vue-tui/.agents/docs/ink-divergences.md

vue-tui - Intentional Divergences from Ink

vue-tui started as a Vue 3 port of Ink, and it still tracks Ink closely: the aim is behavioral parity except where a difference is deliberate. It is no longer only a port, though. It has its own design decisions, additive features, and Vue-native choices.

This document records the places vue-tui intentionally differs from Ink by design. A difference that is not listed here is treated as a bug, or simply unverified behavior, not a design choice.

Reference baseline: Ink v7.0.4 (commit 40b3a7578811fd616341ca4e31cc7748aeeff12f). When bumping the target Ink version, re-validate every entry below against the new source.

Why align to Ink — and when not to

Aligning to Ink is a means, not an end. Ink is a mature, battle-tested implementation, so matching its public surface and behavior lets vue-tui inherit years of bug-fixes and edge-case handling for free. That — reducing bugs by reusing proven behavior — is the entire point of alignment.

It follows that alignment is not the top priority. When Ink's behavior is itself a defect, is unreasonable, or is un-idiomatic for Vue, conformance to Vue's philosophy and the plain reasonableness/correctness of the behavior outrank parity. There vue-tui deliberately diverges, and records it here so the choice is conscious and blessed, not drift.

This guards against two opposite failure modes:

• Blind alignment — copying Ink even where Ink is wrong, or where matching would force un-Vue machinery, merely to match. (Rejected e.g. in the useCursor corner-zombie, the resolve-on-throw exit, and the paint-time invalid-input crash — Ink behaviors vue-tui treats as defects, not contracts.)
• Lazy divergence — inventing a different behavior and rationalizing it as "Vue's way is better" with no genuine Vue-philosophy or correctness reason. Mere presence in this file is not a blessing; every kept divergence needs a real reason and a maintainer decision.

So the test for any difference is never just "does it match Ink?" but "is this the most reasonable, most Vue-idiomatic behavior — and where it diverges from Ink, is that because Ink is wrong or un-Vue, recorded with a maintainer decision?" Reasonableness and Vue idiom come first; alignment is simply the cheapest way to get there whenever Ink is already right.

How to Classify a Divergence

Classify each divergence by the first rule that applies. The order matters: earlier sections are narrower, while later sections are broader fallbacks.

1. If Ink's supported subset still behaves the same and vue-tui only accepts more inputs, supports more contexts, or exposes an extra capability, put it in Additive Supersets.
2. If the primary reason is alignment with Vue's API shape, framework model, mental model, or user expectations, put it in Vue API and Mental Model Divergences.
   • Use Model-Implied Differences when the difference comes from the React/Vue framework-model boundary. Matching Ink would require React-shaped machinery inside Vue, changing a core Vue-facing contract, or dealing with a React-only concept that has no Vue equivalent.
   • Use Vue-Idiomatic Choices when Ink could be copied, but vue-tui chooses the behavior or public surface that better fits Vue's reactivity, lifecycle, component boundaries, current-props model, or API conventions.
3. If the divergence is intentional but is not additive and is not primarily Vue-aligned, put it in Intentional Divergence Choices.
4. If the note is not a divergence, put it in Non-Behavioral Notes.

Each divergence entry states what Ink does, what vue-tui does, and why the difference is deliberate. Some entries also record consequences, costs, tests, or maintainer decisions where those details are needed to understand the decision.

---

Additive Supersets

vue-tui supports more than Ink in these cases. Ink-supported inputs and common use cases remain compatible; vue-tui only adds accepted inputs, contexts, or capabilities.

Multiple <Static> regions

• Ink: keeps a single staticNode; only one <Static> is honored.
• vue-tui: findStatics(root) renders every <Static> in the tree.
• Why: a tree with two <Static> regions renders both. Maintainer decision (2026-05-30): KEEP.

Ctrl+C exits under the kitty protocol too

• Ink: exits only on the legacy \x03 byte (in App), so a kitty-protocol Ctrl+C (\x1b[99;5u) parses fine but never exits. Its guard is byte-specific, not Ctrl+C-specific.
• vue-tui: one encoding-agnostic exit in the always-on stdin controller (emitInput), via parseKeypress. It matches Ctrl+C in both the legacy and kitty forms (but not Ctrl+Shift+C), so it fires no matter which composable holds raw mode (useInput / useFocus / usePaste, or none).
• Why: exitOnCtrlC is defined in terms of Ctrl+C, not one byte encoding. Keeping the exit at the single always-on layer avoids splitting the behavior across two places. Opt out with exitOnCtrlC: false. Maintainer decision (2026-05-30): KEEP. Tests: usePaste-only app exits on {legacy,kitty} Ctrl+C in input-kitty.test.ts.

