ci: add build-prerelease workflow for CUDA RPMs on rpm.lair.cafe

Adds a manually-triggered workflow that builds CUDA-flavoured neuron
binaries and a CPU cortex binary, packages them as Fedora RPMs, signs
them, and rsyncs to the unstable channel at
https://rpm.lair.cafe/fedora/43/x86_64/unstable/. Mirrors the build
pipeline used by grenade/mistralrs-package.

Pipeline:
- prepare: derive {version,short_sha,commit_date} from the checkout;
  the prerelease Release stamp "0.1.YYYYMMDDgitSHORTSHA" sorts below
  the eventual "1" stable release.
- build-cortex: cargo build --release -p cortex-cli on a rust runner.
- build-neuron: matrix over ada (sm_89) and blackwell (sm_120) on
  cuda-13.0 runners; cargo build with features "cuda cudnn flash-attn"
  and CUDA_COMPUTE_CAP set per flavour.
- package-{cortex,neuron}: rpmbuild on the rpm runner against the new
  prebuilt-binary specs in rpm/.
- publish: import signing key, sign RPMs, rsync to oolon, createrepo_c
  --update, then regenerate packages.json for the UI.

New specs are prebuilt-binary variants — they consume the artifact
from the build job rather than running cargo at rpmbuild time. Each
helexa-neuron-{flavour} package Conflicts with the other flavours and
with helexa-neuron (the future source-build stable package) so one
flavour is installed at a time on a given host.

neuron crate gains cudnn and flash-attn feature flags forwarding to
the corresponding candle features, so the CI build command compiles
those kernels into the binary.

sccache is intentionally NOT used in the prerelease jobs — CUDA
compute cap isn't in its cache key, so flavours would mis-hit each
other. Each prerelease build is a clean cargo build.

Required Gitea secrets (already in place for cortex.spec / COPR
workflow):
- RPM_SIGNING_KEY, RPM_SIGNING_KEY_ID
- RSYNC_SSH_KEY

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.gitea/workflows/build-prerelease.yml

@@ -0,0 +1,319 @@
+name: build-prerelease
+
+# Manually-dispatched workflow that builds CUDA-flavoured neuron binaries
+# (and a single cortex binary), packages each as a Fedora RPM, signs
+# them, and publishes to the `unstable` channel at rpm.lair.cafe.
+#
+# Trigger from the Gitea UI: Actions → build-prerelease → Run workflow.
+# Optionally provide a `ref` to build from a non-default branch.
+#
+# The published packages are versioned as e.g.
+#   helexa-neuron-blackwell-0.1.16-0.1.20260518gitabcdef0.fc43.x86_64
+# so they sort BELOW the eventual 0.1.16-1 stable release.
+
+on:
+  workflow_dispatch:
+    inputs:
+      ref:
+        description: "Git ref to build (branch / tag / commit). Defaults to the workflow's branch."
+        required: false
+        default: ""
+
+concurrency:
+  group: prerelease-build
+  cancel-in-progress: true
+
+env:
+  CARGO_INCREMENTAL: "0"
+
+jobs:
+  prepare:
+    name: Resolve version stamps
+    runs-on: rust
+    outputs:
+      version: ${{ steps.info.outputs.version }}
+      release: ${{ steps.info.outputs.release }}
+      short_sha: ${{ steps.info.outputs.short_sha }}
+      commit_date: ${{ steps.info.outputs.commit_date }}
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref }}
+          fetch-depth: 0
+
+      - id: info
+        run: |
+          set -eux
+          VERSION=$(awk -F\" '/^version[[:space:]]*=/ { print $2; exit }' Cargo.toml)
+          SHORT_SHA=$(git rev-parse --short=7 HEAD)
+          COMMIT_DATE=$(git log -1 --format=%cd --date=format:%Y%m%d HEAD)
+          # Prerelease release stamp sorts before "1" (the stable release).
+          RELEASE="0.1.${COMMIT_DATE}git${SHORT_SHA}"
+          echo "version=${VERSION}" >> "$GITHUB_OUTPUT"
+          echo "release=${RELEASE}" >> "$GITHUB_OUTPUT"
+          echo "short_sha=${SHORT_SHA}" >> "$GITHUB_OUTPUT"
+          echo "commit_date=${COMMIT_DATE}" >> "$GITHUB_OUTPUT"
+
+  build-cortex:
+    name: Build cortex binary
+    needs: prepare
+    runs-on: rust
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref }}
+
+      - name: Install/update Rust toolchain
+        run: |
+          if command -v rustup &> /dev/null; then
+            rustup update stable
+          else
+            curl --proto '=https' --tlsv1.2 --silent --show-error --fail https://sh.rustup.rs | sh -s -- -y
+          fi
+          echo "${HOME}/.cargo/bin" >> "$GITHUB_PATH"
+
+      - name: Build cortex (release)
+        run: cargo build --release -p cortex-cli
+
+      - name: Stage binary
+        run: |
+          mkdir --parents artifacts
+          cp target/release/cortex artifacts/cortex
+          ./artifacts/cortex --version || true
+
+      - uses: actions/upload-artifact@v3
+        with:
+          name: cortex-fc43
+          path: artifacts/cortex
+          retention-days: 1
+
+  build-neuron:
+    name: Build neuron-${{ matrix.flavour }}
+    needs: prepare
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - flavour: ada
+            compute_cap: "89"
+            runner: cuda-13.0
+            cuda_home: /usr/local/cuda-13.0
+            build_jobs: 8
+            nvcc_threads: 4
+            cargo_features: "cuda cudnn flash-attn"
+          - flavour: blackwell
+            compute_cap: "120"
+            runner: cuda-13.0
+            cuda_home: /usr/local/cuda-13.0
+            build_jobs: 8
+            nvcc_threads: 4
+            cargo_features: "cuda cudnn flash-attn"
+    runs-on: ${{ matrix.runner }}
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref }}
+
+      - name: Install/update Rust toolchain
+        run: |
+          if command -v rustup &> /dev/null; then
+            rustup update stable
+          else
+            curl --proto '=https' --tlsv1.2 --silent --show-error --fail https://sh.rustup.rs | sh -s -- -y
+          fi
+          echo "${HOME}/.cargo/bin" >> "$GITHUB_PATH"
+
+      - name: Build neuron with CUDA (${{ matrix.flavour }})
+        run: |
+          set -eux
+          export PATH="${{ matrix.cuda_home }}/bin:${PATH}"
+          export LD_LIBRARY_PATH="${{ matrix.cuda_home }}/targets/x86_64-linux/lib:${{ matrix.cuda_home }}/lib64:${LD_LIBRARY_PATH:-}"
+          export LIBRARY_PATH="${{ matrix.cuda_home }}/targets/x86_64-linux/lib:${{ matrix.cuda_home }}/lib64:${LIBRARY_PATH:-}"
+          cargo build --release -p neuron --features "${{ matrix.cargo_features }}"
+        env:
+          CUDA_COMPUTE_CAP: ${{ matrix.compute_cap }}
+          CARGO_BUILD_JOBS: ${{ matrix.build_jobs }}
+          NVCC_THREADS: ${{ matrix.nvcc_threads }}
+
+      - name: Stage binary
+        run: |
+          mkdir --parents artifacts
+          cp target/release/neuron artifacts/neuron-${{ matrix.flavour }}
+          file "artifacts/neuron-${{ matrix.flavour }}"
+
+      - uses: actions/upload-artifact@v3
+        with:
+          name: neuron-${{ matrix.flavour }}-fc43
+          path: artifacts/neuron-${{ matrix.flavour }}
+          retention-days: 1
+
+  package-cortex:
+    name: Package cortex RPM
+    needs: [prepare, build-cortex]
+    runs-on: rpm
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref }}
+
+      - uses: actions/download-artifact@v3
+        with:
+          name: cortex-fc43
+          path: artifacts/
+
+      - name: Build RPM
+        run: |
+          set -eux
+          rm -f ~/.rpmmacros
+          rpmdev-setuptree
+          cp artifacts/cortex ~/rpmbuild/SOURCES/
+          cp data/cortex.service ~/rpmbuild/SOURCES/
+          cp data/cortex-sysusers.conf ~/rpmbuild/SOURCES/
+          cp data/cortex-firewalld.xml ~/rpmbuild/SOURCES/
+          cp cortex.example.toml ~/rpmbuild/SOURCES/
+          cp models.example.toml ~/rpmbuild/SOURCES/
+          cp LICENSE ~/rpmbuild/SOURCES/
+          rpmbuild -bb rpm/cortex-prerelease.spec \
+            --define "cortex_version ${{ needs.prepare.outputs.version }}" \
+            --define "cortex_prerelease ${{ needs.prepare.outputs.release }}" \
+            --undefine dist \
+            --define "dist .fc43"
+
+      - uses: actions/upload-artifact@v3
+        with:
+          name: rpm-cortex-fc43
+          path: ~/rpmbuild/RPMS/x86_64/*.rpm
+          retention-days: 7
+
+  package-neuron:
+    name: Package helexa-neuron-${{ matrix.flavour }} RPM
+    needs: [prepare, build-neuron]
+    runs-on: rpm
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - flavour: ada
+          - flavour: blackwell
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref }}
+
+      - uses: actions/download-artifact@v3
+        with:
+          name: neuron-${{ matrix.flavour }}-fc43
+          path: artifacts/
+
+      - name: Build RPM
+        run: |
+          set -eux
+          rm -f ~/.rpmmacros
+          rpmdev-setuptree
+          cp artifacts/neuron-${{ matrix.flavour }} ~/rpmbuild/SOURCES/
+          cp data/neuron.service ~/rpmbuild/SOURCES/
+          cp data/neuron-sysusers.conf ~/rpmbuild/SOURCES/
+          cp data/neuron-firewalld.xml ~/rpmbuild/SOURCES/
+          cp neuron.example.toml ~/rpmbuild/SOURCES/
+          cp LICENSE ~/rpmbuild/SOURCES/
+          rpmbuild -bb rpm/helexa-neuron-prerelease.spec \
+            --define "neuron_version ${{ needs.prepare.outputs.version }}" \
+            --define "neuron_flavour ${{ matrix.flavour }}" \
+            --define "neuron_prerelease ${{ needs.prepare.outputs.release }}" \
+            --undefine dist \
+            --define "dist .fc43"
+
+      - uses: actions/upload-artifact@v3
+        with:
+          name: rpm-neuron-${{ matrix.flavour }}-fc43
+          path: ~/rpmbuild/RPMS/x86_64/*.rpm
+          retention-days: 7
+
+  publish:
+    name: Publish to rpm.lair.cafe (unstable)
+    needs: [package-cortex, package-neuron]
+    runs-on: rpm
+    concurrency:
+      group: rpm-publish
+      cancel-in-progress: false
+    env:
+      RPM_REPO_HOST: oolon.kosherinata.internal
+      FEDORA_VERSION: "43"
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref }}
+
+      - name: Download all built RPMs
+        uses: actions/download-artifact@v3
+        with:
+          path: rpms/
+          pattern: rpm-*-fc43
+
+      - name: Flatten RPM artifacts
+        run: |
+          set -eux
+          find rpms/ -name '*.rpm' -exec mv --target-directory=rpms/ {} +
+          find rpms/ -mindepth 1 -type d -empty -delete
+          ls -la rpms/
+
+      - name: Check for sequoia-sq
+        run: |
+          if ! command -v sq &> /dev/null; then
+            echo "ERROR: sequoia-sq is not installed. Install with: sudo dnf install sequoia-sq"
+            exit 1
+          fi
+
+      - name: Import signing key
+        run: |
+          echo "${{ secrets.RPM_SIGNING_KEY }}" | gpg --batch --import
+          fpr=$(gpg --batch --with-colons --list-keys "${{ secrets.RPM_SIGNING_KEY_ID }}" | awk -F: '/^fpr:/ { print $10; exit }')
+          echo "${fpr}:6:" | gpg --batch --import-ownertrust
+          sed "s/@GPG_NAME@/${{ secrets.RPM_SIGNING_KEY_ID }}/" rpm/rpmmacros > ~/.rpmmacros
+
+      - name: Sign RPMs
+        run: |
+          set -eux
+          for rpm in rpms/*.rpm; do
+            echo "signing ${rpm}..."
+            rpm --addsign "${rpm}"
+          done
+
+      - name: Set up SSH for rsync
+        run: |
+          install --directory --mode 700 ~/.ssh
+          echo "${RSYNC_SSH_KEY}" | install --mode 600 /dev/stdin ~/.ssh/id_ed25519
+        env:
+          RSYNC_SSH_KEY: ${{ secrets.RSYNC_SSH_KEY }}
+
+      - name: Test SSH connectivity
+        run: |
+          ssh -o StrictHostKeyChecking=accept-new "gitea_ci@${RPM_REPO_HOST}" exit
+
+      - name: Ensure unstable repo directory exists
+        run: |
+          ssh "gitea_ci@${RPM_REPO_HOST}" \
+            "mkdir --parents /var/www/rpm/fedora/${FEDORA_VERSION}/x86_64/unstable"
+
+      - name: Sync RPMs to unstable repo
+        run: |
+          rsync \
+            --archive \
+            --verbose \
+            --chmod D755,F644 \
+            rpms/*.rpm \
+            "gitea_ci@${RPM_REPO_HOST}:/var/www/rpm/fedora/${FEDORA_VERSION}/x86_64/unstable/"
+
+      - name: Update unstable repo metadata
+        run: |
+          ssh "gitea_ci@${RPM_REPO_HOST}" \
+            "cd /var/www/rpm/fedora/${FEDORA_VERSION}/x86_64/unstable && createrepo_c --update ."
+
+      - name: Generate packages.json manifest
+        run: |
+          scp script/generate-packages-json.py "gitea_ci@${RPM_REPO_HOST}:/tmp/"
+          ssh "gitea_ci@${RPM_REPO_HOST}" \
+            "python3 /tmp/generate-packages-json.py \
+              --repodata-dir /var/www/rpm/fedora/${FEDORA_VERSION}/x86_64/unstable/repodata \
+              --output /var/www/rpm/fedora/${FEDORA_VERSION}/x86_64/unstable/packages.json \
+              --base-url https://rpm.lair.cafe/fedora/${FEDORA_VERSION}/x86_64/unstable"

.gitea/workflows/ci.yml

@@ -8,6 +8,7 @@ on:
     branches: [main]
 
 env:
+  CARGO_INCREMENTAL: "0"
   RUSTC_WRAPPER: sccache
   SCCACHE_BUCKET: sccache
   SCCACHE_ENDPOINT: http://caveman.kosherinata.internal:9000
@@ -17,44 +18,38 @@ env:
   AWS_SECRET_ACCESS_KEY: ${{ secrets.SCCACHE_S3_SECRET_KEY }}
 
 jobs:
-  check:
+  fmt:
-    name: Format, lint, build, test
+    name: Format
-    runs-on: fedora
+    runs-on: rust
     steps:
       - uses: actions/checkout@v4
+      - run: cargo fmt --check --all
 
-      - name: Ensure sccache with S3 support
+  clippy:
-        env:
+    name: Clippy
-          RUSTC_WRAPPER: ""
+    runs-on: rust
-        run: |
+    steps:
-          if sccache --version 2>/dev/null && sccache --show-stats 2>/dev/null; then
+      - uses: actions/checkout@v4
-            echo "sccache with S3 support already installed"
+      - run: cargo clippy --workspace -- -D warnings
-          else
+      - run: sccache --show-stats
-            cargo install sccache --features s3 --locked
-          fi
 
