Changelog

1.3.0

Adds a third way to supply Elixir: an already-compiled distribution.

Added

• prebuilt_elixir_from_http_archive and prebuilt_elixir_from_hex_builds (bzlmod/extensions.bzl, repositories/elixir_config.bzl, new repositories/BUILD_prebuilt.tpl, new elixir_prebuilt in private/elixir_build.bzl).

  The mirror of what rules_erlang 3.20.0 added for OTP, and for the same reason: neither existing path can consume a downloadable prebuilt. external_elixir_from_path needs an elixir_home that already exists on the machine running the action, and internal_elixir_from_http_archive always feeds its URL to elixir_build, which runs make.

  elixir_build extracts a source archive, runs make, and copies the resulting bin/ and lib/ into the release directory. A precompiled distribution is those two directories, so staging one is extraction plus the same validation step -- run iex --version and record what it reports.

  elixir_config_ext.prebuilt_elixir_from_hex_builds(
      name = "elixir_1_19_4",
      version = "1.19.4",
      otp_major = "28",
      sha256 = "8fd7b5705b756c0e1ec71f9e8281b4b75801b9564f0205b5035319e8505ad2b4",
  )
  

  Unlike OTP, no relocation step is needed. An OTP installation bakes ROOTDIR into its start scripts and ships an Install script to repoint it; Elixir's launchers resolve their own root relative to argv[0], so a precompiled distribution works from wherever it is unpacked.

  Also unlike OTP, one archive serves every platform. Elixir compiles to BEAM bytecode, which is architecture-independent -- the published archive contains no native objects -- which is why prebuilt_elixir_from_hex_builds takes an otp_major rather than an arch/os pair, and why there is no exec_compatible_with attribute here.

  The rule accepts both zip and tar archives, because hex.pm publishes a zip while other distributions are tarballs, and GNU tar cannot read zip. It fails with an explicit message when unzip is absent for a zip input, and when the extracted archive has no bin/ and lib/ -- the likely mistake being an Elixir source archive, whose failure would otherwise surface as a confusing missing-file error after extraction.

  As with the OTP change, a prebuilt installation registers under the existing :elixir_internal constraint and keeps the :elixir_build target name, so moving a repository between the source and prebuilt paths is a MODULE.bazel edit and nothing else.

1.2.0

First release from https://github.com/bazelverse/rules_elixir.

Continues from the v1.1.0 tag of the unmaintained upstream, carrying the fixes that were being kept downstream as a vendored copy. Requires rules_erlang 3.18.0.

Every change is a separate commit, so git diff v1.1.0..v1.2.0 is exactly this list and nothing else.

Fixed

• Skip system Elixir (repositories/elixir_config.bzl). Honour RULES_ELIXIR_SKIP_SYSTEM=1, so a build can use a hermetic toolchain instead of probing the host for an Elixir install.

• Portable tarball extraction (private/elixir_build.bzl). The same GNU-tar assumption fixed in rules_erlang 3.18.0. --transform does not exist in the bsdtar that macOS ships, so every Elixir target failed on macOS before compiling anything.

• Allow an empty LICENSE glob (elixir_app.bzl). elixir_app hard-failed on any package without a LICENSE file.

• Compile-time data (elixir_app.bzl, private/elixir_bytecode.bzl). elixir_bytecode could not declare compile-time file inputs, so a package reading a file during compilation, or declaring @external_resource, failed in the sandbox.

• ExUnit test headers (private/ex_unit_test.bzl). ex_unit_test staged its dependencies without their include/ directories, so a test using Record.extract(from_lib: ...) died before running a case. Compilation rules stage headers; tests now do too.

• ExUnit workspace layout (private/ex_unit_test.bzl). ex_unit_test flattened srcs and data by stripping the package prefix, which breaks any test resolving a repository-relative path off __DIR__.

• Stage test inputs into TEST_TMPDIR (private/ex_unit_test.bzl). ex_unit_test copied every srcs and data file into TEST_UNDECLARED_OUTPUTS_DIR and ran there. Bazel treats that directory as artifacts the test produced, so it stats and mime-types every entry to build the manifest, then uploads them. Each target shipped a few thousand of its own inputs to the CAS per run. On a remote executor carrying no file(1) it also emitted roughly 2,150 "command not found" lines per test. Nothing was ever collected from there on purpose.

• Migrate off the deprecated Windows condition (ex_unit_test.bzl). @bazel_tools//src/conditions:host_windows is deprecated and warned on every ex_unit_test target; it is now the @platforms//os:windows constraint. This is also a semantic correction. The old key matched the host, while the flag decides whether the generated runner is a batch file or a shell script, a property of the platform the test executes on. For a test rule that is the target platform.

  The key is a Label, not a bare string. A select() key written as a string resolves against the repo mapping of the package that instantiates the macro, so a string would have obliged every consumer to add bazel_dep(name = "platforms") of their own and failed with "No repository visible as @platforms" if they did not. The question never arose while the key was @bazel_tools//..., because bazel_tools is an implicit dependency of every module and is visible from everywhere.