parseKeypress filters kitty query-responses

• Ink: filters kitty keyboard-protocol query-responses (ESC[?Nu) in exactly one place: the auto-detection lifecycle in ink.tsx (stripKittyQueryResponsesAndTrailingPartial on a private onData buffer). Its parse-keypress.ts has no query-response branch.
• vue-tui: mirrors that detection layer (in kitty-keyboard.ts) and adds a parser-level filter: parseKeypress returns { ignore: true } for ESC[?Nu, which useInput then drops.
• Why: the detection layer does not cover the runtime input pipeline (stdin 'data' -> inputParser -> emitInput -> useInput -> parseKeypress). In enabled mode it never runs; in auto mode its onData listener and the stdin controller's handleData both subscribe to the same 'data' event, so stripping its private buffer cannot stop the chunk reaching handleData; and after detection settles the listener is gone. Empirically (Layer 2 removed, rebuilt) a stray query-response reaches a useInput handler as spurious "[?1u" input in all of those cases, including a response split across two reads, which inputParser reassembles before dispatch. The parser-level filter is therefore intentional, not redundant. Introduced 2026-05-31. Tests: "kitty query-response - end-to-end filtering" in kitty-lifecycle.test.ts (RED without it).

useAnimation() outside a render tree drives a standalone animation

• Ink: the default AnimationContext.subscribe() is a no-op subscription with startTime: 0, so a useAnimation rendered outside an Ink tree never ticks.
• vue-tui: useAnimation falls back to a freshly created standalone scheduler (inject(AnimationSchedulerKey, null) ?? createAnimationScheduler()), so frame/time/ delta advance even with no surrounding app.
• Why: the composable still does useful work in isolation, such as a unit test or a non-rendered driver. Additive; inside a tree the injected scheduler is used exactly as Ink's. Contrast with the terminal-bound composables in the outside-render-tree entry, which throw because they have no meaningful standalone mode.

measureElement / useBoxMetrics also accept a Vue component-instance ref

• Ink: measureElement(node: DOMElement) and useBoxMetrics(...) read node.yogaNode directly: a host DOMElement only.
• vue-tui: the ref is resolved through $el as well: a ref bound to a Vue component (whose root host node is on $el), not just a host-node ref, resolves to the underlying yoga node.
• Why: in Vue a template ref on a component yields the component instance, and its host node is reached via $el. Because <Box> is a defineComponent, the $el path is in fact the primary path a normal ref on <Box> takes — the bare host-node ref is the rarer raw-host case. Supporting both is a strict superset that matches how Vue refs behave; a bare host-node ref still works identically to Ink. Maintainer decision (2026-06-13): KEEP — a reasonable Vue-idiomatic adoption (the component-instance ref is the natural Vue path; the bare host-node ref stays Ink-identical).

Two apps sharing one stdin both receive input

• Ink: raw-mode count and the input listener are per-App (useRef), and Ink reads via the 'readable' event + stdin.read() (pull, App.tsx:278-313). Two render()s to different stdout but one stdin each attach a readable listener, but the first-registered listener's read() loop drains the buffer every tick, so the second app receives no input until the first unmounts. And because counts are per-App, the first app's unmount calls stdin.setRawMode(false), dropping raw mode while the second still needs it.
• vue-tui: the terminal raw-mode toggle is refcounted per-stdin (a shared WeakMap), so one app's unmount cannot drop raw mode while another holds it; and the 'data' input listener is per-controller. Each app attaches its own handleData -> own parser -> own emitter. Since 'data' (push) broadcasts to every listener, both apps receive every keystroke, and the second keeps receiving after the first unmounts.
• Why: this covers a combination vue-tui already allows: two createApps to different stdout. The same-stdout no-op is keyed on stdout, not stdin. The push model has no drain race, and a shared raw-mode refcount matches the ownership model when several renderers share one input. The common one-app-to-terminal flow is unchanged: one controller's localRefs equals the shared refs. Test: raw-mode-lifecycle.test.tsx ("two apps sharing one stdin both receive input..."). Maintainer decision (2026-06-07): KEEP.

Vue API and Mental Model Divergences

These divergences come from choosing Vue's API shape, framework model, mental model, and user expectations as the source of truth while tracking Ink. Some are model-implied: matching Ink would require React-shaped machinery inside Vue, changing a core Vue-facing contract, or handling a React-only concept that has no Vue equivalent. Others are idiomatic choices: Ink could be copied, but vue-tui chooses the behavior or public surface that better fits Vue's reactivity, lifecycle, component boundaries, current-props model, or API conventions.

Model-Implied Differences

Reactive composable state is a shallowRef, not a plain snapshot

• Ink/React: a hook re-runs on every render of its component, so it can return a plain value and the caller always reads the latest one. useFocusManager().activeId, for instance, is a bare string | undefined, re-read fresh each render.
• vue-tui: a composable's setup() runs once, so reactive state cannot be a plain snapshot: it would freeze at setup time. vue-tui returns a shallowRef whose .value updates and re-renders the template; read these as .value. Every stateful composable follows this — useFocusManager().activeId is a single ref read as .value, while a composable may instead return an object of refs (useWindowSize() returns { columns, rows }, read as .columns.value / .rows.value). An empty single-ref state holds null (Vue's convention for an empty ref: a template ref is ref<T | null>(null)), where Ink's plain value is undefined.
• Why: the two frameworks track a changing value differently. React reads the newest value by re-running the hook; Vue wraps it in a ref the template subscribes to. This is the general rule, not a per-API choice. useFocusManager().activeId is just one instance. This follows Vue's philosophy: changing state is exposed as a reactive source, not as a one-time snapshot. Maintainer decision (2026-06-07): KEEP.

