diff --git a/CLAUDE.md b/CLAUDE.md
index 03898844..89490918 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -286,7 +286,7 @@ Codeman is a Claude Code session manager with web interface and autonomous Ralph
 
 **Ctrl+V paste trap** (`image-input.js`): `Ctrl+V` routes through `_handleImagePaste()`, which focuses a hidden `contenteditable` trap and reads the clipboard from the paste event landing there; images upload and their paths are typed in, text goes through `terminal.paste()` so bracketed-paste markers survive. ⚠️ **The trap must consume exactly ONE paste event** (Firefox delivers two per keypress: the `execCommand('paste')` event and the keydown's default action); the one-shot flag lives on the trap, never on a browser check. ⚠️ Do not remove the `execCommand('paste')` call: on some mobile engines it is the only route into the trap, and the trap is the only place image blobs are read. Tests: `test/image-paste-trap.test.ts`. → [architecture-invariants#terminal-paste-ctrlv](docs/architecture-invariants.md#terminal-paste-ctrlv)
 
-**Terminal scrollback strip + wheel/touch forwarding**: codex/claude/gemini get the FULL strip (alt-screen, `3J`, mouse DECSETs); tmux-backed opencode gets the MIDDLE strip (alt-screen toggles + mouse DECSETs, `3J` kept); every other tmux-backed mode (shell/antigravity/pi/grok/deepseek/omp) gets the NARROW strip (alt-screen toggles only). Table: `CliCapabilities.altScreen` JSDoc, pinned in test/claude-scrollback-strip.test.ts. ⚠️ Gated on `useMux`: direct-PTY sessions must keep the alt screen. Wheel and touch forward to the CLI for **claude ≥ 2.1.187 ONLY, and only while it has mouse tracking on** (`cliMouseTracking`: fullscreen claude sets it, its default inline renderer does not and scrolls locally like codex); ⚠️ never re-add codex without a fresh measurement (it ignores SGR wheel reports). ⚠️ `getClaudeCliVersion()` must never cache a FAILED probe. ⚠️ Hand-report clicks only while the CLI has mouse tracking on: `_shouldReportMouseToCli()` gates all three report sites on `cliMouseTracking` (from `_recordStrippedMouseMode()`, session.ts), or a plain shell prints the reports as literal text. ⚠️ A `TerminalTile` (grid tile, split Pane B) pages the wheel and reports clicks through these SAME gates with itself as the target (its terminal, its session, never `activeSessionId`); never copy the mode lists into terminal-tile.js. Read `_logScrollRouting()` before diagnosing a scroll report. → [architecture-invariants#terminal-scrollback-strip-flavors-and-wheeltouch-forwarding](docs/architecture-invariants.md#terminal-scrollback-strip-flavors-and-wheeltouch-forwarding)
+**Terminal scrollback strip + wheel/touch forwarding**: codex/claude/gemini get the FULL strip (alt-screen, `3J`, mouse DECSETs); tmux-backed opencode gets the MIDDLE strip (alt-screen toggles + mouse DECSETs, `3J` kept); every other tmux-backed mode (shell/antigravity/pi/grok/deepseek/omp) gets the NARROW strip (alt-screen toggles only). Table: `CliCapabilities.altScreen` JSDoc, pinned in test/claude-scrollback-strip.test.ts. ⚠️ Gated on `useMux`: direct-PTY sessions must keep the alt screen. Wheel and touch forward to the CLI for **claude ≥ 2.1.187 ONLY, and only while it has mouse tracking on** (`cliMouseTracking`: fullscreen claude sets it, its default inline renderer does not and scrolls locally like codex); ⚠️ never re-add codex without a fresh measurement (it ignores SGR wheel reports). ⚠️ `getClaudeCliVersion()` must never cache a FAILED probe. ⚠️ Hand-report clicks only while the CLI has mouse tracking on: `_shouldReportMouseToCli()` gates all three report sites on `cliMouseTracking` (from `_recordStrippedMouseMode()`, session.ts), or a plain shell prints the reports as literal text. ⚠️ A `TerminalTile` (grid tile, split Pane B) forwards the wheel as SGR reports, pages it and reports clicks through these SAME gates with itself as the target (its terminal, its session, never `activeSessionId`), and scrolls Shift+wheel itself (xterm makes it a horizontal no-op); never copy the mode lists into terminal-tile.js. Read `_logScrollRouting()` before diagnosing a scroll report. → [architecture-invariants#terminal-scrollback-strip-flavors-and-wheeltouch-forwarding](docs/architecture-invariants.md#terminal-scrollback-strip-flavors-and-wheeltouch-forwarding)
