Skip to main content

Roadmap

Status date: 2026-09-03. Living document; the issue tracker is authoritative for day-to-day state, this file is authoritative for shape and sequence.

Caveat on that split, recorded because it has already misled readers: five milestone issues are closed as completed while their exit criteria here are not met, because the implementation landed and the validation named in the issue's own scope did not. Where the two disagree, this file is currently the more accurate. See "Unvalidated paths" below and #186.

The same caveat applies to the GitHub milestone view, more sharply. All six milestones are open with zero open issues and no due date, and none of the repository's open issues is attached to a milestone — including #187, which is M3's own missing exit criterion. Read on its own, that view says the project is finished; M2, M3, M4 and M5 below say otherwise, and they are right. Reconciling the two is tracked as #237.

Vision​

One engine that moves a bootc system between images, backends, bootloaders, base families, and desktops — safely, with a staged+rollback contract, and with the same evidence-first discipline at every step: report before acting, verify after acting, keep the previous state bootable.

Three deliverables share the code:

DeliverableWhat it isStability contract
bootc-migrateThe proven OSTree→ComposeFS migrator (protected MVP)CLI surface, output, and behavior frozen; its four E2E cells are untouchable regression gates
bootc-migrate-coreThe capability library: phases, preflight, /etc merge, transaction, registry streaming, stores, scan, remap, boot audit, DE stash/restoreAdditive growth; everything new lands here first
bootc-rebaseThe universal CLI: routing table × strategiesWhere new user-facing capability ships

Where we are​

M0 (MVP hardening) and M1 (same-backend re-base engine) are done. Every issue under both milestones is closed. bootc-rebase truthfully routes all four backend pairs (all four implemented since #260 landed the composefs→ostree route as Strategy::OstreeInstall, exploratory and covered by one non-gating cell) and the capability scan (#24) covers every proposed probe.

In progress, with an explicit boundary between what's landed and what's deliberately deferred — each of M2, M3, and M5 shipped a pure/unit-testable "skeleton" slice first. M2 stopped there, ahead of boot-critical live-system mutation CI cannot validate; M3 and M5 have since grown their live halves, but those run on paths no CI cell reaches (no cross-base E2E cell, no NVRAM mutation; the TUI-assertions gap now has headless render/state tests plus the tui-migrate cell, which is now gating and green), so they are landed-but-unproven rather than done. See each milestone below for the specific line and why.

Next release gate — first post-rename release​

The next release is a trust checkpoint, not a date-only cut. It must restore one coherent public identity after the repository and binary rename while keeping the protected migrator's proven contract distinct from the broader, partly unvalidated re-base engine.

Before tagging, the release owner must record these decisions and evidence:

  • Identity: Cargo workspace/crate versions, changelog heading, Git tag, binary names, archive names, and container tags agree. The published quick start downloads those exact artifacts; no pre-rename bootc-migrate-composefs path remains in the primary install flow.
  • Scope: release notes state whether bootc-rebase is excluded, included as an experimental preview, or supported. Experimental routes must not inherit the protected MVP's stability language.
  • Validation: all protected-MVP E2E cells pass on the release commit, and a manual dispatch of the release workflow successfully builds both target archives and the container without publishing them.
  • Safety boundary: release notes enumerate every live or interactive path that lacks automated coverage (including cross-base, cross-DE, NVRAM, and checklist paths) and link to recovery/undo guidance.
  • Ownership: one named release owner and target date are attached to the roadmap issue; the owner verifies checksums and install commands from the immutable GitHub Release before announcing it.

This gate deliberately does not require unfinished M2–M5 work to graduate. It allows the proven, renamed migrator to ship while making the newer engine's evidence level visible to adopters. After this release, cadence and the bootc-rebase graduation gate should be tracked separately.

