Skip to content

Release v0.2.0 - #245

Merged
BattermanZ merged 59 commits into
mainfrom
dev
Oct 12, 2026
Merged

BattermanZ merged 59 commits into
mainfrom
dev

Conversation

@BattermanZ

Copy link
Copy Markdown
Member

Releases v0.2.0: merges dev into main.

The release was reviewed on the version-bump pull request, #244. just release-publish 0.2.0 merges this pull request, and refuses while any box in that pull request's checklist is unticked. Publishing waits for the maintainer to approve the release title and notes.

BattermanZ and others added 30 commits September 28, 2026 21:40
The nine prototype and archive branches were kept as tags of the same
name and deleted as branches. Three places still said "branch": the
visual identity record's pointer to its prototypes, the comment at the
top of its theme.js, and the #36 note in app.css. The record is
otherwise untouched, since what it chose has not changed.

Co-Authored-By: Claude Code <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011aJcQGaXqh6r8L9zZ3ddo8
The groundwork for the wide layout (ADR 0044). A window is phone, wide
(at least 700 wide and 560 tall) or roomy (wide, and at least 1000 wide).
Nothing on screen changes yet.

The figures are written once, as two Tailwind variants in ui/src/app.css.
The same variants set --room on the page, and the app reads its Room back
from that property on load and on every resize. A `wide:` class and the
Room therefore cannot disagree, and a test compiles the stylesheet to show
that moving a figure moves both.

Screens take the Room from context. A test hands one to the provider, and
without one it is the phone's.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011aJcQGaXqh6r8L9zZ3ddo8
On the wide layout the shell draws an 88-wide indigo rail down the left
edge: the mark with Kamosu's name under it, Home, Recipes, Shopping and
Cooked as the tab bar draws them, and Settings at the foot. The header
and the tab bar are not drawn there. Aurélien chose the rail over a
240-wide paper sidebar and over keeping the header (ADR 0044).

Every screen keeps its own column, centred beside the rail. Bottom
sheets and the unsaved tray keep to that column while their dimming
covers the whole window.

The cooking screen and the Story stay bare. Signing in, an Invite, a
Cookbook Invite, a recovery link and a Share Link's import draw no rail.
They carry the mark and the name centred above the page. The device
remembers whether somebody was signed in, so a stranger arriving at /
or /import never sees the rail come and go, and signing out takes it
away with the Session. Any other page keeps the rail whoever is there,
so a wide window always has a way to /.

The phone layout is unchanged. The navigation list is called places,
since a Section is a part of a recipe.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…yout (#195)

A round button holding a bare arrow sits at the top left of the page,
just right of the sidebar, and stays there while the page scrolls. A
recipe, its History and Share, a Food, an import Report, Foods and
Brought in carry it. Home, Recipes, Shopping, Cooked and Settings do
not, and the phone layout has none.

It goes back in the browser's history, so a search on Recipes comes
back as #191 made it. With nothing of Kamosu behind the page it goes to
the page's parent instead, and it never walks back out to another site.
The tab remembers where each visit began to tell the two apart.

On the wide layout the arrow replaces the small "‹ Settings" line an
import Report and Brought in draw. The phone keeps that line.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
Thirteen sheets each drew their own frame. One component draws them all
now, SheetFrame: a bottom sheet in a phone Room, as before, and in a
wide Room a window 480 wide in the middle of the room beside the rail,
with the page lightly dimmed behind it. Aurélien chose that window on
the prototype over one as wide as the page's column.

The frame owns what the sheets repeated or forgot. Escape closes every
one, where six had no Escape. The caret moves in when a sheet opens and
goes back to what opened it. Enter confirms where a sheet asks one
thing: a confirmation with no gentler act beside it, a false start at
the stove, a copy or a translation in the editor, the paste check.

A Step's photograph still covers the whole screen on every layout.
History's two sheets were not among the twelve the ticket counted and
became windows with the rest.

The phone is unchanged: every sheet photographed before and after at
three phone sizes is byte-identical.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…shows the next Step (#197)

On the wide layout the Step is 38px where the phone's is 26, the amounts
27px, and Back and Next are 72 tall. The next Step is written small under
the Step on an upright tablet, held to three lines, and beside it where
the window is roomy. The last Step shows none, and nor does a Step the
cook is writing on. Where the window is roomy the amounts sit in two
columns, and the screen stops widening at 1120.