A setup()-throwing component emits a dev-only [Vue warn] on stderr

• Ink: a component that throws during render surfaces only through the error overview / exit path; React emits no extra framework warning on stderr (verified: stderr stays empty).
• vue-tui: in a development build, a component whose setup() throws additionally produces Vue's own [Vue warn] lines on stderr (for example, the missing-render-function warning) that Ink has no analog for. While console patching is active (the default; disabled by patchConsole: false or debug, independent of interactive mode), vue-tui treats the [Vue warn] prefix as Vue's framework-diagnostics channel and drops those stderr lines. The patch is installed before the first mount (matching Ink, which patches before the first render), so a setup() throw during the initial mount is filtered too. With patching off, every [Vue warn] surfaces.
• Why: these warnings come from Vue itself and are dev-only (stripped in production builds); they never enter the stdout frame and do not change the exit path. Documented so the stray warn is not mistaken for vue-tui behavior: it is Vue's framework diagnostics. The prefix filter may also drop user-authored stderr logs that intentionally reuse the reserved [Vue warn] prefix; use a different application prefix when that output must be preserved. Maintainer decision (2026-06-06): KEEP.

React concurrent mode

• Ink: built on React; useTransition / useDeferredValue work as ordinary React hooks. Ink v7 also exposes a concurrent?: boolean render option (default false) with two distinct halves (run-verified vs v7.0.4): the root-tag half is inert — under the pinned reconciler (react-reconciler 0.33.0 / React 19) every root is overwritten to ConcurrentRoot, and hook/preemption probes behave identically in both modes — but the dispatch half is live: the default commits the first frame synchronously inside render() / rerender() (bytes reach stdout before the call returns), while concurrent: true schedules the commit asynchronously on a later tick.
• vue-tui: no equivalent — no such composables, and MountOptions has no concurrent flag. mount() commits the first frame synchronously, matching Ink's default dispatch.
• Why: React scheduling concepts with no Vue counterpart; N/A rather than a parity gap. The absent MountOptions.concurrent is this entry, not an unlisted difference: the one observable behavior the flag adds (deferring the first paint past the mount call) has no Vue-side demand, and vue-tui already matches Ink's default.

<Transform> treats all-comment children as no children

• Ink: Transform returns null only for undefined / null children. React's false child and a literal [] child are not nullish, so each creates an empty ink-text node that consumes a flex-gap slot, and in screen-reader mode that node still announces accessibilityLabel (a false or [] child with a label reads the label).
• vue-tui: after slot resolution, null / false / undefined / v-if="false" / a false-yielding && all materialize as the same Comment vnode — React's false !== null edge has no Vue equivalent. <Transform> treats an absent slot, an all-comment slot, or an empty slot array (() => []) as no renderable children and returns null: the node is omitted, no gap slot is consumed, and a <Transform accessibilityLabel> whose children all resolve this way announces nothing in screen-reader mode (the guard runs before label substitution, as in Ink). Boundary parity: '' and 0 children are text vnodes, not comments — both engines render a node ('' takes a gap slot; 0 prints 0) — and a Vue JSX {[]} child is a Fragment vnode that still renders a node, matching Ink; only the bare () => [] slot collapses.
• Why: a child set that renders nothing equals omitting the child — letting framework anchors occupy layout slots would be worse than matching Ink's React-only false !== null edge, which Vue cannot see. The same forcing covers the screen-reader case: Ink announces the label for false but not null children; vue-tui sees identical comments, cannot honor both, and consistently takes the null side. The () => [] collapse alone is not model-forced (Vue can see the empty array); it is a deliberate consistency rider — in each engine () => [] and () => [false] behave alike (Ink renders a node for both, vue-tui omits both), and aligning only [] would create an asymmetry that exists in neither engine without reaching parity. Maintainer decision (2026-06-07): KEEP. Test: transform.test.tsx.

Vue-Idiomatic Choices

Entry point - createApp() instead of render()

• Ink: render(<App/>, options?): options is RenderOptions; returns an Instance.
• vue-tui: createApp(App) returns a TuiApp; app.mount(options?) takes MountOptions.
• Why: mirrors Vue's own createApp mental model. A Vue developer expects an app object (TuiApp) they mount, not a one-shot render call. The mount-options bag and the app handle are therefore Vue-shaped (MountOptions / TuiApp), not render()-shaped (RenderOptions / Instance). Do not add Ink-compatible aliases here: aliases would make the public API look render-shaped while the actual runtime contract is app-shaped. Maintainer decision (2026-06-07): KEEP.

Removing display resets to the default (visible)

• Ink: applyDisplayStyles (styles.ts) calls setDisplay(DISPLAY_NONE) whenever the prop diff carries a display that is not 'flex', and Ink's reconciler diff emits a withdrawn key as display: undefined. So clearing a previously-set display ('none' or 'flex' → removed) hides the box: Ink treats the withdrawn prop as none, neither keeping the prior value nor restoring the default. (A box that simply omits display stays visible — 'display' in style is false, so no setDisplay runs; but an explicit display={undefined} is itself applied as DISPLAY_NONE and hides, like any non-'flex' value.) In the common toggle display={hidden ? 'none' : undefined}, Ink stays hidden on the undefined branch; you must set display="flex" to show it again.
• vue-tui: a removed/undefined display resets to the Box default DISPLAY_FLEX (visible): the same state as if the prop had never been set.
• Why: render = f(current props): no display set means the default (visible). Persisting a withdrawn prop, or flipping it to hidden, does not match that model. Maintainer decision (2026-05-31): KEEP.

