- resolve the NVIDIA branch and kernel-module flavor from PCI device IDs
(Turing+ open modules; Maxwell/Pascal/Volta proprietary, masked below 595;
Fermi/Kepler fall back to nouveau, reported; unknown IDs reported)
- capture pci_id/pci_address in the probe; drive VIDEO_CARDS, package.use,
package.accept_keywords, package.mask, modprobe.d and dracut from the plan,
preserving hybrid Intel/AMD
- add the nvidia_drivers install stage (after the kernel, before boot) that
emerges the driver, regenerates the initramfs, records /etc/phoebus/nvidia and
fails the install on failure; sys update rebuilds modules and the initramfs
- refuse AI changes to x11-drivers/nvidia-drivers
- live ISO: signed non-free components, stable-kernel pin (testing 7.2 breaks
the 550 driver), DKMS build for the exact kernel, gated pre-greetd load,
fail-safe entry, gpu-report, and a --verify-live-nvidia build gate that runs
on cached rootfs builds and fails on any missing component
- docs: ADR 0012, STATUS/CHANGELOG/HANDOFF, live README
- also carries the pending OpenCode model-variant/streaming changes and fixes
the GUI bridge install thread join that could hang
Update the changelog, status, handoff, verification register and VM boot
record with the 2026-09-22 graphical journey work: the fixed step sequence,
the interaction-defect fixes (readiness refresh, catalog sync/staleness, AI
revalidation, keyboard suggestion navigation, truthful keyboard-step label,
link vs service readiness, busy/failure handling), the new gui-install
harness mode, and the final-ISO results.
The gui (28/28) and opencode (18/18) journeys pass; gui-install passes every
journey assertion and starts a real install, but the four-hour harness timeout
fired during the kernel build (base packages ~116 min), so two completed
fresh-disk graphical installs + detached boot remain BLOCKED on host build
time, not on a code defect.
Rework the graphical installer into a fixed first-time-user sequence
(Welcome, Language and keyboard, Internet and readiness, AI connection,
Destination disk, Account and encryption, Review and confirmation,
Installation, Completion) with configuration, validation, plans, approval
and execution still authoritative in Rust.
Fix the audited interaction defects: readiness `Re-check` is a read-only
probe that no longer resets the step or clears the disk; model catalogs are
fetched after syncing the edited settings and credentials and stale replies
are ignored; editing a connection field invalidates the validated state;
suggestion lists support arrow/Enter/Escape with a visible highlight and
accessibility; the keyboard step labels the live-session limitation
truthfully; readiness separates the local link from reachable install
services; and busy states always clear, with the bridge failing outstanding
requests when the backend exits.
Add human-readable language/keyboard names, an explicit AI change review
(Accept/Discard), backend-fact review with edit links, real stage progress
with elapsed time and activity, and clearer failure and completion guidance.
Extend the VM harness with a `gui-install` mode that drives the whole
journey through the visible controls, plus regressions for readiness
refresh, disk survival, keyboard suggestion selection and AI invalidation.
Add Rust tests for readiness, the network probe, error kinds and labels.
A fully-validated OpenCode install was refused at the final step because
selecting a disk or setting the account/LUKS secrets cleared the AI
validation. Bind the validation to a digest of only the AI connection
settings: non-AI config changes preserve it, while an AI provider/model/
endpoint/server or AI-credential change invalidates it. Add regression
tests and password-length placeholder hints.
The region and keyboard fields were free-text, so users had to know zone
names and layout names from memory, and the AI model had to be typed.
Add an 'options' bridge op (live timezones/locales/keymaps) and an
'ai_catalog' op (the OpenCode provider/model catalog, or an
OpenAI-compatible /models list), and a searchable PhoebusCombo picker.
Also name provider/model in OpenCode errors instead of an opaque HTTP
code. Record the user-trial findings U1-U5.
Add an 'opencode' harness mode with a guest-loopback relay (the tunnel
shape), keyboard selection of the OpenCode chip, UI entry of the server/
provider/model (and a 0600 server password), and assertions on the
frontend's secret-free state (tested, validated, applied). Record the VM
evidence and the remaining not-run full-install gate.
The shipped DeepSeek models occasionally key a change or rationale wrongly.
Retry once with a corrective turn; the strict parser remains the final
authority and a retry that still fails is still rejected. Add scripted
transport tests for both the proposal and assistant paths.
Add an 'OpenCode agent' choice with server/password/provider/model fields,
a Test connection action, and a secret-free frontend state line the
harness asserts. Require exactly one installer session and assert the
current step and AI state instead of inferring from a screenshot. Add a
live-OpenCode acceptance make target and document the adapter, transport
policy, states and workflows (ADR 0011 updated).
Select the OpenCode adapter from canonical config in ai_check/ai_test;
add --ai-backend/--opencode-* flags to the installer and sys. Require a
validated interaction (not a provider name) for the mandatory-AI gate in
both the bridge and the CLI, invalidated by config, model or credential
changes. Tests cover the gate, invalidation and the OpenCode backend.