-      - name: Check formatting
+  test:
-        run: cargo fmt --check --all
+    name: Test
-
+    runs-on: rust
-      - name: Clippy
+    steps:
-        run: cargo clippy --workspace -- -D warnings
+      - uses: actions/checkout@v4
-
+      - run: cargo test --workspace
-      - name: Build
+      - run: sccache --show-stats
-        run: cargo build --workspace
-
-      - name: Test
-        run: cargo test --workspace
-
-      - name: Show sccache stats
-        run: sccache --show-stats
 
   srpm-cortex:
     name: Build cortex SRPM
-    runs-on: fedora
+    runs-on: rpm
-    needs: check
+    needs: [fmt, clippy, test]
     if: startsWith(github.ref, 'refs/tags/v')
     steps:
       - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
 
       - name: Determine version
         id: version
@@ -68,6 +63,12 @@ jobs:
           sed -i '/\[workspace\.package\]/,/\[/{ s/^version = ".*"/version = "'"${VERSION}"'"/ }' Cargo.toml
           sed -i "s/^Version:.*/Version:        ${VERSION}/" cortex.spec
 
+      - name: Generate changelog entry
+        uses: https://git.lair.cafe/actions/rpm-changelog@v1
+        with:
+          spec: cortex.spec
+          version: ${{ steps.version.outputs.VERSION }}
+
       - name: Generate source tarball
         run: |
           set -ex
@@ -102,11 +103,13 @@ jobs:
 
   srpm-neuron:
     name: Build neuron SRPM
-    runs-on: fedora
+    runs-on: rpm
-    needs: check
+    needs: [fmt, clippy, test]
     if: startsWith(github.ref, 'refs/tags/v')
     steps:
       - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
 
       - name: Determine version
         id: version
@@ -118,31 +121,37 @@ jobs:
         run: |
           VERSION="${{ steps.version.outputs.VERSION }}"
           sed -i '/\[workspace\.package\]/,/\[/{ s/^version = ".*"/version = "'"${VERSION}"'"/ }' Cargo.toml
-          sed -i "s/^Version:.*/Version:        ${VERSION}/" neuron.spec
+          sed -i "s/^Version:.*/Version:        ${VERSION}/" helexa-neuron.spec
+
+      - name: Generate changelog entry
+        uses: https://git.lair.cafe/actions/rpm-changelog@v1
+        with:
+          spec: helexa-neuron.spec
+          version: ${{ steps.version.outputs.VERSION }}
 
       - name: Generate source tarball
         run: |
           set -ex
           VERSION="${{ steps.version.outputs.VERSION }}"
-          tar czf /tmp/neuron-${VERSION}.tar.gz \
+          tar czf /tmp/helexa-neuron-${VERSION}.tar.gz \
-            --transform "s,^\.,neuron-${VERSION}," \
+            --transform "s,^\.,helexa-neuron-${VERSION}," \
             --exclude='./target' \
             --exclude='./.git' \
             --exclude='*.tar.gz' \
             --exclude='*.src.rpm' \
             .
-          mv /tmp/neuron-${VERSION}.tar.gz .
+          mv /tmp/helexa-neuron-${VERSION}.tar.gz .
 
       - name: Vendor Rust dependencies
         run: |
           VERSION="${{ steps.version.outputs.VERSION }}"
           cargo vendor vendor/
-          tar czf neuron-${VERSION}-vendor.tar.gz vendor/
+          tar czf helexa-neuron-${VERSION}-vendor.tar.gz vendor/
           rm -rf vendor/
 
       - name: Build SRPM
         run: |
-          rpmbuild -bs neuron.spec \
+          rpmbuild -bs helexa-neuron.spec \
             --define "_sourcedir $(pwd)" \
             --define "_srcrpmdir $(pwd)"
 
@@ -154,7 +163,7 @@ jobs:
 
   copr-cortex:
     name: Publish cortex to COPR
-    runs-on: fedora
+    runs-on: fedora-43
     needs: srpm-cortex
     steps:
       - name: Download SRPM
@@ -162,17 +171,16 @@ jobs:
         with:
           name: srpm-cortex
 
-      - name: Configure copr-cli
+      - name: Publish to COPR
-        run: |
+        uses: https://git.lair.cafe/actions/copr-publish@v1
-          mkdir -p ~/.config
+        with:
-          echo "${{ secrets.COPR_CONFIG }}" > ~/.config/copr
+          project: helexa/helexa
-
+          srpm: "*.src.rpm"
-      - name: Submit build to COPR
+          copr-config: ${{ secrets.COPR_CONFIG }}
-        run: copr-cli build helexa/cortex *.src.rpm
 
   copr-neuron:
     name: Publish neuron to COPR
-    runs-on: fedora
+    runs-on: fedora-43
     needs: srpm-neuron
     steps:
       - name: Download SRPM
@@ -180,35 +188,56 @@ jobs:
         with:
           name: srpm-neuron
 
-      - name: Configure copr-cli
+      - name: Publish to COPR
-        run: |
+        uses: https://git.lair.cafe/actions/copr-publish@v1
-          mkdir -p ~/.config
+        with:
-          echo "${{ secrets.COPR_CONFIG }}" > ~/.config/copr
+          project: helexa/helexa
-
+          srpm: "*.src.rpm"
-      - name: Submit build to COPR
+          copr-config: ${{ secrets.COPR_CONFIG }}
-        run: copr-cli build helexa/neuron *.src.rpm
 
   bump-version:
     name: Bump version in source
-    runs-on: fedora
+    runs-on: rust
     needs: [copr-cortex, copr-neuron]
     steps:
       - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
 
-      - name: Stamp version and push
+      - name: Determine version
+        id: version
+        run: echo "VERSION=${GITHUB_REF#refs/tags/v}" >> "$GITHUB_OUTPUT"
+
+      - name: Stamp version
+        run: |
+          VERSION="${{ steps.version.outputs.VERSION }}"
+          sed -i '/\[workspace\.package\]/,/\[/{ s/^version = ".*"/version = "'"${VERSION}"'"/ }' Cargo.toml
+          sed -i "s/^Version:.*/Version:        ${VERSION}/" cortex.spec
+          sed -i "s/^Version:.*/Version:        ${VERSION}/" helexa-neuron.spec
+          cargo check --workspace 2>/dev/null || true
+
+      - name: Generate cortex changelog entry
+        uses: https://git.lair.cafe/actions/rpm-changelog@v1
+        with:
+          spec: cortex.spec
+          version: ${{ steps.version.outputs.VERSION }}
+
+      - name: Generate helexa-neuron changelog entry
+        uses: https://git.lair.cafe/actions/rpm-changelog@v1
+        with:
+          spec: helexa-neuron.spec
+          version: ${{ steps.version.outputs.VERSION }}
+
+      - name: Commit and push
         env:
           GITEA_TOKEN: ${{ secrets.GITEA_TOKEN }}
         run: |
-          VERSION="${GITHUB_REF#refs/tags/v}"
+          VERSION="${{ steps.version.outputs.VERSION }}"
-          sed -i '/\[workspace\.package\]/,/\[/{ s/^version = ".*"/version = "'"${VERSION}"'"/ }' Cargo.toml
-          sed -i "s/^Version:.*/Version:        ${VERSION}/" cortex.spec
-          sed -i "s/^Version:.*/Version:        ${VERSION}/" neuron.spec
-          cargo check --workspace 2>/dev/null || true
           git config user.name "Gitea Actions"
           git config user.email "actions@git.lair.cafe"
-          git add Cargo.toml Cargo.lock cortex.spec neuron.spec
+          git add Cargo.toml Cargo.lock cortex.spec helexa-neuron.spec
           if git diff --cached --quiet; then
-            echo "Version already at ${VERSION}"
+            echo "Nothing to commit for ${VERSION}"
           else
             git commit -m "chore: bump version to ${VERSION}"
             git remote set-url origin "https://gitea-actions:${GITEA_TOKEN}@git.lair.cafe/helexa/cortex.git"

.gitignore

@@ -5,3 +5,4 @@
 .vscode/
 cortex.toml
 doc/plan/*
+script/deploy.sh

CLAUDE.md

@@ -125,7 +125,8 @@ automatically. Clippy warnings must be resolved, not suppressed with
   - One or more GPU nodes running mistral.rs on port 8080
   - Optionally a metrics-only node (no GPU) for Prometheus/Grafana
 - Each node runs `mistralrs serve` on port 8080
-- Gateway listens on port 8000 (API) and 9100 (metrics)
+- Gateway listens on port 31313 (API) and 31314 (metrics)
+- neuron listens on port 13131 on each GPU host
 - TLS terminated at gateway or via nginx; internal traffic is plaintext over WireGuard
 
 ## Conventions
@@ -380,7 +381,7 @@ processes (one process per loaded model, each on its own port).
 
 ## neuron API
 
-neuron exposes an HTTP API on port 9090 that cortex polls and calls.
+neuron exposes an HTTP API on port 13131 that cortex polls and calls.
 
 ```
 GET  /discovery
@@ -424,8 +425,8 @@ endpoint. cortex.toml shrinks to:
 
 ```toml
 [gateway]
-listen = "0.0.0.0:8000"
+listen = "0.0.0.0:31313"
-metrics_listen = "0.0.0.0:9100"
+metrics_listen = "0.0.0.0:31314"
 
 [eviction]
 strategy = "lru"
@@ -433,15 +434,15 @@ defrag_after_cycles = 50
 
 [[neurons]]
 name = "beast"
-endpoint = "http://beast.hanzalova.internal:9090"
+endpoint = "http://beast.hanzalova.internal:13131"
 
 [[neurons]]
 name = "benjy"
-endpoint = "http://benjy.kosherinata.internal:9090"
+endpoint = "http://benjy.hanzalova.internal:13131"
 
 [[neurons]]
 name = "quadbrat"
-endpoint = "http://quadbrat.hanzalova.internal:9090"
+endpoint = "http://quadbrat.hanzalova.internal:13131"
 ```
 
 On startup and periodically, cortex calls `GET /discovery` and
@@ -521,7 +522,7 @@ cortex/
 │   │       └── metrics.rs      # prometheus exporter (unchanged)
 │   ├── neuron/                 # node plane (replaces cortex-agent)
 │   │   └── src/
-│   │       ├── main.rs         # binary entrypoint, axum server on :9090
+│   │       ├── main.rs         # binary entrypoint, axum server on :13131
 │   │       ├── discovery.rs    # nvidia-smi, device enumeration
 │   │       ├── health.rs       # runtime GPU polling
 │   │       ├── api.rs          # HTTP handlers for /discovery, /models, etc.
@@ -595,70 +596,65 @@ placement matching can be added incrementally.
 Completed. Both packages have RPM specs, systemd units, and example configs.
 CI builds parallel SRPMs on tag push and publishes to separate COPR repos.
 
-- `cortex.spec` → `helexa/cortex` COPR: binary, systemd unit, config files
+- `cortex.spec` — installs the `cortex` binary. Package name keeps the
-- `neuron.spec` → `helexa/neuron` COPR: binary, systemd unit, config
+  short `cortex` because no Fedora package collides with it.
+- `helexa-neuron.spec` — installs the `neuron` binary under package name
+  `helexa-neuron`. Renamed from bare `neuron` to avoid collision with
+  Fedora's NEURON neural-simulation package
+  (https://src.fedoraproject.org/rpms/neuron); binary, systemd unit,
+  system user, and config dir all stay named `neuron` since those are
+  project-local contexts.
 - `data/cortex.service`, `data/neuron.service` — systemd units
 - `cortex.example.toml`, `neuron.example.toml`, `models.example.toml`
-- CI: parallel `srpm-cortex` + `srpm-neuron` jobs, then parallel COPR publish
+- CI: parallel `srpm-cortex` + `srpm-neuron` jobs, then parallel COPR
+  publish to a single project `helexa/helexa` hosting both packages.
 
 Install:
 ```sh
-dnf copr enable helexa/cortex && dnf install cortex    # gateway host
+dnf copr enable helexa/helexa
-dnf copr enable helexa/neuron && dnf install neuron    # GPU nodes
+dnf install cortex                # gateway host
+dnf install helexa-neuron         # GPU nodes
 ```
 
-### Phase 11: llama.cpp harness stub
+## 2026-05-18 addendum: candle-native pivot
 
-**Goal:** Prove the harness abstraction works with a second engine.
+Phases 11 (llama.cpp harness) and 12 (mistral.rs COPR) below are
+**superseded**. The project no longer treats mistral.rs or llama.cpp as
+dependencies — both are conceptually out of scope. neuron becomes a
+candle-native inference daemon, with `Harness` retained as an
+internal seam for adding future engines (vision/audio/diffusion) but
+its only implementation being in-process candle.
 
-**Steps:**
+The full staged plan for this pivot lives at
-1. `crates/neuron/src/harness/llamacpp.rs` — implement the `Harness`
+`~/.claude/plans/create-a-more-aggressive-calm-naur.md`. Summary:
-   trait for llama.cpp's `llama-server`.
-   - `start()` — launch `llama-server` with the correct model path,
-     `--port`, `--n-gpu-layers`, `--tensor-split` args. Track the
-     child process.
-   - `stop()` — send SIGTERM to the child process.
-   - `list_models()` — llama-server serves one model per process, so
-     return a single-element list.
-   - `load_model()` — start a new llama-server process for this model.
-   - `unload_model()` — stop the process.
-   - `inference_endpoint()` — return `http://localhost:{assigned_port}`.
-2. Port allocation: neuron assigns ports from a range (e.g. 8100-8199)
-   to llama-server instances.
-3. Register in `HarnessRegistry` when configured:
-   ```toml
-   [[harnesses]]
-   name = "llamacpp"
-   binary = "/usr/local/bin/llama-server"
-   port_range = [8100, 8199]
-   ```
-4. Tests: mock llama-server (simple HTTP server returning canned
-   responses), test load/unload/endpoint lifecycle.
 
-**Done when:** A model with `harness = "llamacpp"` in `models.toml` can
+- **Stage 1 (this commit):** delete `mistralrs.rs` and `llamacpp.rs`,
-be loaded and served through cortex. Tests pass with mock llama-server.
+  scaffold inert `CandleHarness`, drop `endpoint`/`systemd_unit` from
+  `HarnessConfig`, default no-op `start`/`stop` on the `Harness` trait.
+- **Stages 2–4:** wire up candle model load/unload (quantized Qwen3
+  first), add OpenAI-compatible inference endpoint in neuron, then SSE
+  streaming.
+- **Stages 5–6:** load-on-activation (default models in config) and
+  unload-on-deactivation (graceful shutdown).
+- **Stages 7–8:** multi-GPU tensor parallelism and broader model/quant
+  coverage.
 
-### Phase 12 (lower priority): mistral.rs COPR packaging
+Sections of this document that describe mistral.rs HTTP behaviour
+("mistral.rs API gotchas") are retained as historical context for
+Phases 1–10 — they document what was true while the project depended
+on mistral.rs. They do not describe current behaviour.
 
-**Goal:** Fedora RPMs for mistral.rs built against specific CUDA versions.
+---
 
-**Steps:**
+### Phase 11 (superseded): llama.cpp harness stub
-1. `mistralrs-cuda.spec` — RPM spec that clones a pinned mistral.rs git
-   tag, builds with `--features cuda`, links against the system CUDA
-   toolkit. Produces `mistralrs-cuda13-server` (CUDA 13.x / sm_120) and
-   `mistralrs-cuda12-server` (CUDA 12.x / sm_89). Install binary to
-   `/usr/local/bin/mistralrs`.
-2. COPR build config: enable the NVIDIA CUDA repo as a build dependency.
-   Pin the CUDA toolkit version in `BuildRequires`.
-3. Gitea Actions or manual workflow: bump the mistral.rs tag in the spec,
-   trigger COPR rebuild.
-4. neuron's mistralrs harness config references which binary/package
-   provides the mistral.rs binary. neuron could warn at startup if the
-   installed mistral.rs CUDA version doesn't match the discovered driver.
 
-**Done when:** `dnf install mistralrs-cuda13-server` on beast provides a
+~~Originally planned as a second engine to prove the harness
-working `mistralrs` binary built for Blackwell GPUs. `dnf install
+abstraction.~~ Replaced by the candle harness work in the 2026-05-18
-mistralrs-cuda12-server` on benjy provides one built for Ada GPUs.
+addendum above. llama.cpp's any-model/any-hardware breadth is no
+longer in scope for helexa.
 
-This is a separate repo/spec — not part of the cortex workspace — but
+### Phase 12 (superseded): mistral.rs COPR packaging
-tightly coupled operationally. Track it as a sibling project.
+
+~~Originally planned to ship CUDA-versioned mistral.rs RPMs.~~ Replaced
+by the candle harness work in the 2026-05-18 addendum above. With
+mistral.rs out of the dependency tree, there is nothing to package.

Cargo.toml

@@ -8,7 +8,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.1.0"
+version = "0.1.16"
 edition = "2024"
 license = "GPL-3.0-or-later"
 repository = "https://git.lair.cafe/helexa/cortex"
@@ -27,7 +27,7 @@ serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 toml = "0.8"
 
-# http client (for proxying to mistralrs backends)
+# http client (for proxying to neuron backends)
 reqwest = { version = "0.12", features = ["json", "stream"] }
 
 # observability

README.md

@@ -1,22 +1,23 @@
 # cortex
 
-A Rust reverse-proxy and fleet management layer for multi-node
+A Rust reverse-proxy and fleet management layer for multi-node GPU inference
-[mistral.rs](https://github.com/EricLBuehler/mistral.rs) inference clusters.
+clusters. Cortex sits in front of one or more `neuron` daemons (each running
+candle-based inference on a local GPU host) and presents a unified OpenAI +
+Anthropic compatible API surface.
 
 ## Problem
 
 Running local LLMs across multiple GPU nodes (different VRAM tiers, different
 model affinities) requires a unified API surface that:
 
-- Presents a **single `/v1/models` catalogue** merging every model across every
+- Presents a **single `/v1/models` catalogue** merging every model that can be
-  node.
+  served by any neuron in the fleet.
-- **Routes requests** to the correct node based on where a model is loaded (or
+- **Routes requests** to the correct node based on where a model is loaded
-  *can* be loaded).
+  (or can be loaded), handling cold-load and eviction transparently.
-- Manages **model lifecycle** — unload cold models, reload on demand, pin
+- Manages **model lifecycle** — load on demand, unload cold models, pin
-  critical ones — using the mistral.rs
+  critical ones — by calling each neuron's `/models/{load,unload}` API.
-  `/v1/models/{unload,reload,status}` HTTP API (PR #1828+).
 - Translates between **OpenAI and Anthropic** request/response envelopes so
-  every client in the homelab speaks whichever dialect it prefers.
+  every client speaks whichever dialect it prefers.
 - Captures **per-request metrics** (tokens, tok/s, TTFT, latency) and exposes
   them as Prometheus counters/histograms.
 
@@ -30,18 +31,17 @@ model affinities) requires a unified API surface that:
        └────────────────┴──────┬───────┴───────────────┘
                                │
                     ┌──────────▼──────────┐
-                    │   cortex     │
+                    │      cortex         │
-                    │   (cortex-gateway)      │
+                    │  (cortex-gateway)   │
                     │                     │
                     │  Router · Metrics   │
                     │  Evictor · Translate│
                     └──┬──────┬────────┬──┘
                        │      │        │
             ┌──────────▼┐  ┌──▼─────┐  ┌▼──────────┐
-            │ gpu-large │  │gpu-med │  │ gpu-small │
+            │  neuron   │  │ neuron │  │  neuron   │
-            │ mistralrs │  │mistral │  │ mistralrs │
+            │  :13131   │  │ :13131 │  │  :13131   │
-            │ serve     │  │rs serve│  │ serve     │
+            │  candle   │  │ candle │  │  candle   │
-            │ :8080     │  │ :8080  │  │  :8080    │
             └───────────┘  └────────┘  └───────────┘
                   private network (.internal)
 ```
@@ -50,70 +50,48 @@ model affinities) requires a unified API surface that:
 