Nullish flexDirection / flexWrap reset to Box defaults

• Ink: the public <Box> injects flexDirection:'row' and flexWrap:'nowrap' before spreading user style. If a previously-set prop is truly omitted, that default reaches the host and both engines reset (column -> omitted renders "A\nB" -> "AB"; wrap -> omitted stops wrapping). But an explicit flexDirection={undefined} / {null} or flexWrap={undefined} / {null} overwrites the default before the host layer. Ink's applyFlexStyles has no reset branch for these two props, so a dynamic nullish value preserves the previous Yoga value. On first mount, nullish flexDirection leaves Yoga's column default; nullish flexWrap happens to match nowrap.
• vue-tui: nullish current values reset to the public Box defaults (row / nowrap) in the same way as true omission (G19). A conditional spread that removes the key remains parity with Ink; a live binding whose value becomes null or undefined intentionally resets instead of preserving prior Yoga state.
• Why: render = f(current props): a Vue binding with no current flexDirection / flexWrap value means "use the Box default", not "keep whatever Yoga had last render". Preserving the prior value would make layout depend on history rather than current props. The cost is limited to explicit nullish public bindings; true omission remains Ink-parity. Tests: prop-reset.test.tsx.

Public composable naming follows Vue conventions

• Ink/React: public APIs are hooks (useFocus, useInput, ...), but return-type naming is mixed: the stream/app hooks return exported context types named XProps (useStdin → StdinProps, useStdout → StdoutProps, useApp → AppProps), newer hooks return exported non-XProps types — result-named (useBoxMetrics → UseBoxMetricsResult, useAnimation → AnimationResult) or a bare data name (useWindowSize → WindowSize) — and the rest fit neither: useInput/usePaste return void, useFocus/useFocusManager return an unexported type, useCursor an inline shape, useIsScreenReaderEnabled a bare boolean.
• vue-tui: public APIs are Vue composables (useFocus, useInput, ...). Where a composable's return type is exported under a name, the name always follows VueUse's UseXReturn convention (UseAppReturn, UseStdinReturn, UseStdoutReturn, UseStderrReturn, UseAnimationReturn, UseBoxMetricsReturn); the remaining composables return void, plain boolean, or small unexported inline shapes — never an XProps type. XProps is reserved for component props (BoxProps/TextProps, derived via ExtractPublicPropTypes).
• Options types follow the same principle: Ink names a composable's options type locally Options / Props and usually does not export it (e.g. useAnimation's Options is internal — only the return AnimationResult is exported, use-animation.ts:14,30). vue-tui exports each composable's options type under VueUse's UseXOptions name: UseInputOptions, UsePasteOptions, UseFocusOptions, UseAnimationOptions. useAnimation's options type originally shipped as AnimationOptions — the lone holdout — and was renamed to UseAnimationOptions (a hard rename, no alias) while the package is pre-1.0 (0.0.x, no stability promise yet). Maintainer decision (2026-06-13): export composable options types as UseXOptions; renamed AnimationOptions → UseAnimationOptions.
• Why: the public surface should read like Vue code: named composable return types get a single convention (UseXReturn) instead of Ink's mix of XProps, result names, and bare names, and XProps keeps its Vue meaning (component props). Return shapes still mirror Ink field-for-field where the same public state exists; reactive state is represented as refs for the model-implied reason documented above. Do not export Ink-compatible alias names for these types: Vue-first naming is more important than making type imports look portable across React and Vue. Maintainer decision (2026-06-07): KEEP.

Function-valued composable inputs use MaybeRef, not getters

• Ink/React: useInput and usePaste use React's current-props model: a hook can keep a stable event listener and still call the latest handler after a re-render.
• vue-tui: setup() runs once, so passing a function prop's current value directly (useInput(props.onInput)) captures a one-time snapshot. When a composable should follow a function-valued prop, pass a live prop ref instead: useInput(toRef(props, "onInput")) / usePaste(toRef(props, "onPaste")). A wrapper closure that reads props.onInput(...) at event time is also correct.
• Why: this is Vue's standard reactive-source boundary: pass the source, not a value read from it in setup. The handler parameter accepts MaybeRef<Handler> and resolves it with unref() when input/paste occurs. It deliberately does not accept MaybeRefOrGetter<Handler> because a handler is itself a function: useInput(() => {}) must remain an input handler, not be reinterpreted as a getter that returns one. Maintainer decision (2026-06-06): KEEP.

<Static> uses a scoped slot object instead of positional render arguments

• Ink/React: <Static> receives a function-as-children render callback and calls it as render(item, absoluteIndex).
• vue-tui: <Static> exposes a Vue scoped slot with { item, index }, so template users write v-slot="{ item, index }" and TSX users pass a slot function that receives one props object.
• Why: this is the framework-native match for React render children. Vue scoped slots pass one props object, not multiple positional arguments, and that object form is what Vue users expect for slot payloads. The rendered item/index values remain equivalent. Maintainer decision (2026-06-06): KEEP.

