inline-chat-kit 0.49.0

package/CHANGELOG.md

@@ -0,0 +1,2111 @@
+# Changelog
+
+Dates are the day the work landed on `main`.
+
+The versions before 1.0 follow the pre-release convention: **a breaking change
+or new public API bumps the minor**, and the patch is for fixes. Anything that would break an
+existing install is called out under **Breaking**, with what to do about it.
+
+## 0.49.0 — 2026-09-02
+
+### Changed
+
+- **New defaults for the composer's motion**, settled on the dial rather than
+  argued about. Four numbers moved:
+
+  | | was | now |
+  | --- | --- | --- |
+  | `bubble.damping` | 21.5 | 22 |
+  | `bubble.mass` | 0.2 | 0.3 |
+  | `button.staggerEnter` | 0.12 | 0.06 |
+  | `enterButton.bounce` | 0.3 | 0.2 |
+
+  A heavier, better-damped pill and a tighter row: the controls now return
+  half a beat apart rather than a whole one, and the send glyph arrives with
+  less overshoot.
+
+  Nothing about the API changed. A host passing its own `animationConfig` is
+  unaffected; one relying on `defaultInlineAnimConfig` gets the new feel and
+  can pin the old values by passing them.
+
+### Added
+
+- **`wrap.exitDuration`** on `InlineAnimConfig` — how long the controls take
+  to leave when the composer grows onto a second line. It had been a literal
+  `0.15` in three places, which made it the one part of the choreography no
+  configuration could reach.
+
+## 0.48.0 — 2026-09-01
+
+### Added
+
+- **Dictation.** Pass `onTranscribe` to `<ChatInput>` or `<ChatTurnRow>` and
+  the composer offers a microphone beside its plus; leave it out and there is
+  no microphone at all.
+
+  The kit records — permission, `MediaRecorder`, the level of the incoming
+  signal — and hands over a `Blob`. What turns audio into words is a service,
+  and a kit that chose one for its consumers would be wrong for most of them,
+  so the handler is yours:
+
+  ```tsx
+  <ChatInput
+    onTranscribe={async function* (audio, { signal, mimeType }) {
+      const res = await fetch("/api/transcribe", { method: "POST", body: audio, signal });
+      for await (const chunk of res.body.pipeThrough(new TextDecoderStream())) yield chunk;
+    }}
+  />
+  ```
+
+  A string, a promise of one, or an async iterable of deltas — the same three
+  shapes `onSend` takes. The transcript is inserted **at the caret**, so
+  dictating into a half-typed sentence finishes it rather than replacing it.
+
+  `useVoiceInput` is exported for a host building its own composer.
+
+  **`SpeechRecognition` is out of reach through this**, because it holds the
+  microphone itself and will not accept a recording. Stated here rather than
+  discovered: it is the only free transcriber there is.
+
+- **`--ick-voice-meter` and `--ick-voice-meter-spread`** — the ring around the
+  microphone while it listens, and how far it swells at full volume. A wash of
+  the marker rather than a colour of its own.
+
+## 0.47.1 — 2026-09-01
+
+### Changed
+
+- **The card whose pane is open wears a neutral ring, not the accent.** The
+  marker means "this one *chosen*" — a picked option, a badge, a citation — and
+  an artifact being shown beside the conversation is not a choice anybody made
+  about it, it is where the reader is. A ring all the way round instead of an
+  edge on one side, in ink at 45% so it inverts: `--ick-border-bright` is
+  literally `rgb(paper)`, a visible edge on a dark card and nothing at all on a
+  white one.
+
+## 0.47.0 — 2026-09-01
+
+### Changed
+
+- **One column down an answer.** New `--ick-answer-column`: everything that
+  draws in a turn starts on it, and only the boxes bleed wider.
+
+  An answer had two families, each consistent with itself. Things in its flow
+  began at the turn's own edge — the prose, the actions row, and the four
+  disclosures that carry no surface at all (`<Reasoning>`,
+  `<ChainOfThought>`, `<TaskList>`, `<Sources>`). Anything in a box began at
+  its ground's padding plus its card's column. Two families 32px apart down one
+  answer, which is what "the text does not line up with the card" was.
+
+  The column is the value the boxed family already reached — a ground's padding
+  plus a card's — so `<Tool>`, `<QuestionGroup>`, `<Approval>` and
+  `<SystemMessage>` did not move at all. Measured after: all nine components,
+  the prose and the actions row on 32.
+
+- **`<ArtifactCard>` has its ground back.** It was taken away in 0.46.0 while
+  the prose still began at the turn's edge, where a ground under a single card
+  was 32px of indent putting the card's words on a different line from the
+  answer's. With one column it is the opposite: the ground is what puts them
+  *on* it, the same way a tool call's does.
+
+### Fixed
+
+- The actions row is inside `.answer`, so giving it a column of its own put the
+  icons at 64 — an indent applied twice. Guarded, since nothing about the
+  stylesheet says which of the two is the parent.
+
+## 0.46.1 — 2026-09-01
+
+### Fixed
+
+- **The pane has 24 of padding everywhere**, inside and out. It is the only
+  surface in the kit that holds a whole document rather than rows of a
+  component, and a document wants a margin — the nesting chain's 16 is a card's
+  gap around its rows, which is a different job at a different size. It was
+  also running into the top and bottom of the viewport, and a card that runs
+  into the window is not one. Measured: 24 from every edge.
+
+## 0.46.0 — 2026-09-01
+
+### Changed
+
+- **The pane takes a third of the room and can be widened to two thirds.** A
+  fixed 420px was a strip on a wide monitor and half the screen on a laptop;
+  a proportion keeps its proportion. There is a floor either way — a document
+  needs a line length before anything else.
+
+  The control is in the pane's header and the width is `<ChatLayout>`'s, which
+  is the same split as everything else here: how much room the pane takes is a
+  fact about the layout, and the button that changes it has to be somewhere
+  somebody can find it. `pane` is handed `expanded` and `toggleExpanded`
+  alongside `narrow`. There is no widening while it is covering the
+  conversation, because there is nothing left to take.
+
+- **The pane sits away from the edges.** New `--ick-artifact-pane-inset`, on
+  the layout's slot rather than on the pane — how far a card is from the edge
+  of the window is a fact about the layout. It ran into the top and bottom of
+  the viewport, and a card that runs into the window is not one. The covering
+  variant drops it along with its corners.
+
+### Removed
+
+- **An artifact card's ground.** Every other ground in this kit carries
+  something besides its card — a tool call's carries the header, a question
+  group's the title, an approval's the asking and the answering. This one held
+  a single card and nothing else, which is not a surface but an indent, and it
+  read as one: the answer's own prose ran along the turn's left edge while the
+  card's words sat 32px inside it, so the two halves of one answer were two
+  columns. Measured after: card, prose and the actions row all at 312.
+
+## 0.45.1 — 2026-09-01
+
+### Fixed
+
+- **The conversation lost 96px of width the day `ChatLayout` went around it.**
+  `.chat` is a flex column, and a child with `margin: 0 auto` is a flex item
+  that has opted out of stretching — so the playground's centred page was sized
+  by its content instead of by its `max-width`, 624px against 720. Documented
+  on `ChatLayout` as well, since any host wrapping an existing centred page
+  will meet it.
+
+## 0.45.0 — 2026-09-01
+
+### Added
+
+- **The artifact pane** — `<ArtifactCard>`, `<ArtifactPane>`, `<ChatLayout>`,
+  `useArtifacts`, and an `artifact` part so a stream can send one.
+
+  **The kit decides the layout.** On the right, with the conversation making
+  room; below `<ChatLayout>`'s own width it covers the conversation instead. A
+  preview pane is one of the few patterns every AI chat now has, and the worth
+  of a pattern is that it is the same every time.
+
+  **The kit does not decide the content.** `<ArtifactPane>` takes children,
+  because a plan, a table and a diagram are the part that differs between
+  products — and the only part.
+
+  The card is a window: eight lines with a fade at the cut, because an artifact
+  poured out in full is a long message rather than an artifact. It arrives
+  before its content does, which is how one really arrives, and shimmers its
+  own name while it is being written.
+
+  `modal` is the one prop that changes behaviour, and it is not about position.
+  On open, focus moves to the pane's heading — not into its first control,
+  which would skip what the thing is — and it is **not** trapped unless the
+  pane is covering the conversation, where trapping it would lock a reader out
+  of the chat they are still reading. `<ChatLayout>` sets it from its width.
+
+  A container query rather than a media query: the pane answers to the width it
+  has. A kit inside a 480px column on a wide page would otherwise put a 420px
+  pane beside a 60px conversation.
+
+### Fixed
+
+- Two things the axe pass caught in the pane and reading did not:
+  `role="dialog"` is not allowed on an `<aside>`, and a `<header>` inside a
+  landmark is a banner landmark claiming to be the page's.
+
+## 0.44.0 — 2026-09-01
+
+### Added
+
+- **`<SystemMessage>`** — the conversation saying something about itself, and a
+  `notice` part so a stream can send one. `<Context>` warns from 80% that the
+  oldest messages will start dropping out and then goes quiet at the moment
+  they do; this is what says so.
+
+  Three absences, each of them the component rather than a corner cut. **No
+  icon**: every picture in this kit carries a state the words beside it also
+  carry, so one here would say "something is being announced" beside a sentence
+  announcing it. **No dismiss and no action**: it is a line of the transcript,
+  not a toast. **No live region**: the kit has one, written to on a later tick,
+  and a second says everything twice.
+
+  Two tones — `notice` and `danger` — because `<Context>` already settled that
+  an amber in the middle makes somebody learn a scale to read a state they can
+  read in words.
+
+  Five tokens, all pointing at the stack: the ground it sits on, the row corner,
+  the danger tint the tool's error section already uses.
+
+- The playground says it when the window actually fills, and announces it
+  through `announce` — which is what a host streaming a `notice` should do.
+
+## 0.43.1 — 2026-09-01
+
+### Fixed
+
+- **A tool call inside an approval had no room above its header.** Zeroing
+  `--ick-tool-ground-pad` put the header's glyph on the column the title and
+  the buttons stand on, and put the header flush against the card's top edge.
+  One axis was measured and the other was not.
+
+  It is `var(--ick-nest-pad) 0 0` now: nothing sideways, so the column holds;
+  8 above, so the header has room; nothing below, because the tool's body ends
+  in its own gap already. Measured in both states — open and collapsed — the
+  card gives 8 above the header and 8 below whatever it ends with.
+
+  The guard reads the shorthand and states the rule in both directions rather
+  than pinning the value: sideways must be nothing, above must not be. Verified
+  to fail each way.
+
+## 0.43.0 — 2026-09-01
+
+### Removed
+
+- **The shield.** It said "this is about permission" in a picture, beside a
+  title that says it in words, above a card showing the thing and three buttons
+  that are visibly a decision. Every other glyph in the kit carries *state* —
+  queued, running, failed, allowed, denied — and pairs with a word for anybody
+  the picture is not reaching. This one carried a category nothing else could be
+  mistaken for.
+
+  It was also drawn at **24 against everything else's 14**: `.glyph` set the
+  badge's width and height on the icon itself rather than on a box around it,
+  so the `size={15}` in the component was never what the browser used. Which is
+  how it came up at all — it looked too big, and the answer to "how big should
+  it be" turned out to be that it should not be there.
+
+  `--ick-approval-glyph` goes with it, and the title now starts on the column
+  directly rather than after a badge slot.