 **Detached start + service install**: `codeman web -d` relaunches the same entry script `detached:true` (setsid); `nohup` is not what makes it survive. ⚠️ Both `-d` and `service install` must REFUSE when a server is already up on this data dir (pidfile + `/api/status` probe), or a second instance attaches to the first one's live sessions. ⚠️ Never report success not observed: poll `/api/status` until the child answers or dies. `--stop` must verify the pid still looks like Codeman (`ps -o command=`) before signalling. Unit/label names live only in `config/service-names.ts`. `service install` bakes the installing shell's PATH into the unit and never writes `CODEMAN_PASSWORD` into it. → [architecture-invariants#detached-start-and-service-install](docs/architecture-invariants.md#detached-start-and-service-install)
 
 **Self-update** (App Settings → System → Updates): in-app updater for git-clone installs under a supervisor (`systemd`, `launchd`, `launchd-daemon`, `docker-compose`, else `none`). The work runs in a DETACHED `scripts/self-update.sh` writing `update-status.json`, polled across the restart; pure helpers in `src/web/self-update.ts`. ⚠️ Compose: the restart kills the script, so nothing may be appended after the `restarting` marker; the repo must stay a host bind mount over `/opt/codeman` and the image must keep devDependencies + toolchain. ⚠️ `evaluateEnvironmentGate()` refuses releases that change `server.Dockerfile`/`docker-compose.yaml` or add `.env.example` keys, re-evaluated on `POST /api/system/update`; unknowns fail OPEN, but the exit-to-restart needs `--restart-by-exit 1` (`CODEMAN_RESTART_BY_EXIT=1` only in the Compose file). ⚠️ Keep the agent CLIs in `server.Dockerfile` pinned. → [docs/docker-self-update.md](docs/docker-self-update.md), [architecture-invariants#self-update](docs/architecture-invariants.md#self-update)
diff --git a/docs/architecture-invariants.md b/docs/architecture-invariants.md
index 5e269f95..aed0a1c4 100644
--- a/docs/architecture-invariants.md
+++ b/docs/architecture-invariants.md
@@ -225,7 +225,7 @@ Further detail: the `<prefix>: <title>` form (`w3-myapp: fix the login redirect`
 
 **Wheel/touch forwarding is NOT gated on viewport-at-bottom** (#205, `terminal-ui.js:_shouldForwardWheelToApp`): for sessions verified to scroll their own transcript on SGR wheel reports (claude ≥ 2.1.187 while `cliMouseTracking` is true, i.e. fullscreen; version via the local/docker/remote `--version` probes), the plain wheel AND touch drags forward as coalesced SGR reports (`_forwardScrollToApp` → `_sendSyntheticSgrWheel`, 40ms batches, 5-tick cap, 512-byte queue bound). It used to gate on the viewport being at the bottom so both scrollbacks stayed reachable, but a repaint-mode CLI keeps NO terminal scrollback of its own — xterm's buffer holds only replayed repaint frames, so local scrolling drags the CLI's pinned prompt box up the screen over stale frames; and `scrollToLastNonEmptyLine()` routinely parked the viewport off-bottom, silently pinning the wheel to local. Forwarding now snaps the viewport home first (SGR coordinates address the LIVE screen — a report computed from a scrolled-up viewport would hit-test the wrong row). Local scrollback remains on Shift+wheel and the `terminalWheelLocalScrollback` opt-out (both also cover touch via the shared gate; touch has no Shift, so the setting is its only local pin). `_wheelScrollLines()` normalizes `deltaMode` (Firefox fires LINE deltas ≈3/notch — read as pixels that rounded to 0 and fell to the ±1 fallback, ~4× too slow; PAGE deltas scale by `terminal.rows`) while keeping the #154 Shift-axis trap (macOS trackpads put Shift+scroll magnitude on deltaX). Tests: `test/terminal-touch-tap.test.ts`.