ARIA props are typed camelCase; kebab still works but is not type-checked

Full design, type-safety findings, and precedent survey: accessibility-api.

• Ink: kebab string-literal prop keys ('aria-label', 'aria-hidden', 'aria-role' union, 'aria-state' object); JSX keys never camelize.
• vue-tui: the same vocabulary as typed camelCase props (ariaLabel/ariaHidden/ ariaRole/ariaState; AriaRole/AriaState exported, identical to Ink's). Ink's kebab still works at runtime (Vue camelizes onto the declared prop), so aria-role ports unchanged.
• Why (Vue idiom + reasonableness > parity — see "Why align to Ink"): Vue's prop-name-casing mandates camelCase, and — run-verified with tsc/vue-tsc — camelCase is the only spelling type-checked (value/typo/compound mistakes compile-error in both TSX and templates), while kebab aria-* is not (Vue/Volar treat it as a global attr). So ariaRole is the type-safe spelling and aria-role the runtime-only porting escape; the rejected kebab-only $attrs alternative loses typing + Boolean coercion for nothing the checker doesn't already give. Maintainer decision (2026-06-14): KEEP.
• Edges: a future compound aria word must be declared as its mechanical camelize (ariaHaspopup, not ariaHasPopup) or folded into ariaState; aria-hidden is modeled boolean (bare → true), but the string aria-hidden="false" wrongly hides (recorded edge).

Intentional Divergence Choices

These divergences are deliberate, but they are not strict supersets and are not primarily driven by Vue's framework model or API conventions. vue-tui intentionally chooses a different runtime behavior, ownership rule, or out-of-contract handling.

Non-Error thrown values: uniform show-the-error-and-reject

• Ink: accepts the throw, but its handling is non-uniform (run-verified vs v7.0.4): for a truthy non-Error it renders an ErrorOverview showing a string .message (blank for a bare string, since 'boom'.message is undefined) and RESOLVES waitUntilExit() with the raw thrown value — a throw looks like a clean exit. For a falsy throw (0 / '' / null) it renders no overview and leaves waitUntilExit() PENDING (recoverable — a later unmount resolves it with undefined).
• vue-tui: any thrown value renders an ErrorOverview (message = a string .message if present, else String(value)) and REJECTS waitUntilExit() with an Error whose .message EQUALS the displayed message — one messageForNonError(value) helper feeds both the overview header and the reject-wrap, so display and reject can't drift (e17). No synthetic stack (a value with no .stack renders only the header).
• Why: aligning to Ink reduces bugs only where Ink is correct. Ink resolving the exit promise with a thrown value, and silently hanging on a falsy throw, are abnormal, so vue-tui deliberately diverges to one uniform contract: show the error, reject the exit. Same recover-vs-crash family as the invalid-input-validation divergence. Showing a real message (string .message else String(value)) is useful for the lint-discouraged non-Error throw, and matching the rejected message to it removes a confusing internal inconsistency (throw {message:'x'} once displayed x but rejected [object Object]). Introduced 2026-05-31; consistency fixed 2026-06-12. Maintainer decision (2026-06-12): KEEP.

Second mount() on a live stdout is an inert no-op

• Ink: render() keeps one instance per stdout (WeakMap<WriteStream, Ink>); a second render(node, {stdout}) on a stream that already has a live instance warns on stderr but reuses that instance and rerenders the new tree into it.
• vue-tui: a second mount() on a still-live stdout warns on stderr, wires no second renderer, renders nothing, and returns an empty placeholder object (the real controls — unmount(), waitUntilExit() — live on the app, not on mount()'s return value). The first app's tree stays on screen. The skip is scoped to that one guarded call — derived from what the app actually wired, never sticky: a guarded different app's unmount() settles only its own exit promise and never touches the owner's stream or registry entry; the owner double-firing mount() on its own stdout keeps a fully working unmount() (the warning's recovery path); an app that once hit the guard can later mount — and cleanly unmount — on a free stdout; and a live app that merely targeted another app's busy stream stays fully killable.
• Why: a second mount() on a live stdout is a misuse (forgot to unmount(), a re-render glitch fired mount() twice, or expecting mount() to re-render — it doesn't; update reactive state for that). Ink treats it as unsupported and warns too. vue-tui fails safe: it ignores the second mount, keeps the live app rendering, and warns with the two recovery paths. It deliberately doesn't copy Ink's reuse-and-rerender: there's no clean public path to it (createApp binds the tree to the app, so an Ink-style rerender would mean reaching into the live app's container or tearing it down first), and on a misuse path keeping the running app stable beats auto-tearing it down (which would churn on a re-render glitch). Maintainer decision (2026-06-04): KEEP. Test: instance-reuse-guard.test.tsx.

