Skip to content

Package installation - #434

Open
anshisax wants to merge 32 commits into
masterfrom
package-installation
Open

Package installation#434
anshisax wants to merge 32 commits into
masterfrom
package-installation

Conversation

@anshisax

@anshisax anshisax commented Jun 17, 2026 •
edited
Loading

Copy link
Copy Markdown

Problem

Users want specific packages pre-installed so they don't have to run pip install or install.packages() every time they create an app. Custom app options are too complex for simple cases like "R Analysis with these 15 packages."

Solution

Natural language package installation - users can ask Claude for packages in plain English:

Examples:

  • "I need ggplot2" → Claude generates R installation command
  • "Set me up for machine learning" → Installs scikit-learn, xgboost, tensorflow, etc.
  • "I want tools for genomics analysis" → Installs appropriate bio packages

Two approaches:

  1. Pre-install (common-packages feature) - packages available at app startup

    {
  "features": {
    "ghcr.io/verily-src/workbench-app-devcontainers/common-packages": {
      "pythonPackages": "tensorflow scikit-learn pandas",
      "rPackages": "Seurat,DESeq2,ggplot2"
    }
  }
}

  2. On-demand (natural language with Claude) - install as needed while working

    • User: "I need plotly for visualization"
    • Claude: !pip install plotly

What's Added

  • features/src/common-packages/ - Feature for pre-installing packages
  • features/src/llm-context/skills/INSTALL_PACKAGES.md - Natural language skill for Claude
  • PACKAGE_INSTALLATION_GUIDE.md - User documentation

Impact

Before: 69 file changes (duplicate templates, embedded features)
After: 6 net file additions (removed 63 redundant files)

Users can pre-install packages with natural language OR explicit lists - simpler and more flexible than custom app options.

Examples

Machine learning:

{"pythonPackages": "scikit-learn xgboost tensorflow torch pandas numpy matplotlib"}

Genomics:

{"rPackages": "Seurat,DESeq2,GenomicRanges,tidyverse,ggplot2"}

On-demand in Jupyter:
User: "I need deep learning packages"
Claude: !pip install tensorflow torch transformers keras

https://verily.atlassian.net/browse/BENCH-8796

anshisax and others added 4 commits June 17, 2026 12:12
@anshisax @claude
Add comprehensive package installation guide for Workbench apps
6fe52b6 
Add PACKAGE_INSTALLATION_GUIDE.md with three approaches for pre-installing
packages in Workbench apps:

1. Devcontainer Features (easiest for R packages)
   - Use ghcr.io/rocker-org/devcontainer-features/r-packages
   - Simple comma-separated package list
   - Auto-installs system dependencies

2. Custom Dockerfile (best for Python packages)
   - Add RUN pip install commands
   - Cached in image layers for fast startup
   - Full version control

3. Post-create scripts (most flexible)
   - Supports mixed R + Python environments
   - Dynamic installation from requirements.txt

Includes:
- Complete working examples (R Analysis, Jupyter ML/AI stack)
- Comparison table showing when to use each approach
- FAQ section addressing common questions
- Copy-paste templates for quick setup

This addresses the common user need to avoid running pip install or
install.packages() every time they create an app, without requiring
the complexity of custom app options.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
@anshisax @claude
Add common-packages feature for easy package pre-installation
7755652 
Add a devcontainer feature that solves the common use case: users want
specific packages pre-installed to avoid running pip install or
install.packages() every time they create an app.

Features:
- Curated package presets (basic, ml, bio, full)
- Works for both Python and R
- Custom package support on top of presets
- No Dockerfile needed - just add to .devcontainer.json

Example usage:
{
  "features": {
    "ghcr.io/verily-src/workbench-app-devcontainers/common-packages": {
      "pythonPackages": "basic",
      "customPythonPackages": "mypackage",
      "rPackages": "basic",
      "customRPackages": "zoo,forecast"
    }
  }
}

This is simpler than custom app options for the common case where users
just want "R Analysis with these 15 packages".

