Skip to content

fix(iso): get the LUKS end-to-end test green on all four flavors - #94

Merged
castrojo merged 7 commits into
mainfrom
claude/great-allen-pt5nwv
Sep 18, 2026
Merged

castrojo merged 7 commits into
mainfrom
claude/great-allen-pt5nwv

Conversation

@hanthor

@hanthor hanthor commented Sep 17, 2026 •

Copy link
Copy Markdown
Member

Why

The six-phase LUKS end-to-end test has been red on every run since 2026-09-16. Run 35326466386 (on 68d0432, after the GSConnect fix in #89) shows the two remaining causes, one per pair of flavors:

utah, utah-nvidia clear all six phases and die on the last extension assertion:

Extension search-light@icedman.github.com: TypeError: this.set_shader_source is not a function
FAIL: extension search-light@icedman.github.com did not reach ACTIVE on GNOME 51 (state=ERROR)

utah-gaming, utah-nvidia-gaming die in phase 2, in the live session, on a black screen:

graphical.target: active
gdm.service: active
FAIL: no gnome-shell running for liveuser

The serial logs explain the second. main's kernel prints [drm] Initialized simpledrm 1.0.0 for simple-framebuffer.0; the OGC kernel 7.1.8-ogc1 prints no [drm] Initialized line at all. install-ogc-kernel.sh starts from x86_64 defconfig, which enables DRM and i915 only, so on any GPU that is not an Intel iGPU, QEMU's included, GDM starts and mutter has no KMS device to open.

What changed

  1. Search Light is dropped (d862f13). GNOME 51 removed the Clutter shader API its blur effect calls. The submodule, its schema compile, its entries in the source and installed extension checks, and its uuid in the e2e default list go. Common still ships 05-bluefin-searchlight-extension (only the extension's own settings), so the Containerfile removes it after copying Common in and the desktop contract no longer lists it. just check reports eight extensions declaring GNOME 51.
  2. The OGC kernel gets a KMS display (a5d9c17). SYSFB_SIMPLEFB and DRM_SIMPLEDRM turn the firmware framebuffer into a KMS device on any hardware; DRM_VIRTIO_GPU and DRM_BOCHS as modules cover VMs. DRM_SIMPLEDRM joins required_config, checked after olddefconfig, after install and on cache extraction. Editing the script changes the kernel-cache tag, so CI rebuilds the cache first. Native amdgpu/xe stay out until linux-firmware is in the image (parity: bare-metal WiFi dead, linux-firmware not in install set #97).
  3. The e2e test reports every failing extension, not just the first, so the next layered break is found in one run rather than two. Unreadable state stays fatal; screenshots are named per extension.

The learning is recorded in docs/skills/kernel-cache.md and docs/skills/desktop-contract.md.

Verification

  • just check equivalents locally: desktop contract syntax valid, source-mode extension contract passes with 8 extensions, bash -n clean, 65 unit tests pass.
  • The real proof is the next post-testing-e2e run on testing after this merges: all four flavors should reach the fastfetch screenshot.

🤖 Generated with Claude Code

https://claude.ai/code/session_01W6BYLUD6KBzF8rBFXKjP7M

The six-phase LUKS end-to-end test has never passed in CI -- seven runs,
every one red -- and run 35159067119 shows it clears all six phases and dies
on one assertion:

    extension gsconnect@andyholmes.github.io: state=ERROR
    gnome-shell[2744]: Extension gsconnect@andyholmes.github.io:
      Error: Cannot inherit from a final type
    FAIL: extension gsconnect@andyholmes.github.io did not reach ACTIVE

That is the GNOME 51 break #88 describes, and PR #89 fixes it. The problem
this commit addresses is what happens next: the loop called fail() on the
first bad extension, so search-light was never evaluated. Its guard (PR #46,
open since 2026-09-06) is also absent from main, so it is very likely broken
too -- and the way the test is written, the only way to find out is to fix
GSConnect, wait another ~35 minutes, and be told about the second one.

This is the same shape as the failures 718d4e4 and 60e6a6e kept hitting in
the factory: each defect hidden behind the one before it, one full run per
layer. A check whose job is to report what is broken should report all of it.

The loop now records each failure and continues, prints them together, and
fails once with the count. Unreadable state stays fatal for the reason the
existing comment gives -- a check that cannot run reads as green, and did
exactly that once. Per-extension screenshot names so two failures in one run
do not overwrite each other's evidence.