 
-**A false gate on a hollow pane must not mean a DEAD gesture** (#205 round 2, `_maybePageCliTranscript`): every way `_shouldForwardWheelToApp()` returns false leaves a repaint-mode pane scrolling a buffer that has nothing in it (`baseY === 0`) — the version probe came back empty, the CLI really is older than 2.1.187, the `cliMouseTracking` flag is unset (inline claude, or fullscreen right after a server restart), or the user turned on `terminalWheelLocalScrollback`. **opencode is the fifth case, and the gate is false there by design**: its TUI runs on the ALTERNATE SCREEN (1.18.31 measured: tmux `alternate_on=1`, `history_size=0`), so the buffer is hollow, and it IGNORES SGR wheel reports entirely (six `\x1b[<64;…M` reports against an idle pane left the capture byte-identical) while still paging its transcript on PageUp/PageDown (`messages_page_up/down`) — so paging is the only gesture that can reach it, and without it the wheel was silently dead in every opencode tab. The 1.12.0 retest reported exactly that shape for Claude: a wheel that did nothing at all while Fn+Up (PageUp) paged back through intact text, which is the proof that the CLI's own history and the PTY input path were both fine. So under the guard (`_localScrollbackIsHollow()` — `claude` or `opencode`, gate false, `baseY === 0`) wheel and touch travel is translated into coalesced `\x1b[5~` / `\x1b[6~` through the same 40ms queue as the SGR reports, at half a screen of travel per page key (the key jumps a whole screen; a 1:1 mapping was unusably slow with a discrete wheel). ⚠️ `shell`/`pi` own real terminal scrollback and are never paged, and codex/gemini/antigravity/grok/deepseek/omp page-key behaviour is unverified (`docs/scrollback-fix-plan.md`). ⚠️ Shift is excluded on purpose — it is the explicit "give me local scrollback" gesture and must keep that meaning. ⚠️ `terminalWheelLocalScrollback` is deliberately NOT scoped away from repaint-mode CLIs even though it is a footgun there: that would silently override an explicit user choice, so the fallback catches it instead. **Server-side counterpart**: `getClaudeCliVersion()` caches SUCCESS for the process lifetime but must never cache FAILURE — it used to, so one timed-out or PATH-starved probe at the first Claude session start disabled wheel-forwarding for every Claude session until the server restarted (a dead wheel on phone, tablet and laptop at once, the signature of a server-side cause). Failures now retry with a 1/2/4…15min backoff; the policy is the pure `resolveClaudeCliVersion()`. **Tiles** (a grid tile, the split's Pane B: `TerminalTile`) page the wheel through these same gates, called with the tile's own terminal and session (`_localScrollbackIsHollow(target)`, `_shouldForwardWheelToApp(ev, target)`) and the pure `CodemanTerminalInput.pageKeysForTravel`, so the mode list stays in terminal-ui.js alone. ⚠️ A tile's `baseY` is rarely 0 even when hollow: its first capture is taken at the PTY's previous, taller size and its row-shrinking fits push rows up, so it passes `localRows` with those rows discounted (`TerminalTile._localRows`). ⚠️ So a tile can be hollow with its viewport still up in those rows (Shift+PageUp, a scrollbar drag, a wheel during the first replay), and it pages only while `_terminalViewportAtBottom(tile.terminal)` holds: otherwise every wheel, wheel-down included, was paged and swallowed, the stale rows stayed on screen and the click report (which refuses an off-bottom viewport) went dead too. Off the bottom the wheel stays xterm's, so a wheel-down brings the viewport home. The primary pane needs no such gate: hollow there means `baseY === 0`, which is always at the bottom. Tiles have no touch path and do not forward SGR wheel (tile-grid-plan follow-up 4). Tests: `test/terminal-scroll-routing.test.ts`, `test/terminal-tile-scroll.test.ts`, `test/claude-cli-version-cache.test.ts`.