Status of that gate, 2026-09-03: not started, and the cost is now user-visible. The contract exists (RELEASING.md, #171) and no clause of it has been executed. Cargo.toml's workspace version is 0.5.0; the newest GitHub Release is v0.2.0 from 2026-07-04, carrying only pre-rename bootc-migrate-composefs-* archives. So the README quick start's releases/latest/download/bootc-migrate-x86_64-unknown-linux-gnu.tar.gz returns 404, the same URL under the old asset name returns 200, and ghcr.io/tuna-os/bootc-migrate:latest refuses anonymous pulls while ghcr.io/tuna-os/bootc-migrate-composefs still serves them. Both documented install paths therefore fail for a new adopter, and the only artifact that does resolve is the pre-rename binary the README steers people away from.

No release owner and no target date are recorded anywhere, which is the one gate clause that blocks all the others. Tracked as #236.

Unvalidated paths (single list for release notes)​

Consumed by RELEASING.md, which records the release contract (#171): what ships, how versions are chosen, and the pre-tag checklist.

The next release gate requires enumerating "every live or interactive path that lacks automated coverage". Assembling that from five milestone narratives is error-prone, so it is collected here. #186 tracks the fact that each of these had its implementation issue closed as completed while the validation named in that issue's own scope never shipped.

PathStateValidation gapTracking
migrate-bootloader live GRUB2→sd-bootimplemented: stages ESP kernels/BLS entries, leaves GRUB as fallback, arms a one-boot BootNext trial, and supports guarded --promote, --undo, plus a kernel-install resync hookthe gating bootloader-trial cell swaps OVMF BootOrder around the trial reboot and asserts resync, promotion, and undo#65, #189, #382
Boot-entry cleanup (efibootmgr executor)implemented, dry-run default, typed confirmation, NVRAM snapshot + --undolive rename + snapshot restore run in the gating OSTree re-base cell; real-hardware validation remains advisable#31, #189, #204
Cross-base remap + /etc conflict policyimplemented in OstreeDeploy#191 covers the gate; same-family E2E remains; #259 covers the cross-family path#67, #187, #259
DE stash/restore (--de-migrate)implemented, detection table-testedthe non-gating Bluefin→Aurora cell passes --de-migrate and asserts the stash; evidence depends on the exploratory cell and target registry scan succeeding#68, #188
Identity-DB merge across familiesimplemented in cross_family for conversion, OstreeDeploy and ImageSwapcomposefs and OstreeDeploy cells cover identity preservation; the same-family/native gap remains#80, #256, #259
composefs → ostree (ostree_install, Strategy::OstreeInstall)implemented: alongside install through the target's bootc, atomically journaled boot transaction, hash-verified ESP snapshot/restore, /etc merge, /var copy, rollback-first BootOrder, one-shot target BootNext, and target promotion only after an observed boot; the classifier and snapshot/journal round trip are table-testedtwo route cells plus two non-gating fault cells assert resume before arming and rollback after the destructive ESP wipe; real-hardware validation remains advisable#260, #387
Package-host rollback artifacts (package_rollback, bootc-rebase undo)implemented as a reusable snapshot/stage/undo core; rollback wiring into Strategy::PackageInstall remains separateno E2E cell exercises package-host installation, reboot into Previous system, or retention after bootc upgrade; wire through #373 and cover in the package-based matrix #380#374, #373, #380
package → ostree (Strategy::PackageInstall)implemented on the shared alongside-install engine: / is bound at /target with rslave, root-mount kargs are retained, and the old kernel/initramfs/BLS/GRUB paths are snapshotted before bootc empties /boot; argv, karg and snapshot selection are table-testedpackage-host E2E cells are tracked by #380; the old-system boot entry, package-aware /etc, selective data carry and commit remain separate #374–#378 slices#373, #380