Nothing is weakened: every case that failed before still fails. Verified by
splicing the real block into a harness with stubbed helpers and driving all
four states -- both ACTIVE exits 0; both ERROR reports both and exits 1; one
ERROR reports that one and exits 1; unreadable reports both and exits 1.
23 unit tests pass.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W6BYLUD6KBzF8rBFXKjP7M
@hanthor

hanthor commented Sep 17, 2026

Copy link
Copy Markdown
Member Author

The red contract check here is not this PR's, and it will fail every open PR until packages/bluefin.toml is synced.

just check passes completely — desktop contract, GNOME extension contract, download integrity, 73 parity packages, workflow outputs, skill index, 23 unit tests. The failure is the next step, just check-parity, which curls Bluefin's build_files/packages/base.toml live and diffs it against the committed copy:

packages/bluefin.toml has drifted from projectbluefin/bluefin
error: recipe `check-parity` failed with exit code 1

Verified against main, not inferred from the job. My diff touches one file, iso/scripts/luks-e2e.sh, and the identical drift reproduces against main's own copy:

$ git show origin/main:packages/bluefin.toml > /tmp/main-bluefin.toml
$ curl -fsSL -o /tmp/upstream.toml \
    https://raw.githubusercontent.com/projectbluefin/bluefin/main/build_files/packages/base.toml
$ diff -u /tmp/upstream.toml /tmp/main-bluefin.toml
+    "firefox",            # ours has it in [fedora]; upstream moved it
-    "libgda",             # upstream added (Copyous runtime deps, bluefin#1210)
-    "libgda-sqlite",
-    "ppp",                # upstream added
-    "firefox",            # upstream now lists it lower down

The contract job last passed at 2026-09-16T22:27Z (run 35157737556), so upstream moved some time after that. Because the check compares against a mutable upstream branch, it goes red on its own schedule rather than on any change here.

No fix PR exists to port, and I'm deliberately not syncing the file in this PR. The drift isn't mechanical: it re-adds firefox to the parity list, and that's the one name Utah deliberately does not ship as an RPM — #34 moved it to the org.mozilla.firefox Flatpak and into [unavailable], and the factory dropped the recipe. Landing that inside an unrelated E2E test change would bury a contested parity decision in a commit nobody would look for it in.

The proposed patch, for whoever picks it up:

--- a/packages/bluefin.toml
+++ b/packages/bluefin.toml
@@ [fedora]
-    "firefox",
+    "libgda",
+    "libgda-sqlite",
+    "ppp",

with firefox staying out and the existing [unavailable] rationale left in place — i.e. sync the three genuinely new upstream names, and keep the deliberate Firefox divergence documented rather than silently re-adopting it. That belongs in its own PR so the Firefox call is reviewed as a Firefox call. Related: #60 proposes replacing this whole comparison with one derived from Bluefin's effective transactions, which would make this class of breakage less frequent.

Not re-running the job: this is deterministic, not a flake, and re-running against a still-drifted upstream would just fail again.


Generated by Claude Code

@hanthor hanthor left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is right, and I could not find a way to make it lose a failure it used to catch. This would be an approval, but GitHub will not let the review be submitted as one against your own PR (Can not approve your own pull request), so it is filed as a comment — treat the verdict as approve.

What I verified

Base gate first, at the real merge-base (56185bb, which is also current origin/main):

$ git merge-base origin/main pr94
56185bb719ba6e684e210cf8d82a5bb2cd3b29a3
$ just check   # on 56185bb
Ran 23 tests in 0.454s
OK
exit=0

Then the PR head (d054dd6):

$ just check   # on d054dd6
Ran 23 tests in 0.594s
OK
exit=0
$ bash -n iso/scripts/luks-e2e.sh   # exit 0

So the gate is green on both sides and the 23 unit tests in your description is the pre-existing suite, unchanged — the diff is one file, iso/scripts/luks-e2e.sh, exactly as claimed. Justfile:71 already runs bash -n iso/scripts/luks-e2e.sh, so the syntax side of this is covered by just check.

The two greps in your description hold on the PR tree:

$ grep -n "DBusImplementation" scripts/build-gnome-extensions.sh ; echo rc=$?
rc=1
$ grep -n "set_shader_source"  scripts/build-gnome-extensions.sh ; echo rc=$?
rc=1

Neither guard is present, so your reasoning that search-light is very likely the next layer is sound rather than speculative.

