Skip to content

Sync upstream quattro through 4fa93e9f - #93

Merged
mark-groves merged 34 commits into
quattrofrom
cursor/upstream-changes-pr-4fa9
Oct 25, 2026
Merged

mark-groves merged 34 commits into
quattrofrom
cursor/upstream-changes-pr-4fa9

Conversation

@cursor

@cursor cursor Bot commented Oct 24, 2026

Copy link
Copy Markdown

Upstream quattro moved since #92 (cbd44ba1). This brings it through 4fa93e9f (2026-10-24): 27 first-parent commits, 30 non-merge commits cherry-picked. The two merge commits were not replayed; their changes arrived with the commits they merged.

What changed

Notifications. Toasts follow Emil Kowalski's motion: they animate in and out, unnamed senders each get their own toast, and omarchy-notification-send can add action buttons. The keyboard can drive a notification. The saved screen recording toast has Open and Omacut buttons. The notification panel, snooze, and the notifications settings object stay as they were after #92.

Panels, bar, and dock. Panels grow out of the bar, close faster, morph between each other, and grow again when content arrives after they open. The network panel stays open after a DNS choice. Buttons and bar icons give a little under the pointer. Bar colors fade in 250ms on a theme change. A tooltip that is already up is replaced at once. The dock eases, and it stops when animations are off. Every shell animation duration goes through Style.duration, and the shell tests fail a bare number. Dragging a bar widget off the bar removes it with a poof, and the bar surface stays full-size so that drag survives an empty workspace. The indicator peek opens only to the left of the clock.

Lock screen. A video wallpaper shows its cached still frame, blurred like any other wallpaper. The lock no longer polls Hyprland for each panel's DPMS. Face still pauses while the lock itself has blanked the displays.

Fingerprint. Removing fingerprints deletes the reader's own copy with fprintd-delete before the local files go. A platform can turn its reader on or off with the owner's agreement (enable-fingerprint / disable-fingerprint). Setup and removal still go through omarchy-pam-pair-add and omarchy-apply-polkit-pam.

Git identity and setup. First login asks for a name and email afterwards, with GitHub sign-in, rather than during install. Setup > Config > Git runs it again. The setup form is translated from Omarchy's catalogs, and a caller can pass a timezone fallback.

Agents. The agents panel and the default agent cover the rest of the coding agents and providers: Qwen, DeepSeek Harness, MiniMax, Kiro, Freebuff, Kilo, Devin, Amp, Command Code, and the providers behind pi, omp, and OpenCode (Copilot, Muse, Antigravity, Kimchi, Hermes, Ollama, OpenRouter, DeepSeek, Fireworks, and the rest). Collectors share one module. Grok Bot's weekly allowance is its own collector, read from the same Cursor sign-in as Cursor's plan. Cursor itself stays on the official Linux tarball. Face, Grok Bot, and Cursor stay in the menu.

Other. Ethereal gains the Roadster Starlight background. Hermes is seeded where ARM rebuilds it. Migration tests are gone; a migration runs once and is frozen once it ships. The shell's motion rules are written down for Omarchy and plugin work.

Fork overlay

  • Cursor's installer is omarchy-install-cursor-agent. cursor-agent is not in etc/mise/conf.d/omarchy-tools.toml (20 auto_update, 30 lazy).
  • Face lock chrome stays, including the blank pause. Its fades use Style.duration.
  • shell/Ui/qmldir keeps FaceChromeCanvas immediately before FilterInput.
  • The power-profile menu provider stays out.
  • The notification panel stays out.

The fork's combined Cursor usage record (one X Subscription pool, plus Top-up Credit) is replaced by upstream's Cursor collector and the separate Grok Bot collector. The scanner test that drove the old collector is removed; agent-usage-cursor-test.sh and agent-usage-grokbot-test.sh cover the new ones.

Tests

Passed: motion-test.sh, mise-provision-test.sh, fingerprint-package-test.sh, security-fingerprint-remove-test.sh, lock-poster-test.sh, default-agent-test.sh, agents-panel-test.sh, agent-usage-cursor-test.sh, agent-usage-grokbot-test.sh, menu-test.sh, i18n-test.sh.

Open in Web View Automation 

bjarneo and others added 30 commits October 24, 2026 23:05
…omacom#16403)

* Delete the reader's own copy of the prints when removing fingerprints

A match-on-chip reader keeps the enrolled prints in the device as well as
in /var/lib/fprint, and fprintd finds the device's copies only through its
local records. So removal now deletes them with fprintd-delete while fprintd
still runs, before the local files go. It authenticates sudo in view first,
since the delete's output is captured; a refusal changes nothing. A delete
that finds no reader is said plainly, and one that fails is reported without
stopping the removal.

The enrolment instructions also move to short lines, since the setup
terminal wraps a long one mid-word.

Split out of omacom#16360, which builds the Touch ID hooks on it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Drop the width check that measured the test's own text

The loop read the expected instruction the test passes in, not what setup printed, so it could only fail if someone edited the test's string; the exact-output match above already holds setup to that text.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Codex Medium <noreply@openai.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Codex Medium <noreply@openai.com>
Migrations run once per machine and are frozen once they ship, so their tests
stop guarding anything. Drop the migration test files and the migration blocks
inside other tests, and update the testing docs to say migrations get no suite
test.

Co-authored-by: Mark Groves <mark-groves@users.noreply.github.com>
Only an armed drag poofs, so the poof's armed flag was always true. Its nine
root properties collapse into one barPoof object, seeded at random instead of
by a serial. A drag commits any pending removal when it starts, so
removeBarDrag doesn't need to. Moves and removals share writeLayout.

Drop the source-grep assertions; the removeEntry and dragDistanceOutside
tests stay.
…ent (omacom#16360)

A Mac keeps Touch ID off until its owner agrees, so fingerprint setup and
removal gain two optional lifecycle operations, enable-fingerprint and
disable-fingerprint, which omarchy-mac implements on Apple Silicon:

- Setup resolves enable-fingerprint as the user once sudo is
  authenticated and, when it resolves, runs it with sudo before fprintd is
  asked about the reader. The hook asks the owner; a refusal or a failure
  stops setup with nothing enrolled. sudo is checked again afterwards,
  since the owner may read past its timeout.
- A platform that can't be told sets up as one without the hook, with a
  warning: the reader its platform keeps off stays off, so fprintd doesn't
  find it, and setup never stops anywhere else.
- Removal runs disable-fingerprint once the prints and local files are
  gone and before the packages go. A failure is reported, not fatal.

