#69 added `useAppContext()` as a Vue-native rename of Ink's `useApp()`,
qualified to avoid reading as the Vue application instance. On reflection the
"Context" suffix borrowed the name of an internal grab-bag context and slightly
mislabels the hook — it returns app lifecycle controls, not that context. The
collision worry doesn't hold up: Vue has no `useApp()`, the returned
`{ exit, waitUntilRenderFlush }` is clearly not the Vue app instance, and "App"
in vue-tui already means the `TuiApp` from `createApp()`.
Rename to `useApp()` for full Ink fidelity (same name + same shape), and drop
the now-defunct "App composable" entry from ink-divergences.md — it ceases to
be a divergence.
Internal context cleanup (the grab-bag `AppContext` + the `StdinContext`
duplication) is intentionally out of scope here, tracked separately.
BREAKING CHANGE: `useAppContext()` is renamed to `useApp()`. Replace
`const { exit } = useAppContext()` with `const { exit } = useApp()`.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
The "Accessibility props — AriaRole / AriaState" entry claimed "Ink: no
equivalent" and filed aria/screen-reader support under additive features. That
is incorrect: Ink has full accessibility support at this doc's own baseline
(v7.0.4, commit 40b3a75) — aria-label/aria-hidden/aria-role/aria-state on
Box/Text, useIsScreenReaderEnabled, and screen-reader linearization in
render-node-to-output. vue-tui's implementation is a faithful port of it; the
parity commits G03–G59 align the linearization with Ink (e.g. G22's parent-role
dedup matches Ink's `role !== parentRole`).
Accessibility is parity-by-default, so it does not belong in a doc that records
only intentional divergences. Remove the entry; the feature itself is unchanged.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Ink's `useApp()` returns `{ exit, waitUntilRenderFlush }`. vue-tui previously
exposed only `exit()` via `useExit()` and kept `waitUntilRenderFlush` on the
`TuiApp` handle alone. Align the public surface with Ink: add `useAppContext()`
returning the same pair, and remove `useExit()`.
- thread `waitUntilRenderFlush` into the injected `AppContext` via a hoisted
impl shared by the `TuiApp` handle and the composable, so both resolve
identically
- add `useAppContext()`; delete `useExit()`; migrate all call sites, PTY
fixtures, examples, READMEs and the public-API surface test
- port Ink's two "useApp waitUntilRenderFlush" tests; Ink's third relies on
React concurrent mode (N/A in Vue)
- rewrite the ink-divergences entry: this is now a *naming* divergence
(`useAppContext` vs `useApp`, mirroring `createApp` vs `render`), not a
surface one — and fix the prior wrong claim that Ink's `useApp` returns
stdin/stdout/stderr
The name is qualified (`useAppContext`, not `useApp`) so it doesn't read as the
Vue application instance (`createApp`/`app.mount`) — the same Vue-native naming
choice vue-tui already makes with `createApp()` vs Ink's `render()`.
BREAKING CHANGE: `useExit()` is removed. Replace `const exit = useExit()` with
`const { exit } = useAppContext()`.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
vue-tui withheld Ink's named prop types (BoxProps, TextProps, StaticProps,
TransformProps, NewlineProps) and the WindowSize/CursorPosition data shapes
under a blanket "avoid React-shaped type names" rule. That rule over-reached:
a <Box> has props in Vue exactly as in React, so those names carry no
React-vs-Vue content — there's no reason to rename them. Re-export them under
Ink's names so a consumer can name a component's props the same way as in Ink.
- Derive each XProps from the component's runtime `props` object via Vue's
`ExtractPublicPropTypes`, so the public type can never drift from the real
props. Pin `required: true as const` on Static.items / Transform.transform:
a standalone `const` widens `true`→`boolean`, which would otherwise drop them
from the required keys — in both the exported type AND the component's own
`setup(props)` typing.
- Add `WindowSize { columns, rows }` and `CursorPosition { x, y }`, anchored to
their real usage in useTerminalSize / useCursor. (The composables still return
reactive refs of these — the data shape matches Ink; the ref wrapper is the
framework difference.)
- Keep the genuinely-divergent names as-is: DOMElement→TuiNode (real
DOM-emulation vs host-node difference), RenderOptions/Instance→MountOptions/
TuiApp (downstream of createApp()), App/Stdin/Stdout/StderrProps = N/A. Rewrite
the ink-divergences doc to record what's now aligned vs still divergent.
- Add a tsc-checked type-level test (public-types.test-d.ts) asserting the
exported shapes; it is excluded from vitest's runtime run by naming.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Ink keeps its `measure-text` module internal and never re-exports it. vue-tui
exported `measureText` / `measureTextNatural` from the public index under the
mistaken belief — stated verbatim in commit 0e7d775's own message — that doing
so "matched Ink's public API". It does not; Ink keeps that module internal. A
later design doc then rationalized the leak post-hoc as an intentional
divergence. It was neither intentional nor a divergence — it was a mistake.
Align with Ink:
- Drop both from the public index. `measureTextNatural` stays as an internal
helper (yoga.ts uses it). `measureText` had zero production callers (yoga uses
`wrapText` + `measureTextNatural`, never `measureText`) and is removed.
- Integration tests that used `measureText(stripAnsi(x), 9999).width` as a
line-width helper now use `stringWidth(stripAnsi(x))` directly.
- public-api.test.ts gains a regression test asserting neither is exported.
