Repository navigation
Conversation
399431f to
21992f0
Compare
The lock service blanks the display with a one-shot five-second timer that is re-armed only by input reaching the lock surface or by a password check finishing. Hyprland lights DPMS on its own for any pointer motion or key press (mouse_move_enables_dpms, key_press_enables_dpms), and when that input produces no event on the lock surface - one stray sensor count from a mouse at 03:29, in the captured case - the spent timer stays spent and the panel is on until someone touches the machine. On an OLED that is the lock screen burning in overnight. Listen to ext-idle-notify while locked, the protocol the idle service already uses, and re-arm the blank on every resume unless a password check is in flight. Inhibitors are not respected here on purpose: a monitor that respects them treats one appearing as activity, which is the wake omacom#9152 proposes removing from the idle service. The monitor's timeout is shorter than the blank countdown so it is already idle by the time the display goes dark and the next input is a transition it reports. The change never runs the wake and never shortens anything, so the worst a spurious resume can do is delay a blank by five seconds; slow panels are no worse off than before. Log the blank when its process starts and the re-arm when the timer was spent, and expose blankArmed and activityIdle on `omarchy-shell lock status`; until now the shell logged nothing about display state at the lock screen. The five-second budget and the monitor re-add after an HPD drop (omacom#7749,
21992f0 to
be16366
Compare
|
Reviewed at be16366 and reproduced on a worker built from the Omarchy ISO, under a real Hyprland session. Nothing found that needs changing. Reproduction. A kernel uinput mouse parks the cursor in the bottom-right corner, then the session is locked. Once the lock has blanked the panel, the device sends a single
Tests on the worker. The full Related. #8581 re-arms the blank when an output is re-added after an HPD drop (#8580). This PR covers input with no output change. They are different triggers and neither covers the other, so they are not competing fixes. I found no open issue that describes this exact trigger, so I have not linked one. Who checked it. Claude Opus 5.5 reviewed the diff and ran the reproduction. The configured second opinion, Codex Medium, did not run: the bot's daily review budget was spent. This verdict rests on one model's reading plus the worker reproduction. What happens next. It is labelled |
quattro added the fingerprint availability property beside blankArmed and the fingerprint sleep and reach timers beside the activity monitor; both sides are kept whole, so nothing either branch added changes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The idle monitor reports transitions, not input, so a burst of input the lock surface never sees that lasts longer than the five-second countdown resumes it once: the blank fires mid-burst, the next count relights the panel while the monitor is still active, and when the input stops nothing re-arms the spent timer, leaving the panel lit. Seven seconds of outward pointer counts at a screen corner did exactly that on a worker. Re-arm on the idle transition too, when the timer is spent and no password check is in flight. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Co-Authored-By: Codex Medium <noreply@openai.com>
[Medium impact] The PR appears safe to merge; the fixture no longer risks stranding the caller's desktop. SummaryThe PR re-arms lock-screen blanking after activity that never reaches the lock surface.
Reviews (3) · Last reviewed commit: "Isolate lock activity fixture in a dispo..." · Reviewed by Greptile |
|
Reviewed at be16366 and reproduced on a worker built from the Omarchy ISO, under a real Hyprland session. Nothing found that needs changing. Reproduction. A kernel uinput mouse parks the cursor in the bottom-right corner, then the session is locked. Once the lock has blanked the panel, the device sends a single
Tests on the worker. The full Related. #8581 re-arms the blank when an output is re-added after an HPD drop (#8580). This PR covers input with no output change. They are different triggers and neither covers the other, so they are not competing fixes. I found no open issue that describes this exact trigger, so I have not linked one. Who checked it. Claude Opus 5.5 reviewed the diff and ran the reproduction. The configured second opinion, Codex Medium, did not run: the bot's daily review budget was spent. This verdict rests on one model's reading plus the worker reproduction. What happens next. It is labelled |
|
The comment just above was posted by mistake: it repeats the 3 October review of be16366 and no longer describes this branch. This one replaces it. Reviewed again at 204d960. The branch had stopped merging cleanly into The merge (d6e0dd6). The fix (204d960). The idle monitor reports transitions, not individual inputs. Input the lock surface never sees, lasting longer than the five-second countdown, therefore produced only one resume. The blank fired while input was still arriving, the next count relit the panel while the monitor was still active, and when the input stopped nothing re-armed the spent timer. On a worker, seven seconds of outward pointer counts at the screen corner left the panel lit for good, which is the end state this PR exists to prevent. Reproduction. This ran on a worker built from the Omarchy ISO, under a real Hyprland session, using a kernel uinput mouse parked in the bottom-right corner after the lock had blanked.
