TUI core renderer — the append-only contract

What you are dealing with before you touch the rendering engine. This is the companion to tui-runtime-internals.md: that doc maps the flow (input → component tree → render); this doc explains the render contract, why it is shaped this way, and the invariants you must not violate. Scope is the core engine only:

• packages/tui/src/tui.ts — frame pipeline, commit ledger, window math, emitters, cursor placement.
• packages/tui/src/terminal.ts — ProcessTerminal, capability probes, private-CSI reassembly.
• packages/tui/src/terminal-capabilities.ts — TERMINAL profile, sync-output / DECCARA / image detection.
• packages/tui/src/stdin-buffer.ts — escape-sequence reassembly.
• packages/tui/src/utils.ts — width/slice/wrap (the width model).
• packages/tui/src/kitty-graphics.ts + components/image.ts — inline images.
• packages/tui/src/deccara.ts — rectangular-fill optimizer.

Application-layer renderers (transcript, tool calls, session tree, editor, widgets) are out of scope — they live in packages/coding-agent. The one app-layer file that is load-bearing for this contract is transcript-container.ts, which implements the commit-boundary seam described below.

1. The one thing to understand first

The renderer cannot observe the terminal’s scroll position (ConPTY’s probe lies; POSIX has no API at all). The previous engine tried to guess when it was safe to rewrite native scrollback, and every policy choice over that unobservable variable traded one failure family for another (yank ↔ flash ↔ corruption ↔ invisible-until-resize — see the git history of this file for the full war journal). The current engine removes the guess entirely: native scrollback is append-only.

We keep the transcript on the normal screen (native scrollback, native selection, transcript persists after exit). The engine maintains one ledger:

• committedRows (C) — frame rows [0, C) have been physically scrolled into terminal history. They are immutable: the engine never rewrites them, and components must never change them.
• windowTopRow (W) — the frame row mapped to grid row 0. The visible window is frame rows [W, W + height), repainted in place with relative cursor moves.
• commit boundary — reported by the component tree per frame (NativeScrollbackLiveRegion) as two nested ends:
  • byte-stable end (B) — commitSafeEnd ?? liveRegionStart ?? frame.length. Rows below B are asserted never to re-layout and stay under the committed-prefix audit.
  • durable end (D) — max(B, snapshotSafeEnd ?? B). Rows in [B, D) may still drift bytes later (a streaming markdown table re-aligning columns) but are durable — their current snapshot is permanent content, so dropping them when they scroll off is forbidden. They commit audit-exempt: later drift becomes a frozen stale row in history, never a re-anchor.

Per ordinary frame: W = max(C, L − height), C' = max(C, min(D, W)), and the only bytes that ever touch history are the chunk frame[C, C') written at the scrollback seam. The engine also tracks auditRows (A ≤ C) — the byte-stable leading prefix [0, A); the committed-prefix audit (§2) samples only that prefix, so the durable suffix [A, C) drifting never triggers a re-anchor. Scrollback therefore equals frame[0..C) — every row exactly once, in order, with its content at commit time. There is nothing to guess, nothing to defer, and nothing to reconcile: the scroll position is irrelevant because ordinary updates never rewrite anything a scrolled reader could be looking at.

What this costs (the accepted tradeoffs)

• A block that has scrolled past the window top cannot reflow in place. A byte-stable block stays in the live region (below B) until final; a durable block (below D) commits its scroll-off snapshot, so a late layout change of an already-committed row is a frozen stale row in history (duplication never loss), not a dropped row.
• A component tree that reports no seam gets shell semantics: whatever scrolls off is final. Shrinking such a frame into its committed prefix re-anchors the window and leaves the stale copy in history (§3).
• Inside ED3-unsafe multiplexers, a width change terminates the physical-row coordinate epoch. The renderer captures an opaque NativeScrollbackWidthEpoch marker from the last emitted source state before SIGWINCH, then resolves that same logical boundary after the settled-width render. Host-reflowed history stays immutable. Output queued during settlement is emitted only from the resolved old boundary to the current source boundary at the terminal-owned viewport bottom; the settled viewport then repaints in place. No old-width and new-width row counts are compared, and no old viewport row is recommitted. Components without the source contract retain the conservative physical-row fallback. Visible overlays freeze the seam and pinned live regions clip advancement at their final boundary. Height-only resizes retain the existing ledger.
• Direct HerdR panes are not in that category. HerdR’s Ghostty core implements ED3, so OMP clears and replays its source-owned transcript after settlement, matching direct terminals. Repainting host-reflowed rows in place would harden soft wraps and leave residual rows after a later width reversal.

