truefoundry-notebooks（truefoundry 笔记本）

Routing note: For ambiguous user intents, use the shared clarification templates in references/intent-clarification.md.

Jupyter Notebooks

Launch Jupyter Notebooks on TrueFoundry with GPU support, persistent storage, auto-shutdown, and VS Code integration. Write a YAML manifest and apply with tfy apply. REST API fallback when CLI unavailable.

When to Use

• User asks "launch a notebook", "start jupyter", "create notebook"
• User needs a development environment with GPU access
• User wants to explore data or prototype ML models
• User asks about notebook images, auto-shutdown, or persistent storage

When NOT to Use

• User wants to deploy a production service → prefer deploy skill; ask if the user wants another valid path
• User wants to deploy a model → prefer llm-deploy skill; ask if the user wants another valid path
• User wants an SSH server → prefer ssh-server skill; ask if the user wants another valid path

Prerequisites

Always verify before launching a notebook:

1. Credentials — TFY_BASE_URL and TFY_API_KEY must be set (env or .env)
2. Workspace — TFY_WORKSPACE_FQN required. Never auto-pick. Ask the user if missing.
3. CLI — Check tfy --version. Install if missing: pip install 'truefoundry==0.5.0' && tfy login --host "$TFY_BASE_URL"

For credential check commands and.env setup, see references/prerequisites.md.

CLI Detection

tfy --version

Launch Notebook via UI

The fastest way is through the TrueFoundry dashboard:

1. Go to Deployments → New Deployment → Jupyter Notebook
2. Select workspace and configure resources
3. Click Deploy

Launch Notebook via tfy apply (CLI — Recommended)

Configuration Questions

Before generating the manifest, ask the user:

1. Name — What to call the notebook
2. GPU needed? — CPU notebook (default) or GPU notebook (for ML/training)
3. Home directory size — How much persistent storage in GB (default: 20)
4. Auto-shutdown — Enable auto-shutdown after inactivity? If yes, how many minutes? (default: 30 minutes). Set cull_timeout: 0 to disable.

CPU Notebook

1. Generate the manifest:

# tfy-manifest.yaml — Jupyter Notebook
name: my-notebook
type: notebook
image:
  image_uri: public.ecr.aws/truefoundrycloud/jupyter:0.4.5-py3.12.12-sudo
home_directory_size: 20
cull_timeout: 30
resources:
  node:
    type: node_selector
    capacity_type: on_demand
  cpu_request: 1
  cpu_limit: 3
  memory_request: 4000
  memory_limit: 6000
  ephemeral_storage_request: 5000
  ephemeral_storage_limit: 10000
workspace_fqn: "YOUR_WORKSPACE_FQN"

2. Preview:

tfy apply -f tfy-manifest.yaml --dry-run --show-diff

3. Apply:

tfy apply -f tfy-manifest.yaml

GPU Notebook

# tfy-manifest.yaml — GPU Jupyter Notebook
name: gpu-notebook
type: notebook
image:
  image_uri: public.ecr.aws/truefoundrycloud/jupyter:0.4.5-py3.12.12-sudo
home_directory_size: 20
cull_timeout: 30
resources:
  node:
    type: node_selector
    capacity_type: on_demand
  cpu_request: 4
  cpu_limit: 8
  memory_request: 16000
  memory_limit: 32000
  ephemeral_storage_request: 10000
  ephemeral_storage_limit: 20000
  devices:
    - type: nvidia_gpu
      name: T4
      count: 1
workspace_fqn: "YOUR_WORKSPACE_FQN"

Launch Notebook via REST API (Fallback)

When CLI is not available, use tfy-api.sh. Set TFY_API_SH to the full path of this skill's scripts/tfy-api.sh. See references/tfy-api-setup.md for paths per agent.

Create Notebook

TFY_API_SH=~/.claude/skills/truefoundry-notebooks/scripts/tfy-api.sh

$TFY_API_SH PUT /api/svc/v1/apps -d '{
  "name": "my-notebook",
  "type": "notebook",
  "image": {
    "image_uri": "public.ecr.aws/truefoundrycloud/jupyter:0.4.5-py3.12.12-sudo"
  },
  "home_directory_size": 20,
  "cull_timeout": 30,
  "resources": {
    "node": {"type": "node_selector", "capacity_type": "on_demand"},
    "cpu_request": 1,
    "cpu_limit": 3,
    "memory_request": 4000,
    "memory_limit": 6000,
    "ephemeral_storage_request": 5000,
    "ephemeral_storage_limit": 10000
  },
  "workspace_fqn": "WORKSPACE_FQN"
}'

GPU Notebook (REST API)

$TFY_API_SH PUT /api/svc/v1/apps -d '{
  "name": "gpu-notebook",
  "type": "notebook",
  "image": {
    "image_uri": "public.ecr.aws/truefoundrycloud/jupyter:0.4.5-py3.12.12-sudo"
  },
  "home_directory_size": 20,
  "cull_timeout": 30,
  "resources": {
    "node": {"type": "node_selector", "capacity_type": "on_demand"},
    "cpu_request": 4,
    "cpu_limit": 8,
    "memory_request": 16000,
    "memory_limit": 32000,
    "ephemeral_storage_request": 10000,
    "ephemeral_storage_limit": 20000,
    "devices": [
      {"type": "nvidia_gpu", "name": "T4", "count": 1}
    ]
  },
  "workspace_fqn": "WORKSPACE_FQN"
}'

Available Base Images

Default: public.ecr.aws/truefoundrycloud/jupyter:0.4.5-py3.12.12-sudo

Full image registry: https://gallery.ecr.aws/truefoundrycloud/jupyter

Security: Use pinned image versions from references/container-versions.md. Do not dynamically fetch image tags from external registries. Only use official TrueFoundry base images or images built from them.