+
+### Changed
+
+- **The approval draws the card; the subject brings what goes in it.** A tool
+  call inside one brings no surfaces of its own — not its ground, not its card,
+  not its shadow.
+
+  This is the third arrangement and the reason is worth writing down. The
+  approval's card wrapping all three (title, subject, buttons) made an approval
+  holding a tool call two papers for one thing. Letting the tool bring its own
+  card instead left its header stranded on the ground beside the title — two
+  headers over one card, and the second of them a stray row. One card, holding
+  everything that belongs to the subject.
+
+  `.subject` takes no padding, deliberately: what is inside is padded to the
+  column already, so a gap there would push all of it eight pixels off the line
+  the title and the buttons stand on. Measured: title, tool glyph and overview
+  all at 48; chevron and the primary button both ending at 504.
+
+## 0.42.0 — 2026-09-01
+
+### Changed
+
+- **`<Approval>` asks above the card and answers below it.** The title and the
+  buttons are on the ground; the card between them is the subject. They are not
+  part of the thing being approved — they are the asking and the answering.
+
+  The card is gone with them, because the subject already brings one. An
+  approval holding a tool call was a card inside a card: two papers, two
+  shadows, one thing. That is also why the tool no longer has to be flattened
+  into a row on it — an approval sets `--ick-tool-ground: transparent` and
+  nothing else, so a tool call inside one is a tool call.
+
+- **Everything lines up on the subject's own column.** The shield where the
+  tool's glyph is, the buttons ending where its panels end. Measured: shield
+  box, tool glyph and overview text all at 48, chevron and the primary button
+  both ending at 504, against a card at [32, 520].
+
+- **Deny is pulled back by its own padding.** A ghost button is *text* — what
+  you see is the word, not the invisible box around it — and the word sat 16px
+  inside the line the shield above it is on. The two that say yes are the other
+  way round: filled, so the box is the visible thing and its edge is what lines
+  up. A guard ties the pull-back to `Button`'s own padding rather than to a
+  number that happens to look right.
+
+### Added
+
+- **`--ick-nest-column`**, the column a card's content starts in, stated once.
+  Three components have to agree on it now — a question's rows, a tool call's
+  panels, and an approval's title and buttons standing over one — so
+  `--ick-question-pad`, `--ick-tool-pad` and `--ick-approval-column` all read
+  it rather than three numbers happening to match.
+- An approval repoints `--ick-code-pad` to that column, so a subject that is a
+  bare code block lands on it too. Its own 12 would have sat four pixels inside.
+
+### Removed
+
+- **The approval's own `.card`.** With it goes the nesting row that checked its
+  corner, and the guards that required a tool call inside one to be a row: both
+  described the arrangement that caused the doubled paper.
+
+## 0.41.0 — 2026-09-01
+
+### Changed
+
+- **`<Tool>` is a question group's shape now: a header on the ground, a card
+  under it.** The header was the card's own top edge, which meant it had to
+  move whenever the card grew, and — the reason it was raised — that everything
+  in it was measured against the **card's edges**. That is what put the chevron
+  hard into the corner with nothing beneath it to agree with.
+
+  On the ground it takes `--ick-disclosure-inset`, so the glyph starts on the
+  column the card's panels start on and the chevron ends where they end.
+  Measured: glyph left 48 and chevron right 504, against a card at [32, 520] —
+  16 in from each edge, the same column a question card uses.
+
+- **The card says something when it is shut.** One row — the call's summary, or
+  the state in words when it has no summary — padded to the card's column and
+  40 tall, which is a folded question group's row exactly, so a stack of the
+  two lines up.
+
+  It stays when the call opens, and that is a deliberate difference from the
+  question group. There the two bodies swap, and because both are anchored to
+  the same top edge they want the same band and cannot be crossfaded without
+  drawing one over the other. Here nothing swaps: the summary is the card's
+  first row in both states and the detail opens underneath it. Same shape, one
+  fewer thing to go wrong.
+
+- **No hover wash on a tool's header**, for the reason a question group has
+  none: the wash is a rounded box the width of the header, and on a ground
+  there is nothing under it for that shape to agree with.
+
+### Removed
+
+- **`DisclosureHeader`'s `filled` variant.** `<Tool>` was its only consumer, so
+  moving the header onto the ground left the prop, two CSS rules and a guard
+  that "guarded" a variant nothing rendered. The guard now makes the same
+  claim — a tool shut is a folded question's row — against the row that
+  actually draws it.
+
+### Added
+
+- `--ick-tool-pad`, the column a tool card's content starts in, the same job
+  and the same number as `--ick-question-pad`. A guard computes it from how the
+  card actually reaches that column — its own padding plus a panel's — so if
+  either moves this fails rather than quietly meaning something else, and
+  requires it to equal the question card's. One rule, not two that agree.
+
+## 0.40.0 — 2026-09-01
+
+### Fixed
+
+- **The two bodies are never drawn over each other.** Recorded and looked at
+  frame by frame, what read as a flicker was a superimposition: the folded row
+  and the first card are anchored to the same top edge, so they always want the
+  same 40px band, and crossfading them put
+
+  ```
+  3 answers   About them · Household · Support needed
+  1  About them        Milica Stevanović  84  +1
+  ```
+
+  on top of each other at half opacity each. No timing fixes that, because the
+  overlap **is** the crossfade — a dissolve between two different sentences is
+  mush whatever its duration.
+
+  The arriving body waits for the leaving one to be gone. Measured across both
+  directions: the two never carry ink at the same time. What carries the eye
+  across the handover is the box, which is growing throughout — so the pause
+  costs nothing, and there is no bare frame at real speed.
+
+### Added
+
+- **`FoldMotion.fadeInDelay`**, back with a job this time: it is what holds the
+  two bodies apart, not a stylistic pause. `fadeOut` is 0.08 and the delay
+  matches it. A guard requires `fadeInDelay >= fadeOut`.
+
+### Changed
+
+- **`FoldingSlowly` scales every number measured in seconds**, derived from
+  `defaultFoldMotion` rather than listed by hand. Listing them is how the story
+  started lying: it scaled the two fades and not the delay between them, so it
+  showed a long crossfade the component does not do. A slowed-down story that
+  shows something else is worse than none, because it is the one you trust to
+  see the detail — and it disagreed with the instrumented measurement, which is
+  how it was caught.
+
+- Two fold guards match loosely on the source instead of pinning the exact
+  spelling of a call. Both broke on a refactor that changed nothing about the
+  behaviour, which is a guard failing on a rename while it would pass on a
+  wrong number. The number they were standing in for is asserted directly now.
+
+## 0.39.0 — 2026-09-01
+
+### Changed
+
+- **Collapsing is expanding run backwards.** A row leaving was a plain tween
+  while a row arriving was sprung, and the two bodies staggered in opposite
+  directions — last out, first in. The reverse is right when something is being
+  *dismissed*, because it unwinds the way it was built; this is not a
+  dismissal, it is one body replaced by another on the same edge, holding the
+  same answers. One spring builder now serves both states, and both stagger
+  forwards.
+
+  Sampled frame by frame in both directions: the first row leads either way,
+  and `y` runs 0 → -10 on the way out exactly as it runs -10 → 0 on the way in.
+
+- **`rowBounce` is 0.12**, down a third from 0.18. Measured, `y` no longer
+  passes its target at all on the way in — it settles rather than arriving and
+  correcting.
+
+## 0.38.4 — 2026-08-31
+
+### Fixed
+
+- **Text beside a badge takes the badge's line box.** A badge is 24 tall with
+  `line-height: 1`; text beside it inherited whatever line height its type
+  gave it, and two line boxes of different heights, both centred in the same
+  row, land half a pixel apart. Measured with a `Range` over the text itself,
+  the folded group's summary sat at 86.5 against the count's 87.
+
+  `.title` and `.optionTitle` had the rule already — the line box takes the
+  badge's height rather than the badge taking a margin, because a margin is a
+  guess that has to be re-guessed whenever the type changes. `.summaryList`,
+  `.collapsedTitle` and `.upcomingLabel` take it now too, and a guard walks all
+  five.
+
+  Only `.summaryList` moved: the card's own rows were already baseline-aligned,
+  the badge and the title simply being different type sizes. The rule is stated
+  on all of them so the next one does not have to be found by eye.
+
+## 0.38.3 — 2026-08-31
+
+### Fixed
+
+- **A folding row gets a compositor layer while it moves, and gives it back.**
+  Motion animates `y` as an independent transform, and an independent transform
+  does not promote the element on its own — sampled through a whole fold, every
+  row read `will-change: auto`. Motion's own guidance is to name the properties
+  being animated and then take the hint away again, since a permanent hint is a
+  permanent layer. `[data-moving]` on the ground is on for the length of one
+  fold: measured, the layers exist from the click to about 460ms and are gone
+  by 500.
+
+  The taking-away needed care. Both bodies report completion and the leaving
+  one finishes first, so clearing on the first report took the layers back at
+  165ms with the rows still travelling until 300. It clears on the *arriving*
+  body now.
+
+### Changed
+
+- The stylesheet guards look up a rule by its selector at the start of a line.
+  A plain substring found `.summary` inside `.group[data-moving] .summary` and
+  read a `will-change` rule when it wanted the padding one.
+
+## 0.38.2 — 2026-08-31
+
+### Fixed
+
+- **Every badge in a question card is one size.** `--ick-chip-height` was 22
+  against the badge's 24 — measured across every state the card has, the chip
+  was the only box in it that was not 24, two pixels shorter than the numbered
+  badge at the other end of its own row. It points at `--ick-badge-size` now,
+  so a host that resizes one resizes both.
+
+## 0.38.1 — 2026-08-31
+
+### Fixed
+
+- **An answer chip corners like the badge beside it.** New `--ick-chip-radius`,
+  pointed at `--ick-nest-inner` — the chain's innermost corner, the one a badge
+  already takes. They sit in the same row, the number at one end and the answer
+  at the other, and a fully round pill beside an 8px badge is two shapes for one
+  level of the nesting. `--ick-radius-pill` still means pill; buttons and the
+  composer are what want one.
+
+## 0.38.0 — 2026-08-31
+
+### Changed
+
+- **The fold's bodies arrive on springs, one row at a time.** 0.37.0 fixed the
+  distortion and left the swap itself a crossfade — two blocks dissolving
+  through each other. A dissolve is what you reach for when two things are
+  unrelated, and the summary row and the stack of cards are the same answers in
+  two states. It read as the box moving while the content sat there bleeding
+  through itself.
+
+  Rows now enter a little above their place and settle into it, one after the
+  next, and leave the same way; the ground follows them. Both sides of the fold
+  are lists of rows, even the folded side that holds one, so both arrive by the
+  same rule.
+
+- **Springs for what travels, a tween for opacity.** Not a preference. A spring
+  describes where a thing is going and how it arrives, and opacity has nowhere
+  to go — bounded at 0 and 1, so a spring with any bounce overshoots into a
+  clamp and spends the overshoot sitting still. Position and size have no
+  ceiling, which is what makes them worth springing.
+
+### Breaking
+
+- **`FoldMotion` lost `fadeInDelay` and gained four.** `rowDuration`,
+  `rowBounce`, `rowOffset` and `stagger` describe a row arriving; the delay is
+  gone because the stagger is what sequences them now. `visualDuration`,
+  `bounce`, `fadeIn` and `fadeOut` are unchanged in meaning.