Reproducing your four-state table

I did not take the table on faith. I spliced the live block out of the PR tree and drove it against stubbed ssh_target / shot / fail:

$ sed -n '/^EXT_CHECK=/,/^fi$/p' iso/scripts/luks-e2e.sh > block.sh   # 48 lines

with ssh_target synthesising State: <value> per uuid. All four of your rows reproduce, plus four I added:

states result
both ACTIVE exit 0
both ENABLED exit 0
both ERROR both reported, FAIL: 2 extension(s), exit 1
gsconnect ERROR / search-light ACTIVE only gsconnect reported, FAIL: 1 extension(s), exit 1
gsconnect ACTIVE / search-light ERROR only search-light reported, exit 1
both unreadable both reported, exit 1
gsconnect unreadable / search-light ACTIVE gsconnect reported, exit 1
UTAH_E2E_EXTENSIONS= block skipped, exit 0

Unreadable state really does stay fatal — it is collected rather than immediate, but nothing lets it reach the end of the script. That was the case I was most worried about, given the comment about a check that once read as green while asserting nothing.

Two smaller things I checked because they are the usual way an accumulate-then-report rewrite goes wrong:

  • ext_failures=() followed by (( ${#ext_failures[@]} > 0 )) is safe under the script's set -euo pipefail — ${#arr[@]} on an empty array does not trip set -u, so the all-green path does not die on the guard itself.
  • the screenshot rename is not referenced anywhere else. grep -rn 'installed-ext' . returns only the two lines this PR writes, and post-testing-e2e.yml:134 uploads screenshots/ as a directory glob, so per-extension names land in the artifact without a workflow edit.

Relationship to #89 — no conflict

$ git merge-tree 56185bb pr94 pr89 | grep -c '<<<<'
0

They are disjoint (#89 does not touch luks-e2e.sh) and semantically complementary: once #89's try/catch degrades the clipboard portal to an inert object, GSConnect reaches ACTIVE and this loop simply carries on to search-light instead of stopping. I also went looking for #89's specific weakness here — the vacuous node --check on a .js file — and it has no analogue in this PR, because this PR adds no static assertion at all.

Two nits, neither blocking

The summary line is slightly wrong for one of the four cases. fail "${#ext_failures[@]} extension(s) did not reach ACTIVE on GNOME 51" is what an unreadable state prints too, and "did not reach ACTIVE" is not what happened there — the state could not be read, which the older code was careful to word differently. The per-failure FAILED: lines above it are accurate, so the evidence is there, but the headline a reader greps for is not. Something like N extension(s) failed the GNOME 51 load check would cover both shapes.

Worst-case wall clock roughly doubles in the unreadable case. Each extension polls seq 1 24 × sleep 5, so two unreadable extensions now spend ~4 minutes where the old code spent ~2 before exiting. That is the price of the thing you are buying and it is bounded, but it is a real change to the tail of a 35-minute test.

One heads-up outside this diff

Your description leans on #89 landing next, so it is worth knowing that #89 is not currently based on main:

$ git merge-base origin/main pr89
3162105a6edeb108f581df4ca47bb37602b41fb2
$ git log --oneline -1 --diff-filter=A origin/main -- iso/scripts/live-kernel.py
56185bb fix: resolve OGC kernels when assembling live ISOs (#87)

#89 branched before #87, and its diff against main deletes iso/scripts/live-kernel.py and reverts build-iso.sh:67 to the plain usr/lib/modules/.../vmlinuz path. Merged as-is it would take #87 back out. That is #89's problem to fix by rebasing, not yours — but since this PR's value depends on #89 landing cleanly first, better said now than after a 35-minute run.

mergeable_state is blocked (from the per-PR get, which the list endpoint omits), so this still needs whatever branch protection is holding it; nothing in the diff is what is blocking it.


Generated by Claude Code

@hanthor

hanthor commented Sep 17, 2026

Copy link
Copy Markdown
Member Author

Correcting the last section of my review above — the "#89 would revert #87" heads-up is wrong. Please disregard it; #89 does not need a rebase on that account.