Tests on the worker. Who checked it. Claude Opus 5.5 reviewed the diff and ran the reproductions. The second opinion was Codex Medium, over two rounds. In the first round it raised the long-burst gap; I reproduced it before fixing it, and Codex is credited as co-author on that commit. It also noted that the fixture calls the handlers directly rather than going through a real idle transition. That is true, but I left it as is: the fixture runs in a developer's live session, where generating input is not safe, so the worker reproduction is the evidence for the compositor path. In the second round, on 204d960, it found nothing; its independence from my reasoning is not guaranteed. Greptile's check passed with no findings. What happens next. It is labelled |
|
Looked at this again because it no longer merges cleanly into The conflict. Since then Classification. I classified this as a bug fix. The second model, Codex Medium, asked separately from the item alone, called it an enhancement. When the two disagree the item gets no new class label and stops here, so this run did not push the merge and ran nothing on a worker. The What happens next. This is waiting on the maintainer to settle the classification. After that the merge above gets pushed, verified on a worker again and reviewed again. |
Keep upstream's removal of battery and DPMS polling state while preserving the activity monitor and both re-arm handlers. Shadow the renamed stranded-lock helper in the fixture. Drop source-text assertions in favor of runtime checks and report unavailable checks as skips. Clarify that the idle protocol reports transitions, not individual input events.
|
Refreshed this branch at e27b814, merging The conflicts were resolved as suggested: keep upstream's battery/DPMS-polling removals and retain the activity monitor and both re-arm handlers. The fixture now shadows Fresh reproduction in a disposable VM (Hyprland 0.56.2, Quickshell 0.3.2.r7.g2b4501a): a kernel uinput mouse parked at the bottom-right corner, under a real session lock, sends outward
The burst also reproduced the intermediate blank followed by a re-light while input was still arriving; the idle handler then restored blanking. After each final blank, the output stayed off. Checks: all 11 targeted test files passed in the VM, without skips: Classification: this restores the lock's existing automatic-blanking behavior after input that the compositor handles but the lock surface does not receive. It adds no user option and does not change the five-second policy. The same failure is still reproducible on current GPT-6 ( |
|
Looked at the refresh at e27b814. I read it and did not run it: the classification is still unsettled, so nothing ran on a worker this time. The merge. It resolves the conflict the way the last comment described. One finding in the test, from Greptile's review. The fixture can take a real session lock on the machine running it. It sets Classification. I classified it as a bug fix again. Codex Medium, asked separately from the item alone, called it an enhancement again. While the two disagree, the item gets no class label and stops here. What happens next. This waits on the maintainer to settle the classification. After that, this head gets verified on a worker and reviewed again, along with the fixture fix if it has landed by then. |
|
Addressed the fixture safety finding in bed2c0a. The test now starts its own headless Sway compositor with private sockets/runtime paths and an allowlisted environment; it never attaches to the caller’s session. No production code or test-disable switch was added. The fixture now deliberately exercises output addition → screensChanged → real secure protocol-lock acquisition. Disposable-VM verification confirmed that only private Sway was locked and the caller Hyprland session was unchanged. Killing the secured fixture and terminating the runner produced failures and cleaned the owned processes/runtime paths. Startup failure is not treated as a pass. Sway is a test-only dependency: environments without it explicitly skip this runtime fixture; no production/CI package change is included. The original behavior was also exercised inside the VM: one outward raw mouse count reblanked in 5.157 s; a seven-second burst reblanked about 6.37 s after input stopped, with the real session lock held and the output staying off. Verification limitation: 10/11 targeted lock files passed with no skips. lock-appear-budget failed at 2555 ms versus 2340 ms; a single unchanged-parent snapshot comparison passed at 1675 ms versus 1796 ms. Production, shared test helpers and that budget test are byte-identical. The cause remains unproven; no unrelated repair or retry loop, and no full-suite or physical OLED pass is claimed. Full details are in the updated PR description. GPT-6 Astra authored/verified the patch; a separate Astra coordinator reviewed the isolation diff and evidence and independently checked syntax, whitespace and unchanged production sources. The disposable VM and owned processes are stopped. This only updates Omarchy; no Hyprland changes/contributions were made. The maintainer classification decision remains outstanding. |
