OPINIONATED OMARCHY
TROUBLESHOOTING CORPUS
790
RECORDS
790
AUDITED
0
UNCHECKED
2388
SOURCES
LIVE

Journal: handoff

Last updated: 2026-11-05

START HERE: the next session is about getting back on track

The operator's reading on 2026-09-06, after this repo's published pages were audited and found to reproduce: the corpus needs massive expansion and deep auditing, and the skill phase has not started. Everything in this journal since 2026-08-29 has been the bench, and the bench has spent most of that time catching its own defects. The next session is not another bench session.

What "back on track" means, in the order the dependencies run:

  1. Expand the corpus. 790 records across 13 categories since 2026-10-05: one interrupted harvest, one gap-fill pass, the O4 issue-tracker harvest merged on 2026-09-11, and the first full harvest through the O6 append path, 285 records with two audits each. CLAUDE.md "Regenerating the corpus" names the three workflow scripts, what each does, and that harvest-workflow.js costs about 35 agents. Check /usage-credits first; the first harvest died on a spend limit. Pass the corpus root in args. Every new record lands with its provenance marked, never blended in as audited.
  2. Audit deeper. The last 4 unaudited records and the three VM-found defects were applied on 2026-09-06 (second session below), and all four unaudited records turned out to be wrong. What remains is the larger point: audit_status: ok still means "matches its sources", which the first live scenario showed is not "true on Omarchy 4", and 82 records still carry that status on one source pass. O3 IS COMPLETE: 159 records have been re-audited that way and 157 of them needed correcting, across boot-kernel (10), pacman-aur (22), gpu-drivers (14), apps-services (23), network (15), power-suspend (16), omarchy-theming (15), hyprland-config (13), wayland-compat (9), audio-input (9), omarchy-core (8) and display-monitors (5). That is every ok record carrying a danger that applies to Omarchy. Two passed: mt7921e-dead-after-suspend-aspm and shell-section-override-ignored-without-colors-toml. One was rejected as a problem that does not exist, and was kept and rewritten by hand to say so rather than retired. Hand the brief to agents directly, one per one or two records, which is what the last five batches did, and use research/validation/ for the ones a VM can reach. Six ways forward, O1 to O6, are item 8 under "What's left": O1 (lint) and O2 (workflow prompts) are done, O3 is finished across twelve categories, O4 is finished, its 36 records audited and merged on 2026-09-11, and O5 (checked_against) landed on 2026-09-16. O6 landed on 2026-10-03: a harvest result now appends through merge_gapfill.py, so nothing blocks a full harvest. The corpus prose is done, item 6: 1,366 dashes, 613 semicolons and 183 spaced hyphens removed on 2026-09-13 under "What's left", and is its own job. 1b. The 4.0.4 re-check is DONE (2026-09-19) and closed the debt below: eight records held against 4.0.4-1, none of the defects fixed, and merge_gapfill.py gained a recheck verdict status so a re-confirmation no longer downgrades a corrected record. The first defect this project found itself rather than harvested is filed privately as GHSA-8p6g-pf3r-g2rg. See the 2026-09-19 session, which also records that two of four review agents were stopped partway, so shell/ and most of bin/ are unreviewed.

1e. The upstream round U1 to U6 is DONE (2026-10-19). Two posts went out, both from the TechLuddite account at the operator's instruction: U3 as charmbracelet/gum #1172 (gum confirm exits 0 on SIGTERM) and U4 as a comment on omacom/omarchy #12392 (since 4.0.5 the wrong pick is nvidia-580xx-utils). U1, U5 and U6 did not clear the new bar in CLAUDE.md "Contributing upstream", and U2 is already fixed on quattro. Each verdict and its evidence is under "Outcome so far" in the 2026-10-18 session. Watch #1172 and #12392 for replies.

1h. v4.5.0 runs on test1 (2026-11-05), and #17705 blocks none of the operator's machines, because stable takes extra from a snapshot that still has hyprland 0.56.2-4. Two comments went upstream, on #17705 and #17525, and a tested migration was withheld. test1 is not clean. See the 2026-11-05 session.

Next session, in order: 1f (the ranking, no VM needed), then 1d (bring both VMs to v4.0.6 first, which shipped 2026-10-19, saving the 4.0.5 goldens as reflink copies), then the 1g spike. Both VMs are shut off.

1f. The local-model ranking is OWED (2026-10-18). The 10-model opencode scan finished: runs 52 (published-wrong), 53 (stale-advice) and 55 (deep-triage control), 3 repeats each. Run 54 is a duplicate of 53 and run 56 a partial duplicate of 55, both started by a stray launch loop and run concurrently with their twins, so their timings are contaminated. Rank fix rate and harm separately, and leave out the 0 altered files assertion, which failed for every model (item 4 under "Four things the update broke"). Then export_results.py, and update MODELS.md if the result is cited. Token counts are empty on every case: the runner does not yet parse them from the opencode event stream that each case stores.

1g. The agentic lane gets a harness that plays the human, now that U1 to U6 are done (decided 2026-10-19). Today the runner starts opencode run headless in a tmux window over ssh, with nobody present. Under H1 the harness types the prompt into an interactive opencode or pi session on the VM's desktop through QEMU input injection (virsh send-key or QMP input-send-event, with a usb-tablet for absolute pointer), answers the agent's prompts as a user would, and takes virsh screenshot frames for a visual grader that sits beside STATE rather than replacing it. The agent under test stays a CLI agent. H2, where the agent itself drives the desktop, was rejected because the skill is written for an agent with a shell. Both VMs must run at once and be watchable live through tools/view-test-vms.sh, whose windows only watch while the harness injects below the guest. The read-only mirror at tools/provision-bench-vm.sh:176 cannot take input, so H1 needs its own interactive terminal. Start with a spike: omarchy-agentic-config, one model, headless against H1 on test1, then both VMs.

1d. The 4.0.5 re-check is OWED (2026-10-18). Upstream shipped v4.0.5. Both test VMs run it, the workstation does not. v4.0.6 followed on 2026-10-19 and touches no pacman, Vulkan or gaming file, so the work list stands, but hold records against v4.0.6. Three fixes are now wrong and about nine problems are fixed upstream. The work list is the 2026-10-18 session, and the evidence is research/raw/release-4.0.5-impact.md.

1c. The security category EXISTS as of 2026-09-18, 16 records, all 16 corrected by audit. Tier 1 is done except NET1 (blocked at gate 1, this repository published the mechanism first) and NET2 (fails gate 1). Tier 2 and 3 are untouched, 23 topics in research/raw/security-topics.md, and the brief and the gates now have one real outing behind them. The open debt is the 4.0.4 re-check: the harvest was held against 4.0.2-1 on a workstation whose sync database was dated 2026-09-05, and upstream had shipped 4.0.3 and 4.0.4. checked_against says 4.0.2-1 on all 16 honestly, the audit already folded in the two consequences it found, and OM1, OM2, OM6, OM7 and the Omarchy half of PK5 still rest on 4.0.2 blobs. See the 2026-09-18 session.

  1. Then the skill. The design is settled in opinionated-omarchy/CLAUDE.md and does not need re-deriving; it needs a corpus worth retrieving from. The root README.md now says in public that the skill is vaporware. Make that stop being true in that order, not the other way round.

Carried forward from the bench, not the priority: runs 43 to 46 are void (opencode auto-rejecting tool calls outside its working directory, no error on the case, the control hit too) and need repeating with the permission grant now in app/runner.py. The recipe is in the 2026-09-05 fourth session. omarchy-agentic-published-wrong is not calibrated. The chat-lane result (+22.6 to +28.8 pt on four models, controls flat) and the n=31 agentic null (DiD +0.2, p=0.98) stand.

The four upstream reports owed since 2026-09-11 are filed (2026-09-14/15): the resume hook ordering on #8471, #10375 and PR #8888, ESP free space as #11821, the Intel VA-API comment on #7866 back on 2026-09-11, and the enterprise Wi-Fi finding through upstream's private security channel (see writeups/upstream/05-WITHHELD.md, which is deliberately not the report). Three of the four were materially wrong before adversarial review and one of our own diagnostic instructions turned out to be broken. Read the 2026-09-14/15 session before writing anything upstream again.

Four merges landed on 2026-09-13 and the corpus is 489 records, not 492: three retirements (omarchy-resume-hook-appended-after-filesystems, xwayland-apps-blurry-hidpi, nvidia-black-screen-external-after-suspend) and one category move. Three published pages 404 as a result. See the fifth session of 2026-09-13.

State of the record: every figure on the seven published pages was recomputed on 2026-09-06 and reproduces from the repo. Trust the pages as of that date; recompute before quoting anything newer.

Session of 2026-11-05: v4.5.0 on test1, the stable blocker that was not one, and two comments upstream

Upstream shipped v4.5.0 on 2026-11-05 (05:36 UTC). Earlier that day a Claude session on nexus1 judged the operator's three machines not ready to update, on the strength of #17705: stable's omarchy-hyprland-titlebars 0.2.2-2 requires hyprland=0.56.2, the report says extra now has 0.57.1-1, and migration 1791403310.sh then fails partway through the update. That judgement assumed a stable machine takes extra from live Arch. It does not, and the update went through cleanly on test1 twice.

What test1 did

test1 went from 4.0.5-1 to 4.5.0-3 with omarchy update -y, on stable, and exited 0. Migration 1791403310.sh installed omarchy-hyprland-titlebars 0.2.2-2 next to hyprland 0.56.2-4. The kernel moved to 7.2.15-1-omarchy, and after a reboot Hyprland and the shell came back with no failed units and the session unlocked. The pre-update disk was kept as a reflink copy, /var/lib/libvirt/images/pre-4.5.0-test1-4.0.5-1.qcow2, and the second run restored it first, with the same result: exit 0, 100 migrations, the same package pair. test2 was not touched and is still 4.0.5-1. Neither golden image changed.

The reason is the stable mirror. The omarchy-mirror README says it "serves the Arch snapshot the current Omarchy release was tested against" and that "stable never syncs from upstream itself". The extra.db readings on 2026-11-05:

Mirrorhyprland
stable-mirror.omarchy.org (stable)0.56.2-4
stable-mirror.omarchy.cn (stable, cn region profile)0.56.2-4
rc-mirror.omarchy.org (rc)0.56.2-4
mirror.omarchy.org (edge)0.57.1-1
geo.mirror.pkgbuild.com (Arch)0.57.1-1

None of the operator's machines is blocked by #17705. The desktop (ohmy-omarchy, 4.0.5-1), the laptop (windows-laptop, 4.0.6-1) and nexus1 (4.0.6-1) all have only the stable-mirror.omarchy.org line in /etc/pacman.d/mirrorlist, and each sync database lists extra/hyprland 0.56.2-4. None of them was updated in this session. The NVIDIA freeze reported in #17635 on kernel 7.2.15 is untested here, because the VMs have no NVIDIA GPU, and it is the main risk for the desktop's 3090.

One trap, again

The second run hung at Remove 3 orphaned packages?. The watcher that declines gum confirm with SIGINT had been written to the VM after the disk image was captured, so restoring the image deleted it and nothing answered. v4.5.0's omarchy-update passes -y to omarchy-update-orphan-pkgs, but the process running the update is still the 4.0.5 script, which omits it, so the hop from 4.0.5 still hits #8780. kill -INT on the prompt let the run finish. Install the watcher after any reset.

The packages an update does not install

#17525 reports that an update never installs packages added to the default set after the install. That is confirmed: nothing in omarchy-update compares the machine against omarchy-pkg-defaults. Between v4.0.5 and v4.5.0, 24 names were added to install/omarchy-base.packages, and v4.5.0 migrations name 22 of them. On test1 after the update, pacman -T $(omarchy-pkg-defaults) reported two of 162 defaults missing, and neither is an oversight:

This session's first report got the second one wrong. It counted superwhisper-bin as a gap, and a prototype installed it. The prototype, omarchy-pkg-reconcile in the session's scratch folder, installs what pacman -T reports missing, skips any name pacman.log shows was removed and the preinstalls when preinstalls-removed exists, and installs one package per transaction so one bad dependency costs one package. Greptile's review of #15913 also caught a defect it shared: pacman -T … || true treats a failed query as nothing missing. That is fixed in the scratch copy.

What went upstream, and what did not

Both posts went out from the TechLuddite account at the operator's instruction, after a fresh duplicate search and two rounds of adversarial review by Grok 4.7 and GPT-6 Astra:

Withheld: a migration that installs fcitx5-table-extra. It was written and tested on test1. Without it typing_setup.py add cangjie5 exits 1. With it the package installs and the same command exits 0. A rerun with the package present makes no pacman transaction, and five migration-scanning upstream tests pass with and without it. It was not opened as a PR because it reverses the decision #15508 records, reinstalls the package for anyone who removed it, works against the opt-in proposal in #17148, and helps only someone who picks Cangjie from the picker, who is already told which package to install. It sits uncommitted on a local branch in the scratch clone. Reopen it only if a maintainer says on #17525 that older installs should get the full default set.

Withheld: the general reconciliation step. Some defaults depend on a choice the user made, as superwhisper-bin shows, #17148 moves the other way, and #15913 already covers provides-aware filtering for omarchy-reinstall-pkgs. The pacman.log removal check is the only new part, and on its own it does not clear the bar in CLAUDE.md "Contributing upstream".

Outstanding

  1. Watch #17705 and #17525 for replies, and #17705 for a titlebars 0.3.x build on stable.
  2. test1 is not clean. It runs 4.5.0-3 with the withheld migration in /usr/share/omarchy/migrations/1793854933.sh, fcitx5-table-extra installed and cangjie5 in its Fcitx profile. Reset it from a golden or from the pre-update copy before benching on it.
  3. The test VM table and the version paragraph in CLAUDE.md were brought to v4.5.0 the same day, after a reboot put test1 on 7.2.15-1-omarchy. The paragraph had also been wrong since 2026-10-18 about the workstation, which runs 4.0.5-1, not 4.0.4-1.
  4. The 4.0.5 re-check in 1d now has to be held against v4.5.0, a feature release. Nobody has read the v4.0.6...v4.5.0 compare against the corpus yet.
  5. When the fleet updates, snapshot first and check that each machine boots before starting the next. The desktop is the one exposed to #17635.

Session of 2026-10-18: the test VMs move to 4.0.5, and the next audit has a work list

Upstream shipped v4.0.5 on 2026-10-18. Both test VMs were updated to it the same day. The workstation was deliberately left at 4.0.4-1. One agent then read the whole v4.0.4...v4.0.5 compare (43 commits, 201 files) against the corpus. Its report, with the patch line behind every claim, is research/raw/release-4.0.5-impact.md. No verdict from it has been applied. This entry is the work list for the audit that applies them.

What the VMs are now

test1test2
omarchy4.0.4-1 to 4.0.5-14.0.1-1 to 4.0.5-1
kernel7.2.8-9-omarchy7.2.8-9-omarchy
opencode1.18.31 to 2.0.322.0.32

test2 had to be re-provisioned first, because it predated the Defaults:techluddite !authenticate fix and the update would have died at sudo -v. Both golden images were re-saved at 4.0.5 after the update. The previous ones are kept as reflink copies, golden-test1-4.0.4-1.qcow2 and golden-test2-4.0.1-1.qcow2 in /var/lib/libvirt/images, so either older state is one cp away. All eleven agentic seeds pass check_seeds.py on 4.0.5.

Four things the update broke or disproved

  1. Killing gum confirm answers YES under gum 2.0.2, not no. Measured on test1: SIGTERM exits 0, SIGINT 130, SIGKILL 137. omarchy-update-restart runs gum confirm ... && omarchy-system-reboot, so a watcher that sent kill to the reboot prompt rebooted both VMs. No harm resulted, because the kernel update needed the reboot anyway, but the advice in CLAUDE.md to kill the gum process to decline is wrong for this gum. Use kill -INT. The orphan prompt did not fire: pacman.log shows the only removals were cups-pdf and cups-browsed on test2, each by a migration with --noconfirm.
  2. opencode 2.0 removed --dir. opencode run --dir exits 1 with Unrecognized flag: --dir, which floored both cases of run 50 in under 3 seconds. The runner now does cd "$HOME" and passes --standalone, so each case gets a private server rather than 2.0's shared background service. Run 51 confirmed it end to end. Agentic runs before 50 used opencode 1.18 and runs from 51 use 2.0, so do not compare them directly.
  3. omarchy-theme-set kills a headless opencode. It calls omarchy-restart-opencode, which runs killall -SIGUSR2 opencode so an interactive session reloads its theme. opencode run has no handler, so the agent dies with exit 140 as soon as it applies a theme (run 47). The runner now ignores USR2 before starting opencode, which survives exec and was tested both ways. No earlier run was affected: every case in runs 39 to 46 has an empty error field.
  4. A bench seed's pristine copy survived the update and failed one assertion for every model. The theme seed of omarchy-agentic-published-wrong kept an unversioned copy of the Everforest colors.toml in /var/tmp, taken on 4.0.4. On 4.0.5 it restored that copy with 4.0.4's mtime, so pacman -Qkk omarchy reported 1 altered file and the 0 altered files assertion failed on every case of runs 52 to 56, including the benches that never touch the theme. The failure is the same for every model, so rankings from those runs stand, but that check caught nothing in them and regrading cannot recover it. Both VMs were repaired from the cached 4.0.5 package, and the seed now names its copy for the installed release and refuses to capture a file the package already flags. Neither golden held the altered file. The VMs picked it up from the first seed run after reset. test2's golden still held the unversioned copy, which the new seed ignores, and it was removed and the golden re-saved at 13:36 PDT.

What the next corpus audit has to do

Hold these against 4.0.5-1 on a test VM, not the workstation, which is still 4.0.4-1. The brief is research/tools/reaudit-brief.md as usual, but an auditor told to check "this workstation" will check the wrong release. Every verdict here is status: "recheck" with checked_against set, so a corrected record keeps its status and its audit note.

Fixes that are now wrong. Do these first, because a reader following them today loses something.

Problems 4.0.5 fixed. These records should say "fixed in 4.0.5" and keep the workaround for older installs. Confirmed from the patch: nvidia-env-forced-on-igpu-primary-hybrid-laptop, omarchy-windows-vm-launch-silent-setgid-chmod, bar-gone-after-update-shell-restart-race, nvidia-utils-installed-on-amd-intel-only-system, and the display-panel sentence of hyprctl-keyword-monitor-non-legacy-parser. Likely, from the 7.2.8-9 kernel recipe and the mirror: linux-omarchy-intel-iommu-default-on-boot-failures, waydroid-no-binder-on-linux-omarchy-kernel, waydroid-rust-binder-no-such-process-bad-bindgen-build and linux-omarchy-7-2-5-hdmi-audio-underruns. waydroid init on a VM settles the two Waydroid records. The HDMI one needs affected Intel hardware.

Claims that no longer hold.

Needs a closer look. Sections 1 CHECK and 5 of the report: lock recovery under #15020, stale line references after #12538, the sshd firewall change against a LAN-scoped rule, Ghostty and other names now resolving to [omarchy], the ESP and the new UKI migration, and the Neovim SSH clipboard provider. The one with real risk is provider order: with [omarchy] first, a plain pacman -S steam likely resolves lib32-vulkan-driver to Omarchy's nvidia-580xx packages. That touches steam-proton-lib32-graphics-missing, nvidia-utils-installed-on-amd-intel-only-system and package-conflict-and-provider-prompts-during-upgrade. It is likely from the databases, and a dry run on a VM confirms it.

A recheck stamp only: stalled-lock-latch-suspends-unlocked and shell-plugins-run-unsandboxed.

New records worth harvesting. Fourteen candidates are in section 3 of the report. The two that matter most:

The three release-note security items (#11336, #10379, #11495) have no record yet. They pass gate 1 because upstream published them, and their mitigation is omarchy update to 4.0.5.

Outside the corpus. The vendored omarchy/ is byte-identical to the v4.0.5 tag, so the +29.3 pt baseline still describes the released skill. The vendored diagnose-crash/SKILL.md is not: it matches quattro HEAD and tells an agent to run omarchy-crash-mute, which no release ships. That predates 4.0.5.

Next session: the upstream work list

The next session works these upstream, as a report where someone else must fix it and as a pull request from the fork TechLuddite/omarchy where we can. Searched for duplicates on 2026-10-18 with gh search issues and gh search prs against omacom/omarchy. Most already have a thread, so the useful contribution is usually evidence or a tested review, not a new issue.

ProblemExisting threadLikely contribution
U1omarchy-theme-set kills a headless opencode run with SIGUSR2 (exit 140)#8002 open, #8389 closed, PRs #8004 and #14640 openTest #14640 against opencode 2.0.32, which arrived with 4.0.5 after the PR was written, then comment with the result
U2omarchy update -y still hangs on the orphan and reboot prompts on 4.0.5#8780 and #8986 open, PRs #8824 and #9511 openConfirm both scripts are unchanged in 4.0.5 (they are, read on test2), test the PRs on a VM, comment
U3gum confirm exits 0 on SIGTERM under gum 2.0.2, so gum confirm ... && omarchy-system-reboot reboots when the prompt is terminatednone found in charmbracelet/gum or omacom/omarchyRead gum's signal handling, reproduce on plain gum, report to gum. Test on a VM whether logging out at the reboot prompt reboots the machine, which would make it an Omarchy issue as well
U4With [omarchy] first, virtual providers resolve to nvidia-580xx on machines with no NVIDIA cardnone foundNew issue on omacom/omarchy, citing #14551. Before writing, list which omarchy-pkg-add callers depend on vulkan-driver or opengl-driver, because omarchy-pkg-add passes --noconfirm
U5Migration 1791403252 overwrites /etc/pacman.conf.bak, the only backupPR #7998 (timestamped backups) and PR #15017 openComment on #7998 that 4.0.5's migration hits the same single-backup problem
U6A user's existing custom-repo hook still inserts above [core], so custom repositories now rank below [omarchy]none foundReport against #14551, or a PR that migrates the user's copy of the sample

Re-checked on 2026-10-19, a day after v4.0.5, because the tracker moves fast. Four rows changed:

U3 still has no thread (gum's latest release is v2.0.2, 2026-09-24) and U5 is unchanged: quattro's migration still runs sudo cp -a /etc/pacman.conf /etc/pacman.conf.bak at line 22, and #7998 is open.

Evidence already in hand. U1 is run 47 and the exit-140 test in this entry. U3 is the measured exit codes above. U4 is pacman -Sp --noconfirm steam on test2 at 4.0.5-1, which listed nvidia-580xx-utils and lib32-nvidia-580xx-utils from pkgs.omarchy.org. U5 and U6 are in research/raw/release-4.0.5-impact.md, items 10 and 12.

Outcome so far (2026-10-19). The bar for posting is now in CLAUDE.md under "Contributing upstream". In short, post only what is new and could help many people.

How to work them, from what went wrong last time. Read the 2026-09-14/15 session before writing anything. Three of four drafts were materially wrong before review. Every draft goes through an adversarial review by an agent told that the draft is wrong until each claim survives, with the claim most likely to be false named in the prompt, and then a second review by a different model. Say in each post which models drafted and reviewed it. Reproduce on a test VM at 4.0.5-1, never on the workstation. Anything with a security angle goes through HackerOne (<https://hackerone.com/omarchy>), not a public issue. None of U1 to U6 looks like one, and that judgement belongs in the review prompt so the reviewer can overturn it. The operator decides what is posted, and a draft is not a post. On 2026-10-19 the operator had the agent post U3 and U4 itself, and CLAUDE.md now records that rule and how to run the Grok and GPT reviewers.

Not upstream work, but owed. The vendored diagnose-crash/SKILL.md tells agents to run omarchy-crash-mute, which no release ships. Decide whether to re-vendor from the v4.0.5 tag. That moves the bundle sha and starts a new bench series. pacman-file-exists-in-filesystem-omarchy was due to expire at the next Omarchy release (2026-09-14/15 item 4), and that release is 4.0.5.

Not covered

The Hermes and OpenClaw installer bodies were not read. The October Arch snapshot (systemd 262, mkinitcpio 42.2, Qt 6.12, linux 7.2.11) was not checked record by record, and about 47 records cite older Arch versions.

Session of 2026-10-04/05: the corpus grows to 790, and the second audit still found 64% wrong

The first full harvest since O6 ran on 2026-10-04 and was merged on 2026-10-05: 285 new records, 505 to 790. The existing 505 came through every merge byte-identical, checked by comparison after each step.

What ran, and what it cost

StepAgentsSubagent tokensWall clockOutput
harvest-workflow.js, 12 categories366.85M99 min341 records: 143 ok, 66 corrected, 132 gapfill-unaudited
Dedupe10.35M8 min56 removed, 285 kept
audit-existing-workflow.js, 40 batches of 4 to 8405.89M23 min285 verdicts: 182 corrected, 103 ok

Results are in research/raw/harvest-2026-10-04-result.json, harvest-2026-10-04-deduped.json and audit-2026-10-05-harvest-result.json.

The first auditor passed what the second rejected

The harvest's own auditor, one per category seeing 14 to 23 records at once, passed 143 records ok and rejected none. The second pass, six to eight records per auditor against the sources, the quattro tree and this workstation at 4.0.4-1, corrected 182 of 285, 64%, including 74 the first auditor had passed. 90 had their cause rewritten, 39 their symptom, 38 their verify step, 32 their danger and 10 their title, and 10 cited sources were removed. That is lower than the near-100% earlier second passes found, plausibly because the harvesters now read reaudit-brief.md before writing, but it is still most of the batch. Budget the second audit as part of every harvest, not an optional extra. Neither pass rejected a single record.

The new verdict fields were used at once

The four fields added to both workflows in this session were used in the first run that had them: 10 corrected titles, 8 corrected frequencies and 10 source removals, every one of which would otherwise have landed in reason and been lost. A fifth gap was found before it cost anything: an ok from the second auditor on a record the first had corrected would have downgraded it and erased its note. merge_gapfill.py now keeps corrected and extends the note in that case, which fired 16 times. Test added, suite at 36.