Builds on the removal change that deletes the reader's own prints first.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…macom#16469)

* Stop holding the notification surface mapped after the last card leaves

The 800ms linger kept the canvas up for an exit animation that no longer
exists, so a fullscreen window lost direct scanout for nothing each time
the last card closed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Skip notification card fades when animations are off

The card's crossfades used a bare 150ms, so they kept playing after
`omarchy toggle animations` had zeroed every other shell animation.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Open and close the notification deck on a crisper, unbouncing curve

The deck moved for 320ms on a curve that overshot its rest. It now opens
in 250ms and closes in 200ms on a strong ease-out, and the card's
crossfades ease instead of running linear: out as the deck's text
appears, in-out for the hover swaps.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Drop notifications in from under the bar and lift them out again

An arriving card falls the last 24px into place as it fades in (250ms),
and a clicked or expired one rises the same way as it fades out (200ms)
while the cards around it close up, all on the deck's strong ease-out.
Keyboard dismissals stay instant, and a leaving card takes no pointer.
Text giving up the front fades in half the time, so an arrival never
shows both cards' words at once.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Lift a leaving notification out from behind the stack, not over it

A departing card was drawn on top, so in a collapsed deck its fading
plate and words lay over the card stepping forward. It now slips under
the deck, its text going first so it doesn't read through; the last
card keeps its words, with nothing to cover it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Let departing notifications go on their own deadline, past the shortcuts

A leaving card was removed only when the scene clock finished, and
arrivals and updates restart that clock, so a card updating every tick
kept closed ones on screen and counted against the limit indefinitely.
Each departure now ends on a deadline the scene cannot push back.

The activation shortcut also picked the first row, which during an exit
is the card already leaving, and could run its command a second time.
It now takes the newest card still live, as dismissal already did.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Bring a notification back when its sender replaces it mid-exit

The update landed on the card, but the card stayed marked as leaving,
so its deadline then removed the fresh notification and reported it
expired. A replacement now cancels the departure.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Let omarchy-notification-send add action buttons

`--action <label> <program> [words...]` adds a button, up to three, and
`--with <args...>` gives them the data they act on. An action's words are
call-site literals that end at the next flag; the untrusted data goes once,
after --with, which takes the rest of the line like --exec, so a value that
reads like a flag stays data. The first action is also the card click.

The shell runs the buttons itself from the omarchy-actions hint and stores
them with the row, so the sender exits at once and restored or replayed
cards keep their buttons.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Draw notification buttons as a bar along the card's bottom edge

Bordered buttons inside the bordered card read as boxes in boxes. The
buttons now sit in a bar flush with the card's edges: a hairline above,
borderless segments split by hairlines, three to a line, with a short last
line stretching across. A segment fills with the theme's hover colour.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Mirror the notification button bar and unwrap its docs

The bar placed its segments by explicit x, which LayoutMirroring does not
flip, so the first action stayed on the left in right-to-left languages.
The new docs paragraphs are joined onto full lines, as AGENTS.md asks.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Control notifications with the keyboard

Super+Ctrl+comma selects the newest card, opens its deck and gives that
display's notification surface the keyboard: number keys press its buttons,
Enter clicks it, arrows move between cards, Delete dismisses, and Escape,
any choice, a mouse click, the pointer moving to another deck, or ten idle
seconds hand the keyboard back. Notifications never take focus on their
own. Silencing moves to Super+Ctrl+Alt+comma.

The card and the keys share one button list, Delivery.deeds, so the numbers
always match. invokeAction <n> presses a button on the newest card. The
button bar's segments draw their own hairlines, mirrored in right-to-left
languages, so the selection tint covers the whole card, and a label is only
cut short when it cannot fit.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep notification keyboard selection on screen and release it on links

The deck does not scroll, so moving the selection down could select a card
below the bottom of the screen and act on it unseen. The keyboard now stops
at the last card that would fit with its deck open, and a dismissal falls
back to the nearest reachable card. Clicking a link in a card's text now
ends keyboard mode too, so the browser it opens gets the keyboard.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Let notification keyboard selection pass long decks and keep its buttons on screen

Down stopped at the first card of a long deck that did not fit, so the decks
below it, which fit once it collapses, were out of reach. It now skips to the
next card that fits. Selecting a card shows its buttons and can open its
body, so it grows after the check passed; once its height arrives, a card
that no longer fits gives way to the nearest one above.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Skip a notification card that gives way under keyboard selection

A card that fits until selection shows its buttons gave way to the card
above, hid its buttons, and was picked again by the next Down, so later
decks were never reached. It is now remembered for the rest of the keyboard
session and skipped.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…omacom#16470)

* Ease the panels' rotating status line out, never in

The swap faded the old phrase out over 180ms and the next in over 260ms
on an ease-in, which starts slowest just when the eye is on it. Both
halves now ease out, out in 150ms and in over 200ms.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Open and close panels by keyboard without motion

A keybinding or Escape is repeated too often to wait on an animation.
Style.byKey holds a keyed flag for the length of the keypress's answer,
during which Style.duration() is 0, the same as with animations off.
The shell's summon, hide and toggle IPC that keybindings call and the
panel key catcher's Escape run inside it.

The key catcher test loads PanelKeyCatcher under qmltestrunner, which
can't load Quickshell, so it gets a one-function qs.Commons stub.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Grow panels out of the bar icon, and let them go faster than they came

A panel only faded, the same 140ms each way. It now grows from just
under full size, out of the point on its bar-facing edge nearest the
icon it hangs from, as it fades in over 180ms on a strong ease-out, and
closes in 120ms. Switching between panels still holds the card still.
The tray and media popups get the same motion, and the curve becomes
Style.easeOut, which the notifications now share.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Animate panels the same whether a hotkey or the pointer opens them

Hotkeys jumped panels open and shut. They now grow out of the bar and
fade like a click does, so the panel looks the same however it's
summoned. This takes back the keyed flag and Style.byKey, and the key
catcher test's qs.Commons stub with them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Morph one card between panels instead of swapping two

Switching from one panel to another hid the old card at once and put
the new one up in its own place, with a frame or two of nothing while
it mapped. Each panel now tells the bar where its card stands, and the
panel taking over starts its card there, at that size, and moves it into
its own place over 220ms on the strong ease-out, its content fading in
along the way while the old card fades out under it. Third-party panels
reach the card's place through the bar facade as a detached copy.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Morph between panels only within one section of the bar

