Skip to content

v4.2.0

Choose a tag to compare

@github-actions github-actions released this 13 Nov 23:01
· 5 commits to main since this release
79b10cf

A paint run with nobody to answer used to come out empty: a title, a button, and nothing only
this product could show. The thesis now has to name a first screen, the job of each colour and
the one motion, and to say what stands in place of anything an assumption takes out. When the
brief says to invent no figure, the example on that first screen carries none. On our one minimal
landing brief the pages are no longer empty. On SwiftUI, the modules now teach the system zoom
transition for a row that opens a screen, which keeps the navigation stack and the swipe back,
and cast reads its thesis against the loaded modules before any code. Of five runs of one case on
those modules, four used the system transition and one built it by hand, and nothing shows yet
that the new text caused the four. A run costs more than before, not less.

Changed

  • paint clears the ground before the theses. It names the page this category publishes by
    routine and its predictable opposite, lists five worlds the design could draw from, sets the
    first two aside and writes both theses from one of the other three. What the user brought or
    confirmed is kept as given. The outcome is one line, Set aside: ... Kept: ..., under the
    theses and in the final report. Light scope and preserve mode skip it.
  • The visual thesis has a fifth point, First screen:, a labelled line under the identity
    sentence: the one thing the screen shows that only this product could show, what it is made of
    (the project's own material, or an example labelled as an example), what takes the width on a
    wide frame, the form it takes on the narrowest frame, and that it already shows the idea at
    rest. The colour point gives each colour its job, the typography point asks for the reason
    behind a system face, the spacing point names the edges and columns every section answers to.
  • An example on the first screen takes its values from the brief or the project, and is
    named in Allowed patterns:. When the brief says to invent no figure, the example carries no
    figure the brief does not give: it shows the mechanism with the brief's own material and words
    (the names of the things compared, the kinds of mismatch, the states), unless the thesis says
    otherwise with its reason.
  • The interaction thesis has a fifth point, The one motion:: the moment where the interface
    does what the product does, with its trigger, what is on screen if it never runs and its
    reduced-motion form, or none with what states the idea in its place. Not asked on light scope.
  • Phase 4 has two more rules, for what the run writes. A block wider than a line of text gets
    its narrowest-frame form in the same edit that writes it. A stretch left empty is a decision
    the thesis states, or it is closed, and a section that only restates an earlier one is merged
    into it. A section that was already in the project is never merged or closed by these rules.
  • Phase 5 reads the page from the code before the checklist, outside light scope: six
    labelled entries, each with its file:line (First screen:, Sections:, Narrowest frame:,
    Empty stretches:, Colour roles:, The one motion:). What comes back not found is built
    from the thesis, in one round, on the lines the run wrote. With someone answering, merging a
    validated section or taking a decision the validated thesis does not hold is proposed first.
    The entries open the checked group of the report and are not counted in its checks.
  • The audit of cast, paint and bunshin has one more check, "Loaded modules against code".
    For each module the run loaded, it walks the rules that bear on the lines the run wrote and
    reports each departure with its file:line. A departure is licensed by a sentence of the
    validated thesis, quoted, or it is a problem. A thesis marked UNVALIDATED licenses nothing.
  • cast and paint read the thesis against the modules right after loading them, before any
    code.
    The thesis is written before the modules load, so the audit check above could only
    list a departure once the code existed. Where the thesis builds by hand what a module says the
    platform or the library provides, or gives up something a module rule protects (a system
    gesture, reduced motion, a frame budget), the run now says so first. A value the thesis chose
    inside a module's range is a decision, left to the audit. With someone there it asks once,
    revise or keep per gap, shows the thesis with every change in it and waits for a yes: an
    approved thesis is never changed silently. With nobody answering, a thesis the request gave
    or approved is built as approved and its gaps are listed; otherwise the run rewrites the
    thesis on the module's way, prints it with a Revised after LOAD: line and builds that. A
    departure stays only when the request asks, in its own words, for a result the module's way
    cannot give, and words that rule a technique out ("not a push") are read as what the person
    sees and can do, not as an API name. The step runs once. In paint, a revision that changes a
    value MASTER.md or a token file holds is written into them in the same step. bunshin is
    unchanged.
  • Two lines of the scan no longer fail under zsh. The legacy-bridge line and the Apple
    project line were an ls on several globs, which zsh refuses as soon as one matches nothing:
    a project with an .xcodeproj and no .xcworkspace lost its Apple line. Both are a find
    now. The Gradle lines and the pbxproj line still pass globs to grep.
  • The headless region, shared by cast, paint and bunshin, has one more sentence. A
    prohibition in the brief is read at the scope it states, never wider, and when an assumption
    takes something out (a picture, a motion, a chosen typeface), the thesis says what stands in
    its place.
  • tells loads once the thesis is written and, when someone is there to answer, validated;
    never before.
    The load region of the three orchestrators and the module's protocol say so.
    The region also said the module weighs "about 20,000 characters" with its web reference: it
    is close to 25,000.
  • tells: a tell taken out leaves a place. An empty place is a default too, so the thesis says
    what stands there, or that nothing does and why. The module still names no replacement. "Fake
    product drawn in divs" now means boxes passed off as the product, and its question asks first
    whether the project's own material shows what the product does.
  • paint lost text other rules already state: eight lines of the red-flag table (one new line
    replaces them: "the safest thesis is the one that promises least"), the two one-line thesis
    examples, the --persist paragraph and the long form of the MCP tools section. SKILL.md
    still grew, from 68,075 to 78,596 characters.