Raw mode is owned for the interactive lifetime by default (rawMode option)

• Ink: raw mode is lazy / reference-counted to input hooks. useInput / useFocus / usePaste enable it on mount and release it when the last one unmounts, so a screen with no input handler falls back to cooked mode. There is no option to hold it.
• vue-tui: the rawMode mount option defaults to 'always'. Raw mode is enabled at mount and held for the whole interactive run (when interactive and stdin is a TTY), regardless of which input composables are mounted. rawMode: 'auto' opts back into Ink's exact lazy behavior.
• Why: for a long-running interactive app (a full-screen TUI, a coding agent), Ink's lazy model makes raw mode toggle as the user moves between input and no-input screens. The main consequence is echo: on a no-input / streaming screen the terminal is back in cooked mode, so typed keys echo into the half-drawn frame (and line-buffer). Ctrl+C also changes path: on a no-input screen it is a kernel SIGINT rather than the app's own \x03 intercept. Note exitOnCtrlC defaults to true in both Ink and vue-tui, so by default Ctrl+C exits either way; the divergence is only the exit path/code (a clean exit 0 vs a re-raised SIGINT 130). It matters for an app that sets exitOnCtrlC: false to handle Ctrl+C itself: under the lazy model its opt-out is bypassed on a no-input screen (the SIGINT still exits). Holding raw for the lifetime keeps echo and Ctrl+C handling identical on every screen. This matches the cross-framework norm: Bubble Tea, Textual, Ratatui, and prompt_toolkit all own the terminal for the program lifetime. Ink's hook-driven model differs: its "cooked on a no-input screen" behavior follows from refcounting input hooks rather than from an explicit no-input-screen contract.
• Consequence: owning raw mode ref()s stdin, so an 'always' app stays alive until you explicitly unmount() / exit(). It does not auto-exit when idle (the same way an Ink app holding a useInput already does not). The "render and auto-exit" pattern (Ink's inline-output use) is rawMode: 'auto'. Tests: raw-mode-lifecycle.test.tsx ('always' holds raw with no input hook; 'auto' stays cooked; no mid-session oscillation). Maintainer decision (2026-06-13): KEEP.

useCursor() re-asserts the declared caret every commit (persistent declaration)

• Ink: useCursor's no-deps useInsertionEffect (use-cursor.ts:27-32) re-marks the cursor dirty only when the cursor's React component re-renders, and log-update emits the caret only on a dirty commit. React's render cascade re-renders the child on an ancestor-driven commit (so Ink re-asserts there), but on an unrelated sibling/leaf repaint the cursor component does not re-render — Ink does not re-assert and the caret is dropped, zombieing to the bottom-left corner (run-verified vs v7.0.4: a sibling spinner tick ends …> hello\n with no caret suffix, leaving the caret at row 2 col 0).
• vue-tui: the runtime re-emits the last-declared caret at the end of every commit (until the declaration changes or is cleared), so a focused input's caret stays at its edit point across unrelated repaints (spinner / log line / progress bar) in all component topologies. The hide-before-erase / show-at-resolved-position flicker discipline is preserved (no corner streak), and the re-emitted position is clamped to the visible region (y to the line count, x to the width) so a post-resize/shrink stale coordinate never moves out of range.
• Why: this is a deliberate divergence FROM Ink toward correct terminal-app behavior, not Ink-alignment. Real terminal programs that own an edit point re-place the caret at that point every frame (vim emits \e[<row>;<col>H after each repaint; readline re-lands the buffer offset on SIGWINCH; nano homes to its edit cell) — they never leave the caret where the repaint dragged it. Aligning to Ink exists to reduce bugs, not to preserve abnormal behavior; matching Ink's topology-conditional zombie would preserve abnormal behavior, so vue-tui diverges. Per the classification flow this is not a Model-Implied difference (Vue is not forced here — fine-grained reactivity could also be made to re-run the child; the runtime simply chooses to be more correct than Ink at the commit level), and it is not Vue-API-shaped, so it lands in Intentional Divergence Choices. The {x,y} setCursorPosition surface is unchanged — it remains the right low-level IME primitive (a composing glyph offset deliberately decoupled from the buffer point); the fix is an internal per-commit re-emit, so the public API stays compatible. A cleared declaration (setCursorPosition(undefined), e.g. useCursor's onScopeDispose on unmount) re-emits no caret, so teardown still ends with the cursor shown and handed back (\x1b[?25h); the persistent re-emit runs before the unmount clear, so it cannot resurrect a torn-down caret. The one accepted residue is a stale-but-in-range absolute position if an app declares a fixed {x,y} and then shrinks content without re-declaring (it parks at a plausible spot, strictly better than a corner-zombie); a future Stage 2 focus-owned, content-tracking caret (recomputed from the focused widget's layout each frame) would dissolve that residue. Tests: PTY cursor-sibling-repaint.test.ts (a sibling-topology spinner tick re-asserts the caret, not the corner) and unit frame-writer.test.ts (a non-dirty changed-output re-render re-emits the declared suffix; D5 clamp). Maintainer decision (2026-06-12): OVERRIDE prior KEEP — adopt per-commit re-assert. The prior KEEP (2026-06-01) had kept the reactivity-tied behavior to avoid diverging from Ink in the sibling direction; that rationale was overturned when running real terminal apps showed Ink itself zombies the caret there, so matching Ink was matching a defect, not parity.

