go-tools¶
Go CI image: go, golangci-lint, goreleaser (+pro), syft, govulncheck, ONNX Runtime. Derives from ci-base.
Release v0.5.11, 2026-11-20. Built on ci-base. Source: phpboyscout/images/go-tools.
Pin it¶
Pin the full version. Each release also moves its minor tag and latest, which change under you.
| Variant | Tags |
|---|---|
| default | v0.5.11, v0.5 |
About this image¶
The Go track's CI image.
Contains go, golangci-lint, goreleaser (OSS + Pro), syft,
govulncheck, gotestsum and the ONNX Runtime shared library, on top of
ci-base.
Serves go-lint, go-test, go-security and the goreleaser component.
Why it exists¶
dev-tools carried Go, Rust and Node together. A Go repository therefore
pulled 2.7 GB, of which 920 MB was a Rust toolchain it never invoked —
and inherited every one of that image's 66 gate-visible findings (as measured
for cicd spec 0074 in August 2026), not one of which came from Rust.
| measured 2026-09-14 | compressed pull | gate-visible findings |
|---|---|---|
dev-tools:v0.3.4 (archived, frozen) |
958 MB | 73 |
go-tools:v0.2.1 |
484 MB | 15 |
Compressed pull is the sum of the manifest's layer sizes. Findings are what
scan:tools counts — trivy 0.74.0, --pkg-types library --severity
HIGH,CRITICAL --ignore-unfixed — and the number moves whenever an upstream
rebuilds a binary, so re-measure before repeating it. Zero of them are
OS-layer, and that is the one figure that holds: they sit in the upstream Go
binaries, and only their maintainers can clear them.
It runs as root¶
Unlike ci-base, tofu-tools and docs-tools, which all end as the non-root
ci user. This is deliberate, not inherited: GitLab's helper checks out
/builds as root, the Go module cache is written there, and rust-tool-base
seeds /etc/machine-id for a telemetry test. The Go and Rust tracks assume
root; the Tofu and docs tracks do not.
No mise¶
Tools come from upstream release assets, each pinned by an annotated ARG that
Renovate tracks, each verified against the checksum its upstream publishes.
Measured on tofu-tools, mise lost on every axis: 23s against 9s cold build,
two Renovate managers against one, GitHub API version-resolution against plain
release downloads, and no checksum verification at all.
Two details that cost a build each, recorded so they don't again:
- Go's checksum comes from the JSON index, not a
.sha256sidecar. That URL returns HTTP 200 — and an HTML redirect page. A status-code check says the file exists; only reading the body says what it is. - goreleaser Pro's release tag is the plain version —
v2.17.1, notv2.17.1-pro. The wrong guess 404s quietly.
govulncheck is built from source¶
It is the only tool here with no published binary, so go install builds it —
and govulncheck -version reports the Go this image pins (GO_VERSION in the
Dockerfile), current, because we compiled it. The others carry whatever Go
their maintainers used. That contrast is the whole reason
scan:tools is advisory rather than a gate.
go clean -cache afterwards is load-bearing: dev-tools shipped 141 MB of
exactly that build cache — nine times the 15 MB binary it produced — until it
was removed.
gotestsum is built from our fork¶
go-test uses gotestsum for its JUnit report and its end-of-run summary of
passed, skipped and failed tests (cicd spec 0101). It is built from source
too, but from
phpboyscout/gotestsum at a
pinned commit rather than from upstream.
Upstream's v1.13.0 release binary carries 49 advisories by trivy, 45 of them
in the Go standard library it was built with (go1.25.1), and upstream main
still pins golang.org/x/text, x/mod and x/sys at vulnerable versions.
The fork is upstream main with every dependency, direct and indirect, at
its latest release, and its go directive raised to the Go this image pins.
Built here it scans clean on trivy, govulncheck (reachable and unreachable,
test code included) and osv-scanner. The security fixes are not yet offered
upstream; once an upstream release carries them, the build switches back.
GOTESTSUM_COMMIT is the pin and has no Renovate tracker, since the fork
only moves when we move it. GOTESTSUM_VERSION names the upstream release
the fork is based on and is tracked, so a new upstream release opens a
draft MR that never automerges: that MR is the prompt to check whether
upstream has taken the fix, not something to merge as it stands.
ONNX Runtime¶
From the estate artefact channel, not from the vendor: everything else this estate ships is signed and mirrored, and pulling a runtime straight from upstream at build time was the exception. The channel mirrors the release byte-for-byte against a human-approved digest, which is pinned here too.
It lives in this image because the consumer is a Go test suite loading the library through purego at run time, so library and toolchain must ship together. Without it those tests skip rather than fail — a green pipeline for work that never ran.
Note the install glob is lib/*.so*, not libonnxruntime.so*: the narrower
one silently misses libonnxruntime_providers_shared.so.
ONNXRUNTIME_LIB names the path, deliberately generically — the image serves
the estate, so a consumer maps it to whatever variable it reads.
Do not reference $ONNXRUNTIME_LIB from a variables: block¶
This does not work:
An image ENV is not a CI/CD variable. GitLab expands a job's variables:
from the CI/CD variable set only, so the reference resolves to nothing and your
tests report a missing runtime on an image that has one.
Use the literal path instead:
$ONNXRUNTIME_LIB is right inside a script: — the shell in the container can
see it — and wrong in variables:. Duplicating the path is safe provided your
suite fails rather than skips when the gate is on and the library is
missing; if it skips, a disagreement between the two is silent.
Found by phpbotscout while wiring this image up, so the next consumer does not pay for it twice.
The scan is split by who can fix it¶
| Job | Scope | Gate |
|---|---|---|
scan:os |
OS/apk packages — the base and anything we install | hard |
scan:tools |
Go binaries from upstream releases | advisory |
If scan:os is red, we can fix it: bump the ci-base digest or the package.
scan:tools is advisory because its findings are compiled into upstream
binaries at their latest versions and clear only when those maintainers ship
rebuilds. dev-tools soft-fails its whole scan for that reason, which also
silences the half that is actionable.
Design record¶
Everything installed¶
Contents not yet published. This image's releases do not carry an SBOM yet. Until they do, the README above describes what it adds, and the Dockerfile is the full record.