+**A false gate on a hollow pane must not mean a DEAD gesture** (#205 round 2, `_maybePageCliTranscript`): every way `_shouldForwardWheelToApp()` returns false leaves a repaint-mode pane scrolling a buffer that has nothing in it (`baseY === 0`) — the version probe came back empty, the CLI really is older than 2.1.187, the `cliMouseTracking` flag is unset (inline claude, or fullscreen right after a server restart), or the user turned on `terminalWheelLocalScrollback`. **opencode is the fifth case, and the gate is false there by design**: its TUI runs on the ALTERNATE SCREEN (1.18.31 measured: tmux `alternate_on=1`, `history_size=0`), so the buffer is hollow, and it IGNORES SGR wheel reports entirely (six `\x1b[<64;…M` reports against an idle pane left the capture byte-identical) while still paging its transcript on PageUp/PageDown (`messages_page_up/down`) — so paging is the only gesture that can reach it, and without it the wheel was silently dead in every opencode tab. The 1.12.0 retest reported exactly that shape for Claude: a wheel that did nothing at all while Fn+Up (PageUp) paged back through intact text, which is the proof that the CLI's own history and the PTY input path were both fine. So under the guard (`_localScrollbackIsHollow()` — `claude` or `opencode`, gate false, `baseY === 0`) wheel and touch travel is translated into coalesced `\x1b[5~` / `\x1b[6~` through the same 40ms queue as the SGR reports, at half a screen of travel per page key (the key jumps a whole screen; a 1:1 mapping was unusably slow with a discrete wheel). ⚠️ `shell`/`pi` own real terminal scrollback and are never paged, and codex/gemini/antigravity/grok/deepseek/omp page-key behaviour is unverified (`docs/scrollback-fix-plan.md`). ⚠️ Shift is excluded on purpose — it is the explicit "give me local scrollback" gesture and must keep that meaning. ⚠️ `terminalWheelLocalScrollback` is deliberately NOT scoped away from repaint-mode CLIs even though it is a footgun there: that would silently override an explicit user choice, so the fallback catches it instead. **Server-side counterpart**: `getClaudeCliVersion()` caches SUCCESS for the process lifetime but must never cache FAILURE — it used to, so one timed-out or PATH-starved probe at the first Claude session start disabled wheel-forwarding for every Claude session until the server restarted (a dead wheel on phone, tablet and laptop at once, the signature of a server-side cause). Failures now retry with a 1/2/4…15min backoff; the policy is the pure `resolveClaudeCliVersion()`. **Tiles** (a grid tile, the split's Pane B: `TerminalTile`) page the wheel through these same gates, called with the tile's own terminal and session (`_localScrollbackIsHollow(target)`, `_shouldForwardWheelToApp(ev, target)`) and the pure `CodemanTerminalInput.pageKeysForTravel`, so the mode list stays in terminal-ui.js alone. ⚠️ A tile's `baseY` is rarely 0 even when hollow: its first capture is taken at the PTY's previous, taller size and its row-shrinking fits push rows up, so it passes `localRows` with those rows discounted (`TerminalTile._localRows`). ⚠️ So a tile can be hollow with its viewport still up in those rows (Shift+PageUp, a scrollbar drag, a wheel during the first replay), and it pages only while `_terminalViewportAtBottom(tile.terminal)` holds: otherwise every wheel, wheel-down included, was paged and swallowed, the stale rows stayed on screen and the click report (which refuses an off-bottom viewport) went dead too. Off the bottom the wheel stays xterm's, so a wheel-down brings the viewport home. The primary pane needs no such gate: hollow there means `baseY === 0`, which is always at the bottom. Tiles also FORWARD the wheel to Claude's fullscreen renderer as SGR reports (`TerminalTile._maybeForwardWheelToCli`: the same `_shouldForwardWheelToApp(ev, target)` gate, the cell from `_clientPointToCell(x, y, tile.terminal)`, the bytes from the pure `CodemanTerminalInput.sgrWheelReports` the primary pane's `_sendSyntheticSgrWheel` also uses, flushed through the tile's own coalescer to the tile's session, never the primary queue, which flushes to `activeSessionId`); before that a fullscreen Claude tile left the wheel to xterm, which scrolled only stale replayed frames while Claude's transcript never moved. ⚠️ Shift+wheel in a tile is scrolled by the tile itself (`_maybeScrollLocalOnShift`, `scrollLines()` on the dominant axis), as in the primary pane: xterm's own scroller turns a Shift+vertical wheel into a horizontal one off macOS and Chrome on Windows sends Shift+wheel as `deltaX`, so left to xterm the gesture was dead. Tiles have no touch path. Tests: `test/terminal-scroll-routing.test.ts`, `test/terminal-tile-scroll.test.ts`, `test/claude-cli-version-cache.test.ts`.