Flying the card from the center to the right reads as travel, not as
one panel giving way to its neighbour. The bar now records which section
the open card's widget sits in, the center's pieces counted as one as
keyboard switching counts them, and hands the card over only to a panel
in the same section. Across sections the old card fades and the new one
grows out of its own icon as usual. Plugins ask for the card through
popoutCardFor(owner) rather than reading it directly.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Grow panels to fit content that arrives after they open

A Wi-Fi scan or Bluetooth discovery fills its list a beat after the
panel is up, and the card snapped taller when it did. The card now
follows the content's height over 200ms on the strong ease-out,
cropping the new rows as it grows, and is placed by that drawn height so
a bottom or side bar's panel grows away from the bar too. For the first
60ms after opening it snaps instead, while freshly built content
measures itself, so a panel never grows in from a stale height.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Hand a growing panel's drawn height to the panel taking over

The bar was told the card's final height, so switching to a neighbour
while the card was still growing started the next card taller than
what was on screen, and it jumped before morphing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Picking DHCP, Cloudflare or Google closed the panel, so the change
couldn't be seen taking hold or another provider tried without opening
it again. The panel now stays open, and the pill moves to the choice
the moment it's made, falling back if the switch fails, the way the
band pills do. Custom still closes it, since it hands off to a terminal.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…aying it (omacom#15289)

* Show a video wallpaper's still frame on the lock screen instead of playing it

The lock played video wallpapers through OWE's lock feed and, to pause it
while the displays were off, asked Hyprland for every panel's DPMS state every
3 seconds for as long as the session stayed locked. The lock now shows the
video's cached poster, blurred like any still wallpaper, so nothing plays and
nothing polls. The feed module leaves the base packages and the battery
service drops the power-saver flag only the feed read.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep an offscreen check that the lock shows a video's poster and releases it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Drop the lock source checks the poster test already covers

The rewritten asserts matched LockView's binding text and the absence of names this branch deletes, which AGENTS.md rules out as restating the implementation. lock-poster-test.sh proves poster selection and release through the real bindings, and the image effect wraps the one wallpaper unconditionally, so there is no behaviour left for a source check to guard. The summary lines that named the removed DPMS tracking and the lock feed module go with them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Let buttons give a little under the pointer

A button only changed its fill when pressed. Held down it now also
scales to 0.97 over 160ms on the strong ease-out and springs back the
same way, so a click is felt as well as seen. It follows Style.duration,
so it stays still with animations off, and keyboard activation doesn't
press it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep a pressed button's hit area at full size

Scaling the button shrank its MouseArea with it, so a press near an
edge, held until the button gave, and let go there no longer counted
as a click. The MouseArea now undoes the press scale, and only what's
drawn shrinks.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…om#16486)

A new agents/skills/motion.md, linked from AGENTS.md, sets out when
something animates, the Style.duration and Style.easeOut tokens, the
durations, the patterns the shell already has (panels growing from
their icon and morphing within a bar section, late content growing the
card, button press, the notification scene clock) and the pitfalls hit
while building them. Adapted from Emil Kowalski's animation skills to
Qt Quick and to the choices Omarchy has made.

visual-verification.md gains how to check motion a frame at a time:
running a branch's shell, driving it without the mouse, warming panels,
tiling the recording with ffmpeg, and the environment a recorder needs
outside the session. The shipped plugins.md gets a standalone Motion
section for plugin authors, and docs/omarchy-shell.md lists the tokens.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…udget (omacom#16489)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…#16490)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…om#16288)

* Ask for the git identity after first login, with GitHub sign-in

The installer and the first-boot owner form asked for a full name and email before the machine was even set up, for nothing but git. First run now shows a Setup Git notification instead, which opens omarchy-setup-git in the same floating terminal as Update System: it sets user.name and user.email, fills the Compose shortcuts for them, and offers gh auth login. The notification is skipped when an unattended install or an upgrade already set the identity, and Setup > Git in the menu runs it again later.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep the identity prompt for older ISOs, and write XCompose through symlinks

An ISO source from before this change still calls omarchy_prompt_identity from the setup form it vendors from this package, and would abort the install without it. omarchy-setup-git now edits the file ~/.XCompose points at, so a dotfiles-managed symlink stays one.

Co-Authored-By: Codex Medium <noreply@openai.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Go straight to the GitHub sign-in question

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Drop the skip hint from the git identity prompts

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Move Git setup under Setup > Config

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Shorten the Setup Git notification

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Ask before gh installs itself, and keep XCompose permissions

Checking `gh auth status` ran the lazy mise shim, which downloaded and installed gh silently before the sign-in question appeared, and for people who then declined. Reading gh's hosts.yml answers whether someone is signed in without running gh, so gh now installs only after a yes, with its output visible.

Renaming a rewritten ~/.XCompose into place reset its mode to the umask default; copying back over the original keeps it, as Greptile pointed out.

Co-Authored-By: Codex Medium <noreply@openai.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Take the git identity as options, and drop it from installs

omarchy-setup-git --name and --email set the identity without asking anything, so an unattended machine gets it after install, over SSH or from a script, rather than through installer inputs. With options it skips the GitHub question, which needs a person.

Installs no longer carry a git identity at all, so install/user/git.sh, which applied the ISO's OMARCHY_USER_NAME and OMARCHY_USER_EMAIL, goes. The Quattro upgrade sets git config itself. The unattended-install manual drops user_full_name.txt and user_email_address.txt and points at the command instead.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Set only the given fields, replace XCompose atomically, recheck GitHub sign-in

--name or --email alone left the other field's Compose shortcut rewritten from git config, replacing a hand edit nobody asked to change; options now touch only what they name.

Copying the new ~/.XCompose over the old one truncated it first, so an interrupted write could leave it empty. It is renamed into place again, with the original's mode copied onto it beforehand.

A github.com entry in gh's hosts.yml was taken as a working sign-in, so a revoked token skipped the question and Setup Git could not repair it. gh wrote that file, so it is installed and can check the token without the lazy shim installing anything first.