- Remove the now-obsolete entry from .agents/docs/ink-divergences.md.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* chore: remove Ink-parity design docs (.agents/docs), keep the code
Removes the Ink-parity loop documentation — ink-parity-loop.md, ink-parity.md,
and the parity-ledger.md — and drops the now-dangling 'Current docs:' list from
AGENTS.md's Context Engineering section (the convention itself is kept). All the
merged parity CODE fixes remain on main; only the documentation/ledger artifacts
are cleared, to re-orchestrate the documentation with a different approach.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* docs: add focused Ink intentional-divergences design doc
Replaces the removed sprawling parity docs with one focused doc that records
ONLY where vue-tui deliberately differs from Ink (API surface, additive
features, unavoidable Vue-vs-React semantics, N/A React concepts, framework
idioms) — leaving placeholders for the maintainer to supplement.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
13 mediums fixed+merged across sweeps 1-10 (G32, G33, G39, G44, G45, G46,
G52, G58, G59, G63, G64, G68) plus the ~22 earlier gaps. Maintainer paused
the open-ended loop at diminishing returns: latest sweeps surface byte-level/
edge-case divergences (G68 ANSI order is visually identical; G67 refuted as an
unavoidable Vue-vs-React semantic). Open mediums G69/G70 and the LOW tail are
recorded but deferred by decision.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Ink's Text.tsx transform applies each enabled style as a SEPARATE nested
chalk wrap, in the exact order dim -> color -> backgroundColor -> bold ->
italic -> underline -> strikethrough -> inverse. vue-tui's applyChalk built
ONE chained ChalkInstance (color -> bg -> dim -> bold -> ...) and invoked it
once, producing a different, non-Ink byte sequence for any multi-style Text:
e.g. color+bold emitted [31m[1mX[22m[39m vs Ink's [1m[31mX[39m[22m, and
dim+bold dropped the bold re-open after dim's SGR-22 reset.
Rewrite applyChalk to mirror Ink: apply each style as its own nested chalk(...)
call in Ink's order, reusing the existing color/background resolution. Chalk
level handling (FORCE_COLOR / level 0 -> no codes) is preserved and ANSI codes
remain zero-width, so styled-text measurement is unchanged. Multi-style Text now
produces byte-identical ANSI to Ink.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Ink content-sizes the position:absolute, auto-width static box: Static.tsx
sets `{position:'absolute', flexDirection:'column', ...customStyle}` with no
width, ink.tsx calculateLayout never sets the static node's width, and
renderer.ts reads node.staticNode.yogaNode.getComputedWidth() — the computed
width of a yoga absolute, auto-width node, which shrinks to its CONTENT. So
flex-fill children (Spacer/flexGrow/justifyContent/percent) inside a Static
item collapse to content width instead of expanding to the terminal width.
The G44 fix over-forced the iso root to full terminal width (setWidth(columns)),
so a Spacer/flexGrow child expanded to fill the terminal. This refines G44: for
an AUTO-width static node we now leave the iso root width auto and cap it at the
available columns with setMaxWidth(columns), content-sizing exactly like Ink
while still wrapping anything wider than the terminal. The explicit-width
(POINT/PERCENT) copyStyle path from G44 is unchanged and still wins.
Verified against the built Ink reference (v7.0.4, 40b3a75): a Static item Box
row [LEFT][Spacer][RIGHT] at cols=80 renders "LEFTRIGHT" (Spacer collapses),
and [A][flexGrow][B] renders "AB".
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Ink's output.ts maps sliceAnsi(line, from, to) (the horizontal clip) over the
lines BEFORE the lines.entries() loop that runs transformer(line, index), and it
never re-clips the transformer's output. vue-tui's paint write() did the reverse
(commit 2c99431 deliberately moved clip AFTER transform), so a width-sensitive
transform inside an overflowX:"hidden" box received the FULL line and had its
result sliced — corrupting gradients (wrong char count) and OSC-8 hyperlinks
(closing sequence sliced off).
Reorder so the per-line horizontal clip runs first, then the transformers apply
to the already-clipped span. The post-vertical-clip line index passed to each
transformer is unchanged, so transform index/nesting behavior (G21/G32/G52/G58)
is untouched. The pre-existing overflow test that asserted the old re-clip order
(transform-returns-wide-char dropped at the boundary) is rewritten to Ink's
clip-then-transform output (the widened glyph overflows), verified against the
pinned Ink reference (7.0.4 / 40b3a75) via renderToString.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Ink's onRender screen-reader branch (ink.tsx:573-625) writes the wrapped SR
transcript with a RAW stdout.write using manual ansiEscapes.eraseLines(prev) +
(inline static, if any) + the wrapped output, sets lastOutput/lastOutputToRender/
lastOutputHeight, and RETURNS before the normal interactive frame path. It emits
NO clearTerminal, does NOT accumulate/replay fullStaticOutput, does NOT go
through log-update, and does NOT hide the cursor.
vue-tui routed SR frames through renderInteractiveFrame, so a tall/overflowing
SR transcript (outputHeight >= viewportRows, then previousOutputHeight >
viewportRows) hit the clearTerminal branch — wiping the SR user's scrollback,
replaying accumulated fullStaticOutput, and the mount-time hide left the cursor
hidden.
This adds a dedicated `if (isScreenReaderEnabled) { ... return; }` branch in
commit(), before fullStaticOutput accumulation (now gated off for SR) and before
renderInteractiveFrame, mirroring Ink's SR branch byte-for-byte (eraseLines +
inline static + wrapped output, lastOutputToRender = wrapped output with no
trailing "\n", height = split count). It never clears the terminal, never
replays static, never uses the log-update writer, and the mount-time cursor-hide
is now skipped for SR mode. The normal (non-SR) interactive path is unchanged —
clearTerminal-on-tall-frame still applies there. G17/G46 SR behavior preserved.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
In Ink, <Transform> IS an ink-text host node (Transform.tsx renders
<ink-text internal_transform={fn}>), and the reconciler sets isInsideText
for ink-text. So bare-string and <Newline> children of a standalone
<Transform> render inline within that ink-text, with the transform applied
per line by the Output, and a nested <Text> child is squashed inline too.