+
+  Passing the old shape still type-checks for every key it kept and the rest
+  fall back to `defaultFoldMotion`, so nothing breaks silently — a `fadeInDelay`
+  in an override is simply ignored.
+
+### Fixed
+
+- **`package.json` was published empty in 0.37.0.** A `open(path, "w")` in the
+  script that bumped the version truncated the file before the read that was
+  supposed to fill it. Restored from 0.36.1 with the version applied.
+
+## 0.37.0 — 2026-08-31
+
+### Fixed
+
+- **The question group's fold no longer distorts everything inside it.** Motion
+  animates a size the only way it can: it puts the new one in the DOM and
+  scales the box back. Everything in that box that is not itself a layout child
+  rides the scale — and the section title was a plain `<div>`. Sampled frame by
+  frame over one open: the header went from 23px to **11.67 in a single frame**
+  and stretched back over the next 450ms, the title with it, the cards too.
+  The header wrapper and the body are `layout="position"` now, on the ground's
+  own transition, and the header measures a flat 23 through the whole thing.
+
+- **The leaving body stays where it was.** `AnimatePresence mode="popLayout"`
+  makes it `position: absolute`, and with no positioned parent it resolved
+  **67px down the page** and faded out somewhere it had never been. The body is
+  `position: relative`.
+
+- **And it stays inside the ground.** While the ground shrank, the popped-out
+  list stayed pinned at full height and hung three rows out of the bottom over
+  whatever was underneath. The ground clips; its own 16 of padding is more than
+  the float shadow reaches, so card shadows are unaffected.
+
+- **The fold is a crossfade, not a relay.** The arriving body waited 100ms for
+  the leaving one — a tenth of a second of grown, empty box, which is most of
+  what read as the flicker. Measured at the midpoint, the two now sum to about
+  0.9 of an opaque body; they summed to 0.4.
+
+### Added
+
+- **`defaultFoldMotion` and a `foldMotion` prop** on `<QuestionGroup>`, passed
+  through `<ChatTurnRow>`. Five numbers: `visualDuration` and `bounce` for the
+  box, and `fadeIn` / `fadeInDelay` / `fadeOut` for the bodies.
+
+  `visualDuration` rather than stiffness and damping on purpose. The two of
+  them describe the same spring without either one answering "how long is
+  this", which is the only question anybody tuning it is asking.
+
+- **A `FoldingSlowly` story**, the same fold at a quarter speed — because
+  everything that goes wrong in it goes wrong in about eighty milliseconds.
+- `src/__tests__/foldMotion.test.ts` — the four rules above, each verified to
+  fail by putting the fault back.
+
+## 0.36.1 — 2026-08-31
+
+### Fixed
+
+- **The count on a folded question group is a badge.** It stands in the badge
+  column, on a card, one fold away from the numbered badges it stands in for —
+  and it was a pill: 2px of padding, the smallest type in the kit and the
+  ground's fill, which came out about half their height. Same box now, and only
+  the width differs, because this one holds words rather than a numeral.
+
+## 0.36.0 — 2026-08-31
+
+### Changed
+
+- **The dark theme's ground goes down now.** `--ick-ground` is a wash of ink,
+  and ink in the dark is white — so the recess a group of cards sits in came
+  out *lighter* than the page, with the card lighter again. Three surfaces
+  stacked on the same side of the page, and a card that read as a slightly
+  different patch of its own ground.
+
+  It never showed up as a number, because the number was fine: ground to card
+  measured 1.11 in the light and 1.115 in the dark, which is as matched as two
+  themes get. What the light has and the dark cannot is the shadow. A dark
+  shadow on a white ground is plainly visible and does half the work of lifting
+  a card; the same shadow on a near-black ground does almost nothing. So in the
+  dark the tone has to carry both jobs, and 1.11 is one job's worth.
+
+  New `--ick-dark-ground`, a shade rather than a wash of ink, which puts the
+  ground at 10 against the card's 38 and the pair at 1.32. Every ground in the
+  kit reads it — question groups, tool calls, approvals.
+
+- **A section title stands on the column its cards' words start on.** It was
+  sixteen pixels left of the numbers under it, and its chevron four pixels off
+  the pencils. New `--ick-disclosure-inset` moves a header's own content inward
+  from both ends at once; `<QuestionGroup>` sets it to `--ick-question-pad`.
+
+- **A disclosure header's box is its container's content column**, widened by
+  its own padding so only the hover wash reaches past it. It was `width: 100%`
+  pulled back on the left alone, which left the chevron six pixels shy of an
+  edge every other row in the kit sits on. Affects every header that folds —
+  `<Tool>`, `<Reasoning>`, `<ChainOfThought>`, `<TaskList>`, `<Sources>`.
+
+- **A question's folded rows are padded to the same column on both sides.** The
+  right was 8 against the left's 16, so the pencil sat half a column nearer its
+  edge than the badge sat to its own. New `--ick-question-pad` names that
+  column, and the rows, the summary and the title all read it.
+
+- **No wash under a section title on hover.** It is a rounded box the width of
+  the header and there is nothing under it for that shape to agree with, so it
+  read as a stray highlight off the card grid. The chevron lights up instead —
+  which every header now does, wash or no wash. New `--ick-disclosure-hover`,
+  set to `transparent` by `<QuestionGroup>`.
+
+### Added
+
+- `src/__tests__/alignment.test.ts` — the columns, with the browser's own
+  arithmetic done on them rather than a check that particular tokens were
+  spelled a particular way. Two rows can reach the same column through
+  different tokens and both be right.
+- A surfaces guard that composites the dark ground over the page, takes its
+  contrast against the card, and requires more of it than the light theme has —
+  because the light theme has a shadow helping and the dark does not.
+
+## 0.35.0 — 2026-08-31
+
+### Added
+
+- **`<QuestionGroup>` takes a `title`** — what the step is about, at the top of
+  the group — and `TurnPart` of kind `question` carries it through.
+
+### Changed
+
+- **The fold control is that header, and it does not move.** It was the summary
+  card itself: folded, a card at the top; expanded, a row at the bottom. One
+  control in two places, so opening the group meant one shape leaving and
+  another arriving somewhere else, and both directions read badly however the
+  animation was tuned.
+
+  A header that stays put turns the whole thing into one control with two
+  bodies — the questions, or the one row they fold into. Which is also the
+  shape `<Tool>`, `<TaskList>` and `<Sources>` already fold in, so the kit has
+  one folding idiom rather than two.
+
+  Without a title, a foldable group names its control with the count. Without
+  either, there is no header and the group is the list, as before.
+
+### Removed
+
+- `labels.hide` on `<QuestionGroup>`. It worded a pill that no longer exists —
+  the header carries the section's name in both states, so there is nothing
+  left to word.
+
+## 0.34.0 — 2026-08-31
+
+### Fixed
+
+- **Shadows are dark in the dark.** `--ick-shadow-1`, `-2`, `-3` and `-inset`
+  are built on `--ick-ink-rgb`, which is near-black in the light and
+  `245 245 245` in the dark — so every `primary` and `secondary` button was
+  casting a **white glow**. Measured: the outer layer of `shadow-2` lifted the
+  page from 18 to 40, brighter than the card it sat on.
+
+  Four `--ick-dark-shadow-*` counterparts, made of the shade, which is black in
+  both themes. The light values are untouched. A test now says every ink-based
+  shadow needs a dark counterpart and every dark one is made of the shade — it
+  found `-3` on its first run, which nothing in the kit draws with and which
+  would have been wrong the day somebody reached for it.
+
+- **`<QuestionGroup>` folds with one control instead of two.** It had a
+  full-width card when folded and a centred pill underneath the list when
+  expanded — two shapes for one job, so the fold cross-faded a button through a
+  div and neither knew where the other had been.
+
+  The same element now, kept mounted in both states, so Motion moves it rather
+  than replacing it. Label left, chevron hard right, and the chevron **turns**
+  rather than being swapped for a second glyph.
+
+- **The chevron had no gap.** It was passed as a child, so it landed inside the
+  button's label span rather than in the trailing icon slot — "Hide answers⌃"
+  with the caret against the word. There is no `<Button>` there at all now, but
+  the same mistake is worth naming: `iconRight` is the slot.
+
+## 0.33.2 — 2026-08-31
+
+### Fixed
+
+- **A badge sits in the middle of the line it belongs to.** The badge is 24
+  tall and a title line was 16, both top-aligned — which put the number **four
+  pixels** below the middle of the question, and a letter **two** below the
+  middle of its option. Measured in a browser; at four pixels it reads as wrong
+  long before anybody can name it.
+
+  The text's line box takes the badge's height, so the two centre together and
+  stay centred if the type changes — rather than a margin on the badge, which
+  is a number that has to be re-guessed every time either side moves. Field
+  rows were the only ones already right, because `.fieldLabel` had been doing
+  exactly this all along.
+
+  On the first line, deliberately: a question long enough to wrap keeps its
+  badge on the line it starts on.
+
+- **The corner stops breathing during a morph.** A `layout` animation does not
+  resize a box, it **scales** one — and a browser scaling a box scales the
+  corner with it. Measured on a group opening a card: `scaleY` ran 0.932 → 1
+  while `border-radius` stayed a flat 40px, painting a 40 × 37 ellipse that
+  eased back to a circle.
+
+  Motion has a corrector for this and it was not running: it only touches
+  values Motion is *managing*, and a radius living in a CSS class is invisible
+  to it. The number is now read off the element and handed back through
+  `style`, where Motion can see it — read rather than hard-coded, since the
+  corner is a token a host may retune. Both `<QuestionGroup>` and
+  `<QuestionCard>` do it; the card is the one that scales furthest, morphing
+  between a tall card and a 40px row.
+
+## 0.33.1 — 2026-08-31
+
+### Fixed
+
+- **The float shadow is softer in the dark.** It was black at 50% and 35% — six
+  times the light theme's alpha, on the reasoning that a shadow in the dark has
+  to darken something already dark. It overshot: measured on a rendered frame,
+  the ground under a card is `26 26 26` and the shadow's core was taking it to
+  **13**, half its brightness. That is not a card lifted off a page, it is a
+  card cut out of one, and the ring around it was the first thing you saw.
+
+  At 0.28 and 0.18 the core lands at **22** — a four-level dip, the same read
+  the light theme gets going 255 → 235. `<Tool>` and `<Approval>` take the same
+  token, so all three lift the same way again.
+
+## 0.33.0 — 2026-08-31
+
+### Added
+
+- **`<Branch>`, and regenerating no longer throws the old answer away.**
+  Re-submitting a turn overwrote `ai` and `parts` — so the answer you were
+  comparing against was gone the moment the second one started, and comparing
+  is the only reason anybody presses regenerate.
+
+  - `ChatTurn.versions: TurnVersion[]` and `versionIndex`, both optional, so a
+    turn a host built by hand still renders.
+  - `useChatTurns` returns **`showVersion(id, index)`**. Out of range is
+    ignored rather than clamped: asking for version 7 of a turn with two is a
+    bug, and quietly showing the last one hides it.
+  - `<ChatTurnRow>` takes `onShowVersion` and draws the control beside the
+    answer actions.
+
+  `<Branch>` draws **nothing** below two answers. A control reading "1 of 1"
+  offers to take you nowhere, so a turn answered once looks exactly as it did.
+
+  `ai`/`parts` are still the answer on screen and stay equal to
+  `versions[versionIndex]` through one function that both writers — the batched
+  stream flush and the turn patcher — go through. The first version of this had
+  two, and the archive stayed empty while the screen filled.