|
Read the update at bed2c0a, including the diff, discussion and the lock service's screen-change path. Compared with e27b814, the production lock service is unchanged. The test runner now starts the fixture inside private headless Sway with an allowlisted environment and private sockets, addressing the previously reported connection to the caller's compositor in the source. Runtime isolation has not been verified by this review. I classified the change as a bug fix; deepseek-v4p1-flash, asked separately to classify this head, returned enhancement. The classification remains unsettled, so no class label was added and no worker tests or verification ran this time. The author's reported lock-appear-budget failure has not been investigated in this run. GPT-6 in Codex read the code; deepseek-v4p1-flash provided classification only, not a code review. Reasoning effort was not reported for either. This waits on the maintainer to settle the classification before worker verification continues. |
Summary
The lock screen blanks the display with a one-shot five-second timer that is re-armed only by input reaching the lock surface or by a password check finishing. Hyprland lights DPMS on its own for any pointer motion or key press (
mouse_move_enables_dpms,key_press_enables_dpms, both on indefault/hypr/input.lua), and when that input produces no event on the lock surface - one stray sensor count at 03:29, in the captured case - the spent timer stays spent and the panel is on until someone touches the machine. On an OLED that is the lock screen lit for 4 h 49 min, reported as "same defect, opposite symptom" on #7749. The idle service did get the ext-idleresumedfor that motion, but it has no display-off command of its own; the only blank in the tree is the lock's timer.Caught with an evdev + DPMS witness (Hyprland 0.56.2, Omarchy 4.0.2, LG UltraGear OLED on DP-1):
This listens to ext-idle-notify while a lock is requested, the protocol the idle service already uses, and re-arms the blank on every resume unless a password check is in flight. The idle service's own monitor cannot serve here: it respects inhibitors and its timeout is the user's screensaver setting.
respectInhibitors: falseis load-bearing. A monitor that respects inhibitors is paused while one is held, so it never reaches idle and never sees the next input as a transition, and an inhibitor appearing counts as a resume in its own right. Ignoring them costs nothing: the lock blanks five seconds after locking no matter what inhibitors are held, so this monitor darkens nothing the lock was not already darkening.blank-rearmed: activity) only when the timer was actually spent; input the surface already saw and re-armed on earns no line.blank-startedwhen it spawns, andomarchy-shell lock statusgainsblankArmedandactivityIdle. Nothing in the shell logged display state at the lock screen before; the line Re-arm lock blank timer with backoff on screen changes #8581 proposes is the same idea.In #9152 I argued that re-blanking is the wrong fix for a wake the shell itself causes: the flash still happens, so the wake should not fire. That still holds, and the two are complementary rather than the same fix twice. #9152 is for a wake the idle service causes and can skip; this is for lights the shell did not cause and cannot prevent.
Scope, stated up front: this covers the raw-input trigger of the never-blanks state. It never calls
runWake()and never shortens anything, so a panel that is slow today is no slower; the blanks it adds happen only in the state where the panel would otherwise stay lit indefinitely. The five-second budget itself (#7749) and the output re-add after an HPD drop (#8580) are separate defects and untouched, and a device that jitters more often than every five seconds keeps the panel lit, same as today, since Hyprland relights on every count regardless.armBlankTimer,runWake,runBlankand the interval literal are not modified, so the open PRs that change them (#7643, #7673, #7386, #8581, #8561, #12245) still apply; #7643, #7673 and #12245 touch adjacent lines and will need a one-hunk resolution whichever lands second. #12245 makesarmBlankTimertake an interval and gives deliberate wakes 30 s; the re-arm here passes no interval, so it keeps the 5 s default and the two compose without either changing the other's behaviour.The video-wallpaper polling introduced by #6792 has since been removed from the lock service in
quattro. This PR retains that removal; its event-driven activity monitor remains necessary for raw input that wakes the display without reaching the lock surface.No separate issue: the report is the comment on #7749 linked above, and #8580 already describes the same end state from a different trigger.