The canonical README pattern `<Transform transform={fn}>Hello World</Transform>`
(no inner <Text>) works standalone.
vue-tui's transform host was a non-text yoga carrier: its direct text-leaf
children hit the paint no-op leaf branch and were silently dropped, a
<Newline> inside it rendered as a standalone "text" node (name-based
isInsideText saw no <Text> ancestor), and the SR squash ignored bare
text-leaf children. So `<Transform>ab</Transform>` rendered nothing and
`<Transform>a<Newline/>b</Transform>` applied the transform to empty lines.
Fix (treat a standalone <Transform> as a text context, matching Ink):
- paint.ts: when a transform has no yoga-carrying child, squash its inline
children to a string and write it like a text node, pushing the transform
as a per-line Output transformer (applied at paint, never at squash —
matching Ink). Transforms wrapping a yoga child keep the recursion path.
- yoga.ts: bind a text-style measure func on a standalone transform (its
squashed children, WITHOUT its own fn — matching Ink measureTextNode), and
toggle it off/on as yoga children are inserted/removed.
- Newline.ts / Text.ts: treat a <Transform> ancestor as a text context, so a
<Newline>/<Text> directly inside a transform emits inline virtual-text.
- node-ops.ts: allow bare text-leaf children of a transform (text-context
check) and mark a standalone transform dirty on inline child change; the
<Box>-in-<Text> guard still excludes transform (Ink renders Box-in-Transform
empty, no throw).
- screen-reader.ts: include bare text-leaf/virtual-text children of a
transform in SR output (Ink squashTextNodes includes #text).
All transform-in-text (G21/G32/G52), <Transform><Text>…</Text></Transform>,
nested transforms, plain <Text>/<Newline>, and Box-in-Transform cases verified
against Ink 7.0.4. Expected outputs (columns=40) confirmed by running the Ink
reference: `<Transform transform={s=>`<${s}>`}>ab</Transform>` + after →
"<ab>\nafter"; `<Transform>a<Newline/>b</Transform>` → "<a>\n<b>" (per-line);
SR → "ab" / "a\nb".
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Vue materializes a null/v-if/false render as a COMMENT host node that occupies
a positional slot in node.children. React never produces a childNode for such
children, so Ink's squash loop (squash-text-nodes.ts:13) never advances `index`
past them — empirically <Text>A{null}<Transform>(s,i)=>`${i}:${s}`>B</Transform>
yields "A1:B". vue-tui's three squash loops used the raw positional loop counter,
so a preceding comment took a slot and shifted the Transform to "A2:B".
The fix maintains a separate transform index that advances only for children
React would have produced — i.e. skips comment nodes — applied IDENTICALLY in
the paint, measure, and screen-reader paths so all three agree and match Ink.
G21 (which switched these loops from a hardcoded 0 to the positional counter)
introduced the precondition; its real-sibling positional indexing and G32's
transform-in-transform recursion remain intact.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Ink lays the <Static> node out via its OWN yoga node: Static.tsx merges
`{position:'absolute', flexDirection:'column', ...customStyle}` onto the
internal_static <ink-box>, and renderer.ts:48-56 reads
node.staticNode.yogaNode's computed layout directly. So every caller-supplied
layout style prop on `<Static style={{...}}>` (flexDirection, padding, margin,
gap, justifyContent, alignItems, width) governs how the static children are
laid out and written.
vue-tui's isolated paint replayed only the STYLE_PROPS subset (visual
color/border/overflow) that node-ops stores in `el.props`, then hard-defaulted
the fresh iso root to FLEX_DIRECTION_COLUMN. Every other layout style on
<Static> was a silent no-op: `flexDirection:'row'` painted as stacked lines,
`paddingLeft` was dropped, etc.
Fix: copyStyle the static node's yoga — which already holds every resolved
layout prop (including the column default) via node-ops applyYogaProp — onto
the iso root, instead of iterating the incomplete props bag. The static node's
own yoga is display:none (so it occupies no main-tree space) and
position:absolute; both are reset to flex/relative on the iso root since it is
the standalone paint root. An explicit `<Static style={{width}}>` is honored;
otherwise the iso root is constrained to the available columns. The live static
node's own yoga children are never reparented, so the main-tree layout/measure
is untouched.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Ink's screen-reader branch (ink.tsx:617-621) writes the wrapped output
verbatim — `stdout.write(erase + wrappedOutput)` with
`lastOutputToRender = wrappedOutput` (NO appended "\n" in ANY case) and
`lastOutputHeight = wrappedOutput === "" ? 0 : wrappedOutput.split("\n").length`.
The earlier G17 fix only suppressed the trailing newline for the EMPTY SR
frame (`output === ""`); a non-empty multi-line SR frame still fell through
to `output + "\n"`, which (1) parked the cursor on a spurious blank line
below the content and (2) made log-update count the previous frame as N+1
lines, emitting `eraseLines(N+1)` instead of `eraseLines(N)` on every
subsequent multi-line SR frame (off-by-one erase).