+
+## 0.32.2 — 2026-08-31
+
+### Fixed
+
+- **A tool call's sections all start on one column.** A fenced value is a
+  `CodeBlock` — its own panel, with its own inner padding. A string value was a
+  bare stack: no surface, no padding of its own, so its label and its text sat
+  flush at the body's 8px while the block's sat 12px inside the block. Two
+  sections of one tool call at two left edges.
+
+  A text section is the same panel now, and both read `--ick-code-pad`. `Tool`
+  repoints that to 8, so every label, every value and the code itself land
+  **16 from the card's edge** — the column the header's glyph is in. It was 20,
+  which is the kind of four-pixel near-miss that reads as wrong without being
+  nameable.
+
+- **A failed section is that panel tinted, not a second panel inside it.** The
+  error box sat inside the section's box, at a third inset again.
+
+### Added
+
+- `--ick-code-pad`, the inner padding of a `<CodeBlock>` — its label's left
+  edge and its code's. A box that nests one repoints it, the way it already
+  repoints the fill and the corner.
+
+## 0.32.1 — 2026-08-31
+
+### Fixed
+
+- **The composer's attachment section is back.** Making attachments survive
+  the send should not have touched how they look, and it did: the tray lost its
+  surface and its corner and became a bare padded row. It is a section at the
+  top of the composer again — its own paper, its own corner, and transparent
+  once the bubble goes glass, which is how it has looked since it held one
+  image. The thumbnail keeps its `--ick-radius-md` corner and its lift.
+
+  `--ick-attachment-radius` names that corner so it is a token rather than a
+  number somebody has to remember not to change.
+
+## 0.32.0 — 2026-08-31
+
+### Added
+
+- **`<Attachments>`, and attachments that survive the send.** The composer
+  could already pick an image and show it. It could not send it: `onSubmit`
+  took the text and nothing else, so the message went and the picture did not.
+  Offering an attach button and then losing what it attached is worse than not
+  offering one.
+
+  - `onSubmit` is `(value, attachments)`.
+  - `SendContext` carries `attachments`, so a handler gets the file and the
+    sentence together.
+  - `ChatTurn.attachments`, and `<ChatTurnRow>` hands them back to the composer
+    — the bubble *is* the message, so the same component that held them shows
+    them afterwards, read-only.
+  - A picture on its own is a message: an empty box with something attached
+    still sends.
+
+  `<Attachments>` draws them. An image with a `url` shows itself; everything
+  else gets a glyph, its name and its size, because a thumbnail of a PDF at
+  64px is a grey rectangle with a corner turned down. `onRemove` is the whole
+  difference between the composer's copy and the sent message's.
+
+- **`<ChatInput>` takes `attachments`, `onAttach`, `onRemoveAttachment`,
+  `accept` and `multiple`.** Controlled if given, its own otherwise — the rule
+  `useDisclosure` already follows. `onAttach` is for a host that wants to
+  upload first and attach the URL it gets back.
+
+### Fixed
+
+- **The object URLs are revoked.** `createObjectURL` pins the file in memory
+  until it is, and nothing was revoking: attaching and removing an image ten
+  times leaked ten of them, replacing one leaked the one it replaced, and
+  navigating away with one attached leaked that. Now on remove, on replace, and
+  on unmount — and only the ones the composer made, since a URL the host passed
+  in is the host's.
+
+## 0.31.1 — 2026-08-31
+
+### Fixed
+
+- **A citation in a sentence takes the width of its number.** It was the same
+  16px square the source list uses — right in a list, where the numbers sit in
+  a column and have to line up, and wrong in running prose. A single digit is
+  about five pixels wide, so five and a half pixels of empty badge sat either
+  side of it and whatever followed — a comma, most often — came after that gap
+  and read as detached from the number it belongs to.
+
+  Padding in `em` so it tracks the number, and no minimum: what makes it a
+  badge is a fill, a corner and the line's height, not a width. The one in the
+  list keeps its square.
+
+## 0.31.0 — 2026-08-31
+
+### Changed
+
+- **An approval is not tinted any more.** It stands on the same neutral ground
+  a question group does, because it *is* a question: "may I run this" is the
+  same species as "who else lives in the household" — the assistant asking,
+  blocking, with an answer.
+
+  The tint was there on the reasoning that this is the one thing in an answer
+  allowed to look like a box that stops you. The reasoning is real; colour was
+  the wrong instrument for it. Everywhere else in the kit the marker is a small
+  mark pointed at one thing and means **this one** — the badge on the option
+  you picked, the stroke you drew, a citation, a source's number. A wash across
+  a whole box meaning "this kind of box" is a second job for the one accent.
+
+  It also did not carry its weight: at 8% it was too weak to be an alarm and
+  too strong to be nothing, and in the dark it came out olive, which is not in
+  the palette. What says "this is unanswered" is that the box has **buttons in
+  it**, and nothing else in a turn does.
+
+  `--ick-approval-surface` is `var(--ick-ground)`. `--ick-approval-edge` is
+  gone with the ring it drew.
+
+- **Settled, nothing about the box changes** — the buttons simply leave. It
+  used to shed its background and pick up a border, because the tint was the
+  thing saying "unanswered" and had to stop saying it. A question card keeps
+  its paper and its shadow when it folds into an answered row; this does the
+  same.
+
+- **The shield rides in a badge's box**, so the title starts on the line a
+  question card's title starts on.
+
+### Breaking
+
+- `--ick-approval-edge` is removed. If you were overriding it to re-tint the
+  ring, set `box-shadow` on the component instead — or override
+  `--ick-approval-surface`, which is still the ground it stands on.
+
+## 0.30.0 — 2026-08-31
+
+### Changed
+
+- **An approval is a ground holding a card, like everything else.** It was a
+  tinted box with the tool call as the only thing on paper — so the title, the
+  description and the three buttons sat directly on the ground. A question card
+  puts its header *and* its Next button on the card; this was the same three
+  surfaces in a different arrangement, which beside one read as a different
+  object.
+
+  ```
+  ground 40 (tinted)  →  card 24  →  row 16
+  ```
+
+  The tool call on that card is now a **row** rather than a second card: inset,
+  at the row corner, no shadow. Two cards stacked is one surface more than
+  there is depth for — the same fault as two grounds, from the other end. What
+  the row holds takes the corner inside that, and its fenced value is paper,
+  the way a question's badge is paper on a grey row.
+
+  Settled, the card stops floating along with the tint: nothing is being asked
+  any more.
+
+- `--ick-approval-radius` is derived from `--ick-nest-card` rather than from
+  `--ick-tool-radius`. The tool's corner was standing in for the card's back
+  when a tool call was the only thing on the ground.
+
+## 0.29.1 — 2026-08-31
+
+### Fixed
+
+- **The card is opaque in the dark too.** `--ick-dark-card` was
+  `rgb(var(--ick-ink-rgb) / 0.09)` — a wash. On the page that is
+  indistinguishable from an opaque mix; on anything tinted it is not. A tool
+  call inside an `<Approval>` sits on the marker at eight percent, and the card
+  and the code block inside it both came out **olive** — the panel wearing its
+  ground, which is the one thing "the card is paper and opaque" exists to
+  prevent.
+
+  It held in the light, where the card is `#fff`, and was never true in the
+  dark, which is where it is easiest to see. `--ick-dark-card` is a
+  `color-mix` of paper and ink now, and `--ick-inset` is mixed **from the
+  card** rather than washed over whatever is behind it. Same colour on the
+  page; committed to rather than borrowed.
+
+  `--ick-ground` stays a wash on purpose — it is the bottom of the stack, so
+  there is nothing under it to pick up but the page.
+
+  The guard that was meant to catch this checked the **wiring** — that the card
+  points at paper in the light and at its own token in the dark — and never
+  looked at the value. It reads the value now.
+
+## 0.29.0 — 2026-08-31
+
+### Changed
+
+- **One header for everything that folds.** `Tool`, `Reasoning`,
+  `ChainOfThought`, `TaskList` and `Sources` each had a row you click to open
+  something, written out five times — with two pairs byte-for-byte identical
+  and the rest differing in ways nobody had decided on. The shimmer under the
+  label was in two of them, twenty lines of gradient each. The reveal beneath
+  was the same ten lines of `motion` props in all five.
+
+  `DisclosureHeader` and `DisclosureBody` are internals, not exports. **Nothing
+  moved on screen** — the side-by-side story is pixel-identical.
+
+  What is *not* unified is the shape, because the difference is real. A `band`
+  is a full-width row with a right edge to push the meta and the chevron to;
+  `inline` is a label that hugs its own words, for a header that sits in the
+  flow of an answer. An inline header's chevron pushed to a right edge 500px
+  away floats alone in white space with nothing beside it.
+
+- **`--ick-disclosure-*` replaces the per-component header tokens.** `-label`,
+  `-font`, `-weight`, `-size`, `-glyph`, `-meta`. A component repoints them
+  from its own root when it means something different — `Sources` and
+  `TaskList` head a list so their label is ink at medium; `Tool`'s is an
+  identifier so it is mono.
+
+  Repointed rather than overridden with a class, deliberately: two `.label`
+  rules in two stylesheets have equal specificity and which one wins is import
+  order, which is not a thing to build on.
+
+- **The stylesheet is 4.1 kB smaller** (72.81 kB, 12.48 kB gzip), and the entry
+  is 4.2 kB smaller.
+
+### Removed
+
+- `--ick-reasoning-glyph` and `--ick-chain-header`, which said what
+  `--ick-disclosure-glyph` and `--ick-disclosure-label` now say. The kit's
+  `<Reasoning>` and `<ChainOfThought>` no longer set a `.header` of their own
+  for anything to point at.
+
+## 0.28.0 — 2026-08-31
+
+### Added
+
+- **`data-active-input` on the row that owns the live composer.** Which turn
+  that is, is something `<ChatTurnRow>` knows and a host does not — the class
+  it used to set was a hashed CSS-module name nothing outside the package could
+  target, and no rule in the kit used it either, so it did nothing at all.
+
+  The case that wanted it: a page fading its conversation off the bottom edge
+  has no way to exempt the composer. This kit's composer lives at the *end of
+  the feed*, because it morphs into the message it sends — so after a long
+  answer it lands inside the gradient and comes out washed to nearly nothing.
+  It was never actually blocked; a fade is `pointer-events: none`. It just
+  looked unavailable, which is enough.
+
+  ```css
+  .feed [data-active-input] { position: relative; z-index: 6; }
+  ```
+
+## 0.27.0 — 2026-08-31
+
+### Changed
+
+- **The syntax grammars are a chunk, fetched when something needs one.**
+  `lowlight` and eleven grammars were 25 kB gzipped in everybody's bundle for a
+  thing most conversations never show.
+
+  ```
+  entry   60.10 kB gzip  →  35.05 kB gzip     −42%
+  chunk                     25.10 kB gzip     only if an answer has a fence
+  ```
+
+  A code block paints its code plain on the first paint and colours in when the
+  chunk lands. Every block after the first is coloured from its first paint,
+  because the loaded highlighter is kept. No layout shift either way: the text
+  is identical, only the colour arrives late.
+
+  Call `loadHighlighter()` at start-up if that trade is wrong for you — a docs
+  tool where every answer is code. Idempotent; concurrent callers share the one
+  fetch. `canHighlight(lang)` still answers without loading anything.
