nebari-pi-pack

Helm chart that extends Nebari's data-science JupyterHub stack with a Pi coding-agent workflow.

What this chart adds

• Pi launcher service in Hub (/services/pi-launcher/, port 10300)
• Pi named-server profiles (pi-small, pi-medium, pi-large)
• Pi home/quick-access UI customizations
• Optional Pi session sharing RBAC + handlers
• Optional Pi Session Viewer service (/services/pi-session-viewer/, port 10400)
• Optional relay subsystem (core + dummy/slack/whatsapp adapters)
• Shared skills delivered from the Pi runtime image (no PVC sync job)

Architecture

• JupyterHub remains the control plane.
• User requests Pi via launcher.
• JupyterHub spawns/controls named server pi per user.
• Pi pod serves browser terminal via jhsingle_native_proxy + ttyd + pi CLI.
• Pi startup includes --skill /opt/nebari/baked-skills/shared-skills.
• Pi startup loads the OpenAI Codex device-code OAuth override extension (-e /opt/nebari/extensions/openai-codex-device-auth.ts).
• Pi startup loads session sharing commands extension (-e /opt/nebari/extensions/session-share.ts).
• Pi startup also loads pi-self-learning extension (-e /usr/local/lib/node_modules/pi-self-learning/extensions/self-learning.ts).

Install

helm dependency update .
helm upgrade --install data-science-pack . \
  --namespace data-science \
  --wait --timeout 40m

Example values

# Local baseline
helm upgrade --install data-science-pack . \
  --namespace data-science \
  -f examples/values-local.yaml \
  --wait --timeout 40m

# Local reliable
helm upgrade --install data-science-pack . \
  --namespace data-science \
  -f examples/values-local-reliable.yaml \
  --wait --timeout 45m

# Relay profile
helm upgrade --install data-science-pack . \
  --namespace data-science \
  -f examples/values-relay.yaml \
  --wait --timeout 50m

# Production-style baked skills profile
helm upgrade --install data-science-pack . \
  --namespace data-science \
  -f examples/values-baked-skills.yaml \
  --wait --timeout 45m

Important values

• pi.image.*
• pi.sharedSkills.enabled
• pi.sharedSkills.mode (image)
• pi.sharedSkills.imagePath
• jupyterhub.custom.pi-image
• jupyterhub.custom.pi-skills-path
• jupyterhub.custom.pi-coding-agent-dir
• jupyterhub.custom.pi-session-viewer-* (enable/port/S3 settings)
• relay.*

When relay is enabled, keep these aligned:

• relay.core.piAgentDir
• jupyterhub.custom.pi-coding-agent-dir

Build Pi image with baked skills

images/pi-agent/Dockerfile now bakes skills directly from local repository files:

• skills/shared-skills/jhub-deploy/SKILL.md
• skills/shared-skills/observability/SKILL.md
• skills/shared-skills/web-browser/SKILL.md
• skills/shared-skills/gog/SKILL.md
• skills/shared-skills/notion/SKILL.md
• skills/shared-skills/video-frames/SKILL.md
• skills/shared-skills/markdown-converter/SKILL.md

Example:

docker build \
  -f images/pi-agent/Dockerfile \
  -t quay.io/openteams/pi-coding-agent-demo:baked-<tag> \
  .

Then set your Helm values to the pushed image tag/digest:

• pi.image.repository
• pi.image.tag
• jupyterhub.custom.pi-image (prefer digest pin)

Notes

• Legacy PVC-based shared-skills sync is removed from runtime flow.
• Existing running Pi sessions keep old image/skills until pod restart.