Building the native inference worker

task build builds the portable Contenox control and harness binary. The inference worker is a separate native executable, built by the existing Make pipe. Its transport contract connects it to Contenox without linking inference libraries into the CLI.

Run make help for native build and packaging targets. A local native build needs Go, GNU Make, Bash, GNU coreutils, CMake, a C++ compiler, the pinned llama.cpp sources, and the OpenVINO SDK on platforms that include that backend:

make deps-modeld
task modeld:build

On macOS, use make deps-llamacpp-ref in place of make deps-modeld; the worker uses llama.cpp and Metal there. The pinned versions and native flags live in mk/ and the backend Makefiles. Windows native builds use the existing MinGW or Clang/MSVC dependency-bundle path; they are not Linux cross-builds.

OpenVINO runtime, GenAI and tokenizer wheels must match the pinned C++ headers. make deps-openvino installs those versions; make -f Makefile.openvino check-sdk checks an existing SDK. A newer wheel can link successfully against older headers and crash during inference. Dependency bundles carry the verified SDK version; rebuild older bundles that lack it rather than copying the marker into them.

On Linux hosts with a RAM-backed /tmp, use a disk-backed temporary directory for native compilation, for example TMPDIR=/var/tmp make deps-modeld. Keep build directories on disk too. CUDA compilation can exhaust a small tmpfs even when the destination filesystem has ample free space. The CUDA toolkit must provide nvcc on PATH; use a host C++ compiler supported by that toolkit. After building, verify inference on the GPU: a successful build alone does not establish driver compatibility or GPU offload.

task dev-install is the complete checkout installation. It builds a relocatable modeld bundle with every backend supported on the current platform, installs it below the managed modeld data root, updates the same current pointer used by a downloaded release, and exposes the newly built CLI at the path the shell actually resolves. On Linux it includes CUDA automatically when nvcc is on PATH or at /usr/local/cuda/bin/nvcc. NVCC_CCBIN selects a compatible host compiler when the system compiler is newer than the CUDA toolkit supports.

task dev-link remains the fast CLI-only loop. When using an unpackaged worker instead, set CONTENOX_MODELD_BIN to the absolute path of bin/modeld (bin/modeld.exe on Windows). contenox modeld status probes its lease and endpoint. Models live below the worker data root, which defaults to ~/.contenox; CONTENOX_DATA_ROOT selects another root.

Native release pipe

The native release path remains separate from the pure-Go CLI release. It reuses the dependency bundles, checksums, package manifests, and release index from modeld. It does not upload anything during a build.

On each native build host:

make build-modeld
make bundle-modeld-deps
make push-modeld-deps MODELD_DEPS_S3_URI=s3://YOUR-BUCKET/modeld-deps

To consume a prepared dependency bundle on its matching platform:

task modeld:package -- MODELD_DEPS_ROOT=/absolute/path/to/dependency-bundle

The packager checks required libraries, links the worker, runs the packaged version --json smoke probe, and creates the archive, checksum, and index metadata under dist/. It refuses to silently omit OpenVINO when required by the selected profile. MODELD_RELEASE_OPENVINO=0 explicitly selects a llama-only profile.

To publish through the existing store mechanism:

make push-modeld-release MODELD_RELEASE_S3_URI=s3://YOUR-BUCKET/modeld

Store arguments also accept local directories for offline assembly and testing. The release prefix must be exposed over HTTP for installation. Then users run:

contenox modeld install --base-url https://YOUR-RELEASE-HOST/modeld

The installer reads index.json, chooses a compatible stable worker, verifies its checksum and transport/backend report, and updates the managed installation pointer. Compatibility is based on transport protocol and backend capability; the worker need not have the same version string as the CLI. The command’s default URL retains the existing modeld release prefix; copying source does not publish new artifacts there.

The platform matrix in .github/workflows/release.yml remains the supported CLI matrix. task modeld:check compiles worker control paths for that same matrix without native engines. This is a compile check, not evidence that native libraries load or inference works on those machines. Native package smoke and backend tests must run on the corresponding platform with its actual SDKs and models before publishing that worker.

To verify the actual package installation using a local release store:

make push-modeld-release MODELD_RELEASE_S3_URI=/tmp/modeld-release-store
MODELD_TEST_RELEASE_DIR=/tmp/modeld-release-store go test ./internal/modeldinstall -run '^TestSystem_EnsureInstalled_NativeRelease$' -count=1

This starts a local HTTP server, downloads and checks the archive, probes the worker after relocation into a temporary managed installation, and checks reuse. It does not load a model or establish inference correctness.