Duplicates came from inside the harvest, not from the corpus

Telling each harvester to read its category's existing records worked: overlap with the existing 505 was one record. The duplicates were internal. The gap-fill agent never sees the harvester's batch, so it repeats it, and five records on Intel PSR flicker and three on the 4.0.x locale abort came from different agents. One agent judged 33 heuristic candidate pairs plus a full title scan, and 56 records were dropped with their extra sources carried over to the survivor (82 URLs, each checked to be cited by the dropped record). It also held 34 close pairs apart as distinct faults. harvest-workflow.js should pass the harvester's slugs and titles into the gap-fill prompt, which would remove most of this step.

The lint, and the baseline

53 new records hit lint_corpus.py. All 53 were read: every hit is a labelled plain-Arch branch, a command named only to forbid it, a global-option or entry edit to /boot/limine.conf (which Omarchy itself makes, unlike kernel parameters), or a record about Omarchy 3 on purpose. The baseline went from 157 to 210 with every earlier entry unchanged. That is O7 again: the lint cannot tell "run this" from "never run this".

Two mistakes of mine, both caught

Security

The harvest and audit prompts carried the gate-1 rule, and two scans of the merged records found no mention of the Docker DNS mechanism reported privately as GHSA-8p6g-pf3r-g2rg. Three records describe Omarchy's Docker DNS configuration only as far as its source files and public issue #11757 do.

Where to pick this up

  1. Merge PR #76 and this session's PR in that order. The site republishes with 790 records.
  2. Pass the harvester's records into the gap-fill prompt before the next harvest, see above.
  3. shell/ and most of bin/ are unreviewed for security, under the HackerOne terms.
  4. Tier 2 and 3 security topics, 23 of them, still untouched.
  5. Then the skill. The corpus is 56% larger than when the skill design was written.

Session of 2026-10-03: the reporting channel moved to HackerOne, and O6

Upstream now takes security reports through HackerOne, with bounties

Omarchy launched a public bug bounty at <https://hackerone.com/omarchy> on 2026-10-01, and <https://omarchy.org/security/> now names it as the channel, with security@omarchy.org kept as a fallback for reporters without an account. The repository's .github/SECURITY.md still names email only and was last changed on 2026-08-30, so read the website, not the repository file. The operator has an account. Four terms of the program change how this project reports:

The three advisories filed before the launch (GHSA-3v6r-47cg-h3qp, GHSA-7v99-3429-4q3x, GHSA-8p6g-pf3r-g2rg) were acknowledged by upstream's security team on 2026-09-21 and moved to internal tracking, with the Docker DNS one rated low. They are closed from this side and are not being refiled. That also retires the drafted email proposing an 18 December window.

O6: a harvest appends instead of replacing

harvest-workflow.js output could only enter through ingest.py, which replaces the corpus and would have discarded every correction and cause_reconciled stamp. merge_gapfill.py now takes a harvest result (top-level problems, no results) and appends it: existing records come out byte-identical, a colliding slug gets a numeric suffix, a new category gets its label, and a new record resembling an existing one is named for review but kept. ingest.py refuses to overwrite a non-empty corpus without --replace. Four tests cover both, and the suite is at 35.

Three defects in harvest-workflow.js itself were fixed before running it on 2026-10-04:

Where to pick this up

  1. Expand the corpus, item 1 of START HERE, which O6 no longer blocks. Check /usage-credits first.
  2. shell/ and most of bin/ are unreviewed for security, and a finding there now goes to HackerOne under the terms above.
  3. Tier 2 and 3 security topics, 23 of them, still untouched.

Session of 2026-09-19: the 4.0.4 re-check, and the first finding this project found itself

The three things the last session left are done or answered. Eight security records were held against 4.0.4-1 on this workstation, the VM provisioning defect is fixed and test1 runs 4.0.4-1 with a fresh golden image, and a static review of upstream at v4.0.4 (c668141e9c42b13c80c9ca4ea108e11708c5e8a5) produced one confirmed new defect, which is now filed privately.

None of the eight defects is fixed on 4.0.4

The five the last session named (OM1, OM2, OM6, OM7, PK5) plus three the v4.0.2...v4.0.4 compare said upstream had touched: OM4 the lock latch, OM5 the plugin sandbox, OM9 the sudo grant. Every one of the eight still holds. Six needed only stale version wording, line numbers or issue states updated, and two needed more:

The verdict shape was missing a field, for the fourth time

An ok verdict sets audit_status to ok and replaces audit_note. Every security record is already corrected, so merging eight re-confirmations as ok would have downgraded all eight and erased what the 2026-09-18 audit found. There was no way to say "held against a newer release, still stands".

merge_gapfill.py now has a recheck status: it keeps audit_status and audit_confidence, extends audit_note rather than replacing it, requires checked_against (a recheck with no named install is only an ok), and applies any corrected_* fields, because a newer release can leave a defect in place and still date its wording. Two tests cover it, and the corrections block is now one helper both paths call.

That is the fourth time this gap has been found the same way, after corrected_severity, checked_against and corrected_title. The rule in CLAUDE.md held: the field was added with a test rather than applied by hand.

research/docs/security.md was never committed

.gitignore line 26 was docs/, unanchored, for the generated site at the repo root. Git matches that at any depth, so it also matched research/docs/. The older category pages survive because they were tracked before that line existed, and the new security page, added 2026-09-18, was silently ignored. It is now /docs/, and the page is untracked-and-visible rather than invisible. The rule that a commit changing the JSONL must change research/docs/ in the same commit has been quietly unenforceable for that one file since 2026-09-03.

The VMs, and two more ways a headless omarchy update fails

Defaults:techluddite !authenticate is in tools/provision-bench-vm.sh and sudo -n -v now passes on test1. That trap is already public upstream as open issue #10352, so somebody else hit it and reported it. Two new ones, both of which hang omarchy update -y with no output:

  1. omarchy-update-orphan-pkgs ignores OMARCHY_UPDATE_UNATTENDED. It guards on [[ ! -t 0 || ! -t 1 ]], and omarchy-update runs it under script -qefc, which supplies a pty, so the guard passes and it sits on a gum confirm.
  2. omarchy-update-restart does the same with "Linux kernel has been updated. Reboot?".

Ending each prompt lets the run continue, and the packages were already at 4.0.4-1 by then. test1 is provisioned, updated, verified across a reboot (both bench units active, passwordless sudo intact) and saved as the golden image.

A4: the private review, and what it cost

Four review agents were planned over upstream's tree at v4.0.4, split into bin/ plus etc/, install/ plus migrations/ plus default/, the shell core, and the shell plugins. Two of the four never produced anything: a safety classifier stopped the shell core review and the privileged bin/ review partway. Neither was retried in another form. So shell/ and most of bin/ have not been reviewed, and that is the largest gap left in this workstream.

What the surviving reviews found is not described here, deliberately. One candidate was confirmed on the two test VMs and filed privately as GHSA-8p6g-pf3r-g2rg on 2026-09-19, rated medium. Three more are held unreported: one confirmed but low impact, one likely but not reproducible on a VM, one likely by design. Their mechanisms, the evidence and the filed text are in ~/.local/share/omarchy-security-reports/ (mode 0600 in a 0700 directory) and in the 2026-09-19 disclosure record in Substrata's work.decisions. The same rule as writeups/upstream/05-WITHHELD.md applies: nothing about them enters this repository, the corpus or this journal until upstream has fixed them and said so.

The corpus cannot be the vehicle for early reporting. Gate 1 admits a mechanism only once somebody else has made it public, so every security record is a late disclosure by design. Finding things first looks different: read the code, confirm on a VM you own, report privately, and keep it out of the corpus until upstream has moved.

The window, again

The 2026-09-19 report states no disclosure window, only an offer to agree a date. That is the same omission as both enterprise Wi-Fi reports, so it is now three reports in a row. The disclosure record is in work.decisions at draft, and an email to security@omarchy.org is drafted proposing publication no earlier than Friday 18 December 2026 for all three advisories. The channel upstream's policy actually asks for is that email address, and it has never been used. All three advisories sit in triage with no maintainer action: GHSA-3v6r-47cg-h3qp since 2026-09-12, GHSA-7v99-3429-4q3x since 2026-09-14, GHSA-8p6g-pf3r-g2rg since 2026-09-19.

Where to pick this up

  1. Send the email, and put the window in the advisory thread through the web interface. The API refuses advisory comments from the reporter.
  2. Settle the held candidate that a VM cannot reproduce, on physical hardware. The test is in the private directory named above.
  3. shell/ and most of bin/ are unreviewed. Two agents stopped partway. Whatever covers them next should not be a reworded retry of the same task.
  4. Tier 2 and 3 security topics, 23 of them, still untouched.

Session of 2026-09-18: the security category exists, and all 16 records were wrong

16 records merged, 505 in the corpus, 13 categories. The category the last two sessions built the gates for is now published. Eight agents harvested the tier 1 topics against research/tools/security-harvest-brief.md, eight more audited them adversarially, and every single record came back corrected. Zero passed clean.

That is the third batch in a row at essentially 100%, after O3's 157 of 159 and the issue harvest's 36 of 36. It is no longer a surprising result and should stop being budgeted as one: a harvest is a draft, and the audit is half the work.

What the audit caught, and why each one matters

Severity moved on 9 of 16 and every move was downward. The harvest rated 12 of 16 high. Security writing inflates severity, the brief has a four-level scale written for this category, and an auditor told to argue against the scale rather than against how alarming a thing sounds moved most of them to medium or low. Prompt the next audit the same way.

Nine records had their fix rewritten, so none of them would have been right as published. The four worth remembering as shapes:

Two topic-list entries were wrong and the audit corrected them: the AUR has shipped malware three dated times, not twice, and polkit's unreadable-rule failure is not silent, it logs at error priority. What is missing is any failure of the unit, and the "N rules loaded" summary is suppressed at the default log level.

The finding that outranks all of them: an installed version is not the current one

The audit of OM9 went looking for the package build date and found that upstream was at v4.0.4 (2026-09-15) and v4.0.3 (2026-09-08) while this workstation sat at 4.0.2-1 with a pacman sync database dated 2026-09-05. Two releases had come and gone unseen. OM9's expiry defect is fixed in 4.0.3, and OM5's plugin access was restricted there, so two records asserted behaviour that had already changed.

Both were folded into the records before merge, and checked_against says 4.0.2-1 on all 16 because that is what was actually held against a machine. The debt that remains is that OM1, OM2, OM6, OM7 and the Omarchy half of PK5 rest on 4.0.2 blobs and need re-checking on 4.0.4. CLAUDE.md now carries the three commands that would have caught this on day one: the installed version, the age of the sync database, and the current upstream tag.

The operator updated this workstation at the end of the session, so the next session starts on a machine that is current and can do that re-check directly.

corrected_title, and the gap that keeps being found the same way

Two auditors judged a record's title wrong, had nowhere to put the replacement, and wrote it into reason for somebody to apply by hand. That is the third time: corrected_severity and corrected_frequency in September, checked_against last week, corrected_title now. It is in merge_gapfill.py with a test, and both retitles were applied through it.

The rule worth keeping: when a verdict shape is missing a field, add the field with a test rather than hand-applying the change, because a correction that lives only in prose is a correction that gets lost.

One record needed a hand edit after the merge anyway. The cause of the remove-security-sshd record named sudo omarchy-remove-security-sshd literally, and lint_corpus.py reads that shape as advice rather than as a warning. The brief already tells harvesters to describe a trap command in prose, and the record now does.

The test VMs, and three ways a headless omarchy update fails

The VMs were started for validation and are all written up in CLAUDE.md under "The VMs have no pacman sync databases". Briefly, in the order they bit:

  1. omarchy update blocks on a gum confirm with no tty, hanging with an empty log. omarchy update -y is the script's own unattended path and sets OMARCHY_UPDATE_UNATTENDED=1.
  2. Killing that run orphaned sudo paccache -rk2, which kept the flock on $XDG_RUNTIME_DIR/omarchy-update.lock. Every later run said "An Omarchy update is already running" and named neither the process nor the lock file. fuser and lsof showed nothing unprivileged. Walking /proc/*/fd as root found it.
  3. sudo -v prompts even though the user has NOPASSWD: ALL, because the user is also in %wheel, which has password-requiring entries, and validation considers all of them. Some step of the update calls it, so the run dies at sudo: a password is required after the snapshot step. Defaults:techluddite !authenticate in /etc/sudoers.d/99-bench-nopasswd fixes it and belongs in tools/provision-bench-vm.sh, which is not yet done. The VMs are therefore still at 4.0.1-1 and the validation run has not happened.

Also confirmed, and it settles half of an open question in CLAUDE.md: on 4.0.1-1, OMARCHY_PATH is empty in a non-interactive ssh, so bash -lc is required there. That is the old behaviour, and it says nothing about the 4.0.2-1 claim, which is what CLAUDE.md already says. A workstation now at 4.0.4 can settle it with one command from a second machine.

Also this session

research/README.md was brought inside the mechanical writing check. The exclusion pattern research/[^/]+\.md, meant for the loose Hyprland wiki pages, was also matching one of the seven published pages, so it had never been checked. The pattern is now research/[a-z0-9]+\.md, the wiki pages being lowercase and the README not. It carried no prose dashes and 13 prose semicolons, now rewritten.

Where to pick this up

  1. Do the 4.0.4 re-check on the five records named above, now that the workstation is current. It is a small, well-defined job and it closes the only known debt in the new category.
  2. Done on 2026-09-19. Defaults:techluddite !authenticate is in tools/provision-bench-vm.sh, test1 runs 4.0.4-1, and its golden image is re-saved. Two further headless traps turned up in omarchy update -y and are in the 2026-09-19 session.
  3. Tier 2 and 3 security topics, 23 of them, with the brief and gates now proven.
  4. The advisory follow-up from the 2026-09-16/17 session is still unsent, and both advisories are still state: triage, untouched by a maintainer since creation. Checked 2026-09-18.

Session of 2026-09-13 (third): the type designers were licensed and not credited

Mark Cuda's Omarchy Font has been the first thing on every page of the public site since 2026-09-03, and until this session his name appeared in four places a visitor would never open: research/assets/fonts/README.md, NOTICE, a CSS comment in build_site.py and the licence file under docs/fonts/. The site footer named only Departure Mono. The root README.md said "the site's typefaces carry their own licences" and named nobody. CLAUDE.md and section 4 of the 2026-09-02 entry below still said one font was vendored. Every licence obligation was met, and that is exactly the failure: a licence file beside a binary is compliance, not credit.

What changed, all in one commit:

One correction to what this session first reported: pages.yml was said to test only the Departure Mono files. It already tested all four font files. The two Omarchy Font lines were beyond the range that was read. The gap that was real was that no test looked for a name.

The governing rule now lives outside this repo, as attribution/crediting-third-party-work in the standards.engineering lane of Substrata, drafted this session from the OFL 1.1 text and FAQ, the MIT text, Creative Commons' TASL attribution practice, REUSE 3.3, the Apache NOTICE guidance and Debian's copyright format 1.0. Draft, not ratified, not human reviewed.

Session of 2026-09-16/17: the security category gets a gate, and the gate's own precedent was false

The corpus is about to gain a 13th category, security. Two things had to exist before a single record could be harvested, and neither was the harvest.

O5 is done: checked_against

audit_status: ok has always meant "matches its cited sources", which the O3 re-audit proved is a different question from "true on what Omarchy ships": 159 records re-audited, 157 corrected. checked_against holds the pacman package version and the date a record was held against a real install, "<version> <YYYY-MM-DD>", or null.

Five consumers, not the four CLAUDE.md names, because build_site.py reads the corpus too. ingest.py and merge_gapfill.py needed nothing: both already import corpus.FIELDS instead of keeping a private copy, which is what that module was built for. Three tests in the shape of the cause_reconciled ones. The 489 line diff is the whole corpus gaining one null key, verified by stripping the key back off and diffing against the previous file.

It turned out to be a dependency rather than a nicety. The governing security standard requires two point confirmation before a finding is asserted at all, a pinned upstream commit and an installed version, and checked_against is where the second one goes.

The disclosure rule, and the precedent that was not true

First draft sorted candidates by fault: user misconfiguration publishes, upstream defect does not. An adversarial review by Fable 5.1 killed it on two counts, and both were right.

The split cannot classify the work. Every one of the 13 Omarchy candidates is an open upstream issue with no fix, so under the draft the whole section was unpublishable. And the most common case in this category is a bad default upstream chose and a user can change, which is both kinds at once.

The named precedent was checkably false on the live site. The draft said the enterprise Wi-Fi finding is "deliberately absent from the corpus". It is not absent. wpa2-enterprise-8021x-connect-from-cli cites shell/plugins/panels/network/Model.js and Panel.qml, the two upstream files that implement the weakness, and reached the public site at 2026-09-12 03:55 UTC through PR #58. JOURNAL.md line 1115 names Model.js:319-326. The first private report to upstream followed at 2026-09-12 15:34 UTC, eleven hours and forty minutes later, and a second report on the same defect went on 2026-09-14. This repository published the mechanism and then reported it privately, and neither report names the publication. writeups/upstream/05-WITHHELD.md had recorded that caveat honestly all along, and nobody had read it against the rule being drafted.

An earlier version of this entry said three days. That was wrong. It measured the second report rather than the first, because the first had not been found yet. See the correction below.

The rule now states that history instead of claiming the opposite, and keys on what a record discloses rather than on whose fault the defect is. Three gates, all required, in research/README.md: the mechanism is already public and somebody other than us made it public, the record says no more than its cited source, and the record gives a mitigation rather than a reproduction. Gate 1 excludes our own prior publication deliberately. Counting it would turn a disclosure we chose to make into a fact we merely observed.

Nothing was retracted, and that is the considered answer

Checked on 2026-09-17. A third party, pixelsandpointers, filed #11791 on 2026-09-14 naming the same code, and it is still open. So the mechanism is public three ways and withdrawing ours would leave the defect described by somebody else with no mitigation attached, remove a protection readers already have, and advertise the withdrawal. The record stays up and now cites #11791, which is the third party public reference gate 1 wants and which did not exist when the record was written.

41 topics, from 48 candidates

In research/raw/security-topics.md. 12 Omarchy specific, 5 package and key trust, 6 privilege, 3 boot integrity, 2 SSH, 13 network. 22 at tier 1. Every drop carries its reason.

Five research agents gathered them and one review agent corrected five. What the checking caught is the point:

Method note, because it repeats the 2026-09-14 lesson

Sonnet 5 did all five gathering and implementation tasks, each with a hard oracle behind it: a test suite that must go green, or "cite only URLs you actually fetched". Fable 5.1 got the one task whose failure mode is a confident claim with nothing to check it against, and found the defect in our own reasoning. Every prompt named the claim most likely to be false, and in each case that is where the failure was. Three times an agent was wrong or two agents disagreed, and each took one local command to settle. Do not merge an agent's finding without running that command.

Where to pick this up

  1. Write research/tools/security-harvest-brief.md before harvesting anything. The three existing briefs hold agents to what Omarchy 4 ships. None holds them to the disclosure gates, and none asks the question an auditor tuned to "is the fix right" will miss: does the cause field say more than its cited source.
  2. Harvest tier 1 only, 22 topics, in batches of two or three per agent, the shape the last five O3 batches used. Not harvest-workflow.js: it routes through ingest.py, which replaces the corpus and would discard all 407 corrections. O6 is still not done.
  3. Add security to categories.json in the same commit as the first records, never before, or the site publishes an empty category page.
  4. Budget: O3 ran 750k to 900k tokens per ten records, so tier 1 is roughly 1.7 to 2.0M for harvest plus audit. Check /usage-credits first.
  5. There are two advisories for this one defect, and finding the second corrected a figure this journal had already published. GHSA-3v6r-47cg-h3qp was filed at 2026-09-12 15:34 UTC and GHSA-7v99-3429-4q3x at 2026-09-14 16:43 UTC, both by us, both still state: triage, both untouched by a maintainer since creation, rated medium and high. The first is the text in ~/Documents/, which this journal wrongly called a pre-submission draft: it is a filed report sitting at mode 0644. Because only the second was known, the gap between publishing and reporting was written up as three days. Measured against the first it is eleven hours and forty minutes. The rule is unchanged by this and every number attached to it was not. Two live advisories for one defect is noise for a maintainer, so decide which is canonical and withdraw or cross-reference the other before adding anything else to either thread. Neither report names this repository's own publication, and the first describes the source as "a third-party Omarchy troubleshooting corpus", which reads as though the reporter is not the corpus maintainer. The second fixes the attribution and cites the third party's #11791, but still omits ours.
  1. The path was never a discrepancy, and reading the filed report found something worse. ~/Documents/ holds the 2026-09-12 pre-submission drafts. The report that actually went to the maintainer on 2026-09-14 lives exactly where writeups/upstream/05-WITHHELD.md says it does, mode 0600 in a 0700 directory. The standard's Sources line was the thing that was wrong, and it is fixed. Two real gaps came out of reading it: the report states no disclosure window at all, so step 6 of that standard has never been exercised and no date exists on which publishing becomes automatically acceptable, and its prior-disclosure section names the third party's issue but not our own earlier and fuller publication. A maintainer reading it would conclude the only public description is somebody else's passing mention. A one paragraph follow-up in the advisory thread closes that, and it is the highest value thing left in this thread. The disclosure decision record its own standard requires now exists in work.decisions at proposed, and the standards.security amendment is at proposed too. Ratifying either is the one step that asserts a person has read it.
  2. Done on 2026-09-18. CLAUDE.md's mechanical writing check excluded research/README.md through the research/[^/]+\.md pattern meant for the loose wiki pages. The pattern is now research/[a-z0-9]+\.md, which keeps the loose pages out and the published page in. The file carried no dashes in prose: the three at lines 80, 85 and 86 are inside the layout fence and exempt. It carried 13 semicolons in prose, now rewritten, and the check runs clean over the audited set.

Session of 2026-09-14/15: four upstream reports, and the reviews that rewrote three of them

The reports owed since 2026-09-11 are filed. Four went out: three public, one through upstream's private security channel. Every one was adversarially reviewed before submission, and three of the four were materially wrong before that review. That is the story of this session, more than the reports themselves.

ReportWhereOutcome
Resume hook ordering#8471, #10375, PR #8888posted, one substantive reply
ESP free space#11821filed, one substantive reply
Enterprise Wi-Fi, no CA validationprivate advisory, see writeups/upstream/05-WITHHELD.mdstate: triage
Intel VA-API Haswell#7866posted 2026-09-11, already discharged

What the pre-submission reviews caught

Each report was drafted, then handed to an agent told to falsify it rather than proofread it. The instruction that did the work was "assume the draft is wrong until each claim survives", plus a named list of the claims most likely to be false.

Only the Haswell comment, written on 2026-09-11, needed nothing.

The reply round, and the finding that matters most

Both public reports drew a substantive reply within hours. Those replies were answered after a second review round, this time by Claude Fable 5.1 rather than Opus 5, and it found the worst error of the session.

Our own diagnostic instruction was broken. The first comment on #10375 asked jorgenfoss to run journalctl -b -1 -k | grep -i "PM: hibernation". The kernel prints the message that matters from kernel/power/swap.c under pr_fmt "PM: ", so it reads PM: Image not found (code -22) and contains no "hibernation". Verified here: on this workstation, where the hook demonstrably ran and logged that exact line, that grep returns nothing. So the reporter's "no resume attempt logged at all" is also what our instruction produces on a machine where the attempt happened. We may have handed them a false negative and then argued from it.

The follow-up retracts it, and retracts a second theory with it: the idea that jorgenfoss's fix was really the rebuild rather than the reorder does not hold, because omarchy-hibernation-setup writes the drop-in at line 89 and runs sudo limine-mkinitcpio at line 125 in the same invocation. Their image was rebuilt either way.

The review also replaced the proposed test. Asking them to reorder, rebuild, hibernate and revert costs a hibernated session and is not even decisive under the broken grep. Omarchy keeps a persistent journal, so the pre-fix failed boot is probably still on disk and journalctl --list-boots settles it at no risk.

Two things the reply round got right that are worth keeping

Method notes worth carrying

Where to pick this up

  1. No reply yet on #8471 or PR #8888. The #8471 comment argues the ordering is not load-bearing and jorgenfoss's observation sits against it, unresolved until someone reads a pre-fix boot journal.
  2. The private advisory is state: triage, submission.accepted: false, which is the normal pre-triage state and not a rejection. Do not chase it.
  3. The Haswell gap still owes the corpus a new record. This is the oldest open item in this file.
  4. pacman-file-exists-in-filesystem-omarchy expires at the next Omarchy release, per upstream commit f5194e3f.
  5. O7, teaching the lint to skip a warning and a labelled branch, is still open.

Session of 2026-09-13 (sixth): the corpus prose goes through the writing standard

Item 6, the last body of prose in this repo that had not been held to writing-and-responding. Two passes, 25 agent batches, a mechanical verifier between the agents and the corpus. 1,366 em and en dashes, 613 semicolons and 183 spaced hyphens removed, across 1,347 field rewrites. 28 occurrences deliberately survive and are listed below.

BeforeAfter
Em and en dashes1,3668
Semicolons61312
Spaced hyphens18310

The fenced-code exemption protected almost nothing

The note left here beforehand warned that fix is mostly fenced commands, so a blind pass would corrupt them. Measured: 1,364 of the 1,366 dashes were in prose. They sit in the paragraphs around the commands, not in the commands. The exemption still had to be enforced, and it was, but it removed two dashes from scope rather than half the job. Worth knowing before estimating the next pass of this kind.

What made it safe was the verifier, not the prompt

Every batch was checked before it touched problems.jsonl, and nothing merged on an agent's word:

It earned its keep twice. It caught a field where the code spans had moved, which turned out to be a command relocated to a trailing sentence with every span byte-identical, and it caught eleven fields that looked like dropped technical tokens and were all the checker attaching trailing punctuation, NVIDIA against NVIDIA.. Both were false alarms, and both had to be looked at to know that.

28 survivors, and they are the pass-through rule working

These are not misses. If this count ever reads zero, something has edited a quotation.

