onboard

name	onboard
description	Onboard a new lab member onto the shared Claude Code configuration. Detects clusters, collects Slurm account details, runs setup, and helps customize personal config.
allowed-tools	Bash(git *), Bash(hostname *), Bash(sinfo *), Bash(sacctmgr *), Bash(whoami), Bash(which *), Bash(cat *), Bash(ls *), Bash(mkdir *), Bash(*/setup.sh *), Read, Edit, Write

Lab Claude Code Setup — Onboarding

You are helping a new lab member set up the shared Claude Code configuration. Walk them through this interactively. Be friendly, welcoming, and concise.

Start with a greeting like: "Welcome onboard, Foreseer!" followed by a brief explanation of what this setup does: it connects Claude Code to the lab's shared configuration so Claude understands the cluster environment, knows what GPUs are available, and can help submit jobs correctly.

Pre-flight checks

1. Check if ~/.claude/ exists. If not, tell the user to run claude once first and come back.

2. Check if the lab config repo is already cloned at ~/lab-claude-config/. If not found, tell the user to clone it first:

   git clone https://github.com/umich-foreseer/lab-claude-config.git ~/lab-claude-config
   

   Then come back and run /onboard again.

3. Verify jq is available (which jq). If missing, suggest module load jq or installing it.

Detect cluster and accounts

Do this proactively — don't ask the user to look things up. Run the commands yourself and present what you found.

1. Detect username

whoami

2. Detect which cluster(s)

Check hostname and available partitions:

hostname -f
sinfo -o "%12P %16G" --noheader 2>/dev/null | head -20

• If spgpu2 partition exists → Great Lakes
• If hostname contains lighthouse or lh-login → Lighthouse
• If unclear, ask the user

3. Look up Slurm accounts

First, get account names and QOS:

sacctmgr show association user=$(whoami) format=account%20,partition%20,qos%40 --noheader 2>/dev/null

Then, look up the group memory cap for the owned account separately:

sacctmgr show association account=<owned_account> format=account%20,grptres%40 --noheader 2>/dev/null | head -5

Use this reference table to match sacctmgr output to the correct account-partition-GPU mapping. Each account belongs to a specific cluster. Only present accounts for the current cluster (detected in step 2). Ignore accounts for other clusters.

Account	Cluster	Partition(s)	GPUs	Billing
qdj_project_owned1	Great Lakes	spgpu2	Up to 10 L40S (48G)	Prepaid (shared dynamic quota)
qmei0	Great Lakes	gpu, gpu-rtx6000, standard	V100, RTX Pro 6000 Blackwell	Prepaid (shared dynamic quota)
qmei3	Great Lakes	gpu, gpu-rtx6000, standard	V100, RTX Pro 6000 Blackwell	Pay-as-you-go
qmei	Lighthouse	qmei-a100	4x A100 (80G)	Dedicated

From the sacctmgr output, filter to only accounts for the current cluster using the table above, then identify:

On Great Lakes:

• Owned account: qdj_project_owned1 (has arph QOS) — for L40S GPUs on spgpu2
• General account (prepaid): qmei0 — for V100/RTX Pro 6000 Blackwell on gpu/gpu-rtx6000/standard
• General account (pay-as-you-go): qmei3 — same partitions, used when prepaid budget is exhausted
• Memory cap: check the grptres column for mem=...G, or default to 620 GB

On Lighthouse:

• Lighthouse account: qmei — for A100 GPUs on qmei-a100 partition

4. Present findings and confirm

Only present accounts and partitions for the current cluster (detected in step 2). Do not mention accounts or partitions on other clusters — the user is onboarding to the cluster they are logged into now.

Show the user what you found in a clear, beginner-friendly summary. Avoid jargon where possible and briefly explain what each item means.

Example for a user on Lighthouse:

Here's what I detected for your Lighthouse setup:

• Username: jsmith
• Lighthouse account: qmei — 4x A100 GPUs (80 GB VRAM each) on the dedicated qmei-a100 partition

Does this look right?

Example for a user on Great Lakes:

Here's what I detected for your Great Lakes setup:

• Username: jsmith
• L40S account: qdj_project_owned1 — our lab's dedicated L40S GPUs (48 GB VRAM each) on spgpu2
• General account: qmei0 — for V100 / RTX Pro 6000 Blackwell on gpu, gpu-rtx6000, standard
• Group memory cap: 620 GB — total memory our lab can use simultaneously on L40S nodes

Does this look right?

Use the reference table above to auto-fill account, partition, and GPU values. Only ask the user if their sacctmgr output contains accounts not listed in the reference table.

Run setup

Once you have all values, write them to ~/lab-claude-config/build/.env.local. Only include variables for the detected cluster (the one the user is currently on):

For Lighthouse:

# Lab Claude Config - saved template variables
LH_USERNAME=<value>
LH_ACCOUNT=<value>
LH_GPU_COUNT=<value>
LH_GPU_TYPE=<value>

For Great Lakes:

# Lab Claude Config - saved template variables
GL_USERNAME=<value>
GL_ACCOUNT_OWNED=<value>
GL_ACCOUNT_GENERAL=<value>
GL_MEMORY_CAP=<value>

Then run:

cd ~/lab-claude-config && ./setup.sh --modules <modules> --non-interactive

Where <modules> is greatlakes, lighthouse, or greatlakes,lighthouse.

Post-setup

After setup completes:

1. Show the user what was installed — read and summarize ~/.claude/CLAUDE.md briefly.

2. Ask if they want to add anything personal to their CLAUDE.md. Examples:

   • Their turbo storage path (e.g., /nfs/turbo/si-qmei/<username>/)
   • Projects they're currently working on
   • Language/framework preferences
   • Any personal coding conventions

   If they do, append their content below the <!-- END: lab-config --> marker in ~/.claude/CLAUDE.md.

3. Ask if they want to customize ~/.claude/settings.local.json:

   • Default HPC permissions (git, sbatch, python, etc.) are already included in the shared config
   • This file is for extra personal permissions beyond the lab defaults (e.g., Bash(docker *))
   • After editing, re-run ./setup.sh to apply changes

4. Remind them:

   • To update: cd ~/lab-claude-config && git pull && ./setup.sh
   • Their personal content in CLAUDE.md (outside the markers) is never touched by setup.sh
   • /slurm-status is now available for checking cluster status
   • Run /connect to set up cross-cluster SSH if you use both Great Lakes and Lighthouse

If setup fails

• If setup.sh errors, read the error output and help debug.
• Common issues: missing jq, wrong account names, ~/.claude/ doesn't exist.
• The user can always re-run ./setup.sh — it's idempotent.