+
+  The grammars live in their own module with **static** imports, and the guard
+  that keeps them there says why: `lowlight`'s entry re-exports `all` (190
+  grammars) beside `createLowlight`, so importing the *package* dynamically
+  materialises the whole namespace and nothing can shake it back out — 301 kB
+  gzip, thirteen times what deferring it saves. The first attempt did exactly
+  that. Static named imports inside a deferred module shake normally.
+
+### Added
+
+- `loadHighlighter` and `canHighlight` are public.
+
+### Breaking
+
+- `highlightCode(code, lang)` is gone from the module's surface; it was never
+  exported from the package. The highlighter now arrives through
+  `await loadHighlighter()`, which returns the same function.
+
+- A bundler-less consumer loading the ESM directly now sees a second request
+  the first time an answer contains a fence.
+
+## 0.26.0 — 2026-08-31
+
+### Added
+
+- **`chain` and `sources` are turn parts.** `<ChainOfThought>` and `<Sources>`
+  were in Storybook and unreachable from a conversation — nothing a stream
+  could send drew them. A `SendHandler` yields
+  `{ kind: "chain", id, steps, state }` and
+  `{ kind: "sources", id, sources, title?, collapsible? }` now, and
+  `<ChatTurnRow>` draws both.
+
+- **`[^1]` in the prose is a citation.** The number is a position in the turn's
+  `sources` part, and the marker draws as an `<InlineCitation>` carrying that
+  source's title.
+
+  This is the kit's one extension to the markdown grammar, and it exists for a
+  specific reason: a citation is a component, a stream sends text. Before this,
+  `<InlineCitation>` could only be written by hand in JSX — so the component
+  existed and no actual conversation could reach it. GFM spells footnotes the
+  same way but wants a `[^1]: …` definition in the document; a model emits the
+  marker and sends its sources beside the text, never below it.
+
+  The marker is kept **out of the tokens**, so a highlight drawn across the
+  sentence — and the text a thread quotes back — do not contain a stray `[1]`.
+  A marker whose source has not arrived yet still draws, and starts working
+  when the list lands.
+
+- **`<TextHighlighter>` takes `sources` and `onSelectSource`**, which is how the
+  citations reach it. `<ChatTurnRow>` passes the turn's first `sources` part.
+
+## 0.25.1 — 2026-08-30
+
+### Fixed
+
+- **`highlight.js` is declared.** `CodeBlock` imports eleven grammars from it
+  one at a time, and the package listed only `lowlight`. It resolved because
+  `lowlight` depends on `highlight.js` and npm hoists it into reach — so under
+  a strict resolver, or the day `lowlight` picks a different highlighter, it is
+  a consumer's build breaking on an import this package wrote. It is a
+  dependency now, at the version `lowlight` already pulls.
+
+  Nothing new is installed and the bundle does not move; it was always there.
+  What changes is that it is there on purpose.
+
+### Added
+
+- **A test for what installing this costs.** The runtime dependencies, the
+  peers, and every bare import in `src` checked against both — so an import
+  cannot outrun the manifest again, and the playground's dev tooling (DialKit,
+  the perf HUD) cannot cross into the package by being convenient. It found the
+  `highlight.js` gap on its first run.
+
+## 0.25.0 — 2026-08-30
+
+### Changed
+
+- **One nesting chain for the whole kit.** A tool call and a question card are
+  the same object holding different things, and now they are built out of the
+  same four numbers.
+
+  ```
+  --ick-nest-inner   8    a badge, a chip, a fenced value
+  --ick-nest-row    16    an inset panel: an option, a field, a block of output
+  --ick-nest-card   24    paper, lifted off the ground
+  --ick-nest-ground 40    what the paper stands on
+  ```
+
+  with `--ick-nest-pad` (8) inside a card and a row, and
+  `--ick-nest-ground-pad` (16) around a card on its ground.
+
+  Each component used to derive its own. A tool call was a 6px block in a 14px
+  card on a 22px ground, beside a question's 8 / 16 / 24 / 40 — both internally
+  concentric and the two nothing like each other. The same three surfaces at
+  two scales reads as two systems, not as one object holding different things.
+
+  The reason the tool seeded a tighter chain was that a row folded shut is only
+  40px tall, and a 24px corner on it clamps to a pill. It does. So does a
+  question folded shut, which is 40px tall and carries the same 24. Both come
+  out the same shape, which was the point.
+
+- **A folded tool call is a folded question, to the pixel.** 40 tall, 16 in
+  from the left, 8 from the right, and the state glyph rides in the 24px box a
+  question's badge sits in — so the name and the title start on one line when
+  the two are stacked. There is a story for it: **Side by side → The same row,
+  twice**.
+
+- **A tool call's ground is padded like a question group's** — 16 rather than
+  8, so more of it shows around the card. Its output panels take the row corner
+  (16) rather than a 6, which is what the card's 24 and its 8 of padding imply.
+
+- **An `<Approval>` is padded like a ground**, because it is one: 16 rather than
+  12, and its corner follows to 40 — the same box a question group is.
+
+### Fixed
+
+- **The ground in the question stories was cornered by hand** — `padding: 16,
+  borderRadius: 24`, a ground padded like a ground and cornered like the card
+  standing on it. It is `card + padding`, or 40. A test now walks every story
+  for a hand-built ground, since a style object is not a stylesheet and the
+  chain cannot reach into one.
+
+### Breaking
+
+- `--ick-tool-radius` is **24** (was 14), `--ick-tool-inner-radius` **16**
+  (was 6), `--ick-tool-ground-pad` **16** (was 8), `--ick-approval-pad` **16**
+  (was 12). Every one still resolves through the same token name, so an
+  override you had set still works — but if you set one to match the old
+  numbers, it is now out of step with the chain around it. Point it at
+  `--ick-nest-*` instead, or move the seed and let all of it follow.
+
+## 0.24.1 — 2026-08-30
+
+### Changed
+
+- **A tool call gets the ground its card stands on.** The last release made it
+  a card; a card needs something to be a card *on*, and on the page alone it
+  was only paper with a shadow. It is a ground holding a card holding its input
+  and output now — the same three the question card has, in the same order.
+
+  `--ick-tool-ground` and `--ick-tool-ground-pad`, with the corner chain
+  running through both: **inset 6 → card 14 → ground 22**.
+
+  Inside an `<Approval>` the tool brings neither a ground nor a shadow, because
+  the approval already is the ground. Two grounds is one more than there is
+  depth for.
+
+## 0.24.0 — 2026-08-30
+
+### Changed
+
+- **A tool call is a card, the same object a question is.** Paper, lifted off
+  the page, with what it holds inset into it.
+
+  It was the other way round — a recessed grey strip with white panels inside —
+  which is the same three surfaces stacked backwards. On its own that reads
+  fine; beside a question card it reads as a different system. Lined up in the
+  new *Side by side* story, it was the first thing you saw.
+
+  `--ick-tool-shadow` is new, and an `<Approval>` sets it to `none`: a card
+  that already has a ground under it does not need to float off it as well.
+  Which also retires the three tokens the approval was repointing — saying a
+  tool call is a card is the tool's job now, not something done to it from
+  outside.
+
+- **A `Side by side` story.** Everything the kit draws, next to everything else
+  it draws. Not a demo: every fault worth fixing so far was found by looking at
+  two things at once, and none of them were visible in one component alone.
+
+## 0.23.1 — 2026-08-30
+
+### Changed
+
+- **Three rules removed.** A tool call drew one under its header and a code
+  block one under its label, so an open tool with an input and an output had
+  three stacked rules in a component the size of a paragraph.
+
+  The question card underlines nothing — a row is an inset panel with space
+  around it, and that is the whole of it. Two ways of saying the same thing was
+  one too many, and the surface was already saying it. A test now fails if a
+  rule goes back between stacked boxes.
+
+  A rule down the **side** is a different device and stays: `Reasoning` marks
+  an aside with one, and `ChainOfThought`'s says each step follows from the one
+  above it. Neither is separating stacked boxes.
+
+## 0.23.0 — 2026-08-30
+
+### Changed
+
+- **The kit parses markdown itself.** `unified` + `remark-parse` +
+  `remark-gfm` were 31.6 KB gzip — a third of the package — and the slowest
+  thing in it. They are gone from `dependencies`; `lowlight` is the only one
+  left.
+
+  | | before | after |
+  | --- | --- | --- |
+  | bundle | 90.8 KB gzip | **58.2 KB gzip** |
+  | parse, 1.7k chars | 1.95 ms | **0.049 ms** |
+  | parse, 5.2k chars | 5.84 ms | **0.110 ms** |
+
+  That second column is the one that mattered: parsing runs once per frame
+  while an answer streams, so 5.8 ms was a third of a frame spent re-reading an
+  answer that grew by one word.
+
+  The kit only ever used a thin slice — nine block types and nine inline ones,
+  walked straight into a flat token list — so it parses that slice in about
+  three hundred lines. **It is not a CommonMark implementation and does not
+  claim to be.**
+
+  What makes that safe: `remark` stays a **dev** dependency, and a test parses
+  a corpus of real answers through both and compares the finished documents —
+  including every prefix of one, which is what streaming actually parses and
+  where the two first disagreed. Three real differences came out of that and
+  were resolved deliberately: trailing whitespace at the end of a paragraph
+  (matched), a delimiter row that must match the header's cell count before a
+  half-typed table becomes a table (matched), and short table rows, where this
+  pads and `remark` leaves it to the renderer (kept, so a streaming table's
+  last column does not pop in and out).
+
+## 0.22.0 — 2026-08-30
+
+### Changed
+
+- **The concentric corner rule is the kit's, not the question card's.** It was
+  worked out for one component and stayed there, and the same fault turned up
+  immediately in a tool call inside an approval: the tool and the code block in
+  it both had an 8px corner at two different insets, which is the crescent the
+  rule exists to prevent.
+
+  Read inward here rather than outward, because the outer box is short — a tool
+  row folded shut is about 40px tall, and a corner derived from a generous
+  inner gap would make it a pill. So the gap is tight and the block inside
+  takes the small corner: **code 6 → tool 14 → approval 26**, gaps 8 and 12.
+
+  `<CodeBlock>` now reads `--ick-code-radius`, so a box that nests one hands it
+  a smaller corner the way `<Tool>` already handed it a different fill. The
+  guard covers every nesting in the kit — adding a box that holds a box means
+  adding a row to it.
+
+- **One stack of surfaces, named once.** `--ick-ground`, `--ick-card`,
+  `--ick-inset`: what a group sits on, an opaque panel raised on it, a row set
+  into the panel. The question card had worked this out and kept it to itself.
+
+  Which is how a tool call inside an approval came out **pale green**: the
+  tool's own surface is a translucent grey, and a translucent panel on a tinted
+  ground wears the tint. An approval is a ground now, so the tool on it is a
+  card and the code panel in that is inset — the same three steps the question
+  card uses.
+
+- **Three weights for an approval's three answers.** Filled for the narrow yes,
+  outlined for the standing one, flat for no. Two outlined buttons beside each
+  other said the last two were equals, which they are not. Left to right: Deny,
+  Always allow, Allow once — and Deny is still first in the DOM, so a keyboard
+  lands on the safe answer without tabbing past two that say yes.
+
+## 0.21.0 — 2026-08-30
+
+### Added
+
+- **`<Context>`.** How full the context window is, as a small ring.
+
+  It is a gauge, not a feature, and it earns its place for one reason: it is
+  the only honest way to explain why a long conversation starts forgetting.
+  Without it the forgetting looks like the model being stupid rather than the
+  window being full.