The phone's cooking screen is what it was. Recorded in ADR 0044.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
Where the window is roomy the page's column grows from 672 to 1040. The
Ingredients take two fifths and the Method three fifths, and the
Ingredients stay in view while the Method scrolls. What follows the
Method goes back to one 672 column. The editor has the same layout, and
a line dragged to the edge of the Ingredients' column scrolls that
column. A recipe read against another Branch keeps one column.

The part that stays in view is a box inside the left column. As the
column itself it stayed for the whole page and covered everything under
the Method, which a screen test now guards.

A phone and an upright tablet keep the one column they had. Recorded in
ADR 0044.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…ers (#199)

On the wide layout the Foods list is 300 wide and stays where it is while
one Food after another is opened beside it. Its title and search box stay
at its top and its rows scroll under them. A row is the Food's name with
the number of lines that point at it, and the open Food is a card no
wider than 600. Before a Food is open the right side says to choose one.

The up and down arrows open the Food above or below in place of the open
one, so going back afterwards is one step. A click on a row adds an entry
as any link does. The arrows are the caret's own while it is in a field.

The list is drawn by the layout of the two routes, so it keeps its scroll
and what was typed in it. A Food corrected on its card changes its row at
once, and a save that answers after the next Food is open is no longer
drawn under that Food's address.

The address of a Food is the same on every layout. The phone keeps its
two full pages and its rows. Recorded in ADR 0044.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…blets and computers (#200)

On the wide layout a list of every import is 300 wide and stays where it is
while one Report after another is opened beside it. Each source has a
heading with what it remembers, and under it each time something came in:
the date and time, with what happened under it. The heading leads to the
source's screen, which stays a full page and is still where forgetting
lives. The open Report is a card no wider than 600. Before a Report is open
the right side says to choose one.

The up and down arrows open the Report above or below in place of the open
one, so going back afterwards is one step. A Report that watches its import
end tells the list, and a row that says running is read again when it is
opened, so no row goes on saying an import is running after it ended.

Each import gets a Report of its own, so one opened after another never
shows the first one's answered pairs.

Arrows pressed quickly showed the "Kamosu went wrong" card, on Foods too.
A navigation overtaken by the next rejects its promise, and nothing caught
it. Both lists now open in place through one helper that expects this, and
the page transition handles its own rejected promises.

The address of a Report is the same on every layout. The phone keeps its
three full pages. Recorded in ADR 0044.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…rs (#201)

On the wide layout the three browsing screens start at the rail and run to
the window's right edge. Recipes takes as many columns as fit with no card
narrower than 160, the size it has on the phone: 4 across on an upright
iPad Pro 11", 6 on its side, 8 on a MacBook 14". Its search box stops at
672 and the filters stay under it. Home's shelves are still sideways rails
and their cards keep the phone's 168, so an iPad on its side shows two
whole shelves before scrolling.

On Cooked a month is a grid of cards no narrower than 224. A cooking opens
inside its own card, and a tap anywhere on a card opens it. Cards in a row
share a height until a cooking in that month is open. The two buttons of
the delete question stack where their words do not fit side by side, which
is French and Spanish in a narrow card.

Aurélien chose this on prototypes over cards at least 224 wide. The phone's
three screens are what they were. Recorded in ADR 0044.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…on each recipe (#202)

On the wide layout the page is as wide as a recipe's. The recipes chosen take two fifths and stay in view while what to buy scrolls in the other three. Each recipe has − and + either side of how much of it is being shopped for, and the amount between them still opens the two boxes. The amount changes as it is tapped, so two quick taps are two servings. A recipe with no Yield, one like "1½ loaves", and one shopped for in another noun than its own keep the sentence. A recipe that has moved says what it is written for. With nothing chosen the recipes' side still stands. The phone's screen is unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…pe search (#203)

Under the pointer the paper darkens, which Aurélien chose on 6 October 2026
over drawing an indigo line. An outlined button, a tag, a filter or a list row
fills with the darker paper. An indigo button turns a lighter indigo, the one
colour this adds to the palette. A text link gets a patch a little wider than
its words, and a recipe's card brightens. The pointer is a hand over anything
that can be clicked. All of it sits inside (hover: hover), so a touch screen
shows none of it.