All three found by Greptile.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: David Heinemeier Hansson <david@hey.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Codex Medium <noreply@openai.com>
…macom#16492)

Seven animations had drifted to bare numbers before omacom#16490 brought
them back, one of which kept fading after animations were turned off.
motion-test.sh now fails on any `duration:` given a bare number in the
shell's QML. The last three, Button's spinner and the dock's launch
pulse, go through Style.duration too; both loops were already held
still with animations off, and the motion guide says so for loops.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…com#16499)

* Let a notification show a wide picture preview under its text

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Preview the saved screen recording with Open and Omacut buttons

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Show the recording's frame in the toast's thumbnail, not a wide preview

Drops --preview again: the frame sits in the normal image slot.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Show a notification's buttons on the front card of its deck

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Show buttons only on the focused card in an open deck

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say how long and how big the saved screen recording is

The buttons open it now, so the toast no longer explains how to.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Drop the click mark from notification titles

The current card shows its buttons, which say it can be acted on.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Write the recording's length and size in the user's language

Pass each number through omarchy-i18n and keep the units in the catalog.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Name each recording size key literally so the catalog check finds it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…lt agent (omacom#14040)

* Read Grok's ledger and settings of the wrong shape as empty

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Pause Grok's limits on a lapsed access token, as Claude does

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read a Grok settings cache of the wrong shape as no plan in account state

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read Copilot's session store in place, read-only, by its file URI

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Ask gh for its github.com token, never GH_HOST's

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read a body short of its Content-Length as a reply cut short

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Never open a fork's parentSession that isn't a regular file

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Hold a registry's switch threshold to 50-100, as account state does

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read Claude's stats-cache and history lines of the wrong shape as nothing

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Send the Cursor sign-in to CURSOR_API_ENDPOINT only over HTTPS or loopback

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Sign only the primary Codex home in with CODEX_ACCESS_TOKEN

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep and reuse a Codex answer without windows like any other

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Reuse a Kimi answer without windows like any other

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Send the Muse token to MUSE_KEY_ENDPOINT only over HTTPS or loopback, sharing Cursor's check

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read zoneless times at the calendar's edge as undated in Muse and Vibe

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Start Muse's prompt counts over after a gap of days, not on today

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep Kimchi's out-of-credits signal in the help, not a status outside the contract

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Match Kimchi's region whatever its case

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Date Antigravity's rows only within 2000-2100, as the shared reader does

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Run omp usage without stdin, read odd window ids, and keep an anonymous answer for nobody

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Mark Antigravity's limits stale only when some were kept

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count a Hermes state.db without a usage table by its sessions and prompts

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count only Hermes times within today as today's

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Leave negative, text and huge Hermes counts out of its sums

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Floor Devin's fetch stamp rather than round it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Date an undated Qwen record once, by its transcript, not as today's every day

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Describe the collectors' reviewed behaviour in the agents README and contract

Grok's paused limits, the https-only endpoint overrides, OpenCode's and
DeepSeek's pi and omp keys, Z.ai's hourly no-plan recheck, MiniMax's own-id
keys and per-source host, and the shared trusted_endpoint() helper.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Hold every collector to the cut-short answer in the contract checks

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Run collectors in the caller's TZ and drop mise's variables in the test sandbox

run_collector and the partition test's runner start collectors under env -i,
so they dated records by /etc/localtime whatever TZ the test ran in. The
cursor and fireworks tests now restore the zone they borrowed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Floor the time Qwen dates an undated record by

local_day rounds a float of seconds, so a transcript written in the last
half second of a day dated its undated records as the next day's.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Hold "today" to one day in tests that compare today's totals

A run straddling local midnight failed the Qwen plan count, Ollama's scan
reuse and the shared scan-lock check. freeze_today() stops the shared
module's clock, and the rows the fixtures stamp now, at the end of the day
the run began; check_scan_cache freezes it for every collector.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Skip an OpenCode credential nested too deeply to parse

json.loads raises RecursionError, not ValueError, on a deeply nested value,
so one such auth.json or credential row crashed the collector.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep a failed Codex recheck quiet after a plan without windows

A saved answer with no windows has no limits to lose, so a later failed
probe said "Codex limits unavailable" in the urgent colour for a plan
the contract says is no trouble, as Z.ai's no-plan handling already does.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say OpenRouter counts reasoning tokens as output, and assert the split

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Hold Antigravity to opencode's google-vertex rows in the partition test

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Pin Kimi's look-alike domain refusal and its refusal help text

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Cover Z.ai's Zhipu host, raw-key header, per-platform cache and key sources

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Cover Ollama's saved-key fallback, its Bearer header and its full-window cap

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Hold MiniMax to a limits cache per region, and compare kept limits by value

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Cover OpenRouter's opencode key, a refused management key, and kept meters by value

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Cover DeepSeek's opencode key and its balance cache per API host

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Cover Copilot's exported-token order and the remote: true override

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Assert a rate-limited Claude check advises no sooner retry

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Cover OpenCode keeping a later refusal's windows and a no-plan key's quiet

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Prove Claude's index shortcut never opens an unchanged transcript

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Age Muse's own scan schema, not the envelope's, in the dedup rescan check

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Make Devin's newest-fetch check discriminate, and cover hideWeekly and a future fetch

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Make Antigravity's scan reuse and step-model checks discriminate

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Cover a Hermes profile linked to its home counting once

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Cover Fireworks account discovery, its reset on a moved key, the billing summary, and a cut-short answer

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Compare Kimchi's kept limits by value and date its sprint budget from now

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Date Antigravity's old transcript step in the zone the test runs in

run_collector now hands the collector the caller's TZ, so a step at noon
UTC is the next day's in Kiritimati.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Floor a float epoch in the shared local_day instead of rounding it

number() rounds, so a time in the last half second of a day read as the next
day's; it now only screens out untrusted values, and the time is floored.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Start the partition test on the new day when it would cross local midnight

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Describe the agent test harness's TZ, cut-short step, and frozen day in the testing docs

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep Copilot's last allowance when GitHub's quota answer lacks an amount

A missing or unreadable entitlement or credits_used read as 0, which dropped
the meter or showed an untouched allowance and was saved over the last good
answer. It is now a failed check; an explicitly unlimited allowance still
needs no amounts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Don't cache Copilot's transcript totals when a transcript couldn't be read

An unreadable events.jsonl was skipped without marking the scan incomplete,
so its short totals were cached for every run after. The transcript scan now
says whether it read everything, and the cache decision includes it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep Cursor's known limits when a usage section comes back the wrong type

A planUsage or spendLimitUsage that wasn't an object read as an empty one,
giving no meters that were saved over the known limits. It is now a failed
check; a section left out or null is still an empty plan.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Show Kimchi's out-of-credits warning as an Out of credits status

Moving it into the help with an empty status hid it: the panel only shows the
help for a status, and a balance with no funded figure raises no low-balance
alarm of its own. Out of credits joins the documented statuses.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Translate the agents panel's Out of credits status

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Use the official agent marks in the Default Agent menu