 | Crate | Purpose |
 |---|---|
-| `cortex-core` | Shared types: config, node/model state, metrics, OpenAI/Anthropic request/response envelopes |
+| `cortex-core` | Shared types: config, node/model state, metrics, OpenAI/Anthropic envelopes, harness trait, discovery types |
-| `cortex-gateway` | Axum HTTP server: proxy, router, evictor, metrics exporter |
+| `cortex-gateway` | Axum HTTP server: proxy, router, evictor, poller, metrics exporter |
-| `cortex-agent` | Per-node sidecar: polls local mistralrs, reports to gateway, handles restart/defrag |
+| `neuron` | Per-node daemon: GPU discovery, in-process candle inference, model lifecycle API |
 | `cortex-cli` | CLI entrypoint (`cortex serve`, `cortex status`, etc.) |
 
 ## Node setup
 
-Each GPU node runs `mistralrs serve` with a multi-model config. Models are
+Each GPU node runs `neuron` (listening on `:13131`). Neuron uses
-declared but start **unloaded** — mistral.rs lazy-loads on first request and
+huggingface/candle for in-process inference — there is no external
-the gateway can explicitly unload/reload via the HTTP API.
+inference subprocess to manage.
 
-Example node systemd unit:
+The neuron RPM (`helexa-neuron`) ships a systemd unit:
 
-```ini
+```sh
-# /etc/systemd/system/mistralrs.service
+dnf copr enable helexa/helexa
-[Unit]
+dnf install helexa-neuron
-Description=mistral.rs inference server
+systemctl enable --now neuron
-After=network-online.target
-Wants=network-online.target
-
-[Service]
-Type=simple
-ExecStart=/usr/local/bin/mistralrs serve \
-    --from-config /etc/mistralrs/config.toml \
-    --port 8080
-Restart=on-failure
-RestartSec=5
-Environment=CUDA_VISIBLE_DEVICES=0,1
-
-[Install]
-WantedBy=multi-user.target
 ```
 
 ## Gateway config
 
 ```toml
-# cortex.toml
+# /etc/cortex/cortex.toml
 [gateway]
-listen = "0.0.0.0:8000"
+listen = "0.0.0.0:31313"
-metrics_listen = "0.0.0.0:9100"
+metrics_listen = "0.0.0.0:31314"
 
 [eviction]
 strategy = "lru"        # lru | priority
 defrag_after_cycles = 50
 
-[[nodes]]
+[[neurons]]
-name = "gpu-large"
+name = "beast"
-endpoint = "http://gpu-large.internal:8080"
+endpoint = "http://beast.internal:13131"
-vram_mb = 49_152        # e.g. 2x RTX 4090
-pinned = ["your-org/large-model"]
 
-[[nodes]]
+[[neurons]]
-name = "gpu-medium"
+name = "benjy"
-endpoint = "http://gpu-medium.internal:8080"
+endpoint = "http://benjy.internal:13131"
-vram_mb = 24_576        # e.g. RTX 4090
-pinned = ["your-org/medium-model"]
-
-[[nodes]]
-name = "gpu-small"
-endpoint = "http://gpu-small.internal:8080"
-vram_mb = 12_288        # e.g. RTX 3060
-pinned = ["your-org/embedding-model"]
 ```
 
+Model placement profiles live in `models.toml` — see `models.example.toml`.
+
 ## Building
 
 ```sh
@@ -131,19 +109,20 @@ cargo clippy --workspace -- -D warnings   # warnings are errors
 cargo test --workspace                     # all tests must pass
 ```
 
-Tagged releases (`v*`) additionally build an SRPM and publish to COPR.
+Tagged releases (`v*`) additionally build SRPMs for both `cortex` and
+`helexa-neuron` and publish to COPR.
 
 ## Running
 
 ```sh
 # start the gateway
-cortex serve --config cortex.toml
+cortex serve --config /etc/cortex/cortex.toml
 