• ExUnit result detection under Elixir 1.20 (private/ex_unit_test.bzl). The runner asserted on the summary text, matching 0 failure and [0-9] test. Elixir 1.20 replaced N tests, M failures with Result: N passed, so under it every passing suite failed its own assertion.

  Failure detection now rests on the exit code, which the set -eo pipefail at the top of the runner already carried through the tee, and which does not change between releases.

  One assertion on the summary text remains, because there is exactly one thing the exit code cannot express: a suite that executed no tests reports success, which would let a target whose sources stopped matching any test pass forever. It fires only when nothing ran and nothing was excluded. A target that filters by tag and excludes everything it has is doing what it was asked to, and is a normal way to shard a suite, so an exclusion count means the summary is honest and the target passes. Both spellings are read while the supported window spans Elixir 1.19 and 1.20.

  This is the one behavioural difference from 1.1.0. A suite that defines no tests at all, and excludes none, passed before and now fails. That is the case //:empty_test has always asserted should fail.

Added

• mix_archive_build (mix_archive_build.bzl, private/mix_archive_build.bzl): build a .ez Mix archive. The rules could already compile an Elixir app but could not produce an archive, and mix archive.install is the only way to give Mix something it needs before it can resolve a project. Adapted from rabbitmq-server's bazel/elixir/mix_archive_build.bzl (MPL-2.0), the same lineage as this ruleset.

• hex module extension (bzlmod/extensions.bzl): builds Hex itself from source as a Mix archive, exported as @hex//:archive.

  Every consumer needs this and cannot discover it. Mix aborts with "Could not find an SCM for dependency" on any dependency entry in a mix.exs unless Hex is installed as an archive, including in a build that supplies all dependencies from Bazel and passes --no-deps-check. Hex has no dependencies of its own, so it bootstraps with an empty dependency graph. The version defaults to DEFAULT_HEX_VERSION and is overridable with the from_github_release tag, mirroring internal_elixir_from_github_release.

Removed

• Bazel 7 support, declared as bazel_compatibility = [">=8.0.0"]. This follows rules_erlang 3.18.0, which requires Bazel 8 and which this release requires in turn, so 7 was already unreachable in practice.

Changed

• Dependencies moved to current stable.

  	from	to
  rules_erlang	3.16.0	3.18.0
  bazel_skylib	1.7.1	1.9.2
  platforms	-	1.1.0 (new)

  platforms is new because ex_unit_test selects on @platforms//os:windows; see the Windows condition entry above for why the dependency belongs here rather than in every consumer.

• sh_test loads from @rules_shell (test/BUILD.bazel, test/MODULE.bazel). It stopped being a native rule in Bazel 8, so the test module gained both the load statement and a rules_shell dependency.

• @platforms//host:host replaces @local_config_platform//:host (examples/internal-elixir). Bazel 8 removed the autoconfigured local_config_platform repository. The example's platforms also register themselves as execution platforms, which Bazel 9 requires of any target platform that has to run tests.

• The internal-elixir example builds OTP 29.0.5 and Elixir 1.20.3, up from OTP 26.2.5 and Elixir 1.16.1. It is the only from-source coverage there is, and it was pinned two OTP majors behind the supported window.

• CI covers the latest two Elixir minors against the latest two OTP majors, currently 1.20 on OTP 29 and 1.19 on OTP 28. Elixir supports a moving window of OTP majors, so the matrix tests pairs rather than crossing the two lists. The Bazel matrix lives in the registry presubmit, which covers 8.x and 9.x.

• The registry presubmit installs Elixir, which it never did. It built @rules_elixir//... with only Erlang on the machine, leaving the external Elixir toolchain nothing to resolve. It also moves from the platforms: mapping to the tasks: schema, and activates the kerl build so the Elixir archive finds an ERTS to run on.

  The structure here comes from rabbitmq/rules_elixir#7, merged from upstream's final state; the versions are this project's.

• Fix the registry source template's archive name (.bcr/source.template.json). It asked for rules_elixir-{TAG}.tar.gz, but the release workflow uploads rules_elixir-{VERSION}.tar.gz, and this repository tags with a leading v. The two are only equal when the tag carries no prefix, so the published URL would have been rules_elixir-v1.2.0.tar.gz and returned a 404.

• Publishing targets the Bazel Central Registry. The workflow that pushed to rabbitmq/bazel-central-registry@erlang-packages, a private registry fork this project cannot write to, is replaced by bazel-contrib/publish-to-bcr. The registry entry for that fork is also dropped from every .bazelrc; it served rules_erlang 3.15.x and nothing reads it now.

Compatibility

For a bzlmod consumer, 1.2.0 is a drop-in replacement for 1.1.0: no rule, macro, provider or attribute changed. The rules_erlang dependency moves from 3.16.0 to 3.18.0, which is itself a drop-in replacement.

One behaviour differs, described under the ExUnit entry above: an ex_unit_test that runs no tests and excludes none now fails rather than passes. A target that excludes everything by tag is unaffected.

Requires Bazel 8 or newer. Tested on Bazel 8.7.0 and 9.2.0.

Neither module is on the Bazel Central Registry yet, so a consumer needs an override for rules_elixir and one for rules_erlang. Overrides only take effect in the root module, which is why both are needed even though only one is a direct dependency. See the Installation section of the README.