The rules are written once at the foot of ui/src/app.css for every control.
A surface says what darkens on it, and a control with a fill of its own is
named fill by fill. A screen test fails on a hover: class written on a screen
and on a fill with no line there.

/ puts the caret in the search box on Recipes and opens Recipes with the caret
there from anywhere else. It types a slash inside a field, waits while a sheet
is open, and does nothing on the cooking screen, the story or a page read
before signing in.

Enter now saves a Food's names, a Food's cup weight, a line's Reading, the
Operator's public address and the address asked before a first Share Link.
Each was a field beside a button with no form round them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
The domain-modeling skills now look for GLOSSARY.md. Every pointer to the
file follows: AGENTS.md, README, the ADRs, and the source comments that
cite a term. No behaviour changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SzxAWjgZaUEQ3fr6kbhS1m
…204)

Each drop is the act its row under the + is, run on the +'s own code. While
something is held over Recipes the page is covered by a sheet of the paper
with a dashed frame, and words in the middle say what letting go will do.
Aurélien chose this on #204 over a bar across the top. Something Recipes
cannot take turns the frame red and leaves one sentence under the search box.
Offline a drop says what its button says and sends nothing.

Everywhere else a dropped file or link does nothing, and the browser is kept
from opening it in place of Kamosu. Both halves are in ui/src/lib/drop.svelte.ts,
which #205 builds on.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…aid once

#192, the parent spec of the wide layout, carried ready-for-agent and was
the oldest open issue in the queue. Eight sessions in a row fetched it whole
before passing it over. The label is off, and the tracker notes say why it
stays off a parent.

The CodeGraph section repeated what each host's own agent instructions
already say, so every session read it twice. The index is machine-local, so
the host is where the instruction belongs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
CODING_STANDARDS.md is what a reviewer reads, and holds only the calls no
check can make. Its first rule is the one #202 needed: the option Aurélien
chose is part of the spec, so the built screen is compared with that page,
control by control. On #202 the − and + were built square and 48 where the
chosen page drew them round and 40, and no review caught it because no
reviewer was shown the page.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
The file is for anyone who runs Kamosu, so it says "the person who chose"
where it named one, and "the machine that built it" where it said the dev
host.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…205)

A photograph dropped on a recipe's photograph, on a Step's or on a cooking's
sets it, and a Crouton export dropped on its import in Settings starts that
import. Each drop is the act its button already does, run on the button's
own code, and offline it says what the button says.

The recipe's photograph is big enough for the look chosen on #204 and is
covered by it. A Step, a cooking's photographs and the Crouton import are
small places, so nothing is covered there: a dashed frame goes round the
place and its button turns solid and says what letting go will do, the
option chosen on #205. A file of the wrong kind turns the frame and the
button red and leaves one sentence beside the button.

FilePlace in ui/src/lib/drop.svelte.ts is a zone for one file of one kind.
DropCover and DropHere draw the two looks.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
The repository is for anyone who runs Kamosu, and 141 tracked files named
the person who runs the first instance. None does now.

The invented cook in the tests is Stéphane: `Stéphane`, `p_stephane`,
`stephane.example`, the variable `stephane`. The ticket said Clémence. The
tests call that cook "he" about 150 times, often to tell him from Camille
or Marie, so a man's name of the same length, bytes and single accent was
chosen while this was built. The choice is a comment on the ticket.

Comments and records say what was chosen and where it is recorded, and
leave out who. ADRs and the design and research records change in the
naming alone: "the first Operator" where a person is needed, "the person
who chose" for a choice. No decision, date, figure or reasoning moves.