Changed, on SwiftUI

  • swiftui-motion teaches the system zoom transition first. When a row or card opens a pushed
    or presented screen, it starts from navigationTransition(.zoom(sourceID:in:)) with
    matchedTransitionSource(id:in:) (iOS 18+), which keeps the stack, the bar and the back
    gesture. matchedGeometryEffect stays the tool inside one hierarchy. The module lists what a
    hand-built overlay gives up, and how to gate the modifiers below iOS 18.
  • A drag that moves its own view reads in a stable coordinate space. A DragGesture inside
    the .offset its translation drives takes coordinateSpace: .global or a named ancestor. In
    the default .local space the view flickers between two positions. The examples of
    swiftui-motion and mobile-principles now do so.
  • mobile-principles: giving up the swipe back is a stated cost. "Never override" became
    "keep it by default". A thesis that replaces the navigation stack says what is lost (the edge
    swipe, the bar, its back button) and what replaces each.
  • VERSIONS.md has two new rows, read in Apple's documentation on 2026-11-13: the zoom
    navigation transition and the DragGesture coordinate space, with their availability.

Measured, on the web

One brief, minimal-saas: a landing page for an invoice-matching product, with an explicit
request for a minimal, restrained style, run headless with claude plugin eval on
claude-opus-5-5. Two runs per arm. The 4.1.3, frontend-design and no-plugin pages are the
baseline runs of the same week. The pages were scored blind on stills, out of 10, by one model
in three reviewer roles. No human panel.

Rounds 1 to 3, the last one reported. The new arm ran on this branch at 5e3f9c4, before
the review edits and before the rule on example values.

Set This branch 4.1.3 Anthropic's frontend-design No plugin
M1 (new run 1, 4.1.3 run 2, the others run 1) 7.50 5.17 7.33 5.83
M2 (new run 2, 4.1.3 run 1, the others run 2) 7.67 4.33 7.17 5.67
  • Above the 4.1.3 pages by 2.3 and 3.3 points, above the no-plugin pages by 1.7 and 2.0. All
    six votes agree on both.
  • Level with frontend-design. The new pages lead by 0.2 and 0.5 points, and the same
    unchanged pages moved by up to 0.8 between judging rounds. That lead is inside the noise.
  • Both new pages showed an example invoice with invented quantities and prices, labelled as
    illustrative, on a brief that asked to invent no other figure. The rule on example values
    above was written for that.
  • The run behind the M2 score failed the paint-loaded grader (score 0.67): its trace does
    not show that the modules paint asks for were loaded.
  • Two other briefs, one run each, were held to watch for a regression (etale-landing 7.50
    against 7.17 for the earlier page, nocturne-planetarium 6.83 against 5.33). The judges saw
    none. That is too few runs to say more: the earlier nocturne page, unchanged, moved by 1.5
    points between judging rounds.

Round 4, the rule on example values. Two arms, two runs each, on the same brief: the paint
text of this release without the rule, and with it. This time the three reviewer roles were told
to weigh the brief's "invent no figure" as its author would, and saw two stills per page where
the earlier rounds saw three. Rounds 1 to 3 did not give that instruction, so these scores do
not compare with the table above, and the instruction counts against any page that invents
figures. Each set holds one page of each arm and two baseline pages. Means recomputed from the
raw votes, three per page.

Set With the rule Without the rule 4.1.3 Anthropic's frontend-design No plugin
M1 (run 1 of each arm, the others run 1) 8.00 6.50 not in this set 6.67 7.33
M2 (run 2 of each arm, 4.1.3 run 1, frontend-design run 2) 7.83 6.83 7.00 6.33 not in this set
  • No invented figure on the two pages run with the rule. Each shows an example table made
    of words (the kinds of mismatch, the states), and every figure on the page is in the brief.
    The two pages run without it, and the two frontend-design pages, invent quantities and prices.
  • The pages with the rule come first in both sets, on all three votes, 1.5 and 1.0 points
    above the pages without it, and 0.67 and 0.83 above the next page.
  • Judged that way, the text without the rule is below the no-plugin page in M1 (6.50 against
    7.33) and no higher than the 4.1.3 page in M2 (6.83 against 7.00, one vote of three apart).

    The lead over 4.1.3 and no plugin in rounds 1 to 3 does not hold for a reader who
    counts the invented figures.
  • All four runs built on the first try, with no human edit, and passed the three graders.

