Building the Capsem VM runtime

Overview

Capsem builds one VM runtime: a kernel, initrd and rootfs per architecture. Applications come from OCI images (images/, resolved through images/catalog.toml), never from the runtime rootfs, so every build of a commit produces the same asset set.

• config/docker/image/ is the runtime's source: build.toml (package set, EROFS settings, size budgets, kernel pin, base images), manifest.toml, and the kernel/, security/ and vm/ inputs.
• guest/artifacts/ holds the core guest payloads: capsem-init, doctor, diagnostics, bench, the container launcher.
• capsem-admin image build copies those two inputs into a generated backend workspace under cache/target/build/image-workspace/<arch|all>/ and runs the build there.
• The Python builder backend renders Docker templates and emits assets, build ledgers, and OBOMs only when invoked by the admin build rail. Do not add product truth directly to the backend image-spec path.

Source Layout

config/
  docker/                 Dockerfile templates (*.j2)
  docker/image/
    build.toml            Runtime package set, EROFS, rootfs budgets, kernel pin
    manifest.toml         Image manifest metadata and changelog
    kernel/               defconfig.<arch> and ordered kernel patches
    security/             Built-in web security inputs
    vm/                   Guest shell environment
guest/artifacts/          Core guest payloads: init, doctor, diagnostics, bench
images/                   OCI application images and their catalog
cache/target/assets/            Generated VM assets
cache/target/packages/          Generated native packages

The materialized backend workspace is an implementation detail, not an authoring surface. It is never a config root.

capsem-admin is a tool, not a config authority. It validates, materializes, builds, and checks contracts; it must not grow scaffolding commands that invent MCP, AI provider, package, or rule truth. Do not add admin config roots, guest config roots, settings metadata, provider registries, or backend-owned catalogs as product truth.

CLI commands

just build-assets [arch]                      # The runtime for one arch, or every one
just _build-assets [arch]                     # Same rebuild, CI-facing name
just _build-kernel arm64                      # Kernel slice
just _build-rootfs arm64                      # Rootfs slice
uv run --project build_system --frozen capsem-builder audit                  # Parse trivy/grype vulnerability output

Underneath, the gate runs capsem-admin image build [--config-root config] [--guest-dir guest] [--output assets] [--arch arm64|x86_64] [--template all|kernel|rootfs] [--clean] [--json].

Use admin/just recipes for all runtime image work. capsem-builder is a backend helper only; it must not expose or document public build, validate, inspect, mcp, render-only, or dry-run rails for image authoring. capsem-admin image build may call private Python modules such as capsem_builder.image.image_build_backend; agents must not make those modules public CLI contracts.

Per-arch asset layout

cache/target/assets/
  manifest.json          Version, checksums, asset list
  B3SUMS                 BLAKE3 checksums
  arm64/
    vmlinuz              Kernel
    rootfs.erofs         Root filesystem
    initrd.img           Initial ramdisk (repacked by just exec)

The approved release default is EROFS with lz4hc compression level 12 (config/docker/image/build.toml [build.erofs]).

config/docker/image/build.toml [build.rootfs] also owns two independent release budgets: the exported tar ceiling and the final EROFS ceiling. The builder checks the first before compression and the second before writing the artifact ledger. Never raise a size ceiling to accommodate an unexplained image jump; inspect the rootfs composition first.

Build Ledger

Each per-arch build emits build-ledger.log JSONL. The rootfs.config_inputs record captures the rendered rootfs package list and the EROFS config. Installed-package/component truth belongs in the CycloneDX OBOM, not the build ledger.

Adding a guest tool

The runtime package set is config/docker/image/build.toml [build.rootfs].runtime_apt_packages. Add a package there only when Capsem's own guest machinery needs it: the container launcher (runc, umoci, iptables, iproute2), capsem-init (auditd, e2fsprogs), the trust store the Capsem CA lands in, or what capsem-doctor, capsem-bench and the network diagnostics run with.

Anything a user or an agent works with -- an AI CLI, a language toolchain, a vendor binary -- belongs in an OCI image under images/, not in the runtime.