One corpus test matched a real recipe by a title that opens with the name.
It now matches that title by how it ends. CODING_STANDARDS.md gains the
rule that keeps the name out.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
A dated record of the audit, numbered as the tickets #218 to #228 refer
to it.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
#224 #225 #226 #227)

The phone's bar takes the rail's look and the two draw a place with one
component; the header's You card is a gear, and the header scrolls away
(#218). The round back arrow stands on the phone too, in the page's flow
and sticking to the top (#219). Cook this is under the servings and the
other acts are a list of rows (#220). Settings is a menu and four pages,
each reading only what it shows (#221). A phone turned sideways cooks in
two columns, the top row's two statements fold under one ⋯, and the
how-much chip wraps rather than truncating (#222). Shelves get earlier
and later under a pointer only, and tag chips wrap on the wide layout
(#223). One date helper in the app's Language, a title on every page,
plurals, the Food note un-reddened, the device named right (#224). Words
that act are buttons, labels 12px, a visible caret, the editor's tools as
marks (#225). The shopping row, the notice cards' width, no install card
signed out or on the wide layout, a compact stepper (#226). The opened
diary card is one card on the wide layout (#227).

ADR 0044 records what the phone took from the wide layout.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
The small buttons that replaced underlined words, the shelf arrows and
the editor's per-line marks keep the size the chosen pages drew them at,
and each grows an invisible round itself to 48, as the cooking screen's
chips have since #88. Two side by side keep a gap at least as wide as
their two overhangs. Recorded in ADR 0044 (#223, #225).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…220)

Option 1 of the audit of 9 October 2026, chosen on 10 October over the
list of rows built first: Cook this and Edit side by side under the
servings, the other acts as a grid of two-column tiles at the foot, and
Delete apart from them as the small underlined word it was. A tile whose
act has happened is filled: the recipe on the shopping list, the Sheet
ready to share.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
Edit this recipe goes back among the acts at the foot, first in the
grid, so Cook this stands alone under the servings. Every button in the
grid is one line and one height: the line under History, how many saves
and when, made it taller than its neighbours, and it goes; the History
page's own first line says it. The module that built that line goes with
it.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…check asks

The last commit on #220 stopped using a grid class and left its rule in the
committed stylesheet, and left two files Prettier would lay out otherwise.
`just check` failed on both. This is `just css` and `just format`, nothing
written by hand.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
`Oignon jaune : 2`, `Eau : 800 ml` and `Pâtes orzo 200 g` were each stored
whole as a Food with no amount, so every distinct amount named a Food of its
own. The reader now reads what follows a colon by itself. A measure there
measures the Food before the colon, words that only describe leave that Food
with no measure, and a name makes what stood before it a label. An amount and
its Unit at the end of a line with no colon are read as its measure. A bare
number at the end stays in the name, which keeps `Pastis 51` whole.

A label is dropped only where what follows reads cleanly. Where an amount has
no Unit the reader knows, or the words end as a sentence does, the line reads
as it did before, because the paste reader puts every pasted line through
here and lost a recipe of its corpus otherwise.

`à convenance` is dropped as `to taste` is, and `cm` is a Unit of the open
set. Lines already stored are read again by `reread_ingredient_lines`.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
`100g Crème` and `800ml d'Eau` were each stored whole as a Food with no
amount, because the reader cut a line at its spaces and `100g` is no number.
A word made of an amount and a Unit of the closed set is now read as the two
words it would be with a space, wherever an amount is read: at the start of a
line, after a colon, and at the end of a line. A number against anything else
stays in the name, so `7up`, `3oeufs` and `2cm piece ginger` read as before.

`cs` and `c.s.` are tablespoons, `cc` and `c.c.` teaspoons, and `pince` is a
Unit as `pincée` is. After an amount a lone `t` is a teaspoon and a lone `T` a
tablespoon, the one Unit whose case matters, and the shopping list keeps the
two apart where it adds Units by their spelling.

`un peu de`, `a little`, `a bit of` and `un poco de` opening a line are
dropped, as `quelques` is.

The reading corpus test scores agreement with Crouton's own unit field, and
Crouton filed the new spoons as bare counts. Seven lines that now read a Unit
count against that share, so its floor moves from 0.95 to 0.94.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
A cooking whose As Cooked was kept as a Version showed, with its rating,
under the Version it started from. The Version that was word for word what
was eaten looked untried.

get_thread now answers eaten_version_id on every Attempt: the As Cooked's
Version once it is on the Thread, otherwise the Version the cooking started
from. The History files cookings by it. The Core works it out because it
withholds a Kitchen-mate's As Cooked, so a screen could not.