Broaden the suppression to ALL screen-reader frames so the written output
and its height match Ink's SR branch exactly. Non-SR interactive frames are
untouched — they still append "\n" as before.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
In Ink (App.tsx v7.0.4, 40b3a75) the isFocusEnabled guard lives ONLY in
handleTabNavigation, not in focusNext/focusPrevious. So after disableFocus()
pressing Tab/Shift-Tab is a no-op, but a programmatic
useFocusManager().focusNext()/focusPrevious() still moves focus.
vue-tui previously short-circuited focusNext/focusPrevious on the internal
`enabled` flag, making the programmatic API a no-op while focus was disabled —
divergent from Ink. Move the enabled-check out of focusNext/focusPrevious and
into the Tab/Shift-Tab input listener (mirroring handleTabNavigation), keeping
the focusables.length === 0 short-circuit so focusing an empty/unmounted tree
stays a harmless no-op. The now-redundant local `enabled` is dropped in favor of
ctx.enabled as the single source of truth.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
A plain <Box> (no explicit flexDirection) lays out as a row — yoga's Box
default is FLEX_DIRECTION_ROW (host/yoga.ts) — so its screen-reader children
must be joined with a space, matching Ink (which hardcodes flexDirection:'row'
in Box.tsx and derives the SR separator from style).
Root cause: flexDirection is a pure yoga prop. node-ops applies it to the yoga
node but does NOT mirror it into node.props (it is not in STYLE_PROPS), so the
yoga row default was never reflected there. screen-reader.ts read
node.props["flexDirection"], which was undefined for every live-rendered box
(both default AND explicit column/row), wrongly defaulting the separator to
"\n" in all cases.
Fix: resolve the direction from the yoga node (preferring an explicit
props.flexDirection for the direct-built unit fixtures), mirroring
static-channel.ts's resolvedFlexDirection so both SR linearization paths derive
the separator identically. Root keeps undefined → "\n" (Ink's column-default
root). row-reverse/column-reverse reversal is unaffected.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
The first exit() call now captures its value/error and initiates teardown
synchronously; subsequent exit() calls are complete no-ops, so waitUntilExit
resolves/rejects with the FIRST value rather than the last. This mirrors
Ink's handleAppExit guard (isUnmounted || isUnmounting → early return).
Previously each exit() queued a microtask that overwrote pendingExitResult/
pendingExitError before resolveExit ran, making it last-wins. An exitInitiated
flag set at the top of exit() now guards the value capture and re-resolve,
while the deferred microtask teardown (needed because exit() is called from
inside the Vue update cycle) is preserved.
Reverses the sweep-1 refutation: exit() was NOT already first-wins guarded.
Also guard unmount-in-progress (isUnmounting parity) + value→error test.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): record sweep-4 (G32-G38) + G32 pr-open
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): recurse into nested <Transform> in <Text> squash (Ink parity, G32)
paint.ts renderTextWithInlineStyles and text-measure.ts flattenLeaves now
factor their per-child text squashing into a recursive squashTransformChild
helper that recurses GENERICALLY into transform-typed children to any nesting
depth, applying each transform with its positional sibling index and the
innerText.length > 0 guard — matching Ink squashTextNodes generic recursion
(squash-text-nodes.ts:22-39). Previously a <Transform> nested directly inside
another <Transform> (inside a <Text>) was dropped: its grandchild loop only
handled text-leaf/virtual-text/text, so a transform grandchild contributed
nothing — silent total content loss in paint and 0-width measurement (broken
layout). Paint and measure stay behaviourally identical so layout and output
agree.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): record sweep-3 — 7 new gaps (G25-G31, all LOW severity)
Third re-audit of main (all 22 prior fixes merged) vs Ink 40b3a75. Trend:
sweep-1 found 14 (incl med), sweep-2 found 8 (incl 1 HIGH), sweep-3 found 7 —
ALL low severity. The high/medium-impact divergences are fixed; what remains
is a diminishing long tail of niche edge cases:
- G25 truncate-wrap multi-line edge; G26 kitty key-release input (reverses the
sweep-1 refutation); G27 raw-mode release input-detach timing; G28 useStdin
public-surface narrowing; G29 cursor-position-during-render; G30 SR nested
Transform-in-Transform inner fn (refines G23); G31 useAnimation interval
reactivity.
G07 + G24 candidates still pending the maintainer's decision.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* docs(parity): promote G07 + G24 candidates to allowlist (maintainer: keep)
G07 (kitty Ctrl+C exits): gated on internal_exitOnCtrlC (default true) — the
exitOnCtrlC option is intentionally honored under kitty protocol too, not just
\x03. G24 (multiple <Static>): vue-tui is strictly more capable. Both kept.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
EOF
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): concatenate children in screen-reader <Transform> (Ink parity, G23)
The screen-reader transform branch joined children with "\n", producing
newline-separated text where Ink concatenates. In Ink a <Transform> is an
`ink-text` node, so the screen-reader path squashes it via squashTextNodes
(squash-text-nodes.ts:42), which concatenates child text with "". Join with
"" instead of "\n" to match.
The transform node's own fn is intentionally NOT applied: verified empirically
against Ink 7.0.4 that squashTextNodes only applies the internal_transform of
*child* nodes (squash-text-nodes.ts:34-39), never of the top-level node handed
to it. A <Transform> directly under a <Box> is squashed as the top-level node,
so its own transform is skipped — yielding the bare concatenated children.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): ledger — G23 pr-open, reconcile G22 merged, log G23 spec correction
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): dedup screen-reader role against immediate parent only (Ink parity, G22)
Drop the `?? options.parentRole` grandparent fallback so a role-less intermediate
resets inherited parentRole to undefined for its children, matching Ink's
immediate-parent-only role dedup (render-node-to-output.ts:68-69 passes only
`node.internal_accessibility?.role`, no fallback).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): ledger — G22 pr-open, reconcile G21 merged
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): pass child sibling index to nested <Transform> (Ink parity, G21)
paint.ts renderTextWithInlineStyles and text-measure.ts flattenLeaves squash
loops now pass the child's position index to child.transform instead of a
hardcoded 0, matching Ink squash-text-nodes.ts:13,38 where internal_transform
receives the plain loop counter over node.childNodes (all siblings). A nested
<Transform> that is the Nth child of a <Text> therefore gets index = N. Both
spots use the same index basis so paint and measurement agree. Refines the
earlier G06 refutation — the inline/squash path was the real gap.
