fix(ci): wire the computed release version into the built binaries #55

Merged

rcsheets merged 1 commit from fix/version-wiring into main

2026-07-23 09:42:34 +00:00

rcsheets commented

2026-07-23 09:40:27 +00:00

Owner

The containerized builds report gllm unknown on the index page (and in /status and the gllm_server_info metric). Two independent breaks:

The semver never reached the build. CI computes the release version in the "Set image tags" step, uses it to tag the image, and throws it away -- no build arg, no ldflags. The binary and the image tag were never connected.
The SHA fallback was dead too. .dockerignore excludes .git (deliberately, for stale-artifact hygiene), so go build inside the container has no repo to stamp vcs.revision from, and BuildVersion() fell through to "unknown".

The fix threads the version through:

internal/engine/status.go: link-time buildVersion var; BuildVersion() prefers it, keeping the VCS-revision fallback for local make build.
Dockerfile / Dockerfile.cuda: ARG VERSION= injected via -ldflags -X.
build-push.yaml: both image builds pass --build-arg VERSION=${{ steps.tags.outputs.version }}.
Makefile: image/image-cuda pass git describe so local images report provenance instead of "unknown".

Both CPU and CUDA images get the plain version (no -cuda suffix) -- the backend is already a separate /status field and the flavor lives in the image tag. Dockerfile.cuda-prerelease is untouched (build-only canary, never pushed); the 13.4 GA probe inherits the empty ARG default harmlessly.

Verified: go vet/go build/engine tests pass; a scratch test run with the -X flag reports the injected version and without it falls back as before.

🤖 Generated with Claude Code

The containerized builds report `gllm unknown` on the index page (and in `/status` and the `gllm_server_info` metric). Two independent breaks: 1. **The semver never reached the build.** CI computes the release version in the "Set image tags" step, uses it to tag the image, and throws it away -- no build arg, no ldflags. The binary and the image tag were never connected. 2. **The SHA fallback was dead too.** `.dockerignore` excludes `.git` (deliberately, for stale-artifact hygiene), so `go build` inside the container has no repo to stamp `vcs.revision` from, and `BuildVersion()` fell through to `"unknown"`. The fix threads the version through: - `internal/engine/status.go`: link-time `buildVersion` var; `BuildVersion()` prefers it, keeping the VCS-revision fallback for local `make build`. - `Dockerfile` / `Dockerfile.cuda`: `ARG VERSION=` injected via `-ldflags -X`. - `build-push.yaml`: both image builds pass `--build-arg VERSION=${{ steps.tags.outputs.version }}`. - `Makefile`: `image`/`image-cuda` pass `git describe` so local images report provenance instead of "unknown". Both CPU and CUDA images get the plain version (no `-cuda` suffix) -- the backend is already a separate `/status` field and the flavor lives in the image tag. `Dockerfile.cuda-prerelease` is untouched (build-only canary, never pushed); the 13.4 GA probe inherits the empty `ARG` default harmlessly. Verified: `go vet`/`go build`/engine tests pass; a scratch test run with the `-X` flag reports the injected version and without it falls back as before. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

rcsheets added 1 commit

2026-07-23 09:40:27 +00:00

fix(ci): wire the computed release version into the built binaries

ci / test_and_build (pull_request) Successful in 23s

Details

0cedf31b46

The index page (and /status, and gllm_server_info) reported "gllm
unknown" from the containerized builds: BuildVersion() only knew how to
read the Go toolchain's VCS stamp, but .dockerignore excludes .git from
the build context so there was nothing to stamp -- and the release
version CI computes was only ever used to tag the image, never passed
into the build.

Thread it through: a link-time buildVersion var in the engine package
(VCS revision stays as the fallback for local `make build`), a VERSION
build arg in both Dockerfiles, and CI passing its computed version to
both image builds. Local `make image`/`image-cuda` pass git describe so
they report provenance instead of "unknown" too.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

pr-reviewer-bot commented

2026-07-23 09:40:28 +00:00

Collaborator

Automated review by pr-reviewer v0.37.0 | Safety Check | Claude Sonnet 5 | tracking id r-61e18c-61f25d
This is an AI-generated review and may contain mistakes.

Status: ✅ Completed

⚠️ The primary review (Mistral Small) failed. The review below is from Claude Sonnet 5, which reviewed the same diff.

✅ Verdict: LGTM — Straightforward, low-risk build/CI wiring fix with backward-compatible fallback logic and no security or breaking-change concerns.

The change is a straightforward, low-risk CI/build-wiring fix:

internal/engine/status.go: adds a package-level buildVersion var, overridable via -ldflags -X, with fallback to existing VCS-based logic. Logic is correct and backward compatible (empty string falls through to old behavior).
Dockerfiles: ARG VERSION= with empty default is safe — harmless if unset, matches stated intent.
build-push.yaml: adds --build-arg VERSION=... to both CPU and CUDA build invocations, sourced from the already-computed steps.tags.outputs.version. No secrets involved, straightforward plumbing.
Makefile: adds VERSION ?= $(shell git describe ...) with sane default and passes it through to local image builds.

No security concerns (no secrets, no injection vectors — version strings come from git tags/CI, not user input). No accidentally committed files. No breaking changes: the fallback path is preserved, so binaries built without the new ldflag continue to behave as before. Diff is small and matches the PR description precisely; nothing looks tacked on or suspicious.

Minor note: no new tests added for the status.go change, but the logic is trivial (string override with fallback) and this is acknowledged in the PR's own context notes.

*Automated review by [pr-reviewer](https://git.brooktrails.org/brooktrails/pr-reviewer) v0.37.0 | Safety Check | Claude Sonnet 5 | tracking id `r-61e18c-61f25d`* *This is an AI-generated review and may contain mistakes.* **Status:** ✅ Completed --- > ⚠️ The primary review (**Mistral Small**) failed. The review below is from **Claude Sonnet 5**, which reviewed the same diff. **✅ Verdict: LGTM** — Straightforward, low-risk build/CI wiring fix with backward-compatible fallback logic and no security or breaking-change concerns. The change is a straightforward, low-risk CI/build-wiring fix: - `internal/engine/status.go`: adds a package-level `buildVersion` var, overridable via `-ldflags -X`, with fallback to existing VCS-based logic. Logic is correct and backward compatible (empty string falls through to old behavior). - Dockerfiles: `ARG VERSION=` with empty default is safe — harmless if unset, matches stated intent. - `build-push.yaml`: adds `--build-arg VERSION=...` to both CPU and CUDA build invocations, sourced from the already-computed `steps.tags.outputs.version`. No secrets involved, straightforward plumbing. - `Makefile`: adds `VERSION ?= $(shell git describe ...)` with sane default and passes it through to local image builds. No security concerns (no secrets, no injection vectors — version strings come from git tags/CI, not user input). No accidentally committed files. No breaking changes: the fallback path is preserved, so binaries built without the new ldflag continue to behave as before. Diff is small and matches the PR description precisely; nothing looks tacked on or suspicious. Minor note: no new tests added for the status.go change, but the logic is trivial (string override with fallback) and this is acknowledged in the PR's own context notes.

rcsheets merged commit baf40f5a0a into main

2026-07-23 09:42:34 +00:00

rcsheets referenced this pull request from a commit

2026-07-23 09:42:34 +00:00

Merge pull request 'fix(ci): wire the computed release version into the built binaries' (#55) from fix/version-wiring into main

rcsheets deleted branch fix/version-wiring

2026-07-23 09:42:35 +00:00