2. The frame pipeline (what you are editing)

#doRender per frame:

1. Compose the frame, collecting the first root child’s getNativeScrollbackLiveRegionStart() and optional pinning policy.
2. Audit the committed exact prefix (findCommittedPrefixResync, skipped on geometry frames). The detector samples the prefix tail (up to 8 non-blank rows in the last 24, SGR-stripped). A single in-place mismatch is accepted as stale history; a structural shift re-anchors at the first changed row, favoring duplication over content loss. An in-place width change does not audit or re-slice the prior epoch’s physical coordinates; it resolves the captured logical source marker in the settled-width frame.
3. Classify the frame as a gesture-driven full paint, an opt-in divergence rebuild, or an ordinary update and calculate the window/commit chunk. Overlays freeze commits. A pinned live region clips its offscreen mutable suffix instead of snapshotting it.
4. Extract cursor markers, prepare width-safe lines, slice the window, and composite overlays into the screen-coordinate window only.
5. Emit:

ED3 (CSI 3 J) is emitted in exactly one place — #emitFullPaint({ clearScrollback: true }). The normal callers are explicit user gestures: session replace/branch/resume (requestRender(true, { clearScrollback: true })), resize outside a multiplexer, and resetDisplay() (the display-reset chord, Alt+L by default). It clears native history without ED2 first; the replay overwrites every row from home so terminals without synchronized output do not expose a blank viewport. A gesture pins the user to the tail, so the history snap is acceptable.

The second caller is an ordinary-render divergence when tui.scrollbackRebuild is enabled: if the committed prefix structurally resynchronizes or the current frame collapses into committed rows, the renderer clears and replays the current frame to replace stale preview history with the final form. This path is disabled by default and never runs after the first paint, during an explicit replacement/geometry frame, or inside a multiplexer. Multiplexers never get ED3 (it is a no-op there and a replay would duplicate pane history). The ordinary update path never emits ED2/ED3 or an absolute cursor home — several terminal families snap a scrolled reader to the bottom on those.

The commit-boundary seam (the load-bearing app contract)

NativeScrollbackLiveRegion has one boundary and one optional policy:

• getNativeScrollbackLiveRegionStart() returns the first local row that may still mutate. Rows before it are declared byte-stable at the current width.
• isNativeScrollbackLiveRegionPinned() keeps the mutable suffix viewport-local rather than recording frozen snapshots as it scrolls off. This is for replacing dashboards, not append-shaped transcript content.
• Reporting no seam gives shell semantics: rows commit as they scroll.

When multiple root children report a seam, the topmost seam wins because commits are prefix-only. NativeScrollbackCommittedRows lets containers pass the committed count down to children, and NativeScrollbackReplay lets components release layout locks before a destructive replay.

NativeScrollbackWidthEpoch is the cross-width source contract. Capture reads only state that produced the last emitted frame. Resolve projects that source boundary into the newly rendered width, while the current-boundary method identifies the logical suffix queued during settlement. Containers propagate the marker through nested sources; Markdown snapshots its last rendered source text, so a streaming update received before SIGWINCH cannot masquerade as already-emitted output.

TranscriptContainer implements the application seam. It scans for the first unfinalized transcript block. Finalized blocks before it are exact; that live block may extend the exact boundary through getTranscriptBlockSettledRows(). Assistant messages derive those settled rows from completed content blocks and markdown’s frozen-token prefix, while constructs that can re-layout asynchronously (for example Mermaid) defer settling. Pinning is propagated from the first live block; tool execution uses it for replacing preview/dashboard states.