That claim came from a two-dot diff (git diff main pr89), which shows everything main has that the PR branch doesn't — including #87, which landed after #89 branched. That is not what merging does. The three-dot diff is what the PR actually changes:

$ git merge-base origin/main pr89
3162105a6edeb108f581df4ca47bb37602b41fb2

$ git diff --name-status origin/main pr89          # two-dot — misleading
...
M	iso/scripts/build-iso.sh
D	iso/scripts/live-kernel.py
D	tests/test_live_kernel.py

$ git diff --name-status $(git merge-base origin/main pr89) pr89    # three-dot — the real change set
M	docs/skills/desktop-contract.md
M	docs/skills/index.json
M	docs/skills/index.md
M	scripts/build-gnome-extensions.sh
A	tests/test_gnome_extensions.py

Five files, none of them under iso/. GitHub's own file list for #89 agrees. And the merge preview keeps #87 intact:

$ T=$(git merge-tree --write-tree origin/main pr89)
$ git ls-tree -r $T --name-only | grep -c iso/scripts/live-kernel.py
1

Merges clean, live-kernel.py survives. A three-way merge only applies what #89 changed relative to its own base, so #87 is not touched.

The rest of the review stands — the state-table reproduction, the set -u check, the screenshot-rename grep, the merge-tree ... | grep -c '<<<<' → 0 disjointness with #89, and the two nits. This was my error in one section, not a problem with your PR.


Generated by Claude Code

@hanthor

hanthor commented Sep 18, 2026

Copy link
Copy Markdown
Member Author

@castrojo this one is waiting on #108 (or #110) to land first — its only red check is the inherited check-parity failure from main. Once that merges I'll bring main in and it should go green; then it needs the same last-pusher approval as the others.


Generated by Claude Code

hanthor and others added 5 commits September 18, 2026 14:21
The ISO end-to-end test has failed on utah and utah-nvidia since GSConnect
was fixed, on the next extension in line:

    Extension search-light@icedman.github.com:
      TypeError: this.set_shader_source is not a function
    FAIL: extension search-light@icedman.github.com did not reach ACTIVE
      on GNOME 51 (state=ERROR)

GNOME 51 removed the Clutter shader API the extension's blur effect calls.
Rather than carry a fork of the effect, the extension goes: the submodule,
its schema compile in build-gnome-extensions.sh, its entry in the source
and installed extension checks, and its uuid in the e2e default list.

Common still ships /etc/dconf/db/distro.d/05-bluefin-searchlight-extension,
which holds only the extension's own settings (shortcut, colours, size), so
the Containerfile removes it after copying Common in and the desktop
contract no longer lists it. Nothing enables the extension anywhere in this
tree, so no enabled-extensions key changes.

just check's source-mode extension contract now reports eight extensions
declaring GNOME 51.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W6BYLUD6KBzF8rBFXKjP7M
The gaming and nvidia-gaming ISO end-to-end runs fail earlier than the
extension check, in the live session:

    graphical.target: active
    gdm.service: active
    FAIL: no gnome-shell running for liveuser

with a black screenshot. The two serial logs explain it. main's kernel
prints

    [drm] Initialized simpledrm 1.0.0 for simple-framebuffer.0 on minor 0

and the OGC kernel (7.1.8-ogc1) prints no [drm] Initialized line at all.
install-ogc-kernel.sh starts from x86_64 defconfig, which enables DRM and
i915 only, so on any GPU that is not an Intel iGPU, QEMU's stdvga included,
the kernel has no KMS device. GDM starts, mutter has nothing to open, and
no shell ever runs. The installed system would behave the same way.

The script now enables SYSFB_SIMPLEFB and DRM_SIMPLEDRM, which turn the
EFI/VESA framebuffer into an unaccelerated KMS device on any hardware, and
DRM_VIRTIO_GPU and DRM_BOCHS as modules for VMs. DRM_SIMPLEDRM joins the
required_config list checked after olddefconfig, after install and on
cache extraction, with a require_config gate that names it if olddefconfig
drops it. Editing the script changes the kernel-cache tag, so CI rebuilds
the cache before the next gaming image.