 # check fleet status
 cortex status
 
 # list all models across nodes
-curl http://localhost:8000/v1/models
+curl http://localhost:31313/v1/models
 ```
 
 ## License

cortex.example.toml

@@ -3,22 +3,22 @@
 # Copy to cortex.toml and adjust for your environment.
 #
 # Environment variable overrides use CORTEX_ prefix with __ separators:
-#   CORTEX_GATEWAY__LISTEN=0.0.0.0:9000
+#   CORTEX_GATEWAY__LISTEN=0.0.0.0:31313
 
 [gateway]
-listen = "0.0.0.0:8000"
+listen = "0.0.0.0:31313"
-metrics_listen = "0.0.0.0:9100"
+metrics_listen = "0.0.0.0:31314"
 
 [eviction]
 strategy = "lru"
-# Restart mistralrs after this many load/unload cycles to defragment VRAM.
+# Restart neurons after this many load/unload cycles to defragment VRAM.
 # Set to 0 to disable.
 defrag_after_cycles = 50
 
 # -- Nodes ---------------------------------------------------------------
-# Each [[nodes]] entry declares a mistral.rs instance in the fleet.
+# Each [[nodes]] entry declares a neuron daemon in the fleet.
-# Models are discovered by polling the node's /v1/models endpoint.
+# Models are discovered by polling the neuron's /models endpoint.
-# Pinned models are never evicted.
+# Pinned models (see models.toml) are never evicted.
 
 [[nodes]]
 name = "gpu-large"

cortex.spec

@@ -1,5 +1,5 @@
 Name:           cortex
-Version:        0.1.0
+Version:        0.1.16
 Release:        1%{?dist}
 Summary:        Inference gateway for multi-node GPU clusters
 
@@ -13,9 +13,24 @@ ExclusiveArch:  x86_64
 BuildRequires:  rust >= 1.85
 BuildRequires:  cargo
 BuildRequires:  gcc
+BuildRequires:  gcc-c++
+BuildRequires:  cmake
+BuildRequires:  perl-interpreter
+BuildRequires:  pkgconfig(openssl)
 BuildRequires:  systemd-rpm-macros
 
 Requires(pre):  shadow-utils
+Requires:       systemd
+Requires:       firewalld-filesystem
+
+# systemd-rpm-macros ships a unit dep generator that parses User=/Group=
+# from our .service file and emits Requires: user(cortex)/group(cortex).
+# rpm's sysusers provides-generator emits the unversioned form for groups
+# but only a versioned user(cortex) = <base64> for users with GECOS/home/
+# shell. Provide the unversioned user(cortex) explicitly so dnf can resolve
+# the auto-generated Requires. Without this, dnf5 silently filters the
+# package and reports "Nothing to do".
+Provides:       user(cortex)
 
 %description
 Cortex is a Rust reverse-proxy that sits in front of multiple inference
@@ -41,13 +56,14 @@ cargo build --release -p cortex-cli
 %install
 install -Dm755 target/release/cortex %{buildroot}%{_bindir}/cortex
 install -Dm644 data/cortex.service %{buildroot}%{_unitdir}/cortex.service
-install -dm750 %{buildroot}%{_sysconfdir}/cortex
+install -Dm644 data/cortex-sysusers.conf %{buildroot}%{_sysusersdir}/cortex.conf
-install -Dm640 cortex.example.toml %{buildroot}%{_sysconfdir}/cortex/cortex.toml
+install -Dm644 data/cortex-firewalld.xml %{buildroot}%{_prefix}/lib/firewalld/services/cortex.xml
-install -Dm640 models.example.toml %{buildroot}%{_sysconfdir}/cortex/models.toml
+install -dm755 %{buildroot}%{_sysconfdir}/cortex
+install -Dm644 cortex.example.toml %{buildroot}%{_sysconfdir}/cortex/cortex.toml
+install -Dm644 models.example.toml %{buildroot}%{_sysconfdir}/cortex/models.toml
 
 %pre
-getent group cortex >/dev/null || groupadd -r cortex
+%sysusers_create_compat %{_builddir}/%{name}-%{version}/data/cortex-sysusers.conf
-getent passwd cortex >/dev/null || useradd -r -g cortex -d /var/lib/cortex -s /sbin/nologin cortex
 
 %post
 %systemd_post cortex.service
@@ -63,10 +79,22 @@ getent passwd cortex >/dev/null || useradd -r -g cortex -d /var/lib/cortex -s /s
 %doc README.md
 %{_bindir}/cortex
 %{_unitdir}/cortex.service
-%dir %attr(750,root,cortex) %{_sysconfdir}/cortex
+%{_sysusersdir}/cortex.conf
-%config(noreplace) %attr(640,root,cortex) %{_sysconfdir}/cortex/cortex.toml
+%{_prefix}/lib/firewalld/services/cortex.xml
-%config(noreplace) %attr(640,root,cortex) %{_sysconfdir}/cortex/models.toml
+%dir %{_sysconfdir}/cortex
+%config(noreplace) %{_sysconfdir}/cortex/cortex.toml
+%config(noreplace) %{_sysconfdir}/cortex/models.toml
 
 %changelog
-* Tue Apr 15 2026 Rob Thijssen <grenade@rob.tn> - 0.1.0-1
+* Thu Apr 16 2026 Gitea Actions <actions@git.lair.cafe> - 0.1.16-1
+- chore: ignore local deploy script
+- chore: move default ports out of common-collision ranges
+- ci: drop actions/cache for cargo registry and target
+
+* Thu Apr 16 2026 Gitea Actions <actions@git.lair.cafe> - 0.1.14-1
+- ci: publish both packages to a single helexa/helexa COPR project
+- fix(rpm): rename neuron package to helexa-neuron
+- ci: commit generated %changelog entries back to main
+
+* Wed Apr 15 2026 Rob Thijssen <grenade@rob.tn> - 0.1.0-1
 - Initial package

crates/cortex-cli/src/main.rs

@@ -5,7 +5,7 @@ use tracing_subscriber::EnvFilter;
 
 #[derive(Parser)]
 #[command(name = "cortex")]
-#[command(about = "Unified inference gateway for multi-node mistral.rs clusters")]
+#[command(about = "Unified inference gateway for multi-node GPU clusters")]
 #[command(version)]
 struct Cli {
     #[command(subcommand)]
@@ -23,7 +23,7 @@ enum Commands {
     /// Print the fleet status (models, nodes, health).
     Status {
         /// Gateway API endpoint to query.
-        #[arg(short, long, default_value = "http://localhost:8000")]
+        #[arg(short, long, default_value = "http://localhost:31313")]
         endpoint: String,
     },
 }

crates/cortex-core/src/anthropic.rs

@@ -2,7 +2,7 @@
 //!
 //! These mirror the `/v1/messages` format used by the Anthropic API.
 //! The gateway accepts these, translates to OpenAI format, proxies to
-//! mistral.rs, then translates the response back.
+//! the inference backend (neuron), then translates the response back.
 
 use serde::{Deserialize, Serialize};
 use serde_json::Value;

crates/cortex-core/src/config.rs

@@ -22,9 +22,9 @@ fn default_models_path() -> String {
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct GatewaySettings {
-    /// Address to listen on for API requests (e.g. "0.0.0.0:8000")
+    /// Address to listen on for API requests (e.g. "0.0.0.0:31313")
     pub listen: String,
-    /// Address to listen on for Prometheus metrics (e.g. "0.0.0.0:9100")
+    /// Address to listen on for Prometheus metrics (e.g. "0.0.0.0:31314")
     pub metrics_listen: String,
 }
 
@@ -50,7 +50,7 @@ pub enum EvictionStrategy {
 pub struct NeuronEndpoint {
     /// Human-readable node name (e.g. "beast")
     pub name: String,
-    /// Base URL of the neuron daemon (e.g. "http://beast.internal:9090")
+    /// Base URL of the neuron daemon (e.g. "http://beast.internal:13131")
     pub endpoint: String,
 }
 
@@ -70,8 +70,8 @@ impl Default for GatewayConfig {
     fn default() -> Self {
         Self {
             gateway: GatewaySettings {
-                listen: "0.0.0.0:8000".into(),
+                listen: "0.0.0.0:31313".into(),
-                metrics_listen: "0.0.0.0:9100".into(),
+                metrics_listen: "0.0.0.0:31314".into(),
             },
             eviction: EvictionSettings {
                 strategy: EvictionStrategy::Lru,

crates/cortex-core/src/harness.rs

@@ -9,13 +9,13 @@ use async_trait::async_trait;
 use serde::{Deserialize, Serialize};
 
 /// Configuration for a harness instance on a neuron.
+///
+/// All current harnesses are in-process (candle); per-harness tuning
+/// (cache paths, device policies, etc.) lives in dedicated config
+/// blocks rather than on this struct.
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct HarnessConfig {
     pub name: String,
-    /// Base URL of the harness (e.g. "http://localhost:8080" for mistral.rs).
-    pub endpoint: Option<String>,
-    /// Systemd unit name, if the harness is managed via systemd.
-    pub systemd_unit: Option<String>,
 }
 
 /// Health status of a harness process.
@@ -47,16 +47,24 @@ pub struct ModelInfo {
 }
 
 /// What an inference harness must do, from neuron's perspective.
+///
+/// All current harnesses are in-process — they share neuron's address
+/// space and lifecycle. `start`/`stop` therefore default to no-ops; a
+/// future process-supervising harness would override them.
 #[async_trait]
 pub trait Harness: Send + Sync {
-    /// Human-readable name (e.g. "mistralrs", "llamacpp", "comfyui").
+    /// Human-readable name (e.g. "candle").
     fn name(&self) -> &str;
 
-    /// Start the harness process if it is not already running.
+    /// Start the harness. Default no-op for in-process harnesses.
-    async fn start(&self, config: &HarnessConfig) -> Result<()>;
+    async fn start(&self, _config: &HarnessConfig) -> Result<()> {
+        Ok(())
+    }
 
-    /// Stop the harness process gracefully.
+    /// Stop the harness. Default no-op for in-process harnesses.
-    async fn stop(&self) -> Result<()>;
+    async fn stop(&self) -> Result<()> {
+        Ok(())
+    }
 
     /// Health check. Returns the harness process status.
     async fn health(&self) -> HarnessHealth;

crates/cortex-core/src/node.rs

@@ -6,7 +6,7 @@ use std::collections::HashMap;
 #[derive(Debug, Clone)]
 pub struct NodeState {
     pub name: String,
-    /// Base URL of the neuron daemon (e.g. "http://beast.internal:9090").
+    /// Base URL of the neuron daemon (e.g. "http://beast.internal:13131").
     pub endpoint: String,
     pub healthy: bool,
     pub models: HashMap<String, ModelEntry>,

crates/cortex-core/src/openai.rs

@@ -3,7 +3,7 @@
 //! These are a subset sufficient for chat completions (streaming + non-streaming).
 //! Fields not relevant to proxying are captured as `serde_json::Value` via
 //! `#[serde(flatten)]` so we forward them without needing to enumerate every
-//! extension field mistral.rs supports.
+//! extension field a backend might support.
 
 use serde::{Deserialize, Serialize};
 use serde_json::Value;
@@ -22,7 +22,7 @@ pub struct ChatCompletionRequest {
     pub max_tokens: Option<u64>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub stream: Option<bool>,
-    /// All other fields (tools, response_format, mistral.rs extensions, etc.)
+    /// All other fields (tools, response_format, backend extensions, etc.)
     #[serde(flatten)]
     pub extra: Value,
 }

crates/cortex-gateway/src/proxy.rs

@@ -1,4 +1,4 @@
-//! Streaming HTTP reverse proxy to mistral.rs backends.
+//! Streaming HTTP reverse proxy to neuron backends.
 //!
 //! For streaming requests, SSE chunks are forwarded as they arrive.
 //! The proxy captures timing information for metrics but does not

crates/cortex-gateway/tests/common/mod.rs

@@ -22,6 +22,7 @@ use tokio::net::TcpListener;
 /// - GET /models/:id/endpoint (returns the inference URL)
 /// - POST /models/unload (accepts unload requests)
 /// - GET /v1/chat/completions + POST /v1/chat/completions (inference)
+///
 /// Returns the neuron base URL.
 pub async fn spawn_mock_neuron() -> String {
     let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
@@ -54,7 +55,7 @@ pub async fn spawn_mock_neuron() -> String {
 
 async fn mock_neuron_list_models() -> Json<Value> {
     Json(json!([
-        {"id": "test-model", "harness": "mistralrs", "status": "loaded", "devices": [0], "vram_used_mb": 8000}
+        {"id": "test-model", "harness": "candle", "status": "loaded", "devices": [0], "vram_used_mb": 8000}
     ]))
 }
 

crates/cortex-gateway/tests/poller.rs

@@ -12,8 +12,8 @@ use std::sync::Arc;
 async fn test_poller_discovers_models() {
     // Mock neuron reports 2 models via /models endpoint (neuron format).
     let mock_url = common::spawn_mock_neuron_with_models(json!([
-        {"id": "model-a", "harness": "mistralrs", "status": "loaded", "devices": [0], "vram_used_mb": 8000},
+        {"id": "model-a", "harness": "candle", "status": "loaded", "devices": [0], "vram_used_mb": 8000},
-        {"id": "model-b", "harness": "mistralrs", "status": "unloaded", "devices": [], "vram_used_mb": null}
+        {"id": "model-b", "harness": "candle", "status": "unloaded", "devices": [], "vram_used_mb": null}
     ]))
     .await;
 
@@ -63,8 +63,8 @@ async fn test_poller_discovers_models() {
 #[tokio::test]
 async fn test_poller_updates_gateway_models_endpoint() {
     let mock_url = common::spawn_mock_neuron_with_models(json!([
-        {"id": "model-x", "harness": "mistralrs", "status": "loaded", "devices": [0], "vram_used_mb": null},
+        {"id": "model-x", "harness": "candle", "status": "loaded", "devices": [0], "vram_used_mb": null},
-        {"id": "model-y", "harness": "mistralrs", "status": "loaded", "devices": [1], "vram_used_mb": null}
+        {"id": "model-y", "harness": "candle", "status": "loaded", "devices": [1], "vram_used_mb": null}
     ]))
     .await;
 
@@ -152,8 +152,8 @@ async fn test_poller_marks_unreachable_node_unhealthy() {
 #[tokio::test]
 async fn test_poller_removes_stale_models() {
     let mock_url = common::spawn_mock_neuron_with_models(json!([
-        {"id": "keep-me", "harness": "mistralrs", "status": "loaded", "devices": [0], "vram_used_mb": null},
+        {"id": "keep-me", "harness": "candle", "status": "loaded", "devices": [0], "vram_used_mb": null},
-        {"id": "drop-me", "harness": "mistralrs", "status": "loaded", "devices": [0], "vram_used_mb": null}
+        {"id": "drop-me", "harness": "candle", "status": "loaded", "devices": [0], "vram_used_mb": null}
     ]))
     .await;
 
@@ -183,7 +183,7 @@ async fn test_poller_removes_stale_models() {
 
     // New mock with only one model.
     let new_mock_url = common::spawn_mock_neuron_with_models(json!([
-        {"id": "keep-me", "harness": "mistralrs", "status": "loaded", "devices": [0], "vram_used_mb": null}
+        {"id": "keep-me", "harness": "candle", "status": "loaded", "devices": [0], "vram_used_mb": null}
     ]))
     .await;
 

crates/cortex-gateway/tests/streaming.rs

@@ -51,18 +51,18 @@ async fn test_streaming_sse_passthrough() {
     }
 
     assert!(
-        chunks.len() >= chunk_count + 1,
+        chunks.len() > chunk_count,
-        "expected at least {} chunks (got {}): {:?}",
+        "expected more than {} chunks (got {}): {:?}",
-        chunk_count + 1,
+        chunk_count,
         chunks.len(),
         chunks,
     );
 
     assert_eq!(chunks.last().unwrap(), "[DONE]");
 
-    for i in 0..chunk_count {
+    for (i, chunk) in chunks.iter().enumerate().take(chunk_count) {
         let chunk_json: serde_json::Value =
-            serde_json::from_str(&chunks[i]).expect("chunk should be valid JSON");
+            serde_json::from_str(chunk).expect("chunk should be valid JSON");
         assert_eq!(
             chunk_json["choices"][0]["delta"]["content"],
             format!("token{i}")

crates/neuron/Cargo.toml

@@ -12,6 +12,30 @@ path = "src/lib.rs"
 name = "neuron"
 path = "src/main.rs"
 
+[features]
+default = []
+# Enables CUDA acceleration in candle. Without this feature, candle
+# compiles for CPU only and Device::new_cuda calls fall back to CPU.
+cuda = [
+    "candle-core/cuda",
+    "candle-nn/cuda",
+    "candle-transformers/cuda",
+]
+# Use cuDNN for convolution / attention kernels. Requires CUDA.
+cudnn = [
+    "cuda",
+    "candle-core/cudnn",
+    "candle-nn/cudnn",
+    "candle-transformers/cudnn",
+]
+# FlashAttention kernels. Requires CUDA.
+flash-attn = [
+    "cuda",
+    "candle-transformers/flash-attn",
+]
+# Reserved for GPU-only integration tests in later stages.
+cuda-integration = ["cuda"]
+
 [dependencies]
 cortex-core.workspace = true
 tokio.workspace = true
@@ -24,9 +48,19 @@ tracing-subscriber.workspace = true
 anyhow.workspace = true
 async-trait.workspace = true
 clap.workspace = true
+thiserror.workspace = true
 figment.workspace = true
 toml.workspace = true
 
+# candle for in-process inference. CUDA support is gated behind the
+# crate's `cuda` feature (default off) so the workspace builds on