Antigravity, Claude, Crush, Kimi and Muse Code drew Nerd Font stand-ins;
they now use their own marks from the icon font. Amp's old chevrons are
replaced with its current wordmark tile, and Codex uses the Codex app's
mark instead of the plain OpenAI blossom.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Use the official Antigravity, Fireworks and OpenRouter marks in the agents panel

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Harden the shared collector helpers against hostile replies and stamps

A reply nested too deep, a key with a line break, or a server dripping
bytes no longer crashes or holds a collector; an overflowed number no
longer makes the record invalid JSON; a limits cache stamped before the
clock went back gets replaced; and local_day ignores digit strings that
can't be a time.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read Grok session days through local_day so an out-of-range time skips its row

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Skip a Claude transcript line that isn't an object rather than the whole file

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count a Claude transcript line with a falsy time nowhere, not as today

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Give Claude's stats-cache fallback the record's recentDays and modelUsage shapes

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Key the single Claude home's limits cache by the account it is signed in as

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Skip a wrong-shaped Codex rollout line instead of the rest of the file

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep Codex limits live when reset credits or account/read are malformed

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Treat a Codex config.toml that isn't UTF-8 as naming no local endpoints

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Skip untrusted Copilot times and costs instead of failing the record

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Treat Antigravity answers of another shape as failed checks

A quota or omp answer with a non-list where a list belongs no longer
crashes the collector or replaces the kept limits with none, and a bucket
without a remaining fraction is skipped rather than read as untouched.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Send Kimchi's key only to a trusted gateway

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep Kimi's last limits when a reply isn't a quota answer

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Leave Kimi's reset time empty when it won't parse

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Take over a Kimi CLI lock lease nobody renewed

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Name no Hermes active day for start times outside 2000-2100

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say Devin is unavailable when its session database can't be read

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Skip a Z.ai limit row tagged with a list and a .env that isn't UTF-8

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep OpenCode's reset times ISO and its retry and stale flags honest

A reset past the calendar's end no longer crashes the collector, a
non-ISO resetsAt falls back to resetInSec, an answer that breaks the
reader is no longer read as unreachable, and with nothing kept no
limits are called stale.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Skip a Qwen line nested too deep and a runtimeOutputDir holding a NUL

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Send the Synthetic key only to a trusted SYNTHETIC_API_BASE_URL

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep DeepSeek's key on trusted endpoints, reject overflowing balances and skip caching partial dsh scans

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep MiniMax's key on trusted endpoints and skip caching partial MiniMax Code scans

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep Fireworks' last answers past odd billing replies and its key on trusted endpoints

Only the billing-role gate now reads as no live balance, an amount past a
float or a misshapen list is a failed check, an edge-of-calendar bucket has
no day, an overdrawn balance reads as nothing left, and an auth.ini that
isn't UTF-8 reads as no key.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Run a user's own agent install ahead of mise's lazy shim for every agent

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count a Codex rollout line with a falsy time nowhere, not on the file's day

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Prefer mise's installed agent over a leftover ~/.local/bin copy

A ~/.local/bin install still wins over mise's lazy shim, but once mise has
the tool installed the launcher runs the same copy a terminal does. Older
one-line mise wrappers are recognized as wrappers too, and an OpenClaw
prompt that happens to be its name is never rewritten.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Don't call limits stale when a failed check kept none

Grok, Claude and per-account entries said limitsStale whenever a check
failed, even with no limits to show; the other collectors already only
mark kept limits stale.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Treat a reply that never finishes as cut short, not unreachable

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Reread Codex's cached per-file totals after the falsy-time change

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Install node alongside an npm-backed default agent

Node is itself a lazy mise tool, and mise refuses to install an npm tool on
a machine without it, so choosing Qwen Code, Amp, Kilo or another npm agent
on a fresh install failed with "Could not install".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read an Antigravity bucket with no fraction as fully spent

Google's proto3 JSON leaves a zero remainingFraction out, so skipping such
a bucket hid the meter exactly when its window ran out.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count a Copilot line with an untrusted time nowhere, not on the file's day

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep reading a zoneless Grok session time as UTC

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say what Codex's file cache schema 3 retires

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read Copilot's used allowance from what GitHub says remains

GitHub's quota snapshot carries remaining, quota_remaining and
percent_remaining but no credits_used, so every limited account lost its
meter. credits_used is still read first where one is sent.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Title Cursor's included meters Weekly on a weekly plan

GetPlanInfo's included_usage_period and included_usage_resets_at say when
a plan's included allowance resets each week; the on-demand budget stays
on the billing cycle.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read a unified-billing Grok account's monthly quota instead of 0%

xAI leaves creditUsagePercent out of the credits answer for unified-billing
accounts and gives their monthly quota on the plain billing answer, so ask
that. A period already over without a percentage is no answer either.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read an Antigravity bucket's remaining amount and skip disabled buckets

agy's QuotaSummaryBucket holds its remainder as a oneof of a fraction and
an amount, so a bucket on the amount arm has no fraction without being
spent. It now reads as spent only at a zero amount, and a disabled bucket
is skipped. A bucket with neither keeps reading as spent.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Ask Antigravity's quota summary for the project loadCodeAssist names

agy's RetrieveUserQuotaSummaryRequest carries a project, and omp sends the
cloudaicompanionProject from loadCodeAssist there. An answer naming none
still leaves the body empty.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read a Muse window reset given as an ISO time

omp's schema for Muse's subscription windows takes resets_at as a string
or a number and reads a string as ISO; the collector read only numbers.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Name Devin's Pro Ultimate and pooled Enterprise tiers

TeamsTier also numbers PRO_ULTIMATE (8) and ENTERPRISE_SAAS_POOLED (11).
DevinPlanInfo's field 2 is the canUseCli bool, never a plan name, so stop
reading it as one.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Mistral Vibe sessions from the Unified Harness store

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read pi sessions from the sessionDir its settings.json names

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say that Claude's stats-cache days are UTC dates

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Resolve Hermes' home and profiles root as Hermes does

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Never send Cursor's API key as a bearer

The Cursor CLI trades CURSOR_API_KEY for an access token and stores it
beside the key; the dashboard takes that token, not the key. Use the
traded token only while it belongs to the exported key, honour
CURSOR_AUTH_TOKEN as a bearer, and otherwise ask as nobody rather than as
another account.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Stop adding pi's and omp's reasoning to output, which already holds it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Leave out the copies an opencode fork makes in its original message store

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count the tokens pi spends on compactions, summaries, and cache warming

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count typed prompts in pi, omp, and opencode, not each model step

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say that pi's and omp's reasoning is already part of output