Native drivers for real gaming GPUs (amdgpu, xe) are left out on purpose:
they need linux-firmware in the image, which #97 records as missing. The
kernel-cache skill records the diagnosis and where to look first.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W6BYLUD6KBzF8rBFXKjP7M
@hanthor hanthor changed the title test(e2e): report every failing extension, not just the first fix(iso): get the LUKS end-to-end test green on all four flavors Sep 18, 2026
@hanthor

hanthor commented Sep 18, 2026

Copy link
Copy Markdown
Member Author

@castrojo this PR now carries the two fixes that make the ISO end-to-end test pass on all four flavors, not just the reporting change: Search Light is dropped (GNOME 51 removed the shader API it calls) and the OGC kernel gains a simpledrm KMS device (the gaming flavors were booting to a black screen because x86_64 defconfig ships i915 and nothing else). Title and body are rewritten to match. It needs an approval from someone other than hanthor, who is the last pusher.


Generated by Claude Code

castrojo pushed a commit that referenced this pull request Sep 18, 2026
* fix: restore the bluefin.toml parity contract

packages/bluefin.toml is meant to be a byte-for-byte copy of upstream
Bluefin's build_files/packages/base.toml. utah.toml says so in as many words
-- "it is the parity contract, and any drift there is a parity bug" -- and
just check-parity exists to make that loud. It has drifted, so that check is
red on main today and every open pull request inherits the failure, #94 and
#105 among them.

The drift is four hunks: firefox moved into [fedora], and libgda,
libgda-sqlite and ppp missing.

Syncing the file alone would have been worse than the failure it fixes. All
three missing names were measured against both repositories the image
actually enables -- the digest-pinned factory metadata layer from
Containerfile, and public-hummingbird-x86_64-rpms:

    libgda           absent from both
    libgda-sqlite    absent from both
    ppp              absent from both, and nothing Provides it

So a bare sync hands three unresolvable names to the install transaction and
turns a failing lint into a failing image build. The sync and the
[unavailable] entries have to land together, which is why they are one
commit.

Each entry carries a tracking issue, as that list requires:

  libgda, libgda-sqlite (#106)
    Upstream added both for the maintained Copyous extension
    (bluefin#1210). Hummingbird carries no package whose name contains gda
    at all. One factory recipe covers both, because libgda-sqlite is a
    subpackage of the libgda SRPM rather than its own source -- Fedora's
    libgda.spec declares %package sqlite alongside ui, tools, sqlcipher,
    mysql and postgres.

  ppp (#107)
    Worth more than dial-up trivia. Hummingbird ships NetworkManager-ppp,
    both 1.58.0-1.hum1 and 1.58.1-1.hum1, and its metadata carries
    Requires: ppp. With ppp nowhere in the enabled set that plugin is
    uninstallable from Hummingbird's own repository, so NetworkManager's
    PPPoE and PPTP paths cannot be installed at all. It is the same shape as
    avahi (#104) and nautilus (#100): a plugin shipped without the daemon
    behind it.

firefox needed no handling. It is already in [unavailable], and contract()
filters that set out of every section, so moving it between sections cannot
put it back in the image -- which matters, because Utah ships
org.mozilla.firefox as a Flatpak and re-adding the RPM would regress that
decision.

This change installs nothing new and removes nothing. Proven rather than
argued: contract() over main's two files and over this branch's both return
the same 74 packages, with no additions and no removals, and libgda,
libgda-sqlite, ppp and firefox are each absent from the resulting set.

23 tests pass, 45 subtests pass. The parity diff against upstream is now
empty, the RPM contract check reports 58 Bluefin, 10 GNOME and 5 service
packages, and install-packages.py's own guard confirms no [unavailable]
package appears in the install set.

Noted but deliberately not changed: evolution-ews-core, fish and zsh predate
this and carry no issue reference, so the "every entry MUST carry a tracking
issue" rule is not currently enforced by anything. Worth a guard, but not in
a commit whose job is restoring the contract.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W6BYLUD6KBzF8rBFXKjP7M

* fix(e2e): capture the fastfetch shot at 1080p, with its logo

The screenshot on the repository front page is the project's end-to-end
evidence, and it was about to lose Utah's branding. #84 added

    fastfetch --logo none

to the session bashrc, which suppresses the logo outright. The committed
screenshot predates that change and still has the Utahraptor in it, so the
next passing run would have quietly replaced it with a logo-less one.

The workaround was reasonable, because the logo genuinely did not fit.
fastfetch needs about 90 columns with it -- roughly 33 for the logo column
plus a ~57 character detail line -- and Ghostty opens at 80x24, so lines like

    KVM/QEMU Standard PC (Q35 + ICH9, 2009) (pc-q35-11.1)

wrapped to column 0 and pushed fastfetch's own header off the top of the
window. The committed screenshot shows exactly that: its first visible line is
the tail of a wrapped one, and the OS identity line is scrolled away.