composefs → ostree with /boot on an encrypted root (ostree_install::esp_resident)implemented: the route detects /boot without its own partition on a dm-crypt or rd.luks. root, copies the new kernel and initrd to the ESP with a BLS entry that keeps the ostree= argument, makes that entry the systemd-boot default and puts systemd-boot first in BootOrder; detection, entry parsing, the ESP plan and the staging are table-testedno E2E cell installs a composefs host with /boot on LUKS; the hand-made equivalent was staged on a real host, but its boot is not yet recorded; after bootc upgrade, the ESP copy does not follow the new kernel#305
composefs → composefs image swap (Strategy::ImageSwap)implemented; cross-family swaps apply policy after finalization and preserve first-boot repairone ordinary non-gating cell runs; dedicated cross-family image-swap coverage remains#66, #259
Cross-family migration on the bootc switch routes (cross_family)implemented for OstreeDeploy and ImageSwap; target defaults, identity remap and first-boot repair are coveredfive non-gating cells cover refusal, staged /etc and booted identity; community images remain unvalidated#256, #259, #261
First-boot repair (firstboot_repair, L3)implemented on the CoreMigration and OstreeInstall routes; each class and the unit are table-tested on fixture rootsthe CoreMigration cells break a canary tree after staging and assert the ownership repair after the reboot; the flatpak and selinux-label classes run in no cell, and selinux-label gets input only when the verify probe reports that class#309
package → ostree (PackageInstall)source detection, route planning, package-host readiness and hardware reports implemented; TUI exposes an OSTree-only read-only previewinstall, /etc classification, data carry, rollback, software report, commit, first-boot verification and every E2E cell remain open; no package-host source is supported#370–#381
NativeStore (composefs-native)behind a feature flag, off by defaultdefault path still pins a legacy-CLI builder#13
TUI composefs-host route step (#312)implemented: background scan, dual-capable route choice, honest notes, bootc-rebase executioncovered by state-machine tests only; tui-migrate cell runs on an ostree host and does not exercise the composefs-host screens or launch bootc-rebase#312
Package-managed source → OSTree (Strategy::PackageInstall)harness support is implemented: checksum-pinned Fedora Cloud, synthesized Workstation, and Ubuntu Cloud disks, cloud-init SSH, package fixtures and Previous system assertionsthe three #380 cells are exploratory until the package route and its readiness/report/bootloader pieces (#373–#379) land; the Fedora Cloud cell must be promoted to gating after its first green run on main#373–#380
postmarketOS-family ARM64 laptops (apk, LUKS root, ESP at /boot, devicetree)implemented: apk host detection, LUKS root preflight, whole-ESP snapshot and restore, bind plan for ESP at /boot, devicetree & arm64 kargs policy, per-board DTB match against the host's compatible, and a Previous system rollback entry that also restores the pre-install BootOrderno arm64 E2E cell exists, and none can run in the lanes this repo has: the whole matrix needs /dev/kvm, GitHub's arm64 hosted fleet does not surface nested virtualization (actions/runner-images#14062), and .github/runs-on.yml defines x86_64 families only. The cell (LUKS2 argon2id ext4, ESP at /boot, systemd-boot with a devicetree entry) therefore needs an arm64 KVM host — a self-hosted arm64 runner or an EC2 Graviton bare-metal RunsOn profile — before it can be written; X13s hardware qualification remains open behind it#461