Also updated PACKAGE_INSTALLATION_GUIDE.md to feature this approach first.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
@anshisax @claude
Simplify common-packages feature - just list your packages
7d87039 
Remove presets. Users just specify their package list:

{
  "features": {
    "common-packages": {
      "pythonPackages": "pandas numpy scikit-learn",
      "rPackages": "tidyverse,ggplot2,dplyr,plotly,shiny"
    }
  }
}

This solves the exact use case: 'R Analysis with these 15 packages
pre-installed' without the complexity of custom app options.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
@anshisax
Add test app for common-packages feature
3dbde28
@anshisax
anshisax requested review from a team as code owners June 17, 2026 21:28
anshisax and others added 7 commits June 17, 2026 15:31
@anshisax @claude
Add Claude skill for generating R apps with custom packages
4d702c6 
Enables workflow:
User: 'I want R with tidyverse, ggplot2, plotly'
Claude: Generates complete devcontainer directory with those packages pre-installed

Output can go in user's own repo - doesn't need to be in workbench-app-devcontainers.

Skill teaches Claude to use the common-packages feature for package pre-installation.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
@anshisax
Rename anshicustomapp to customapp
69c3999
@anshisax @claude
Update skill to support ALL app types (Jupyter, VSCode, R)
0f0cc86 
- Renamed skill to CREATE_CUSTOM_APP_WITH_PACKAGES (not just R)
- Feature works for Jupyter, VSCode, RStudio, etc.
- Supports both Python and R packages
- Simplified documentation

Users can request any app type with any packages:
- 'I want Jupyter with pandas and numpy'
- 'I want VSCode with tensorflow'
- 'I want R with tidyverse'

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
@anshisax
Fix common-packages feature path to be inside .devcontainer
f9d738f
@anshisax
Add llm-context feature to R app for Claude
68ebb5e 
  support
@anshisax @claude
Fix invalid JSON in .devcontainer.json - remove line breaks in strings
17f38c5 
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
@anshisax
Simplify package installation to natural language approach
2aef6f4 
- Remove 63 files (duplicate app templates and embedded features)
- Add INSTALL_PACKAGES.md skill for natural language package requests
- Simplify PACKAGE_INSTALLATION_GUIDE.md to focus on two approaches:
  1. Natural language with Claude (easiest)
  2. Pre-install with common-packages feature
- Keep core common-packages feature for pre-installation
- Enable Claude to understand requests like 'I need ggplot2' or 'set me up for machine learning'

This reduces the PR from 69 files to 6 files while enabling more flexible package installation through natural language.
pantherman594
pantherman594 reviewed Jul 7, 2026
View reviewed changes
Comment thread features/src/common-packages/devcontainer-feature.json
Comment thread features/src/llm-context/skills/CREATE_CUSTOM_APP_WITH_PACKAGES.md
Comment thread features/src/llm-context/PACKAGE_INSTALLATION_GUIDE.md
anshisax and others added 3 commits July 9, 2026 14:29
@anshisax @pantherman594
Update features/src/llm-context/skills/CREATE_CUSTOM_APP_WITH_PACKAGE…
f6ba93f 
…S.md