Columns are the lever, and point size on its own is not. Ghostty sizes its
window in cells, so a smaller font yields a smaller window with the same 80
columns and the same wrap -- it would have traded the wrap for worse OCR and
fixed nothing. What buys the columns is the display: both VMs now ask for
1920x1080 instead of the 1280x800 the machine default gave, and at that size a
110-column window at 12pt is about 880px of 1920. The logo fits with room to
spare and the font goes UP from the default rather than down, which helps the
tesseract gate that reads UTAH-E2E-FASTFETCH and Kernel out of the image.

xres/yres are the EDID preferred-mode hints QEMU passes to the guest. -vga
none replaces the machine default with the same stdvga device rather than a
different adapter, so the graphics path is unchanged -- this is a mode hint,
not a driver swap. Applied to the live VM as well as the installed one, since
live-desktop.png is published in the same record.

Ghostty is a flatpak here, and which config path it honours depends on whether
the sandbox exposes xdg-config or keeps its own per-app directory, so the
geometry is written to both ~/.config/ghostty/config and
~/.var/app/<app>/config/ghostty/config rather than guessing. Everything is
overridable: UTAH_E2E_DISPLAY_WIDTH, UTAH_E2E_DISPLAY_HEIGHT,
UTAH_E2E_FONT_SIZE, UTAH_E2E_COLUMNS, UTAH_E2E_ROWS.

Verified what can be verified without a hypervisor. The remote script was
rendered through a stub ssh_target and executed against a throwaway HOME: it
writes both ghostty configs with the expected three keys, writes the autostart
entry, and leaves `fastfetch && echo UTAH-E2E-FASTFETCH` in bashrc with no
--logo none. VGA_ARGS expands to exactly four argv entries, so no word
splitting. All shell scripts parse; 23 tests and 45 subtests pass.

NOT verified here: that QEMU accepts VGA.xres/yres in this build, and that the
guest picks 1920x1080 from the EDID hint. There is no hypervisor in this
environment. Both fail loudly rather than silently if wrong -- a rejected
device property aborts the VM at startup, and a display that stayed at
1280x800 would still wrap and fail the OCR gate.

Separately noted, not fixed: docs/verification/screenshots/installed-desktop.png
and installed-fastfetch.png are currently byte-identical (cd6577d6...). The
desktop shot is taken after the terminal autostart has already mapped
fastfetch, so two named checkpoints record one frame. That is a question about
what installed-desktop is meant to prove, not about font size, so it is left
alone here.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W6BYLUD6KBzF8rBFXKjP7M

---------

Co-authored-by: Claude <noreply@anthropic.com>

@castrojo castrojo left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Reviewed with a focus on the two things that could go wrong quietly here: the vendored submodule and the kernel cache.

Search Light removal is done properly. It goes through .gitmodules and the gitlink together — git diff --submodule=log shows the e1d9fff gitlink and its URL/branch stanza both deleted, with no submodule content edited and no orphaned pin. That matters because AGENTS.md puts those submodules off-limits; removing an extension GNOME 51 cannot load is legitimate, but only if it goes through .gitmodules rather than poking at vendored trees. After git submodule update --init --recursive, verify-gnome-extensions.py --source reports GNOME extension contract passed: 8 extensions declare GNOME 51 and exits 0, so the contract follows the removal instead of breaking on it.

The kernel cache invalidation is real and intended. Of the four inputs that hash into the cache tag, only scripts/install-ogc-kernel.sh changes; Containerfile.kernel's BASE_IMAGE line, install-nvidia.sh, and packages/hummingbird.repo are untouched. scripts/kernel-cache-tag.sh prints b547145c52fa7afe on main and 9d502d2049a1e93e here — so the ~45-minute rebuild is incurred deliberately and not silently bypassed, which is the correct behaviour when you change what the kernel is built with. That is also why this PR's CI is slow; that is the cost, not a problem.

The fix itself. Enabling built-in SYSFB_SIMPLEFB/DRM_SIMPLEDRM plus DRM_VIRTIO_GPU/DRM_BOCHS and asserting them across the compile, install, and cache paths is the right shape: the assertion is what stops a future config change from silently producing a kernel with no display in a VM, which is exactly the failure this is fixing for the gaming flavors.