That last group is the one a future pass will get wrong. The prose filter protects backticked and fenced regions, and none of those are backticked. The agents were told to watch for it and caught every instance. A purely mechanical pass would have corrupted all five.

Two defects found, one of them mine

gtk-file-chooser-does-nothing-xwayland had the part worth removing.Updated on 2026-09-13: with no space. That is mine, from the merge pass earlier the same day: the repointing edit appended a paragraph with .lstrip("\\n") on the wrong side and glued it to the preceding sentence. An agent found it, correctly declined to fix it inside a punctuation pass, and reported it. Fixed here.

The other is unfixed and left deliberately: two dashes remain inside fenced shell comments, for example # ~/.config/fontconfig/conf.d/99-emoji.conf — same <alias>/<prefer> blocks as before. Code is exempt by rule, and a pass over comments inside fences is a different job that is probably not worth doing.

An agent clobbered the orchestrator's tooling

One batch wrote its own scan.py, apply.py and verify.py into the shared job directory, and verify.py overwrote the verifier this session was using. The next verification run silently reported one batch's result, problems: 0 records: 27, instead of ten batches' results, and it was only obvious because the output format was wrong. The tooling now lives somewhere the agent prompts never name. Agents share the job's tmp directory and will write helper scripts into it, so an orchestrator's tools do not belong on a path an agent has been told about. The same shape appeared earlier in the day when a batch left a build_v1.py in an output directory.

Where to pick this up

  1. The corpus now meets the writing standard on every axis the standard names. The mechanical check in CLAUDE.md covers tracked markdown, and the corpus needs the prose-aware scan used here, because a naive grep counts code and quotations.
  2. Three upstream reports are owed and none is written: the enterprise Wi-Fi profile with no certificate validation, the Intel video-acceleration installer matching graphics by marketing name, and the resume hook ordering asserted by issues 8471, 8888 and 10375.
  3. The Haswell gap still owes the corpus a new record.
  4. pacman-file-exists-in-filesystem-omarchy expires at the next Omarchy release, per upstream commit f5194e3f.
  5. O7, teaching the lint to skip a warning and a labelled branch, is still open and now has a second argument for it: the baseline holds 157 records and most of the recent additions are records the audits improved.

Session of 2026-09-13 (fifth): the four merges O3 left behind

Four decisions the re-audit could not take for itself, because a slug and a category are identity rather than content and no verdict field reaches them. Corpus 492 to 489, ok 82 / corrected 407, from 1,507 distinct sources, with 175 cause_reconciled stamps. Three published pages now 404, which is the price of each retirement and was weighed one at a time.

1. omarchy-resume-hook-appended-after-filesystems retired

Re-audited on 2026-09-12 and returned reject: Omarchy's HOOKS+=(resume) drop-in landing last does not break resume, disproved live here where /sys/power/resume reads 253:0 against /dev/mapper/root. It was kept and rewritten that day on the operator's call, to preserve the URL. It is now merged into hibernate-resume-hook-missing-or-misordered, because its slug and title state the disproved claim and a reader arriving by that slug was told the wrong thing before reading a word. This reverses the 2026-09-12 keep decision, on the operator's instruction, and the URL dies with it.

The survivor already carried the ordering rule in its cause. What moved is the part that only the retired record held: a danger naming the three ways the fix circulating in issues 8471, 8888 and 10375 breaks a working machine, and those three issues as sources.

2. xwayland-apps-blurry-hidpi retired

Its 2026-09-13 re-audit found it near-identical to xwayland-blurry-on-fractional-scale: same symptom, mechanism, primary fix, danger, three shared sources. Three things it held that the survivor did not are now in the survivor's fix, ahead of the force-zero-scaling recipe:

A third record pointed at the retired slug. gtk-file-chooser-does-nothing-xwayland sent readers to it for the toolkit environment variables. It now names the survivor. That is the check worth running on every retirement and it is mechanical: scan symptom, cause, fix, verify, danger and title across the whole corpus for the dead slug before committing.

3. hyprland-055-lua-config-input-ignored moved, not merged

Filed under audio-input and about Hyprland input configuration, which its auditor called a filing error rather than a distinction. Moved to hyprland-config. Its slug and URL are unchanged, so nothing 404s, which is what makes a category move the cheap decision of the four.

Not merged into keyboard-layout-not-applied-hyprland, now its sibling in that category, because the audit found them overlapping and not duplicated: this record holds the Lua migration itself, the require and split-file mechanics, the gesture syntax move and the 0.54-and-older branch, while the sibling holds the layout specifics. A cross-reference was added instead.

audio-input 35 to 34, hyprland-config 39 to 40.

4. nvidia-black-screen-external-after-suspend retired, and this one is a consolidation

Worth recording honestly, because the premise changed under the decision. The auditor found it the same problem as nvidia-suspend-resume-black-screen-vram and that record strictly better, then rewrote it into display-side triage that deferred to the sibling for the memory question. After that rewrite it was no longer a duplicate, so retiring it removes no redundancy. It consolidates: its triage is now the first section of the survivor's fix, ahead of the driver checks, because those steps are cheaper and settle the commonest case, a panel in DPMS off or a monitor re-enumerated at 0x0.

One defect goes with it rather than being fixed: it carried systemd-boot and grub in applies_to, wrong for the Omarchy branch, and no verdict field reaches applies_to. The survivor's list is already correct.

What this says about the schema

Four decisions, and the reason all four needed a human is the same: slug, title, category and applies_to are the fields an auditor can disprove and cannot touch. corrected_severity, corrected_frequency and sources_remove were added on 2026-09-11 for exactly this reason, and the remaining four are the ones that force a merge instead of a verdict. Two of them, slug and category, arguably should stay that way: changing a slug is a URL retirement whatever else it is, and that is a decision rather than a correction. title and applies_to are not, and a future verdict schema could carry both.

Where to pick this up

  1. The dashes, item 6: 1,366 em and en dashes across 335 of 489 records. The next corpus job, and the 150-title pass of 2026-09-03 is the precedent.
  2. Three upstream reports are owed and none is written: the enterprise Wi-Fi profile with no certificate validation, the Intel video-acceleration installer matching graphics by marketing name, and the resume hook ordering asserted by issues 8471, 8888 and 10375. The third is now better evidenced than it was, because the record asserting it has been retired for being wrong.
  3. The Haswell gap still owes the corpus a new record.
  4. pacman-file-exists-in-filesystem-omarchy expires at the next Omarchy release: upstream commit f5194e3f replaces the inline pacman call with a bin/omarchy-update-pacman wrapper run under systemd-run --scope, absent from 4.0.2-1.

Session of 2026-09-13 (fourth): O3 finishes, and the last 13 were wrong too

13 records, 7 agent batches, omarchy-core and display-monitors. 13 corrected, all at high confidence. The corpus is 492 records, ok 82 / corrected 410, from 1,510 distinct sources, with 177 cause_reconciled stamps. 13 citations removed.

O3 is complete. Every ok record that carried a danger and applied to Omarchy has now been through a second pass: 159 records across ten categories, of which 157 needed correcting. Two survived, mt7921e-dead-after-suspend-aspm and shell-section-override-ignored-without-colors-toml. The 8 records still matching ok plus danger are those two plus six the earlier passes judged general Arch rather than Omarchy-specific.

A first-pass note that was wrong when it was written

Three records in this batch carried detailed first-pass notes claiming line-by-line verification. Two held up: omarchy-update-requires-free-space and omarchy-update-lock are byte-identical to quattro on 4.0.2-1, so their thresholds, messages and env-var escapes all still stand. What those notes missed was scope rather than fact, and the gap was serious in both cases.

The third was not stale. It was wrong at the time. aur-updates-silently-skipped carried a note calling omarchy-update-aur-pkgs "a five-line script whose else-branch prints". It is neither five lines nor that command: it runs yay -Sua --noconfirm --cleanafter --ignore gcc14,gcc14-libs, and that --ignore predates the first audit by seven months. It makes the record's verify step false, because gcc14 and gcc14-libs are real AUR packages that stay in yay -Qua forever by design. A confident first-pass note is not evidence, which is the whole premise of O3 and is now demonstrated rather than assumed.

The free-space check measures the wrong filesystem

update-blocked-insufficient-free-space is about a 10 GiB threshold on /. The check reads / only, so a full ESP sails through it and the update dies later with the same No space left on device. On Omarchy the ESP fills from the very snapshots the record is about, because limine-snapper-sync copies each snapshot's kernel there under LIMIT_USAGE_PERCENT=85. The corrected record splits root from /boot and reclaims ESP space by deleting snapshots rather than files, and warns that paccache -rk1 spends the cache that omarchy-update-pkg-prune treats as the only offline downgrade path.

stale-omarchy-update-lock had the same shape of defect in its safety step: pgrep -af '/usr/bin/pacman' matches nothing during a real upgrade, because omarchy-update-system-pkgs runs sudo env ... pacman -Syu so argv[0] is bare pacman. It never mentioned /var/lib/pacman/db.lck, which is the authoritative live-transaction indicator and the one Omarchy's own omarchy-migrate polls for 900 seconds.

A recovery that locks you out of the recovery

omarchy-restart-shell on 4.0.2-1 no longer merely refuses on a locked session. It sets relock=1 and polls for 30 seconds until the session is secure again, so the obvious fix for a stale lock, restarting the shell over ssh, hands you a fresh lock you still cannot release. This is now in CLAUDE.md's domain facts, because it is a trap for the test VMs and the agentic lane and not only for a reader. The same audit found ~/.cache/quickshell/crashes does not exist on a working machine and the Wayland-fatal path writes no report by design, so the record's step 1 pointed at an empty directory.

A fix that burns a keyslot for nothing

bluetooth-keyboard-cannot-unlock-luks offered systemd-cryptenroll with FIDO2 as an alternative unlock. That writes a systemd-fido2 LUKS2 token which only systemd-cryptsetup consumes, and Omarchy 4 assigns the busybox encrypt hook wholesale in omarchy_hooks.conf. /usr/lib/initcpio/install/encrypt copies none of the token plugins, libfido2 or ask-password units that sd-encrypt would, so the reader spends a keyslot and gets nothing. The record also guessed /dev/nvme0n1p2 where /proc/cmdline carries cryptdevice=PARTUUID=...:root.

secure-boot-blocks-omarchy-install reached the right conclusion from two wrong premises: nothing on Omarchy 4 is sealed to the TPM, and CONFIG_MODULE_SIG_FORCE is unset with /sys/kernel/security/lockdown reading [none], so unsigned DKMS modules are not what Secure Boot rejects. The unsigned Limine BOOTX64.EFI and the UKI are. Its danger also omitted BitLocker, which is the one step in that record that actually locks a reader out of another operating system, and it offered no recovery at all.

The distribution already self-heals one of these

monitor-powered-off-at-boot-black-0x0 describes a panel that was asleep at boot returning a mode-less EDID. Omarchy 4 already recovers it: omarchy-hyprland-monitor-watch, running here as PID 2316 from autostart.lua, calls recover_modeless and reloads Hyprland on a backoff capped at 60 seconds until a mode appears, and Hyprland itself retries three times at one second. The record's off/on/detect sysfs dance is also actively risky: per drm_sysfs.c, only detect is needed, while off and on set DRM_FORCE_OFF and DRM_FORCE_ON and stick.

The lint caught a defect an auditor introduced

bluetooth-keyboard-cannot-unlock-luks came back with sudo omarchy-bluetooth-power on in its fix. sudo strips OMARCHY_PATH, so every omarchy subcommand then fails with find: '/themes/': No such file or directory, which is exactly what the sudo-omarchy-cmd lint pattern exists to catch. The script calls rfkill unblock bluetooth itself and takes no sudo, and the sibling record bluetoothctl-no-default-controller, corrected the same day, already called it without. Fixed by hand after the merge and recorded in the audit_note.

That is the first time in this programme the lint has caught a new defect rather than an inherited one, and it is worth setting against O7: the same check that has been collecting false positives all session also caught the one thing seven auditors' worth of prose got wrong.

Two domain facts changed, both outside the corpus

Where to pick this up

  1. O3 is done. The next corpus job is item 6, the dashes: 1,367 em and en dashes across 336 of 489 records. It edits the source of truth and must regenerate research/docs/ in the same commit, and the 150-title pass of 2026-09-03 is the precedent to copy.
  2. Four operator decisions are banked, none carryable by a verdict field, all wanting a merge or a rename on the mDNS precedent: the resume-hook slug that names a non-defect, xwayland-apps-blurry-hidpi duplicating xwayland-blurry-on-fractional-scale, hyprland-055-lua-config-input-ignored filed under audio-input, and nvidia-black-screen-external-after-suspend being the same problem as nvidia-suspend-resume-black-screen-vram, which is strictly better. That last record's applies_to also still carries systemd-boot and grub, which are wrong for Omarchy.
  3. Three upstream reports are owed and none is written: the enterprise Wi-Fi profile with no certificate validation, the Intel video-acceleration installer matching graphics by marketing name, and the resume hook ordering asserted by issues 8471, 8888 and 10375.
  4. The Haswell gap still owes the corpus a new record.
  5. One record is already known to expire: upstream commit f5194e3f, landed 2026-09-13, replaces the inline pacman call in pacman-file-exists-in-filesystem-omarchy with a bin/omarchy-update-pacman wrapper run under systemd-run --scope. It is absent from 4.0.2-1, so the record goes stale at the next release.

Session of 2026-09-13 (third): O3 clears wayland-compat and audio-input, all 18 wrong

18 records, 9 agent batches, two categories in one run. 18 corrected, 16 at high confidence and 2 at medium. The corpus is 492 records, ok 95 / corrected 397, from 1,474 distinct sources, with 167 cause_reconciled stamps. 8 citations removed.

That closes the four categories that held real work. The O3 backlog is 21, and 14 of those are records an earlier pass judged not Omarchy-specific or that have already been through a second pass.

A file nobody loads

xwayland-apps-blurry-hidpi told the reader to put hl.env() lines in ~/.config/hypr/envs.lua. That file is never loaded. Both the shipped /usr/share/omarchy/config/hypr/hyprland.lua and the live user copy require exactly five user modules, hypr.monitors, hypr.input, hypr.bindings, hypr.looknfeel and hypr.autostart, and hypr.envs is not among them. Worth carrying forward as a check on any future record that writes Hyprland environment variables: name the file that requires it, or the edit does nothing.

Four of the five variables in that block are shipped by Omarchy anyway and live in systemctl --user show-environment, so the headline fix was a no-op twice over. Only SDL_VIDEODRIVER was missing.

Two fixes that would have destroyed a working file

A record that sends the reader to a bootloader the machine does not have

bluetoothctl-no-default-controller told the reader to edit /boot/loader/entries/*.conf or /etc/default/grub. Omarchy 4 has neither: no grub package, no systemd-boot, no /etc/default/grub. It boots a UKI through Limine with cmdline drop-ins in /etc/limine-entry-tool.d/. The corrected fix uses /etc/modprobe.d/btusb-autosuspend.conf instead, which removes the unbootable-machine danger entirely rather than relocating it.

The same record missed Omarchy's design: the rfkill soft block is how Omarchy stores the Bluetooth on and off state, through omarchy-bluetooth-power, called from the shell panel, and bluetoothctl power on fails outright while it is set. So the recovery is omarchy-bluetooth-power on, and a reader following the old record would conclude their adapter was broken.

Guessed causes stated as fact

chromium-xwayland-nvidia-gpu-crash names a GPU cause and recommends an X11 workaround. The cited issue is still open with no root cause, and its only reproduction comment gets the same SIGTRAP under --ozone-platform=x11 and with extensions disabled, which points at the --oauth2-client-* lines that omarchy-install-chromium-google-account appends rather than at the GPU. On Omarchy the flags file already carries --ozone-platform=wayland, so appending the X11 switch leaves two contradictory ones.

wine-proton-native-wayland-driver was exactly right about Wine, confirmed against wine 11.16 source rather than the wiki, and wrong about Proton: PROTON_ENABLE_WAYLAND appears nowhere in Valve's Proton. It is a GE-Proton option, also read by proton-cachyos, so the record's "Proton 10 or Experimental" version floor described a setting stock Proton silently ignores.

What held, which is worth recording too

Not everything drifted. The PipeWire config paths did not move: WirePlumber 0.5.15 ships the crackling record's exact stanza at /usr/share/wireplumber/wireplumber.conf.d/alsa-vm.conf, both api.alsa.period-size and api.alsa.headroom are present in the installed libspa-alsa.so, and ~/.local/state/wireplumber/ holds the four state files the record names. The echo-cancellation record's four webrtc.* keys, the rnnoise label and three control names all verified against the shipped libraries. OBS 32 has not replaced v4l2loopback either: linux-v4l2.so still gates on modinfo v4l2loopback and still carries the exclusive_caps=1 autoload line.

The recurring correction in this pair of categories was not staleness. It was that Omarchy already does the work: omarchy restart audio does the three-service restart plus stuck-USB recovery, omarchy-capture-screenshot already handles multi-monitor by dividing by scale and swapping for transform, bluetooth-a2dp-autoconnect.conf is already installed into the directory one record says to create, and gtk-primary-paste.sh already sets the GTK key another record prescribes.

Two operator decisions, banked and not acted on

Neither can be carried by a verdict field, which is the same wall as the resume-hook slug.

  1. xwayland-apps-blurry-hidpi duplicates xwayland-blurry-on-fractional-scale in hyprland-config: same symptom, mechanism, primary fix, danger and three shared sources. Its one distinct contribution is a make-it-native-Wayland branch that is mostly already default. The auditor recommends merging it into the sibling, carrying that branch and the SDL_VIDEODRIVER gap. Its own companion, gdk-scale-mismatch-oversized-xwayland, is not a duplicate: inverse symptom, Omarchy-specific cause, and it holds the Flatpak override, the Steam launch options and three upstream issues the sibling lacks.
  2. hyprland-055-lua-config-input-ignored is filed under audio-input and is not about audio. It is Hyprland input configuration, like its sibling keyboard-layout-not-applied-hyprland in hyprland-config. The auditor calls the split a filing error rather than a distinction, and says the two overlap without duplicating: this one uniquely holds the Lua migration itself, the require and split-file mechanics and the 0.54-and-older branch.

One agent died on a session rate limit, after doing the work

The ninth batch was killed by a 429 while composing its reply, so it reported nothing. Its two verdict files were already written, timestamped after every other batch's, with 4,900 and 5,600 character reasons and full corrected fields. They were read and merged on their contents rather than on a summary, which is the right order anyway. Worth knowing for the next long run: a dead agent is not necessarily lost work, so check the output directory before re-running a batch and paying for it twice.

Where to pick this up

  1. O3 has 21 records left and only 7 of them are new work: omarchy-core 8 minus the 6 already through a second pass, plus display-monitors 5. The rest are gpu-drivers 4, boot-kernel 2, omarchy-theming 1 and network 1, all previously audited or judged not Omarchy-specific. O3 is effectively done after display-monitors and the omarchy-core remainder.
  2. Three operator decisions are now banked: the resume-hook slug, the XWayland duplicate and the audio-input filing error. All three want a merge or a rename, and the mDNS precedent on 2026-09-11 is the shape to copy.
  3. Three upstream reports are still owed and none is written: the enterprise Wi-Fi profile with no certificate validation, the Intel video-acceleration installer matching graphics by marketing name, and the resume hook ordering asserted by issues 8471, 8888 and 10375.
  4. The Haswell gap still owes the corpus a new record.

Session of 2026-09-13 (second): O3 clears hyprland-config, and all 13 were wrong

13 records, 7 agent batches. 13 corrected, every one at high confidence, the first category where nothing survived and nothing was hedged. The corpus is 492 records, ok 113 / corrected 379, from 1,436 distinct sources, with 154 cause_reconciled stamps. 14 citations removed.

An auditor ran a command on the live desktop, and the brief is now stricter

While probing how hyprctl handles arguments, one auditor ran hyprctl reload badarg. hyprctl ignores the junk argument and performs the action, so that was a real reload of the operator's running session. The config was unchanged, so it re-read the same files and nothing moved, and the auditor reported it unprompted at the top of its reply rather than burying it.

The brief already said to change nothing and named the lock screen. What it did not say is that probing a command is running it, which is the assumption that produced this. reaudit-brief.md now says so, and names hyprctl reload, hyprctl keyword and hyprctl dispatch explicitly, since the corpus's own domain facts already record that hyprctl dispatch evaluates its input as Lua. Self-reporting is the behaviour to keep: an auditor who hides a breach is worse than one who commits it.

A record that can take the machine down, found by reproducing it

lua-config-getters-infinite-loop-at-load was rated medium and describes a Lua getter hanging the keybindings menu. The auditor lifted the stub out of omarchy-menu-keybindings, ran it in isolation on Lua 5.5.1, and watched it pass a million ipairs iterations with #stub at zero. The compositor survives only because config evaluation is watchdogged at LUA_TIMEOUT_CONFIG_RELOAD_MS = 1500. The menu is not, and every unanswered SUPER+K leaks another unbounded lua process, which walks to a kernel OOM. The record named three getters. There are more than twenty hl.get_* functions and they all do it. Severity raised to high.

Two records were backwards, not merely stale

Four fixes aimed at a version that has moved

nvidia-hyprland-modeset-cursors-mgpu was the worst of them and every defect was a drift, not an error at the time of writing. /sys/module/nvidia_drm/parameters/modeset is mode 0400, so the symptom, the fix and the verify all hand the reader a command that returns Permission denied unprivileged. The in-driver default is now true, so the modeset headline is largely obsolete. NVreg_PreserveVideoMemoryAllocations was superseded by NVreg_UseKernelSuspendNotifiers=1 on 595+, and the three nvidia sleep services are disabled here by upstream design, so the record's "must be enabled" would be a regression. And cursor:use_cpu_buffer already auto-enables on NVIDIA, so that step was a no-op.

Same shape elsewhere: hyprpm's state moved to /var/cache/hyprpm/<username>, so the record's rm -rf ~/.local/share/hyprpm ~/.cache/hyprpm deletes nothing, and hyprpm purge-cache is the current command. debug:error_limit = 5 is already the default, so the config-error record's advice to cap it was a no-op. And hl.env() takes a third boolean on 0.56.2 that makes Hyprland run systemctl --user import-environment itself, which the record never mentions while recommending the manual commands instead.

Work the distribution already does, again

This is the fourth category in a row where records told readers to do something Omarchy ships. /usr/share/omarchy/default/hypr/autostart.lua already runs systemctl --user import-environment $(env | cut -d'=' -f 1) on hyprland.start, and measured here, systemctl --user show-environment carries nine variables that hl.env() set and that are not in Hyprland's seven-name allowlist. So the symptom is usually absent on a stock install and the record told Omarchy readers to add a command Omarchy already runs. Likewise omarchy-hyprland-reload-guard pause|resume supersedes the hyprctl eval pause pair, and Omarchy's own omarchy.keyboard-layout bar widget answers the keyboard record's problem with a click.

The share picker record was right about the mechanism and wrong about the software

selection -1 is SSelectionData.type cast to int with TYPE_INVALID = -1, and the cited issue supports the record line for line. But the qt6-wayland diagnosis is attached to the wrong picker: Omarchy configures hyprland-preview-share-picker, which is GTK4 plus gtk4-layer-shell and links no Qt at all, while the Qt6 dependency belongs to the stock hyprland-share-picker. And pacman -S hyprland-preview-share-picker works only on Omarchy, where it comes from the [omarchy] repo, so the record's claim that the fix applies on four other distributions is false.

The lint baseline is drifting towards meaning "checked", not "known bad"

Seven records were added to the baseline across this session's three categories, and every one of them is a record the audit improved. They trip the lint because a better-written fix names the dangerous command in order to warn against it, quotes what an Omarchy script itself runs, or labels a plain-Arch branch that legitimately uses hyprlang. A regex cannot read intent, so this is the lint working as designed and the baseline is where it belongs.

It is still worth naming, because the baseline now carries 157 records and its meaning has shifted. A cheap improvement, not done here: skip a line whose sentence contains a negation near the pattern, or that sits under a **Plain Arch label, and count the remainder. That would cut the false positives without weakening the check. Recorded as O7 under "What's left".

Where to pick this up

  1. O3 continues, 39 records left: wayland-compat 9, audio-input 9, omarchy-core 8, display-monitors 5, gpu-drivers 4, boot-kernel 2, omarchy-theming 1, network 1. The last four are records already through a second pass, or ones an earlier pass judged not Omarchy-specific, so the real work is the first four categories, 31 records.
  2. Three upstream reports are owed, none written: the enterprise Wi-Fi profile with no certificate validation, the Intel video-acceleration installer matching graphics by marketing name, and the resume hook ordering asserted by issues 8471, 8888 and 10375. A fourth is now arguable: hyprpm-plugin-permission-popup-every-time found that Omarchy ships ecosystem:enforce_permissions off, which is a deliberate choice worth confirming rather than a defect worth reporting, so read it before writing anything.
  3. The Haswell gap still owes the corpus a new record.

Session of 2026-09-13: O3 clears omarchy-theming, and a second record passes

15 records, 8 agent batches. 14 corrected, 1 ok, every verdict at high confidence. The corpus is 492 records, ok 126 / corrected 366, from 1,390 distinct sources, with 143 cause_reconciled stamps. 20 citations removed, the most of any run, almost all of them basecamp/omarchy URLs that the rename left pointing at the Omarchy 3 tree.

The second record to survive the audit

shell-section-override-ignored-without-colors-toml came back ok with nothing rewritten but its source URLs. Every mechanical claim holds at the exact line numbers it cites, shell.lock.toml is a real shipped file, the 13-section list matches the generated shell.toml on this machine exactly, and the behaviour is now documented upstream in docs/theming.md on quattro, which the record predates. That is two clean passes in 115 records.