+# non-CUDA hosts and CI runners.
+candle-core = "0.10.2"
+candle-nn = "0.10.2"
+candle-transformers = "0.10.2"
+tokenizers = { version = "0.22", default-features = false, features = ["onig"] }
+hf-hub = { version = "0.4", features = ["tokio"] }
+
 [dev-dependencies]
 tokio = { workspace = true, features = ["test-util"] }
 reqwest.workspace = true

crates/neuron/src/api.rs

@@ -1,6 +1,7 @@
 //! HTTP API handlers for the neuron daemon.
 
 use crate::harness::HarnessRegistry;
+use crate::harness::candle::{CandleHarness, InferenceError};
 use crate::health::HealthCache;
 use axum::Router;
 use axum::extract::{Path, State};
@@ -9,6 +10,7 @@ use axum::response::{IntoResponse, Json};
 use axum::routing::{get, post};
 use cortex_core::discovery::{DiscoveryResponse, HealthResponse};
 use cortex_core::harness::ModelSpec;
+use cortex_core::openai::ChatCompletionRequest;
 use serde_json::{Value, json};
 use std::sync::Arc;
 use tokio::sync::RwLock;
@@ -18,6 +20,10 @@ pub struct NeuronState {
     pub discovery: DiscoveryResponse,
     pub health_cache: Arc<HealthCache>,
     pub registry: RwLock<HarnessRegistry>,
+    /// Typed handle to the candle harness for inference routes. Cached at
+    /// startup so `/v1/chat/completions` doesn't have to hold the registry
+    /// read lock or perform dyn-Trait dispatch per request.
+    pub candle: Option<Arc<CandleHarness>>,
 }
 
 /// Build the neuron API router.
@@ -29,6 +35,7 @@ pub fn neuron_routes() -> Router<Arc<NeuronState>> {
         .route("/models/load", post(load_model))
         .route("/models/unload", post(unload_model))
         .route("/models/{model_id}/endpoint", get(model_endpoint))
+        .route("/v1/chat/completions", post(chat_completions))
 }
 
 async fn discovery_handler(State(state): State<Arc<NeuronState>>) -> Json<DiscoveryResponse> {
@@ -102,3 +109,40 @@ async fn model_endpoint(
             .into_response(),
     }
 }
+
+/// OpenAI-compatible chat completions. Non-streaming for Stage 3; the
+/// streaming path is added in Stage 4.
+async fn chat_completions(
+    State(state): State<Arc<NeuronState>>,
+    Json(req): Json<ChatCompletionRequest>,
+) -> impl IntoResponse {
+    let Some(candle) = state.candle.as_ref().map(Arc::clone) else {
+        return (
+            StatusCode::SERVICE_UNAVAILABLE,
+            Json(json!({"error": "candle harness not enabled on this neuron"})),
+        )
+            .into_response();
+    };
+
+    if req.stream.unwrap_or(false) {
+        return (
+            StatusCode::NOT_IMPLEMENTED,
+            Json(json!({"error": "streaming responses arrive in Stage 4"})),
+        )
+            .into_response();
+    }
+
+    match candle.chat_completion(req).await {
+        Ok(resp) => Json(resp).into_response(),
+        Err(InferenceError::ModelNotLoaded(id)) => (
+            StatusCode::NOT_FOUND,
+            Json(json!({"error": format!("model '{id}' not loaded on this neuron")})),
+        )
+            .into_response(),
+        Err(InferenceError::Other(e)) => (
+            StatusCode::INTERNAL_SERVER_ERROR,
+            Json(json!({"error": e.to_string()})),
+        )
+            .into_response(),
+    }
+}

crates/neuron/src/config.rs

@@ -6,7 +6,7 @@ use figment::{
     providers::{Env, Format, Toml},
 };
 use serde::{Deserialize, Serialize};
-use std::path::Path;
+use std::path::{Path, PathBuf};
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct NeuronConfig {
@@ -14,10 +14,29 @@ pub struct NeuronConfig {
     pub port: u16,
     #[serde(default)]
     pub harnesses: Vec<HarnessConfig>,
+    /// Per-harness configuration. Currently only `candle` is recognised.
+    #[serde(default)]
+    pub harness: HarnessSettings,
+}
+
+/// Settings for individual harness implementations. Each harness owns
+/// its own sub-table so users only configure the harnesses they enable.
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
+pub struct HarnessSettings {
+    #[serde(default)]
+    pub candle: CandleHarnessConfig,
+}
+
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
+pub struct CandleHarnessConfig {
+    /// HuggingFace cache directory for model weights.
+    /// When unset, defers to hf-hub's default (~/.cache/huggingface).
+    #[serde(default)]
+    pub hf_cache: Option<PathBuf>,
 }
 
 fn default_port() -> u16 {
-    9090
+    13131
 }
 
 impl NeuronConfig {
@@ -33,8 +52,9 @@ impl NeuronConfig {
 impl Default for NeuronConfig {
     fn default() -> Self {
         Self {
-            port: 9090,
+            port: 13131,
             harnesses: vec![],
+            harness: HarnessSettings::default(),
         }
     }
 }

crates/neuron/src/harness/candle.rs

@@ -0,0 +1,441 @@
+//! Candle harness — in-process inference using huggingface/candle.
+//!
+//! This is the sole `Harness` implementation. Inference runs inside
+//! the neuron process; there is no external subprocess.
+//!
+//! - Stage 2 wired GGUF (Qwen3 only) load/unload via `quantized_qwen3`.
+//! - Stage 3 (this) adds `chat_completion` — a non-streaming OpenAI
+//!   compatible chat completion routed to the loaded model's forward
+//!   pass on a per-model serialised generation loop.
+
+use anyhow::{Context, Result};
+use async_trait::async_trait;
+use candle_core::quantized::gguf_file;
+use candle_core::{Device, Tensor};
+use candle_transformers::generation::{LogitsProcessor, Sampling};
+use candle_transformers::models::quantized_qwen3::ModelWeights as QuantizedQwen3Weights;
+use cortex_core::harness::{Harness, HarnessHealth, ModelInfo, ModelSpec};
+use cortex_core::openai::{
+    ChatCompletionChoice, ChatCompletionRequest, ChatCompletionResponse, ChatMessage,
+    MessageContent, Usage,
+};
+use std::collections::HashMap;
+use std::path::PathBuf;
+use std::sync::Arc;
+use std::time::{SystemTime, UNIX_EPOCH};
+use tokenizers::Tokenizer;
+use tokio::sync::{Mutex, RwLock};
+
+/// In-process candle harness. Owns the loaded model registry.
+pub struct CandleHarness {
+    models: Arc<RwLock<HashMap<String, Arc<LoadedModel>>>>,
+    hf_cache: Option<PathBuf>,
+    bind_url: String,
+}
+
+/// A loaded model with its tokenizer, device placement, and architecture-
+/// specific weights. The `arch` is `Arc<Mutex<>>` so the lock guard can be
+/// moved into `spawn_blocking` for synchronous candle forward passes.
+pub struct LoadedModel {
+    pub model_id: String,
+    pub arch: Arc<Mutex<ModelArch>>,
+    pub tokenizer: Tokenizer,
+    pub device: Device,
+    pub quant: Option<String>,
+    pub devices: Vec<u32>,
+}
+
+/// Architecture-specific weights. Stage 3 still supports only Qwen3
+/// quantized; Stage 8 broadens this to additional families and
+/// non-quantized variants.
+pub enum ModelArch {
+    Qwen3Quantized(QuantizedQwen3Weights),
+}
+
+impl CandleHarness {
+    pub fn new(bind_url: String, hf_cache: Option<PathBuf>) -> Self {
+        Self {
+            models: Arc::new(RwLock::new(HashMap::new())),
+            hf_cache,
+            bind_url,
+        }
+    }
+
+    /// Pick a candle `Device` for the requested indices. Without the
+    /// `cuda` feature, or if CUDA initialisation fails, falls back to CPU.
+    fn pick_device(devices: &[u32]) -> Result<Device> {
+        let _idx = devices.first().copied().unwrap_or(0) as usize;
+        #[cfg(feature = "cuda")]
+        {
+            match Device::new_cuda(_idx) {
+                Ok(d) => return Ok(d),
+                Err(e) => tracing::warn!(
+                    device = _idx,
+                    error = %e,
+                    "CUDA device unavailable, falling back to CPU"
+                ),
+            }
+        }
+        Ok(Device::Cpu)
+    }
+
+    /// Resolve a model spec to local GGUF and tokenizer file paths via
+    /// hf-hub. Downloads on first use; subsequent calls are cached.
+    async fn resolve_files(&self, spec: &ModelSpec) -> Result<(PathBuf, PathBuf)> {
+        let mut builder = hf_hub::api::tokio::ApiBuilder::new();
+        if let Some(cache) = &self.hf_cache {
+            builder = builder.with_cache_dir(cache.clone());
+        }
+        let api = builder.build().context("build hf-hub API")?;
+        let repo = api.model(spec.model_id.clone());
+
+        let info = repo
+            .info()
+            .await
+            .with_context(|| format!("fetch HF repo info for {}", spec.model_id))?;
+
+        let quant = spec.quant.as_deref().unwrap_or("");
+        let quant_lc = quant.to_lowercase();
+        let gguf_filename = info
+            .siblings
+            .iter()
+            .map(|s| s.rfilename.as_str())
+            .filter(|name| name.to_lowercase().ends_with(".gguf"))
+            .find(|name| quant_lc.is_empty() || name.to_lowercase().contains(&quant_lc))
+            .ok_or_else(|| {
+                anyhow::anyhow!(
+                    "no GGUF file matching quant {:?} in repo {}",
+                    spec.quant,
+                    spec.model_id
+                )
+            })?
+            .to_string();
+
+        tracing::info!(
+            model = %spec.model_id,
+            file = %gguf_filename,
+            "resolving GGUF (may be cached)"
+        );
+        let gguf_path = repo
+            .get(&gguf_filename)
+            .await
+            .with_context(|| format!("fetch GGUF {gguf_filename}"))?;
+        let tokenizer_path = repo
+            .get("tokenizer.json")
+            .await
+            .context("fetch tokenizer.json")?;
+        Ok((gguf_path, tokenizer_path))
+    }
+
+    /// Run a non-streaming chat completion against a loaded model.
+    ///
+    /// Returns a typed `InferenceError` when the model isn't loaded so the
+    /// handler can map to an appropriate HTTP status without string-matching.
+    pub async fn chat_completion(
+        &self,
+        request: ChatCompletionRequest,
+    ) -> Result<ChatCompletionResponse, InferenceError> {
+        let loaded = {
+            let models = self.models.read().await;
+            models.get(&request.model).cloned()
+        };
+        let loaded = loaded.ok_or_else(|| InferenceError::ModelNotLoaded(request.model.clone()))?;
+
+        let prompt = format_qwen3_prompt(&request.messages);
+
+        let encoding = loaded
+            .tokenizer
+            .encode(prompt.as_str(), true)
+            .map_err(|e| InferenceError::Other(anyhow::anyhow!("tokenize: {e}")))?;
+        let prompt_tokens: Vec<u32> = encoding.get_ids().to_vec();
+        let prompt_len = prompt_tokens.len();
+
+        let temperature = request.temperature.unwrap_or(0.7);
+        let top_p = request.top_p;
+        let max_new = request.max_tokens.unwrap_or(512) as usize;
+        let seed = unix_subsec_nanos();
+
+        let eos_id = loaded
+            .tokenizer
+            .token_to_id("<|im_end|>")
+            .or_else(|| loaded.tokenizer.token_to_id("<|endoftext|>"));
+
+        let arch_arc = Arc::clone(&loaded.arch);
+        let device = loaded.device.clone();
+        let model_id = request.model.clone();
+
+        let (generated_ids, finish_reason) =
+            tokio::task::spawn_blocking(move || -> Result<(Vec<u32>, String)> {
+                let mut guard = arch_arc.blocking_lock();
+                run_inference(
+                    &mut guard,
+                    &device,
+                    &prompt_tokens,
+                    max_new,
+                    temperature,
+                    top_p,
+                    seed,
+                    eos_id,
+                )
+            })
+            .await
+            .map_err(|e| InferenceError::Other(anyhow::anyhow!("inference task panicked: {e}")))?
+            .map_err(InferenceError::Other)?;
+
+        let completion_text = loaded
+            .tokenizer
+            .decode(&generated_ids, true)
+            .map_err(|e| InferenceError::Other(anyhow::anyhow!("detokenize: {e}")))?;
+
+        let usage = Usage {
+            prompt_tokens: prompt_len as u64,
+            completion_tokens: generated_ids.len() as u64,
+            total_tokens: (prompt_len + generated_ids.len()) as u64,
+        };
+
+        Ok(ChatCompletionResponse {
+            id: format!("chatcmpl-{:x}", unix_subsec_nanos()),
+            object: "chat.completion".into(),
+            created: unix_now_secs(),
+            model: model_id,
+            choices: vec![ChatCompletionChoice {
+                index: 0,
+                message: ChatMessage {
+                    role: "assistant".into(),
+                    content: MessageContent::Text(completion_text),
+                    extra: serde_json::Value::Object(Default::default()),
+                },
+                finish_reason: Some(finish_reason),
+                extra: serde_json::Value::Object(Default::default()),
+            }],
+            usage: Some(usage),
+            extra: serde_json::Value::Object(Default::default()),
+        })
+    }
+}
+
+#[async_trait]
+impl Harness for CandleHarness {
+    fn name(&self) -> &str {
+        "candle"
+    }
+
+    async fn health(&self) -> HarnessHealth {
+        HarnessHealth {
+            name: "candle".into(),
+            running: true,
+            uptime_secs: None,
+        }
+    }
+
+    async fn list_models(&self) -> Result<Vec<ModelInfo>> {
+        let models = self.models.read().await;
+        Ok(models
+            .values()
+            .map(|m| ModelInfo {
+                id: m.model_id.clone(),
+                harness: "candle".into(),
+                status: "loaded".into(),
+                devices: m.devices.clone(),
+                vram_used_mb: None,
+            })
+            .collect())
+    }
+
+    async fn load_model(&self, spec: &ModelSpec) -> Result<()> {
+        if spec.harness != "candle" {
+            anyhow::bail!("expected harness=candle, got harness={}", spec.harness);
+        }
+
+        {
+            let models = self.models.read().await;
+            if models.contains_key(&spec.model_id) {
+                anyhow::bail!("model '{}' already loaded", spec.model_id);
+            }
+        }
+
+        let devices = spec.devices.clone().unwrap_or_else(|| vec![0]);
+        let device = Self::pick_device(&devices)?;
+
+        let (gguf_path, tokenizer_path) = self.resolve_files(spec).await?;
+
+        let tokenizer = Tokenizer::from_file(&tokenizer_path)
+            .map_err(|e| anyhow::anyhow!("load tokenizer: {e}"))?;
+
+        // File I/O + GGUF parsing + tensor materialisation are CPU-bound,
+        // so run them on a blocking task to avoid stalling the runtime.
+        let device_for_load = device.clone();
+        let gguf_path_for_load = gguf_path.clone();
+        let model_id_for_log = spec.model_id.clone();
+        let arch = tokio::task::spawn_blocking(move || -> Result<ModelArch> {
+            tracing::info!(model = %model_id_for_log, path = ?gguf_path_for_load, "loading GGUF");
+            let mut file = std::fs::File::open(&gguf_path_for_load).context("open GGUF file")?;
+            let content = gguf_file::Content::read(&mut file)
+                .map_err(|e| anyhow::anyhow!("parse GGUF: {e}"))?;
+
+            let architecture = content
+                .metadata
+                .get("general.architecture")
+                .and_then(|v| v.to_string().ok().cloned())
+                .unwrap_or_default();
+            tracing::info!(architecture = %architecture, "GGUF architecture");
+
+            match architecture.as_str() {
+                "qwen3" => {
+                    let weights =
+                        QuantizedQwen3Weights::from_gguf(content, &mut file, &device_for_load)
+                            .map_err(|e| anyhow::anyhow!("from_gguf qwen3: {e}"))?;
+                    Ok(ModelArch::Qwen3Quantized(weights))
+                }