Resize unconditionally cancels the pending trailing commit

• Ink: resized() paints synchronously via onRender() but does not cancel a pending throttled onRender; when that trailing commit re-runs and shouldClearTerminalForFrame clears (because the previous frame overflowed), Ink emits a second clearTerminal.
• vue-tui: onResize calls scheduler.cancel() as its first, unconditional step on every resize (not only narrowing ones), dropping any pending throttled commit before its synchronous commit, so the redundant trailing clear never fires.
• Why: the synchronous resize commit already reflects the current tree, so the pending commit would repeat the same clear. The dedup is triggered by an overflowing frame plus a pending commit — not by narrowing specifically; a widening resize cancels the pending commit too. Emitting one clear instead of two has no visible behavior difference (issue #26). The separate narrowing-only writer.clear() frame reset is a distinct mechanism.

Degenerate boxes do not lay out or paint children when the content area is gone

• Ink: its size model is border-box-like: width/height are handed to Yoga as the element's outer size, while border and padding consume space inside that size. Ink has no boxSizing / content-box prop. During paint, though, Ink only clips children when overflow is hidden. With default visible overflow, children can leak into border rows or outside the box when border/padding squeeze the content area to zero, and a bare width={0} Box can still let zero-width text wrapping create extra visible rows (B\nA beside a sibling). Examples in Ink v7.0.4 include: width={3} height={2} borderStyle="single" painting the child on the bottom border row; width={2} height={3} leaking the child past the right border; width={4} height={3} paddingX={1} borderStyle="single" leaking into the bottom border; and bare width={0} text reserving rows through wrap-ansi's width-0 layout.
• vue-tui: layout computes each Box's inner content size by subtracting computed border and padding from the outer box size, clamps it to {width >= 0, height >= 0}, and temporarily removes that Box's in-flow yoga children from the layout when either dimension is zero. Paint applies the same inner-content gate, so the in-flow child subtree neither reserves invisible rows nor writes glyphs outside a nonexistent content area. Border and background are still painted as far as the outer area permits. Absolutely-positioned children are exempt — their containing block is the padding box (inside the borders), not the content rect, so they still lay out and paint (clipped only by overflow:hidden), matching Ink. Positive-size content areas keep the existing overflow behavior; this is not a blanket overflow:hidden.
• Layout model guidance: primitive Box should preserve the Yoga/flexbox model rather than paper over it with ad-hoc layout corrections. Defaults such as flexShrink: 1 are part of that model, and a child resolving to zero width or height can be a valid layout result. Higher-level components are where stronger user intent belongs: scroll, list, and viewport abstractions should keep their content at natural size (flexShrink: 0, or an equivalent encapsulated default) and let a bounded viewport clip or offset what is visible. Paint containment is the renderer invariant underneath both cases: whatever Yoga resolves, in-flow children may only paint inside their owning Box's content rectangle; if that rectangle has no positive width or height, the in-flow child subtree does not paint. Absolutely-positioned children are the exception — they paint against the containing block and are suppressed only by overflow:hidden, matching Ink.
• Why: children need a real content rectangle to lay out and paint into. If the resolved content width or height is zero, rendering child text or nested borders on top of the frame, outside the box, or on later rows is an implementation artifact, not useful output. This follows the common TUI box model: Ratatui renders child widgets into an inner Rect, Textual reduces content space from the assigned box, and Rich panels render children with child width/height after subtracting the border. Bubble Tea's viewport follows the same separation at the component level: content keeps its natural size while the viewport exposes a bounded visible window and offsets. The behavior also prevents negative repeat/count math and paint crashes in tiny legal boxes. Maintainer decision (2026-06-07): KEEP. Tests: text-wrap-width.test.tsx, flex.test.tsx, text.test.tsx, absolute-in-degenerate-box.test.tsx.
• Future content-box: this does not block adding an explicit content-box option later. That option would change how a requested size is expanded into an outer box size before layout. Once an outer box exists, the paint invariant remains the same: a child subtree only lays out and paints when the resolved inner content rectangle has positive width and height. The default remains border-box-like, matching Ink's current public sizing model.

Out-of-type style values are forwarded, not defensively coerced

• Ink: several flex/align setters coerce an invalid runtime value to a default: flexShrink non-number -> 1; alignItems/alignSelf/alignContent/ justifyContent falsy ("") -> their default (STRETCH / AUTO / FLEX_START); and an out-of-set value matches none of Ink's if-chain branches, so no setter runs and the previous/default value persists.
• vue-tui: these setters trust the typed prop surface and forward the raw value to yoga: a non-number flexShrink is passed through; toAlign("")/toJustify("") look up "" and pass undefined to the setter; and out-of-set values that yoga happens to accept (space-*/auto on alignItems) reach yoga rather than being ignored.
• Why: every one of these is reachable only via a TS-bypass. The public prop types forbid them. Within the typed contract Ink and vue-tui are identical. Ink's per-value coercion is defensive code for runtime values vue-tui's types already exclude. Duplicating those typeof/falsy guards would add checks for inputs the public types reject. (flexGrow is not in this set: both only coerce null/undefined -> 0.) If a reviewer shows any case is reachable in-type, it becomes a bug to fix, not a divergence.