Also adds the `innerText.length > 0` guard in text-measure.ts flattenLeaves so
measurement skips transforms on empty text (matching paint.ts and Ink:34), and
converts screen-reader.ts squashTextContent to use forEach-with-index so the
nested-Transform index is correct in SR mode too (was hardcoded 0).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): ledger — G21 pr-open, reconcile G20 merged
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): guard writeToStdout/writeToStderr against post-teardown writes (Ink parity, G20)
Return early if teardownStarted, mirroring Ink ink.tsx:673/702, so a write
after unmount (e.g. a stray useStdout().write or console.log routed through
writeToStdout after teardown) cannot run clear()/write/restore on an
already-torn-down renderer and corrupt the restored terminal state.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): ledger — G20 pr-open, reconcile G19 merged
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): reset yoga props to default on dynamic removal (Ink parity, G19)
Removed style props now reset to the yoga default (margin/padding/min/gap/flexGrow→0, flexShrink→1, flexBasis→auto, flexDirection→ROW, flexWrap→NO_WRAP, alignItems→STRETCH, alignSelf→AUTO, justifyContent→FLEX_START, position→RELATIVE) instead of keeping a stale value — matches Ink's reconciler diff + styles.ts.
The fix threads the previous prop value (prev) from patchProp into applyYogaProp so that resets only fire on genuine removals (prev is a real value, not null/undefined from Vue's initial-mount or never-set patches). RESETTABLE_PROPS is extended with all newly resettable keys.
Follow-up blocker fixes:
- marginX/marginY/paddingX/paddingY now map to Yoga.EDGE_HORIZONTAL/EDGE_VERTICAL (matching Ink styles.ts) instead of concrete EDGE_START/END/TOP/BOTTOM. They compose with the specific edges per yoga precedence, so removing an axis shorthand no longer clobbers a surviving marginLeft/etc.
- applyYogaProp and all setters now treat null the same as undefined (value == null) for the removal/reset path. Vue's host renderer passes next=null (not undefined) when a key disappears from a spread props object (e.g. Static spreads style into host props), which previously bypassed the reset and forwarded raw null into yoga (NaN/0 corruption).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): ledger — G19 pr-open, reconcile G18 merged
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): restore terminal on signal exit via signal-exit (Ink parity, G18)
Previously nothing routed a process signal to teardown(): SIGINT-as-signal,
SIGTERM or SIGHUP killed the process with the cursor hidden, the alternate
screen active and raw mode on, leaving the terminal corrupted.
Mirror Ink (ink.tsx:426): register signal-exit's onExit(teardown,
{alwaysLast:false}) at interactive mount, storing the unsubscribe fn, and
call it first thing in teardown() (ink.tsx:765) so the handler is removed on
unmount()/exit() and can't leak or double-run. teardown() stays idempotent
(teardownStarted guard) so a signal-triggered teardown plus a later unmount
won't double-run, and we don't prevent the process from exiting. Only the
live interactive, non-debug mount registers — render-to-string /
non-interactive paths never touch process signal handlers; registration is
guarded against double-registration.
Uses signal-exit v4 (named onExit export; ships ESM + types, so no
@types/signal-exit needed). PTY test sends SIGINT/SIGTERM/SIGHUP to a mounted
alt-screen app and asserts the captured output ends with show-cursor
(\x1b[?25h) + leave-alt-screen (\x1b[?1049l).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Review follow-ups (3 fixes): register signal-exit whenever interactive
(drop the !debug gate so debug-but-interactive apps, which still enter the
alt-screen/hide the cursor, restore on signal — Ink ink.tsx:426); add
!teardownStarted to the registration so a spent app instance does not
re-register on a same-instance remount (the next unmount() returns early at
the teardownStarted guard before it could unsubscribe — a leak); and make
the PTY test prove the SIGNAL drove teardown (fixture never self-unmounts, so
restore bytes can only come from the signal path) with a debug-mode signal
test, an exit-anchored waitForOutput drain, and a bounded retry for the
async-flush race under saturated runners.
Review follow-ups (2 fixes): synchronous restore flush on signal — the
signal-exit teardown path now writes the restore escapes (show-cursor,
leave-alt-screen, disable-kitty) via fs.writeSync to the stdout fd so they
reach the terminal before signal-exit re-raises the signal (a buffered async
stream.write could be lost on abrupt exit); the normal unmount path keeps async
writes. Removed the config-wide retry:3 from vitest.pty.config.ts (it masked
the whole PTY suite) and scoped a retry:2 to the signal-teardown describe only,
for the residual parent-side node-pty onData read-race under a saturated runner.
* chore(parity): ledger — G18 pr-open
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Re-audit of main (all 16 sweep-1 fixes merged) vs Ink 40b3a75 found 8 confirmed
divergences the first sweep missed:
- G18 (HIGH): no signal-based teardown — terminal corrupted on SIGINT/SIGTERM/SIGHUP.
