Rewrite dmr as a standalone, engine-independent binary

[ericcurtin]

Summary

This rewrites cmd/dmr (previously a ~70-line hack combining the CLI command
tree with a bolted-on serve subcommand) into a small, documented,
genuinely standalone package, and adds the build/release plumbing needed to
distribute it outside this repo — laying the groundwork for adoption by
tools like sbx that
don't want a Docker Desktop or Docker Engine dependency.

What changed

• cmd/dmr/root.go builds the full model management command tree
  (run, ls, pull, rm, ps, inspect, logs, ...) and always forces
  MODEL_RUNNER_HOST to the local daemon first. This guarantees every dmr
  invocation takes the "manual host" path in cmd/cli/desktop/context.go
  and never probes for Docker Desktop or a Docker Engine connection. A
  regression test (TestServePersistentPreRunEOverridesParent) locks in the
  cobra semantics this relies on: a child command's own PersistentPreRunE
  replaces, rather than chains after, its parent's.
• cmd/dmr/serve.go runs pkg/server in-process (the same daemon used
  by the docker-model-runner container and Docker Desktop's bundled
  runner), with --port/--socket/--models-path flags in addition to the
  existing env vars. --port/--socket are mutually exclusive
  (MarkFlagsMutuallyExclusive), and an explicit flag clears the other
  transport's environment variable first, since pkg/server prefers
  MODEL_RUNNER_PORT over MODEL_RUNNER_SOCK whenever both happen to be
  set — otherwise an inherited env var could silently override an explicit
  flag. Covered by TestApplyServeEnv.
• Docker-Engine-only commands that manage a docker-model-runner
  container (install-runner, start-runner, stop-runner,
  restart-runner, reinstall-runner, uninstall-runner) are now hidden
  from the dmr command tree — dmr serve replaces the container
  entirely, so these no longer apply and only added confusion.
• pkg/envconfig.LlamaServerPath() no longer defaults to the Docker Desktop
  bundle path on non-macOS platforms (that path can never exist there); it's
  now only checked as a last-resort lookup on macOS itself, where Desktop
  vendors its own copy.
• Confirmed dmr has zero cgo dependencies, so make build-dmr now builds
  with CGO_ENABLED=0, versioned from the dmr-vX.Y.Z tag scheme
  (git describe --match 'dmr-v*', stripped of its prefix) rather than the
  container image's v* tags. A new make build-dmr-cross target
  cross-compiles it for every platform we intend to package: macOS
  (arm64), Linux (amd64/arm64), Windows (amd64) — all verified locally.
• Adds .github/workflows/release-dmr.yml: on a dmr-vX.Y.Z tag, the build
  job runs make build-dmr-cross directly (rather than re-encoding
  GOOS/GOARCH/ldflags in YAML, so there's a single source of truth for
  dmr's build flags), publishes a GitHub release, opens the Homebrew
  formula PR against docker/homebrew-tap, and submits the WinGet manifest
  to microsoft/winget-pkgs via wingetcreate — so dmr becomes
  installable via brew install docker/tap/dmr / winget install Docker.dmr
  on its own release cadence, independent of Docker Desktop's or Docker
  Engine's pipelines. Modeled on docker/sandboxes's publish-brew.yml /
  publish-winget.yml.
• Documents all of the above in README.md, cmd/dmr/README.md, and the
  new packaging/README.md.

Testing

• go build ./..., go vet ./..., gofmt -l — clean
• make lint (golangci-lint run ./...) — 0 issues
• go test -race ./... — all packages pass, including new tests for
  cmd/dmr (TestRemoveEngineOnlyCommands,
  TestServePersistentPreRunEOverridesParent, TestApplyServeEnv) and
  pkg/envconfig (LlamaServerPath platform behavior)
• go mod tidy — no changes
• Manually ran dmr serve --port 19434 with no Docker installed/running,
  confirmed /models responds and the daemon downloads llama.cpp from the
  registry on demand
• Cross-compiled and ran dmr version/dmr --help for all four target
  platforms (darwin/arm64, linux/amd64, linux/arm64, windows/amd64 build
  only, not run), and exercised the release workflow's packaging step
  locally against real make build-dmr-cross output
• Rebased on current main (no conflicts) and squashed to a single commit

Not in scope here

Linux .deb/.rpm packaging is unaffected — it continues through the
existing docker-model-plugin pipeline referenced in release.yml.