Transcript assembly also reports RenderStablePrefix: unchanged component array references at unchanged offsets let the engine skip work over the byte-identical prefix. Components that discard or lock committed material must honor the committed-row and replay hooks. Freezing/settling is a correctness contract, not a terminal-specific optimization. —

3. Invariants — MUST / NEVER

1. NEVER add a new CSI 3 J (ED3) callsite. ED3 flows only through #emitFullPaint({ clearScrollback: true }), only for gestures, never inside multiplexers.
2. NEVER rewrite a committed row. No emitter may touch frame rows < C, and W ≥ C always (re-showing a committed row on the grid duplicates it for a scrolling reader — the historical corruption family). When a component violates immutability, the audit (§2) degrades to duplication — never silently skip rows, never erase history.
3. Commits are exactly the chunk. Any byte shape that scrolls the screen must scroll only rows accounted for by the commit advance.
4. An ED3-unsafe multiplexer width resize NEVER advances history. The old committed physical-row coordinate is opaque after reflow. The resize leaves the host-reflowed viewport in place and establishes a complete-frame baseline independent of the native commit count. Subsequent growth writes the exact current-width rows newly crossing the seam—not blank scroll commands—then repaints the bounded viewport; only that slice advances commits. Visible overlays advance neither the baseline nor the seam ledger; overlay exit backfills the exact hidden slice. Pinned live regions advance only through their final boundary; finalization releases the deferred mutable slice. During a height shrink, only occupied old-frame rows actually moved into history by the host are excluded from the append-owned seam; empty viewport rows do not consume content-driven movement. Height-only resizes do not terminate the epoch. Direct HerdR uses ED3 source replay instead.
5. NEVER probe the viewport position or fork on platform in the update path. win32 behaves like POSIX. The probe APIs are gone; do not reintroduce them.
6. Only declare rows exact when their bytes are stable. Mutable transcript content may commit as an unpinned frozen snapshot, but rows before the seam remain under the exact-prefix audit.
7. Park the hardware cursor at real content bottom, not the padded window bottom, or height shrinks scroll live rows into history and duplicate them per resize step.
8. Cursor writes live inside the synchronized-output frame, before ESU — never as a second frame after it.
9. NEVER throw in the render hot path. Clamp over-wide lines (truncateToWidth); a width mismatch is cosmetic, not fatal.
10. Multiplexers get no destructive clear and no history rewrap on resize — repaint the window in place; pane history keeps its old wrap.
11. Any change to the ledger math, the emitters, or the seam must be validated by the stress harness (§6) across its full scenario matrix, not by a single-terminal smoke test.

4. Terminal capability detection

TERMINAL (terminal-capabilities.ts) is resolved once at import from TERMINAL_ID plus environment sniffing; detection helpers are pure over (env, platform) and unit-testable.

• shouldEnableSynchronizedOutputByDefault(env, id) → DEC 2026 default. Precedence: user opt-out (PI_NO_SYNC_OUTPUT/PI_TUI_SYNC_OUTPUT=0) → user force-on (PI_FORCE_SYNC_OUTPUT=1/PI_TUI_SYNC_OUTPUT=1) → TERM_FEATURES advertises Sy → WT_SESSION → known direct terminals → off for risky multiplexers and unknowns. Reconciled at runtime by the DECRQM mode-2026 report; a user override still wins.
• detectRectangularSgrSupport(id, env) → DECCARA fills: kitty only, off in multiplexers and under PI_NO_DECCARA.
• supportsScreenToScrollback → kitty’s ED22 (used once, on the initial paint, to preserve the pre-existing shell screen).

The old ED3-risk classifier (eagerEraseScrollbackRisk, PI_TUI_ED3_SAFE, submitPinsViewportToTail) is gone: behavior no longer depends on which terminal is rendering, so there is no risk class to detect. Env sniffing now only selects optimizations (sync output, DECCARA, images), where a miss is cosmetic, not corrupting.

5. Width model

visibleWidth / truncateToWidth / sliceByColumn / wrapTextWithAnsi (utils.ts) all agree on one UAX#11 width model. Slicing, truncation, wrapping, and segment extraction run on the native engine (@musepi/pi-natives, Rust unicode-width); visibleWidth measures with Bun.stringWidth pinned to that same model (STRING_WIDTH_OPTS: countAnsiEscapeCodes: false, ambiguousIsNarrow: true) — a JSC builtin that shares the native width tables without the per-call N-API box the native scanner traps on under Bun 1.3.x. The two must never disagree; mixing unpinned width models in measure-vs-slice produced crashes.

