Dakota Overview#

Load when you need context on what dakota is, how it differs from production Bluefin, what unique features it has, or when planning new package additions.

When to Use#

Use when you need product/repo context before planning work, when someone asks what Dakota is, or when deciding whether a package or feature fits Dakota's scope.

When NOT to Use#

• Authoring .bst element files → use buildstream.md or add-package.md
• Debugging CI pipeline failures → use ci.md
• Looking up specific packaging patterns → use the relevant packaging-*.md skill

Core Process#

1. Confirm whether the question is about Dakota's identity, architecture, or scope.
2. Use the hard facts here to avoid importing bluefin or OSTree assumptions.
3. Route into a packaging or CI skill once the repo context is clear.

Hard Facts — Violations Recorded 3+ Times#

Dakota runtime is composefs. It is NOT OSTree.

• The BST export includes /sysroot/ artifacts (OSTree build leftovers) — --prune /sysroot/ strips them at rechunking time.
• The booted system uses the composefs-oci backend. There is no ostree runtime on a running dakota system.
• Never suggest OSTree-specific tooling (bootupd, ostree admin, rpm-ostree) for a running dakota system.

zstd is disabled. Plain podman push is correct.

• zstd fails with bootc composefs ("unexpected EOF reading tar entry") regardless of flags.
• After chunkah rechunking (chunkah build → podman load), plain podman push is the correct path.
• Do not reintroduce skopeo, --compression-format=zstd:chunked, or any oci-dir workaround for post-chunkah pushes. Read issue projectbluefin/dakota#119 before asserting anything about push compression.

Hardware confirmation is required before upstream PRs.

• The validation gate is: bootc upgrade on test hardware succeeds + reboot + GDM active.
• CI green is not sufficient. Hardware confirmation is.

Production image = ghcr.io/projectbluefin/dakota:stable.

• Streams: :testing (nightly, e2e-gated), :stable (weekly promotion)
• Rolling nightly stream: :next / :btw (GNOME 51 master — see below)
• When someone says "is X in the image", check the GHCR image via skopeo inspect or podman run --rm — not a local machine unless explicitly asked.

Verify hypothesis before stating root cause.

• State missing packages as a hypothesis ("likely cause is X") until confirmed by live evidence.

What Is Dakota?#

Dakota is Project Bluefin's CoreOS-model bootc image — built entirely from source using BuildStream 2. It follows the same architecture as CoreOS and GNOME OS: bootc-native, composefs runtime, no OSTree, no rpm-ostree, no dnf.

freedesktop-sdk provides glibc/systemd/kernel, gnome-build-meta provides GNOME Shell/Mutter/GTK, and dakota adds Bluefin-specific packages on top.

Key positioning: Dakota is a curated subset of production Bluefin, not a 1:1 clone. It intentionally includes things production Bluefin doesn't have (sudo-rs, uutils-coreutils, GNOME nightly) and intentionally omits things that don't make sense for a from-source build (Nvidia drivers, ZFS, enterprise AD/Kerberos).

Published image: ghcr.io/projectbluefin/dakota:{testing,stable,next,btw}

Image Streams#

:next / :btw — Rolling GNOME 51 stream#

:next is dakota's fastest variant — tracks gnome-build-meta master (GNOME 51
development branch). It is positioned as the arch-competitor stream: newest
GNOME Shell, newest GTK, newest everything.

Key differences from main:

• gnome-build-meta.bst tracks master (not gnome-50)
• No bootc override (GNOME 51 master ships current bootc)
• Junction bumps are fully automated — track-next-junctions.yml at 20:00
  UTC opens auto-merge PRs; no human review required
• No promotion to :stable — ever. This stream does not feed the weekly
  release pipeline.
• Fixes from main (sbom, CI) must be manually cherry-picked to next —
  they do not land automatically.

When working on next:

git checkout upstream/next -b fix/my-next-fix
git push upstream fix/my-next-fix:next # push directly, no PR needed for fixes

Sync fixes from main:

git log upstream/next..upstream/main --oneline -- Justfile .github/
git cherry-pick <shas>

Historical path note: the repo still uses bluefin in key filenames such as
elements/bluefin/* and oci/bluefin.bst. Those are Dakota paths, not proof
that Dakota follows bluefin's dnf/Containerfile overlay model.