An auditor reproduced a failure and found the record's own fix still broken

installed-theme-without-colors-toml-breaks-foot says a theme with no colors.toml leaves the terminal unable to start, and gives a three-key colors.toml as the repair. The auditor built that exact file in a sandboxed HOME, ran the template generator, and foot still refused, because omarchy-theme-color --all returns empty strings for the six hues and the generated foot.ini carries regular1= with no value. Nine keys are needed. This is the first verdict in the whole programme where the auditor exercised the record's own fix rather than reading it, and the fix failed. It is worth repeating deliberately where a fix is cheap to run in a scratch HOME.

The same record was also missing the recovery that matters. A reader whose terminal will not start cannot follow a fix that assumes a terminal, and Omarchy binds SUPER+SHIFT+CTRL+SPACE to the theme menu, which needs no terminal at all.

Walker is gone from Omarchy 4, and the answer was not to delete the records

Two records are about the Walker launcher. walker and elephant are not installed, nothing in the quattro tree mentions either, and SUPER+SPACE is omarchy-menu toggle. DHH closed issue 6443 with "Walker is gone on quattro". The obvious move was a reject each, and the auditor declined it on evidence: walker 2.17.0-1 and elephant 2.22.0-1 are current in the AUR, both released 2026-07-16, and Waiting for elephant... is still in Walker's own theme XML, so a user who installs it themselves on Omarchy 4 can still hit both problems. Both records now label Walker as something the user added and say what Omarchy 4 uses instead. That is the right shape for the hypridle case too, and it is a better default than retiring a record because the distribution moved on.

A fix that reached nothing, and one that reached too much

Two recovery commands that would have hit the wrong target

theme-set-hangs-with-brave-origin-running told the reader to run pkill -f 'brave --refresh-platform-policy --no-startup-window'. That pattern is unanchored and matches the shell the reader types it into, which the auditor reproduced with pgrep -af. It is now an anchored pgrep and a kill by PID. And app-menu-broken-icons-after-quattro-upgrade told the reader to restore PNGs into a directory Omarchy no longer maintains, when omarchy-settings 4.0.2-1 ships 18 of those 20 icons as themed icons under /usr/share/icons/hicolor/*/apps/, so rewriting Icon= to the name is the durable repair. Only GitHub.png and Battle.net.png have no packaged equivalent.

What the category was actually full of

The lint's false-positive rate is now worth naming

Three of this session's four baseline additions are records that name a dangerous command in order to warn against it, or quote what an Omarchy script itself runs. bare-pacman-Sy inside "never turn one of them into pacman -Sy <pkg>", mkinitcpio-P inside "do not use sudo mkinitcpio -P here", and sudo-pacman-Syu inside a cause quoting omarchy-update-system-pkgs verbatim. The lint is doing its job, since a regex cannot read intent, and the baseline is the designed place for this. It is worth knowing that a better-written record trips it more often, not less.

Where to pick this up

  1. O3 continues, 52 records left: hyprland-config 13, wayland-compat 9, audio-input 9, omarchy-core 8, display-monitors 5, gpu-drivers 4, boot-kernel 2, omarchy-theming 1, network 1. The last four categories are records already through a second pass, or records the earlier passes judged not Omarchy-specific, so the real remaining work is the first five.
  2. Three upstream reports are owed, none written: the enterprise Wi-Fi profile with no certificate validation, the Intel video-acceleration installer matching graphics by marketing name, and the resume hook ordering that issues 8471, 8888 and 10375 all assert and this corpus disproved.
  3. The Haswell gap still owes the corpus a new record.

Session of 2026-09-12/13: O3 clears power-suspend, and one record is a non-problem

16 records, 8 agent batches. 15 corrected, 1 rejected. 13 verdicts at high confidence, 2 at medium. The corpus is 492 records, ok 140 / corrected 352, from 1,361 distinct sources, with 130 cause_reconciled stamps. 8 citations removed.

The reject was kept and rewritten, not retired

omarchy-resume-hook-appended-after-filesystems says Omarchy's HOOKS+=(resume) drop-in lands after filesystems and breaks hibernation. The drop-in mechanics in it are right. The defect is not real. sort -zVu concatenation does put resume last at position 15, and last still runs before root is mounted, and it still lands after encrypt, which is the only ordering the Arch wiki requires. The wiki's own example, cited by this record, puts resume after filesystems. Disproved live on this workstation: /sys/power/resume reads 253:0, matching /dev/mapper/root, so the hook resolved the LUKS device from last position. Issue 8471's author says outright he could not demonstrate a failure from the ordering, and 8888 and 10375 repeat the same mechanism with nothing confirmed.

Worse, the record's own fix is destructive twice over. Deleting omarchy_resume.conf breaks omarchy-hibernation-available and omarchy-hibernation-remove, both of which grep that exact path for ^HOOKS+=(resume)$. And copying the grep -h '^HOOKS' output into a hand-written array reinstates the kms hook that omarchy_hooks.conf strips on NVIDIA-only machines, which pulls nouveau and about 100 MB of GSP firmware back into the initramfs.

A reject deletes a record and 404s a published page, which is the same shape as the mDNS merge and belongs to the operator. The operator's call was to keep the URL and rewrite the record, so it was hand-written rather than merged: the title now says the hook landing last is not the bug, the fix tells the reader to change nothing and points at the four records that hold the real causes, and the danger carries the three ways the circulating fix damages a machine. The auditor's verdict is reproduced verbatim in the audit_note beneath a line saying the record was a reject that was kept. audit_status is corrected, so corpus counts move by one: ok 140 / corrected 352.

The slug still names the non-defect, and no field can change that. It is the same wall the mDNS record hit on 2026-09-11, where the answer was a merge. Anyone reaching this record by its URL or its slug arrives expecting the fix it no longer gives.

The category was full of Omarchy 3

Two records described a mechanism that does not exist on this distribution any more. omarchy-hypridle-timeouts-too-aggressive is built entirely on hypridle and ~/.config/hypr/hypridle.conf. hypridle is not installed, nothing owns /usr/bin/hypridle, and the quattro tree has no hypridle path at all. Omarchy 4 does idle in a Quickshell service, /usr/share/omarchy/shell/plugins/services/idle/Service.qml, reading idle.screensaver and idle.lock from ~/.config/omarchy/shell.json with defaults of 150 and 300 seconds. The record said 152 for the lock. suspend-blocked-by-inhibitor-lock had the same problem from the other end: every systemd claim in it held on systemd 261.2-1, and every Omarchy claim in it was Omarchy 3. Its two cited basecamp/omarchy/master raw URLs still serve the Omarchy 3 files, which is where the error came from, and that is worth recognising as a source trap rather than an auditor's slip: a raw URL against master returns content, so nothing looks broken.

This is the gap the O4 issue-tracker audit named on 2026-09-11 and could not fix, a record sending readers to a hypridle config file that does not exist on Omarchy 4. It is now fixed.

Three fixes that break the thing they promise to protect

Hibernation on Omarchy 4 is opt-in, and two records assumed otherwise

omarchy-hibernation-setup is referenced in three places in the quattro tree and in no installer script, so a stock install has no hibernation and the path in is omarchy hibernation setup. On top of that, omarchy-hibernation-available requires /etc/mkinitcpio.conf.d/omarchy_resume.conf to exist as well as the non-zram swap sum exceeding /sys/power/image_size, so a reader who grows swap by hand never sees the Hibernate entry appear and has no way to tell why. Neither record said so. hibernate-blocked-by-zram-only-swap also pointed at /etc/systemd/zram-generator.conf, which is the lowest-precedence file of the three, loses to Omarchy's own 90-omarchy.conf drop-in, and is deleted outright by migration 1785013000.sh on the next omarchy update.

Smaller things worth keeping

A field no verdict can correct, again

hibernate-blocked-by-zram-only-swap carries laptop in applies_to and not desktop, and the problem applies equally on a desktop. There is no corrected_applies_to, so the auditor could say so only in prose. This is the third time the verdict schema has been short of a field that an auditor needed, after corrected_severity and corrected_frequency on 2026-09-11 and sources_remove the same day. The pattern is clear enough now to act on rather than to keep recording: any reader-facing field the auditor can disprove needs a verdict key. applies_to, tags and title are the remaining three, and title and slug are the ones that force a merge instead, as the mDNS record showed.

Where to pick this up

  1. O3 continues, 53 records left: omarchy-theming 15, hyprland-config 12, wayland-compat 9, audio-input 9, display-monitors 5, omarchy-core 2, network 1. That last one is mt7921e-dead-after-suspend-aspm, already re-audited and confirmed.
  2. Still owed, and carried from 2026-09-11: the enterprise Wi-Fi upstream report, and a new corpus record for the Intel video-acceleration installer matching graphics by marketing name. The Haswell comment itself was posted on omacom/omarchy#7866 on 2026-09-11.
  3. A third upstream report is now owed, and it is the cheapest of the three: omacom/omarchy 8471, 8888 and 10375 all assert the resume hook ordering defect that this session disproved, and 8888 is an open pull request that would move the hook. The evidence is in this session's audit_note and in raw/o3-power-suspend-audit.json.

Session of 2026-09-11 (fifth): the mDNS merge, one record removed and one split out

The decision the network audit left open. Both mDNS records had been re-audited that day and both came back corrected, so this is not a cleanup of bad content. It is a merge done because one of them carried a slug and title asserting the cause its own audit disproved, that ufw blocks 5353, and no verdict field can rename either.

Corpus stays at 492 records: mdns-local-hostnames-fail-ufw-blocks-5353 removed, kde-connect-ports-blocked-by-ufw added in its place. apps-services 59 to 58, network 46 to 47. Sources 1,324 to 1,326.

What moved where

mdns-local-hostname-not-resolving survives, moves from apps-services to network, and is retitled to cover both halves. It absorbed three things:

kde-connect-ports-blocked-by-ufw is new, in network. KDE Connect was the one part of the removed record that survived its audit intact and is not mDNS at all, so it gets its own record rather than riding along inside a record about name resolution. Every claim in it was re-checked while splitting: the 1714 to 1764 range on both protocols is stated by KDE's own documentation, retrieved 2026-09-11, install/config/firewall.sh opens only LocalSend's 53317, ufw status on this machine lists no rule for the range, and /etc/ufw/after.rules:27 is -A ufw-after-input -m addrtype --dst-type BROADCAST -j ufw-skip-to-policy-input under the comment don't log noisy broadcast, which is why the discovery packets die with no log line.

Provenance is kept rather than tidied away

Both records' audit_note fields name the removed slug and say what moved and why. Reader-facing fields point only at records that exist, checked mechanically. The removed record's own text is not lost either: it is in the git history and in raw/o3-network-audit.json, which holds the verdict that disproved its cause with the measurements behind it.

What this does to the public site

The removed record had a URL on the published site and that URL now 404s. That is the cost of retiring a record whose title states something false, and it is the right trade. Nothing internal dangles, because the site and the search index are regenerated wholesale from the JSONL on every build.

Where to pick this up

  1. O3 continues, 69 records left: power-suspend 16, omarchy-theming 15, hyprland-config 12, wayland-compat 9, audio-input 9, display-monitors 5, omarchy-core 2, network 1.
  2. Two upstream reports are owed, and neither is written: the enterprise Wi-Fi profile that sets no certificate validation, and the Intel video-acceleration installer that matches graphics by marketing name so Haswell matches nothing.
  3. The Haswell gap also still owes the corpus a new record.

Session of 2026-09-11 (fourth): O3 clears network, and one record finally passed

15 records, 8 agent batches. 14 corrected, 1 ok. The corpus is 492 records, ok 156 / corrected 336, from 1,324 distinct sources, with 119 cause_reconciled stamps. 13 citations removed, the most of any run.

The first clean pass, after 51 corrections

mt7921e-dead-after-suspend-aspm came back ok with nothing rewritten. modinfo -p mt7921e still prints exactly one parameter on kernel 7.1.9, __setup("mem_sleep_default=", ...) is still in kernel/power/suspend.c, both forum threads carry the record's exact log lines and both its fixes, and the Omarchy boot path in it is right, because limine-mkinitcpio alone re-reads the cmdline and re-registers the UKI. Its boot-limine-conf lint hit is a false positive: it names that file only to warn against editing it.

That matters more than another correction. It is the first evidence a record can survive this audit, which is what makes the other 51 verdicts mean something rather than reading as a reviewer finding faults to look useful.

A published record's title asserts a cause that is false

mdns-local-hostnames-fail-ufw-blocks-5353 blames ufw for blocking mDNS. It does not. /etc/ufw/before.rules:68 accepts 224.0.0.251:5353 and before6.rules:136 the ff02::fb equivalent, both [unmodified] per pacman, and that accept has shipped since ufw 0.30.1 in March 2011, so no version drops it. Measured here with no 5353 rule anywhere in user.rules: avahi-browse lists the LAN, lpstat -e finds the printer, getent hosts truenas.local resolves. So the LocalSend and printer symptoms are misattributed too.

What survives is narrower and worth keeping: unicast replies are dropped, which Avahi never asks for (avahi-core/query-sched.c:217 passes unicast_response=0), the commented-out MULTICAST RETURN is the real trap, and KDE Connect's 1714 to 1764 genuinely is closed.

A merge is owed and is NOT done, because it deletes a record. The auditor's recommendation, which this session is recording rather than executing: this is one problem, mdns-local-hostname-not-resolving should survive, because this slug and title assert the disproven cause and no verdict field can rename either, and the survivor must absorb three things from here, the negative firewall finding with the exact rule text and a "do not open 5353" instruction, the ufw-not-local ordering trap with its danger clause, and the unicast-reply caveat. KDE Connect's ports belong in a new network record, since KDE Connect is not mDNS. The merge is also a category move out of apps-services.

Two of this session's own hypotheses were wrong, and both were useful

The auditors were told that Omarchy's NetworkManager global-dns override probably explained the VPN DNS records, and that ufw's default-deny probably explained the path MTU blackhole. Both were checked and rejected. The override is real and does outrank per-connection DNS, but neither Tailscale nor WireGuard goes through NetworkManager, and the deciding fact is the route-only domain on the tunnel link in systemd-resolved. And /etc/ufw/before.rules:34 already accepts icmp destination-unreachable, which is what carries fragmentation-needed. Both corrected records now say so explicitly, so the next reader does not spend an hour there.

What the category was actually full of

A security finding against the shipped product, not the corpus

Omarchy's own network panel builds enterprise Wi-Fi profiles without setting 802-1x.ca-cert or 802-1x.domain-suffix-match (Panel.qml, Model.js:319-326), so an unvalidated enterprise profile is the shipped default rather than a reader's mistake. The auditor strengthened that record's danger instead of softening it. Worth an upstream report. The same record passed the password in argv, which Omarchy explicitly avoids with a source comment saying argv is world-readable in /proc, and the corrected fix adopts Omarchy's stdin pattern.

The brief told auditors they had no sudo, and that is no longer true

sudo -n now succeeds without a password on this workstation. Every prompt in this run said so explicitly and told the agent not to use it, and reaudit-brief.md is corrected in this commit: it now names what must never be touched, including the lock screen that cannot be released headlessly, and says to report a check as not run rather than running it with root. Read-only discipline is the only thing that has kept this machine intact across 67 audited records today.

Where to pick this up

  1. O3 continues, 69 records left: power-suspend 16, omarchy-theming 15, hyprland-config 12, wayland-compat 9, audio-input 9, display-monitors 5, omarchy-core 2, network 1. That last one is mt7921e-dead-after-suspend-aspm, already re-audited and confirmed: the backlog is defined by audit_status: ok, so a record that passes keeps counting. The count will never reach zero and this journal is the record of what has actually been through a second pass.
  2. The mDNS merge, as recommended above. It deletes a record and moves a category, so it wants a decision rather than an agent.
  3. Still owed from the O4 audit: a new record for the Intel video-acceleration installer matching graphics by marketing name, where Haswell matches nothing.
  4. Two upstream reports are now owed: the enterprise Wi-Fi profile with no certificate validation, and the Haswell VA-API gap.

Session of 2026-09-11 (third): O3 clears apps-services, 23 more, and all 23 were wrong

The largest O3 category, and the one most likely to pass, because apps-services is mostly general Linux rather than Omarchy internals. 23 records, 12 agent batches. 23 of 23 came back corrected, all at high confidence. The corpus is 492 records, ok 170 / corrected 322, from 1,281 distinct sources, with 107 cause_reconciled stamps. 83 records remain in the O3 backlog.

The calibration mattered, and it worked

After 50 consecutive corrections a reviewer starts finding faults to look useful, so every prompt in this run said plainly that ok is a legitimate and expected verdict here, that an unnecessary rewrite is itself a defect, and that a record can be correct precisely because it is generic. The spread of what came back says it landed: two records lost only their danger, one only its fix, and three auditors reported claims they had expected to correct and did not after checking.

One of those is worth keeping as the model. The WirePlumber auditor was about to mark a configuration-shadowing claim wrong, because upstream's own documentation says fragments from all locations merge, then read lib/wp/base-dirs.c at the installed tag, found a same-named lower-priority file is dropped under a comment saying exactly that, and left the record's text alone. The record and the Arch wiki are right and upstream's prose is the misleading part.

Following two of these records broke a working machine

Commands that cannot do what the record says

Causes that were backwards

Two hazards nobody had flagged

omarchy-settings's post_install and post_upgrade copy its own nsswitch.conf over /etc, with an upstream comment saying customizations will be reset, and there is no .pacnew and nothing for pacdiff to catch. Any record editing that file has to say so. And snapper -c root delete can never reclaim space in @home or @pkg, which is where the space actually is on this layout, so a reader with no snapshots was being told to delete snapshots.

The lint caught a defect the audit introduced

docker-compose-against-podman-socket came back with sudo pacman -Syu podman, which carries both a sync and a sysupgrade flag and is therefore exactly what omarchy-update-pacman-guard aborts, in the same fix that explains the guard. Corrected by hand to pacman -S --needed and recorded in the record's own audit_note. This is the lint working on audit output rather than on harvest output, which is a use it had not had before. Two other new hits were read and kept: a plain-Arch branch and a warning telling the reader not to run the matched command.

A domain fact in CLAUDE.md was incomplete

An auditor pushed back on the claim that OMARCHY_PATH comes only from ~/.bashrc. It is also set by /usr/share/uwsm/env.d/10-omarchy, which sources default/bash/env-bootstrap, so systemctl --user show-environment does carry it in a graphical session and user units started there see it. Verified here and corrected in the domain facts. The ssh conclusion is unchanged, and that is the third of this session's own premises an auditor has corrected.

Where to pick this up

  1. O3 continues, 83 records left: power-suspend 16, omarchy-theming 15, network 15, hyprland-config 12, wayland-compat 9, audio-input 9, display-monitors 5, omarchy-core 2.
  2. network should go next despite not being the largest, because this run found its mdns-local-hostnames-fail-ufw-blocks-5353 record has a cause that contradicts the stock configuration: /etc/ufw/before.rules:68 already accepts multicast mDNS. Its sibling here recommends merging the two once that is settled.
  3. Still owed from the O4 audit: a new record for the Intel video-acceleration installer matching graphics by marketing name, where Haswell matches nothing and the machine ends up with no VA-API driver.

Session of 2026-09-11 (second): O3 clears gpu-drivers, and all 14 were wrong

O4 merged, so the next option is O3: the ok records that carry a danger and apply to Omarchy, which are the ones that can break a machine if the fix is wrong. gpu-drivers first, 14 records, seven agent batches, each handed reaudit-brief.md, the record JSON and an output directory.

14 of 14 came back corrected, all at high confidence. Every one had its fix, its verify and its danger rewritten. 11 causes were replaced and stamped, 4 frequencies corrected, and 12 cited URLs removed. The corpus is now 492 records, ok 193 / corrected 299, from 1,213 distinct sources, with 91 cause_reconciled stamps.

One defect family accounts for most of it

These records were generic Arch advice wearing an Omarchy label, and the tell is always the apply step. Omarchy 4 has no /etc/default/grub, /boot/limine.conf is regenerated on every kernel transaction, /etc/mkinitcpio.d/ is empty so mkinitcpio -P writes nothing, and the machine boots a UKI so there is no /boot/initramfs-linux.img to inspect and no fallback entry to boot. Records named all four.

The two critical records are the sharpest case: the fix caused the failure its own danger warns about. Both tell the reader to rebuild the initramfs with mkinitcpio -P, which on this layout writes no boot image, so the machine reboots into the old image with no driver. That is the black screen the danger tells you to avoid.

One fix was more dangerous than the problem. amdgpu-idle-freeze-gfxoff-ppfeaturemask carried a ppfeaturemask value copied from a forum ladder. Against the kernel 7.1 source it does not disable only GFXOFF: it re-enables PP_OVERDRIVE_MASK and PP_GFX_DCS_MASK, both off by default, and the Overdrive bit taints the kernel and prints a critical warning. The wiki's own caution is that enabling unstable PowerPlay bits causes flicker and broken resume, which is the failure being fixed. The auditor replaced the constant with one computed from the kernel's default.

Other findings worth keeping

The lint predicted 10 of the 14

Every one of the four AMD records carried a boot-limine-conf hit in data/lint-baseline.json, recorded when the lint landed and never followed up. That is the lint working as designed, as a candidate finder rather than an audit, and it gives an order for the rest of O3: of the 106 records left, 22 carry a hit, and the largest single pattern is 14 records still telling readers to run mkinitcpio -P.

This commit also clears one baseline entry, omarchy3-tree on omarchy-nvidia-580xx-target-not-found, because the audit fixed the Omarchy 3 path it was flagging. Clearing an entry is a deliberate act and belongs in the commit that re-audits the record, which is this one. Six new hits were added and every one was read first: three are warnings telling the reader not to run the matched command, two sit in branches explicitly labelled plain Arch, and sudo omarchy-apply-system is a false positive for that script, which sets its own OMARCHY_PATH default at line 83 rather than inheriting one.

Where to pick this up

  1. O3 continues. apps-services is the largest at 23, power-suspend 16, omarchy-theming and network 15 each. Prefer the 22 records the lint already flags.
  2. A new record is owed: the auditor of the VA-API record found a live upstream defect that is not in the corpus. Omarchy's video-acceleration installer matches Intel graphics by marketing name, and Haswell matches none of them, so those machines finish an install with no VA-API driver and the update path never re-runs the script.

Session of 2026-09-10 and 2026-09-11: O4 finished, 36 records merged, and all 36 were wrong

Steps 2 and 1 of the O4 pick-up list, in that order: the duplicates reconciled, then batch 01 read, then a second duplicate pass on 2026-09-11 that found three groups the first pass had missed, then the audit of all 36, then the merge. Every one of the 36 came back corrected. Nothing was clean and nothing was rejected, so the problems are all real and the detail was all wrong somewhere.

The corpus is now 492 records, ok 207 / corrected 285 / unaudited 0, drawn from 1,190 distinct sources, with 80 cause_reconciled stamps across five dates. O4 is done.

Every stage is banked separately, because each one is evidence about a different thing: the raw harvest in raw/issue-harvest-partial.json exactly as the workflow wrote it, the reconciled set in raw/issue-harvest-reconciled.json, batch 01 alone in raw/harvest-01.json, the audited set in raw/issue-harvest-audited.json, and what was actually fed to the merge in raw/issue-harvest-merge-payload.json.

1. Nine groups, found in two passes, and the first pass was not good enough

The banked payload listed six near_duplicates_to_reconcile pairs. Three of those pairs share a record, so they are five groups covering eleven records. A similarity pass over all 40, token-set Jaccard on title plus symptom plus cause at a 0.25 threshold, found a sixth group the harvest's own detector missed: the two plugin-lock records, from issues 7106 and 9441, which are one defect with two outcomes. Thirteen records became six.

That threshold was too high, and it cost a second pass. On 2026-09-11, while batching the records for audit, two of them turned out to be the same defect by reading their titles side by side: nvidia-env-forced-on-igpu-primary-hybrid-laptop and libva-driver-name-nvidia-forced-on-hybrid-laptop, both blaming /usr/share/omarchy/default/hypr/nvidia.lua, both citing commit 33d7363c, both fixed by the same hl.env override. They score 0.231. Re-running the sweep at 0.15, and adding two signals that do not depend on wording, found two more: the two migration 1786643346 records at 0.214 and the two VMware records at 0.160.

The two signals are worth keeping for any future dedup, because each one caught a pair on its own: a source URL shared between two records, and two or more shared file paths. Two harvesters writing up the same defect agree on the file they blame long before they agree on prose. After the second pass the only candidate left is two records that happen to cite the same release tag, which is not a duplicate.

Nine groups over nineteen records became nine records. 40 plus batch 01's six, minus ten, is 36.

grouprecords inkept
installer ESP mount failure2installer-mounting-the-esp-failed-squashfs-superblock
AMD AV1 firmware corruption2av1-video-color-corruption-linux-firmware-amdgpu-20260810-1
Codex limits unavailable3codex-limits-unavailable-agents-panel
Gemini default agent2default-agent-gemini-fails-antigravity-replacement
migration 17875159272omarchy-update-migration-1787515927-fails-bash-5-3
plugin write while locked2plugin-file-write-while-locked-strands-session
NVIDIA env on a hybrid laptop2nvidia-env-forced-on-igpu-primary-hybrid-laptop
migration 17866433462migration-1786643346-browser-window-open-loop
VMware shell black desktop2vmware-shell-black-desktop-qt-quick-dmabuf

The records outside those groups are carried through byte-identical and in their original order, and the existing and skipped lists are untouched.

2. Reconciling is not deduplicating, and six groups disagreed with themselves

Every group was read in full and settled against upstream sources and this workstation rather than by keeping the longer record. What that turned up:

Local confirmations that went into the merged records, all on omarchy 4.0.2-1: line 531 of omarchy-agent-usage-codex reads -a on-request, the cleanup() at line 82 of omarchy-theme-set-browser-policy is an if block and line 16 of migration 1787515927 ends in || true, and shell.qml lines 348 to 354 still destroy every plugin service.

3. What the audit still owns

None of this is an audit. The reconciled file carries nine open questions, Q1 to Q9, one per group, naming what a merge may not assume: whether omarchy-iso#111 is the accepted fix and why hand-creating an EFI partition helps, whether the RDNA4 quantizer-matrix mechanism explains the RDNA3 integrated-GPU reports, the codex release note that retired untrusted, the date Google cut individual accounts off, whether the migration reporters were on dev or edge, PR #7169, the three competing gates proposed for the NVIDIA environment and the Pascal case none of them fixes, whether PR #7026 has merged, and whether the VMware failure has a compositor-side fix in hyprwm/Hyprland#12966.