1. Add the package to runtime_apt_packages (no duplicates; the model refuses an empty or duplicated list).
2. Rebuild with just build-assets (or just _build-rootfs <arch>).
3. Verify with capsem-doctor inside a booted VM.

Do not edit generated Dockerfiles. Docker templates live under config/docker/.

Dockerfile templates

Templates live in config/docker/:

• Dockerfile.rootfs-dependencies.j2 -- snapshot-selected Debian packages;
• Dockerfile.kernel-dependencies.j2 -- snapshot-selected kernel toolchain and the SHA-256-verified kernel archive;
• Dockerfile.rootfs.j2 -- network-denied first-party rootfs assembly;
• Dockerfile.kernel.j2 -- network-denied kernel compilation, initrd assembly, and vmlinuz extraction.

asset-dependencies is the visible resumable frontier between those pairs. The gate materializes one input-keyed helper for every selected architecture/template, validates its platform and identity label, and passes only its exact image ID to the source build. Source builds always use BuildKit network none and never use the remote CI cache. A carried frontier must revalidate every helper; it must not silently rebuild inside the sealed lane. The dependency helpers reuse the one checked-in Debian snapshot authority rather than the mutable sources inherited from the base image.

Private build-helper generations

Local qualification and hosted qualification use one materialization path. Whenever an exact input-keyed helper is absent locally, its owning materializer first tries the platform-scoped generation under ghcr.io/google/capsem-ci. It validates the pulled image's platform, input-key label, and org.opencontainers.image.source label before retagging it locally. An unavailable, unauthenticated, or invalid remote generation falls through to the same deterministic network-open local build; there is no cache-only build path and no opt-in environment switch.

The private inventory contains the host builder, Linux parity base, install builder, both package builders, both guest Rust builders, the host asset-tools builder, and the host architecture's kernel and rootfs dependency builders. capsem-gate private-images --publish materializes that inventory before it publishes immutable generation tags. .github/workflows/private-images.yaml runs it weekly and on defining-input changes, independently of the release and official OCI image workflows, and retains the latest three tagged generations per helper and platform.

Developers who can read the repository's private packages may authenticate Docker to GHCR for warm local qualification. Missing credentials only select the local fallback; they do not change the resulting helper identity. Keep all helper Dockerfiles under build_system/docker/ and dependency templates under config/docker/; add any new network-open helper to the single inventory and the repository-source-label guard.

Keep the verified registry alias when retagging an imported helper. BuildKit can fetch inherited lazy layers from that origin; a local-only FROM name can redirect those fetches to Docker Hub. Resolve the registry digest through the shared Docker boundary and retain exact image-ID/platform checks. Cache retirement removes all owned aliases and accounts for the image bytes once. Every child build includes the Linux parity base and source lane: use the shared exact-reference check before building and verify the identity afterward. Without a qualified registry origin, keep the verified local tag even if Docker records local RepoDigests. BuildKit can resolve those digests through Docker Hub instead of the local image store, breaking sealed source builds.

Templates use Jinja2 with variables from the admin-materialized image workspace. Do not add a second preview rail for product truth; if a build input needs validation, add it to the normal admin validation path.

Every architecture's base_image is required to be its immutable repository@sha256:<child-manifest> identity. Do not use a mutable tag or the multi-platform index. capsem-gate materializes a missing exact child through the Docker daemon before the cross-execution probe and build lanes; keep this on the ordinary guarded Docker runner, not the host-process egress broker.

Builder Internals (for modifying the builder itself)

Architecture: config/docker/image -> admin workspace -> Pydantic -> context dict -> Jinja2 -> Dockerfile

1. Runtime source (config/docker/image/ and guest/artifacts/).
2. capsem-admin copies both into a backend build workspace.
3. Pydantic models (build_system/builder/image/models.py) parse that workspace.
4. Context dict (build_system/builder/image/docker.py) feeds Jinja2 templates.
5. Jinja2 templates (config/docker/) produce Dockerfiles.