Integration tests drive the curl client against a local TCP server
(status preservation, chunked multibyte bodies, oversized bodies,
redirects, abrupt disconnects, authentication). The live tests obtain a
real model response when PHOEBUS_OPENCODE_LIVE_* is set and fail as
BLOCKED when a live run is required but not configured.
OpenCode is an agent server, not an OpenAI endpoint. Add a session/
message client over the established curl transport that implements
ModelTransport, so the strict proposal parser, facts and review are
reused. The probe enforces the loopback/TLS transport policy, pins the
resolved address, supports basic auth, checks HTTP status before parsing,
and reports truthful connection states (ready is reserved for a validated
response). Accept only completed assistant text parts and delete the
transient session on every path. Tests cover endpoint/policy, auth, a
misleading 500, credential non-leakage and the request shape.
Add ai.backend (direct|opencode) plus opencode_server/provider/model
fields, distinct from the direct provider. Require the change-object
shape (action+value) in the proposal prompt: the shipped models keyed
changes by the action name otherwise.
Update the changelog, status matrix, handoff and VM boot verification with
the 23/23 graphical journey, the hardened harness, the confirmed fixes and
the remaining blocked/unverified gates (full GUI install past AI, rollback,
installed desktop, physical second-computer OpenCode transport).
The tiled welcome window was not keyboard-focused, so the primary action
had no focus ring and Return did nothing; request activation and force
focus. Refuse Escape during the destructive phase and say the operation
cannot be cancelled. Show the AI connection guidance beside the controls.
Guard the fixes with live_iso tests.
Validate every writable drive (qcow2, inside PHOEBUS_VM_DISKS, no /dev/*,
block/char devices, raw images, escaping symlinks or outside backing
chains), give each run its own OVMF variable file, QMP socket and state
directory, and use QMP plus argument arrays instead of shell interpolation.
Return nonzero on any failed assertion, timeout, EOF or missing result and
record PASS/FAIL/BLOCKED/SKIPPED/NOT RUN. Add 16 unit tests, a report
generator and a findings register; make test runs the harness suite.
OpenCode is an agent application, not an OpenAI-compatible endpoint. Add
a typed probe over /global/health and /config/providers that distinguishes
disconnected, authentication required, incompatible, reachable-no-model and
ready, and never returns provider credentials. Expose it as
phoebus-installer --opencode-probe. Document the four components, host and
tunnelled remote workflows, and the Phoebus assistant boundary (ADR 0011).
Reject a symlinked path component while writing beneath TargetRoot
instead of following it, so a planted symlink cannot redirect an
installer write to the host. Cover escaping directories and leaves.
Accept only http(s) endpoints and refuse URLs that could inject curl
config directives. Restrict the transfer to http,https, keep redirects
disabled (so a header cannot follow one to another origin) and bound the
response with max-filesize. Add endpoint and transport regression tests.
Consume the approval before destructive work and refuse a second
start_install from a retry, a second frontend or a reconnect. Bound
request lines and secret values; reject deeply nested JSON. Add
regression tests for replay, oversized input and nesting.
probe_disks parsed lsblk human-readable sizes as bytes, so every disk
reported 0 B and the size-based identity check was blind; pass -b.
Re-probe the selected disk immediately before the destructive stage and
refuse if it disappeared, became protected (mounted/read-only/live) or
changed size/model since the review. Add regression tests for each case.
Add a branded live welcome window and a Quickshell/QML installer frontend
over a narrow, fail-closed JSON bridge (phoebus-installer --gui-bridge)
that reuses the same InstallConfig, plan, validation, authorization and
staged executor as the TUI/CLI. Add the Phoebus brand system (emblem,
wordmark, wallpaper, tokens), suppress DMS first-run dialogs on the live
medium, identify the live medium as protected, make the installer
shortcut live-only, and fix the installed boot default so systemd-boot
never selects the recovery entry.
Verified in a disposable graphical VM (welcome/installer journey,
duplicate prevention, keyboard navigation, blocked AI prerequisite) plus
a real non-interactive install and detached boot (INSTALL-RC=0,
NO_SECRET_LEAK, systemd running, 0 failed units, LUKS2, Btrfs, network).
- desktop.profile defaults to mango with Flatpak on; explicit none/minimal/dwl/
full still win. The plan shows the selected session.
- Installed login uses greetd + tuigreet (authentication required, Mango default,
dwl selectable); live autologin is a separate live-only configuration.
- installer/live/build.sh builds a rootless Debian testing rootfs with Mango
(pinned source build), DankMaterialShell (pinned SHA-256 release), Quickshell,
scenefx, Qt Wayland and greetd, then splices it into the base Debian-Live boot
equipment. Live autologin, an Install Phoebus launcher and a serial/VT
recovery console are included and the build fails clearly if a desktop
component is missing. Verified in a graphical seat-providing VM (Mango + DMS +
Quickshell; DMS panel captured).
- Shared desktop assets emitted by phoebus-installer --emit-asset, a graphical
VM driver (live-driver.py), and regression tests for defaults/overrides,
login, live/installed separation and the live ISO guards.
- Docs updated: ADR 0005/0009, STATUS, HANDOFF, CHANGELOG, ROADMAP, DEVELOPMENT
and VM-BOOT-VERIFICATION.