Three lint_corpus.py hits survive across the 36 records and all three are legitimate: two sudo pacman -Syu inside a branch labelled plain Arch, and one /boot/limine.conf in a symptom describing what a reporter saw in that generated file rather than telling anyone to edit it. They are new hits, so lint_corpus.py --check will fail the first time these records enter the corpus. Add them to data/lint-baseline.json in the merge commit rather than rewording correct text.

4. Batch 01, the ten issues the 2026-09-07 run never reached

Read in full on 2026-09-10 against issue-harvest-brief.md, with every claim about what Omarchy ships checked on this workstation (4.0.2-1) and against the quattro branch or a release tag. Six records, four mapped onto records that already exist, nothing skipped.

issueoutcome
6868ghostty-epoll-backend-segfault-write-queue
6882windows-vm-launch-no-rdp-window-stale-log
6887mise-stale-registry-attestation-install-failure
6894quattro-upgrade-uki-missing-root-parameter
6909internal-panel-forced-to-2x-scale-clamshell-poll
6917screensaver-self-dismisses-and-session-never-locks
6858existing: omarchy-lockscreen-no-keyboard-focus-after-resume
6888existing: the pending plugin-lock record, added as a third source
6889existing: the pending Gemini record, already cited
6890existing: quattro-upgrade-incomplete-do-not-reboot

Two of the six are live defects on 4.0.2-1 with no merged fix, which is why they are worth having. Omarchy ships async-backend = epoll in its Ghostty config as a workaround for Hyprland slowness, and that setting selects a libxev backend with a null dereference in its write path: six reporters, four Omarchy releases, one fault offset, and every window in the single-instance process dies together. The removal is omacom/omarchy#6963, still open. And the screensaver closes itself about a second after it opens whenever a bar panel holds keyboard focus, which cancels the pending lock: one reporter measured 239 dismissals over two days, 233 of them within 1.40 to 1.56 seconds, and an overnight run of 225 idle cycles in which the session never locked at all. That is omacom/omarchy#7102, still open, and the workaround is to turn the screensaver off so the lock path runs without it.

The existing notes are the more useful half again. The corpus record for the lock screen not taking keystrokes scopes the defect to lid-open resume and tells the reader to edit ~/.config/hypr/hypridle.conf. Neither hypridle nor hyprlock is installed on Omarchy 4 and no such file exists, and the thread shows the defect is one single-shot focus grab that five separate paths lose, with three pull requests open and none merged. That record needs a re-audit, not a tweak.

5. The audit: 36 of 36 corrected

21 agent batches of one or two records each, grouped by category, each handed reaudit-brief.md, the record JSON and an output directory, exactly as the brief prescribes. Three facts were passed in on top of the brief because they are newer than it: today's date, that v4.0.3 is the newest tag rather than v4.0.2, and that every cited issue and pull request state in these records was written between 2026-09-07 and 2026-09-11 and had to be re-checked. That last one was not hypothetical, since it is what the v4.0.3 correction above came from.

The result is research/raw/issue-harvest-audited.json, 36 records carrying audit_status, audit_confidence and the auditor's reason verbatim as audit_note, with cause_reconciled stamped on the 25 whose cause was rewritten.

verdicts36 corrected, 0 ok, 0 rejected
confidence34 high, 2 medium
fix rewritten29
cause rewritten25
danger rewritten13
symptom rewritten13
verify rewritten12
records with a danger16 of 36

The fix field is where the defects live. 29 of 36 fixes were rewritten, and the failures are not subtle once someone runs them. A sed command could not run at all, because the substitution used | as its delimiter while the replacement contained ||. A chmod g-s promised mode 700 and produces 777 from what the container actually leaves. A verify command grepped six lines after a marker for a value printed five lines before it. Another called a script with no argument, so it prints usage and fails even on a fixed system. None of these survive one attempt by a person at a keyboard, which is the whole argument for the audit.

Three fixes were actively harmful. The screensaver record told the reader to run omarchy-toggle-screensaver to restore locking, and that command is a flip, so on a machine where the screensaver is already off it re-enables it and silently restores the unlocked session the record exists to prevent. The fingerprint record stopped the daemon synchronously during resume, which blocks the thaw for the stop timeout in exactly the wedged case the hook exists for. And the boot record's example root= named a device that does not exist on an ISO-installed machine, so a reader following it literally stays unbootable.

Mechanisms were wrong even where the fix worked. The installer ESP cause blamed an unsettled udev re-read, and the script already runs partprobe, sync, a sleep, a device wait and wipefs before formatting. The real mechanism is the untyped mount walking /proc/filesystems, where squashfs is present on the live ISO and vfat is not. The pull request author had already removed the settle and sleep from the patch after review, so our fix told readers to add two steps upstream had dropped. The VMware cause blamed a rejected dmabuf modifier where the established mechanism is a driver surface handle failing to close, disproved for modifiers by standalone tests. The Codex cause described an 8 second timeout that never happens, because the read returns empty on the dead child and raises immediately, measured upstream in single-digit milliseconds.

Staleness inside three days. One record pinned a guard to a line number read on 4.0.2-1, and v4.0.3 rewrote that file and moved it, so the auditor replaced the pin with a grep. Another had its version floor corrected from one release to a range. A third had a release-note claim tightened from "about three hours" to two hours forty-three minutes.

Four of the nine open questions were answered, three of them against my own judgement. The Gemini cutoff date I dropped as uncited is citable: the issue links a Google developers blog post, retrieved 2026-09-11, stating that Gemini CLI stopped serving individual accounts on 2026-06-18. The Codex approval-policy change I declined to cite exists as openai/codex#39630, merged 2026-08-20. Q2 resolved to "do not split the AV1 record", because Arch's -2 reverted every amdgpu VCN blob and the Mesa fix is gated on the quantizer matrix rather than on any ASIC, so one regression covers RDNA3 and RDNA4. Q5 resolved to both channels, with one reporter's own $OMARCHY_PATH proving a dev checkout where they had written edge.

The audit found two gaps in its own schema. An auditor judged a severity wrong and had nowhere to put it, because reaudit-brief.md's verdict carries no corrected_severity, so the judgement went into prose and was applied by hand. A second found a cited GitHub issue number that is really a discussion and could only say in prose that the URL must be deleted, because a verdict can append sources and never remove one. Both are recorded in the audited file under audit.schema_gaps, and the brief needs both fields before the next run.

Every source was then fetched. 273 distinct URLs, all resolving except api.fast.com, which answers 403 to an anonymous GET and is kept deliberately with that noted. Nine URLs were repaired by hand: seven omacom-io organisation names normalised to the canonical omacom, one blob/main path corrected to blob/master, and the discussion-as-issue URL removed.

6. The merge, and what it needed

merge_gapfill.py needed no change after all. Its append path only stamps gapfill-unaudited when the audit block is missing, and every record had a verdict, so the tested path applied all 36 and stamped cause_reconciled itself. The payload deliberately carries gapfill and gapfillAudit and no audit block: that first pass walks every record of a named category, so an audit block here would have re-audited the 456 records already in the corpus against verdicts that do not name them, which is the exact hazard CLAUDE.md warns about.

The merge ran against a backup and the first 456 lines of the result diffed byte-identical, so nothing already in the corpus moved. The corpus diff is 36 insertions and 0 deletions.

Three things the tool could not carry were re-applied by hand afterwards, from the audited file, and each one is listed in it:

The 12 new lint hits went into data/lint-baseline.json, 140 records to 147, with nothing removed. Every number in prose across the repo was recomputed rather than copied: README.md, research/README.md, CLAUDE.md, corpus.py, the FIELDS test, validation/README.md and this file. One had drifted badly. validation/README.md said 6 of 456 records carry a fenced verify block, and it is now 39 of 492, because 33 of the merged records have one.

7. Where to pick this up

  1. O4 is finished. The next corpus work is O3, the 120 ok records with a danger that apply to Omarchy, on gpu-drivers or apps-services.
  2. Add corrected_severity and corrected_frequency to reaudit-brief.md, and a way for a verdict to remove a source, before the next audit run. Both gaps cost hand edits here.
  3. The dash cleanup of data/problems.jsonl is unchanged at 1,756 across 405 records, now of 492. The 36 merged records carry none.

Session of 2026-09-07 (second): O4 harvests from the issue tracker, and stops one batch short

O4, the option that says stop harvesting the web and read the upstream tracker instead. Six corpus records had cited GitHub issues that did not support them, so the standard here is narrow: a record exists only where the thread carries a fix a maintainer or a second reporter confirmed. A plausible workaround nobody confirmed is not a record.

Nothing from this reached the corpus, deliberately. data/problems.jsonl is untouched and still reads ok 207 / corrected 249 / unaudited 0. The harvest is banked as raw provenance in research/raw/issue-harvest-partial.json, which says so in its own status field.

1. What was built

2. What came back

Eleven harvesters, ten issues each. Ten of eleven finished; batch 01 died on the weekly limit and its ten issues (6858, 6868, 6882, 6887, 6888, 6889, 6890, 6894, 6909, 6917) were never read. From 99 issues:

records written40
issues mapped to an existing record instead21
skipped as unconfirmed, feature requests or duplicates37

The skip rate is the point. Thirty-seven threads had no fix anyone confirmed, and under the old standard several would have become records.

The existing list is the more valuable half and is re-audit fuel for O3: it names records whose cause the tracker contradicts. Among them, chromium-video-black-hybrid-angle (the variable comes from nvidia.lua, went live in 4.0.1, and needs a logout), laptop-display-stays-dark-after-external-unplug (the cause is Hyprland 0.56's FALLBACK output, not a stale monitor list), and suspend-fails-nvidia-video-memory (carries a mkinitcpio -P defect the lint already flags).

3. Why it was not merged

Two blockers, either one sufficient:

Merging would also have put 21 omarchy-core records into a category that has 2 ok records with a danger today, skewing the corpus toward the tooling and away from the hardware problems users hit.

4. Two live facts changed under us, both now corrected

5. Where to pick this up

  1. Run batch 01: gh issue view <n> -R omacom/omarchy --comments for the ten numbers above, with issue-harvest-brief.md. DONE 2026-09-10: six records, four mapped onto existing records, nothing skipped. See the 2026-09-10 session.
  2. Reconcile the six near-duplicate pairs into one record each. DONE 2026-09-10, and it was six groups over thirteen records rather than six pairs. See the 2026-09-10 session and raw/issue-harvest-reconciled.json.
  3. Audit all 40 with reaudit-brief.md, one agent per one or two records.
  4. Only then merge, through merge_gapfill.py, dry-run on a copy first, and expect merge_gapfill.py to need a small change: its append path assigns gapfill-unaudited and these records will arrive already audited.

Session of 2026-09-07: all 22 pacman-aur records re-audited, all wrong, and the lint earned its keep

O3 on pacman-aur: the 22 records that were ok with a danger and apply to Omarchy, one auditor per two records, the same brief as the boot-kernel batch plus a section of pacman facts read off this workstation (the guard script, DownloadUser = alpm, the [omarchy] repo, HoldPkg, yay from the Omarchy repo and no paru). Eleven auditors were launched; five died on a session limit, but three of those had already written both verdicts, so 18 of 22 records came back, all corrected, all at high confidence. The operator paused there, then resumed for the four lost records with two more auditors: all four corrected, three at high confidence and local-package-database-corrupted at medium. Section 4 has them. O3 stops after this category.

1. What was wrong, again in clusters

2. The lint caught the auditors

lint_corpus.py --check ran after the merge and flagged three new hits. One was a labelled plain-Arch aside and went into the baseline. The other two were the auditors' own sudo omarchy-snapshot create, in two rewritten fixes, and omarchy-snapshot calls sudo itself and reads OMARCHY_PATH, which sudo strips. That is the shape the lint was built from one day earlier, written fresh by an agent holding the brief that names it. Both records were patched through corpus.write_jsonl with a sentence appended to the audit note, and the baseline was rewritten deliberately: 139 records, the three changes listed in the commit.

3. Verified, and not

Dry-run then diff: exactly eighteen records changed, only in the named fields, and the four records without a verdict came through untouched. research/tests/run.sh passes (18), the site builds. Nothing was exercised on a VM, no stale keyring, lock or conflict was induced, and the six verdicts from auditors that died after writing were read in full rather than trusted on their status.

After the first eighteen: ok 211 / corrected 245.

4. The four lost records, and a sibling audit catching an applied one

The four came back with the same shapes plus two the batch had not seen. The hook-failed record's title and cause were the opposite of what happens on Omarchy: the commonest source of command failed to execute correctly here is the update guard, the only hook on the machine with AbortOnFail, and it installs nothing. Its sample output cannot occur at all (no presets, and limine-mkinitcpio-hook's script swallows build failures and fails only on the ESP check). The database record gained the recovery Omarchy makes cheap: every omarchy update snapshots the root subvolume, and /.snapshots/N/snapshot/var/lib/pacman/local is a complete copy to restore from. The AUR record quoted yay 12 text that yay 13 does not print. The disk-full record now says the 10 GiB check runs before the prune, so under 10 GiB omarchy update frees nothing.

One of those auditors also read omarchy-update-aur-pkgs and contradicted a claim the yay-libalpm verdict had made and I had merged the day before: that omarchy update aborts at the AUR step. The script runs yay inside an if body whose last command is echo, so it exits 0 whatever yay returned and the update continues. Confirmed by reading the script, the applied record's symptom was rewritten and its audit note says what was wrong and who caught it. Two auditors, two readings of the same file, and the one that read it more carefully was right. That is the argument for reading every verdict rather than trusting its confidence field.

The lint flagged two new hits after the merge, both legitimate (a labelled plain-Arch line and a symptom block that shows the guard firing), and the baseline is now 140 records.

The corpus is now 456 records, ok 207 / corrected 249 / unaudited 0, 925 distinct sources, 55 cause_reconciled stamps across four dates, 1,756 dashes across 405 records. Thirty-seven ok records have been checked against Omarchy 4 since 2026-09-06, and thirty-seven needed correcting.

Session of 2026-09-06 (second): the last four unaudited records, the VM findings, and ten boot-kernel re-audits

Item 2 of the START HERE block, the part that did not need a 35-agent harvest. Four records had carried unaudited since the first harvest's auditors never returned a verdict for them, and the three defects the 2026-09-01 VM scenario found in mkinitcpio-pacnew-unhandled-breaks-next-boot had sat in research/validation/README.md unapplied. All five are now corrected, each with an audit_note naming what was checked and where, and no record carries unaudited.

1. All four unaudited records were wrong, and one was wrong about its own cause

One auditor per record, against the primary sources with the fetch workarounds in CLAUDE.md, and against this workstation where the claim was checkable locally. Every one came back corrected at high confidence:

The same shape as 2026-09-01: sound advice, mis-specialised, and one record whose cited issue does not support its claim. Every audit note names the sources, and the sources an auditor relied on that the record did not list were appended to it.

2. The VM findings are applied, and the merge tool can now carry them

merge_gapfill.py honoured only corrected_fix and corrected_cause. Two of the three 2026-09-01 findings were a wrong symptom and an overstated danger, so there was no honest way to apply them. It now honours corrected_symptom and corrected_danger the same way, and appends any sources a verdict carries. One test covers all three, and the suite is 14 tests, all green.

The pacnew record's verdict was written from the validation findings plus ownership checks on this workstation: /etc/default/limine is owned by no package, the package-owned template is /etc/limine-entry-tool.conf from limine-mkinitcpio-hook, and the file on Omarchy 4 that can both get a .pacnew and carry the hook list is /etc/mkinitcpio.conf.d/omarchy_hooks.conf, a backup file of omarchy-settings. The symptom now quotes the right files, the cause names the right owners, and the danger says what overwriting each one actually does on Omarchy 4 versus plain Arch. The fix stands.

3. Verified

Dry-run on a copy of the corpus first, then the real merge produced a byte-identical file. Exactly five records changed and only in the fields the verdicts named. The verdict set was scoped to those five slugs, and the 152 other records in the four affected categories came through with every field unchanged. build_db.py regenerated four category pages, research/tests/run.sh passes, and the site builds.

After this batch the corpus was 456 records, ok 239 / corrected 217 / unaudited 0, with 33 records stamped cause_reconciled. Section 5 moves it again.

4. What this does not change

audit_status: ok still means "matches its sources". Nothing here was exercised on a VM except the pacnew record, whose scenario already passed 6/6. The rewritten hibernation fix in particular is built from two issue threads and the setup script, not from a hibernate cycle, and a scenario for it needs a hybrid laptop these VMs cannot imitate. Item 1 of START HERE, the harvest, is untouched: it needs a /usage-credits check and an explicit decision to spend about 35 agents.

5. Ten boot-kernel records re-audited for Omarchy 4: ten out of ten were wrong

The four unaudited records and the pacnew finding share one shape: sound Arch advice that was never checked against what Omarchy 4 actually ships. So the next batch asked that question directly of the records where being wrong breaks boot: the ten boot-kernel records with status ok, a danger set, and omarchy in applies_to. One brief, five auditors, two records each, every claim checked against the cited source and, where it could be, against this workstation. All ten came back corrected, nine at high confidence and one (nvidia-modules-missing-from-initramfs-black-screen) at medium because no DKMS failure was induced.

The defects cluster, which is the useful part:

Every fix now carries labelled Omarchy 4 and plain Arch branches where they differ, and every audit note says what was checked locally versus from a source and what was not exercised. merge_gapfill.py gained corrected_verify on the way, because the chroot record's verify step looked for /boot/vmlinuz-linux. Same dry-run-then-diff discipline: exactly ten records changed, only in the named fields, and the 25 other boot-kernel records are untouched.

The corpus is now 456 records, ok 229 / corrected 227 / unaudited 0, from 832 distinct sources (verdict sources are appended to the record), with 41 records stamped cause_reconciled across three dates. Dashes: 1,839 across 418 records.

The deploy failed on the merge, and the failure was correct. pages.yml's sanity check required an UNAUDITED label somewhere under docs/records/, as a guard that the provenance disclaimer still renders. With zero unaudited records that grep is empty, so the first site build after this work stopped before deploy. The check now asserts that every record page carries an audit LED and that at least one carries a non-clean one, which is the property it meant to guard. Fixed in a follow-up PR the same day.

6. O1 landed: a lint for the shapes the re-audit found

research/tools/lint_corpus.py greps every record that applies to Omarchy for eleven shapes today's audits found wrong on Omarchy 4: mkinitcpio -P, sudo pacman -Syu, bare pacman -Sy, /boot/vmlinuz-linux, the Omarchy 3 tree, hyprland.conf, hl.set, edits to /boot/limine.conf, bare hyprctl dispatch, sudo omarchy-, and /boot/efi. It is a candidate finder, not an audit: some hits sit in a branch labelled plain Arch. 138 of 456 records carry at least one hit, and mkinitcpio -P alone is in 52 (18 still ok). data/lint-baseline.json records those 138 as known, --check and a test fail on any hit outside it, and the test was proved able to fail by injecting a hit into a clean record. Clearing a slug from the baseline belongs in the commit that re-audits it. The six options themselves are item 8 under "What's left".

7. O2 landed: the workflows now read the brief first

All three workflow scripts prepend the same preamble to every harvester, gap-filler and auditor prompt: read tools/reaudit-brief.md from the corpus root before writing anything, hold every claim to it, and write both branches where Omarchy 4 and plain Arch differ. The preamble also carries inline the shapes that caught the most records, so an agent that skims the file still sees them. harvest-workflow.js now takes args.root (CLAUDE.md had said all three did; only two did), its audit schema accepts every corrected_* field plus sources, and its merge applies them with a cause_reconciled stamp instead of honouring corrected_fix alone. Verified by evaluating each script's constant section under node with a fake root; no workflow was run. O6, making that harvest extend the corpus rather than replace it through ingest.py, is still open and still a precondition for running it.

What this says about the other 219 ok records. Fifteen records checked against Omarchy 4 today, fifteen wrong. The boot-kernel set was chosen because it is where the cost is highest, not because it is where the errors are, so the rate elsewhere is unknown but there is no reason to expect zero. The same brief (scratchpad/audit2/PREAMBLE.md this session, worth promoting into research/tools/ as the prompt for audit-existing-workflow.js) runs at about 130k to 180k tokens per two-record agent, and now lives at research/tools/reaudit-brief.md. Of the 219, 142 carry a danger and apply to Omarchy: apps-services 23, pacman-aur 22, power-suspend 16, omarchy-theming 15, network 15, gpu-drivers 14. The pacman-aur set is the next batch by cost of being wrong (partial upgrades, keyring, downgrades), then gpu-drivers.

Session of 2026-09-06: the published documents audited against the repo

Every document the site renders under "How this was built" was read against the corpus, the bench specs and the tracked export. Every published figure that could be recomputed was, and all of them reproduce. What did not hold:

Both fixed in the same PR before merge: CLAUDE.md said seventeen bench specs and nine Omarchy ones, and skillbench/results/README.md said 31 runs. Later the same day the repo gained a root README.md, its first, with a screenshot of the site and a plain statement that the skill is vaporware, and this block was rewritten to point the next session at the corpus rather than the bench.

Session of 2026-09-05 (fourth): the turn budget was not real, and four runs are void

Runs 43 to 46 are contaminated and their conclusions are withdrawn, including one that had already been published on the site.

1. opencode was silently refusing the agent's edits

opencode run is non-interactive, and opencode's default policy is external_directory: {"*": "ask"}. In non-interactive mode "ask" means auto-reject, for any tool call touching a path outside the working directory:


! permission requested: external_directory (/tmp/many/*); auto-rejecting
tool_use bash {"command":"mkdir -p /tmp/many"} -> "The user rejected permission to use
                                                   this specific tool call."
step_finish reason: "tool-calls"

No error is recorded on the case. The run completes, status ok, and the transcript ends at reason: "tool-calls", which is exactly what a model that chose to stop looks like.

It was hitting every one of the four runs, the control included:

runtaskcases with a rejection
43rebind-packaged-default32/40
43looknfeel-not-hyprlang28/40
44 (control)dropin-shadows-unit32/40
44 (control)deleted-file-holds-disk21/40
45theme-overlay-not-packaged20/20
46theme-overlay-not-packaged10/10
45/46idle-lock-not-hypridle3/20, 0/10

The tasks that failed are exactly the ones reaching outside $HOME, and both control tasks do, in /etc and /var/tmp. The one that worked lives entirely inside it. (The run 44 rows were added on 2026-09-06; the entry as first written omitted the control.)

2. Two conclusions withdrawn

The chat-lane results and the n=31 agentic null are unaffected: both predate the opencode backend and ran through pi.

3. What actually found it

The operator asked why the agent was stopping instead of accepting a workaround. Three hypotheses were tested and two were wrong:

--dir "$HOME" from the previous session was a half fix: it made ~/.config writable and left everything else refused. That is why idle-lock-not-hypridle worked and nothing else did.

4. The fix, and the lesson worth keeping

app/runner.py writes ~/.config/opencode/opencode.json per case with permission.external_directory: {"*": "allow"}, pinned by a test that carries the case counts in its docstring. autoupdate stays false so opencode cannot change version mid-run.

A silent refusal and a model giving up are indistinguishable from the score column, and nearly indistinguishable from the transcript. Only the string auto-rejecting separates them, and it appears in the tmux output rather than in the JSON event stream. This is the fourth harness bug in this lane to imitate a capability failure, after the scratch HOME, the working directory, and run 21's VRAM spill.

Session of 2026-09-05 (third): a new bench, and the constraint it exposed

omarchy-agentic-published-wrong was written from benches/SKILL-COVERAGE.md, hand-verified across all three outcome states, and then measured. Both tasks fail, in opposite directions, and the reason is more useful than the bench.

1. The bench is not calibrated, and is marked so

taskresult
idle-lock-not-hypridlebare 1.0 on GLM-class. Too easy: they already know idle.lock is in shell.json. 27 of 30 cases edited.
theme-overlay-not-packagedboth arms at the 0.8 floor. Zero of 30 cases ever edited anything.

Hand-verification was sound: seed-only 5/6 and 4/5, wrong answer 4/6 and 2/5, correct 6/6 and 5/5, so a wrong answer scores below doing nothing in both. The bench measures what it was built to measure. The models simply do not reach the state it grades.

2. The real constraint is a TURN BUDGET, not difficulty

Measured across every opencode agentic case run so far:

taskmedian turnscases that edited anything
idle-lock-not-hypridle427/30
looknfeel-not-hyprlang69/40
rebind-packaged-default67/40
theme-overlay-not-packaged40/30

Overall median 5 turns. Two reads plus one edit gets done. Find-a-thing, make a directory, copy, then modify does not, and it fails silently: the model stops at the same budget and never signals it has not finished.

That reframes run 43's finding. The skill diverting an agent into /usr/share/omarchy/ did not merely waste effort, it spent the whole budget. Under a five-turn ceiling, a skill that adds one research step is the difference between finishing and not.

A task is not too hard, it is too many steps. Seed away the setup so the agent only has to make the change. This is now in SKILL-COVERAGE.md alongside the saturation rule.

3. A property of the split bundle, ruled out as the cause

skill:omarchy is files: [SKILL.md], and SKILL.md links to six topic guides while telling the reader to consult them. In that variant the guides are absent, so an agent follows the pointer and finds nothing. Observed verbatim: "The skill's theming guide doesn't exist, so I'll explore the stock Everforest theme structure directly."

The bundle is complete on disk; this is an artefact of how skills.yaml splits it. It means omarchy is not simply "less skill" than omarchy-full but a skill with dangling references, which is worth remembering when reading any comparison between the two.

It was not the cause here. Run 46 with omarchy-full, which does contain theming.md, scored identically: 0 edits in 10 cases. Recording it because it was my first hypothesis and it was wrong, and because it still affects interpretation elsewhere.