Key files

Context dict (rootfs template variables)

_rootfs_context() in docker.py builds the dict passed to Dockerfile.rootfs.j2 (plus the Debian snapshot keys):

{
    "arch": ArchConfig,           # Per-arch settings (docker_platform, rust_target, etc.)
    "arch_name": str,             # "arm64" or "x86_64"
    "apt_packages": list[str],    # build.toml [build.rootfs].runtime_apt_packages
    "guest_binaries": list[str],  # GUEST_BINARIES in docker.py
}

Kernel context dict

{
    "arch": ArchConfig,
    "arch_name": str,
    "kernel_version": str,  # exact checked-in release, e.g. "6.18.44"
    "kernel_sha256": str,   # verified before the source archive is extracted
    "kernel_patches": list, # applied in order with --fuzz=0
}

How to: Add a new guest binary

Guest binaries are compiled from crates/capsem-agent/. Every architecture on every host goes through container_compile_agent(); native Linux has no second ambient Cargo/rustup rail. The asset preflight first materializes the host platform's config-selected exact Rust child image plus the checked-in Rust toolchain and Cargo.lock; a foreign target also materializes its exact config-pinned C compiler package and both the foreign-target and host-target Cargo dependency sets. The host set covers target-gated build-script and proc-macro dependencies that cargo fetch --target <foreign> does not select. The actual container build runs with --network none and Cargo --locked --offline.

Materialize only helpers the command can consume: every requested architecture on either supported host, and none for a kernel-only build. Immutable Debian guest bases remain architecture-selected separately.

1. Add the binary target in crates/capsem-agent/Cargo.toml
2. Add the binary name to GUEST_BINARIES list in docker.py
3. The template already loops {% for binary in guest_binaries %} to COPY + chmod 555

Verifying Linux builds locally

just _cross-compile [arch] builds everything in a container: agent binaries, frontend, and the full Linux .deb package. Useful for catching system dependency issues before CI.

just _cross-compile           # Build for host arch (arm64 on Apple Silicon)
just _cross-compile x86_64    # Build x86_64 deb

Build pipeline (what build_image() does)

For rootfs:

1. Build guest agent binaries (cross_compile_agent -- every target uses the pre-materialized, network-denied Rust builder; a foreign target cross-compiles on the host CPU)
2. Assemble build context (prepare_build_context) -- copies CA cert, shell configs, diagnostics, agent binaries
3. Render Dockerfile from template
4. docker build
5. Export container filesystem as tar
6. Create EROFS from tar (create_erofs -- runs mkfs.erofs in a container)
7. Clean up container image

For kernel:

1. Read the exact kernel version and SHA-256 from the checked-in build config
2. Assemble build context (defconfig, capsem-init)
3. Render Dockerfile from template
4. docker build, verifying the downloaded source archive before extraction
5. Extract vmlinuz + initrd.img from image
6. Clean up

The guest Rust builder workspace: do not widen the /src/* glob

container_compile_agent mounts the checkout read-only at /src and assembles a writable workspace at /build:

for f in /src/*; do b=$(basename "$f"); \
  [ "$b" != target ] && [ "$b" != crates ] && ln -s "$f" /build/; done

/src/* does not match dotfiles. That exclusion is load-bearing. Widening it breaks the build. It reads like an oversight -- .cargo/config.toml never reaches the container, so the checked-in Cargo configuration is never applied -- and it has been "fixed" on exactly that reasoning.

Why it must stay: .cargo/config.toml declares

[target.x86_64-unknown-linux-musl]
linker = "rust-lld"

On a developer host, x86_64-unknown-linux-musl is a cross target and rust-lld is correct. Inside the Alpine builder that same triple is the host target, so inheriting the file makes every proc-macro crate -- serde_derive, tokio-macros -- link its host .so with rust-lld:

rust-lld: error: unable to find library -lgcc_s
rust-lld: error: unable to find library -lc
error: could not compile `tokio-macros`

The rule this generalizes to: checked-in Cargo configuration is developer-host configuration. The builder container owns its own toolchain settings and receives them as environment on the docker run, never by reading them out of the tree. If a container build needs a linker or a CC, pass it explicitly.

build_system/tests/gate/test_guest_rust_builder_hermetic.py::test_container_workspace_excludes_dotfiles fails if the glob is widened, so this cannot be rediscovered the slow way.