Composables throw outside a render tree

• Ink: the hooks read a React context whose default value is a no-op object, so calling e.g. useStdin() outside an Ink tree returns inert defaults without an error.
• vue-tui: useApp, useStdout, useStderr, useStdin, useWindowSize, useFocus, useFocusManager, useInput, usePaste, useCursor, and useIsScreenReaderEnabled throw when their context is absent ("... must be called inside a vue-tui render tree"). useBoxMetrics and useAnimation do not throw: they fall back. useBoxMetrics reports zero metrics, and useAnimation drives a standalone scheduler. See the additive entry.
• Why: a composable used in the wrong place is a bug, and a thrown error names it at the call site instead of returning a context that quietly does nothing. The two exceptions fall back because they have a meaningful standalone behavior (zero metrics / a working animation), so throwing would remove a useful capability. Required app, terminal, focus, and input context should fail fast when absent; no-op defaults hide bugs. Maintainer decision (2026-06-07): KEEP.

Invalid input is validated at the component layer, not the paint layer

• Principle: vue-tui validates the covered invalid render inputs — a chalk-modifier backgroundColor like "bold", a foreground color key that exists on chalk but is not callable like "level", and an unknown borderStyle — at the component-render layer (box-validate.ts for <Box>, text.vue for <Text>), not down at the paint layer. A bad value therefore throws where the error boundary catches it -> ErrorOverview -> a clean reject of waitUntilExit(), exactly like any other component error. The app reports the error instead of crashing.
• Ink: validates the same covered inputs lazily at paint (colorize / render-border, run from the reconciler's commit hook): outside React's ErrorBoundary, so a bad value is an uncaught crash, not a recoverable error.
• Why: the key constraint is where paint runs. vue-tui's paint runs in a Vue post-flush callback (queuePostFlushCb, decoupled from render), so a throw there escapes onErrorCaptured and wedges the scheduler. Unlike a component error, it cannot be made recoverable. The escape itself is symmetric, not a Vue weakness: a component error boundary (React ErrorBoundary; vue-tui's onErrorCaptured wrapper) covers framework-managed component work, never the renderer's paint callbacks, so a paint-layer throw is uncatchable in both engines. Validating in Box / Text keeps a bad value on that boundary-driven recoverable path; Ink's paint-time check can only crash. This is a deliberately chosen fail-safe given a constraint that is symmetric across React and Vue (recover-vs-crash), not a Vue-model-forced difference — hence an intentional choice rather than a model-implied one. Provenance: the wedge claim rests on the earlier paint-throw investigation; the 2026-06-12 audit could not reach a paint throw from public or raw-host input (paint's if (!chars) return; border fallback intercepts an invalid borderStyle that bypasses component validation), so it stands on that prior record, not an in-audit reproduction.
• Cost: the component-layer check is eager (no paint-time layout/squash info), so it over-throws in a few degenerate, invalid-input-only cases Ink never reaches. For the covered public inputs in normal reachable cases, both libraries error; only the channel (recoverable reject vs crash) differs. Maintainer decision (2026-06-07): KEEP. vue-tui makes the more reliable library choice here: reject the same invalid input with a recoverable, prop-specific error instead of preserving Ink's lower-level paint crash and chalk implementation message. Tests: background-color.test.tsx, plus the borderStyle validation tests.
• Text validates regardless of content (2026-06-13, user-blessed): <Text> now validates color and backgroundColor on every render, not only when its content is non-empty — matching <Box>, which already validates its own colors unconditionally. Ink does not throw for empty text only because its colorize call is lazy (an incidental implementation artifact, not a design choice); an invalid value is invalid regardless of content, and content-gated validation is a latent footgun. Principle: reasonable behavior over incidental Ink parity. The former wouldRenderNonEmptyText gate was removed. Screen-reader-hidden Text still returns before validation (matches Box).

Non-Behavioral Notes

These notes are not divergence entries. They document Vue-facing conventions or internal mechanics so they are not mistaken for parity gaps.

• The exported host-node type is TuiNode (TuiContainer | TuiTextLeaf | TuiComment, from @vue-tui/runtime/internal), not Ink's DOM-emulation DOMElement (nodeName / attributes / childNodes). vue-tui keeps a native host tree, so the exported type names that tree; measureElement and template refs accept it. The type rename itself does not imply a runtime behavior difference from Ink's DOM-emulation node.
• null / false / undefined / v-if="false" children are materialized by Vue as comment vnodes, which vue-tui's host renderer turns into an inert TuiComment: no yoga node, paints nothing, never shifts a sibling, and skipped when counting child positions (the child.type !== "comment" guards in paint.ts, text-measure.ts, and screen-reader.ts; G52). This is renderer mechanics, not a divergence entry by itself. The observable <Transform> literal-false edge is documented above as a model-implied divergence.
• Commit timing is deliberately Ink-aligned: leading+trailing throttle at Math.max(1, Math.ceil(1000/maxFps)) ms behind a maxFps > 0 guard (34ms at the default maxFps=30 — both engines compute exactly this), synchronous resize. The scheduler mirrors the observable timing of Ink's es-toolkit throttle (run-verified vs v7.0.4): the trailing timer re-arms on every deferred call, so the trailing commit fires at lastCall+wait (not windowStart+wait), and a call arriving a full window after the first deferral commits synchronously (es-toolkit's maxWait), keeping a ~wait cadence under sustained updates. This remains true even though re-renders come from Vue's fine-grained reactivity, not a React subtree re-render. One deliberate exception: resize cancels the pending trailing commit — see the divergence entry "Resize unconditionally cancels the pending trailing commit".