• Fast path: printable ASCII is one cell per code unit.
• Anything past the ASCII prefix measures through Bun.stringWidth (CSI/OSC stripped to zero); tabs are added back at the fixed DEFAULT_TAB_WIDTH columns.
• OSC 66 sized spans are added back as scale × (explicit w ?? payload width) — Bun.stringWidth would otherwise strip the whole span to zero.

Rule: any new measuring code routes through these helpers, and the hot path clamps instead of throwing. Known residual: combining-heavy scripts (Arabic harakat) survive painting verbatim, but ghostty-web’s cell readback can migrate non-spacing marks across cells — the stress harness compares those rows with marks stripped (sameLinesAllowingMarkDrift).

6. The fidelity gate (use it)

packages/tui/test/render-stress-harness.ts drives the renderer’s real emitted ANSI into a ghostty-web VirtualTerminal across randomized op sequences and parameterized terminal shapes, and validates the contract with a shadow commit ledger: an independent reimplementation of §1’s math, fed only by observed frames (a render wrap) and observed bytes (a write wrap). Per op it asserts:

• the whole tape (scrollback + grid) equals shadowTape + window slice, row for row, including across resizes;
• scrolled readers stay pinned and visible history rows are never rewritten;
• multiplexer pane history grows by exactly the committed chunk;
• sync-output/autowrap bracket discipline, cursor parking, background columns, duplicate accounting.

Run it — plus render-regressions.test.ts, streaming-scrollback-defer.test.ts, and the issue-*-repro.test.ts files — before changing ledger math, emitters, or the seam. A change that passes one terminal and one seed is not verified.

7. Capability probes & stdin reassembly

ProcessTerminal fuses capability queries with a bare DA1 (CSI c) sentinel so a non-answering terminal is detected when DA1 returns first. Replies can arrive split across a stdin flush, so:

• #privateCsiResponseBuffer accumulates \x1b[?… partials while a sentinel is outstanding, rejoins on the terminator byte, then runs the handlers on the complete reply. A new \x1b mid-reassembly or >256 bytes abandons the partial so real keys still reach input.
• #da1SentinelOwners is a typed FIFO discriminated by kind so a keyboard DA1 cannot be mistaken for an OSC 11 / DECRQM / graphics-probe sentinel.
• DECRQM probes (2026/2048/2031) drive runtime feature gating.

Rule: any new probe must own a typed sentinel and survive a split reply (feed the reply byte-by-byte in a test and assert nothing leaks to input).

8. Inline images & memory

Kitty images are transmit-once, place-many (kitty-graphics.ts). ImageBudget keeps only the most-recent N images live; when the cap is exceeded the demoted image’s pixels are deleted by id (a=d,d=I) and its visible rows re-render as the text fallback through the ordinary window diff — no destructive replay. A demoted placement already committed to history simply loses its pixels (committed rows are immutable), and the text fallback is height-preserving once a graphic has rendered (reserved rows + fallback line), so demotion never shrinks the block and never shifts committed content below it.

Rule: never re-emit full base64 per frame. Kitty Unicode placeholders are default-on only for kitty/ghostty (PI_NO_KITTY_PLACEHOLDERS / PI_KITTY_PLACEHOLDERS).

9. Escape hatches (env vars)

Removed with the old engine: PI_TUI_ED3_SAFE (no ED3-risk lever exists), PI_CLEAR_ON_SHRINK (shrinks always clear exactly), PI_TUI_DEBUG (per-render dump superseded by PI_DEBUG_REDRAW ledger logging and the stress harness replay/reduce tooling).

10. Before you touch the render core — checklist

• Are you about to emit CSI 3 J anywhere other than the gesture-driven clearScrollback full paint? Stop.
• Could any code path rewrite, or re-show on the grid, a frame row below committedRows? Stop.
• Does your byte shape scroll rows that are not the commit chunk? That breaks scrollback == frame[0..C).
• Are you adding a viewport probe, a platform fork, or a terminal-brand branch to the update path? The contract exists so none are needed.
• New mutable UI above the editor? It must report (or live inside) the live-region seam, or it will freeze at first commit.
• Did you run the stress harness and the repro suite across the full scenario matrix — not just one terminal and one seed?
• New probe? Typed sentinel owner + split-reply test.
• New width path? Routed through the shared native engine, clamped (never thrown) in the hot path.