Build Images

Building VM Images

Overview

Capsem image builds are profile-led.

• config/profiles/<profile_id>/profile.toml is the profile ledger.
• Profile sibling files own packages, MCP declarations, rule files, detection files, tips, build-time hooks, and packaged guest root seed files.
• capsem-admin validates profile-owned inputs and materializes the generated backend build workspace.
• 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

Read config/README.md before changing this layout.

config/
  settings/               UI/application preferences and generated UI schema
  corp/                   Corporate source contracts and rule files
  docker/                 Dockerfile/build templates
  profiles/<profile_id>/
    profile.toml          Source ledger; no hash/size pins
    enforcement.toml      Profile enforcement rules
    detection.yaml        Profile Sigma detections
    mcp.json              Profile MCP declarations
    apt-packages.txt      Profile apt package input
    python-requirements.txt / python-requirements.lock
    npm-packages.txt / npm-package-lock.json
    build.sh              Profile image build hook
    tips.txt              Profile guest tips
    root/                 Guest / seed, projected by capsem-init
cache/target/config/            Generated runtime config with asset/file evidence
cache/target/assets/            Generated VM assets
cache/target/packages/          Generated native packages
guest/artifacts/          Core guest payloads: init, doctor, diagnostics, bench

The materialized backend workspace may contain generated package-set files and profile build scripts. Treat those as implementation details, not authoring surfaces. The workspace is never a config root and never a second profile catalog.

capsem-admin is a tool, not a config authority. It validates, materializes, builds, and checks the profile/corp/settings contracts; it must not grow scaffolding commands that invent profile, MCP, AI provider, package, or rule truth outside config/profiles, config/corp, and config/settings. Do not add admin config roots, guest config roots, settings metadata, provider registries, or backend-owned profile catalogs as product truth. schema validates one contract, catalog lists materialized profile instances, and UI metadata only helps render settings.

CLI commands

just _build-assets code [arch]                # Profile-derived asset rebuild
just _build-kernel arm64 code                 # Kernel slice
just _build-rootfs arm64 code                 # Rootfs slice
uv run --project build_system --frozen capsem-builder audit                  # Parse trivy/grype vulnerability output

Use admin/just recipes for all product 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 profile/image authoring. capsem-admin image build may call private Python modules such as capsem.builder.image_build_backend; agents must not make those modules public CLI contracts.

Building assets

Full rebuild (kernel + rootfs):

just _build-assets    # Runs doctor + validate + build for host arch

Individual templates:

just _build-kernel arm64
just _build-rootfs arm64

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)

Rootfs EROFS settings are profile-derived. The approved release default is EROFS with lz4hc compression level 12.

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. It also scans every exported member against the configured forbidden prefixes. Keep Ollama accelerator families forbidden under both /usr/lib/ollama and /usr/local/lib/ollama: vendor archives have moved the bundle between those roots, and the guest exposes no accelerator device. A profile hook should remove the payload for efficiency, while the exported-tree scan remains the fail-closed proof. 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 declared profile package inputs, rendered rootfs package lists, profile root/build-script inputs, EROFS config, git revision, and project version. Installed-package/component truth belongs in the CycloneDX OBOM, not the build ledger.

Profile Source And Generated Evidence

Profile sibling files are ledgered source inputs, but agents must not add or hand-edit hash or size fields in profile.toml. If editing apt-packages.txt, python-requirements.txt, npm-packages.txt, build.sh, rules, MCP declarations, tips, or root seed files makes capsem-admin profile check fail, fix the source contract or the validation/materialization rail with tests. Do not "just fix the hash" in TOML.

Generated runtime asset URLs/hashes belong in cache/target/config after capsem-admin profile materialize, not in checked-in source TOML. Profile materialization must recopy descriptor files and root/ payloads from source on every run; stale generated roots are a release blocker, not a cache.

Adding packages to the VM

1. Edit the profile-owned package file, for example config/profiles/code/apt-packages.txt, python-requirements.txt, or npm-packages.txt. Python and npm direct selections must be exact, and their checked-in integrity locks must be regenerated together; never install from the direct list alone.
2. Run the admin/profile validation path.
3. Run just _build-assets code to rebuild the rootfs.
4. Verify with capsem-doctor inside a booted VM.

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

Adding a guest CLI/tool

There are no image-owned AI providers. A CLI/tool exists only if the active profile declares the package/build hook and any required guest root seed files.

1. Add package input to the profile package files, or add build-time shell work to profile-owned build.sh.
2. Add config files under config/profiles/<profile_id>/root/ so they project into the VM at boot.
3. Add MCP declarations to profile-owned mcp.json when relevant.
4. Add network/model/security behavior through profile/corp rules, not builder provider config.
5. Let the credential broker plugin capture/materialize credentials at runtime; do not add settings-owned boot secrets.
6. Rebuild with just _build-assets code and verify with capsem-doctor.

build.sh is executed only while constructing the rootfs image. It is the right place for digest-bound official binary archives such as Claude, AGY, or Ollama when they cannot be represented as apt/npm/Python package inputs. Never pipe or invoke a mutable installer endpoint. Install verified runtime binaries under system paths such as /usr/local/bin; anything left only under /root can be hidden by the runtime overlay.