Everything not in this table — the OSTree→ComposeFS migrator itself, including rollback and commit — is covered by the seven-cell E2E matrix.

Milestones​

M0 — MVP hardening (continuous; protects everything else) — done​

All exit-criteria issues closed: #72 (cfs CLI drift, resolved short-term by probe+delegation, long-term by #13's NativeStore), #22 (E2E rollback proof),

26 (rollback subcommand), #25 (commit verified fresh-install-identical),

17 (post-migration /var cleanup), #27 (sleep inhibitor), #18 (pre-baked SSH

E2E image), #12 (phase-module unit tests), #29 (Containerfile-initrd alternative — evaluated and superseded by the shipped host-side dracut --rebuild approach, closed as won't-implement).

Exit criteria met: rollback proven in CI; commit/undo produce indistinguishable-from-fresh layouts; no known MVP flakes.

M1 — Same-backend re-base engine (scenarios A / A′) — done​

63 (ostree-rebase E2E cell), #64 (Strategy::OstreeDeploy), #66

(Strategy::ImageSwap), #24 (capability scan: parsers, registry fetch wiring, bootc-rebase scan subcommand, and the Compatible: YES/NO gate) — all closed.

Exit criteria met: bootc-rebase --plan truthfully answers all four backend-pair routes; ostree→ostree proven by its own E2E cell; scan output drives route refusal with evidence.

M2 — Bootloader migration (scenario B) — implemented; live cell added​

#65 — the pure core (BLS entry assembly, kernel-arg carry-over, entry-token derivation) and the standalone live path are implemented in both CLI frontends. The live path copies every existing deployment to the ESP, writes BLS entries, keeps GRUB as a fallback, installs a kernel-install drop-in, and arms a one-boot BootNext trial without changing the persistent BootOrder.

--promote refuses unless BootCurrent identifies the trial entry, then makes that entry persistent. --undo restores the saved firmware order and BootNext, removes the migration-owned NVRAM entry and ESP files, and restores any pre-existing loader configuration and hook.

The bootloader-trial QEMU cell deliberately swaps the saved BootOrder to the trial entry because OVMF does not honor BootNext in this scenario. It reboots through systemd-boot, invokes the installed kernel-install resync hook with a second kernel version, promotes the trial, then undoes it and verifies the original order and GRUB fallback.

Exit criteria: a GRUB2 Bluefin VM stages, trials, resyncs, promotes, and undoes the systemd-boot path in the dedicated QEMU cell.

M3 — Cross-base re-base (scenario C) — both parts landed, neither exercised by CI​

#67 part 1 (remap planner + apply walk over the staged deployment) is done and wired into OstreeDeploy, gated by is_cross_base + --accept-cross-base.

Part 2 (the cross-base /etc conflict policy) was previously recorded here as blocked: OstreeDeploy and ImageSwap delegate /etc merging to bootc switch, not to mergetc, so there was no caller for a mergetc cross-base extension. That is still true of mergetc — and it turned out to be the wrong question. The blocker was stated in terms of the call site; the inputs were never missing. bootc switch stages without rebooting, so afterwards the source's defaults (<booted>/usr/etc), the user's live /etc, and the target's defaults (<staged>/usr/etc) all still sit on disk beside the merge's own output (<staged>/etc).

So the policy landed as bootc-migrate-core::etc_conflict: a narrow post-merge reconciliation pass, not a second merge. It rewrites only the paths where all three inputs disagree — the conflict class the native merge cannot reason about, because within one base lineage "keep the user's value" is the right answer and across two it is not — and leaves every other path exactly as bootc switch produced it. Target defaults win; the displaced value is preserved as a .rebase-old sidecar (the same convention #15 introduced in mergetc); machine-describing paths and the identity DBs are reported but never replaced. This is the same "adjust the staged deployment before first boot" seam part 1's remap already uses.

What is not proven: neither part 1 nor part 2 executes in CI at all. Both are unit-tested (planning, exemptions, sidecar naming, report/JSON, and a collect→plan→apply round trip over real trees) and neither has run on a real cross-base system.

Why, precisely — this went through two wrong explanations before the code was actually read, so the reasoning is recorded rather than the conclusion alone:

  • is_cross_base (scan.rs) is lineage-aware: it returns false when either side's ID_LIKE contains the other's ID. CentOS declares ID_LIKE="rhel fedora", so a CentOS → Fedora re-base is same-lineage by design — see the cross_base_same_family_via_id_like_is_clean test. So "every cell is Fedora-family → Fedora-family" is substantively right, and Bluefin LTS being CentOS Stream 10-based does not by itself make a cell cross-base. (An earlier revision of this file claimed otherwise; that was wrong.)
  • Separately, the three bluefin:lts cells run E2E_MODE=composefs-migrate — the MVP binary, which merges via mergetc and has no is_cross_base gate at all.
  • And when a cell was actually built to exercise this (#187), it uncovered a third blocker that outranked both: inside the E2E guest the target-image scan could not reach ghcr.io, so gate_cross_base degraded to a no-op with only a warning and is_cross_base was never evaluated at all (#191).