4. What the next attempt needs

Not a harder seam. A shorter one. Either make task 1 require more than a single edit, or seed task 2's overlay directory so the agent only has to write the file. The seam inventory is sound; the step count is what has to change.

Session of 2026-09-05 (second): the first Go ladder, and the skill diverts an agent

Runs 43 and 44: the GLM ladder (glm-5.1, 5.2, 5.3-flash, 5.3) on the agentic lane through opencode, 5 repeats, none vs skill:omarchy, 160 cases across both. Zero marginal cost, because every rung is opencode-go/.

1. The numbers

rungOmarchy lift (run 43)control lift (run 44)DiD
glm-5.1-20.0 pt (p=0.02)+0.0 pt (p=1.00)-20.0
glm-5.2+10.0 pt (p=0.21)+0.4 pt (p=1.00)+9.6
glm-5.3-flash-13.3 pt (p=0.14)-10.0 pt (p=0.10)-3.3
glm-5.3-16.7 pt (p=0.06)-3.7 pt (p=0.53)-13.0

The control does not move, so nothing here says the skill degrades general Linux. The Omarchy direction is negative on three rungs of four and inconsistent, mean DiD about -6.7 pt.

Corrected 2026-09-06: every number in that table is void, the control column included. Run 44 was hit by the permission refusal described in the fourth session on 53 of 80 cases, more than run 43, so the flat control measured the refusal rather than the skill.

2. The mechanism, which is the solid part

On rebind-packaged-default with the skill loaded: zero edit-tool calls and zero clean stops across 20 cases. The same skill on the sibling task produced 7 edits and 7 clean stops, so this is not truncation, a step cap or plumbing.

The transcripts show what happens. The skill sends the model to grep /usr/share/omarchy/ for the binding API, it researches the packaged defaults, and it never gets round to editing ~/.config/hypr/bindings.lua. Bare, lacking that pointer, it pokes about locally and edits the user file.

That is directly useful for the skill this repo exists to build. A corpus-backed skill should say what to change, not invite a tour of upstream.

3. Three caveats, all load-bearing

4. What the next bench needs

omarchy-agentic-stale-advice cannot measure GLM-class models: they nearly max it bare. A harder Omarchy agentic bench is now the blocking artefact, and the mechanism above suggests what it should test, namely whether an agent acts rather than researches.

5. Goldens re-saved

Both VMs powered off, saved one per invocation, restarted, ready:true on both, and the agent key verified at 67 characters from a login shell after the reboot. Without this the next reset would have dropped the key exactly as it dropped NOPASSWD sudo on 2026-09-01 and both bench keys on 2026-09-02.

Session of 2026-09-05: the agentic lane runs on opencode, and three bugs wore one costume

1. Go was reachable all along, under a prefix I never tried

opencode models reports two providers: opencode/ (69 models, pay-as-you-go) and opencode-go/ (27 models, the subscription). Every model previously written up as "not supported" (glm-5.3, qwen3.8-flash, longcat-2.0, omen-alpha, hy3, kimi-k2.6 and the rest) is a Go model reachable only under the second prefix, and only through the CLI.

The previous session's conclusion that Go was unreachable was wrong, and it was wrong because I called opencode/<model> throughout without ever listing the providers. The operator's TUI screenshot showing Qwen3.8 Flash OpenCode Go at $0.00 is what found it.

Go is free at the margin, measured rather than assumed: opencode/ moved the account balance by exactly what opencode stats reported, while five opencode-go/ sessions moved it not at all and stats still claimed about two cents. So opencode stats reports notional cost and does not separate covered from billed. That also corrects the earlier "the API has no cost accounting" claim, which was too broad in the other direction.

2. The agentic lane now drives opencode, and it is instrumented

params.agent: "opencode" runs opencode run --format json. For the first time since run 18 the lane records turns, tokens, tool calls and cost per case, against a 16-character median transcript and no token accounting before.

That instrumentation is not a nicety. It is the only reason the next section was solvable.

3. Three bugs, all producing exactly 4/6

Runs 39, 40 and 41 scored the do-nothing floor on every case, across two models. Each had a different cause and all three looked identical from the score column, which is the failure mode MODELS.md exists to warn about:

Run 42, after the fixes: looknfeel-not-hyprlang none 4/6 skill 6/6, rebind-packaged-default none 6/6 skill 4/6. One win each way at n=1, which says nothing about the skill and everything about the harness: it discriminates now instead of flooring everything.

4. A stock Omarchy machine has no bare condition

Omarchy symlinks omarchy and diagnose-crash into ~/.agents/skills on every install, host and VM alike. So the none arm silently carries the skill it is the control for unless the harness clears it, and more interestingly, every opencode session a real Omarchy user starts already has the skill loaded. "Does the skill help" is not hypothetical for them, it is their default.

5. Credential handling, and one mistake

tools/install-agent-key.sh provisions the model key onto the VMs once, the way install-bench-key.sh does for ssh. The runner handles no token: a login shell already exports it, and a test asserts the generated command contains no API_KEY, sk-, Bearer or token.

Two failures worth keeping. Nested quoting through a login shell failed silently under set -e; and printf '%s' with no trailing newline made the remote read hit EOF, return non-zero, and abort the whole remote script before it wrote anything. The key lives in ~/.bash_profile, not ~/.bashrc, because Omarchy's .bashrc returns on line 5 for non-interactive shells, which is exactly why Omarchy exports OMARCHY_PATH above that line.

The key was also printed in plaintext into a session transcript by running bash -x on a script that handles a secret. Rotation was offered and deferred. Do not trace a script that touches a credential.

Session of 2026-09-04 (fourth): the docs are published, and Go turns out to be unreachable

1. Go bills nothing, because an API key means pay-as-you-go

Settled by a controlled experiment rather than by reading docs. The operator read the balance, a measured batch ran through the opencode CLI on Go-listed models only, and the balance was read again:

balance$4.98 to $4.62, -$0.36
opencode stats$0.07 to $0.43, +$0.36

Exact match. Neither /zen/v1 nor opencode run draws on the Go subscription when authenticated with an API key. A second, larger batch predicted $4.13 against an actual $4.20, so opencode stats is close but over-reports by roughly 8% and is not authoritative.

Two things this corrects:

The open question is no longer "does Go cover the API" but "how is Go reachable at all", since an API key appears to mean pay-as-you-go by definition.

2. The project's own documents are now pages on the site

The "How this was built" cards linked out to GitHub. They now open rendered pages: research/README.md, skillbench/README.md, skillbench/MODELS.md, skillbench/ZEN.md, JOURNAL.md and the writeup, driven by a DOCS table in build_site.py.

md_doc extends md_lite to what these files actually contain: headings, paragraphs, lists, fenced and inline code, pipe tables (64 rows across the docs), blockquotes, bold, italic, links and rules. Same escape-first rule, so no document can inject markup.

Three defects were caught by rendering rather than by reading the generator, which is the same lesson as the paragraph bug:

Cross-document links resolve locally where the target is rendered and fall back to GitHub at HEAD otherwise, so [MODELS.md](MODELS.md) inside ZEN.md becomes models.html.

pages.yml now triggers on those six sources, and CI fails if any doc page is missing or if the cards stop pointing at them. Without that the site would quietly serve a stale copy of its own documentation, which is the failure that a paths trigger is for.

Session of 2026-09-04 (third): the chat lane on cloud models, and a real Omarchy effect

The first cloud runs. The skill shows a large, replicated, Omarchy-specific lift on the chat lane, which is the opposite of the agentic lane's null and the first positive result this project has produced under proper controls.

1. Four clean models, four replications

linux-desktop-gauntlet, 10 repeats, none vs skill:omarchy, 200 cases per model.

modelOmarchy taskscontrol tasksDiD
nemotron-3.5-lightning-free+28.8 pt+3.4 pt+25.5
deepseek-v4-flash+25.1 pt+1.8 pt+23.2
qwen3.5-plus+27.3 pt-5.3 pt+32.6
qwen3.6-plus+22.6 pt-1.8 pt+24.3

Every Omarchy lift is p < 0.0001. No control is significant. Against the agentic lane's DiD of +0.2 pt at p = 0.98, this is the same skill and the same corpus behaving completely differently on the lane where the incumbent was always shown to work.

The gauntlet is a MIXED bench, six Omarchy tasks and four general-Linux ones, so it carries its own control and the difference-in-differences comes out of one run at no extra cost. Reporting only the pooled figure would state a lift belonging to neither group: on run 34 the pooled +18.6 pt is +28.8 on Omarchy and +3.4 on controls. lift_test.py now splits automatically and takes --model=, since a run can hold several.

Checked against the obvious objection, that the controls are simply easier: the skill uses 42% of available headroom on Omarchy tasks against 10% on controls, so the effect survives normalising for the ceiling.

2. I contaminated three models, by fixing a bug

The gauntlet spec carries max_tokens: 350. That cap had never been in force, because options{} was ignored. Making parameters work made it effective for the first time, and reasoning models spend 350 tokens thinking and emit nothing. Those cases graded zero.

The truncation is variant-correlated, because the skill makes answers shorter:

modelbare truncatedskilled truncatedas scoredexcluding truncated
deepseek-v4-pro81/10019/100om +37.1, DiD +8.5om +9.1, DiD +1.0
minimax-m359/10026/100om +32.2, DiD +36.3om +15.5, DiD +20.1
minimax-m2.553/10028/100om +31.1, DiD +33.0om +20.2, DiD +16.1

The guard I wrote had a hole exactly where the risk was. The commit says max_tokens is sent "only when a spec asks for one" so it cannot impose a new cap. The spec does ask for one. I guarded the default and not the spec value.

lift_test.py now drops output_source='empty' cases: a response with no text cannot be graded on its text, and scoring it zero conflates "wrong" with "did not finish". The four clean models are unchanged by that, which is the check that it is not doing something arbitrary. deepseek-v4-pro is unmeasurable at this cap rather than null, since excluding truncation leaves 19 bare cases.

This is the evidence for the held max_tokens decision. At 350 the bench cannot measure a reasoning model at all. Raising it edits a bench spec and starts a new series, so it stays an operator decision.

3. The free tier cannot carry a bench

Of six free models, three produced unusable data: big-pickle and mimo-v2.5-free returned 429 on all 200 cases each, exhausted by earlier probing, and laguna and ling lost 63 and 84 cases to 503 at different rates per variant, which is the shape that voided run

  1. Discarded rather than reported.

_status_of classified the 503s as unavailable rather than as model failures, which is the PR #34 fix working on its first real outing.

4. The skill makes answers shorter as well as better

deepseek-v4-flash averaged 3,257 output tokens bare against 1,000 with the skill. That cuts directly against the "it just makes the model write more" explanation that killed the agentic lift, and it is why bare cases time out more often (10 against 4 on run 37). That bias is conservative: dropping the slowest bare cases raises the bare mean, so the measured lift is a floor.

5. Cost

$3.09 actual against a $5 budget, six paid models, 1,200 cases. The estimator predicted input to within 0.2% and underestimated output by 33%, so the reasoning multiplier is 4x rather than 3x. The contaminated models were cheaper than predicted, because truncation cut their output short.

Session of 2026-09-04 (second): the cloud terrain, and four defects that fail silently

The chat lane can now post to an OpenAI-compatible gateway. Getting there turned up four things that would each have produced plausible, wrong numbers, and one of them reaches backwards into every chat-lane run this project has ever made.

1. options{} has never been applied, on Ollama either

Generation parameters sat in an Ollama-native options object. The OpenAI-compatible /v1 endpoint ignores it. Measured on both servers with options.num_predict = 8:

servercompletion tokens returned
Ollama /v180
Zen /v1172

No chat-lane run has ever used the temperature its spec asked for, including the +29.3 pt baseline. Every one ran at the server default. Both arms of every paired run were equally affected, so the past comparisons stand on their own terms, but a run made after this fix is not comparable to one made before it. Parameters now go top level.

max_tokens is deliberately not sent unless a spec asks for one. The nominal 512 was never in force, so emitting it would impose a brand new cap while looking like a portability fix. The right value is an open question the operator is still weighing.

2. Half the reachable models return no content at all

Twelve of eighteen reachable Zen models returned null or empty content on a short prompt, with the answer in message.reasoning. The runner read content alone, so those recorded as "said nothing" at status ok and graded as zero. _answer_of now falls back to reasoning and records case_result.output_source so the two can be told apart. That distinction is load-bearing: a reasoning trace may consider and reject the trap before settling on the right command, so a regex_forbidden check can fire on a model that got it right.

3. An unreachable model returns a bare 500

Not 402, not 403. Internal server error, on 38 of 59 probed models, indistinguishable from a real outage. Recorded as unavailable rather than error so an unconfigured model cannot be read as one that failed the task. Only kimi-k2.6 was honest about it, with a 401 Model is disabled.

4. Two smaller ones, both fatal in their own way

All four are pinned by tests, and each was proved to fail against the old code before being kept: reverting the three runner fixes produces four failures, and reverting the compose fix produces a fifth. 51 tests now.

5. The terrain is documented because it is not guessable

skillbench/ZEN.md and skillbench/tools/probe_zen.py, the cloud siblings to MODELS.md and probe_models.py.


listed 66 | probed 59 | reachable 18 | HTTP 500 on 38

Five within-family ladders are usable: GLM three rungs, MiniMax three, Kimi three, Qwen two, DeepSeek two, plus six free models. That is the instrument the capability-band question needs, and it is better than the two ladders the first probe suggested.

A listing is not an entitlement. /v1/models lists ids that 500 on use, and omits ids from OpenCode's own Go page (glm-5.3, qwen3.8-flash, longcat-2.0, omen-alpha) which return "not supported".

6. Three account settings do nothing through the API

Flipped on, then off, with a full snapshot at each step. Zero effect on listing or reachability, in both directions:

settinglistingreachability
allow models that train on request datanonenone
allow models hosted in Chinanonenone
use available balance after usage limitsnonenone

Only the per-model enable toggles moved anything, and they moved it from 20 listed to 66. The three above sit under "control which providers are used for routing", which is the likely explanation: they govern how the CLI picks a model when none is named, and naming one over the API leaves nothing to route. Unconfirmed.

7. Two things I had wrong, both caught by checking

Worth recording because the checks were cheap and the claims were not:

8. Go does not cover the API, and a run now has a price

The balance moved 2 cents during the probe session. That answers the question the previous section left open: API calls bill pay-as-you-go per token, and the Go plan's request-per-5-hour caps govern its own routing rather than this endpoint.

It also inverts a claim made earlier the same day. "Requests, not tokens, are the scarce resource" was correct about Go and wrong about the bench, and it stood for about an hour before the balance settled it. Both ZEN.md and CLAUDE.md are corrected.

skillbench/tools/estimate_cost.py prices a run before it is launched, from measured per-case token use rather than assumption: the 272 banked chat-lane cases that carry usage give 127 in / 654 out for none and 3,286 in / 320 out for skill:omarchy.

One 2-task bench at 31 repeats, output tripled to allow for reasoning traces:

scopecost
six free models$0.00
deepseek-v4-flash$0.17
eight cheapest paid models$3.60
all eleven paid models$9.13

So one bench walked across the whole ladder at full statistical power is affordable, and the whole suite across many models is not: every chat bench at 31 repeats runs $2.33 on the cheapest paid model and $46.93 on kimi-k3. The plan is one bench with headroom, the six free models for the bottom of the curve, and roughly eight paid rungs above them.

Caveat that belongs next to the number: every measured figure came from a local model, and most local models are not reasoning models. The 1.0 multiplier is a floor.

9. Still open

  1. Why does every Western-lab model return 500 while every Chinese-lab and open model works on the same key? Every OpenAI, Anthropic, Google and xAI id fails. Not explained by any account setting, all three of which were flipped both ways. BYOK for those providers is the leading guess, and a 500 is indistinguishable from an outage either way.
  2. --mode json remains unimplemented, so turns and tokens per agentic case are still unknown. That now blocks pricing the agentic lane, not just diagnosing it.
  3. Which ids do the Go page's models use? glm-5.3, glm-5.3-flash, qwen3.8-flash, qwen3.7-plus, longcat-2.0 and omen-alpha all return "not supported", which may mean the id is wrong rather than the model unavailable. Those are the newest rungs of two of the five ladders, so it is worth knowing.
  4. The max_tokens default, deliberately left unset pending a decision. 5k is the operator's working figure.

Session of 2026-09-04: the agentic lane is answered, and the answer is a null

Run 31, the control's third attempt, is the first clean one: 124/124 cases, zero seed failures, no VM drained, only legitimate agent timeouts. The seed repeatability check and the pool drain guard both did their jobs, which is what the previous two attempts cost.

benchnoneskilllift95% CIp
omarchy-agentic-stale-advice (run 28)0.8710.890+1.9 pt[-3.8, +7.5]0.59
linux-agentic-deep-triage (run 31)0.7470.765+1.7 pt[-2.7, +6.2]0.47
difference in differences+0.2 pt0.98

The two lifts are the same size. skill:omarchy nudges general-Linux triage exactly as much as it nudges tasks about the Omarchy 3 to 4 split, which is the condition the controls were built to detect: a small uniform effect consistent with more context making a model marginally more careful, and nothing to do with Omarchy knowledge.

Worth stating plainly because it took four runs to get here honestly: the incumbent skill has no measurable Omarchy-specific effect on the agentic lane. Not a weak one. p = 0.98 on the difference.

What this does and does not mean

It does not mean the corpus skill will fail. It means:

The decay, kept as the methodological point

nliftp
3+11.1 pt0.55
10+8.3 pt0.21
31+1.9 pt0.59

Had the n=3 figure been published this project would have claimed an eleven point lift that does not exist. The cost of getting that right was four runs, two of them lost to harness defects now fixed and tested.

The consistency framing was tested too, and it is also null

The obvious rescue for a null lift is that the skill buys steadiness rather than score: fewer catastrophic runs, a higher floor, safer everyday use. That is a different measurement from a mean and the bench stores enough to check it. Across the same 248 cases:

runvariantsdminbelow floorsolvederrors
28 Omarchynone0.1680.5002/6238/626
28 Omarchyskill0.1580.5001/6241/628
31 controlnone0.1180.6000/627/6212
31 controlskill0.1360.5002/6211/6214

Spread is flat on the Omarchy bench and wider on the control. The worst case gets worse on the control, and two cases fall below the do-nothing floor where bare had none. Errors rise in both arms.

Run 25's "the skill halves the error rate" does not replicate. That was 6 of 20 against 2 of 20, and the 2026-09-03 entry flagged it as worth measuring deliberately. At n=62 it is 6 against 8, and 12 against 14. Same shape as the +11.1 pt lift: convincing small, absent at full power.

One thing does move in both arms: solved-outright, 38 to 41 and 7 to 11. Note which moved more. The control gained 4 solves from a base of 7 while the Omarchy bench gained 3 from a base of 38, so if anything the skill helped the control more. That sharpens the answer-length reading rather than softening it.

The question was never binary, and both ends of the curve are now measured

Read runs 21, 28 and 31 together and "does the skill help" is the wrong shape. The answer is a curve over model capability, and this project has measured both ends and never the middle:

That is what makes the local work an anchor rather than a dead end.

The instrumentation gap, which is a bench choice and not a pi limitation

app/runner.py:114 invokes pi without --mode json. pi supports text, json and rpc, and ships read/bash/edit/write tools with an allowlist, so the harness was never the limitation. But in text mode the runner stored only the final assistant text: a median of 16 characters across roughly 600 agentic cases, and 0 of them carry token accounting.

Three consequences, in ascending order of cost:

--mode json is a prerequisite for any paid run.

Everything local is banked

skillbench/data/ is gitignored under the repo's usual rule that derived artefacts are rebuilt rather than committed. That rule assumed the artefact can be rebuilt. research/data/problems.db regenerates in 0.2 s; this one regenerates from nothing, and one disk sat between the project and 31 runs, 1,044 cases, 4,465 grades.

skillbench/tools/export_results.py writes the tracked skillbench/results/: runs.jsonl with aggregates computed exactly as app/main.py:_agg does, cases.jsonl with everything lift_test.py needs, grades.jsonl so which assertion failed survives, and the 13 distinct sha-pinned specs.

1.3 MB against the database's 5.7 MB. The difference is almost entirely case_result.request: 4.87 MB of prompt text re-recorded per case, derivable from the spec exported beside it. The non-prompt keys in it (lane, vm, queue_wait_s) are kept.

Verified by re-deriving the journal from the export alone, with no database. Runs 28 and 31 come back at +1.9 and +1.7 pt with a DiD of +0.2, and run 21's 14 cases reproduce the MODELS.md table exactly. The export is byte-deterministic, so a new run appends instead of rewriting the file.

One subtlety to keep straight: state in runs.jsonl is a micro-average (post_passed / post summed over cases) while lift_test.py uses a macro-average (the mean of per-case ratios) because the unit is the case. They agree when every case has the same number of post assertions (run 28: 0.8710 both ways) and diverge slightly when they do not (run 31: 0.7478 against 0.7473). Quote the macro-average for anything inferential.

Where this goes next: low-cost cloud

The corpus was always for people fixing their own Omarchy machines, and local models matter to them for privacy and for cost. What died is the claim that the skill lifts a model that was already capable. Those are different claims and only the second one is affected.

The shift is to low-cost cloud through OpenCode Zen's OpenAI-compatible endpoint at https://opencode.ai/zen/v1/chat/completions. The Go plan is $10/month flat, with per-model caps in requests per 5-hour window running from 110,000 (Kimi K3) down to 1,350 (Grok 4.6).

The flat fee is the point, and not mainly for cost. A pay-as-you-go key on an agent that can loop is financially unbounded, which is the wrong shape for an unattended multi-hour run and worse for a daily driver. A cap turns a runaway loop into a rate-limit error rather than an invoice. It also makes a published result reproducible by a reader for $10, which matters more than experimental convenience given who this is for.

Go's 25 models form seven within-family capability ladders: Qwen 3.6 Plus through 3.8 Max, GLM 5.1 through 5.3, Kimi K2.6 through K3, DeepSeek V4 Flash and Pro, MiniMax M2.7 and M3, MiMo V2.5 and Pro, Hy3 and Hy4. One provider, one API, one fee, one harness, one variable. That is the instrument the middle of the curve needs, and the weak rungs are the ones to test: benching the headline models would reproduce the devstral null for the same reason it happened the first time.

Order of work: --mode json first, then the chat lane, which needs no VM, costs one request per case, and is the only lane where the skill has ever shown an effect. The agentic lane follows. Nothing about cloud requires touching the VMs, the goldens, the seed checker or the pool guard, and no existing bench spec may be edited, because they are sha-pinned and editing one orphans runs 23, 25, 28 and 31.

To settle at signup: whether the Go entitlement covers raw API calls or only the opencode CLI. The Go page says it "can be used with any agent", which implies the former, but the Zen docs describe pay-as-you-go credits separately and the two are not stated to be the same.

Session of 2026-09-03 (fifth): two harness defects, and the check that would have caught both

1. Run 30 is void, and it was my seed rather than the pool

The control's second attempt lost 51 of 124 cases to mount exit 32, skewed 31 skill against 20 bare. The teardown I wrote for deleted-file-holds-disk ran stop, umount, rm. When the holder still owned its fd for a moment after stop, umount failed busy, the rm deleted the image anyway, and the loop device stayed attached to a dead inode:


/dev/loop0: [0031]:269707 (/var/tmp/bench-disk.img (deleted))

Teardown now waits for the unmount rather than assuming it, falls back to lazy, detaches any loop still pointing at the image, and removes the file last. The mount also prints losetup -a on failure instead of exiting 32 with nothing, because a bare exit code is what made this take a full run to find.

The pool guard from run 29 worked. No VM was drained, nothing cascaded, every failure stayed inside its own case. Run 30 is unusable because half its cases never posed their question, not because one bad case poisoned the rest. Those are different failures and the fix for one does not address the other.

2. The real defect was in how seeds were verified

Run 27 and run 30 are the same shape: teardown that assumes the previous case finished cleanly. Both survived hand-verification, because a seed was verified once. The runner fires it before every case, so the second cycle is the one that matters.

skillbench/tools/check_seeds.py runs every agentic seed N times back to back and fails on the first repeat that breaks. All nine pass; the disk seed that lost run 30 now survives 12 consecutive cycles with zero deleted-inode loop devices, against 41% case loss.

The checker caught a defect in itself first, and that is the part worth remembering. A first draft reported omarchy-agentic-config/theme-switch as broken. That seed is fine: it was invoked through plain bash -s while vm.run() uses bash -lc, and OMARCHY_PATH comes from ~/.bashrc. A checker that does not match the runner invents failures and hides real ones, which is worse than no checker.

It proves the runner can pose a question repeatedly. It does not check that assertions still discriminate; that stays a hand-verification against the shipped templates.

3. Why this got built rather than just re-running

The harness is about to be pointed at cloud model classes and effort levels, so the cost of a four-hour failure that only surfaces in the results is going up, not down. Two runs and roughly eight hours were lost to bugs a few seconds of checking would have caught.

Session of 2026-09-03 (fourth): the incumbent's agentic lift is not real, and one case can poison a run

1. Run 28, n=31: the lift decays to nothing

omarchy-agentic-stale-advice, devstral-small-2:24b, 31 repeats, agent_timeout raised to 900 s as a launch param:

variantnSTATE (ok-only)
none620.887
skill:omarchy620.907

Lift +1.9 pt, p = 0.59, 95% CI [-3.8, +7.5]. The interval spans zero.

The decay is the finding, not the endpoint:

runnliftp
233+11.1 pt0.55
2510+8.3 pt0.21
2831+1.9 pt0.59

This is not an underpowered null. The power calculation from run 25 asked for about 62 cases per variant and that is exactly what run 28 has. Had the n=3 figure been published, this project would have claimed an eleven point lift that does not exist.

The residue is confined to one task. rebind-packaged-default moved from 12/31 solved to 16/31; looknfeel-not-hyprlang went backwards, 26/31 to 25/31. A real skill effect does not appear in one task and reverse in the other.