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read Claude's weekly limit from seven_day, as Claude Code does

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read Anthropic's utilization as a percentage always, as Claude Code does

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count the prompts a person typed in Claude Code, not each API request

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep a forked Claude session's parent in the session count

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Key Claude replies by message and request id

A gateway can reuse one message id for every response, so the message id
alone collapsed distinct responses into one. A /btw sidechain's replay of
its parent's message, which carries a new request id, still counts once.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Claude's advisor iterations under the advisor's model

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Leave Claude Code turns another provider served out of the Claude record

Claude Code fronts whatever ANTHROPIC_BASE_URL names, and Z.ai, Ollama
and others document pointing it at their APIs. A reply by a non-Claude
model, or any reply while a known provider's host is configured, is that
provider's, so it no longer counts as Claude usage.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count the prompts typed in Codex rollouts, not the requests answering them

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Name the Codex plan as Codex's own /status does

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Never cache a Codex scan that couldn't read a rollout

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count a local Codex turn with an untrusted time nowhere

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count the prompts typed in local Codex rollouts, not the requests

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Remember Codex provider endpoints so local history survives a config change

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Hermes' default profile while a named profile is active

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count a Qwen Code Responses API reply's reasoning once

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Find the legacy Kimi CLI's sign-in under KIMI_SHARE_DIR

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Kimi Code's local usage from its session wire logs

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count only MiniMax Code replies on MiniMax's own routes

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read MiniMax's window status: exhausted is full, unlimited and not-in-plan show no meter

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Pick mmx's MiniMax host in the order mmx does

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Ask a self-hosted Kimchi login's own gateway, never the public one

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count only Kimchi replies on Kimchi's own gateway routes

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count DeepSeek Harness sessions on its pi-ai deepseek route

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Show the DeepSeek balance in the currency that holds credit

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say Out of credits when DeepSeek's balance no longer covers a call

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read Ollama Cloud limits from its documented balance endpoint

/api/usage only reports usage history. /api/balance gives a legacy plan's
session and weekly windows with their resets, or a current plan's included
credits against the monthly allowance.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read Ollama's plan under its lowercase JSON key

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Show an OpenRouter inference key refused by /credits as healthy

OpenRouter documents /credits as management-key only. Its 403 means no
balance for that key, not a refused credential; a management key reads
the balance instead when there is one.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count OpenRouter activity reasoning once, inside completion tokens

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read Z.ai's body-level authentication codes as a refused key

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Share the Claude Code transcript reader and who-bills-it classifier

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Claude Code turns under the provider that served them

Z.ai, Ollama, MiniMax, DeepSeek, Kimi, OpenRouter and the local collector
now claim the Claude Code replies the claude collector leaves out, by the
shared claude_code_owner() rule, so each reply lands in exactly one record.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Share the Codex rollout reader the local collector keeps

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Codex --oss rollouts on Ollama cloud models under Ollama Cloud

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Share the DeepSeek Harness log reader, its routes chosen by the caller

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count dsh, Kimchi and MiniMax Code routes under the provider that bills them

OpenRouter claims DeepSeek Harness's pi-ai openrouter route and Kimchi
sessions run with --provider openrouter. A MiniMax Code custom provider is
billed by whoever runs its config.yaml baseURL, so the shared reader hands it
to MiniMax, Z.ai, OpenRouter, the local collector or the others by host, and
to nobody for a gateway nothing can place.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count pi and opencode sessions on the Kimi Code plan under Kimi

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Rebuild the shared Claude Code index in place on --force

Other collectors now read the index too, so removing it made each of them
parse every transcript from scratch on a forced refresh.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Name every shared reader in the collector module's docstring

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Drop the translated manuals' claim that Kimi has no local usage stats

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Codex rollouts by their token_usage_record lines

Codex 0.163 writes one token_usage_record per billed response, a remote
compaction's included, which no token_count line carries. Once a rollout
has records they are its usage; earlier token_count lines still count.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count a Codex subagent's thread under the session that started it

Review, guardian and spawned-agent rollouts name their root thread in
session_meta.session_id (else parent_thread_id); their tokens now book
under it instead of adding a session per rollout.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count a subagent's Claude Code reply that shares its parent's message id

A gateway that reuses one message id gives a subagent's reply its parent's
id too, and the /btw replay rule dropped it. Only a whole copy of the
parent's reply is a replay now.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Don't count Claude Code's built-in local commands as prompts

/model, /clear, /compact and the like run locally and ask the model
nothing; Claude Code answers their command line with a local-command
output line, which now takes the prompt back. Custom commands and skills
still count.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say that Claude Code's compactions are missing from its tokens

Claude Code 2.1.308 writes no usage for a compaction's request, only a
compact_boundary line with the context's size before and after.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Build the shared Claude Code index once when collectors run together

Eight collectors read Claude Code's transcripts, and the update command
starts them all at once, so each parsed every transcript and wrote the
same index. One now brings the index up to date under a lock while the
others wait and reuse it; an index current for every transcript is read
without the lock. On a 357 MB tree a cold update took 8.8 s of CPU and
now takes 1.6 s, with warm runs unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Parse only the pi session files that hold a collector's own lines

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count the Z.ai, Kimi Code and Muse plans under every id pi, omp and opencode give them

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read ~/.hermes whole beside a custom Hermes root, its own state.db too

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count a DeepSeek Harness prompt per typed message, not per billed step

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say DeepSeek, Kimi and Local have no local stats when none were found

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Kimi Code, Qwen Code, Hermes and Vibe usage under the provider that bills it

Each harness collector now keeps only the rows its own service billed,
classifying the route each harness records (Kimi Code's config.toml alias,
Qwen Code's auth type and modelProviders baseUrl, Hermes' billing base URL,
Vibe's provider api_base) with claude_code_owner(). The local collector
reads the self-hosted rows through the same readers (harness_rows).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say Qwen Code and Vibe have no local stats when none were found

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Hold harness stores' routes to the partition: each request counted once by its biller

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Grok prompts from turns, a fork's copied turns once

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Show a unified Grok account's monthly quota beside its weekly credits

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count the calls of a Muse child agent that keeps no log of its own

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Copilot prompts from the requests a typed prompt started

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Leave Copilot requests sent to a custom provider out of Copilot's record

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Show Ollama's purchased credits left as a balance

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Ask OpenRouter's /credits once per refresh, management key first

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Moonshot's pay-as-you-go API under Kimi from every harness

Claude Code turns sent to Moonshot already counted under Kimi, while the
same usage from pi, omp and opencode counted nowhere.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read Claude Code's fallbacks only on a machine without transcripts