That third blocker is now fixed, which moves the honest status but does not change the conclusion. #191 shipped on 2026-08-28: build_cross_base_plan returns a tri-state CrossBaseVerdict, and Unknown — the unreachable-scan case — is refused on the same terms as a known cross-base pair rather than waved through. The ostree-rebase cell asserts that refusal (either "Cross-base re-base detected" or "Cannot determine whether …" is accepted; proceeding unguarded fails the cell), and it then opts in with --accept-cross-base to continue, so the gate's wiring now has live coverage in CI.

What still has none is the thing the gate protects: the remap walk and the /etc reconciliation pass. Every pair in the matrix is same-lineage Fedora, so is_cross_base returns false whenever the scan succeeds, and the opt-in is the operator's rather than a real cross-base crossing. So the path is coverable now but still uncovered, and the open question is the one #187 was always about: which available image pair actually qualifies as cross-base under the ID_LIKE rule. Tracked as #187.

Related: #80 confirmed (via reading ostree's merge_configuration_from() source directly) that bootc switch's native merge does plain whole-file 3-way merge with no identity-DB (passwd/group/etc.) key-level reconciliation — the exact class of problem mergetc's union-merge exists to prevent. The etc_conflict pass deliberately does not close that gap: it holds the identity DBs exempt (it has no union-merge to rescue them either) and keeps the existing advisory warning. #80 still needs either an upstream ostree/bootc change or its own compensating logic.

Exit criteria (not yet met): fedora-family → centos-family E2E cell with a populated /var: correct ownership after reboot, report lists every renumbered account, .rebase-old sidecars present where defaults were taken.

M4 — Native store & the generation matrix (the #72 endgame) — not started beyond the feature flag​

#13's NativeStore (composefs/composefs-oci crates, no CLI shelling) exists behind the composefs-native feature flag and is off by default. The default path still probes host/target/builder for a legacy-CLI-capable bootc and pins quay.io/fedora/fedora-bootc:42 as a builder when none of the three has it — visible in every E2E run's Phase 2 log line. Store selection by target generation, and retiring the pinned legacy builder, have not been picked up.

Exit criteria (not yet met): migration succeeds with no legacy-CLI bootc anywhere (host, target, builder); BMC_CFS_BUILDER becomes a no-op escape hatch.

M5 — Desktop & UX (scenario E + human factors) — computable cores landed; interactive/live coverage is partial​

Three issues share the same shape: the pure/reusable core shipped first and is unit-tested. The interactive checklist in #15 is now driven through a full migration by the gating TUI cell. The gating OSTree re-base cell also executes #31's live NVRAM rename and snapshot restore against OVMF and asserts that efibootmgr -v is byte-identical afterwards. That evidence does not cover #65's unimplemented bootloader flip, and real-hardware validation is still advisable before treating firmware-specific behavior as universal.

  • #68 — DE config stash/restore (GNOME dconf/gnome-shell, KDE kdeglobals/plasma, COSMIC, niri, XFCE), a best-effort portable-preference extractor, and the pre-switch.d/post-switch.d hook contract are done, unit-tested, and exposed as bootc-rebase de-migrate stash|restore. Target-image DE detection now streams session files, session binaries, and any display-manager default session out of the registry (de_detect), the same decision function classifies the running host, and rebase --de-migrate (off by default) stashes every human account's outgoing DE config before staging and restores a matching stash on the way back. The decision logic is table-driven-tested offline; what has not run is a cross-DE E2E cell (Bluefin GNOME → an Aurora/KDE image) asserting the stash exists on a real system and survives the reboot — that and the portable subset actually being applied by a hook still need live validation.
  • #15 — the factory-vs-live /etc diff computation is done, exposed as bootc-migrate etc-drift (table or JSON). The interactive checklist UI (etc-drift --interactive, or --review-drift as "Phase 0.5" ahead of a live migration) and its wiring into Phase 4's merge decision (EtcDriftManifest / merge_etc_files_with_overrides, unit-tested) are now implemented. The checklist's terminal event loop — previously the one piece only manual/corral-VM validation could reach — is now covered twice over: headless state-machine + TestBackend render tests (drift_review::tests, and tui::tests for the migration wizard), and the tui-migrate E2E cell, which drives the checklist and then a full migration through the wizard on a pty inside the VM (tests/tui-e2e-driver.py; see docs/testing.md "TUI testing"). The cell is gating as of 2026-08-27 and publishes an asciicast/GIF/screenshot walkthrough artifact on every run. The full seven-cell matrix runs on GitHub-hosted runners (run 33071608765, all green).
  • #31 — the UEFI boot-entry audit (dead/generic-label/duplicate/firmware-managed classification) is done, read-only, exposed as bootc-rebase boot-entries. Interactive selection, live entry removal, and branding-rename are now implemented too, taken on with explicit sign-off on the NVRAM-mutation risk rather than deferred again. The split follows the mitigation below: the whole decision — which entries may be deleted or renamed, and which refusals stop a plan — is a pure, table-tested planner (boot_cleanup::plan), and the efibootmgr executor (boot_cleanup::live) only performs what the planner approved. Dry-run is the default; --apply requires a typed confirmation and writes a restorable NVRAM snapshot first; --undo replays it. The gating OSTree re-base cell now covers the create-before-delete rename and --undo against live OVMF variables, including a byte-for-byte restoration check. Entry deletion and the checklist's terminal event loop are not exercised by that flow and still need targeted VM or real-hardware validation.