See references/container-versions.md for latest versions.

Choosing an Image

• No GPU needed: Use the minimal image (py3.11.14-sudo)
• GPU workloads: Use CUDA image (cu129-py3.11.14-sudo)
• Custom packages: Build a custom image (see below)

Auto-Shutdown (Scale-to-Zero)

Notebooks auto-stop after inactivity to save costs. Default: 30 minutes.

Configure cull_timeout in minutes in the manifest (default: 30). Set to 0 to disable auto-shutdown.

What counts as activity: Active Jupyter sessions, running cells, terminal sessions. What doesn't count: Background processes, idle kernels.

Persistent Storage

• Home directory (/home/jovyan/) persists across restarts
• APT packages installed via apt do NOT persist — use Build Scripts
• Pip packages installed in home directory persist
• Conda environments persist

Recommended Storage by Use Case

Custom Images

Extend TrueFoundry base images to pre-install packages:

FROM public.ecr.aws/truefoundrycloud/jupyter:0.4.6-py3.11.14-sudo

USER root
RUN DEBIAN_FRONTEND=noninteractive apt install -y --no-install-recommends ffmpeg
USER jovyan

RUN python3 -m pip install --use-pep517 --no-cache-dir torch torchvision pandas scikit-learn

Critical: Do NOT modify ENTRYPOINT or CMD — TrueFoundry requires them.

Build Scripts (Persistent APT Packages)

Instead of custom images, add a build script during deployment to install system packages on every start:

sudo apt update
sudo apt install -y ffmpeg libsm6 libxext6

Cloud Storage Access

Via Environment Variables

Set during deployment:

• AWS S3: AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY
• GCS: GOOGLE_APPLICATION_CREDENTIALS

Via IAM Service Account

Attach cloud-native IAM roles through service account integration for secure, credential-free access.

Via Volumes

Mount TrueFoundry persistent volumes for direct data access. See volumes skill.

Git Integration

JupyterLab includes a built-in Git extension. Configure:

git config --global user.name "Your Name"
git config --global user.email "you@example.com"

Use Personal Access Tokens or SSH keys for authentication.

Python Environment Management

Default: Python 3.11. Create additional environments:

conda create -y -n py39 python=3.9

Wait ~2 minutes for kernel sync, then hard-refresh JupyterLab.

Presenting Notebooks

Show as a table:

Notebooks:
| Name          | Status  | Image         | GPU  | Storage |
|---------------|---------|---------------|------|---------|
| dev-notebook  | Running | py3.11 + CUDA | T4   | 50 GB   |
| data-analysis | Stopped | py3.11        | None | 20 GB   |

<success_criteria>

Success Criteria

• The notebook is launched and accessible via its URL in the TrueFoundry dashboard
• GPU resources are allocated as requested and visible inside the notebook (e.g., nvidia-smi works)
• Persistent storage is configured so the user's files survive restarts
• Auto-shutdown is enabled to prevent unnecessary cost from idle notebooks
• The user can install packages and access their data (cloud storage, volumes, or local upload)

</success_criteria>

Composability

• Need workspace: Use workspaces skill to find target workspace
• Need GPU info: Use workspaces skill to check available GPU types on cluster
• Need volumes: Use volumes skill to create persistent storage, then mount
• Deploy model after prototyping: Use deploy or llm-deploy skill
• Check status: Use applications skill to see notebook status

Error Handling

CLI Errors

tfy: command not found
Install the TrueFoundry CLI:
  pip install 'truefoundry==0.5.0'
  tfy login --host "$TFY_BASE_URL"

Manifest validation failed.
Check:
- YAML syntax is valid
- Required fields: name, type, workspace_fqn
- Image URI exists and is accessible
- Resource values use correct units (memory in MB)

Notebook Not Starting

Notebook stuck in pending. Check:
- Requested GPU type may not be available on cluster
- Insufficient cluster resources (CPU/memory)
- Image pull errors (check container registry access)

GPU Not Detected

GPU not visible in notebook. Verify:
- Used CUDA image (cu129-* variant)
- Requested GPU type is available (check workspaces skill)
- CUDA toolkit version matches your framework requirements

Storage Full

Notebook storage full. Options:
- Clean up unused files in /home/jovyan/
- Increase storage allocation
- Mount an external volume for large datasets

REST API Fallback Errors

401 Unauthorized — Check TFY_API_KEY is valid
404 Not Found — Check TFY_BASE_URL and API endpoint path
422 Validation Error — Check manifest fields match expected schema