Architecture Comparison#

Fundamental Difference#

Production Bluefin images are Containerfile-based overlays — they start with FROM base_image and run dnf install. Dakota builds the entire stack from source using BuildStream. With good cache hits from upstream CAS, most is pre-built. But Bluefin-specific Rust packages (bootc, uutils-coreutils, sudo-rs) and GRUB are compiled from source.

What Dakota Has That Others Don't#

Gap Analysis#

Gaps as of 2026-06-03. Y = present, N = absent.

Shell & Terminal Tools#

Networking & VPN#

Containers#

Hardware & Drivers (Intentionally Out of Scope)#

Nvidia drivers, ZFS, Xbox controller (xone), Framework laptop modules — not in scope. Building these from source is not practical.

Notable Gaps (Priority Order)#

Build Optimization Notes#

Heavy build contributors:

1. Rust packages — bootc (~200 crates), uutils-coreutils (~250 crates), sudo-rs. These are the crown jewels of dakota's approach — keep building from source.
2. GRUB — built in 3 variants (i386-pc, i386-efi, x86_64-efi). Required because upstream GNOME OS uses systemd-boot only; Bluefin needs GRUB for bootc compatibility.
3. Junction patches — patches to freedesktop-sdk and gnome-build-meta modify junction identity hashes, which may affect upstream cache hit rates. Upstreaming patches would improve this.
4. Pre-built binary pattern — used for Tailscale, Zig, Homebrew, Go CLI tools. New packages should follow this pattern where possible.

Common Rationalizations#

Red Flags#

• Suggesting RPM, rpm-ostree, or OSTree-specific remediation for Dakota runtime problems
• Treating :next as a stable-promotion stream
• Describing Dakota as a 1:1 clone of production Bluefin

Verification#

• The answer reflects Dakota's actual runtime and build model
• Bluefin/OSTree assumptions were explicitly filtered out
• The scope advice matches Dakota's intended positioning

Lessons Learned#

zstd compression is incompatible with bootc composefs (2026-06-07)#

Attempting to push with --compression-format=zstd:chunked or via skopeo after chunkah rechunking fails with "unexpected EOF reading tar entry" at the composefs layer. After chunkah build → podman load, the only working push path is plain podman push. Do not reintroduce skopeo, oci-dir workarounds, or zstd flags for post-chunkah pushes. See projectbluefin/dakota#119 for the investigation.

Dakota is composefs at runtime — not OSTree (2026-06-07)#

The BST export includes /sysroot/ artifacts (OSTree build tooling leftover). These are stripped via --prune /sysroot/ at rechunking time. The booted system uses the composefs-oci backend — there is no ostree runtime on a running Dakota system. Never suggest rpm-ostree, bootupd, or ostree admin commands for a running Dakota system.

countme infrastructure — reports to Fedora, not a custom endpoint#

Dakota ships /usr/libexec/dakota-countme (enabled via bluefin-countme.timer, weekly). It reports to Fedora's countme infrastructure by curling the metalink URL with a libdnf5-format User-Agent — the same protocol dnf5/rpm-ostree-countme use. No dnf or RPM packages needed.

Key files:

• files/countme/dakota-countme — the script (reads os-release, tracks age cookie, builds the URL)
• files/countme/bluefin-countme.{service,timer} — systemd units
• elements/bluefin/countme.bst — installs everything
• elements/oci/os-release.bst — PLATFORM_ID: "platform:fNN" is the Fedora base version source

The script identifies the OS as Dakota in the User-Agent so it's distinguishable from Bluefin in Fedora's dataset. When gnome-build-meta bumps to a new Fedora base, update PLATFORM_ID in os-release.bst in the same commit (see patch-junctions.md).

Opt-out: create /etc/bluefin-countme-opt-out.

The validation gate is: bootc upgrade on test hardware succeeds + reboot + GDM active. CI passing confirms the image was built and passes bootc container lint. It does not confirm the image boots correctly on real hardware. Do not mark issues resolved without hardware evidence, and do not state something is "fixed" based on CI alone.