A home whose transcripts were all another provider's took its history.jsonl
and stats-cache.json prompts and sessions again, which that provider's
collector had already counted. The fallbacks also counted built-in local
commands as prompts, and prompts sent to another provider's host as Claude's.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Claude Code sent to Alibaba's Model Studio under Qwen

Model Studio's Anthropic-compatible hosts placed nowhere, so Claude Code
sent there counted its Claude-named replies as Claude's and the rest under
no collector. They are now qwen's host for every harness, and the qwen
collector reads Claude Code's and MiniMax Code's turns as zai does.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Key Qwen's scan cache by the routes it classified with

A modelProviders route moved after a scan kept the old partition for up to
900 seconds on --limits-only runs, while another collector already showed
the new one.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read Qwen Code's base URL from its .env files and settings env

A reply no modelProviders route names goes where OPENAI_BASE_URL or
ANTHROPIC_BASE_URL says, which Qwen Code takes from ~/.qwen/.env, ~/.env,
settings.json's env, or the base URL /auth saved. Those replies were all
counted as Qwen's, so a self-hosted or OpenRouter route set up that way,
as Qwen Code documents, showed under Qwen.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep local from counting the zai, kimi and muse ids other collectors claim

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Document that omp's compactions record no usage to count

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Book no Hermes session that never reached a provider under Hermes

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read a Vibe mistral-backend provider's api_base to find who bills it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Key Kimi Code's and Vibe's cached scans by their route config

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep a revoked OpenRouter management key's refusal and balance when the inference key has none

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say Copilot has no local stats when only custom-provider requests were found

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Store MiniMax Code's route match as lists so a cached scan can match it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Rescan local's usage when a harness's route config changes

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say where Qwen's base URL comes from and that Model Studio's Claude Code turns count under Qwen

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count the Claude Code usage cost-state records beyond the transcript replies

Claude Code 2.1.308 writes cost-state lines with each cost lineage's running
per-model totals, which include requests no reply line records: a /compact,
the session title, side queries. Count what they hold beyond the lineage's
replies once per lineage, on the day of the cost-state that first holds it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Name Alibaba's Model Studio among the hosts that take Claude Code's turns

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Codex rollouts on a provider's host under that provider's collector

A Codex model_providers entry whose base_url is a PROVIDER_HOSTS host is
billed by that provider, so codex_route_owner() places it there through the
remembered provider map, now shared from collectors.py, with the map in the
routed scan-cache match. Local and Ollama Cloud apply the same rule.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count a Kimi Code or Qwen Code prompt no request answered where its route goes

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count no Qwen Code OpenAI-compatible reply with no visible base URL under Qwen

Qwen Code sends one to OpenAI's API by default.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say Kimi Code, Qwen Code and Vibe sessions are placed by the current config

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Name the management key when OpenRouter refuses only that key

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say there are no local stats when Muse, Devin, OpenCode or Synthetic counted nothing here

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Move the agent glyphs past quattro's Tailscale mark at U+E90F

quattro gave U+E90F to Tailscale, so the agent marks this branch added
move up one, to U+E910-U+E91F, and the font carries both.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Leave the translated manuals to the sync workflow

docs/i18n.md has a pull request change only the English sources; the
translations follow through the sync workflow.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Book pi prompts on the reply that answered them, on the day they were typed

Every assistant line now reaches the reader, so a prompt another provider
answered no longer lands on the next reply here, and each prompt keeps its
own trusted day instead of taking the reply's.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count MiniMax Code cloud models on a local route, and their prompts, by model

A custom provider on the local Ollama daemon answering with a cloud-tagged
model is Ollama's, not local work, and a prompt takes the model of the reply
that answered it, so a model filter keeps it with its reply.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep a routed scan that missed a Claude Code transcript out of the cache

scan_routed() now counts Claude Code's read-success with MiniMax Code's and
Codex's, through claude_code_usage(), which scan_claude_code() wraps.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Reread every routed Codex rollout on --force in each collector that bills one

Z.ai, Qwen, MiniMax, DeepSeek, Kimi and OpenRouter now pass the refresh
choice to scan_routed(), as Ollama already did, so a forced scan no longer
replays saved Codex file totals.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep Cursor's weekly limits stale when its plan can't be checked

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Bill no one for Vibe sessions on an alias config.toml dropped

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep Claude's own scan out of the cache when a transcript couldn't be read

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Name claude_code_usage and mcode_route_owner's model rule in the collector docs

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Install node for an npm-backed agent after Remove Preinstalls too

Node can stay declared as a lazy tool without being installed, and mise
then refuses an npm agent with mise use -g just as with mise install.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Book an opencode prompt on its first step across providers

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Share one Codex rollout cache across collectors and read grown rollouts on from where they stopped

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count a Kimi Code alias whose provider left config.toml nowhere

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Claude Code and Codex sent to Synthetic under synthetic

Add api.synthetic.new to PROVIDER_HOSTS, give hf:/syn: models with no
override to synthetic, and have the synthetic collector read routed
Claude Code, MiniMax Code and Codex usage like zai does.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say which Claude Code versions write cost-state and what it makes up for

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Point local at the shared Codex rollout cache and drop its own digest

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say which unnamed Kimi Code aliases still count under Kimi

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count pi, omp and opencode prompts no reply answered

A typed prompt whose request failed, or that got no reply, was booked on
no reply and counted by no collector (pi and omp carried it onto the next
answered reply, whatever its provider). It is now a tokenless prompt row on
the model the session was set to use: in pi and omp the failed reply, else
the last model switch or reply; in opencode the user row's model, and in
V2 the reply after it, spent or not. With no model known it counts nowhere.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say where the local collector's harness routes come from

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count a Codex guardian or subagent thread under its parent session in every collector

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Codex's built-in provider where openai_base_url sends it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Kimchi gateway usage from pi, omp, opencode, Claude Code and Codex

Kimchi's gateway bills pi/omp sessions on its provider ids anywhere, opencode's
kimchi provider, and the Claude Code and Codex setups kimchi claudecode and
kimchi codex write. Add kimchi.dev to PROVIDER_HOSTS so those routes are
Kimchi's whatever the model, and have the kimchi collector read them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count each pi and omp request once, and none omp never sent

omp writes a reply that overflowed the model's context window twice under
new entry ids, so a reply with a responseId is now deduplicated by its
provider, response id and time. A prompt whose agent was killed mid-request
(no assistant line) rode on the next reply, whatever its provider, and was
booked twice when a fork copied it; it is now its own row on the model the
session was on. A prompt omp refused before sending (an error with its
errorId and no duration, like Muse Code's invalid credential) counts
nowhere, so it no longer marks the collector in use.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Look up opencode first steps only in the collector's own sessions

