PPBDS devcontainer

Container infrastructure for the PPBDS GitHub organization. This repo publishes one Docker image to the GitHub Container Registry:

ghcr.io/ppbds/devcontainer

It serves everyone: students taking the data science course (via codespace-starter) and developers working on PPBDS packages.

Note: this image replaced the old devcontainers/base, devcontainers/dev, and devcontainers/student trio in v1.0.0. The old packages remain on GHCR at their final tags so existing pins keep working, but they receive no new versions — switch to ghcr.io/ppbds/devcontainer.

Consuming

In a downstream repo's .devcontainer/devcontainer.json:

{
  "image": "ghcr.io/ppbds/devcontainer:X.Y.Z"
}

Replace X.Y.Z with the latest tag from the releases page. Pin to a specific semver tag — :latest floats with main and is not for student-facing (or, since 2026-07, package-repo) config.

Tag families

• :latest — most recent successful build from main. Convenient, not stable.
• :X.Y.Z and :X.Y — semver tags from a GitHub release on this repo. The canonical stable pin.

What's inside

• R — the rocker/tidyverse foundation, the PPBDS course packages (tutorial.helpers, vscode.tutorials, misc.tutorials, primer.tutorials), the arf console, Quarto, httpgd, and pak.
• Modeling — tidymodels plus engines (xgboost, lightgbm, catboost, randomForest, ranger, glmnet, bonsai) and brms for Bayesian regression via Stan.
• Inference reporting & presentation — gt, marginaleffects, patchwork, easystats (used throughout the primer book and tutorials).
• Mapping / census — sf + tidycensus (with tigris) for census-tract maps, plus the Kyle Walker toolkit: mapgl (token-free WebGL vector maps via MapLibre + CARTO styles), crsuggest, and idbr. Live census queries need a free Census API key.
• Interactive web viz + Shinylive — R htmlwidgets (plotly, leaflet, DT, crosstalk) and Python (plotly, altair, folium, itables) emit self-contained client-side JS; Shinylive (R & Python) runs Shiny apps as WebAssembly — everything publishes to static hosting (GitHub Pages).
• Python — a data-science stack installed with uv into /opt/venv from a pinned lockfile: numpy, pandas, matplotlib, seaborn, scikit-learn, statsmodels, jupyter, ipykernel. Default python on PATH, registered Jupyter kernel, group-writable so pip install works.
• Observable — Quarto's {ojs} cells render Observable Plot / OJS out of the box; the Observable Framework CLI is installed for standalone data-app projects.
• Package development — devtools, pkgdown, roxygen2, testthat, usethis, and the R CMD check toolchain (incl. qpdf).
• AI coding assistants — see below.

AI coding assistants

The image ships four AI coding CLIs so users can pick the model that fits their cost and quality needs:

• claude — Claude Code. Anthropic's CLI, single-provider, uses Claude models. Highest quality, highest cost.
• codex — Codex CLI. OpenAI's CLI, uses GPT/Codex models. Included with a ChatGPT Plus/Pro plan.
• agy — Antigravity CLI. Google's terminal coding agent — the successor to the Gemini CLI, which Google retired on 2026-06-18.
• aider — Aider. Multi-provider. Point it at DeepSeek, OpenRouter, OpenAI, Anthropic, or any OpenAI-compatible endpoint. The cost-flexible option.

The image ships no credentials. The recommended way to authenticate is to sign in on first run: start the CLI and complete the account sign-in (a ChatGPT plan for codex, a Claude plan for claude, a Google account for agy). This bills against a flat-rate subscription rather than metered API calls. As a fallback — and the only option for aider — supply an API key via Codespaces user secrets (below); each CLI stays inert until it has either a sign-in or a key.

Setting up API keys (the fallback path, per-user, one-time)

Configure each key as a personal Codespaces secret so it appears as an environment variable in every Codespace you launch.

1. Go to https://github.com/settings/codespaces.

2. Under Codespaces secrets, click New secret.

3. Add the secret name and value for each tool you plan to use:

   Secret name	Used by	Where to get it
   ANTHROPIC_API_KEY	claude (or sign in instead)	https://console.anthropic.com
   ANTIGRAVITY_API_KEY	agy (or sign in instead)	https://aistudio.google.com/apikey
   OPENAI_API_KEY	aider → OpenAI models	https://platform.openai.com/api-keys
   DEEPSEEK_API_KEY	aider → DeepSeek directly	https://platform.deepseek.com
   OPENROUTER_API_KEY	aider → any model via OpenRouter	https://openrouter.ai/keys

   (codex authenticates by signing in with a ChatGPT plan, or via codex login; it does not read OPENAI_API_KEY.)

4. Under Repository access, grant access to the repos you launch Codespaces from.

5. Save. The next Codespace you launch will have those env vars available.

Using cheaper models via Aider

# DeepSeek directly — typically ~$0.02–0.05 per 50K-token coding session
aider --model deepseek/deepseek-chat

# Or via OpenRouter — one key, hundreds of models you can swap between
aider --model openrouter/deepseek/deepseek-chat
aider --model openrouter/qwen/qwen-3-coder
aider --model openrouter/google/gemini-2.5-flash

For frontier-quality answers on hard problems, use claude directly (charged at Anthropic's rates) or aider --model anthropic/claude-sonnet-4-6.

Architecture and design rationale

See CLAUDE.md and the comments in the Dockerfile — the Dockerfile comments are the canonical record of what's installed and why.