Accumulating extension failures instead of stopping at the first, with per-extension screenshots, is a straightforward diagnostics improvement — the previous behaviour hid every failure after the first, which is the worst property for an e2e run you wait 180 minutes for.

Also checked: no literal flavor names introduced (just check enforces that), packages/bluefin.toml untouched, generated skill catalog current (generate_skill_index.py --check exits 0; both skill docs changed body text only, not front matter), and python3 -m unittest tests.test_iso_ci passes 10 tests.

Caveat on evidence, stated plainly: the contract and preflight checks passed and the main image built, but the four-flavor LUKS result this PR is titled after is not observable from a pull_request run — dispatch-iso only fires after a testing build. So I am approving on the mechanism being correct and asserted, not on having watched all four flavors go green. Watch the first post-merge Post-Testing E2E.

Approving.

@castrojo
castrojo merged commit b162d96 into main Sep 18, 2026
14 checks passed
@castrojo
castrojo deleted the claude/great-allen-pt5nwv branch September 18, 2026 16:54

@hivecommons-hive hivecommons-hive Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Reviewed — no findings from this perspective.

— hive: agent=reviewer backend=copilot model=claude-fable-5

castrojo pushed a commit that referenced this pull request Sep 21, 2026
…SO (#162)

* Give the OGC kernel btrfs, and stop the fastfetch gate reading the screen edge

The first ISO end-to-end run after #94 got much further and split the four
flavors into two different failures. Both are fixed here.

gaming and nvidia-gaming now boot the live session -- the simpledrm fix
worked -- and die one phase later, in the installer:

    mkfs.btrfs -f -L root /dev/mapper/fisherman-root          (ok)
    mount -t btrfs /dev/mapper/fisherman-root /mnt/...
    mount: unknown filesystem type 'btrfs'
    fisherman: fatal: mounting root: ... exit status 32

Utah installs to btrfs on LUKS and x86_64 defconfig has no BTRFS_FS, so
the OGC kernel can format a root volume and then not mount it. The gaming
ISO boots the OGC kernel, so this hits the installer as well as the
installed system. required_config was written for the live-boot and LUKS
paths because no gaming flavor had ever reached the phase that formats a
disk; it now also names the filesystem the disk is made of. BTRFS_FS
selects its own crc, zlib, lzo, zstd, raid6 and xor helpers, so
olddefconfig pulls the rest in, and a require_config gate catches it if
it does not. This rotates the kernel-cache tag, so the next gaming build
compiles once.

utah and nvidia go all the way to phase 6 -- installed, unlocked, logged
in, GSConnect ACTIVE, wayland session -- and fail on the screenshot gate
alone. The screenshot is perfect; the gate could not read it. From the
run's own fastfetch-ocr.txt:

    TAH-E2E-FASTFETCH
    Utah (Version: testing-20260918-2216657)
    Linux 7.1.8-100.fc43.x86_64

The marker lost its first glyph, and every fastfetch label -- OS:,
Kernel:, Uptime: -- is missing while their values read perfectly. The
leftmost column of the terminal is what tesseract does not read, so both
assertions were aimed at exactly the text that cannot survive. The marker
is now framed as ===UTAH-E2E-FASTFETCH=== so it does not start at column
0, and the second assertion matches the kernel version fastfetch prints
rather than the label in front of it. Both still prove what they did
before: an interactive shell ran our marker, and fastfetch rendered real
system information beside it. Nothing is skipped and the gate stays
fatal.

Checked against that captured OCR text rather than assumed: the old
patterns both fail on it, the new version pattern passes on it unchanged,
and the framed marker passes once the frame is there.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W6BYLUD6KBzF8rBFXKjP7M

* Retry Flathub pulls, so one timed-out object cannot fail a flavor's ISO

Today's post-testing-e2e run (35432516418) is the first to get two of the
four flavors all the way through: utah and utah-nvidia installed to LUKS,
unlocked, logged in and passed the fastfetch gate. The two failures are
unrelated to each other, and only one of them is a bug in the image.

utah-gaming died composing its ISO, three minutes and forty-four seconds in:

    Error: Failed to install org.mozilla.firefox: While pulling
    app/org.mozilla.firefox/x86_64/stable from remote flathub: While fetching
    https://dl.flathub.org/repo/objects/88/680ed7...commitmeta:
    [28] Timeout was reached

Nothing was wrong with the image. The Flathub pull is the largest network
operation in the whole build -- Firefox alone is around 200 MB -- and a single
timed-out object took down the flavor's entire end-to-end run. Meanwhile the
curl fetching the installer bundle, twenty lines above, has had --retry 3 all
along for exactly this reason.