+                other => anyhow::bail!(
+                    "unsupported GGUF architecture '{other}'; Stage 3 only supports qwen3"
+                ),
+            }
+        })
+        .await
+        .context("blocking load task panicked")??;
+
+        let loaded = Arc::new(LoadedModel {
+            model_id: spec.model_id.clone(),
+            arch: Arc::new(Mutex::new(arch)),
+            tokenizer,
+            device,
+            quant: spec.quant.clone(),
+            devices,
+        });
+
+        let mut models = self.models.write().await;
+        models.insert(spec.model_id.clone(), loaded);
+        tracing::info!(model = %spec.model_id, "model loaded");
+        Ok(())
+    }
+
+    async fn unload_model(&self, model_id: &str) -> Result<()> {
+        let mut models = self.models.write().await;
+        if models.remove(model_id).is_none() {
+            anyhow::bail!("model '{model_id}' not loaded");
+        }
+        tracing::info!(model = %model_id, "model unloaded");
+        Ok(())
+    }
+
+    async fn inference_endpoint(&self, model_id: &str) -> Option<String> {
+        let models = self.models.read().await;
+        models.contains_key(model_id).then(|| self.bind_url.clone())
+    }
+}
+
+/// Errors returned by `CandleHarness::chat_completion`. The
+/// `ModelNotLoaded` variant lets the HTTP handler map cleanly to 404
+/// without string-matching on anyhow messages.
+#[derive(Debug, thiserror::Error)]
+pub enum InferenceError {
+    #[error("model '{0}' not loaded on this neuron")]
+    ModelNotLoaded(String),
+    #[error(transparent)]
+    Other(#[from] anyhow::Error),
+}
+
+/// Apply the Qwen3 chat template:
+///
+/// ```text
+/// <|im_start|>{role}\n{content}<|im_end|>\n
+/// ...
+/// <|im_start|>assistant\n
+/// ```
+///
+/// The trailing `<|im_start|>assistant\n` cues the model to begin a turn.
+/// Non-text content parts (vision blocks) are joined as text only; full
+/// multimodal handling is out of scope for Stage 3.
+fn format_qwen3_prompt(messages: &[ChatMessage]) -> String {
+    let mut prompt = String::new();
+    for msg in messages {
+        let content = match &msg.content {
+            MessageContent::Text(s) => s.clone(),
+            MessageContent::Parts(parts) => parts
+                .iter()
+                .filter_map(|p| p.get("text").and_then(|v| v.as_str()))
+                .collect::<Vec<_>>()
+                .join(""),
+        };
+        prompt.push_str("<|im_start|>");
+        prompt.push_str(&msg.role);
+        prompt.push('\n');
+        prompt.push_str(&content);
+        prompt.push_str("<|im_end|>\n");
+    }
+    prompt.push_str("<|im_start|>assistant\n");
+    prompt
+}
+
+#[allow(clippy::too_many_arguments)]
+fn run_inference(
+    arch: &mut ModelArch,
+    device: &Device,
+    prompt_tokens: &[u32],
+    max_new: usize,
+    temperature: f64,
+    top_p: Option<f64>,
+    seed: u64,
+    eos_id: Option<u32>,
+) -> Result<(Vec<u32>, String)> {
+    let mut logits_processor = {
+        let sampling = if temperature <= 0.0 {
+            Sampling::ArgMax
+        } else {
+            match top_p {
+                Some(p) => Sampling::TopP { p, temperature },
+                None => Sampling::All { temperature },
+            }
+        };
+        LogitsProcessor::from_sampling(seed, sampling)
+    };
+
+    let mut generated: Vec<u32> = Vec::new();
+
+    let mut next_token = match arch {
+        ModelArch::Qwen3Quantized(model) => {
+            model.clear_kv_cache();
+            let input = Tensor::new(prompt_tokens, device)?.unsqueeze(0)?;
+            let logits = model.forward(&input, 0)?;
+            let logits = logits.squeeze(0)?;
+            logits_processor.sample(&logits)?
+        }
+    };
+
+    if Some(next_token) == eos_id {
+        return Ok((generated, "stop".into()));
+    }
+    generated.push(next_token);
+
+    for index in 0..max_new.saturating_sub(1) {
+        next_token = match arch {
+            ModelArch::Qwen3Quantized(model) => {
+                let input = Tensor::new(&[next_token], device)?.unsqueeze(0)?;
+                let logits = model.forward(&input, prompt_tokens.len() + index)?;
+                let logits = logits.squeeze(0)?;
+                logits_processor.sample(&logits)?
+            }
+        };
+        if Some(next_token) == eos_id {
+            return Ok((generated, "stop".into()));
+        }
+        generated.push(next_token);
+    }
+
+    Ok((generated, "length".into()))
+}
+
+fn unix_now_secs() -> u64 {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|d| d.as_secs())
+        .unwrap_or(0)
+}
+
+fn unix_subsec_nanos() -> u64 {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|d| d.as_nanos() as u64)
+        .unwrap_or(0)
+}

crates/neuron/src/harness/llamacpp.rs

@@ -1 +0,0 @@
-// llama.cpp harness implementation — Phase 11.

crates/neuron/src/harness/mistralrs.rs

@@ -1,163 +0,0 @@
-//! mistral.rs harness implementation.
-//!
-//! Wraps the mistral.rs HTTP API for model lifecycle management
-//! and optionally manages the process via systemd.
-
-use anyhow::Result;
-use async_trait::async_trait;
-use cortex_core::harness::{Harness, HarnessConfig, HarnessHealth, ModelInfo, ModelSpec};
-use reqwest::Client;
-use serde::Deserialize;
-
-pub struct MistralRsHarness {
-    endpoint: String,
-    systemd_unit: Option<String>,
-    client: Client,
-}
-
-impl MistralRsHarness {
-    pub fn new(endpoint: String, systemd_unit: Option<String>) -> Self {
-        Self {
-            endpoint,
-            systemd_unit,
-            client: Client::builder()
-                .timeout(std::time::Duration::from_secs(30))
-                .build()
-                .expect("failed to build HTTP client"),
-        }
-    }
-}
-
-/// Response from mistral.rs `GET /v1/models`.
-#[derive(Debug, Deserialize)]
-struct ModelsResponse {
-    data: Vec<ModelEntry>,
-}
-
-#[derive(Debug, Deserialize)]
-struct ModelEntry {
-    id: String,
-    #[serde(default)]
-    status: Option<String>,
-}
-
-#[async_trait]
-impl Harness for MistralRsHarness {
-    fn name(&self) -> &str {
-        "mistralrs"
-    }
-
-    async fn start(&self, _config: &HarnessConfig) -> Result<()> {
-        let Some(unit) = &self.systemd_unit else {
-            anyhow::bail!("no systemd unit configured for mistralrs harness");
-        };
-
-        let output = tokio::process::Command::new("systemctl")
-            .args(["start", unit])
-            .output()
-            .await?;
-
-        if !output.status.success() {
-            let stderr = String::from_utf8_lossy(&output.stderr);
-            anyhow::bail!("systemctl start {unit} failed: {stderr}");
-        }
-
-        // Wait for the health endpoint to respond (up to 30s).
-        let url = format!("{}/health", self.endpoint);
-        for _ in 0..30 {
-            tokio::time::sleep(std::time::Duration::from_secs(1)).await;
-            if self.client.get(&url).send().await.is_ok() {
-                tracing::info!(unit, "mistralrs started and healthy");
-                return Ok(());
-            }
-        }
-        anyhow::bail!("mistralrs started but health endpoint did not respond within 30s");
-    }
-
-    async fn stop(&self) -> Result<()> {
-        let Some(unit) = &self.systemd_unit else {
-            anyhow::bail!("no systemd unit configured for mistralrs harness");
-        };
-
-        let output = tokio::process::Command::new("systemctl")
-            .args(["stop", unit])
-            .output()
-            .await?;
-
-        if !output.status.success() {
-            let stderr = String::from_utf8_lossy(&output.stderr);
-            anyhow::bail!("systemctl stop {unit} failed: {stderr}");
-        }
-        Ok(())
-    }
-
-    async fn health(&self) -> HarnessHealth {
-        let url = format!("{}/health", self.endpoint);
-        let running = self.client.get(&url).send().await.is_ok();
-        HarnessHealth {
-            name: "mistralrs".into(),
-            running,
-            uptime_secs: None,
-        }
-    }
-
-    async fn list_models(&self) -> Result<Vec<ModelInfo>> {
-        let url = format!("{}/v1/models", self.endpoint);
-        let resp = self.client.get(&url).send().await?;
-
-        if !resp.status().is_success() {
-            anyhow::bail!("GET /v1/models returned {}", resp.status());
-        }
-
-        let models_resp: ModelsResponse = resp.json().await?;
-        Ok(models_resp
-            .data
-            .into_iter()
-            .map(|m| ModelInfo {
-                id: m.id,
-                harness: "mistralrs".into(),
-                status: m.status.unwrap_or_else(|| "loaded".into()),
-                devices: vec![],
-                vram_used_mb: None,
-            })
-            .collect())
-    }
-
-    async fn load_model(&self, spec: &ModelSpec) -> Result<()> {
-        let url = format!("{}/v1/models/reload", self.endpoint);
-        let resp = self
-            .client
-            .post(&url)
-            .json(&serde_json::json!({ "model_id": spec.model_id }))
-            .send()
-            .await?;
-
-        if !resp.status().is_success() {
-            let body = resp.text().await.unwrap_or_default();
-            anyhow::bail!("POST /v1/models/reload failed: {body}");
-        }
-        Ok(())
-    }
-
-    async fn unload_model(&self, model_id: &str) -> Result<()> {
-        let url = format!("{}/v1/models/unload", self.endpoint);
-        let resp = self
-            .client
-            .post(&url)
-            .json(&serde_json::json!({ "model_id": model_id }))
-            .send()
-            .await?;
-
-        if !resp.status().is_success() {
-            let body = resp.text().await.unwrap_or_default();
-            anyhow::bail!("POST /v1/models/unload failed: {body}");
-        }
-        Ok(())
-    }
-
-    async fn inference_endpoint(&self, _model_id: &str) -> Option<String> {
-        // mistral.rs routes internally by model name in the request body,
-        // so the inference endpoint is always the base URL.
-        Some(self.endpoint.clone())
-    }
-}

crates/neuron/src/harness/mod.rs

@@ -1,15 +1,22 @@
 //! Harness registry — maps harness names to trait implementations.
 
-pub mod llamacpp;
+pub mod candle;
-pub mod mistralrs;
 
 use anyhow::Result;
 use cortex_core::harness::{Harness, HarnessConfig, ModelInfo, ModelSpec};
 use std::collections::HashMap;
+use std::sync::Arc;
 
 /// Registry of available harness implementations.
+///
+/// Holds an `Arc<dyn Harness>` per harness for generic lifecycle dispatch
+/// (load/unload/list_models). When a candle harness is registered, a typed
+/// `Arc<CandleHarness>` is also cached so inference routes can bypass the
+/// dyn-Trait dispatch and reach harness-specific methods (chat completion,
+/// streaming, etc.).
 pub struct HarnessRegistry {
-    harnesses: HashMap<String, Box<dyn Harness>>,
+    harnesses: HashMap<String, Arc<dyn Harness>>,
+    candle: Option<Arc<candle::CandleHarness>>,
 }
 
 impl Default for HarnessRegistry {
@@ -22,10 +29,11 @@ impl HarnessRegistry {
     pub fn new() -> Self {
         Self {
             harnesses: HashMap::new(),
+            candle: None,
         }
     }
 
-    pub fn register(&mut self, harness: Box<dyn Harness>) {
+    pub fn register(&mut self, harness: Arc<dyn Harness>) {
         self.harnesses.insert(harness.name().to_string(), harness);
     }
 
@@ -34,6 +42,12 @@ impl HarnessRegistry {
         self.harnesses.keys().cloned().collect()
     }
 
+    /// Typed handle to the candle harness, if registered. Used by inference
+    /// routes that need methods beyond the `Harness` trait surface.
+    pub fn candle(&self) -> Option<Arc<candle::CandleHarness>> {
+        self.candle.clone()
+    }
+
     /// List models from all registered harnesses.
     pub async fn list_all_models(&self) -> Result<Vec<ModelInfo>> {
         let mut all = Vec::new();
@@ -81,19 +95,25 @@ impl HarnessRegistry {
     }
 
     /// Build a registry from harness configs.
-    pub fn from_configs(configs: &[HarnessConfig]) -> Self {
+    ///
+    /// `bind_url` is the URL where this neuron serves inference (its own
+    /// listen address). In-process harnesses (currently the only kind)
+    /// return this URL from `inference_endpoint`.
+    pub fn from_configs(
+        configs: &[HarnessConfig],
+        bind_url: &str,
+        settings: &crate::config::HarnessSettings,
+    ) -> Self {
         let mut registry = Self::new();
         for config in configs {
             match config.name.as_str() {
-                "mistralrs" => {
+                "candle" => {
-                    if let Some(endpoint) = &config.endpoint {
+                    let harness = Arc::new(candle::CandleHarness::new(
-                        registry.register(Box::new(mistralrs::MistralRsHarness::new(
+                        bind_url.to_string(),
-                            endpoint.clone(),
+                        settings.candle.hf_cache.clone(),
-                            config.systemd_unit.clone(),
+                    ));
-                        )));
+                    registry.candle = Some(Arc::clone(&harness));
-                    } else {
+                    registry.harnesses.insert("candle".into(), harness);
-                        tracing::warn!("mistralrs harness missing endpoint, skipping");
-                    }
                 }
                 other => {
                     tracing::warn!(harness = other, "unknown harness type, skipping");

crates/neuron/src/main.rs

@@ -37,6 +37,7 @@ async fn main() -> Result<()> {
     });
 
     let port = args.port.unwrap_or(cfg.port);
+    let bind_url = format!("http://localhost:{port}");
     let start_time = Instant::now();
 
     tracing::info!("running hardware discovery");
@@ -47,9 +48,12 @@ async fn main() -> Result<()> {
         "discovery complete"
     );
 
-    // Build harness registry from config.
+    // Build harness registry from config. In-process harnesses (candle)
-    let registry = HarnessRegistry::from_configs(&cfg.harnesses);
+    // need to know neuron's own bind URL so they can return it from
+    // inference_endpoint.
+    let registry = HarnessRegistry::from_configs(&cfg.harnesses, &bind_url, &cfg.harness);
     discovery_result.harnesses = registry.names();
+    let candle = registry.candle();
 
     let health_cache = Arc::new(health::HealthCache::new());
     health_cache
@@ -65,6 +69,7 @@ async fn main() -> Result<()> {
         discovery: discovery_result,
         health_cache,
         registry: RwLock::new(registry),
+        candle,
     });
 
     let app = api::neuron_routes().with_state(state);

crates/neuron/tests/api.rs

@@ -14,6 +14,7 @@ async fn spawn_neuron(discovery: DiscoveryResponse) -> String {
         discovery,
         health_cache,
         registry: RwLock::new(registry),
+        candle: None,
     });
 
     let app = api::neuron_routes().with_state(state);
@@ -135,56 +136,30 @@ async fn test_models_empty_registry() {
     assert!(body.as_array().unwrap().is_empty());
 }
 
-/// Spawn a mock mistral.rs backend and a neuron with the mistralrs harness
+/// Verify the candle harness registers, list is empty by default, and a
-/// pointing at it, then test the full model lifecycle through neuron's API.
+/// load attempt for an obviously-bogus model id returns a 4xx error
+/// without crashing the daemon. Real load/unload exercising actual GGUF
+/// download is covered by `tests/candle_lifecycle.rs` (cuda-integration).
 #[tokio::test]
-async fn test_models_via_mistralrs_harness() {
+async fn test_candle_harness_registers_and_rejects_bogus_model() {
-    use axum::routing::{get, post};
-    use axum::{Json, Router};