- G19 (med): dynamic yoga prop removal doesn't reset to default (stale layout).
- G20: writeToStdout/stderr lack an isUnmounted guard.
- G21: nested <Transform> in <Text> gets hardcoded index 0 (squash path; refines the
earlier G06 refutation — the inline/nested case IS a gap).
- G22: SR role dedup inherits grandparent role.
- G23: <Transform> under <Box> SR-joins children with newline (should concatenate).
- G24 (CANDIDATE): vue-tui supports multiple <Static>; Ink has one staticNode — awaiting
maintainer decision (appended to ink-parity.md candidates, like G07).
Also reconciles G17 -> merged (#45).
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): linearize screen-reader static output + drop empty-SR-frame newline (Ink parity, G17)
(a) The live static channel flushed <Static> via the 2D grid painter
(paintIsolated) even in screen-reader mode, so bordered static items leaked
box glyphs. paintStaticNode now takes an isScreenReaderEnabled flag and
linearizes fresh static children via renderScreenReaderOutput
(skipStaticElements:false) instead — matching Ink's renderer.ts:24, which
renders node.staticNode through renderNodeToScreenReaderOutput. Non-SR static
is unchanged. render.ts commit() and render-to-string.ts thread the flag.
(b) Interactive SR frames went through renderInteractiveFrame, which appends
"\n" even for empty output, leaking a spurious blank line. Ink's SR path writes
the wrapped output directly with lastOutputToRender = wrappedOutput (no appended
newline), so an empty SR frame emits zero lines. We now suppress the trailing
newline for EMPTY SR output only — matching ink.tsx:573-626 — leaving non-SR
and non-empty SR frames untouched.
Follow-up fixes (two review findings):
- renderToString in SR mode no longer DROPS <Static> output: the SR return
branch now prepends the captured/linearized static output like the non-SR
path (Ink's SR renderer returns staticOutput when node.staticNode exists,
renderer.ts:24-33).
- The SR static linearization now honors the <Static>'s resolved flexDirection
for separator + child order (read from yoga via getFlexDirection), matching
screen-reader.ts:73-82 (row/row-reverse → space, *-reverse reverses order);
the default column case is unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): ledger — G17 pr-open, reconcile G16 merged
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): let per-edge borderDimColor=false override general dim (Ink parity, G16)
edgeDim now uses `?? generalDim` (nullish) instead of `|| dimAll`, so an explicit per-edge false wins — matching Ink render-border.ts:54. The five borderDimColor prop declarations in Box.ts are changed from bare `Boolean` to `{ type: Boolean, default: undefined }` so Vue does not boolean-coerce absent per-edge dim props to false (Vue only applies absent→false casting when no explicit default is provided), preserving the undefined sentinel needed for the nullish fallback.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): ledger — G16 pr-open, reconcile G14 merged
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): guard against two renderers on the same stdout (Ink parity, G14)
Make skipped-mount unmount() a pure no-op: add `skippedMount` flag set in
the instance-reuse guard branch; teardown() and resolveExit() return early
when set, so no write-barrier or stdout touch reaches the owner's stream.
Strengthen test: assert app2.unmount() writes nothing to process.stdout
(catches the pre-fix empty write-barrier) and that app1 still owns the
WeakMap entry after app2.unmount() (third mount on same stdout still warns).
* chore(parity): ledger — G14 pr-open, reconcile G13 merged
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): clear+reset on terminal-width decrease during resize (Ink parity, G11)
Track lastTerminalWidth (initialized at mount via resolveSize); onResize clears
writer + resets frameState.lastOutput/lastOutputToRender when the new width is
narrower than the previous width, mirroring the logic in ink.tsx:459-474 that
prevents duplicate overlapping re-renders on terminal narrow. Width increases and
pure height changes are unaffected.
Preserve outputHeight on narrowing (matches Ink ink.tsx:462-466 which leaves
lastOutputHeight intact): zeroing it suppressed the clearTerminal path in
shouldClearTerminalForFrame for overflowing frames (hadPreviousFrame=false).
Adds overflow-path test that was RED with outputHeight=0 and GREEN with the fix.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): ledger — G11 pr-open, reconcile G10 merged
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): throw a descriptive error when raw mode is unsupported (Ink parity, G10)
Previously the raw-mode acquire path silently no-opped on a stdin where raw mode
is unsupported (non-TTY / isRawModeSupported false), so using useInput on such a
stdin did nothing with no diagnostic. Ink's handleSetRawMode (App.tsx:315-327)
instead throws immediately when enabling raw mode is unsupported, with two
distinct messages (default process.stdin vs a custom stdin) both pointing at the
isRawModeSupported docs.