Profile build.sh contract

Remember this rail when touching profile image contents:

• config/profiles/<profile_id>/build.sh is a profile-owned build hook.
• It runs inside the rootfs Docker build, before the EROFS image is produced.
• It does not run during native glow-up, service startup, VM boot, or user session creation.
• It is for image construction work that cannot be cleanly expressed through apt-packages.txt, python-requirements.txt, or npm-packages.txt.
• It may install public runtime tools such as Claude, AGY, and Ollama into stable system paths.
• It is not a second profile format, provider registry, runtime settings file, credential injection path, or local developer repair script.
• It must not bake credentials, per-user state, corp policy, rules, MCP decisions, or runtime settings.
• The owning profile.toml must reference it through [files.build]; the descriptor hash/size is refreshed by the profile-derived build rail, never by hand.
• Changing build.sh changes future rootfs assets only. Rebuild assets through the profile-derived just/admin-tool rail before claiming a VM contains the change.
• The same profile materialization path must be used locally and in CI; no one-off Docker or installer path is release proof.
• Verification must be black-box: boot the rebuilt profile image, run the tool from the VM, and inspect the generated session evidence when the tool should produce network, model, MCP, file, process, or credential events.

Decision rule:

• Normal Debian package: use apt-packages.txt.
• Normal Python package: use python-requirements.txt.
• Normal npm package: use npm-packages.txt.
• Vendor shell installer, binary tarball, wrapper creation, or cleanup that must happen while baking the immutable rootfs: use build.sh.
• Anything that depends on user/corp/runtime state: do not use build.sh.

Dockerfile templates

Templates live in config/docker/:

• Dockerfile.rootfs-dependencies.j2 -- snapshot-selected Debian packages and the profile's network-resolved Python/npm/vendor inputs;
• 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 profile/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.

Templates use Jinja2 with variables from the admin-materialized profile image workspace. Do not add a second preview rail for product truth; if a build input needs validation, add it to the normal profile/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: Profile -> admin materialization -> Pydantic -> context dict -> Jinja2 -> Dockerfile

The data flows through four layers:

1. Profile ledger (config/profiles/<id>/profile.toml) and profile-owned sibling files.
2. capsem-admin validates and materializes a backend build workspace.
3. Pydantic models (src/capsem/builder/models.py) parse that workspace.
4. Context dict (src/capsem/builder/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:

{
    "arch": ArchConfig,           # Per-arch settings (docker_platform, rust_target, etc.)
    "arch_name": str,             # "arm64" or "x86_64"
    "apt_packages": list[str],    # Materialized from profile apt-packages.txt
    "python_packages": list[str], # Materialized from profile python-requirements.txt
    "python_install_cmd": str,    # e.g. "uv pip install --system --break-system-packages"
    "npm_packages": list[str],    # Materialized from profile npm-packages.txt
    "profile_root_seed": bool,    # Whether profile-root/ is copied into the image
    "profile_build_script": bool, # Whether profile-build.sh is executed
    "npm_prefix": str,            # e.g. "/opt/ai-clis"
    "guest_binaries": list[str],  # ["capsem-pty-agent", "capsem-net-proxy", "capsem-mcp-server"]
}

Kernel context dict

{
    "arch": ArchConfig,
    "arch_name": str,
    "kernel_version": str,  # exact checked-in release, e.g. "6.6.130"
    "kernel_sha256": str,   # verified before the source archive is extracted
}

How to: Change a shipped CLI

1. Prefer a profile package file (apt-packages.txt, npm-packages.txt, or python-requirements.txt) when the tool has a normal package manager.
2. Use profile-owned build.sh when the vendor ships an official shell installer. The build hook runs during rootfs construction only.
3. Make sure binaries end up in stable system paths such as /usr/local/bin.
4. Validate and materialize through capsem-admin; if the rail cannot express the change, implement it with tests first.
5. Add or update capsem-admin materialization tests and Docker context tests.
6. Rebuild: just _build-assets code and verify with capsem-doctor.

Ollama is intentionally installed by config/profiles/<id>/build.sh, not by a VM one-off command. That keeps Codex, Claude, AGY, and OpenAI-compatible local testing available in every shipped profile image that declares the hook.

How to: Add a new package to an existing set

1. Edit config/profiles/<profile_id>/apt-packages.txt, python-requirements.txt, or npm-packages.txt.
2. Validate and materialize through capsem-admin.
3. Keep the checked-in profile source free of generated hashes or sizes.
4. Rebuild: just _build-assets <profile_id>.

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

Backend Workspace Schema

The backend workspace is generated by capsem-admin; do not author it by hand for product behavior. Its install inputs are package-set TOML files:

[npm]
name = "Node Packages"
manager = "npm"
install_cmd = "npm install -g --prefix /opt/ai-clis"
packages = ["@scope/package"]

Profiles own CLI/tool selection. If an installer cannot be represented as a package set, put it in config/profiles/<profile_id>/build.sh, reference it from [files.build] in profile.toml, refresh pins with capsem-admin, and rebuild through the admin/just rail. Do not add a provider registry under backend-generated image workspaces.

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. Extract tool versions (extract_tool_versions)
8. 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 profile release run compiles that graph three times -- co-work rootfs, code rootfs, and assets.pack-initrds -- so emulation costs roughly forty 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, npm, and curl-based CLI installers 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 src/capsem/builder/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