Cross-compiling guest binaries instead of emulating

Foreign-target guest builds used to run rustc under QEMU on the target's own platform child. Measured on a 16-core Linux host, cold, --locked --offline, for the six aarch64 guest binaries:

A release run compiles that graph for every architecture's rootfs and again for assets.pack-initrds, so emulation costs tens of minutes per run.

What a cross image needs, materialized at image-build time on the same network-open setup edge that cargo fetch --locked already uses:

• apk add --no-cache ${CROSS_PACKAGES}, where config owns the exact package tuple (currently clang21=21.1.2-r2) and the helper identity includes it. ring is the only crate in the capsem-agent + capsem-bench graph that compiles C -- nothing else pulls cc, cmake or bindgen. Alpine's clang cross-compiles it for a foreign musl target with no external sysroot. The pinned Rust toolchain already supplies rust-lld; no separate ambient linker package is installed.
• rustup target add "${RUST_TARGET}", after which the existing rustup target list --installed assertion proves it landed.
• CC_<target>=clang and CFLAGS_<target>=--target=<target> on the run.
• CARGO_TARGET_<TARGET>_LINKER=rust-lld on the run -- not by linking .cargo/config.toml into the workspace, for the reason in the section above.

The runtime build stays --network none with --locked --offline. The image tag is keyed by the base, target, cross-package tuple, Dockerfile and lockfiles, so a change to any of them is a different image rather than a silent reuse.

Container runtime requirements

On macOS, Docker runs inside a Colima VM with limited resources. The rootfs build runs apt and the guest Rust builds concurrently -- the default RAM allocation may cause OOM kills (exit code 137).

Minimum: 12GB RAM. Recommended: 16GB RAM, 8 CPUs.

# Colima (macOS)
colima stop && colima start --vm-type vz --vz-rosetta --memory 16 --cpu 8

# Linux: Docker runs natively, no memory tuning needed
# sudo apt install docker.io

just doctor owns the product readiness gate. capsem-builder doctor is a backend helper used by the build rail to check container/runtime prerequisites.

The resource check lives in build_system/builder/image/doctor.py:

• check_container_resources() -- checks docker info
• Thresholds: DOCKER_MIN_MEMORY_MB = 4096, DOCKER_RECOMMENDED_MEMORY_MB = 8192

Container image compatibility

Guest cross-build containers use the exact per-platform rust:1.97.1-alpine3.23 child manifests in config/docker/image/build.toml, never a mutable Rust tag. Those children already own the exact toolchain, native musl target, musl headers, and compiler. At the guarded asset-prefetch boundary, build_system/docker/Dockerfile.guest-rust-builder resolves the Cargo.lock graph and, only for a foreign target, adds the exact config-pinned C compiler package and Rust target. Do not add package installation, target installation, index updates, or downloads to container_compile_agent(); its runtime network is deliberately none.

The local helper tag is an input cache key, not an OCI content digest. Two cold Docker builds may have different image IDs because registry/index and layer metadata are materialization outputs. Cargo verifies every registry package against Cargo.lock, and the nightly/release qualification boundary is the specific helper image materialized by that run: after this one guarded fetch edge, the binary build is locked, offline, and network-denied. Do not claim byte-for-byte reproducible helper images unless every remaining registry and layer byte is independently pinned.

The selected image is a minimal Alpine image and has no Bash. Many common utilities (file, less, vim, etc.) are NOT available. Runtime shell commands must be POSIX sh and use only the BusyBox tools already present in the materialized image.

Lesson learned: using file /output/binary to verify compiled binaries failed because file is not in slim images. Replaced with ls -l which is always available and still confirms the copy succeeded. The real validation (existence + non-zero size) is done in Python after the container exits.

Rule: never assume a command exists in a slim container image. Stick to coreutils or install what you need explicitly.

Clock skew workaround

All asset dependency-helper apt-get update calls use -o Acquire::Check-Valid-Until=false -o Acquire::Check-Date=false against the config-owned HTTPS Debian snapshot to handle container VM clock drift. Without this, apt rejects Release files whose timestamp is in the future relative to the VM's clock. This can occur with any container VM backend on macOS.

Files affected:

• Dockerfile.kernel-dependencies.j2
• Dockerfile.rootfs-dependencies.j2