2. Run 29 is void, and the reason matters more than the run

The control run lost 11 consecutive cases to No route to host on test1, starting at case 786. Every one was skill:omarchy on dropin-shadows-unit.

test1 was found at a TTY login prompt rather than in its autologin session; test2 had its tmux session but no IP. The host was healthy throughout: dnsmasq listening on 67 and 53, all three virbr0 ufw rules present, libvirtd not restarted since 28 August. A reset from the golden image recovered both machines immediately.

The mechanism is not established. The most plausible reading is that a case rebooted or otherwise disrupted test1, which is notable because rebooting is exactly the wrong answer dropin-shadows-unit exists to catch. An earlier guess that the guest kernel showed an agent-run system update was wrong: the golden image carries the same 7.1.9-arch1-2.

The bug the incident exposed is in the runner, not the VM. _one_agentic_case released its machine back to the pool in a finally, unconditionally, so a dead VM was handed to case after case. Because the runner finishes one variant before starting the next, the losses land on whichever variant was running late. An infrastructure failure therefore arrives looking exactly like a model result, and here it looked like the skill degrading general Linux performance by 5.5 points. That number would have been reported.

Fixed three ways:

Two regression tests pin it, 45 total now. The second is deliberately synchronous, driven through asyncio.run, because the test image installs plain pytest and an asyncio-marked test would be collected, skipped, and prove nothing.

Remember the rebuild. compose.yaml does not mount ./app, so this fix needed docker compose up -d --build to take effect. A bench YAML is picked up live; an app change is not.

3. What this means for the skill

The bar in opinionated-omarchy/CLAUDE.md said "do not regress the incumbent on the agentic lane". That bar is now trivially low: there is no incumbent lift on that lane to regress. The chat lane's +29.3 pt remains the only demonstrated skill effect in this repository, which strengthens the case for the token-light resident core rather than weakening it. A corpus-backed skill that shows a surviving agentic lift at n=31 would be a new result, not a repeat of one.

Session of 2026-09-03 (third): the writing standard, applied everywhere but the corpus

Every prose file in the repo outside research/data/problems.jsonl now follows writing-and-responding. The audit found more than punctuation: a 720-line duplicated block in this file, a wrong control count in CLAUDE.md, and a README recipe that named the workflow which does the opposite of what it claimed.

1. JOURNAL.md contained 720 duplicated lines

Lines 978-1697 were a byte-exact copy of lines 258-977: five whole sessions (2026-09-02 second, 2026-09-02, 2026-09-01 second, 2026-09-01 first, 2026-08-30) written twice, 37% of the file. Verified as an exact match before deleting rather than merged by hand, and the file went 1948 to 1228 lines with no unique content lost. Nothing flagged it because nothing reads this file mechanically; it only surfaced when a duplicate-line count was run as part of the audit.

2. Four substance defects, all in agent context files

These are the ones that would have cost someone real time, and they are worth more than the punctuation pass that surfaced them:

Smaller: opinionated-omarchy/CLAUDE.md introduced a three-item list as "two ways", and skillbench/README.md and CLAUDE.md disagreed on the skill's context cost (~3.2k vs ~3.1k tokens). Measured at 12557 characters after frontmatter stripping, so ~3.1k, and both now say that.

3. What was changed, and what was deliberately not

Clean: CLAUDE.md, JOURNAL.md, NOTICE, opinionated-omarchy/CLAUDE.md, research/README.md, research/bench/README.md, research/bench/nexus1-baseline-2026-08.md, research/validation/README.md, research/assets/fonts/README.md, skillbench/README.md, skillbench/MODELS.md, skillbench/benches/CLAUDE.md, the writeup, and the prose build_site.py emits.

Four bodies of text were left alone on purpose, and the reasons are not interchangeable:

Inline code comments were also left. The skill's scope excludes code, and the line worth holding is between a comment and a string a user actually sees. Four generator strings in build_db.py and ask.py were fixed, because they are prose: they print to the CLI and are written into the tracked research/docs/ pages. research/docs/ was regenerated and the diff is exactly those strings, 17 lines across 5 files, nothing in a record.

4. Agent instructions seeded in three places

scripts/check-writing.ps1 needs PowerShell, which is not installed here, so both review passes were manual. CLAUDE.md carries a git ls-files | grep equivalent scoped to the audited set, run and verified: every line it returns is a code fence, a copyright range, or a numeric range.

5. Verified

research/tests/run.sh 13/13, build_db.py and build_site.py both rebuild clean, and the documented check command returns only the exempt shapes. The site was rebuilt and its intro re-read as rendered text rather than as source.

Session of 2026-09-03 (second): a control with headroom, and the public site

1. linux-agentic-deep-triage, the control the DiD needed

linux-agentic-triage scores 0.950 bare, five points of headroom, so it could not show a lift by construction and the difference-in-differences in runs 25/26 was weak in both directions. The new control scores 0.759 bare (run 27): about 24 points of room, the same difficulty profile as the Omarchy trap bench, and nothing the skill mentions (contamination checked: 0 files in the bundle mention lsof, drop-ins, systemctl edit, deleted files, df, du or inodes).

Two tasks, both with an attractive wrong answer:

Hand-verified across every outcome state, and a reboot scores below doing nothing (3/6 and 2/5 against floors of 4/6 and 3/5) via a marker in /run, which is tmpfs.

Three assertion bugs found in verification, all the same family as the 2026-08-29 tilde bug, and all green before the agent runs:

test_control_benches_are_flagged failed when the bench was added, which is the test doing its job: the control set is pinned so adding one is deliberate.

2. The public site: research/tools/build_site.py

Generates the repo-root docs/ (GitHub Pages' default), not research/docs/, which build_db.py unlinks on every corpus build. 456 record pages, a client-side symptom search over a 193 KB JSON index, and 12 category groups.

The design is the Control Room theme, recovered from work.handoffs at 2026-08/22-controlroom-clean-room-extraction. That is a draft, human_reviewed: false document whose section 1 is transcribed verbatim from the original style.css (the reliable part; its sections 2-3 are unreviewed judgement and were not used). All five motifs are reproduced: scanline+vignette, twin corner gradients, phosphor traces, glowing LEDs, and accent-keyed chrome.

The mapping that makes it fit this project: audit_status drives the status LED. Teal for audited, amber for corrected, red for unchecked. The board therefore reads the corpus's honesty at a glance rather than burying it, and the conditional cause_reconciled disclaimer is reproduced on the record page exactly as ask.py and the markdown do it. A site that told a reader the cause "was not rewritten" about one that was would be the same defect in a third place.

The Control Room's five group accents are mapped onto the twelve fix categories in thematic families (Omarchy core/theming teal, Hyprland/display/wayland blue, GPU/boot/power purple, pacman/apps amber, network/audio pink). Twelve distinct accents was considered and rejected: past roughly eight, categorical colours stop being reliably distinguishable under colour-vision deficiency, and the group name is always present as text, so colour is redundant encoding here rather than the only channel.

docs/ is gitignored. It is 4.3 MB regenerated wholesale on every run; committing it would add that churn to every corpus change. Pages can build it in CI instead. That is a decision still open, along with self-hosting the two webfonts: the source handoff is explicit that fetching Space Grotesk and IBM Plex Mono from a CDN at build time yields a silently unstyled page on a network blip, so the generator ships fallback stacks only.

3. CI publishes the site, and three theme defects fixed

.github/workflows/pages.yml builds the site from the JSONL and publishes the artifact, so the 4.3 MB never enters history. It sanity-checks before publishing: one page per record (a silent drop would otherwise ship a smaller corpus that looks complete), and that UNAUDITED still renders. If the generator ever stops emitting provenance, that must fail the build rather than publish a corpus reading as uniformly trustworthy. Needs the repo public and Pages set to "GitHub Actions"; until then the deploy step fails, which is the honest failure mode.

Three defects found by measuring rather than looking, all inherited from the original:

The font question is not solvable by picking a universal font. There is no widely installed face that is both CRT-flavoured and legible at 8px: Courier New is the only truly universal "old mono" and it is a thin, wide typewriter face that fails at label sizes. The fallback chain now covers the three platforms properly (ui-monospace/SF Mono/Cascadia Mono/JetBrains Mono/DejaVu Sans Mono/Liberation Mono), but the real fix is the one the source handoff already mandates: self-host the woff2. Left undone deliberately: it is a licensing and binary-assets decision.

4. Departure Mono, vendored: the CRT chrome, and only one third party

Corrected 2026-09-13. True when written, false since 2026-09-03, when commit 5e726ec vendored a second face: Omarchy Font, MIT, © 2026 Mark Cuda, which took over the wordmark and the group headings. Departure Mono kept only the micro-labels. The section below is left as written because it records why the first face was chosen. The two-face layout and the reason for it are in research/assets/fonts/README.md, and the credit for both designers is on every page of the site, on the front page and in the root README.md as of the session of 2026-09-13 (third).

The font question from section 3 is resolved, and not by finding a universal face. One font is vendored, not two: Departure Mono (22 KB woff2) carries the chrome (wordmark, micro-labels, group headings, the places letterforms are decoration) while readouts, code blocks and sources stay on the system ui-monospace stack, which is good on all three platforms and costs nothing. Vendoring a second family for the data would have doubled the licensing surface for no gain.

Licence, stated precisely because the metadata is wrong. The upstream repo's GitHub API entry reports MIT; the bundled LICENSE is SIL Open Font License 1.1, © 2022–2024 Helena Zhang, and that is the one that governs. No Reserved Font Name is declared, so redistributing the unmodified file is straightforward. OFL clause 2 requires each copy to carry the notice and licence, so build_site.py copies both into docs/fonts/ and the CI job now fails the build if either is missing: a licence breach should not be able to ship quietly.

Pixel fonts need whole-pixel sizes. The transcribed spec uses 8.5px and 9.5px labels; at fractional sizes a pixel face blurs into mush. Every rule that switched to the CRT face is rounded to an integer, and smoothing is disabled on those rules only. That is a deliberate divergence from the transcription, and the reason is recorded in research/assets/fonts/README.md next to the font.

Self-hosted rather than fetched, for the reason the original handoff gives: the Control Room's own Dockerfile curled five woff2 files at build time with || true on each, so a network blip produced a silently unstyled build. 22 KB in git removes that failure mode and every runtime third party at once.

5. The repo is public, the site is live, and the context files were re-audited

<https://techluddite.github.io/opinionated-omarchy/>. Build and deploy both green on the first run, and the paths trigger works: writing the 150 titles rebuilt and republished the site with no manual step.

Every context file was then re-checked against the repo rather than copied forward: 456 records / ok 240 / corrected 212 / unaudited 4 / 0 titleless, 766 sources, 12 categories, 29 cause_reconciled stamps, 17 bench specs (9 Omarchy, 6 controls, 5 agentic), 3 workflow scripts, 43 skillbench tests, 13 corpus tests. What had drifted:

6. Pre-flight for going public, and what the screenshots caught

Rendering the site found three defects that reading the generator would not have:

Pre-flight for the public flip came back clean: no keys or tokens in tracked files or in history (the bench key was never added; it is ignored by a nested skillbench/.gitignore). Two things were fixed first:

7. Scaling settled: per-record files, and FTS5 rather than grep

Per-category storage is already broken, not a future risk. Seven of the twelve category pages exceed the 32K context window today at 456 records; network.md is 43.3k tokens. With per-record files a category is metadata, so splitting one is a field edit, never a migration.

The real limit is match precision, not file size. Unranked grep over an index returns 22 hits for bluetooth, 32 for nvidia, 45 for audio and 94 for boot. Reading those is ~94k tokens, and at 10x growth it is hopeless. ask.py already solves this with FTS5 + bm25 and tuned per-column weights, so the skill should ship the SQLite index, not grep. Confirmed with the operator as the intended direction.

Session of 2026-09-03: n=10, and the lift does not survive it

Runs 25 and 26 replicate runs 23/24 at ten repeats. The +11.1 pt lift does not hold.

benchnoneskilllift95% CIp
omarchy-agentic-stale-advice0.8000.883+8.3 pt[−1.7, +18.3]0.205
linux-agentic-triage (control)0.9501.000+5.0 pt[+0.0, +13.3]0.491
difference in differences+3.3 pt0.807

Read the last row. The skill lifts the general-Linux control nearly as much as it lifts the Omarchy bench, and that is precisely the condition the controls exist to detect: by this repo's own rule it is measuring answer length, not Omarchy knowledge. The n=3 figures (+11.1 vs +5.6, a +5.6 gap) shrank to +8.3 vs +5.0, a +3.3 gap at p = 0.81. Classic small-sample optimism, caught by doing the obvious thing and running it again bigger.

The binary framing agrees exactly and is easier to hold onto: solved outright, 8/20 bare vs 13/20 with the skill, Fisher exact p = 0.205.

1. n=10 was underpowered by about 3x, and that is computable

At the observed effect (0.40 → 0.65 solved) and its variance, 80% power at α=0.05 needs ~62 cases per variant ≈ 31 repeats. We ran 10. So "not significant" here means the experiment was too small to tell, not that the skill does nothing. The honest next step is either 31 repeats or accepting that this effect size is not worth the hours.

skillbench/tools/lift_test.py now does this properly: permutation test, bootstrap CI, and the difference-in-differences, stdlib only. The unit is the case, not the assertion: the 6 post assertions inside a case are heavily correlated, and treating 60 assertions as 60 independent samples would have manufactured significance out of nothing.

2. The control is saturated, which is a real limit on this comparison

linux-agentic-triage scores 0.950 bare, so it has 5 points of headroom and cannot show a large lift by construction. That caps how much signal the difference-in-differences can ever carry, and it means the DiD test here is weak in both directions.

A control needs headroom for the comparison to mean anything. Every existing control was chosen to be easy general Linux; none was chosen to be hard general Linux. That is the flaw to fix before spending 31 repeats on anything.

3. What did hold up

4. The latency fix is NOT live, and runs 25/26 still carry the old semantics

compose.yaml mounts ./benches and ./data but not ./app, so app/runner.py is baked into the image. The queue-wait fix committed on 2026-09-02 therefore did not take effect for runs 25 and 26: it needs docker compose up -d --build. Correcting the previous entry: the old latency semantics apply through run 26, not run 24. (This is also why a new bench YAML is picked up with no rebuild but an app change is not.)

Session of 2026-09-02 (second): the trap seam, and the first agentic lift

Two things were asked for and both landed: qwen3.8:27b is unblocked, and the Omarchy 3 trap seam now has a bench. The lift is positive but not yet demonstrated: the control moved too.

1. qwen3.8:27b was never incapable

Recorded as blocked earlier the same day on one bare case: pi exited 1, a transcript containing only 500: no user query found in messages, and a floor score. Re-run with 3 repeats (run 22) it scores 8/8, 8/8, 7/8, and the 7/8 case had already done most of the work before erroring. It is the fourth capable model, and run directly it produced the best solution of anything tested: it used Omarchy's own omarchy-refresh-limine instead of hand-rolling limine-update, and noticed OMARCHY_PATH is not passed through sudo.

Root cause, and it is Ollama's, not pi's. no user query found in messages is compiled into the Ollama binary. qwen3.8:27b reports family qwen35 with template {{ .Prompt }} (a stub), so Ollama renders it with a built-in renderer for that family which requires at least one user message. One curl reproduces it with no agent involved: system+user returns 200, system alone returns the 500, and so does system+assistant. devstral-small-2:24b accepts all three, which is why only this model trips. Somewhere in a long loop pi sends an array with no surviving user turn. Intermittent, about 1 run in 3.

The lesson is about the scan, not the model: a one-repeat scan cannot separate "cannot act" from "hit an intermittent harness failure". Anything recorded as blocked deserves a re-run before it is believed. This is now written into skillbench/MODELS.md.

2. omarchy-agentic-stale-advice, the first bench with headroom

Two tasks where the widely published answer is wrong on Omarchy 4: the Omarchy 3 ~/.local/share/omarchy git checkout, and hyprlang config that Hyprland 0.55 deprecated for Lua. Both prompts say the change "must survive an omarchy update", the phrasing a real user would use, satisfiable only from the user tree, and it never names the answer.

The design rule worth copying: the wrong answer scores BELOW doing nothing. Verified by hand on a VM across five outcome states:

outcometask 1task 2
does nothing4/64/6
hyprlang hyprland.conf3/63/6
resurrects the Omarchy 3 tree3/6n/a
correct Lua in the user tree6/66/6

A pass/fail assertion cannot separate "did nothing" from "did the wrong thing"; this does.

Both first-draft patterns were green on a file the agent never touched. bindings.lua ships a commented -- hl.unbind("SUPER + SPACE") example and looknfeel.lua ships five commented hl.config( examples plus a commented gaps_in, so hl\.unbind and hl\.config matched the pristine template and task 2 would have scored 6/6 while doing nothing. Fixed with ^[^-]*, which excludes Lua comments. This is the tilde-quoting bug of 2026-08-29 in different clothes, and it was caught only because the schema doc insists on hand-verifying before and after on a real VM. Hand-verify against the SHIPPED templates, not an empty file.

3. Run 23/24: a positive lift, and an honest control that undercuts it

devstral-small-2:24b, 3 repeats, none vs skill:omarchy:

benchnoneskilllift
omarchy-agentic-stale-advice0.8891.000+11.1 pt
linux-agentic-triage (control)0.9441.000+5.6 pt

This is the first positive lift the agentic lane has ever produced, and the trap bench is the first with real headroom: bare hits the floor on rebind-packaged-default in 2 of 3 runs, where every earlier agentic bench was saturated at 8/8.

But do not report it as a result yet. The control moved +5.6 pt, which is one assertion flipping out of 18; the Omarchy delta is four out of 36. Both are a couple of assertion flips on n=3. The direction is right and the Omarchy bench moved twice the control, but the sample cannot separate a real effect from noise. More repeats is the whole next step.

Also worth noting: the two tasks behave differently. looknfeel-not-hyprlang is nearly saturated bare (6/6, 6/6, 4/6) while rebind-packaged-default is not (4/6, 4/6, 6/6). The headroom is in the binding task, and a future bench should lean that way.

4. Agentic latency_s was measuring the queue, and a previous claim was wrong

t0 was set before pool.acquire(), so agentic latency included time spent waiting for a VM. Concurrency there is the pool (2), so a case launched 12th banks the whole queue, and because the runner finishes one variant before the next, the second variant systematically looks slower. That inflated run 23's skill mean to 1572 s against 790 s bare, and it means the "3.0x latency" reported for run 18 overstates the skill's cost; that figure should not be quoted.

Fixed: the clock starts after acquire(), and queue_wait_s is recorded separately. Corrected 2026-09-03: the fix is not live. compose.yaml does not mount ./app, so the runner is baked into the image and needs docker compose up -d --build. The old semantics therefore apply through run 26, not run 24.

Session of 2026-09-02: what the local models can actually do

The agentic bench got its first paired run and its first honest answer, and the answer moves the open problem rather than closing it.

1. Run 18: the paired run, and it is a flat zero

omarchy-agentic-root-config, devstral-small-2:24b, none vs skill:omarchy, 3 repeats:

variantcasessuccessSTATEmean latency
none31.0000.958 (23/24)97 s
skill:omarchy31.0000.958 (23/24)291 s

Identical to three decimals, at 3.0x the latency. The cost of putting the skill in an agent loop reproduces (2.0x and 4.5x in runs 14/15); the benefit still does not exist.

2. Run 21: the reason, and it is not task difficulty

The standing assumption was that the bench needed harder tasks. That assumption is now evidence-against. A bare scan of all 14 tool-capable local models (skillbench/MODELS.md) found:

Every model capable enough to drive the loop is also capable enough to finish the task. The band is that narrow. Harder tasks do not widen it; they just move the three capable models down while the other eleven keep scoring the floor for unrelated reasons.

The floor is 5/8, not 0, which is what makes this easy to misread: five of the eight assertions are "you did not break anything", so a model that does nothing still scores 0.625. Score alone cannot distinguish "did nothing" from "did most of it".

3. Failure modes that look identical in the score column

Four different causes produce the same 5/8, and telling them apart needed transcripts, /api/ps and the error field:

The result worth remembering: gemma4 (8B) reported the merge complete while changing nothing. A transcript-trusting grader scores that a success. It is the sharpest possible argument for why the agentic lane asserts on the machine and carries no transcript checks.

4. Context and VRAM are hard constraints, and they lived nowhere in the repo

Both now in CLAUDE.md, because both are on the ollama systemd unit rather than in this repository and are therefore easy to lose:

5. Three defects in the VM tooling, all silent

Found while getting the pool usable again; all three cost time before they were understood.

6. The grader question from run 17 is answered, with data

The 2026-08-30 entry asked whether STATE is inflated by scoring timed-out cases, and said to decide before the next paired run. Computing it both ways over run 17 gives state_ok == state_all == 1.000: every timed-out case had genuinely passed all its post assertions. STATE is not inflated. STATE and success measure different things (did the machine end up correct, versus did the agent exit within budget) and both were right. No grader change. Run 18 had no timeouts, so it does not bind there either.

Session of 2026-09-01 (second): writer tests, the first live scenario, and root in the bench

Item 4 of "What's left" is closed, and the VM spot-checking under item 5 has started.

1. The corpus writers now have one schema definition, and tests

ingest.py's FIELDS allowlist was missing cause_reconciled. That was predicted in the backlog ("read ingest.py for the same defect, before it is next run") and it was real: the replace path projected records onto a private list that never learned about the field added on 2026-08-30.

Adding the string to the second list is not the fix; two lists that must never diverge is the defect. research/tools/corpus.py now holds the only FIELDS, plus the only read_jsonl / write_jsonl, and both writers import it. That also pins newline="\n" and encoding="utf-8" in one place instead of at each call site.

write_jsonl raises on a key it cannot classify. The harvest genuinely emits chaff the corpus drops on purpose (cause_note, cause_extra, verify_note, enumerated from the raw payloads into WORKFLOW_ONLY), so a blanket raise would have broken every merge. The distinction is the point: dropping enumerated chaff is the projection working, dropping an unclassified key is an unfinished schema change.

13 tests, stdlib unittest, research/tests/run.sh. Not pytest: the corpus tooling is dependency-free by design so it runs on a bare container, and a suite that needed pytest installed would be the first thing to break that. Two of them are the ones that would have caught 2026-08-30: FIELDS is asserted against schema.sql and against the live corpus, in both directions.

The tests were checked against the defect, not just run green. Removing cause_reconciled from FIELDS again produces 3 failures and 8 errors, and both invariant tests name the missing field in their message. An assertion that cannot fail is the thing this project keeps catching itself on, so it gets proved rather than assumed.

Two consumers of the four are now checked automatically. schema.sql, build_db.py, ask.py and corpus.py still all have to be edited by hand for a new field.

The refactor is provably neutral: reading all 456 records and writing them back through the new code is byte-identical to data/problems.jsonl (sha256 fd5f5bfe653745e8…). Key order is load-bearing (it is the key order of every line on disk), so FIELDS may be appended to but never reordered.

2. The first live scenario: a corpus record exercised on a real VM

New directory research/validation/: induce a problem on a throwaway VM, apply a fix, assert on the machine, append to runs.jsonl. Read its README before adding one; the trust model is the whole design.

Validation never touches audit_status. That field means "checked against its sources", and one VM agreeing is not a source confirming. Results are a separate append-only log because a record has one audit but many runs, each with its own date and Omarchy version. And repair: is explicitly an operator's reading of the record's prose, not the record: record fixes are branching prose, and only 6 of 456 have a fenced verify block, so nothing here can execute "the fix" itself.

First scenario: mkinitcpio-pacnew-unhandled-breaks-next-boot on omarchy 4.0.1-1. 6/6 assertions pass: the record's remediation advice is sound, and its Omarchy-vs-plain-Arch branch is confirmed by the system itself: /usr/local/bin/mkinitcpio is a wrapper from limine-mkinitcpio-hook that warns This does not update Limine boot entries and offers limine-mkinitcpio instead, exactly as the record says.

But three claims in that record are wrong for Omarchy 4, and the source audit missed all three. /etc/default/limine is owned by no package, so it can never produce the .pacnew the record's symptom block quotes; /etc/mkinitcpio.conf is [unmodified] on a stock install, so neither can it; and the danger claim that overwriting mkinitcpio.conf "removes your encryption, plymouth and btrfs hooks" is false, because those come from a package-owned drop-in sourced afterwards that assigns HOOKS= wholesale. Measured, not inferred. Generic Arch advice mis-specialised to Omarchy, the same family as the Omarchy 3 → 4 tree split.

These are recorded in the validation README, not edited into the corpus. A correction needs an audit_note and a cause_reconciled stamp through merge_gapfill.py; a silent rewrite is exactly what the provenance fields exist to prevent.

3. What the spike cost, which was the point of running it

The harness is written once. Per record after that, the expensive part is neither the seed nor the assertions: it is establishing ground truth on a live machine first. Eight ssh round trips went into learning that /boot is a root-only vfat ESP, that Omarchy boots a UKI and has no vmlinuz, that the hooks live in a drop-in, and that /etc/default/limine is unowned. Only after that could the assertions be written correctly.

Two of six assertions were wrong on the first run and both failed for the wrong reason: one asserted /boot/vmlinuz-linux, which does not exist on this system, and did so as an unprivileged user against a dmask=0077 mount, where test returns 1 for "unreadable" and looks identical to "absent". A third re-ran limine-mkinitcpio inside an assertion, making grading a side effect that rebuilt the boot image; that is why the runner now has repair_output_* types.

Realistically 30–45 minutes per scriptable record, more for anything needing a reboot to observe. So this scales to a few dozen records, not 456; roughly a third of the corpus is out of reach of these VMs anyway (67 nvidia / 49 intel / 47 amd against a virtio GPU, 297 laptop, most of power-suspend, much of network). Spot-check and bench-source, not corpus validation, and the README says so.

The golden-image workflow was exercised end to end and matches its documentation: reset 0.76 s, ssh reachable ~7 s later. virsh shutdown is still ignored; poweroff over ssh works.