Exit criteria (not yet met): bluefin↔aurora-style switch preserves user data untouched, stashes/restores DE state, swaps DE-scoped flatpaks on request; non-experts can read what will happen before it happens.

M6 — Package-host onboarding — detection and readiness only​

#370 extends the engine to regular rpm-, dpkg-, and pacman-managed installs. We use this order because package hosts have varied disk layouts and lack an OSTree factory /etc. Detection alone does not prove that a side-by-side install is safe.

The two read-only pieces have landed:

  • #371 detects Backend::Package, plans PackageInstall, and defaults its target backend to OSTree.
  • #372 reports boot mode, filesystems, LUKS/LVM, btrfs subvolumes, /boot, /home, free space, active package managers, SELinux, and target hardware.

Mutation and completion remain open. Tracked work includes:

  • #373: install engine.
  • #374: old-system rollback and undo.
  • #375: package-database /etc classification.
  • #376: home and data carry.
  • #377: software report.
  • #378: guided commit.
  • #379: first-boot verification.
  • #380: package-source E2E cells.

381 records that boundary and adds evidence rows. It adds an OSTree-only

dry-run entry to the TUI. Live confirmation and software reports need #373 and #377.

Exit criteria (not yet met):

  • The Fedora Cloud → fedora-bootc cell is green on main.
  • The TUI shows target-specific readiness and software reports before live confirmation.
  • The old OS boots after staging and after undo.
  • First-boot verification passes.
  • Guided commit removes only the verified old system.

Deferred extensions (raised on #30, not yet scoped into a milestone)​

Two directions came up in discussion on the generalize-into-a-re-base-engine RFC but never got a milestone or an issue. Recorded here so they aren't lost, not because either is imminent:

  • composefs→ostree (reverse backend switch) — implemented as Strategy::OstreeInstall (#260, crates/bootc-migrate-core/src/ostree_install.rs): the target image's own bootc install to-existing-root runs in a privileged container against the physical root (/sysroot), building an OSTree deployment beside the composefs one. The composefs ESP artifacts are hash-snapshotted before bootc's alongside mode empties the ESP, then restored and verified immediately after that destructive operation; /etc is 3-way merged (cross-family policy included) and /var copied into the stateroot. The cutover is an atomic journaled transaction (#387): "Linux Boot Manager" remains first as rollback, the target is armed once with BootNext, restart classifies observed ESP/NVRAM state and either resumes or repairs, and commit --apply makes the observed-good target persistent. Non-gating route and fault-injection cells exercise both interruption boundaries. undo remains the narrow path for a migration this tool itself performed.
  • Cross-family re-base (Fedora ↔ Ubuntu ↔ Arch) via bootc switch — cross_family.rs covers conversion and both switch routes. Target defaults win; source vendor-only files drop; machine state, user files and identity data stay; displaced edits use .rebase-old. OstreeDeploy applies the policy after finalization and schedules relabeling. ImageSwap preserves sysusers/ldconfig repair. The feature stays behind --accept-cross-base; cells are tracked by #259.