A cooking kept inside the collapse window loses the Version it started
from. It stays on the Thread now, under the Version it was kept as.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
An open cooking in Cooked now has a section, "What you changed", that lists
the lines the cook did not leave alone and offers to keep them. It keeps
through promote_as_cooked, the Operation the recipe page's offer uses, so the
Version that lands is the same one. The first tap saves nothing; the second
does, after the moved-on warning and the Copy sentence where they apply.

list_attempts answers one more field per cooking, `unkept`: the Branch
keeping would land on, whether the recipe has changed since, and whose the
recipe is. It is null where there is nothing to keep. Keeping lands on the
Branch that was cooked, so a cooking of a variation is not written over the
plain recipe the diary opens.

"Not now" on the diary closes and calls nothing, and a cooking answered
"Leave it in the diary" on the recipe page is still offered here. Both were
chosen on the ticket on 10 October 2026, with option A for the layout.

The changed lines are drawn by ChangedLines.svelte, which the recipe page's
offer now uses too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…e marked (#211)

Tapping a Version on the History opened a small sheet with its lines and
nothing about what had changed since. It now opens a page of its own, at
/recipes/<branch>/versions/<version>, laid out as the recipe is.

A new Operation, changed_since, lays an older Version over the newest
Version of its Branch. It answers what divergence answers and shares the
Core's pairing with it. The page marks the lines with the row component
two Branches use: a changed line, a line taken out since, and a Ghost for
a line written since.

A dark bar stays at the top of the window and leads back to the recipe as
it stands (option B of the prototype recorded on the issue). The page can
be cooked from, and whoever writes the recipe can carry a line back into
it and save an ordinary Version. Tags, the amount picker, sharing and the
other acts stay on the recipe as it stands.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
BattermanZ and others added 29 commits October 10, 2026 19:50
Since #184 any run of words inside a Reading's target could stand for the
whole name, so "steamed rice" linked the rice vinegar and "add the chicken"
linked the chicken broth. A part of a name now links only when it holds the
name's main word: the last word in English, the first in French and Spanish.
get_recipe hands each Version's Language to step_uses, and a Branch with no
Language is read from both ends.

A cut, part or form gives way to the food beside it, so "the chicken" still
links chicken thighs and "la vanille" links extrait de vanille. The words
for that are a new list in reading.rs.

On a development library of 96 Branches this removes 46 of 1,038 links and
adds none. The ticket lists them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
The reader compares every joining word with accents folded away, so French
"thé" matched the English article "the" and was dropped. "1 thé vert" named
the Food "vert" and "200 ml de thé" named nothing. "dés" and "dé" were lost
the same way, as "des" and "de".

The three accented spellings are now kept. Unaccented "the", "des" and "de"
still join, and so does "à".

Keeping "thé" broke the Quebec teaspoon, "cuillère à thé", which had read
correctly only because its "thé" was dropped. It is now listed as a
teaspoon, long and short forms.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…214)

The pages somebody reads before they are in the app draw one way on both
layouts: no tab bar, no gear, no sidebar, the mark and the name. #226 built
the phone half. This closes what #214 still held open.

A Cookbook Invite is for a Person who already has an account, so the shell
now treats its address as it treats `/`: bare until this device knows a
Person is signed in, and with their navigation once it does. The page says
what it heard, so a device that knew nothing gets the navigation as soon as
the link is read. It carries no back arrow, being a page opened from a link.

An Invite and a recovery link stay bare whoever opens them. "What is
Kamosu?" keeps its place under the login form and nowhere else.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…tors (#215)

The test looked for its lock in /proc/locks, which lists every lock on the
machine. The kernel hands that list out a piece per read and finds its place
by counting, so a lock released by any process between two reads makes the
reader step over one that is still held. Beside processes opening and closing
SQLite databases the test went red 8 runs in 400, and the whole suite on run
25, always at that first assertion with the lock in fact held.

holds_a_lock_on reads /proc/self/fdinfo for each descriptor open on the
database file. Same company, 1,200 runs, none red, and the assertion still
fails when a plain open and close drops the lock. src/db.rs is unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…ust (#216)

Six recipes replace the Playwright scripts each ticket copied out of the
last one's gitignored folder:

- just dev-session / dev-session-end sign in as the account in .env and
  keep the cookie in .dev/sessions/. They print the file's path, never
  the cookie or the password. Ending revokes the Session the cookie
  itself belongs to, so no other can be ended by mistake.
- just screens photographs each path at every size and fails on a
  console error or a sideways scroll. --webkit runs it in Safari's engine.
- just screens-same compares two sets byte for byte and names what
  differs. Two runs over an unchanged screen write identical files.
- just proto-new captures the real screens as static pages with their
  photographs and a copy of assets/. just proto-serve serves a prototype
  by name to the LAN with caching off, and serves nothing else.