Runs, receipts, stills and the comparison:
genjutsu-examples, baseline-2026-11.

Measured, on SwiftUI

One case, etale-ios: cast asked, headless, to make a list row open into its detail "as one
surface, not a push". Six runs, same prompt, same model, same starting app, on four commits. Each
was written without a compiler, then built with xcodebuild, unchanged, on the first try, and
filmed in the simulator. Every run made is in the table; none was run again.

Run genjutsu Cost Swift lines Transition Stack and swipe back
1 4.1.3, 5e00595 3.42 USD +487 -78 built by hand given up
2 4.1.4 plus the module fixes, 3e0d6d3 1.53 USD +78 -0 system zoom kept
3 this release before the step, 5105f53 2.46 USD +434 -59 built by hand given up
4 this release with the step, 844f375 1.36 USD +15 -0 system zoom kept
5 the same 1.30 USD +13 -0 system zoom kept
6 the same 1.31 USD +22 -1 system zoom kept

"The step" is the one that reads the thesis against the loaded modules before any code.

  • Used the system transition: 0 of 1 on 4.1.3, 1 of 2 on the fixed modules without the step, 3
    of 3 with the step.
  • The step was not seen doing what it was written for. Runs 4, 5 and 6 each named the zoom
    transition in the thesis they wrote before loading any module, as run 2 had. None wrote a
    hand-built thesis, so the step never had one to bring back to the system transition. That path
    is untested.
  • Three of three does not prove the step. At the 1 of 2 seen without it, three in a row
    happens one time in eight by chance. The step's text sits in cast's SKILL.md, which a run
    reads before it writes its thesis, and may have steered the three theses. Nothing here shows
    it.
  • What the step printed. Run 4: Thesis against modules: no gap. Run 5: one revision, the
    Reduce Motion fallback, which left the code with no Reduce Motion guard at all. Run 6: one
    revision, a duration, which the step's own text calls a decision and not a gap; it did not
    raise the Reduce Motion question, which its audit then listed as a problem and left unfixed. In
    all three the output is in the final report only, not between the load and the first line of
    code.
  • Run 3 is why the step exists. On the fixed modules it read them before it wrote code, kept
    its hand-built surface on the brief's "not a push" and wrote what that costs into its thesis.
    Its audit printed "Loaded modules against code" with three problems, all true, the first being
    the hand-built surface, and changed nothing. On screen the time lands on the wrong row, the
    large time sits under the Dynamic Island and the list shows under the detail.
  • The module fixes alone are not proven either. On the fixed modules four runs of five used
    the system transition, and each of the four named it before it read a module. The modules did
    not stop run 3.
  • No drag flickers in the four zoom runs, where run 1's drag to close flickers on every
    frame. They have defects of their own: a white box around the row on return in runs 2, 4 and 5,
    larger in 4 and 5, and in run 6 a line draw the thesis promises that cannot be seen. Reduce
    Motion was filmed for run 2 only.
  • The graders do not tell these runs apart. All six score 5 of 5: the graders check that the
    skill fired, loaded its module and touched the list, not what the transition is.

The six runs, their code, films and receipt:
genjutsu-examples, etale-ios.

What did not move, and what is not measured

  • Cost. A paint run on this brief costs about 2 USD (1.95 and 1.93 with the rule, 1.93 to
    3.07 over the ten candidate runs), against 1.64 to 1.73 on 4.1.3, 0.64 to 0.68 with
    frontend-design and 0.46 with no plugin: about 4.2 times the no-plugin run, where 4.1.3 was at
    about 3.7.
  • The six reading entries do not open the final report. The brief asks for a short report,
    and a line now says a short report keeps them. In round 4, with that line, none of the four
    reports opens on them, and the report of the web smoke run below does not hold them at all.
  • Tuned on this brief. The rules were adjusted over four rounds on minimal-saas, against
    the same baseline pages. Every run of every round is kept. No other brief was run in round 4.
  • Small samples, one judge model. Two runs per arm on the web, six runs of one case on
    SwiftUI (one before the module fixes, two on them without the step, three with it), one model
    in three roles, no human panel. Claude Code was 2.1.325 in round 4 and in SwiftUI runs 3 to 6,
    against 2.1.323 before.
  • The release was run once on each side, before its last step. Round 4 ran paint without the
    new audit check, so a smoke run was made at 5105f53, which holds everything in this release
    but the step that reads the thesis against the modules. On the web (minimal-saas, paint, one
    run, not scored by the judges): 2.37 USD, no figure outside the brief, built on the first try,
    3 of 3 graders. Its report does not print the "Loaded modules against code" check under its
    name, and the six reading entries are absent from it. On SwiftUI it is run 3 above. The step
    itself then ran three times in cast (runs 4 to 6) and never in paint.
  • Not measured at all: a conversation with someone answering, where the thesis to approve is
    now longer; bunshin, and cast on the web, with the new headless sentence; Compose, which has
    no recorded run; motion on the web, which stills do not show; any web brief that is not a
    minimal landing page.