The cross-provider first-step query read every assistant row in the
database on each collector's rebuild, about 0.8 s on a 220k-message
store even for collectors with no rows there. Read the gated rows first
and run it only for their sessions, through opencode's session index.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep local from counting the Kimchi ids the kimchi collector claims

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Choose opencode first steps only among the steps the reader keeps

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count a pi or omp prompt once by its own id when a fork answers it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Date an opencode V2 prompt whose request failed by when it was typed

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read Codex rollouts once per forced refresh, not once per collector

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count a Codex entry on the user's own hardware as local whatever its id

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Rescan Codex when openai_base_url moves between OpenAI and an unplaced host

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep an unselected Codex profile's openai_base_url off plain rollouts

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count Qwen Code replies a server gave a recycled response id

Dedupe branch copies by the record uuid /branch copies verbatim, not by
the server's response id.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Name Synthetic and Kimchi among the Claude Code and Codex routes in the agents README

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep local from counting the kimchi-dev/<upstream> routes kimchi claims

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Book each Codex prompt with the model that answered it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Switch a Codex thread's model at thread_settings_applied

A model picked with -m on a resume or /model is written as
thread_settings_applied ahead of any request on it, so a switch-time
compaction was booked to the model before the switch.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read the Codex collector's rollouts through the shared resumable reader

The collector's own reader read any rollout that changed again from the
start, so every refresh during a long session reread it whole. The shared
codex_rollouts() reads a grown rollout on from where it stopped, and now
reads a rollout naming no session_meta as the built-in provider's.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read each Codex rollout once per forced refresh, whenever each collector starts

A forced read only reused what another forced read wrote after it began, so
a collector reaching the rollouts after another had finished read them all
again. omarchy-agent-usage-update now exports when the refresh began, and a
forced read takes what any forced read of that refresh wrote.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Say only config.toml's openai_base_url moves every built-in Codex rollout

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep a Qwen fork's copied records in the session that wrote them

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Count an opencode prompt refused before sending nowhere

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Book a pi prompt a fork answered on another provider under one collector

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Start no Codex app-server for an API-key login, which has no plan limits

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Note that interactive Codex's thread-title requests are missing from its stats

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Treat a pi fork's deleted parent as no parent, so the scan stays complete

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read a pi fork tree's other files only when they hold a prompt the scan booked

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read a pi session header nested too deep as naming no parent

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read a Codex API key beside ChatGPT tokens as an API-key login, as Codex does

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep a deleted pi session tying its forks, so a prompt they copied counts once

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Let bar icons give a little under the pointer

Bar icons had no press state at all. Held down, a widget's label or
glyph now scales to 0.9 over 160ms on the strong ease-out and springs
back the same way. Only the mark shrinks, never the button, so the hit
area and the bar's layout stay put. A bar glyph is small enough that a
panel button's 0.97 wouldn't show.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Press the icon itself on bar icon buttons

BarIconButton draws its icon in its own content item and turns
WidgetButton's label off, so the press scale never reached the icons
most of the bar is made of: the panels, indicators and tray. The
content item now takes it too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Tell a bar widget when the bar's own press handler holds it

Each bar slot lays a MouseArea over its widget for dragging, and that
area takes the press, so the widget's own MouseArea never reported
itself pressed and the press feedback never showed. The slot now marks
the button under the pointer as held from press to release, cancel, or
the start of a drag, and the press scale follows that too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Press bar icons through an open panel's overlay too

With a panel open, its full-screen overlay takes clicks on the bar and
forwards them, so clicking an icon to close its panel, or to switch to
another, showed no press. The overlay now marks the bar button under
the pointer as held until release, as the bar's own press handler does.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
… them (omacom#16515)

* Show each toast from an unnamed sender on its own instead of stacking them

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Regroup a replayed toast after giving it a fresh key

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…6517)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
… a timezone fallback (omacom#16049)

The ISO installer asks for a language first, but the setup form it shares
with first-boot owner setup only speaks English, so the ISO had to swap
its strings by their English text on the way to gum. The form now reads
its strings from cli.json through omarchy_form_text, in the language the
caller names in OMARCHY_FORM_LANGUAGE, with English filling in. The live
ISO carries the catalogs but no omarchy-i18n and no user language file,
which is why the caller names it. First-boot owner setup leaves it unset
and stays English.

OMARCHY_FORM_TIMEZONE_FALLBACK stands in when geo.omarchy.org can't place
the machine, so the ISO can preselect the language's capital offline.

The translations for the 35 languages the ISO's console can draw come
over from the ISO's catalogs. The git identity prompt stays English: only
ISO configurators from before omarchy-setup-git still call it, and they
never name a language.

Co-authored-by: David Heinemeier Hansson <david@hey.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
cursoragent and others added 4 commits October 24, 2026 23:12
The motion check fails a bare animation duration, and the face card's
opacity fades were still raw numbers.

Co-authored-by: Mark Groves <mark-groves@users.noreply.github.com>
Upstream's Cursor and Grok Bot collectors replace the fork collector
that test drove, and the new collector tests cover them.

Co-authored-by: Mark Groves <mark-groves@users.noreply.github.com>
A failed install was reported and then still saved as the default.

Co-authored-by: Mark Groves <mark-groves@users.noreply.github.com>
Remove Preinstalls now deletes a quoted Cursor mise wrapper, and a
window that hits its reset is marked elapsed.

Co-authored-by: Mark Groves <mark-groves@users.noreply.github.com>
@coderabbitai

coderabbitai Bot commented Oct 24, 2026 •

Copy link
Copy Markdown

Important

Review skipped

Bot user detected.

To trigger a single review, invoke the @coderabbitai review command.

Run configuration
  • Source mode: Auto
  • Repository: Unavailable
  • Branch: Unavailable
  • File: No configuration file from this repository applied
  • Commit: Unavailable
  • Layers: Organization UI
  • Review profile: CHILL
  • Plan: Advanced
  • Run ID: 0ec49bf3-071b-4d56-a7ac-f60a82a9d7e6

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

  • 🔍 Trigger review
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

@cursor
cursor Bot marked this pull request as ready for review October 24, 2026 23:15
@mark-groves
mark-groves merged commit 3008383 into quattro Oct 25, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

7 participants