Playwright is now a dev dependency of ui/. The sizes live in
ui/tools/sizes.mjs, and every script reads them from there.

Review turned up three faults, all fixed here: a path holding ? or & was
pasted bare into the shell, a failing instance was read as an ended
Session and its cookie dropped, and proto-serve would serve any folder,
the repo root with .env in it included.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
… dark mode (#228)

Three questions the audit of 9 October raised, each put on prototypes and
answered on the ticket.

Type follows the reader's own text size, from the sizes it was drawn at up
to 135% (ADR 0046). Every type size is a count of --type-unit, which follows
the root's font size. Chrome on Android sets that from the system once a
page carries the text-scale meta; iOS is given the system's body font on the
root. Only type moves, so the stylesheet no longer writes a length in rem.

A Cover is drawn from sixteen pasta shapes where there were eight. The list
doubled so that each Lineage keeps its shape or takes the one eight places
further on, and none moves between two old shapes.

Kamosu has no dark mode, on purpose (ADR 0045).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…229)

ADR 0005 asks for a quiet mark where a rating came from a different
Version, and it was never built. A "no" from before a recipe was fixed
read as a "no" to the recipe as it stands.

get_recipe now says, per rating, which Version it was given on: the
newest, an older one of this Branch, or one only a variation or a
Translation holds. That Version is the one that was eaten, by the rule
the Thread already hangs a cooking by. A rating whose Version a
collapsing save folded away is older, and is marked and still counted.

The recipe page writes "on an older Version" or "on another Version"
in small words under the name, the option chosen on the ticket. The
words open that Version where the Branch still carries it, and the
Thread otherwise.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…sorted by question (#231)

search_recipes takes a cookbook_id: the reader's own Cookbook or any other
they see. The narrowed shelf holds the recipes with a Branch there and each
card opens that Branch. A Cookbook the caller may not see answers as one that
does not exist. list_tags takes one too, for the Tag row.

The Recipes screen is the design chosen on the ticket. Cookbooks, Kitchens and
Tags are each a headed line of chips, a tap holds one and a second tap lets it
go, so the All chip is gone. One of Cookbook and Kitchen is held at a time.
"My recipes" keeps its meaning and becomes a tick beside the count, "Cooked or
written by me", which combines with everything else. ADR 0027 is amended.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…character stays on a name (#230)

The Step linker folded every accent away before asking whether a word
only joins two others, so "thé", "dés" and "dé" reached it as "the",
"des" and "de" and a short name could not rest on them. It now keeps the
accent on those three words. An unaccented "the" alone still names
nothing, and beside another word of the name it is read as before.

The reader trimmed a combining accent off the end of a Food's name along
with the punctuation. A Reading is now written with each accented letter
as one character, so one name typed on two keyboards is one spelling.
The Ingredient Line itself stays as written.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
dev and main are now protected on GitHub: a pull request is required on
both, admins included, and every merge is a merge commit. AGENTS.md gains
the branching rules, and a workflow closes the tickets a pull request
names when it merges into dev, since GitHub does that only for main.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…w fails loudly

Review of #236. The workflow runs on pull_request_target, so a pull
request from a fork can close its ticket, and an API failure now stops
the run red instead of reading as nothing to close. The tracker notes
give the gh pr forms that work on this host.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
docs: work lands on dev through a pull request
The repository moved from a personal account to the BatterWorks
organisation on 2026-10-11. GitHub redirects the old address, so this
only tidies the 148 links that named it. The Docker Hub image keeps its
name, battermanz/kamosu.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…erworks

docs: Kamosu lives under BatterWorks
…flow reaches main

GitHub reads a pull_request_target workflow from the default branch, so
the closing workflow did nothing on dev. It now also listens for
pull_request, which is read from the pull request itself, and skips that
event for a fork, whose token cannot close an issue.