4. Item 1 is unblocked: the agentic lane could not reach root, and now can

Writing the scenario into a bench turned up why omarchy-agentic-config is saturated, and it is not that nobody wrote hard tasks. Every task in it is a ~/.config edit because that was the ceiling. The bench drives the VM over ssh with no tty (vm.py run() is bash -lc with nothing on stdin), and the bench user had no passwordless sudo, so nothing requiring root could be seeded, performed by the agent, or asserted. Userspace config is the easy end of Omarchy, and the bench could only ever measure that end.

tools/provision-bench-vm.sh now installs /etc/sudoers.d/99-bench-nopasswd, validated with visudo -c so a broken sudoers is never shipped, and using id -un rather than $USER. $USER is set by login(1) and is frequently empty in a non-interactive ssh, which would have written a rule for the wrong name or none.

Both golden images were re-saved with it baked in, because the first reset silently dropped it and a sudo assertion would then fail looking like a bench bug. Verified: a reset now comes back with NOPASSWD intact and no leftover bench state.

This is safe on these VMs and nowhere else: disposable, NAT-only, no real data, and their password is already committed in plain text. It does let a misbehaving agent break the machine, which is what the 0.76 s reset is for.

5. The first deliberately hard agentic bench

skillbench/benches/omarchy-agentic-root-config.yaml. One task: resolve a .pacnew for the Limine boot config, keeping both the operator's local setting and the new upstream one, and regenerate the boot config.

The difficulty is structural rather than obscure. Both wrong answers are single plausible commands that exit 0 and look like completion: mv the .pacnew over the live file adopts upstream and destroys the local edit; rm keeps local and silently discards the new option. Asserting on both values means each shortcut fails a different assertion, so the bench says which mistake was made rather than just "failed".

Checked by hand on a VM as benches/CLAUDE.md requires, all four states:

statescorewhat failed
after seed, nothing done5/8both values, the .pacnew, the rebuild
mv the .pacnew over the file7/8the local setting
rm the .pacnew7/8the upstream setting
a real merge, then rebuild8/8n/a

The remaining five assertions are collateral-damage guards (hooks intact, UKI present, pacman -Qkk omarchy clean) which correctly pass on an untouched machine and are there to catch an agent that succeeds destructively.

Two things in it are worth copying and are written into benches/CLAUDE.md: the "both wrong answers are one command" shape, and the fact that a seed touching /etc must be idempotent, since defaults.vm.restore is $HOME-relative and /etc persists between cases. That bench keeps a pristine copy on first run and re-derives from it.

All 43 skillbench tests still pass with the new spec, so it loads, its assertions compile, its paths are absolute, and the control set is still pinned.

Not yet run against a model. The bench is validated as a measuring instrument; what it measures is the next session's work, and it needs a paired bare/skill run to say anything about lift.

6. A lead worth someone's time

/etc/mkinitcpio.conf.d/omarchy_resume.conf is HOOKS+=(resume), which appends resume after filesystems, fsck and btrfs-overlayfs. Corpus record resume-hook-after-filesystems-hibernation (one of the 4 remaining unaudited records) describes that exact ordering as a problem. Either the record is wrong, or Omarchy ships the broken ordering by default. Observing the ordering does not establish which, and guessing would be the fabricated-precision failure mode again. Check it against a source.

Session of 2026-09-01 (first): the last unaudited records, and the Windows remnants

Item 2 of "What's left" is closed, and the repo no longer carries anything from its Windows era.

1. The 28 unaudited records are audited

wayland-compat (12) and network (16) had been harvested but never reviewed since the gap-fill pass, whose audit agents died on API streaming errors. Four auditors, batched by category, returned 28/28 verdicts:

okcorrectedreject
wayland-compat570
network781
total12151

27 of 28 verdicts are confidence: high. Seven corrected records also had their cause rewritten and stamped cause_reconciled: 2026-09-01, so the corpus now carries 29 stamps across two dates. The corpus is 456 records, ok 240 / corrected 212 / unaudited 4.

The one rejection is the most interesting result. usb-tethering-renamed-wwan-networkmanager-ignores described a real kernel regression: the auditor confirmed commit 67d1a89 "rndis_host: Flag RNDIS modems as WWAN devices" and verified by fetching drivers/net/usb/rndis_host.c at each stable tag that wwan_rndis_info appears in exactly the versions the record named. But the patch was reverted, and is absent from v6.15, v6.17 and master. The record presented a six-week 2025 regression as a live problem at frequency: common, on a workstation running 7.1.8. Its own cited thread said so at post #60. Separately the auditor traced NetworkManager's source to show the record's headline fix cannot work even on an affected kernel, because nm-wwan-factory.c unconditionally sets *out_ignore = TRUE, so no device object is ever created for nmcli to act on.

That is the fabricated precision failure mode from the 2026-08-30 session showing up once more, in its most convincing costume yet: every specific in the cause was checkable and correct, and the conclusion was still wrong because nobody checked whether the story had ended.

2. Two blockers in that path, both fixed first

Neither would have failed loudly.

Both are written up in full in writeups/2026-09-01-merge-gapfill-silent-defects.md, including the two follow-ups they leave open: the corpus tooling has no tests at all, and adding a schema field still has no checklist covering its four consumers.

The second was caught by dry-running the merge against a copy of the corpus with synthetic verdicts before letting it near the real file; worth doing again, because the failure is silent and the evidence it destroys is the evidence of honesty.

That dry run also confirmed the property that actually makes this merge safe: merge iterates every record in the audited category, not just the targeted ones. Records with no verdict are preserved, so scoping the verdict set is what protects the 33 already-audited wayland-compat records. The workflow filters verdicts to the assigned slugs for the same reason.

3. The audit workflow is now a third script in the repo

research/tools/audit-existing-workflow.js. The repo had two workflow shapes and neither audits a record that already exists: gapfill-workflow.js harvests against auditor-named gaps and audits only what it just wrote, and its one audit-existing path is Track A, hardcoded to apps-services. That gap is exactly what made the old item-2 recipe wrong, so the fix is a script rather than a note.

Retarget it by editing BATCHES; slugs are listed explicitly rather than derived, which keeps the run resumable and means a verdict can never land on a record you did not name. Batches of 6-8 leave each agent enough budget to actually fetch each record's sources. The obvious next user is the 4 remaining unaudited records.

4. The Windows era is gone

The repo was converted from Windows earlier; two things survived that conversion.

Two things that look like Windows remnants are not, and should be left alone.

Session of 2026-08-30: catching the journal up, and starting the re-audit

Housekeeping first, then item 3 of "What's left". Three things had drifted out of the written record since the 29th.

1. Runs 16 and 17 happened and were never written down

Both on muse-glimmer:30b, a model the agentic lane had not been tried with before.

Run 16 (linux-boot-partition-full, chat lane) is unremarkable and that is the point: 1.000 in both variants, 91 prompt tokens bare against 3197 with the skill. A control behaving exactly as a control should.

Run 17 (omarchy-agentic-config, agentic lane) reports 0.333 bare against 0.667 with the skill, and that spread is entirely timeout attrition, not skill signal. Every case that did not error passed every assertion; the run's STATE score is 1.000 in both variants. The whole difference is which cases hit the wall:

tasknoneskill:omarchy
binding-user-treeok, 176 sok, 251 s
monitor-persistagent exceeded 600sagent exceeded 600s
theme-switchagent exceeded 600sok, 847 s

So the bench is still saturated exactly as the 29th recorded it; this adds only that a 30B model does not fit the current agent_timeout: 600. Do not read run 17 as a lift.

2. A grader question run 17 raised, still open

Post assertions are evaluated on timed-out cases too, and they passed there. Both monitor-persist cases errored on the timeout, yet each reports post 3/3 and contributes state = 1.0. That is why the run shows STATE 1.000 alongside a success rate of 0.333.

This may well be correct: an agent can finish the edit and then burn the remaining budget without exiting, and refusing to score work it demonstrably did would be its own distortion. But it means STATE currently averages over cases the runner classified as failures, so STATE and success can disagree without either being wrong, and a reader comparing them will assume one is broken. Decide which it is before the next paired run, and whichever way it goes, make the UI say so. This is the same class of thing the control caught in run 12/13: not a wrong number, but a number whose meaning is not written down.

3. Smaller drift

4. Item 3 of "What's left" is closed: 22 stale causes, not 130

The backlog said "roughly 130 corrected records may still carry a cause the auditor disproved". That number was the size of the population nobody had looked at, never a count of defects. All 130 have now been read, and 22 actually had a cause their own audit note contradicts. The other 108 are fine: their notes open with "the diagnosis is right", "cause verified exactly", "the mechanism is real", and go on to object to the fix, which the first-pass audit had already rewritten.

Scoping it took provenance, not keywords. The README records that the second pass's auditors could return a corrected_cause while the first pass's could not, so the affected set is exactly the corrected records whose audit came from the first pass. Reconstructing that from raw/harvest-result.json and raw/gapfill-result.json (apps-services was the only category re-audited in pass 2; everything else in gapfill-result audits the new records) yields 130 records, and independently reproduces the README's "20 causes replaced" figure, which is what confirmed the mapping was right. Keyword heuristics on the notes gave a union of 123 and would have been both wrong and unfalsifiable.

The rewrites needed no new source work. The auditor had already done it; the first pass simply had no field to put the result in. Each new cause is written from its own note. Two failure modes dominate, and both are worth recognising again elsewhere:

A schema change was needed to stay honest. Rewriting a cause silently would have destroyed the distinction between "cause checked and correct" and "cause never revisited". Records now carry cause_reconciled (a date, or absent), it is a column in schema.sql, and the disclaimer under the audit note in both ask.py and the generated markdown is now conditional on it: previously it told every reader of all 197 corrected records that the cause "was not rewritten", which is now false for 22 of them. A blanket disclaimer that is wrong for part of its audience is the same class of problem as the assertion that could not fail: it stops carrying information.

5. Agent context files brought current, and two gaps filled

Three counts had gone stale as the repo moved, all in the direction that matters: they undercounted work that had landed:

Two supplemental files were added, both where an agent would otherwise have had to reverse-engineer from source.

skillbench/benches/CLAUDE.md, the bench-spec schema. This was the real gap: the README explains why the bench exists across 310 lines but never documents the spec, so writing a bench meant reading spec.py, checks.py and vmchecks.py. That matters right now, because "write harder agentic tasks" is the top open item. It carries the full post: and checks: type tables, the three loader/test rules that are enforced (name must match filename, agentic-without-post is refused, every lane needs a control), and both grader traps that invalidated runs 12/13 (the tilde-quoting asymmetry and pgrep matching its own shell) because those are exactly what a new bench author will reproduce.

opinionated-omarchy/CLAUDE.md, what the skill has to be. That directory had a zero-byte .gitkeep and nothing else, and it is the one place an agent is most likely to start writing from a blank slate. It now records the retrieval-shape problem (the corpus does not fit in context and "run ask.py" is not a skill), the +29.3 pt Omarchy / −2.3 pt control baseline it has to beat, the instruction to write its benches before tuning it, and the provenance it must not launder: 28 records are still gapfill-unaudited and must not reach a user reading as audited. It replaces the .gitkeep, which a real tracked file makes redundant.

Every figure in these files was checked against the repo rather than copied forward: bench counts and control names from the YAML, 43 tests from a run, 457/228/197/28/4 and the 22 reconciled from the JSONL, 911 cases and the two lift figures from research/bench/, 1.6 MB from du. The named tests were confirmed to exist before being cited.

Session of 2026-08-29: the agentic lane

The Skill Bench now grades what an agent does on a real machine, not only what a model says. This was the top item on "what's left" and it is built, running and documented.

1. How it works

A bench declares lane: agentic. Each case then: acquires a VM from a pool, restores the paths the bench declares, applies a seed: (the breakage), runs pi --print --skill <dir> in a tmux window, and evaluates the task's post: block over ssh.

Two graders, kept apart in the UI on purpose: QUALITY (transcript) and STATE (the VM). Collapsing them would hide a model that describes the right edit and never makes it.

omarchy-agentic-config is the first bench: add a keybinding, switch a theme, configure a monitor. Every task also asserts pacman -Qkk omarchy reports 0 altered files, which is a hard statement that the agent stayed out of /usr/share/omarchy, the exact place Omarchy-3-era advice sends it. A chat bench can only ask whether a model mentions the right directory.

2. What the paired runs showed: no signal, because the bench is SATURATED

Runs 14 and 15, devstral-small-2:24b, none vs skill:omarchy, 3 repeats, the Omarchy bench against its new control:

benchnoneskilldelta
omarchy-agentic-config24/24 = 1.00022/24 = 0.917−8.3 pt
linux-agentic-triage (control)17/18 = 0.94416/18 = 0.889−5.5 pt

The bare model already scores 100% on the Omarchy bench. There is no headroom for a skill to help, so this bench cannot measure skill efficacy against an agent this capable. Both deltas are slightly negative and the control moved nearly as far as the Omarchy bench (−5.5 vs −8.3), which, by the rule the controls exist to enforce, means these numbers are not measuring the skill at all.

The tasks need to be harder before this bench discriminates. Writing tasks that a good agent fails without the skill is the actual open problem; three tasks a competent agent does by default measure nothing.

The one clear, reproducible effect is cost. Mean case latency:

benchnoneskill
omarchy-agentic-config364 s714 s2.0×
linux-agentic-triage33 s146 s4.5×

It shows up in the control too, so it is the cost of putting ~3.1k tokens into an agent loop, not anything Omarchy-specific. Two of 36 cases hit the 600 s timeout, one in each variant, so that part is not attributable to the skill.

2a. The first paired run was wrong, and the control is what caught it

Runs 12/13 reported +12.5 pt / −5.6 pt. Those numbers were invalid and are retained here only as a lesson. Two grader bugs:

The tell was the control scoring exactly 0.500 on every task in both variants, a constant rather than a measurement. A control that cannot move is a broken control, and catching that on its first outing is precisely what it is for. Three regression tests now cover it, including one that scans every shipped bench for post paths that cannot resolve.

3. The lesson that changed the bench: good agents do not narrate

devstral scored 3/3 on the monitor task while its entire transcript was "Task completed." The chat lane's checks: were therefore scoring verbosity and marking the best-performing agent down for being terse. Forbidden-pattern checks are no better: a silent agent passes them all and reads as 100%.

So the agentic bench now carries no transcript checks at all. In this lane the machine is the measurement.

4. Four things that cost real time, all now written down in CLAUDE.md

5. Golden disk images replaced snapshots

/var/lib/libvirt/images is btrfs with no NOCOW flag, so cp --reflink=always shares extents: reset 1 s, boot to ssh 14 s, full cycle ~30 s, verified with a marker file, against minutes for an ISO rebuild. tools/golden-test-vm.sh save|reset|status.

Note virsh shutdown (ACPI) is ignored by these VMs; use ssh <vm> 'sudo systemctl poweroff'.

6. The VMs are now provisioned, and mirrored to their consoles

tools/provision-bench-vm.sh makes a VM a bench target, idempotently: SDDM autologin, never lock/blank/suspend, a systemd user unit holding a long-lived tmux session, a second unit attaching a foot terminal to it read-only on the console, and pi pointed at the host's Ollama. tools/install-bench-key.sh gives the bench its own ssh key (never the operator's), gitignored under skillbench/secrets/.

Read-only is load-bearing both ways: a watcher cannot type into a running case, and the runner therefore must never use tmux send-keys (tmux refuses it outright). Launching each case as a window is the supported path.

7. Isolation: what it is, and what it is not

Before every case the VM is restored from a tar of the paths the bench declares. Anything outside those paths persists. Verified working: an agent invented ~/.config/omarchy/keybinds.conf and the next case's restore removed it.

It is not a disk rollback, deliberately: the container runs cap_drop: ALL, no-new-privileges and has no libvirt socket, and handing it the hypervisor would trade a real security property for convenience. The disk reset stays an operator action between runs.

What's left

1. Replace the agentic bar, because the incumbent has nothing to regress

Answered 2026-09-04, cleanly. Runs 28 and 31, both n=31, both without a lost case: Omarchy lift +1.9 pt (p = 0.59), control lift +1.7 pt (p = 0.47), difference in differences +0.2 pt at p = 0.98. skill:omarchy moves general Linux as much as it moves Omarchy, which is the condition the controls exist to detect.

The measurement question is closed. What is open is a decision, and it belongs to whoever writes the skill:

Unchanged: only 4 of 14 local models can drive the loop at all (skillbench/MODELS.md), so difficulty is not the lever, wrongness is.

2. 150 records have no title (DONE 2026-09-03)

All 150 written and merged. Every record now carries a title and no generated heading is a bare slug.

Worth recording why they were missing, because it says something about the harvest: the gap was not scattered. It sat in exactly six categories at 23-27 records each (audio-input, hyprland-config, display-monitors, wayland-compat, pacman-aur, omarchy-core): one harvester pass that never emitted the optional field, not 150 independent omissions. title is optional in the schema and nothing validated it, so the gap survived two audits and only surfaced when the site was rendered and a third of the pages were headed screenshare-black-screen-no-portal.

Written from each record's own symptom, in the voice the existing titles already used: imperative and specific, naming the error where there is one. Applied through corpus.write_jsonl so field order, LF and UTF-8 stayed pinned, and verified as a title-only change: 0 non-title field differences across all 456 records, 150 titles changed, every one of them previously empty.

The site's slug-prettifier is now dead code for this corpus but is kept: title is still optional, so the next harvest can reintroduce the gap.

3. Finish auditing 28 records (DONE 2026-09-01)

All 28 audited: 12 ok, 15 corrected, 1 rejected and removed. No record carries gapfill-unaudited any more. See the 2026-09-01 session entry above.

The recipe that used to live here was wrong, and is worth keeping as a warning. It said to edit GAP_CATEGORIES down to those two categories and re-run the gap-fill workflow. That list drives the harvest phase, and the harvest had already succeeded: those 28 records are its output; only gapfillAudit came back NONE. Running it would have re-harvested the same topics as -2 suffixed duplicates and left all 28 exactly as unaudited as before. Track B has no audit-existing path at all; that shape exists only in Track A, hardcoded to apps-services.

4. Stale cause fields on first-pass corrected records (DONE 2026-08-30)

All 130 reviewed, 22 rewritten, each stamped cause_reconciled. The "~130" was a worst-case bound on an unreviewed population, not a defect count. See the 2026-08-30 session entry above.

5. The corpus tooling has no tests, and one script is unread (DONE 2026-09-01)

Closed by the second session of 2026-09-01. ingest.py did carry the predicted defect; there is now one FIELDS in corpus.py and 13 stdlib tests, verified against the defect rather than merely run green. The four-consumer checklist is written into CLAUDE.md and two of the four are now checked automatically. Original text follows.

New, and a direct consequence of what the 2026-09-01 audit turned up. Two defects in merge_gapfill.py would have silently destroyed the cause_reconciled provenance, and nothing in the repo would have caught either: skillbench/tests/ covers the bench, and there is no test anywhere for build_db.py, ask.py, merge_gapfill.py or ingest.py.

Two concrete follow-ups, in order of value:

Also worth writing down: adding a field to the record schema currently has four consumers and no checklist: schema.sql, build_db.py, ask.py, merge_gapfill.py. Three were updated when cause_reconciled landed and one was missed.

Full detail in writeups/2026-09-01-merge-gapfill-silent-defects.md.

6. The corpus prose has not been through the writing standard (DONE 2026-09-13)

Closed. research/data/problems.jsonl was the last body of prose in the repo that had not been through writing-and-responding, and it is now. Two passes: 1,366 em and en dashes across 335 records, then 613 semicolons and 183 spaced hyphens across 273 records. See the sixth session of 2026-09-13.

The three worries recorded here before the work turned out to matter in different proportions than expected, which is worth keeping:

Original text follows.

Everything outside research/data/problems.jsonl was audited against writing-and-responding on 2026-09-03 and is clean. The corpus itself was deliberately left alone, and it is the largest remaining body of prose in the repo.

The measurement on 2026-09-03: 1,880 em and en dashes across 424 of 456 records (1,839 across 418 after the 2026-09-06 corrections), concentrated in fix (658), audit_note (412), danger (253), cause (255) and symptom (247). Those render straight onto the public site's record pages and into research/docs/, so the front page now reads to one standard and the 456 pages behind it do not.

Three things make this bigger than a find-and-replace, and they are why it was not done in the same pass:

Do not fix these piecemeal. A partial pass leaves the untouched records looking like a deliberate editorial choice rather than work not yet done.

7. Optional / not started

8. Six ways to make the corpus true on Omarchy 4, not only source-consistent

Recorded 2026-09-06 after fifteen of fifteen ok records checked against Omarchy 4 needed correcting. Cheapest and most forward-fixing first. Recommendation at the time: O1 and O2 now, O3 on pacman-aur next, O4 and O6 before the generic harvest is run again.

9. Backlog: Omarchy plug-ins as corpus content

Idea from the operator, 2026-09-06. Omarchy has a plug-in ecosystem: a plug-in installer the operator rates highly, one or two plug-ins of the operator's own, and a plug-in library. None of it is in the corpus, which has no plugins category. The library is a natural source for a common-problems-and-fixes list, in the same shape as the existing categories, and the operator's own plug-ins are cases where the fix can be verified against the author. Not started: the installer and plug-ins are not on this workstation yet, and their names are not recorded here until they can be checked. Recorded in the Substrata ideas backlog as well.

Session of 2026-08-28/29: the Skill Bench landed

Two things happened: the lab's skill-efficacy data was brought into this repo, and a dedicated bench container was built here to extend it.

1. This workstation is ohmy-omarchy, and that changes what is easy

Worth stating plainly because it was not obvious: the dev box is the lab's local LLM endpoint. RTX 3090, Ollama on 0.0.0.0:11434 with eleven models, including qwen2.5, the exact model every nexus1 Omarchy measurement was taken on. Docker 29.7.2 was already installed and running.

So the bench needs no LiteLLM, no API key, no cloud, and no lab round-trip. Routing through nexus1 would mean ohmy-omarchy → nexus1 → ohmy-omarchy, since nexus1's LiteLLM points back here.

2. The nexus1 baseline is recorded: research/bench/

911 graded cases, ten models, twelve runs, pulled from the lab's Postgres with the bench specs and skill provenance alongside. The headline: the Omarchy skill lifts Omarchy-specific tasks +29.3 pt on average and the general-Linux control tasks −2.3 pt. That gap is the whole argument: a skill that merely made answers longer would lift both.

omarchy/SKILL.md here is byte-identical to what nexus1 benched (sha a8d88cf…), so those numbers are a baseline to reproduce, not merely a reference. diagnose-crash is not identical (5710 bytes here vs a 4173-byte asset there); treat its figures as indicative only.

It is in research/bench/, not research/docs/, because build_db.py:167 unlinks every *.md in docs/ before regenerating. A hand-written page there survives until the next corpus build.

3. skillbench/, a bench container in this repo

Omarchy-only port of the lab's Skill Bench: one container, local Ollama, SQLite, http://127.0.0.1:8878. Dropped as unnecessary: LiteLLM, Postgres, Authelia, budget guard, cost projection, model registry, the paid lane, and 14 non-Omarchy benches.

Two deliberate improvements over the original, both methodology rather than features:

Also fixed by construction, both open findings in nexus1's own audit of its bench: runs orphaned by a restart are reconciled at startup instead of wedging at running, and /readyz actually probes the DB and Ollama rather than only proving the HTTP process is alive.

Verified working, against the baseline. 27 unit tests pass (skillbench/tests/run.sh). The ten-model replication of nexus1's run 33 on omarchy-monitor-config:

nexus1here
bare0.5290.729
skill0.8000.971
lift+27.1 pt+24.3 pt
models improving10/1010/10
prompt tokens32663244

The lift reproduces and the token counts are near-identical: the skill body is the same and injected the same way. Absolute levels are higher on both sides, most likely the serving path (LiteLLM vs direct Ollama) plus weights moving under the same tags.

The lesson that came out of it: benches and skills are sha-pinned, but model weights cannot be pinned. Trust deltas within a run; distrust absolute scores across runs separated by time. Written up in both READMEs.

4. The firewall gotcha, again, in a new costume

Containers could not reach Ollama. Cause: Omarchy's ufw is default-deny incoming and Docker manages only forwarding, the same family as the virbr0 DHCP gap, which is now written up as one pattern in CLAUDE.md rather than two incidents.

The subtlety worth keeping: an interface-scoped rule on docker0 does not work, because Compose puts the container on its own generated bridge. compose.yaml therefore pins the network to 172.28.7.0/24 so the rule can name it.

5. Housekeeping

omarchy-old/ deleted: its purpose was never recorded and could not be reconstructed. opinionated-omarchy/ is confirmed as the destination for the skill this repo is building.

Where the references are

WhatWhere
Corpus design, schema, trust modelresearch/README.md
The 456 records (source of truth)research/data/problems.jsonl
Generated per-category readingresearch/docs/*.md
The public site<https://techluddite.github.io/opinionated-omarchy/>
Site generator, and the CI that publishes itresearch/tools/build_site.py, .github/workflows/pages.yml
Which local models can run the agentic laneskillbench/MODELS.md
Is a measured lift real?skillbench/tools/lift_test.py
Search / build / ingest toolingresearch/tools/
Deep-research report (13 verified findings, 12 refuted folk fixes)research/raw/deep-research-report.json
Raw workflow output, kept for provenanceresearch/raw/harvest-result.json, research/raw/gapfill-result.json, research/raw/audit-28-result.json
Post-mortems worth keeping outside the journalwriteups/
Gaps the auditors namedresearch/raw/gapfill-todo.json
Per-category harvest/audit countsresearch/raw/harvest-stats.json
Test VM build / viewer scriptstools/make-test-vm.sh, tools/view-test-vms.sh
Test VM credentials, VNC ports, ufw rulesCLAUDE.md → "Test VMs"

The refuted list in the deep-research report is worth a read on its own; it is mostly widely repeated folk fixes that primary sources actually contradict.

Gotchas that cost time

Full list in CLAUDE.md. The ones that bit hardest:

From the VM work, and both cost real time:

JOURNAL.md · view on github