     use cortex_core::harness::HarnessConfig;
-    use serde_json::Value;
+    use neuron::config::HarnessSettings;
 
-    // Mock mistral.rs backend.
+    let registry = HarnessRegistry::from_configs(
-    let mock_app = Router::new()
+        &[HarnessConfig {
-        .route(
+            name: "candle".into(),
-            "/v1/models",
+        }],
-            get(|| async {
+        "http://localhost:13131",
-                Json(json!({
+        &HarnessSettings::default(),
-                    "data": [
+    );
-                        {"id": "test-model", "status": "loaded"},
-                        {"id": "other-model", "status": "unloaded"}
-                    ]
-                }))
-            }),
-        )
-        .route(
-            "/v1/models/unload",
-            post(|Json(_body): Json<Value>| async { Json(json!({"status": "ok"})) }),
-        )
-        .route(
-            "/v1/models/reload",
-            post(|Json(_body): Json<Value>| async { Json(json!({"status": "ok"})) }),
-        );
-
-    let mock_listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
-    let mock_addr = mock_listener.local_addr().unwrap();
-    tokio::spawn(async move {
-        axum::serve(mock_listener, mock_app).await.unwrap();
-    });
-    let mock_url = format!("http://{mock_addr}");
-
-    // Build neuron with mistralrs harness pointing at mock.
-    let registry = HarnessRegistry::from_configs(&[HarnessConfig {
-        name: "mistralrs".into(),
-        endpoint: Some(mock_url.clone()),
-        systemd_unit: None,
-    }]);
 
+    let candle = registry.candle();
     let health_cache = Arc::new(HealthCache::new());
     let state = Arc::new(NeuronState {
         discovery: fake_discovery(),
         health_cache,
         registry: RwLock::new(registry),
+        candle,
     });
 
     let app = api::neuron_routes().with_state(state);
@@ -197,7 +172,6 @@ async fn test_models_via_mistralrs_harness() {
 
     let client = reqwest::Client::new();
 
-    // GET /models — should return models from mock mistralrs.
     let resp = client
         .get(format!("{neuron_url}/models"))
         .send()
@@ -205,45 +179,139 @@ async fn test_models_via_mistralrs_harness() {
         .unwrap();
     assert_eq!(resp.status(), 200);
     let models: Vec<serde_json::Value> = resp.json().await.unwrap();
-    assert_eq!(models.len(), 2);
+    assert!(models.is_empty());
-    assert_eq!(models[0]["id"], "test-model");
-    assert_eq!(models[0]["harness"], "mistralrs");
-    assert_eq!(models[0]["status"], "loaded");
-    assert_eq!(models[1]["id"], "other-model");
-    assert_eq!(models[1]["status"], "unloaded");
 
-    // GET /models/test-model/endpoint — should return mock URL.
+    // Sending a wrong-harness spec should be rejected synchronously
-    let resp = client
+    // without touching the network or the model registry.
-        .get(format!("{neuron_url}/models/test-model/endpoint"))
-        .send()
-        .await
-        .unwrap();
-    assert_eq!(resp.status(), 200);
-    let body: serde_json::Value = resp.json().await.unwrap();
-    assert_eq!(body["url"], mock_url);
-
-    // POST /models/unload — should succeed.
-    let resp = client
-        .post(format!("{neuron_url}/models/unload"))
-        .json(&json!({"model_id": "test-model"}))
-        .send()
-        .await
-        .unwrap();
-    assert_eq!(resp.status(), 200);
-    let body: serde_json::Value = resp.json().await.unwrap();
-    assert_eq!(body["status"], "unloaded");
-
-    // POST /models/load — should succeed.
     let resp = client
         .post(format!("{neuron_url}/models/load"))
+        .json(&json!({"model_id": "definitely/not-real", "harness": "not-candle"}))
+        .send()
+        .await
+        .unwrap();
+    assert_eq!(resp.status(), 400);
+
+    // Registry still empty.
+    let resp = client
+        .get(format!("{neuron_url}/models"))
+        .send()
+        .await
+        .unwrap();
+    let models: Vec<serde_json::Value> = resp.json().await.unwrap();
+    assert!(models.is_empty());
+}
+
+/// `/v1/chat/completions` returns 503 when no candle harness is registered.
+#[tokio::test]
+async fn test_chat_completions_no_candle_harness() {
+    let registry = HarnessRegistry::new();
+    let health_cache = Arc::new(HealthCache::new());
+    let state = Arc::new(NeuronState {
+        discovery: fake_discovery(),
+        health_cache,
+        registry: RwLock::new(registry),
+        candle: None,
+    });
+    let app = api::neuron_routes().with_state(state);
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    let url = format!("http://{addr}");
+
+    let resp = reqwest::Client::new()
+        .post(format!("{url}/v1/chat/completions"))
         .json(&json!({
-            "model_id": "test-model",
+            "model": "anything",
-            "harness": "mistralrs"
+            "messages": [{"role": "user", "content": "hi"}]
         }))
         .send()
         .await
         .unwrap();
-    assert_eq!(resp.status(), 200);
+    assert_eq!(resp.status(), 503);
-    let body: serde_json::Value = resp.json().await.unwrap();
+}
-    assert_eq!(body["status"], "loaded");
+
+/// `/v1/chat/completions` returns 404 when the requested model isn't loaded.
+#[tokio::test]
+async fn test_chat_completions_model_not_loaded() {
+    use cortex_core::harness::HarnessConfig;
+    use neuron::config::HarnessSettings;
+
+    let registry = HarnessRegistry::from_configs(
+        &[HarnessConfig {
+            name: "candle".into(),
+        }],
+        "http://localhost:0",
+        &HarnessSettings::default(),
+    );
+    let candle = registry.candle();
+    let health_cache = Arc::new(HealthCache::new());
+    let state = Arc::new(NeuronState {
+        discovery: fake_discovery(),
+        health_cache,
+        registry: RwLock::new(registry),
+        candle,
+    });
+    let app = api::neuron_routes().with_state(state);
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    let url = format!("http://{addr}");
+
+    let resp = reqwest::Client::new()
+        .post(format!("{url}/v1/chat/completions"))
+        .json(&json!({
+            "model": "definitely/not-loaded",
+            "messages": [{"role": "user", "content": "hi"}]
+        }))
+        .send()
+        .await
+        .unwrap();
+    assert_eq!(resp.status(), 404);
+}
+
+/// `/v1/chat/completions` with `stream: true` returns 501 until Stage 4
+/// wires up SSE.
+#[tokio::test]
+async fn test_chat_completions_streaming_not_yet_implemented() {
+    use cortex_core::harness::HarnessConfig;
+    use neuron::config::HarnessSettings;
+
+    let registry = HarnessRegistry::from_configs(
+        &[HarnessConfig {
+            name: "candle".into(),
+        }],
+        "http://localhost:0",
+        &HarnessSettings::default(),
+    );
+    let candle = registry.candle();
+    let health_cache = Arc::new(HealthCache::new());
+    let state = Arc::new(NeuronState {
+        discovery: fake_discovery(),
+        health_cache,
+        registry: RwLock::new(registry),
+        candle,
+    });
+    let app = api::neuron_routes().with_state(state);
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    let url = format!("http://{addr}");
+
+    let resp = reqwest::Client::new()
+        .post(format!("{url}/v1/chat/completions"))
+        .json(&json!({
+            "model": "anything",
+            "messages": [{"role": "user", "content": "hi"}],
+            "stream": true
+        }))
+        .send()
+        .await
+        .unwrap();
+    assert_eq!(resp.status(), 501);
 }

crates/neuron/tests/candle_lifecycle.rs

@@ -0,0 +1,87 @@
+//! Real model load/unload lifecycle through the candle harness.
+//!
+//! Gated behind the `cuda-integration` feature because it downloads a
+//! real (small) GGUF from HuggingFace and materialises tensors on the
+//! configured device. Run on a host with network access and either a
+//! CUDA GPU (when built with `--features cuda`) or enough CPU RAM to
+//! hold the model.
+//!
+//! Usage:
+//!   cargo test -p neuron --features cuda-integration --test candle_lifecycle
+//!
+//! Optional environment variables:
+//!   NEURON_TEST_MODEL_ID — HuggingFace repo to load (default: a small
+//!     public Qwen3 GGUF repo).
+//!   NEURON_TEST_QUANT    — quant substring matched against GGUF
+//!     filenames (default: "Q4_K_M").
+//!   HF_HOME              — HuggingFace cache directory.
+
+#![cfg(feature = "cuda-integration")]
+
+use cortex_core::harness::{HarnessConfig, ModelSpec};
+use neuron::config::HarnessSettings;
+use neuron::harness::HarnessRegistry;
+use std::path::PathBuf;
+
+#[tokio::test]
+async fn test_candle_qwen3_load_unload_lifecycle() {
+    let _ = tracing_subscriber::fmt()
+        .with_test_writer()
+        .with_env_filter("info,neuron=debug")
+        .try_init();
+
+    let model_id = std::env::var("NEURON_TEST_MODEL_ID")
+        .unwrap_or_else(|_| "Qwen/Qwen3-0.6B-GGUF".to_string());
+    let quant = std::env::var("NEURON_TEST_QUANT").unwrap_or_else(|_| "Q4_K_M".to_string());
+
+    let mut settings = HarnessSettings::default();
+    if let Ok(home) = std::env::var("HF_HOME") {
+        settings.candle.hf_cache = Some(PathBuf::from(home));
+    }
+
+    let registry = HarnessRegistry::from_configs(
+        &[HarnessConfig {
+            name: "candle".into(),
+        }],
+        "http://localhost:13131",
+        &settings,
+    );
+
+    let spec = ModelSpec {
+        model_id: model_id.clone(),
+        harness: "candle".into(),
+        quant: Some(quant),
+        tensor_parallel: None,
+        devices: Some(vec![0]),
+    };
+
+    registry
+        .load_model(&spec)
+        .await
+        .expect("load_model should succeed");
+
+    let models = registry.list_all_models().await.expect("list_all_models");
+    assert_eq!(models.len(), 1, "expected exactly one loaded model");
+    assert_eq!(models[0].id, model_id);
+    assert_eq!(models[0].harness, "candle");
+    assert_eq!(models[0].status, "loaded");
+
+    let url = registry.inference_endpoint(&model_id).await;
+    assert_eq!(url, Some("http://localhost:13131".into()));
+
+    // Re-loading the same model should be rejected.
+    let again = registry.load_model(&spec).await;
+    assert!(again.is_err(), "second load should error");
+
+    registry
+        .unload_model(&model_id)
+        .await
+        .expect("unload_model should succeed");
+
+    let models = registry.list_all_models().await.expect("list_all_models");
+    assert!(models.is_empty(), "registry should be empty after unload");
+
+    // Unloading a model that isn't loaded should error.
+    let err = registry.unload_model(&model_id).await;
+    assert!(err.is_err(), "unload of missing model should error");
+}

data/cortex-firewalld.xml

@@ -0,0 +1,7 @@
+<?xml version="1.0" encoding="utf-8"?>
+<service>
+  <short>cortex</short>
+  <description>Cortex — inference gateway for multi-node GPU clusters</description>
+  <port protocol="tcp" port="31313"/>
+  <port protocol="tcp" port="31314"/>
+</service>

data/cortex-sysusers.conf

@@ -0,0 +1,3 @@
+g cortex - -
+u cortex - "Cortex inference cluster" /var/lib/cortex /sbin/nologin
+m cortex cortex

data/neuron-firewalld.xml

@@ -0,0 +1,6 @@
+<?xml version="1.0" encoding="utf-8"?>
+<service>
+  <short>helexa-neuron</short>
+  <description>Neuron — per-node GPU discovery and harness daemon for cortex</description>
+  <port protocol="tcp" port="13131"/>
+</service>

data/neuron-sysusers.conf

@@ -0,0 +1,3 @@
+g neuron - -
+u neuron - "Neuron GPU node daemon" /var/lib/neuron /sbin/nologin
+m neuron neuron

data/neuron.service

@@ -5,11 +5,11 @@ Wants=network-online.target
 
 [Service]
 Type=simple
-ExecStart=/usr/bin/neuron --config /etc/cortex/neuron.toml
+ExecStart=/usr/bin/neuron --config /etc/neuron/neuron.toml
 Restart=on-failure
 RestartSec=5
-User=cortex
+User=neuron
-Group=cortex
+Group=neuron
 
 [Install]
 WantedBy=multi-user.target

helexa-neuron.spec

@@ -0,0 +1,101 @@
+Name:           helexa-neuron
+Version:        0.1.16
+Release:        1%{?dist}
+Summary:        Per-node GPU discovery and harness management daemon for cortex
+# Package name disambiguates from Fedora's existing "neuron" package
+# (NEURON neural simulation environment from Yale). Binary, systemd
+# unit, and system user are still called "neuron" for brevity.
+
+License:        GPL-3.0-or-later
+URL:            https://git.lair.cafe/helexa/cortex
+Source0:        %{name}-%{version}.tar.gz
+Source1:        %{name}-%{version}-vendor.tar.gz
+
+ExclusiveArch:  x86_64
+
+BuildRequires:  rust >= 1.85
+BuildRequires:  cargo
+BuildRequires:  gcc
+BuildRequires:  gcc-c++
+BuildRequires:  cmake
+BuildRequires:  perl-interpreter
+BuildRequires:  pkgconfig(openssl)
+BuildRequires:  systemd-rpm-macros
+
+Requires(pre):  shadow-utils
+Requires:       systemd
+Requires:       firewalld-filesystem
+
+# systemd-rpm-macros ships a unit dep generator that parses User=/Group=
+# from our .service file and emits Requires: user(neuron)/group(neuron).
+# rpm's sysusers provides-generator emits the unversioned form for groups
+# but only a versioned user(neuron) = <base64> for users with GECOS/home/
+# shell. Provide the unversioned user(neuron) explicitly so dnf can resolve
+# the auto-generated Requires. Without this, dnf5 silently filters the
+# package and reports "Nothing to do".
+Provides:       user(neuron)
+
+%description
+Neuron is a per-node daemon for cortex inference clusters. It discovers
+local GPU hardware via nvidia-smi, runs in-process inference via
+huggingface/candle, and exposes an HTTP API for model lifecycle
+management (load, unload, list, inference endpoint).
+
+%prep
+%autosetup
+tar xf %{SOURCE1}
+mkdir -p .cargo
+cat > .cargo/config.toml << 'EOF'
+[source.crates-io]
+replace-with = "vendored-sources"
+
+[source.vendored-sources]
+directory = "vendor"
+EOF
+
+%build
+cargo build --release -p neuron
+
+%install
+install -Dm755 target/release/neuron %{buildroot}%{_bindir}/neuron
+install -Dm644 data/neuron.service %{buildroot}%{_unitdir}/neuron.service
+install -Dm644 data/neuron-sysusers.conf %{buildroot}%{_sysusersdir}/neuron.conf
+install -Dm644 data/neuron-firewalld.xml %{buildroot}%{_prefix}/lib/firewalld/services/helexa-neuron.xml
+install -dm755 %{buildroot}%{_sysconfdir}/neuron
+install -Dm644 neuron.example.toml %{buildroot}%{_sysconfdir}/neuron/neuron.toml
+
+%pre
+%sysusers_create_compat %{_builddir}/%{name}-%{version}/data/neuron-sysusers.conf
+
+%post
+%systemd_post neuron.service
+
+%preun
+%systemd_preun neuron.service
+
+%postun
+%systemd_postun_with_restart neuron.service
+
+%files
+%license LICENSE
+%doc README.md
+%{_bindir}/neuron
+%{_unitdir}/neuron.service
+%{_sysusersdir}/neuron.conf
+%{_prefix}/lib/firewalld/services/helexa-neuron.xml
+%dir %{_sysconfdir}/neuron
+%config(noreplace) %{_sysconfdir}/neuron/neuron.toml
+
+%changelog
+* Thu Apr 16 2026 Gitea Actions <actions@git.lair.cafe> - 0.1.16-1
+- chore: ignore local deploy script
+- chore: move default ports out of common-collision ranges
+- ci: drop actions/cache for cargo registry and target
+
+* Thu Apr 16 2026 Gitea Actions <actions@git.lair.cafe> - 0.1.14-1
+- ci: publish both packages to a single helexa/helexa COPR project
+- fix(rpm): rename neuron package to helexa-neuron
+- ci: commit generated %changelog entries back to main
+
+* Wed Apr 15 2026 Rob Thijssen <grenade@rob.tn> - 0.1.0-1
+- Initial package

models.example.toml

@@ -6,7 +6,7 @@
 
 [[models]]
 id = "your-org/large-model"
-harness = "mistralrs"
+harness = "candle"
 quant = "Q4_K_M"
 vram_mb = 19000
 min_devices = 2
@@ -15,7 +15,7 @@ pinned_on = ["gpu-large"]
 
 [[models]]
 id = "your-org/medium-model"
-harness = "mistralrs"
+harness = "candle"
 quant = "Q6_K"