StdinController.acquireRawMode now throws that two-message error on
!isRawModeSupported. The unguarded useInput path surfaces it (matching Ink's
use-input.ts, which calls setRawMode(true) ungated). useFocus now guards on
isRawModeSupported before acquiring (matching Ink's use-focus.ts), so focus
degrades to a safe no-op on a non-TTY rather than throwing. releaseRawMode keeps
its no-op guard — Ink only throws when enabling.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): ledger — G10 pr-open, reconcile G09 merged
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* test(runtime): assert full raw-mode error message for exact Ink parity (G10, codex)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): wrap external stdout/stderr writes in synchronized-update markers (Ink parity, G09)
writeToStdout/writeToStderr now emit bsu/esu around clear+write+restore when
shouldSynchronize, matching the render path and Ink ink.tsx:687-728. The sync
variable was already computed at mount time (render.ts:489); the external-write
functions simply lacked the wrapping. For writeToStderr, BSU/ESU go to stdout
(not stderr) because synchronized-update mode is a stdout capability — exactly
mirroring Ink's ink.tsx:717-728 behaviour.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): ledger — G09 pr-open
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): draw box borders per-edge without the min-size guard (Ink parity, G05+G15)
Removed blanket `w<2||h<2` return from drawBorder; replaced it with a
`w<1||h<1` degenerate guard. Vertical sides now start at
`offsetY = top ? 1 : 0` and run for `Math.max(0, h - topRows - bottomRows)`,
matching Ink render-border.ts:133. Fixes: (G05) a 1-cell-tall box with only
side rails rendered nothing; (G15) with borderTop=false the left/right rails
were shifted one row down. Updated 4 existing snapshots that encoded the
old buggy behavior and added 3 new tests that verified red before green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): ledger — G05+G15 pr-open, reconcile G03 merged, log snapshot decision
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): render linear screen-reader output in the live commit path (Ink parity, G03)
commit() previously called paint(tuiRoot) (the 2D grid painter) unconditionally
for both the non-interactive and interactive branches, so an app mounted with
isScreenReaderEnabled emitted the visual frame (box-drawing borders, padded
grid) into the live stream instead of flat linearized text. isScreenReaderEnabled
was only consulted to disable commit throttling.
Add a renderFrame(width) helper that branches on isScreenReaderEnabled: when SR
is enabled it linearizes the tree via renderScreenReaderOutput(tuiRoot,
{ skipStaticElements: true }) and wraps it with wrapAnsi(out, width,
{ trim: false, hard: true }), mirroring Ink's onRender SR branch
(ink.tsx:598-603). Both commit branches now call renderFrame() instead of
paint() directly. The non-SR path is byte-for-byte unchanged (renderFrame
returns paint(tuiRoot)). Static output continues to flush through the existing
paintStaticNode path; full SR-static linearization parity is deferred.
render-to-string.ts already used renderScreenReaderOutput and is unaffected.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): ledger — G03 pr-open, reconcile G02 merged, track G17 (SR edges)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): coalesce useAnimation ticks within the render-throttle window (Ink parity, G02)
useAnimation now coalesces scheduler ticks that fall inside the current
render-throttle window and reports delta as the time since the last
RENDERED tick, so velocity-driven motion (position += speed * delta)
advances at correct wall-clock speed even when the commit throttle is
coarser than the animation interval. Previously delta was ~one scheduler
interval per committed tick, under-integrating velocity at render time.
- animation-scheduler: createAnimationScheduler(renderThrottleMs = 0)
exposes renderThrottleMs on the AnimationScheduler (no-op variant = 0).
- render.ts: derive animationRenderThrottleMs from maxFps using Ink's
Math.max(1, ceil(1000/maxFps)); 0 on debug/screen-reader/unthrottled
paths, mirroring the commit-throttle gate.
- useAnimation: tick() skips while now < nextRenderTime; on an allowed
tick delta = now - lastRenderedTime, then nextRenderTime = now + window.
Also default maxFps to 30 (Ink parity: options.maxFps ?? 30) and derive a
single renderThrottleMs that drives BOTH the commit scheduler and the
animation scheduler, so the coalescing engages on the default non-debug
path (previously it only engaged when maxFps was passed explicitly).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): ledger — G02 pr-open, reconcile G01 merged, log G02 decisions
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(runtime): unmount written <Static> items to match Ink (G01)
Ink's <Static> renders `items.slice(index)` and advances `index` to
`items.length` in a post-commit `useLayoutEffect`, so once an item has been
painted it is removed from the tree and its component unmounts. vue-tui kept
every Static item mounted forever: the component always mapped the full
`props.items`, and write-once was enforced only at flush time via a positional
`writtenCount` slice — the item components never tore down.
Now the <Static> component owns a `cursor` (Ink's `index`) and renders only
`items.slice(cursor)`. The renderer advances the cursor AFTER a commit has
painted the fresh items, via an `onWritten` callback registered on the host
static node — the vue-tui analogue of Ink's post-commit layout effect. This
ordering guarantees items are written before they are sliced out and unmounted,
so no item is ever lost or re-painted.
Write-once bookkeeping moved from a positional `writtenCount` to a
`writtenNodes` Set keyed by host-node identity. A single logical item expands to
several host nodes (the <Text>/<Box> plus empty text-leaf fragment anchors Vue
inserts), so a positional count mis-sliced once the cursor advanced; identity
tracking is anchor-agnostic. The shared `paintStaticNode` helper paints children
not yet in the set, records them, prunes unmounted entries, then fires
`onWritten`; render.ts, render-to-string.ts and flushStatic all use it.
Make the cursor mirror Ink fully so it can DECREASE, not just increase.
`onWritten` now SETS the cursor to items.length (was max-with-current), and a
length watch lowers it on shrink — needed because a shrink that leaves the
already-sliced children empty produces no host mutation, hence no commit/
onWritten to re-sync. Without this, [A,B] (cursor→2) → [A] → [A,C] sliced(2)=[]
and silently dropped C. paintStaticNode now always prunes and calls onWritten
(even on empty commits), painting only when there are fresh children.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore(parity): ledger — G01 pr-open, reconcile G12 merged
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Replace three `stdout.columns ?? 80` / `stdout.rows ?? 24` spots in render.ts with
`resolveSize(stdout).columns/rows`. The `??` guard only falls back on null/undefined,
not on 0 — so non-TTY environments where stdout reports 0 columns would collapse Yoga
layout to width 0. Ink's `getWindowSize` (utils.ts:8-23) uses a truthy guard
(`if (columns && rows)`) and a fallback chain through terminal-size → 80/24 defaults.