So the three flatpak installs now go through a retry_flatpak helper: three
attempts, 15s then 30s apart. Retrying is cheap because flatpak resumes a
partial pull from the local repository, so only what is still missing is
re-fetched, and every install already passes --or-update, which makes a retry
a no-op for refs that completed.

It does not swallow failures. A repository that is genuinely gone still fails
the build after three attempts, and there is a test for that as well as for
the recovery case -- both drive the real helper with a stub flatpak rather
than asserting on the script's text. A third test walks every install line in
the script and fails if one bypasses the retry, so the next install added
cannot quietly reintroduce the gap.

The other failure, utah-nvidia-gaming, is the btrfs one this branch already
fixes: the OGC kernel formats the root volume and then cannot mount it.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W6BYLUD6KBzF8rBFXKjP7M

* Gate the OGC kernel config in milliseconds, not in a 45-minute build

Review finding on 84685d6: the body claimed "new coverage for the BTRFS_FS
gate" and there was no BTRFS reference anywhere under tests/. That was true,
and it is the same mistake I have been flagging on other people's pull
requests this week -- a description asserting coverage the diff does not carry.

Rather than delete the claim, make it true. verify_config is a shell function
taking a config path, so its contract is testable without building anything:

- BTRFS_FS is in required_config;
- a config missing it is rejected, and the message names it;
- a complete config passes, whether the symbols are =y or =m, because
  BTRFS_FS is enabled with --module;
- and every symbol in required_config is genuinely checked -- a subtest drops
  each one in turn and asserts the gate catches it, so a symbol listed but
  never looked at cannot masquerade as a gate.

That last one is the case worth having. The list is the documentation of what
the gaming flavors need to boot, unlock and mount; an entry nobody verifies is
a comment with a misleading shape.

A full kernel build is still the only complete proof and still runs in CI.
This covers the part that can fail in review rather than in 45 minutes.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W6BYLUD6KBzF8rBFXKjP7M

* Make the installer flatpak retry idempotent

The retry wrapper was added so a single timed-out Flathub object could not
fail a whole flavor's end-to-end run. Two of the three installs it wraps
already passed --or-update, which is what makes a retry safe: re-running is a
no-op for a ref that already completed. The installer-local install did not.

That asymmetry is the one case where the retry could make things worse. An
attempt that installs the app but still exits nonzero leaves attempts 2 and 3
to fail with "already installed", so a flaky success becomes a hard failure --
the opposite of the reason the wrapper exists. It is unlikely on a file://
remote rather than the network path the retry was written for, but the helper's
contract should not depend on which call site it is protecting.

The new test asserts every retried install carries --or-update, and it fails
against the previous line, so the uniformity is now a gate rather than a
convention.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W6BYLUD6KBzF8rBFXKjP7M

* Drop the redundant BTRFS require_config now that the gate covers it

With #158 merged, BTRFS_FS is in required_config, and verify_config .config runs
on the very next line iterating that array with the identical pattern
^CONFIG_BTRFS_FS=(y|m)$. So my extra require_config line asserted something
already asserted two lines later.

Worse than merely redundant: require_config prints a fixed diagnostic listing
SCHED_CLASS_EXT, NTSYNC, ANDROID_BINDER, DEBUG_INFO and BPF settings. For a
btrfs failure that is noise pointing at the wrong subsystem, where
verify_config says plainly which symbol is missing.

What is left of this branch over main is the Flathub retry, the test classes,
and the comment recording the btrfs failure verbatim -- the mount error, why the
list had never needed a filesystem before (the missing KMS device stopped the
test at phase 2), and that BTRFS_FS selects its own crc/zlib/lzo/zstd/raid6
helpers so naming it is enough for olddefconfig. OgcKernelConfigGateTests still
drives the real verify_config against generated configs, dropping each
required_config symbol in turn, so it now gates main's code rather than this
branch's.

250 tests pass.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W6BYLUD6KBzF8rBFXKjP7M

---------

Co-authored-by: Claude <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants