Continuous integration

CI uses separate reusable Test, Build, and Documentation workflows, with one final aggregate Gate in the entry workflow. Protect main with that single required check (Gate under the CI workflow). It runs on every pull request, main push, or manual dispatch. The reusable workflows also accept workflow_dispatch so a leg can be re-run on its own. All required jobs must succeed, including every matrix leg. Failures, cancellations and unexpected job skips fail the gate. Linux-only format/vision steps are intentionally skipped on other test hosts. Concurrency groups cancel superseded runs on the same PR or ref; no repository settings are changed by the workflow.

Tests

• Linux x64, Windows x64 and macOS arm64 run locked CPU-only Cargo Clippy and workspace tests. Linux also checks formatting and the CPU vision feature's validation/unit tests. Model-dependent tests keep their existing #[ignore].
• macOS supports Apple Silicon ARM64 only. Both component workflows verify runner.arch and uname -m before any checkout/build/test. There are no Intel macOS matrix entries or archives; the native packager also rejects that host.
• Linux checks Python syntax/undefined names with Ruff's E4,E7,E9,F rules, installs only the locked Python dev dependency group (Ruff), then runs offline data/tokenization/training-safety/evaluation helper tests and evaluator fixture tests. The opt-in training runtime tests remain skipped.
• No CUDA, Metal or Vulkan feature/toolkit, GPU runner, pretrained model, dataset preparation/download, training, inference benchmark or accuracy evaluation is enabled. Offline Hugging Face flags and JET_RUN_TRAINING_RUNTIME=0 reinforce that boundary. There is no Electron deliverable.

Documentation

The Documentation workflow installs the locked Node and pnpm tools, restores the frozen dependency graph, and builds the VitePress site. The build validates page rendering and internal links, then uploads the static output as a GitHub Pages artifact. Pull requests must pass this build. A successful main-branch Gate deploys the same artifact to GitHub Pages.

Packaging

Native CPU hosts produce these CLI archives, retained for 14 days as Actions artifacts:

Host	Artifact
Windows x64 (windows-latest)	jet-cpu-windows-x64.zip
Linux x64 (ubuntu-26.04)	jet-cpu-linux-x64.tar.gz
Linux arm64 (ubuntu-26.04-arm)	jet-cpu-linux-arm64.tar.gz
macOS arm64 (macos-latest)	jet-cpu-macos-arm64.tar.gz

The official runner image list was checked on 2026-10-02: macos-latest maps to macOS 26 ARM64, and windows-latest maps to Windows Server 2025. Ubuntu's newest available image is 26.04; ubuntu-latest still maps to 24.04, so both Linux architectures use the newer explicit 26.04 labels. The single Gate and Python checks also use 26.04.

cargo xtask package builds the actual release CLI with --locked and no optional features; the release optimization profile is preserved. It includes Jet/llama.cpp licenses, README and small JSONL fixtures. It extracts each archive and checks --version, --help, and model-free export-prompts JSONL output. It never publishes a release or signs an artifact. Linux packages use the host glibc/C++ runtime (Ubuntu 26.04 baseline), macOS uses system libraries, and Windows statically links the MSVC runtime. These are CPU packages, not GPU runtime bundles. GPU packaging scripts remain available for local use.

JET_CPU_PORTABLE=1 disables build-host CPU tuning and SSE4.2/AVX/AVX2/BMI2/FMA/F16C for portable x64 artifacts, and sets the ARM CPU baseline to armv8-a. This trades performance for portability; local builds retain normal native tuning unless explicitly opted in. Run cargo xtask package (or mise run package:cpu) with the native platform toolchain installed; CI uses the same Rust entry point with --platform to check the expected host. Rust libraries generate and extract both archive formats without PowerShell, Bash or external archivers. Running the task again replaces its CPU archive.

Dependencies, permissions and caches

Mise 2026.10.0 is pinned separately from action revisions; the already-merged tooling PR #1 supplies the minimum version. Only Rust/CMake or uv are installed with --locked from mise.lock. Cargo uses Cargo.lock; Ruff is installed with uv sync --locked --only-group dev and uses training/.python-version. No lockfile is regenerated in CI. Cargo.lock also locks the xtask archive dependencies. GitHub Actions use floating major tags (for example actions/checkout@v7, jdx/mise-action@v5) rather than commit SHAs or exact patch pins. Renovate watches the github-actions manager with digest pinning disabled so those majors stay current.

All workflows have contents: read; checkout disables persisted credentials. PR setup receives an explicitly empty token, with no downstream token default. Non-PR setup can use the read-only workflow token for tool downloads. No secrets are inherited, and pull_request_target is not used.

Mise caches include tool version, OS/architecture, install subset and config/lock hashes. Cargo caches separate test/release configurations and include OS, architecture, runner image version, tool/dependency locks and the pinned llama.cpp gitlink. These identify the native toolchain and dependency inputs; Cargo/CMake track changed source files. The run ID lets successful trusted runs refresh build caches without hashing workflow/source files. Python's uv cache uses OS/architecture and uv.lock; uv tracks wheel/interpreter compatibility. PRs restore caches but cannot publish them. Build output archives are not cached.

Run apm install to restore the project-local Stop That Shit skill from lennney/stop-that-shit/skills/stop-that-shit at its locked revision.