`resolveSize()` in useTerminalSize.ts already implements this chain; now exported and
used by the renderer. The non-TTY viewportRows → 24 branch is preserved (Ink-aligned).
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
G06 re-verification: the audit claimed <Transform>'s fn gets a hardcoded
index 0 "instead of the childNode index". Ink's index (output.ts:230-239)
is the LINE index, applied per output line — not a child index; the audit
misread it. vue-tui already produces correct per-line line indices for
multi-line transforms (existing tests "transform with multiple lines" and
transform-yoga pass unmodified). paint.ts:314's transform(innerText, 0) is
only the inline <Transform>-in-<Text> path (single logical line, 0 matches
Ink). No observable gap — marked refuted, not fixed.
Also reconciles G08 -> merged (landed in #31).
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Ink derives the focus id via useMemo(() => customId ?? random, [customId])
and keys its add/remove effect on [id], so changing the id prop re-registers
the component under the new id. vue-tui captured `const id = options.id ?? …`
once at setup (and typed id as a plain string), so it never reacted.
- Widen id to MaybeRefOrGetter<string>.
- Track the current registration and re-register (unsubscribe/remove old,
subscribe/add new, re-apply active state) in a watcher keyed on the resolved
id, mirroring Ink's [id] effect. isActive handling unchanged.
Test (test-first, verified red before the fix): focus is driven purely by
focus(id) (no Tab, which would focus by position and mask the bug); changing
the id re-registers under the new value and the old id goes dead.
Also reconciles G04 -> merged (landed in #30).
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Ink's render-border.ts computes each border edge's background from
border<Edge>BackgroundColor ?? borderBackgroundColor only — it never falls
back to the Box's own backgroundColor. vue-tui's colorizeEdge had an extra
`?? bgColor` fallback, so a Box with backgroundColor but no explicit border
background painted its background onto the border glyphs too.
Drop the fallback. Background still fills the inner content area; border
glyphs are now uncolored unless an explicit border background is set.
Tests rewritten to match Ink (per maintainer's align-to-Ink policy; see
.agents/docs/parity-ledger.md Decisions log):
- add failing-first repro "Box backgroundColor does not bleed onto border
glyphs (Ink parity)"
- "wrapped text preserves backgroundColor on every content line": assert
inner rows carry bg, border rows don't (height 4->5 so text fits)
- "Box background with border fills content area": snapshot updated so
border rows have no bg
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Ink-parity audit against v7.0.4 (commit 40b3a75): 10 areas reviewed in
parallel, candidates adversarially verified by 26 agents total.
16 candidates verified → 14 confirmed gaps (3 medium, 11 low), 2 refuted:
- exit() second-wins: vue-tui is already guarded (not last-wins).
- kitty key-release printable-text suppression: Ink behaves the same.
Each gap recorded in the ledger with Ink/vue-tui evidence and a fix sketch;
fixes ship as individual PRs per the loop in ink-parity-loop.md.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Add a "test" script (vp test --passWithNoTests) to @vue-tui/cli and a
ci:test:cli branch (dependsOn ci:build) to the run.tasks graph. No CLI tests
exist yet, but wiring the branch now means future CLI tests are covered
automatically rather than silently skipped. #26 (item 3).
The test asserting process.listenerCount("beforeExit") lived in the
parallel wait-flush.test.tsx, conflicting with the branch's rule that
process-global-state tests live in *.sequential.test.* files. Move it into
leak.sequential.test.tsx (already the home for process exit/SIGINT listener-
count assertions) and extend that file's header comment. #26 (item 2).
The resize handler painted synchronously via commit() but left the commit
scheduler's pending trailing throttle timer armed. If an update was sitting
in that timer, it fired a second doCommit() right after the resize paint —
and because shouldClearTerminalForFrame clears whenever the previous frame
overflowed the viewport, the second commit emitted a duplicate clearTerminal.
Cancel the pending trailing commit before the synchronous paint; the paint
already reflects the current tree, so the pending commit is redundant.
Regression test (test-first) in throttle.sequential.test.tsx reproduces the
double-clear (2 clears) and verifies the fix (1 clear). Closes#26 (item 1).
Review follow-ups on PR #25:
- package.json `ready`: run build BEFORE check:lint, matching the CI graph
where ci:lint dependsOn ci:build. Type-aware lint rules need the built
@vue-tui/runtime types; with lint before build, `vp run ready` on a fresh
checkout (dist removed) could misreport lint. Verified `CI=true vp run ready`
on a clean checkout now passes with 0 lint warnings.
- vite.config.ts + ci.yml comments: corrected the stale claim that "fmt and
lint run immediately alongside build" — only fmt has no build dependency;
lint now waits on build too. Dropped the outdated "~40s vs ~60s" figure.
The concurrency notes described `sequence.concurrent: true`, but that was
rolled back — it starved timing-sensitive render tests on the 4-core CI runner.
Rewrite to reflect what actually ships:
- File-level parallelism (fileParallelism: true), tests within a file serial;
explain WHY in-file concurrency is deliberately avoided (the local-vs-CI
core-count trap) and that PTY needs pool: forks.
- *.sequential.test.* files group process-global-state tests (fake timers,
listener/yoga-node counts).
- New rule: tests must not implicitly depend on host env. CI=true flips
interactive mode off, so both vitest configs force CI:false; inject env
behavior explicitly and reproduce CI with `CI=true vp run ci` on a fresh
checkout.
- FORCE_COLOR must also be set in spawned child envs, not just vitest config.
Yunfei He