Refs #7749, #8580.
Test
test/shell.d/lock-activity-reblank-test.shruns the real lock service only inside a disposable headless Sway compositor with software rendering, private Wayland/IPC/runtime paths and an allowlisted environment. It does not connect to the caller's desktop. The brightness/wake/stranded-lock helpers remain shadowed.The fixture checks the activity monitor, re-arm handlers, password-authentication gate, real timer/blank-command execution and unlocked no-op behavior. It then adds an output and verifies that
screensChangedcan acquire a real protocol lock—even when onlylockRequestedwas set. That lock is confined to the disposable compositor. Process-group teardown and independent deadlines bound the fixture and compositor.Test dependency: Sway is test-only, not a production dependency. Missing Sway or Quickshell produces an explicit skip; present-but-broken compositor startup fails closed. Hosts and CI images without Sway therefore skip this runtime fixture. Never run its QML directly on a live desktop. Sway coverage does not replace the Hyprland raw-input/DPMS reproduction.
Safety update verified 4 November 2026
At bed2c0a, the fixture safety finding is addressed without any production-code changes from e27b814.
All runtime verification ran through
herd-testinside a disposable 2-vCPU Omarchy VM, using Hyprland 0.56.2, Sway 1.12 and the image's installed Quickshell:lock-appear-budgetfailed: 2555 ms against a 2340 ms budget. One comparison with the unchanged parent snapshot passed: 1675 ms against 1796 ms. Production sources, the budget test and its shared helpers are byte-identical; the differing result does not establish a cause. No unrelated repair or retry-until-green was attempted.No full CLI/shell-suite or physical OLED retest is claimed. The current-upstream A/B run was not repeated; the November 1 evidence below remains separately dated. The owned VM and all test processes were stopped. No owner desktop was touched and no Hyprland source or upstream contributions were made.
The timing failure remains visible for maintainer assessment. This update does not resolve the separate bug/enhancement classification disagreement.
Refresh verified 1 November 2026
Merged
quattroat9e0b5e2f2into the existing PR history; refreshed head:e27b814b9. The conflict resolution preserves upstream's removals and this PR's activity handlers. The fixture now shadowsomarchy-hyprland-session-lock-stranded.In a disposable Omarchy VM running Hyprland 0.56.2 and Quickshell 0.3.2.r7.g2b4501a, a kernel uinput mouse was parked at the bottom-right screen edge under a real session lock. A single outward
REL_Y +1and a seven-second burst both woke the output without moving the cursor. With the lock service from currentquattro, it remained lit at least 20 seconds after input stopped. With this head, the single count was followed by a blank about five seconds later; the burst was followed byblank-rearmed: input-stoppedand a blank about six seconds after its final count. The output stayed off afterward. The idle service was disabled in both runs to isolate the lock's behavior.Eleven targeted test files passed in that VM without skips:
lock-activity-reblank,lock-appear-budget,lock-autorepeat,lock-fingerprint-indicator,lock-fingerprint-retry,lock-fingerprint-service,lock-poster,lock-state,system-lock,hyprland-session-lock, andsleep-lock. The activity fixture also failed with the lock service reverted to currentquattro(its activity-monitor API is absent there); the real-input A/B reproduction above is the evidence for the user-visible failure and recovery. The full CLI/shell suites were not rerun for this localized refresh. Physical OLED hardware was not retested.AI provenance
GPT-6 Astra authored and verified the November merge/fixture refresh and the November 4 test-isolation repair. A separate GPT-6 Astra coordinator reviewed the isolation diff and evidence, checked syntax/whitespace and confirmed production code was unchanged before publishing it. No new different-model review was performed for this safety update. Earlier independent reviews apply to their preceding heads, not automatically to this patch. Maintainer classification and review remain outstanding.