 vram_mb = 12000
 min_devices = 1
@@ -23,7 +23,7 @@ pinned_on = ["gpu-medium"]
 
 [[models]]
 id = "your-org/embedding-model"
-harness = "mistralrs"
+harness = "candle"
 quant = "Q8_0"
 vram_mb = 8000
 min_devices = 1

neuron.example.toml

@@ -1,16 +1,24 @@
 # neuron.example.toml — example configuration
 #
-# Copy to /etc/cortex/neuron.toml and adjust for your environment.
+# Copy to /etc/neuron/neuron.toml and adjust for your environment.
 #
 # Environment variable overrides use NEURON_ prefix with __ separators:
-#   NEURON_PORT=9090
+#   NEURON_PORT=13131
 
-port = 9090
+port = 13131
 
 # -- Harnesses ---------------------------------------------------------------
-# Each [[harnesses]] entry declares an inference engine managed by neuron.
+# Each [[harnesses]] entry enables an inference engine. Currently only
+# "candle" is supported — it runs in-process and uses huggingface/candle
+# for inference on local CUDA devices (or CPU when CUDA is unavailable).
 
 [[harnesses]]
-name = "mistralrs"
+name = "candle"
-endpoint = "http://localhost:8080"
+
-systemd_unit = "mistralrs.service"
+# -- Candle harness settings -------------------------------------------------
+# Optional tuning for the candle harness.
+
+[harness.candle]
+# HuggingFace cache directory for model weights. When unset, hf-hub's
+# default (~/.cache/huggingface) is used.
+# hf_cache = "/var/lib/neuron/hf-cache"

neuron.spec

@@ -1,69 +0,0 @@
-Name:           neuron
-Version:        0.1.0
-Release:        1%{?dist}
-Summary:        Per-node GPU discovery and harness management daemon for cortex
-
-License:        GPL-3.0-or-later
-URL:            https://git.lair.cafe/helexa/cortex
-Source0:        %{name}-%{version}.tar.gz
-Source1:        %{name}-%{version}-vendor.tar.gz
-
-ExclusiveArch:  x86_64
-
-BuildRequires:  rust >= 1.85
-BuildRequires:  cargo
-BuildRequires:  gcc
-BuildRequires:  systemd-rpm-macros
-
-Requires(pre):  shadow-utils
-
-%description
-Neuron is a per-node daemon for cortex inference clusters. It discovers
-local GPU hardware via nvidia-smi, manages inference harnesses (mistral.rs,
-llama.cpp), and exposes an HTTP API for model lifecycle management.
-
-%prep
-%autosetup
-tar xf %{SOURCE1}
-mkdir -p .cargo
-cat > .cargo/config.toml << 'EOF'
-[source.crates-io]
-replace-with = "vendored-sources"
-
-[source.vendored-sources]
-directory = "vendor"
-EOF
-
-%build
-cargo build --release -p neuron
-
-%install
-install -Dm755 target/release/neuron %{buildroot}%{_bindir}/neuron
-install -Dm644 data/neuron.service %{buildroot}%{_unitdir}/neuron.service
-install -dm750 %{buildroot}%{_sysconfdir}/cortex
-install -Dm640 neuron.example.toml %{buildroot}%{_sysconfdir}/cortex/neuron.toml
-
-%pre
-getent group cortex >/dev/null || groupadd -r cortex
-getent passwd cortex >/dev/null || useradd -r -g cortex -d /var/lib/cortex -s /sbin/nologin cortex
-
-%post
-%systemd_post neuron.service
-
-%preun
-%systemd_preun neuron.service
-
-%postun
-%systemd_postun_with_restart neuron.service
-
-%files
-%license LICENSE
-%doc README.md
-%{_bindir}/neuron
-%{_unitdir}/neuron.service
-%dir %attr(750,root,cortex) %{_sysconfdir}/cortex
-%config(noreplace) %attr(640,root,cortex) %{_sysconfdir}/cortex/neuron.toml
-
-%changelog
-* Tue Apr 15 2026 Rob Thijssen <grenade@rob.tn> - 0.1.0-1
-- Initial package

rpm/cortex-prerelease.spec

@@ -0,0 +1,102 @@
+# Prebuilt-binary spec for cortex.
+#
+# Unlike cortex.spec (which builds from source via cargo), this spec
+# wraps a pre-built `cortex` binary produced by an upstream CI job and
+# packages it for rpm.lair.cafe. The %build phase is a no-op.
+#
+# Required defines at rpmbuild time:
+#   cortex_version    e.g. "0.1.16"
+#   cortex_prerelease e.g. "0.1.20260518gitabcdef0"   (used as Release)
+
+%global _build_id_links none
+%global debug_package %{nil}
+%global __strip /usr/bin/true
+
+%{!?cortex_version: %global cortex_version 0.0.0}
+%if 0%{?cortex_prerelease:1}
+%global cortex_release %{cortex_prerelease}
+%else
+%global cortex_release 1
+%endif
+
+Name:           cortex
+Version:        %{cortex_version}
+Release:        %{cortex_release}%{?dist}
+Summary:        Inference gateway for multi-node GPU clusters (prebuilt)
+
+License:        GPL-3.0-or-later
+URL:            https://git.lair.cafe/helexa/cortex
+
+Source0:        cortex
+Source1:        cortex.service
+Source2:        cortex-sysusers.conf
+Source3:        cortex-firewalld.xml
+Source4:        cortex.example.toml
+Source5:        models.example.toml
+Source6:        LICENSE
+
+ExclusiveArch:  x86_64
+
+Requires(pre):  shadow-utils
+Requires:       systemd
+Requires:       firewalld-filesystem
+
+Provides:       user(cortex)
+
+%description
+Cortex is a Rust reverse-proxy that sits in front of multiple neuron
+inference daemons and presents a unified OpenAI and Anthropic
+compatible API surface.
+
+This package wraps a binary built upstream in CI; the source-build
+spec (cortex.spec) remains available for stable releases.
+
+%prep
+cp %{SOURCE0} ./cortex
+cp %{SOURCE1} .
+cp %{SOURCE2} .
+cp %{SOURCE3} .
+cp %{SOURCE4} .
+cp %{SOURCE5} .
+cp %{SOURCE6} .
+
+%build
+# Already built in the upstream CI build job.
+
+%install
+install -Dm755 cortex %{buildroot}%{_bindir}/cortex
+install -Dm644 cortex.service %{buildroot}%{_unitdir}/cortex.service
+install -Dm644 cortex-sysusers.conf %{buildroot}%{_sysusersdir}/cortex.conf
+install -Dm644 cortex-firewalld.xml %{buildroot}%{_prefix}/lib/firewalld/services/cortex.xml
+install -dm755 %{buildroot}%{_sysconfdir}/cortex
+install -Dm644 cortex.example.toml %{buildroot}%{_sysconfdir}/cortex/cortex.toml
+install -Dm644 models.example.toml %{buildroot}%{_sysconfdir}/cortex/models.toml
+
+%pre
+getent group cortex >/dev/null || groupadd -r cortex
+getent passwd cortex >/dev/null || \
+    useradd -r -g cortex -d /var/lib/cortex -s /sbin/nologin \
+        -c "Cortex inference gateway" cortex
+
+%post
+%systemd_post cortex.service
+
+%preun
+%systemd_preun cortex.service
+
+%postun
+%systemd_postun_with_restart cortex.service
+
+%files
+%license LICENSE
+%{_bindir}/cortex
+%{_unitdir}/cortex.service
+%{_sysusersdir}/cortex.conf
+%{_prefix}/lib/firewalld/services/cortex.xml
+%dir %{_sysconfdir}/cortex
+%config(noreplace) %{_sysconfdir}/cortex/cortex.toml
+%config(noreplace) %{_sysconfdir}/cortex/models.toml
+
+%changelog
+* Mon May 18 2026 Gitea Actions <actions@git.lair.cafe> - %{cortex_version}-%{cortex_release}
+- Prerelease build from upstream CI binary.

rpm/helexa-neuron-prerelease.spec

@@ -0,0 +1,122 @@
+# Prebuilt-binary spec for helexa-neuron flavoured by CUDA compute capability.
+#
+# Unlike helexa-neuron.spec (which builds from source via cargo), this
+# spec wraps a pre-built `neuron-{flavour}` binary produced by an
+# upstream CI job and packages it for rpm.lair.cafe. The %build phase
+# is a no-op.
+#
+# Required defines at rpmbuild time:
+#   neuron_version    e.g. "0.1.16"
+#   neuron_flavour    e.g. "ada", "blackwell"  — matches the CI build
+#                     matrix's compute_cap label.
+#   neuron_prerelease e.g. "0.1.20260518gitabcdef0"   (used as Release)
+#
+# One flavour can be installed at a time on a given host; flavour
+# packages Conflict with each other.
+
+%global _build_id_links none
+%global debug_package %{nil}
+%global __strip /usr/bin/true
+
+%{!?neuron_version: %global neuron_version 0.0.0}
+%{!?neuron_flavour: %global neuron_flavour blackwell}
+%if 0%{?neuron_prerelease:1}
+%global neuron_release %{neuron_prerelease}
+%else
+%global neuron_release 1
+%endif
+
+Name:           helexa-neuron-%{neuron_flavour}
+Version:        %{neuron_version}
+Release:        %{neuron_release}%{?dist}
+Summary:        Per-node GPU inference daemon (candle, %{neuron_flavour} flavour)
+
+License:        GPL-3.0-or-later
+URL:            https://git.lair.cafe/helexa/cortex
+
+Source0:        neuron-%{neuron_flavour}
+Source1:        neuron.service
+Source2:        neuron-sysusers.conf
+Source3:        neuron-firewalld.xml
+Source4:        neuron.example.toml
+Source5:        LICENSE
+
+ExclusiveArch:  x86_64
+
+# Binary links against the CUDA runtime, cuDNN, NCCL, etc. Suppress
+# auto-detected exact soname deps — users may have CUDA from various
+# sources (rpmfusion, nvidia-direct) at different compatible versions;
+# a runtime dlopen failure surfaces a clearer error than rpm dep
+# resolution would.
+%global __requires_exclude ^lib(cuda|cudart|cudnn|cublas|cublasLt|curand|nvrtc|nccl)
+
+Requires(pre):  shadow-utils
+Requires:       systemd
+Requires:       firewalld-filesystem
+
+Provides:       helexa-neuron = %{neuron_version}-%{neuron_release}
+Provides:       user(neuron)
+
+# Mutual exclusion across flavours and the source-build variant.
+Conflicts:      helexa-neuron
+Conflicts:      helexa-neuron-ada
+Conflicts:      helexa-neuron-ampere
+Conflicts:      helexa-neuron-blackwell
+# (The Conflicts: with self is filtered by rpm at install time.)
+
+%description
+Neuron is the per-node daemon for cortex inference clusters. It
+discovers local GPU hardware via nvidia-smi, runs in-process
+inference via huggingface/candle, and exposes an HTTP API for model
+lifecycle management (load, unload, list, inference endpoint).
+
+This is the %{neuron_flavour} flavour, built for that CUDA compute
+capability. Install the flavour matching the GPUs on this host.
+
+%prep
+cp %{SOURCE0} ./neuron
+cp %{SOURCE1} .
+cp %{SOURCE2} .
+cp %{SOURCE3} .
+cp %{SOURCE4} .
+cp %{SOURCE5} .
+
+%build
+# Already built in the upstream CI build job (with --features cuda).
+
+%install
+install -Dm755 neuron %{buildroot}%{_bindir}/neuron
+install -Dm644 neuron.service %{buildroot}%{_unitdir}/neuron.service
+install -Dm644 neuron-sysusers.conf %{buildroot}%{_sysusersdir}/neuron.conf
+install -Dm644 neuron-firewalld.xml %{buildroot}%{_prefix}/lib/firewalld/services/helexa-neuron.xml
+install -dm755 %{buildroot}%{_sysconfdir}/neuron
+install -Dm644 neuron.example.toml %{buildroot}%{_sysconfdir}/neuron/neuron.toml
+
+%pre
+getent group neuron >/dev/null || groupadd -r neuron
+getent passwd neuron >/dev/null || \
+    useradd -r -g neuron -d /var/lib/neuron -s /sbin/nologin \
+        -G video,render \
+        -c "Neuron GPU node daemon" neuron
+
+%post
+%systemd_post neuron.service
+
+%preun
+%systemd_preun neuron.service
+
+%postun
+%systemd_postun_with_restart neuron.service
+
+%files
+%license LICENSE
+%{_bindir}/neuron
+%{_unitdir}/neuron.service
+%{_sysusersdir}/neuron.conf
+%{_prefix}/lib/firewalld/services/helexa-neuron.xml
+%dir %{_sysconfdir}/neuron
+%config(noreplace) %{_sysconfdir}/neuron/neuron.toml
+
+%changelog
+* Mon May 18 2026 Gitea Actions <actions@git.lair.cafe> - %{neuron_version}-%{neuron_release}
+- Prerelease build from upstream CI binary (%{neuron_flavour} flavour).

rpm/rpmmacros

@@ -0,0 +1 @@
+%_openpgp_sign_id @GPG_NAME@

script/generate-packages-json.py

@@ -0,0 +1,154 @@
+#!/usr/bin/env python3
+"""Parse RPM repodata and emit a packages.json manifest for the UI."""
+
+import argparse
+import gzip
+import json
+import os
+import subprocess
+import sys
+import xml.etree.ElementTree as ET
+from datetime import datetime, timezone
+
+RPM_NS = "http://linux.duke.edu/metadata/common"
+OTHER_NS = "http://linux.duke.edu/metadata/other"
+REPO_NS = "http://linux.duke.edu/metadata/repo"
+
+
+def find_repodata_file(repodata_dir, data_type):
+    """Read repomd.xml and return the path to a specific data type's file."""
+    repomd_path = os.path.join(repodata_dir, "repomd.xml")
+    tree = ET.parse(repomd_path)
+    root = tree.getroot()
+
+    for data in root.findall(f"{{{REPO_NS}}}data"):
+        if data.get("type") == data_type:
+            location = data.find(f"{{{REPO_NS}}}location")
+            if location is not None:
+                href = location.get("href", "")
+                return os.path.join(os.path.dirname(repodata_dir), href)
+
+    return None
+
+
+def open_compressed(path):
+    """Open a gzip or zstd compressed file for reading."""
+    if path.endswith(".zst"):
+        result = subprocess.run(
+            ["zstdcat", path], capture_output=True, check=True
+        )
+        import io
+        return io.BytesIO(result.stdout)
+    else:
+        return gzip.open(path, "rb")
+
+
+def parse_primary(repodata_dir):
+    """Parse primary.xml.{gz,zst} and return package metadata."""
+    path = find_repodata_file(repodata_dir, "primary")
+    if not path:
+        print("error: primary metadata not found in repomd.xml", file=sys.stderr)
+        sys.exit(1)
+
+    packages = {}
+    with open_compressed(path) as f:
+        tree = ET.parse(f)
+
+    for pkg in tree.getroot().findall(f"{{{RPM_NS}}}package"):
+        if pkg.get("type") != "rpm":
+            continue
+
+        name = pkg.findtext(f"{{{RPM_NS}}}name", "")
+        version_el = pkg.find(f"{{{RPM_NS}}}version")
+        ver = version_el.get("ver", "") if version_el is not None else ""
+        rel = version_el.get("rel", "") if version_el is not None else ""
+        arch = pkg.findtext(f"{{{RPM_NS}}}arch", "")
+
+        size_el = pkg.find(f"{{{RPM_NS}}}size")
+        size = int(size_el.get("package", "0")) if size_el is not None else 0
+
+        time_el = pkg.find(f"{{{RPM_NS}}}time")
+        build_time = int(time_el.get("build", "0")) if time_el is not None else 0
+
+        location_el = pkg.find(f"{{{RPM_NS}}}location")
+        filename = os.path.basename(location_el.get("href", "")) if location_el is not None else ""
+
+        key = f"{name}-{ver}-{rel}"
+        packages[key] = {
+            "name": name,
+            "version": ver,
+            "release": rel,
+            "arch": arch,
+            "summary": pkg.findtext(f"{{{RPM_NS}}}summary", ""),
+            "size": size,
+            "buildTime": build_time,
+            "rpmFilename": filename,
+            "changelog": [],
+        }
+
+    return packages
+
+
+def parse_other(repodata_dir, packages):
+    """Parse other.xml.gz and attach changelog entries to packages."""
+    path = find_repodata_file(repodata_dir, "other")
+    if not path:
+        return
+
+    with open_compressed(path) as f:
+        tree = ET.parse(f)
+
+    for pkg in tree.getroot().findall(f"{{{OTHER_NS}}}package"):
+        name = pkg.get("name", "")
+        version_el = pkg.find(f"{{{OTHER_NS}}}version")
+        ver = version_el.get("ver", "") if version_el is not None else ""
+        rel = version_el.get("rel", "") if version_el is not None else ""
+        key = f"{name}-{ver}-{rel}"
+
+        if key not in packages:
+            continue
+
+        for entry in pkg.findall(f"{{{OTHER_NS}}}changelog"):
+            packages[key]["changelog"].append({
+                "author": entry.get("author", ""),
+                "date": int(entry.get("date", "0")),
+                "text": (entry.text or "").strip(),
+            })
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--repodata-dir",
+        required=True,
+        help="path to the repodata/ directory",
+    )
+    parser.add_argument(
+        "--output",
+        required=True,
+        help="path to write packages.json",
+    )
+    parser.add_argument(
+        "--base-url",
+        required=True,
+        help="public base URL for the repo (e.g. https://rpm.lair.cafe/fedora/43/x86_64)",
+    )
+    args = parser.parse_args()
+
+    packages = parse_primary(args.repodata_dir)
+    parse_other(args.repodata_dir, packages)
+
+    manifest = {
+        "generated": datetime.now(timezone.utc).isoformat(),
+        "baseUrl": args.base_url,
+        "packages": list(packages.values()),
+    }
+
+    with open(args.output, "w") as f:
+        json.dump(manifest, f, indent=2)
+
+    print(f"wrote {len(packages)} packages to {args.output}")
+
+
+if __name__ == "__main__":
+    main()