The npm-installed Claude Code runs via Node.js, which causes remote
control to fail with '/usr/bin/node: bad option: --sdk-url'. The native
binary handles subprocess spawning correctly.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
npm was serving a cached older version even with Docker no-cache.
Clear npm cache and use --prefer-online to force a fresh registry fetch.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
GHA cache was serving stale npm install layers despite cache-bust ARG.
Remove all caching — every build is now fully clean.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The 2.0.0-dev tag was only built from the now-merged
feature/serverless-2.0.0 branch. Pushes to main only tagged latest,
so the 2.0.0-dev image in the registry was frozen and missing all
recent fixes. Default to latest and also tag main builds as 2.0.0-dev
for backwards compatibility.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Two fixes:
- Move Claude Code npm install below TOOLS_CACHEBUST ARG so it actually
gets refreshed when the cache-bust value changes
- Make GITHUB_REPO env conditional so an empty Helm value no longer
overrides the value provided via the Kubernetes secret (envFrom)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The direct GCS binary download approach has been unreliable across
multiple attempts. Revert to the proven npm install method. Node.js
is already required for Happy Coder so there is no extra dependency.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
npm install fails in CI due to native dependency compilation issues.
Download the pre-built binary directly from the official GCS distribution
bucket with SHA256 checksum verification. This approach worked previously
(run #135) and avoids npm entirely — Node.js is only needed for Happy Coder.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The native binary installer (both direct GCS download and claude.ai/install.sh)
has been unreliable during Docker builds. Revert to the proven npm approach.
Node.js is already required for Happy Coder, so there's no extra dependency.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The previous native installer approach used a direct GCS bucket download
that was fragile and failing during builds. Switch to the official
install script (claude.ai/install.sh) which handles version discovery,
platform detection, and checksum verification properly.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The legacy GPG key import and .list format was failing with exit code 100
in CI. Switch to the DEB822 .sources format and install -D key method
per Microsoft's current documentation.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The native Claude Code installer (and other tools) fetch "latest" at
build time, but Docker layer caching serves stale layers because the
RUN command text never changes. Add TOOLS_CACHEBUST build arg with
github.run_id so every CI run re-downloads fresh tool binaries.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace npm-based Claude Code installation with the native binary
installer. Downloads directly from Anthropic's distribution bucket to
/usr/local/bin/claude — no Node.js dependency for Claude Code anymore.
Node.js is retained for Happy Coder only.
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
Updates GitHub Actions to build development images from serverless feature branch:
## GitHub Actions Updates
- Trigger builds on feature/serverless-* branches
- Add 2.0.0-dev tag for feature/serverless-2.0.0 branch
- New routing proxy build job for serverless features
- Parallel builds: main devcontainer + routing proxy
## Chart Updates
- Default image tag changed to 2.0.0-dev
- Routing proxy tag updated to 2.0.0-dev
- Ready for development testing
## Build Outputs
When pushed to feature/serverless-2.0.0:
- ghcr.io/cpfarhood/devcontainer:2.0.0-dev
- ghcr.io/cpfarhood/devcontainer-routing-proxy:2.0.0-dev
This enables immediate testing of serverless features without manual builds.
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
Implements a complete serverless development container platform:
## Architecture
- Authentik forward auth for authentication/authorization
- NGINX routing proxy extracts GitHub repo from URL path
- Knative Service auto-scales dev container instances from 0
- Dynamic GitHub repo routing via /github/{owner}/{repo}
## Components
- routing-proxy: NGINX-based service for repo extraction and forwarding
- deployment.yaml: Complete K8s manifests (proxy, Knative, ingress, secrets)
- authentik-config.yaml: Authentik application and provider configs
- serverless scripts: Dynamic repo initialization and startup handling
- Comprehensive documentation and Makefile for ops
## Key Features
- Scale to zero when not in use (cost-effective)
- Per-request isolation (each repo gets own container)
- Built-in file manager for upload/download
- Support for private repos via GitHub tokens
- User attribution via Authentik headers
- WebSocket support for VNC connections
Example usage: https://devcontainer.farh.net/github/microsoft/vscode
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
- Install Helm v3.17.1 in Dockerfile for chart development (closes#49)
- Add fileManager toggle using base image's WEB_FILE_MANAGER (closes#11)
- Wire WEB_FILE_MANAGER env vars in deployment template
- Update CLAUDE.md, README.md with new features and values
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
- Add helm.sh/resource-policy: keep to PVC (prevent data loss on uninstall)
- Add fail guard for empty name value in Helm templates
- Fix Makefile IMAGE_NAME from antigravity to devcontainer
- Pin busybox:1.37, homeassistant:v6.7.1, playwright:v0.0.68 (was latest/stable)
- Set imagePullPolicy: IfNotPresent on pinned sidecars
- Remove fetch/sequentialthinking from .mcp.json (sidecars removed from chart)
- Default storage.className to empty (use cluster default, was ceph-filesystem)
- Default Happy Coder URLs to empty (was private farh.net endpoints)
- Broaden githubRepo schema to accept GitLab/Gitea URLs
- Add unknown IDE warning before VSCode fallback
- Add mkdir -p before credential file write (fix fresh PVC boot)
- Guard app user existence in cont-init-user.sh
- Add NOTES.txt post-install template with port-forward and secret hints
- Add standard app.kubernetes.io/* labels and separate selectorLabels
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add fetch and sequentialthinking MCP servers to allowed list, and enable
voltagent dev-exp and lang subagent plugins.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Fixed the Crush AI assistant installation in Dockerfile by replacing the
single-line tar extraction with --strip-components (which was failing)
with a multi-step approach: download to temp, extract, move binary,
and cleanup. This ensures the binary is properly extracted from the
versioned directory structure in the tarball.
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
PR #45 added fetch and sequentialthinking MCP sidecars to values.yaml
and the deployment template but missed updating values.schema.json.
The schema has additionalProperties: false on mcp.sidecars, causing
Helm upgrade to fail with validation errors.
Also adds resourceProfile to the schema as it was missing.
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Happy <yesreply@happy.engineering>
Remove tag triggers and duplicate release job from build-and-push.yaml.
The release-unified.yaml workflow handles the full release flow (Docker
build, Helm chart, GitHub release) when triggered via workflow_dispatch.
Previously, release-unified.yaml pushing a commit to main AND a v* tag
would trigger build-and-push.yaml up to twice, causing multiple Docker
builds to race for the :latest tag. The stale GHA layer cache in the
racing build could overwrite :latest with an image missing new tools
(e.g., crush, opencode).
Changes:
- Remove tags: ['v*'] trigger (release-unified handles tag-based releases)
- Remove duplicate release job (Helm chart + GitHub release)
- Remove semver tag patterns from metadata (not needed without tag trigger)
- Skip builds from github-actions[bot] to avoid racing with release commits
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
Add two new MCP (Model Context Protocol) sidecars to enable web content fetching
and structured problem-solving capabilities:
- **Fetch MCP**: Web content fetching and HTML to markdown conversion on port 8082
- **Sequential Thinking MCP**: Structured thinking and problem-solving processes on port 8083
Both sidecars are enabled by default and use the official MCP Docker images
(mcp/fetch and mcp/sequentialthinking) with fastmcp SSE transport.
Changes:
- Add fetch and sequentialthinking sidecars to values.yaml
- Add sidecar containers to deployment.yaml template
- Update .mcp.json with new server endpoints
- Update CLAUDE.md documentation with new sidecar details
Closes#43, #44
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
Chris Farhood
github-actions[bot]