1.0 — Universal migrator​

All routes in the table implemented or explicitly refused with evidence; MVP binary either retired into bootc-rebase --target-backend composefs or kept as a thin alias; docs complete (architecture, generations, recovery, hooks). Version and deprecation policy published.

Dependency graph​

graph TD
M0["M0 MVP hardening — done"] --> M1["M1 same-backend engine — done"]
M1 --> M2["M2 #65 migrate-bootloader — live trial, resync, promote/undo landed"]
M1 --> M3P1["M3 #67 pt1 remap — done, wired into OstreeDeploy"]
M3P1 --> M3P2["M3 #67 pt2 /etc conflict policy — landed post-switch, no cross-base E2E cell"]
M1 --> GAP80["#80 identity-DB merge gap — confirmed, tracked separately"]
M1 --> M4["M4 NativeStore selection / retire legacy builder — not started"]
M1 --> M5A["M5 #68 DE stash/restore — detection+wiring landed, cross-DE E2E cell deferred"]
M1 --> M5B["M5 #15 etc-drift report + TUI + Phase 4 wiring — done, TUI unit-tested + tui-migrate E2E cell gating"]
M1 --> M5C["M5 #31 boot-entry audit + cleanup/branding — planner tested, NVRAM path needs a UEFI VM"]
M1 --> M6["M6 #370 package hosts — detection + readiness landed; install, reports, rollback and E2E open"]

Risks & standing mitigations​

  • Upstream drift is the norm, not the exception. bootc replaced its cfs CLI once mid-project; assume it will again. Mitigation: the generation matrix (probe, delegate, native), pinned-builder escape hatch, and the empirical harness in docs/cfs-cli-generations.md to re-verify fast.
  • MVP regression via shared code. Mitigation: MVP protection rule (frozen behavior, additive-only in core, probe-gated divergence), four untouchable E2E cells.
  • Boot-critical work can't be validated by this project's normal loop. Mitigation, applied consistently across M2 and M5: land the pure/testable core, stop before live NVRAM/ESP/interactive-UI mutation, document the exact remaining plan on the issue, and require explicit sign-off + a dedicated E2E cell before attempting it — rather than shipping unvalidated boot-path code just because the rest of a PR's CI run was green.
  • CI capacity. Runner starvation observed both 2026-07-19 (this repo's own concurrency groups) and 2026-07-24 (org-wide GitHub Actions queue congestion across most tuna-os repos simultaneously — not fixable from any one repo's side; just wait it out or check org Actions capacity/billing). Mitigation: everything is verified locally (or via a remote build host) before push; heavy validation designed as unit/loopback experiments where possible; E2E cells narrow and dispatchable individually.
  • Settings translation temptation (#68). Prior art is unanimous that GNOME↔KDE translation fails; the spec forbids it. Stash/restore only.

Decision log (summary — details in issues)​

  • Bootloader on ostree→ostree: migrate to systemd-boot when ready (#64)
  • UID/GID divergence: auto-remap with report (#67)
  • Cross-base /etc conflicts: target defaults win, user value kept as .rebase-old sidecar (#67, part 2 — landed as bootc-migrate-core::etc_conflict, see M3 above; not yet exercised by a cross-base E2E cell)
  • Store format is defined by the reader at boot (target image) — writer selection follows the target's generation (#13/#72)
  • XBOOTLDR GUID-retype: dead (sd-boot ≥258.2 requires vfat) — ESP-copy
    • resync instead (#65)
  • DE settings: stash/restore, never translate (#68)
  • Boot-critical or UI-only remaining scope gets a documented plan on its issue, not a best-effort implementation without validation (#65, #31, #15)
  • When boot-critical scope is taken on (#31's cleanup): split the decision from the execution, so every safety rule is a pure unit test and the executor holds no policy; back up before mutating; make the destructive step opt-in and human-gated; and never remove the path back to a bootable state (#31)