[sourcery-ai]

Hey - I've found 1 issue, and left some high level feedback:

• In the Makefile, DMR_VERSION uses git describe --match 'v*', which won’t pick up dmr-v* tags used by the new release flow; consider changing the match pattern (or dropping it) so local make build-dmr builds get the same versioning as the dmr-vX.Y.Z releases.
• The GitHub release-dmr.yml workflow re-encodes the dmr build flags, ldflags, and target matrix instead of reusing make build-dmr-cross, which risks them drifting; consider calling the Makefile target from the workflow or sharing a single source of truth for build configuration.
• For dmr serve, double-check whether commands.NewRootCmd defines a PersistentPreRun(E) on the root; if it does, that will still run before serve even though serve sets its own PersistentPreRunE, so you may need an explicit override to avoid the Docker CLI plugin initialization path you’re trying to bypass.

Prompt for AI Agents

Please address the comments from this code review:

## Overall Comments
- In the Makefile, `DMR_VERSION` uses `git describe --match 'v*'`, which won’t pick up `dmr-v*` tags used by the new release flow; consider changing the match pattern (or dropping it) so local `make build-dmr` builds get the same versioning as the `dmr-vX.Y.Z` releases.
- The GitHub `release-dmr.yml` workflow re-encodes the dmr build flags, ldflags, and target matrix instead of reusing `make build-dmr-cross`, which risks them drifting; consider calling the Makefile target from the workflow or sharing a single source of truth for build configuration.
- For `dmr serve`, double-check whether `commands.NewRootCmd` defines a `PersistentPreRun(E)` on the root; if it does, that will still run before `serve` even though `serve` sets its own `PersistentPreRunE`, so you may need an explicit override to avoid the Docker CLI plugin initialization path you’re trying to bypass.

## Individual Comments

### Comment 1
<location path="Makefile" line_range="63" />
<code_context>
 build-cli:
 	$(MAKE) -C cmd/cli

+DMR_VERSION := $(shell git describe --tags --always --dirty --match 'v*')
+DMR_LDFLAGS := -s -w \
+	-X main.Version=$(DMR_VERSION) \
</code_context>
<issue_to_address>
**issue (bug_risk):** Version discovery ignores dmr-specific tags and may not match released dmr binaries

DMR_VERSION uses `git describe --match 'v*'`, but dmr releases are now tagged `dmr-vX.Y.Z`. As a result, `build-dmr`/`build-dmr-cross` will keep picking up the generic `v*` tags (container image scheme), so locally built dmr binaries won’t match published dmr release versions. Please either drop `--match` or switch to `--match 'dmr-v*'` (and strip the prefix) so standalone dmr builds follow the dmr tag scheme.
</issue_to_address>

---

Sourcery is free for open source - if you like our reviews please consider sharing them ✨

• X
• Mastodon
• LinkedIn
• Facebook

_{Help me be more useful! Please click 👍 or 👎 on each comment and I'll use the feedback to improve your reviews.}

[gemini-code-assist]

Code Review

This pull request introduces a standalone dmr binary that bundles the inference daemon and model management CLI, removing dependencies on Docker Desktop or Docker Engine. It adds cross-compilation support, documentation, and refactors the CLI command tree to hide engine-specific commands. Feedback focuses on improving the CLI flag handling in cmd/dmr/serve.go by ensuring explicit flags correctly override conflicting environment variables and enforcing mutual exclusivity between the --port and --socket flags.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[ericcurtin]

Addressed all three review points from @sourcery-ai and both from @gemini-code-assist-bot in the latest push:

1. DMR_VERSION tag matching — now uses --match 'dmr-v*' (stripped of the dmr- prefix) instead of v*, so local make build-dmr/make build-dmr-cross builds get the same version as the dmr-vX.Y.Z releases.
2. release-dmr.yml duplicating build config — the workflow's build job now just runs make build-dmr-cross directly (with fetch-depth: 0 so git describe resolves to the pushed tag) instead of re-encoding GOOS/GOARCH/ldflags in YAML, so there's one source of truth for dmr's build flags.
3. PersistentPreRunE override on serve — verified cobra's documented behavior (a child's own PersistentPreRunE replaces rather than chains after its parent's) and added TestServePersistentPreRunEOverridesParent as a regression test locking this in.
4. --port/--socket mutual exclusivity — added cmd.MarkFlagsMutuallyExclusive("port", "socket").
5. Inherited env var silently overriding an explicit flag — an explicit --socket/--port now unsets the other transport's env var first, since pkg/server prefers MODEL_RUNNER_PORT over MODEL_RUNNER_SOCK whenever both are set. Covered by a new TestApplyServeEnv.

Rebased on current main (no conflicts — main hadn't moved) and squashed back into a single commit. make lint and go test -race ./... both still pass across the whole repo.