Co-authored-by: David Shen <2984600+pantherman594@users.noreply.github.com>
@anshisax @claude
Add installsAfter to common-packages feature
0e38da4 
Ensures common-packages installs after Python, R, and workbench-tools features are available.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
@anshisax @claude
Move PACKAGE_INSTALLATION_GUIDE to llm-context feature
e4464cb 
Address review feedback - this file belongs in the llm-context feature directory, not the repository root.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
@anshisax
anshisax requested a review from pantherman594 July 9, 2026 21:40
NBBense
NBBense requested changes Jul 9, 2026
View reviewed changes
Comment thread features/src/common-packages/README.md Outdated
anshisax added 6 commits July 13, 2026 09:39
@anshisax
Fix performance documentation - clarify caching behavior
08c3142 
Addresses NBBense feedback: packages persist across stop/restart,
not across all app creations. Only cached default configs benefit
from persistent caching.
@anshisax
Add test app for package installation feature
30349e2 
Test app validates:
- llm-context feature loads
- common-packages feature installs Python packages
- Packages are available on first launch
@anshisax
Move test app to repo root - fix devcontainer feature paths
8c0e0e3 
Devcontainer CLI requires local features to be children of .devcontainer/
Moving from src/jupyter-test-packages to jupyter-test-packages (root level)
allows ../features/src/* paths to work correctly.
@anshisax
Add symlink to features in test app - fix devcontainer path constraint
b8321e1 
Devcontainer CLI requires local features to be children of .devcontainer/
Symlink allows ./features/src/* paths to work within the constraint.
@anshisax
Add Jupyter to test app - make it a working Workbench app
8e937bf 
Test app now runs Jupyter Lab on port 8888 to satisfy Workbench
proxy requirements while testing package installation features.
@anshisax
Use Jupyter base image - fix externally-managed-environment error
2a88308 
Python 3.11 blocks pip in system Python. Using jupyter/base-notebook
which has Jupyter pre-installed and allows pip packages.
pantherman594
pantherman594 reviewed Jul 13, 2026
View reviewed changes
Comment thread jupyter-test-packages/.devcontainer/devcontainer.json Outdated
anshisax added 2 commits July 13, 2026 11:56
@anshisax
Add common-packages to workbench-jupyter-with-llm for testing
3481871 
Testing package installation feature in a working Jupyter template.
@anshisax
Remove test app and revert template modifications
495f70f 
Responding to @pantherman594 feedback on jupyter-test-packages.

Test app validation complete:
- llm-context feature: natural language package installation working
- common-packages feature: pandas, numpy, scikit-learn, matplotlib,
  seaborn pre-installed successfully
- Domain-based installation validated (genomics, time series packages)

Reverted workbench-jupyter-with-llm to original state - users will
add features to their own apps as needed.

Deliverables:
- features/src/llm-context/ (natural language installation skill)
- features/src/common-packages/ (pre-install packages)
- Documentation and README
anshisax and others added 5 commits July 13, 2026 14:46
@anshisax
Add Gemini support for package installation
19a0259 
Created GEMINI_PACKAGE_INSTALL.md with simplified domain mappings
for Gemini CLI. Users can now use natural language package
installation with both Claude and Gemini.

Same domain mappings as INSTALL_PACKAGES.md but formatted for
Gemini's consumption.
@anshisax
Add Antigravity support for package installation
82257dc 
Created ANTIGRAVITY_PACKAGE_INSTALL.md with domain mappings for
Antigravity AI assistant. Now supports natural language package
installation across Claude, Gemini, and Antigravity.

Same domain mappings as INSTALL_PACKAGES.md but formatted for
Antigravity's consumption.
@anshisax @claude
Fix shellcheck SC2086: quote PYTHONPACKAGES variable
5b03bd0 
Adds quotes around ${PYTHONPACKAGES} in pip install command to prevent
globbing and word splitting issues as flagged by shellcheck.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
@anshisax @claude
Make package skill emit a single self-contained Jupyter devcontainer
04aab9e 
Rework CREATE_CUSTOM_APP_WITH_PACKAGES so the skill outputs one
copy-pasteable .devcontainer.json that a customer points their custom
app config at: prebuilt jupyter image + common-packages feature, both
by ghcr path. Fixes the prior local ./.devcontainer path and the
package-only config that didn't actually launch an app.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@anshisax @claude
Skills: emit a full custom-app folder with pre-installed packages
e5b6088 
Rework the package-installation skills so "create me a custom app with
these packages" produces a complete app folder (.devcontainer.json +
docker-compose.yaml + devcontainer-template.json, +Dockerfile for
VSCode) that drops into src/<app-name>/ and runs as a Workbench custom
app, rather than a lone .devcontainer.json (which can't launch an app).