- Write guru.conf with auto-sync = no so the first emerge --sync (before
dev-vcs/git exists) only syncs the main tree; GURU is synced explicitly.
- In the Mango session, check dms.service is active after starting
mango-session.target and fall back to a direct 'dms run' if not.
- Mango + DankMaterialShell desktop option (desktop.profile = "mango") through
the GURU overlay with scoped ~amd64 keywords, a sealed session entry point and
a dwl fallback session (ADR 0009).
- Coherent rollback/recovery: coordinated @ + @var snapshots, fstab and BLS
root-subvolume rewriting, matching-kernel selection, snapshot recovery entry,
retention/space gates and sys undo (ADR 0007).
- Installed assistant (sys ask) with an OS-level systemd sandbox boundary and
fail-closed typed plans; resolved package effects bound to approval.
- Nix build isolation enforced (sandbox, no fallback) plus lifecycle/cleanup.
- Desktop helpers (menu/lock/idle), transaction log, release gates with atomic
publish and signing helpers, and a Debian-Live appliance VM harness.
Nix is on by default (NIX_ENABLED=no opts out); a desktop VM verified
nix-daemon active and 'nix shell nixpkgs#hello' -> Hello, world!. Create
/nix/var/nix/daemon-socket (the unit's ConditionPathIsReadWrite) and the
.wants links. finalize now runs sync before unmounting so the session script
and exported log are durable across a forced poweroff.
A binary ebuild is impossible (the upstream installer has no DESTDIR and runs
nix-store --load-db against the real root), so bootstrap the pinned,
SHA-256-verified official release with discrete commands in the target chroot
(no curl|sh, no sh -c): store copy, DB load, nixbld users, profile wiring and
nix-daemon units. Gated behind packages.nix (default off) until VM-verified.
Installing Phoebus now requires a connected model: the install gate refuses
with no provider unless the explicit break-glass --no-ai (alias --legacy, or
PHOEBUS_NO_AI=1) is used. Add a 'local' provider kind (Ollama/llama.cpp/
LM Studio) with an optional API key and --ai-endpoint override, so a user can
run a model they downloaded; no model is shipped. Desktop installs now include
xdg-desktop-portal + gtk/wlr backends for Flatpak and Wayland apps.
The no-model fallback is the Phoebus deterministic CLI (sys) running the same
typed actions the assistant emits, exposed as an explicit non-AI (legacy) mode.
Raw emerge/nix/flatpak remain a last resort rather than the intended front-end.
Installing Phoebus now requires a connected model (remote provider or a
user-downloaded local model; none is shipped). The installed system is AI-first
and uses a connected model proactively. It is never required to boot, run or
recover: the deterministic CLI (sys, emerge, nix, flatpak) remains as a
break-glass fallback, and no model ever holds privilege.
The AI assistant stays a defining, always-present subsystem, but Phoebus
ships no model. The user may download a local model or configure a remote
provider; with no model the assistant still explains, plans and applies
deterministic changes and the CLI remains available. Installation requires the
network; the installed system must boot, run and recover offline.
sys install <name>, sys remove <name> and sys search <query> are
deterministic and model-free: Portage resolves a bare name or a full atom.
Adds the remove_package and search_packages typed actions (remove is High risk
and gets the protective snapshot; search is read-only). This is the
old-fashioned fallback so the system stays fully administrable with no LLM.
AI and Nix are now always-present parts of Phoebus (not optional add-ons):
the assistant ships with a local model so it works offline and in recovery,
and Nix is a second package lane shipped and updated through Portage.
Installation needs the network; the installed system must boot, run and
recover without it. The assistant is the preferred front-end, never a gate:
the deterministic CLI (sys, emerge, nix, flatpak) is always available. The
model still never holds privilege. Records ADR 0006 (package lanes).
The appliance ISO firmware-boots, installs (INSTALL_RC=0, NO_SECRET_LEAK) and
the installed disk boots detached as Gentoo-only (Portage, no apt, systemd
running, 0 failed units). Two anomalies are recorded rather than hidden: the
target hostname and a first-boot DHCP timing observation.
Add installer/live/build.sh: overlay the static installer, the OpenCode
manifest and the Debian tools the standard live image lacks (cryptsetup,
btrfs-progs, gdisk, dosfstools, parted, util-linux-extra) onto a Debian-Live
stable base, patch the boot config for a serial console, and preserve the
hybrid BIOS+UEFI boot. The live squashfs is never repacked and libc is never
bundled, so the appliance is unchanged and the target remains Gentoo-only.
Records ADR 0005 and the implementation plan.
Deterministic review now rejects encryption=none, filesystem=ext4 and the
aggressive kernel profile, so the LUKS2/Btrfs/stability posture cannot be
traded for convenience. A live provider that returns an invalid proposal is
reported and the installer continues with the proven defaults (the explicit
--ai-proposal-file test path still fails loudly).
emerge --buildpkgonly refuses while dependencies still need merging, which
is normal for an out-of-date system. Record that as Ready with the fetched
distfiles and the Portage reason instead of a hard Failure, and include the
redacted stderr tail in the state note. The VM harness check timeout is now
configurable.