+
+  Which is why the warning says **what happens next** rather than only that a
+  number is high — "82%" tells somebody nothing they can act on. The whole
+  sentence is the meter's accessible name and its `title`, so it reaches a
+  pointer and a screen reader alike.
+
+  Two colours, not three: quiet until `warnAt`, then the danger colour. A gauge
+  with an amber in the middle makes somebody learn a scale to read a number
+  they can already see.
+
+  `role="meter"` with the value on it, rather than a number a screen reader has
+  to find in a sentence. A total of zero reads as empty rather than full — a
+  window nobody has reported yet is not a full one — and a host summing its own
+  tokens is clamped at 100 rather than drawn past it.
+
+  **This finishes the agent tier.** G1 through G7 are done.
+
+## 0.20.1 — 2026-08-30
+
+### Changed
+
+- **Deny moved to the left of an approval,** with the two that say yes pushed
+  right. It is first in the DOM as well, so a keyboard reaches the safe answer
+  without tabbing past the other two — the gap between it and the affirmative
+  choices is still what stops the mis-click.
+
+## 0.20.0 — 2026-08-30
+
+### Added
+
+- **`<Approval>`.** "It wants to do this. Is that all right?" — the one pattern
+  from a coding agent that generalises to any agent that acts, and the only
+  component here whose whole job is to slow somebody down for a moment.
+
+  **Three answers, not two.** "Yes" and "yes forever" are not the same answer,
+  and one button for both collects the wrong one. Allow once is the primary:
+  the narrow permission should be the easy one to give and the standing one
+  should cost a moment's thought. Deny sits at the far end and only turns red
+  under the pointer.
+
+  Decided, it stops being a set of buttons and becomes a record of what was
+  decided — live controls under a decision already made invite a second one
+  that contradicts the first.
+
+- **`kind: "approval"` on `TurnPart`,** and `onDecideApproval` on
+  `<ChatTurnRow>`. Data like every other part: the tool it names is drawn for
+  it, unrun, rather than passed in as an element.
+
+- **`variant="outline"` on `<Button>`.** The gap between `primary`, which is
+  filled, and `secondary`, which is naked until touched: a button that looks
+  like one before you reach for it.
+
+  `secondary` reading as bold text has now caught this kit twice — the empty
+  state's openers in 0.9.0, and an approval's Deny here. The second time it was
+  a permission control that did not look like a control, so the gap is filled
+  rather than worked around again.
+
+- **`TurnPartUpdate`** — a part with everything optional but `kind` and `id`,
+  which is what a stream and `updatePart` actually send. Saying a tool call
+  finished is now `{ kind: "tool", id, state: "done" }` and nothing else.
+  Repeating a field to satisfy a type is how one that was not meant to change
+  gets overwritten with whatever was easiest to type — which is exactly what
+  happened to an approval's title before this existed.
+
+## 0.19.0 — 2026-08-30
+
+### Changed
+
+- **A question's corners nest.** A box's corner is now the corner of the thing
+  inside it plus the gap between them — the rule that stops a rounded row
+  inside a rounded card leaving a crescent of card between the two curves. It
+  is the thing that reads as "not quite fitting" without anybody being able to
+  name it.
+
+  Measured, from the inside out. The badge seeds the chain at 8, a row is 8
+  plus its own 8 of padding, a card is 16 plus the 8 it puts around its rows,
+  and the group is 24 plus the 16 it puts around its cards:
+
+  | | was | is |
+  | --- | --- | --- |
+  | badge | 8 | 8 |
+  | option / field row | 8 | **16** |
+  | card | 12 | **24** |
+  | group | 24 | **40** |
+
+  Written as arithmetic rather than as four numbers —
+  `--ick-question-radius-badge` seeds it and `-row`, `-card` and `-group` are
+  `calc()` off it — so the chain cannot drift when one of the paddings changes.
+  A test does the same sum and reports the whole chain when a link breaks.
+
+  The folded group row takes a card's corner, since that is what it is. The
+  count beside it is a pill, which is outside the chain: a pill's corner is its
+  own height, not something derived from what contains it.
+
+## 0.18.0 — 2026-08-30
+
+### Added
+
+- **`<Sources>` and `<InlineCitation>`.** A numbered marker in the text and the
+  list underneath.
+
+  **The citation marker and the highlight marker are the same interaction seen
+  twice**, and this is where that pays. The kit already had a way of saying
+  *this run of words is picked out* — the marker somebody draws over an answer
+  to ask about it. A citation is that same statement made by the answer rather
+  than by the reader, so it is drawn the same way rather than in a second
+  visual language nobody has learned. Give `InlineCitation` the passage as its
+  children and the passage is marked, on every line it wraps onto.
+
+  Pressing a marker **marks** the entry rather than scrolling to it. The list
+  is already under the answer, and moving the page under somebody who clicked a
+  marker in a sentence they were reading loses them the sentence.
+
+  The list is open by default: sources are the difference between an answer
+  somebody can check and one they have to trust. An entry is a link only where
+  there is somewhere to go — a keyboard lands on every link, and one that goes
+  nowhere is a stop for nothing. Same for the marker, which is a `<button>`
+  only when it has an `onSelect`.
+
+## 0.17.0 — 2026-08-30
+
+### Added
+
+- **`<ChainOfThought>`.** How the answer was arrived at, step by step.
+
+  Three components now draw a sequence, and the line between them is the only
+  reason there are three: `Reasoning` is prose, `TaskList` is a plan — known up
+  front, fixed order, items changing state — and this is a **derivation**,
+  which grows, and where each step follows from the one above it. That last
+  word is what the line down the glyph column draws; a task list has no line
+  between its items, because a plan's items do not follow from each other.
+
+  A step's `body` takes anything, including a `<Tool>` when the step was a tool
+  call. While it thinks, the header carries the running step's label rather
+  than a count — that is the question somebody watching is asking, and the
+  reason to look at a folded chain at all.
+
+  Folds on the same terms as `Reasoning`: open while thinking, away once the
+  answer starts, and overruled for good by anybody who touches it. Four
+  components share `useDisclosure` now, and three share `StateGlyph`.
+
+  The connector came out of the first screenshot as a nine-pixel stub — the
+  rail was only as tall as its glyph, so there was nothing for the line to span.
+  It stretches to the step now.
+
+### Changed
+
+- **The shimmer guard covers all three shimmers**, not two. A third copy of the
+  same effect landed in `ChainOfThought`, which is exactly the drift the test
+  was written for.
+
+## 0.16.0 — 2026-08-30
+
+### Added
+
+- **A turn carries `parts`, and a `SendHandler` can stream them.** The agent
+  tier existed and was unreachable: `Reasoning`, `Tool`, `TaskList` and
+  `QuestionGroup` all worked, and nothing carried one into a conversation. An
+  answer was one string, so a tool call, a plan and a block of reasoning either
+  flattened into that string or never arrived.
+
+  A streamed item is now either a string — a delta of the answer's prose, as
+  before — or a `TurnPart`, merged into `turn.parts` **by its id**. The merge is
+  shallow, which is what makes streaming one bearable: send the state change on
+  its own and the text that arrived before it is still there.
+
+  `<ChatTurnRow>` draws each kind with the component that owns it, in one block
+  so the row's generous gap sits between the question and the answer — not
+  between a tool call and the sentence it produced.
+
+- **`updatePart(turnId, part)` on `useChatTurns`,** and `onAnswerQuestion` /
+  `onEditQuestion` on `<ChatTurnRow>`. A question the assistant asked is
+  answered by the person reading it, not by the stream, and that answer has to
+  land somewhere. The row never keeps it — the parts are the host's.
+
+- **`mergeParts`** is exported, for a host folding parts into state of its own.
+
+### Breaking
+
+- **`ChatTurn` now has a required `parts` field.** Turns from `useChatTurns`
+  always have it; code that builds a `ChatTurn` literal — a test, a story, a
+  host's own state — needs `parts: []`.
+
+- **`SendHandler`'s iterable widened** to `AsyncIterable<string | TurnPart>`.
+  A handler that only yields strings is unaffected.
+
+## 0.15.0 — 2026-08-30
+
+### Added
+
+- **`<TaskList>`.** What the agent means to do, what it is doing, and what it
+  has finished.
+
+  **The order never changes.** A list that sorted itself as work progressed
+  would move the line somebody is reading out from under them, and the sequence
+  is half of what the list is saying — these steps, in this order. Only the
+  glyphs change.
+
+  A list rather than a box, which is the difference from `Tool`: that is a
+  record of one thing that ran and gets a surface; this is a sequence, and a
+  surface around it would make it a thing beside the answer rather than part of
+  it. Three inks down the list — what is finished recedes, what is queued is
+  quieter still, and the one being worked on is the answer to "where is it up
+  to".
+
+  `collapsible` folds it once every task is done, on the same terms as
+  `Reasoning`: the list's preference, overruled for good by anybody who touches
+  it. The running task carries `aria-current="step"`, so a screen reader can
+  jump to it rather than counting down the list.
+
+### Breaking
+
+- **`--ick-tool-glyph`, `--ick-tool-glyph-done`, `--ick-tool-glyph-error`,
+  `--ick-tool-spinner`, `--ick-tool-spinner-track` and
+  `--ick-tool-glyph-size` are gone.** The four states are shared with
+  `TaskList` now, and so are their colours: `--ick-state-pending`,
+  `--ick-state-running`, `--ick-state-done`, `--ick-state-error`,
+  `--ick-state-track` and `--ick-state-size`. If you had restyled a tool call's
+  glyph, the new names do the same job for both components.
+
+  Two copies of one drawing is how the two start to disagree — which the
+  shimmer had already demonstrated one release earlier.
+
+## 0.14.1 — 2026-08-30
+
+### Fixed
+
+- **The shimmer stood still for two thirds of every cycle.** With the gradient
+  two and a half times the element and swept from `150%` to `-50%`, the bright
+  point travels from one element-width left of the words to two widths right of
+  them — so it is on the text for a third of the pass and off it for the rest.
+  A shimmer that spends most of its time as a static grey line is doing the
+  opposite of its job.
+
+  Sampled frame by frame over one period to confirm it rather than guess: two
+  in every three frames were the word standing still.
+
+  It is `200%` swept `110%` → `-10%` now, which puts the bright stop on the
+  left edge at one end and the right edge at the other, so it crosses the words
+  and little else. 1.8s rather than 2.4.
+
+  Both `Loader`'s shimmer and `Reasoning`'s had it, and both are fixed. A test
+  compares the two, since nothing else would make the next person fixing one of
+  them look at the other.
+
+## 0.14.0 — 2026-08-30
+
+### Added
+
+- **`<Reasoning>`.** What the model worked through before it answered. Open
+  while it thinks, folded away once the answer starts — the one detail every
+  kit shipping this has converged on, and it is right: thinking is worth
+  watching while it happens and worth almost nothing afterwards, but it has to
+  stay reachable, because the times it matters are exactly the times the answer
+  looks wrong.
+
+  Folding is the block's preference, not something done to the reader. Open it
+  and it stays open, however many times the state changes underneath.
+
+  It times itself when no `duration` is given — adjusted during render on a
+  change of state rather than in an effect, which would mean a second pass
+  every time the thinking stopped.
+
+  Prose, not a panel, which is the difference between this and `Tool`: a tool
+  call is a record of something that ran and gets a box; this is the model
+  talking to itself, and a box would give it a weight it has not earned next to
+  the answer it is only explaining. A rule down the side says "aside" and stays
+  out of the way.