- CREATE_CUSTOM_APP_WITH_PACKAGES.md: full rewrite. Per-app-type
  scaffolds (Jupyter/RStudio-Shiny/VSCode) based on the real templates,
  with common-packages injected; complete worked Jupyter example.
- ANTIGRAVITY_PACKAGE_INSTALL.md: add "Create a Custom App (Full
  Folder)" section mirroring the same behavior.
- INSTALL_PACKAGES.md: fix the Creating Apps section that showed a
  conflicting single-file config; defer to the folder-scaffold skill.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
anshisax and others added 5 commits July 21, 2026 09:21
@anshisax @claude
Add common-packages test apps for end-to-end verification
d320569 
- workbench-jupyter-with-llm: add common-packages feature (local path)
  with sample python/R packages so the existing app exercises package
  pre-install.
- custom-jupyter-ml: full skill-generated custom app folder
  (.devcontainer.json + docker-compose.yaml + devcontainer-template.json
  + README) to verify the skill produces every file needed for a custom
  app. Uses the local feature path since common-packages isn't published
  to ghcr yet.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@anshisax @claude
Fix common-packages: don't quote PYTHONPACKAGES (breaks pip install)
90f5ad4 
Commit 5b03bd0 quoted ${PYTHONPACKAGES} to satisfy shellcheck SC2086,
but that makes pip treat the whole space-separated list as one invalid
requirement ("pandas numpy scikit-learn") and the feature install fails,
erroring the app build. Word-splitting is intentional here; leave it
unquoted and disable SC2086 inline instead.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@anshisax @claude
Add jupyter-single-cell custom app (skill-generated, Python/scanpy)
286ca6a 
Full app folder generated by CREATE_CUSTOM_APP_WITH_PACKAGES for a
single-cell (Python) request: .devcontainer.json + docker-compose.yaml +
devcontainer-template.json + README, with common-packages installing
scanpy anndata leidenalg pandas numpy matplotlib. Uses the local feature
path pending ghcr publication.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@anshisax @claude
Surface CREATE_CUSTOM_APP_WITH_PACKAGES in generated CLAUDE.md
78c7467 
The in-app skill routing table only advertised CUSTOM_APP.md, so
requests like "create a custom app with these packages" routed to the
generic guide and never hit the package skill. Add
CREATE_CUSTOM_APP_WITH_PACKAGES.md and INSTALL_PACKAGES.md to the
Available Skills table and add a trigger block that takes precedence
over CUSTOM_APP.md whenever packages/pre-install are involved.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@anshisax @claude
Fix skill+jupyter-single-cell: use prebuilt Jupyter image, not debian
da695c4 
The jupyter-template base defaults to bare debian:bullseye, so generated
Jupyter apps tried to build Python/JupyterLab from scratch and failed at
deploy (sysconfig.get_default_scheme AttributeError, post-startup.sh
exit 1). Base Jupyter apps on the prebuilt app-workbench-jupyter image
instead (custom-workbench-jupyter-template pattern: Dockerfile +
jupyter-common compose), which is the base that deploys successfully.

- CREATE_CUSTOM_APP_WITH_PACKAGES.md: Jupyter now uses the prebuilt
  template; worked example rewritten with Dockerfile + include compose;
  added a warning against jupyter-template.
- src/jupyter-single-cell: regenerated on the prebuilt base.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
pantherman594
pantherman594 reviewed Jul 22, 2026
View reviewed changes
Comment thread src/custom-jupyter-ml/.devcontainer.json

@pantherman594 pantherman594 Jul 22, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hey @anshisax, looks like there are 2 other generated apps that might've been from your testing

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@pantherman594 pantherman594 pantherman594 approved these changes
@NBBense NBBense Awaiting requested review from NBBense

Requested changes must be addressed to merge this pull request.

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@anshisax @pantherman594 @NBBense