Closes #238

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
…om-dev

fix: a pull request merged into dev closes its ticket before the workflow reaches main
…s line (#232)

CHANGELOG.md gains `## Unreleased` above v0.1.0, filled once from the 46
commits on dev since that tag and the issues they close. One line per
change a cook, an Operator or an agent would notice, under After
upgrading, Added, Changed and Fixed, each linked to its issue. A fix to
the wide layout, which is itself new, has no line of its own.

After upgrading holds one act: an Operator reads every line again, since
four fixes to the reader leave saved Readings as they were.

AGENTS.md says a ticket that changes what a person running Kamosu sees or
does adds its line in the same commit, and that a ticket with nothing to
notice adds none.

Closes #232

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
docs: the changelog keeps an Unreleased section, and a ticket adds its line (#232)
The Dockerfile takes VERSION and GIT_SHA as build arguments and writes
them to org.opencontainers.image.version and .revision. Both default to
empty, so a build that passes neither works as it did.

No stage that compiles declares either argument, so a build that changes
only GIT_SHA compiles nothing. Measured under Docker and Podman.

Podman 4.9 reuses a LABEL step by its text and never reads the
argument, so with the layers above unchanged a second build kept the
first one's label. A small stage writes both arguments to a file, and
the image copies it in above the labels. Podman compares a copied file
by what it holds, and writes the labels again.

tests/image_labels.rs reads the Dockerfile and fails if it stops doing
any of this.

Closes #233

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
An image says which version and commit it is (#233)
`just release-prepare <version>` opens the version-bump pull request with
the release checklist beside a draft GitHub Release, and once the bump has
merged, the pull request from dev into main. `just release-publish
<version>` merges it, tags the merge, has the release hook build and push
the images, and publishes the release. Both work out where the release
stands from git and GitHub, so a run after a failure carries on at the
first step not done.

The scripts are Kamosu's own copy of Hatchdoor's, in ui/tools/, without
the deploy step, the What's new page, the docs-freshness check and the
roadmap. Their tests run in `just test`. `just release-library <folder>`
carries a copy of a real library through every migration for the
checklist. docs/releasing.md is the runbook and ADR 0047 the decision.

Closes #234

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
Found on the live rehearsal: `just check` type-checks the screens against
messages a build compiles, so a clone that never built refuses.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
A release is cut by two commands, with a runbook (#234)
…ase (#235)

docs/docker-hub/ holds the overview and the one-line description of
battermanz/kamosu, in the wording the maintainer chose. A test holds both to
Docker Hub's limits, keeps every address in the overview absolute, and keeps
its Compose file level with docker-compose.yml. The runbook adds publishing
the page to the release hook's duties, and the release checklist gains a line
for it.

Closes #235

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bd9L6cqnwnqgjZotYHWUSM
The Docker Hub page is kept in the repo and published by a release (#235)
Sets the version to 0.2.0 in Cargo.toml, Cargo.lock, ui/package.json, ui/package-lock.json, and dates the changelog's Unreleased section as v0.2.0. Made by `just release-prepare 0.2.0`.
rustls moves to 0.23.45, past RUSTSEC-2026-0285, and `npm audit fix` takes
the lockfile-only updates on the interface side. The seven advisories that
arrive with Typst have no upgrade to take and are recorded in deny.toml with
why a Sheet never reaches them. The one npm finding left, braces under the
Tailwind command-line tool, is recorded there too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UC2Kj954nT9iSHXhT8qJPX
…y what 0.2.0 ships

The README and the Compose example name 0.2.0, point at the Settings pages
as they are now, and describe the tablet and computer layout. ADRs 0011,
0027 and 0035 each gain a note for what the release changed around them.
The glossary's Tag belongs to a Cookbook, as it has since #131, and it
gains Divergence and the words the interface uses for a Sheet's namesake,
a Cover's shape, Promotion and a recipe's Steps.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UC2Kj954nT9iSHXhT8qJPX
@BattermanZ
BattermanZ merged commit 89020c3 into main Oct 12, 2026
1 check 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.

1 participant