+
+  The shimmering word is a real word in the button rather than a `<Loader>`.
+  The loader is decorative and marks itself `aria-hidden`; drawn that way the
+  button had no accessible name at all. Caught by the tests before it shipped,
+  and three axe cases keep it that way.
+
+### Changed
+
+- **`formatDuration` moved out of `Tool`** into its own module, and both it and
+  `Reasoning` read it from there.
+
+- **`useDisclosure`** holds the open/shut rule both blocks share: the host if it
+  is controlling the row, then whoever clicked it, then the row's own
+  preference. Derived rather than an effect, which is what lets a reader's
+  decision outlive every state change after it.
+
+## 0.13.0 — 2026-08-29
+
+### Added
+
+- **`questionAlign` on `<ChatTurnRow>`.** `end` by default — the composer is
+  about to become the reader's own bubble, and those sit right. `stretch` fills
+  the row instead, which is what an *opening* composer wants: on an empty
+  conversation it is not a message on its way, it is the box under the openers,
+  and a pill floating at the right edge of a centred block reads as unrelated
+  to the block.
+
+  A flex child sizes to its content, so the input is told to fill as well —
+  stretching the row on its own does nothing.
+
+### Changed
+
+- **The badge on a chosen option is the accent taken down towards ink.** At
+  full strength it was a shade off the tint under it: two yellows that close
+  are not a badge on a row, they are one yellow with a rounded hole in it.
+  Paper was tried first and was no better — pale on pale.
+
+  `--ick-marker-deep` mixes the marker with ink rather than stating a colour
+  per theme, which is the point: ink is near-black in the light and near-white
+  in the dark, so it comes out darker than the row in one and lighter in the
+  other. One requirement — separate from the row — answered correctly twice.
+
+  Measured rather than eyeballed. 55% marker put the letter at 3.5:1; 45% is
+  the first that clears 4.5. It ships at 40%: the letter reads 5.6:1 in the
+  light and 16.3:1 in the dark, and the badge 5.1:1 and 4.0:1 against the row
+  it sits on.
+
+## 0.12.1 — 2026-08-29
+
+### Fixed
+
+- **Picking an option no longer draws a focus ring around it.** The option row
+  rang on `:focus-within`, and a mouse click focuses a button — so choosing one
+  outlined it, which is the exact thing `:focus-visible` exists to avoid.
+  Measured: after a real click, `:focus-visible` was `false` and
+  `:focus-within` `true`, with a 2px outline computed.
+
+  `:focus-within` was there for the "something else" row, which is a label
+  around an input and so has no `:focus-visible` of its own. It is scoped to
+  that row now. A keyboard still rings every row, and the row still lights up
+  when its input is focused.
+
+## 0.12.0 — 2026-08-29
+
+### Added
+
+- **`<Tool>`.** One tool call: what was run, what with, what came back, and
+  which of four states it is in — queued, running, done, failed. Shut by
+  default, because most of the time nobody cares.
+
+  **It opens itself when it failed**, since an error nobody can see has not
+  been reported. Derived from the state rather than forced by an effect, so a
+  call that fails later opens while one somebody deliberately shut stays shut —
+  reopening a row under a reader's hands to show them something they dismissed
+  is not help.
+
+  What it was given decides how it is drawn. A string is text: wrapping
+  `"Belgrade, 24°C"` in a fence puts it in quotes with its newlines spelled
+  out, which is worse than reading it. An object is JSON, in a `CodeBlock` with
+  its copy button, which is most of why anybody opens the row. An element is
+  somebody having decided for themselves, so it is left alone. A value JSON
+  cannot hold prints something unhelpful rather than throwing.
+
+  The state is not carried by colour: the glyph changes shape, and the row says
+  which state it is in in words that only a screen reader hears.
+
+- **`--ick-danger`.** The first hue past ink, paper and the marker. The marker
+  already means "this one", so a failed call wearing the accent for choice
+  would say the wrong thing. A channel triplet like the rest, lifted in the
+  dark, where the light red was nearly black on a near-black page.
+
+### Changed
+
+- **`--ick-tool-code-fill`** points a fenced value inside a tool row at paper
+  rather than at the sunken surface the row itself uses — otherwise the two are
+  one grey with a hairline through them. Lighter than the row in the light,
+  darker in the dark; different in both, which is the requirement.
+
+## 0.11.1 — 2026-08-29
+
+### Fixed
+
+- **Every badge in a question sits on one vertical line.** The number over a
+  question was 12px from the card's edge and the a/b/c beside its rows were
+  16, so the two did not line up. The rows already answered and the ones still
+  to come were at 12 as well. All four are at 16 now.
+
+- **A field row's focus edge stopped taking up room.** It was a 1px border,
+  and a border is part of the box — which put every field row's badge a pixel
+  to the right of every option row's. It is an inset shadow now: the same
+  edge, drawn in no space.
+
+  A test does the sum the browser does — jsdom lays nothing out — and fails if
+  any of the four rules drifts again.
+
+### Changed
+
+- **A question still to come shows its number rather than an asterisk.** It is
+  question three whether or not anybody has reached it, and saying so is what
+  makes the list read as a list. White, since an upcoming row has no card
+  under it: the badge is what stands off the group's own surface.
+
+## 0.11.0 — 2026-08-29
+
+### Added
+
+- **The parts a question is built from are public.** `QuestionShell`,
+  `QuestionOptionRow`, `QuestionFieldRow`, `QuestionOtherRow` and
+  `QuestionBadge`.
+
+  The kit ships three question shapes, and three is not all of them. A fourth —
+  pick one *and* say something, say — is now a composition rather than a fork,
+  and it arrives already wearing the same tokens, focus behaviour and ARIA as
+  the three that ship. Exporting the rows without the shell would have been
+  half of it: the box, the header and the footer would still have been
+  hand-rolled out of numbers that were tokens ten minutes earlier.
+
+  Being public API changed them. Each takes a `className` that is added to its
+  own rather than replacing it, spreads the rest of its props onto the element
+  it ends in, and forwards its ref to the thing worth having one for — the
+  input, in the two rows that have one. `letter` is optional everywhere. And
+  they are named for what they are from outside rather than for where they sat
+  inside the card.
+
+  `QuestionShell` paints the card by default, since a shell standing on its own
+  is the reason it is exported. `QuestionCard` passes `card={false}` — the box
+  that morphs between the three states is its own, and two would nest. The card
+  is now built out of the shell rather than repeating it, so the shipped
+  component is the proof the exported one works.
+
+### Fixed
+
+- **Two field rows with the same label no longer share an id.** The input's id
+  was built out of the letter and the label, which is unique inside one card
+  and not across two — and two inputs holding one id means the second row's
+  label focuses the first row's input. It is `useId()` now, and an `id` of your
+  own still wins.
+
+- **The option row's tap no longer scales under reduced motion.** A 1% squash,
+  but it was the one animation in the question rows that was not asking.
+
+## 0.10.1 — 2026-08-29
+
+### Fixed
+
+- **A question card was clipping its own contents.** `.active` had padding and
+  a 100% width but no `box-sizing`, so the padding was added *outside* the
+  width and every option row and the Next button lost 16px to the card's
+  `overflow: hidden`. Measured: rows 488px wide inside a 488px card, starting
+  8px in.
+
+  This package has no global `box-sizing` reset — one written inside a CSS
+  module leaks into the host's page — so every box with padding has to say it
+  itself. A test now reads the rules and fails if one stops.
+
+- **The card is white.** `--ick-question-card` was `--ick-surface-raised`, a 5%
+  mix, which put it within three percent of the ground it sits on: the same
+  surface in a slightly different shade rather than a card. It is `--ick-surface`
+  now, with the ground moved down to meet it. Dark keeps a raised value, since
+  `--ick-surface` there is the page itself and a card painted with it would
+  sink into the ground instead of sitting on it.
+
+## 0.10.0 — 2026-08-29
+
+### Added
+
+- **`<QuestionCard>`, `<QuestionGroup>` and `<Chip>`.** A structured question
+  inside a conversation — the assistant asks something with a shape to it, and
+  the answer is picked or typed rather than written out. Three question shapes
+  (`inputs`, `single`, `multi`), three card states that morph between each
+  other, and a group that folds to one summary row once it is done with.
+
+  Brought across from another project and rebuilt on this kit's tokens, type
+  and buttons. The accent doing the work is the marker yellow, on the badge of
+  whatever is chosen and on the border of the field being typed into.
+
+  Two behaviours kept because they were right: a single-select waits a beat
+  before committing, or the card is gone before anyone sees what they picked;
+  and the "something else" row is a `<label>`, not a button, because an input
+  inside a button is not reliably focusable.
+
+  Two changed. The collapsed row is one `<button>` naming what it does rather
+  than a click handler on a `<div>` with another button inside it — clickable
+  but not reachable by keyboard, and two tab stops for one action. And the
+  letter badges are `aria-hidden`: the letter is a visual index, and in the
+  tree it turns a field called "Their name" into "a Their name".
+
+- **Question and chip tokens.** `--ick-question-*` for the three stacked
+  surfaces, `--ick-badge-*`, `--ick-chip-*`.
+
+## 0.9.0 — 2026-08-29
+
+### Added
+
+- **`<EmptyState>`.** What is on screen before anybody has asked: an icon, a
+  title, a description and a row of openers, each optional and none drawn in
+  place of what is left out. Openers need an `onSuggestion` — a chip that
+  reports nowhere is a button that does nothing.
+
+  `title` renders as text rather than a heading. It sits inside a conversation
+  the host already owns, and claiming a level in their document is not ours.
+
+- **`<Loader>`.** `dots` for the gap between sending and the first word,
+  `shimmer` for words standing in for something.
+
+  Silent by default: `useChatTurns` already announces that a response is
+  coming, and a second live region means hearing it twice. `label` opts in.
+
+- **`ChatTurnRow` shows the loader** between the question being sent and the
+  first word landing. A sent question with a blank space under it reads as
+  nothing having happened.
+
+- **Two tokens**, `--ick-loader-dot` and `--ick-empty-pad`.
+
+## 0.8.0 — 2026-08-29
+
+### Added
+
+- **`<AnswerActions>`.** Copy, regenerate and a verdict, under a settled
+  answer. The input has had a hover row since the beginning and the answer had
+  nothing, which is backwards — the answer is the part worth keeping.
+
+  Only what has somewhere to report is drawn. A control that calls nothing
+  looks like a feature and behaves like a dead end. Pressing the verdict
+  already given reports `null`, which is how somebody takes it back.
+
+  Copy confirms and announces itself, for the same reason the code block's
+  does: a tick is a picture, and a reader who cannot see it is told nothing
+  happened.
+
+  `reveal` makes the row invisible until hovered or focused — invisible, not
+  absent: it keeps its space so nothing shifts, still hit-tests at zero
+  opacity, and `:focus-within` brings it back for anyone arriving by keyboard.
+
+- **`ChatTurnRow` renders it**, with `onRegenerate`, `onFeedback`, `feedback`
+  and `answerActions`. They appear when the answer settles rather than while it
+  arrives.
+
+## 0.7.1 — 2026-08-28
+
+### Added
+
+- **`<Conversation anchorId>` and `anchorOffset`.** Hold an element at the top
+  of the view instead of following the end of the content.
+
+  0.7.0 only did the second, which is wrong for the way this kit works: a
+  submitted message has to go to the top and stay there while the answer
+  arrives underneath, so what is on screen is the question and its answer.
