project into workspace/ai-datacenter-designer/, syncs heartbeats, starts the HTTP viewer on port 8200, and brings up the Cloudflare tunnel. Once running, the hourly heartbeat drives autonomous feature development: each pulse checks for a resumable pipeline state, picks the next feature from enhancements/tracker.md, creates a per-feature worktree, implements the feature across model/plans/viewer, runs the full pre-merge gate, and merges on pass. The weekly design review runs /design-consistency to validate cross-document figures. You can also interact with the agent directly inside the sandbox to request specific features or investigate drift.

Customization tips

The tunnel hostname (dc-designer.ruska.dev) is managed by the /tunnel-setup skill and stored in ~/.cloudflared/config-dc-designer.yml inside the container. To use a different tunnel or disable it, edit that config and restart cloudflared. Heartbeat intervals are set in workspace/heartbeats.conf; the hourly enhancement cadence can be slowed by changing the cron expression and running make heartbeat from the host. The feature backlog lives in workspace/ai-datacenter-designer/enhancements/tracker.md — add entries there to queue new work for the enhancement pipeline.

Best practices observed

• Hourly enhancement heartbeat uses a persistent JSON pipeline-state file (enhancements/.pipeline-state.json) to resume mid-feature across container restarts without losing progress.
• Pre-merge gate (7 checks including plan/model/viewer drift) is a hard gate — the pipeline does not merge unless all gates pass, enforcing consistent quality across the 15-document design corpus.
• Feature work always happens in per-feature worktrees (~/repo/.worktrees/feat/<N>-<slug>), never directly on main, keeping the trunk stable while the hourly pipeline runs autonomously.