feat: make Docker (DinD/DoD) optional via INSTALL_DOCKER build arg by recoskyler · Pull Request #25 · recoskyler/devcontainers

recoskyler · 2026-06-11T23:11:02Z

Summary

Makes the Docker CLI + Compose plugin (Docker-outside-of-Docker, via a bind-mounted host socket) optional through a new INSTALL_DOCKER build arg on the base image (default true, so existing images are unchanged). Because the arg lives on the base and every variant builds FROM it, this controls all images.

CI now publishes two flavors of every image:

	Docker tag	No-docker tag
each image	`:latest`, `:<version>`	`:latest-nodocker`, `:<version>-nodocker`

Changes

base/Dockerfile — ARG INSTALL_DOCKER=true gates the Docker apt repo and the docker-ce-cli/docker-compose-plugin packages (POSIX if inside RUN; gh/stripe still install). The docker group + socket-fix entrypoint stay (inert without the CLI).
.github/workflows/build.yml / check.yml — doubled matrix: base runs 2 jobs (docker + no-docker), variants run 10 (5 × {docker, no-docker}). INSTALL_DOCKER is passed to the base/VNC rebuild steps; -nodocker tag suffix via metadata-action flavor (onlatest=true); parallel -nodocker GHA cache scopes. Flutter no-docker chain preserved (no-docker base → VNC → flutter).
Docs — README Images table lists both flavors, Build Arguments documents INSTALL_DOCKER, CI/CD notes + CLAUDE.md reflect the doubled matrix and new cache scopes.
Cleanup (already staged on this branch) — removes Context7 MCP + suggestion hook, gstack skills, and the ~/.claude home-merge mechanism.

CI disk exhaustion fix + image slimming

The doubled matrix surfaced No space left on device failures in the variant builds (php, bun, rust, flutter) — GitHub runners (~14 GB free) ran out during image export/load. Investigation of the run logs traced it to two oversized layers, now trimmed:

headroom-ai[all] → headroom-ai[code,mcp,proxy] (scripts/setup-claude.sh): [all] pulled torch + ~2.5 GB of bundled CUDA libs (the 3.4 GB layer), plus onnxruntime/opencv/sentence-transformers via the image/evals/memory extras. The retained extras keep tree-sitter code analysis, MCP, and the proxy server (CPU-only onnxruntime/transformers/magika, ~500–700 MB) with no torch/CUDA. Lives in the base image, so all 5 variants shrink by ~2.5 GB+.
Removed the full Android Studio IDE (~2–3 GB) from the flutter image. Flutter/Android builds use the SDK (cmdline-tools, platform-tools, build-tools) installed via sdkmanager; install the IDE at runtime if GUI dev over VNC is needed.
Expanded the variants' "Free disk space" CI step — full ghcup/hostedtoolcache/swift removal + swapoff + apt-get clean for headroom margin.
Docs/scripts — README drops the Android Studio bullet; validate-flutter-runtime.sh Section 4 now skips the IDE-launch check (manual docker exec script, not wired into CI; verify-tools.sh never checked Android Studio).

Test plan

check.yml passes on this PR (builds + verifies all 12 images, posts per-image verification comments)
No variant build hits No space left on device
No-docker images build and pass tool verification
docker build --build-arg INSTALL_DOCKER=false -f base/Dockerfile . produces an image without docker, with gh/stripe intact
Default build still contains the Docker CLI (:latest unchanged)
headroom CLI works (proxy/code/mcp) in the built image; no torch/CUDA present
flutter build/flutter test work on the flutter image without Android Studio

kubectl install reliability

base/Dockerfile — kubectl now installs from its official static binary (dl.k8s.io/release/<v1.32>/bin/linux/$ARCH/kubectl) with curl --retry-all-errors instead of the pkgs.k8s.io apt repo. That repo's Release.key fetch intermittently returns 403 on CI runners (CDN rate-limiting), which non-deterministically failed the base build (broke the docker base while the no-docker base passed in the same run). Removes the apt repo, GPG key, and apt-get update round-trips.

Add INSTALL_DOCKER build arg (default true) to the base image so the Docker CLI + Compose plugin (Docker-outside-of-Docker, via a bind-mounted host socket) can be omitted entirely. The arg lives on the base and propagates to every variant through FROM. CI now builds every image in two flavors via a doubled matrix in both build.yml and check.yml: - base job runs 2 matrix jobs (docker + no-docker) - variants job runs 10 (5 variants x {docker, no-docker}) - no-docker images publish with a -nodocker tag suffix (metadata-action flavor suffix, onlatest=true so :latest-nodocker never clobbers :latest) - parallel -nodocker GHA cache scopes (base-nodocker, bun-nodocker, ...) - flutter no-docker chain preserved: no-docker base -> VNC -> flutter Docs (README, CLAUDE.md, devcontainer-claude.md) updated: Images table lists both flavors, Build Arguments documents INSTALL_DOCKER, CI/CD notes reflect the doubled matrix and new cache scopes. Also removes unused devcontainer tooling that was already staged on this branch: Context7 MCP + suggestion hook, gstack skills, and the ~/.claude home-merge mechanism (merge-claude-home.sh).

github-actions · 2026-06-11T23:22:45Z