+  Following the end stacks everything downwards instead, and the composer that
+  appears when the answer settles ends up past the fold.
+
+  Point it at the turn that was just sent. It needs room to scroll into — an
+  element cannot be brought to the top of a container that ends just below it —
+  which is what a large bottom padding on the viewport is for.
+
+  With an anchor above the reader, scrolling *up* is how they return to it, so
+  letting go is decided by distance rather than direction.
+
+## 0.7.0 — 2026-08-28
+
+### Added
+
+- **`<Conversation>`.** The scroll container: it keeps up with an answer as it
+  arrives, lets go the instant the reader scrolls away, and offers a button
+  back. Three tokens for the gap and padding.
+
+  It follows the **end of the content**, not the bottom of the container.
+  Those are only the same thing when nothing is padded below, and this kit's
+  demo carries a screen-height pad so a turn can be pulled to the top —
+  scrolling to the bottom there parks the answer above the fold in front of a
+  blank screen. Measuring the content makes one behaviour right for both.
+
+  It reads intent from the **input** rather than the scroll event. A component
+  watching scrolling cannot tell its own from the reader's, and ends up either
+  dragging them back down mid-sentence or never following at all. A wheel
+  upwards, `PageUp`, `Home`, or a drag away from the end, and it stops.
+
+  `ref` is forwarded to the viewport rather than the root, because a ref here
+  is for scrolling and the root does not scroll. `viewportClassName` styles the
+  scroller; `className` styles the box you lay out.
+
+## 0.6.0 — 2026-08-28
+
+### Added
+
+- **`<CodeBlock>`.** The language, a copy button that confirms and announces
+  itself, and code that scrolls sideways rather than widening the answer. The
+  markdown renderer uses it for every fence; it is exported for use on its own.
+
+  Ten languages, chosen and measured: `lowlight/common` is 37 grammars and
+  51.6 KB gzipped — nearly the size of everything else here — for languages a
+  chat will almost never show. These cost a quarter of that. A language outside
+  the list renders unhighlighted rather than throwing.
+
+  The scheme is ink at four weights rather than a syntax palette, because this
+  kit is ink, paper and one acid yellow. Six tokens make it a palette for
+  anyone who wants one.
+
+- **Code tokens.** `--ick-code-size`, `-leading`, `-comment`, `-keyword`,
+  `-string`, `-number`, `-name`, `-attr`, `-addition`, `-deletion`.
+
+### Fixed
+
+- **A code block keeps its own font.** `pre` and `code` are two of the elements
+  a host app or docs tool styles without thinking about it, and an inherited
+  family loses to any direct declaration — so the family is stated on the
+  elements themselves. Ligatures are off with it: `--ick-font-mono` is a token,
+  and a brand setting Fira Code would otherwise get `npm run dev -- --port`
+  drawn as `dev —— ——port` while the clipboard hands over the real characters.
+
+### Dependencies
+
+- `lowlight` and ten `highlight.js` grammars. 56.6 KB → 81.6 KB gzip.
+
+## 0.5.1 — 2026-08-28
+
+### Fixed
+
+- **`<ChatHeader variant="bordered">` is opaque.** It was transparent with a
+  hairline under it, so a sticky header had the conversation scrolling visibly
+  through its own title. It now sits on `--ick-surface`, which is also what
+  separates it from `glass`: a panel with a line under it, rather than a
+  frosted pane.
+
+  Found by fixing the Storybook story rather than the component. The three
+  materials were being shown on a blank page with nothing under them, where
+  all three look identical — so the story now scrolls real content beneath a
+  sticky header, and the bug was in the first frame of it.
+
+  Three tests read the rules out of the stylesheet, since jsdom does not paint
+  and this is the kind of thing that comes back once nobody is looking.
+
+## 0.5.0 — 2026-08-28
+
+### Breaking
+
+- **`TextHighlighter` renders its text as markdown.** An answer containing
+  `*`, `#`, `` ` `` or `-` at the start of a line now renders as emphasis, a
+  heading, code or a list rather than as those characters. If you were passing
+  markdown and relying on it staying literal, escape it.
+
+  Raw HTML in the input is dropped rather than rendered. Model output is
+  untrusted text, and there is no version of injecting it into the host's page
+  that is worth the surface it opens.
+
+### Added
+
+- **Markdown, inside the highlighter rather than around it.** Headings,
+  emphasis, strikethrough, links, inline and fenced code, lists, blockquotes,
+  tables, images and rules.
+
+  The design constraint was the marker. Internally the words stay a **flat
+  array of tokens addressed by index** — hit-testing reads an index off the
+  span under the pointer, the keyboard cursor walks the indices, a highlight's
+  text is a run of them joined back. Markdown only decides which element each
+  token is drawn inside, so a stroke that starts in plain text and ends inside
+  `**bold**` is one run of indices like any other.
+
+  Fenced blocks are the exception: preformatted, so not tokenised, so not
+  markable. `CodeBlock` will own them properly.
+
+- **Markdown tokens.** `--ick-md-gap`, `-item-gap`, `-heading-space`,
+  `-indent`, `-quote-indent`, `-rule` and `-code-fill`. The gaps are in `em`
+  so they scale with the answer rather than drifting away from it.
+
+### Dependencies
+
+- `unified`, `remark-parse` and `remark-gfm`, bundled rather than left to the
+  consumer. The package went from 18.9 KB to 56.6 KB gzip; that is the price
+  of markdown working on install with nothing to configure.
+
+  `react-markdown` was the obvious choice and is not used. Its feature is
+  swapping *components*, and what this needs is every **text node** split into
+  indexed spans — text nodes are strings, not components. It also parses
+  inside its own render, which hands back the tokens a render too late for the
+  keyboard cursor that needs them.
+
+## 0.4.0 — 2026-08-27
+
+Additive.
+
+### Added
+
+- **`<ChatTurnRow>`.** One turn: the question as a composer that has become a
+  bubble, and the answer under it. It existed all along in the playground,
+  which meant anyone installing the package had to rewrite the one thing the
+  package is about — the README could only tell them to, and explain the memo
+  they would need.
+
+  Not named `Message`. The user half is a live input that morphs into its own
+  bubble, not a rendered record of what was typed, and that is the whole idea.
+
+  Memoised, and the memo is load-bearing: `useChatTurns` leaves untouched turns
+  referentially identical, which only pays off if the rows act on it. Every
+  callback takes the turn's id rather than being closed over per row, so a
+  consumer can hoist them and not hand the memo a new prop each render.
+
+- **Turn tokens.** `--ick-turn-gap`, and `--ick-answer-size` / `-leading` /
+  `-tracking` / `-measure` for the answer. The measure is in `ch` rather than
+  pixels, so a comfortable line length follows whatever font a brand sets.
+
+## 0.3.0 — 2026-08-27
+
+Additive. Nothing in 0.2.0 changes behaviour.
+
+### Added
+
+- **`<ChatHeader>`.** The chrome above a conversation: title, subtitle, avatar,
+  back, and a row of actions. Three materials (`plain`, `glass`, `bordered`),
+  three sizes (40 / 48 / 56px), and a centred arrangement for the native/mobile
+  pattern.
+
+  Actions are described rather than passed as children:
+
+  ```tsx
+  actions={[{ id: "share", label: "Share", icon: <Share2 size={16} />, onClick: share }]}
+  ```
+
+  which is what lets `collapseActionsAt` fold them into a keyboard-navigable
+  overflow menu when the header is narrow — a header cannot summarise children
+  it cannot read. Anything with no icon-and-label shape goes in as `children`
+  and stays put.
+
+  `label` is required, and a `count` is folded into it: the badge on the glyph
+  is `aria-hidden`, so without that a reader would never learn the number.
+
+- **Header tokens.** `--ick-header-*` for height, padding, gap, backdrop, blur,
+  border, title, subtitle and the count badge — twelve, plus one that restates
+  the backdrop for the dark theme, where a lighter wash still shows the text
+  scrolling underneath.
+
+## 0.2.0 — 2026-08-26
+
+The package became something other people can install: it has tests, it can be
+themed, and it can be used without a mouse.
+
+### Breaking
+
+- **`ReplyThreadPopup` requires `onSendMessage`.** It used to be optional, and
+  without it the component streamed placeholder copy about the Higgs boson —
+  shipped to every consumer. A component has no answers of its own; that is the
+  host app's job.
+  ```tsx
+  <ReplyThreadPopup
+    onSendMessage={(message, quotedText, { signal }) => callYourApi(message, quotedText, signal)}
+    …
+  />
+  ```
+- **`ChatInput` has no `variant` prop.** There were four; three had no call
+  sites. Remove the prop.
+- **Components read `--ick-` custom properties.** The old unprefixed names
+  (`--ink`, `--surface-hover`, `--font-geist-sans`, …) still resolve, as
+  aliases, but they will go. A host that loads its own fonts should say so:
+  ```css
+  :root {
+    --ick-font-sans: var(--font-geist-sans);
+    --ick-font-mono: var(--font-geist-mono);
+  }
+  ```
+- **The kit now responds to a `.dark` class or `data-theme="dark"` on the root
+  element,** and to `prefers-color-scheme` when neither is set. An app that
+  already uses `.dark` for its own theme will find the kit following it. Pin it
+  with `data-theme="light"` if that is not wanted.
+- `GlassButton` is deprecated in favour of `<Button variant="glass">`. It still
+  works and renders identically — the wrapper translates the old size names.
+
+### Added
+
+- `useChatTurns` — turn state, streaming and the reveal, in one hook. Takes an
+  `onSend` returning a string, a promise or an async iterable, and hands back
+  an `AbortSignal`.
+- A design token layer: three tiers, all prefixed `--ick-`, with a dark theme.
+  See [theming.md](./theming.md).
+- Dark mode, following `prefers-color-scheme` unless told otherwise.
+- Keyboard operation throughout — a word cursor for making highlights, reachable
+  controls for the ones already made, an escapable menu, and a real dialog.
+- Screen-reader announcements when an answer arrives, and `prefers-reduced-motion`
+  honoured across the kit.
+- `announce()` — the shared live region, exported so a host can speak its own
+  errors into it rather than adding a competing one.
+
+### Fixed
+
+- **A finger could not scroll past an answer.** `touch-action: none` on every
+  paragraph made the conversation a dead zone on a phone. Related: a scroll the
+  browser took over was committed as a highlight, leaving a marker behind on
+  whatever word the finger landed on.
+- **The package was styling the host's document.** `ReplyThreadPopup` rendered
+  a `<style>` element with an unscoped `div::-webkit-scrollbar` rule, which hid
+  the scrollbar on every div on the page for as long as a thread was open.
+- **`GlassButton` had no accessible name while loading.** The label was hidden
+  with `visibility: hidden`, which removes it from the accessibility tree, and
+  the spinner is `aria-hidden` — so it announced as a button called nothing.
+- `ChatInput` left a timer running past unmount.
+- Three performance passes, each measured: pointer crossings over a highlight
+  went from ~28 DOM mutations to 0, opening the menu from 407 to 67, and
+  streaming an answer stopped touching the turns above it at all (366 to 0).
+
+### Internal
+
+- 167 tests and CI: lint, tests, build, a Storybook build and a dry-run pack.
+- Storybook, with a token table read live from the stylesheets.
+- Two guards that fail rather than rely on memory: one when a component is
+  exported without a story, one when a literal colour appears outside the token
+  file.
+
+## 0.1.0
+
+Extracted from the website into a package.