Releases and upgrading¶
Every release is built and published by
.github/workflows/release.yml;
the process is documented in Releasing. A release
contains:
- the container image (the MCP server / CLI runtime) on GHCR and Docker Hub,
built for
linux/amd64andlinux/arm64on native runners; - an SBOM for the source tree and for each platform image, in both SPDX and CycloneDX;
- build metadata (
metadata.json) recording the source commit, build date, and the pinned Go, Helm, and Delve versions, plus the image digest; - a version-pinned MCP config for each supported AI tool (OpenCode, Claude
Code, Cursor, VS Code) and the
install-mcp.shhelper.
The image also carries the same information as OCI labels and as
/usr/local/share/helm-debugger/version.txt, and a GitHub provenance
attestation is pushed to the registry for gh attestation verify.
Apple Silicon¶
The linux/arm64 image runs natively on M-series Macs. This matters: the
debugger drives Delve, which uses ptrace, and ptrace does not work under
amd64-on-arm64 emulation (QEMU does not implement it, and Rosetta fails on
register reads). Docker Desktop picks the arm64 image automatically, so no
config change is needed on an Apple Silicon Mac.
Upgrading¶
For Docker use, pick a tag and change one line:
docker pull ghcr.io/jessesimpson36/helm-debugger:v0.2.0
| Tag | Behavior |
|---|---|
vX.Y.Z |
Immutable. Reproducible; upgrade by changing the tag. |
X.Y / X |
Moving. Receive patch/minor updates within a line. |
latest |
Moving. Newest non-prerelease. |
goX.Y.Z-helm...-delve... |
Immutable toolchain tag; changes only when the toolchain does. |
An AI coding tool does not update MCP server configuration on its own, so the
config is shipped as a release asset. Each release attaches a version-pinned
config per supported tool plus install-mcp.sh, which merges only the
helm-debugger server entry into an existing config (leaving other servers and
settings alone):
bash install-mcp.sh v0.2.0 # opencode, project-local
bash install-mcp.sh latest --global # or track latest globally
bash install-mcp.sh v0.2.0 --harness claude # Claude Code (.mcp.json)
bash install-mcp.sh v0.2.0 --harness cursor # Cursor (.cursor/mcp.json)
bash install-mcp.sh v0.2.0 --harness vscode # VS Code (.vscode/mcp.json)
(The download may not keep the executable bit, hence bash.) The script needs
jq and only rewrites the helm-debugger entry. Alternatively, copy the pinned
opencode-ghcr.json, claude-ghcr.json, cursor-ghcr.json, or
vscode-ghcr.json from the release assets (Docker Hub variants are attached
when Docker Hub publishing is configured).
After changing the config, reconnect the server (/mcps → helm-debugger, or
opencode service restart), then confirm the reported version:
docker run --rm ghcr.io/jessesimpson36/helm-debugger:v0.2.0 --version
See MCP server for how each tool consumes the config.