 
 **Why the wheel went where it went is LOGGED** (`_logScrollRouting`): one console line per session per distinct decision — `[scroll] <id> → forward-sgr|page-keys|local-scrollback|repull-refused-downgrade (mode=…, cliVersion=…, localScrollbackOptOut=…, mouseTracking=…, localScrollbackRows=…)`. #205 ran two rounds of remote guesswork over questions this line answers directly; keep it when touching the routing.
 
diff --git a/docs/tile-grid-plan.md b/docs/tile-grid-plan.md
index 4a2573ee..943706b4 100644
--- a/docs/tile-grid-plan.md
+++ b/docs/tile-grid-plan.md
@@ -1011,13 +1011,17 @@ exits green. Use the browser runner for those files and read the file count.
    viewer keeps a stale width and renders garbled output (#464).
 3. WebSocket backpressure (`bufferedAmount` threshold, drop and send `{t:'r'}`
    on drain) for grids over slow links.
-4. Tile parity extras: mouse-wheel forwarding for Claude's fullscreen renderer,
-   a "Load full history" action inside a tile. (Done since: a tile pages a
-   hollow buffer's CLI transcript with PageUp/PageDown, the primary pane's
-   #555 route, and hand-reports a plain click while its session has
-   `cliMouseTracking` on, both through the primary pane's gates aimed at the
-   tile. The SGR wheel forwarding itself is still open: a fullscreen Claude
-   tile leaves the wheel to xterm.)
+4. Tile parity extras: a "Load full history" action inside a tile. (Done
+   since: a tile pages a hollow buffer's CLI transcript with PageUp/PageDown,
+   the primary pane's #555 route, hand-reports a plain click while its session
+   has `cliMouseTracking` on, and forwards the wheel to Claude's fullscreen
+   renderer as SGR wheel reports from its own cells
+   (`TerminalTile._maybeForwardWheelToCli`, encoding shared with the primary
+   pane via `CodemanTerminalInput.sgrWheelReports`), all through the primary
+   pane's gates aimed at the tile. Before that, a fullscreen Claude tile left
+   the wheel to xterm, which scrolled only stale replayed frames. Shift+wheel
+   scrolls the tile's local scrollback itself (`_maybeScrollLocalOnShift`),
+   since xterm turns it into a horizontal no-op off macOS.)
 5. WebGL in tiles, after measuring the DOM renderer with nine busy tiles.
 6. Named grid presets, possibly per owner on the server.
 7. The end state: the main terminal becomes a 1x1 grid of `TerminalTile`,
diff --git a/src/web/public/terminal-tile.js b/src/web/public/terminal-tile.js
index 222ebc65..5ac18139 100644
--- a/src/web/public/terminal-tile.js
+++ b/src/web/public/terminal-tile.js
@@ -11,14 +11,20 @@
  *
  * Deliberately plainer than the primary pane (this.terminal/this._ws in
  * terminal-ui.js): no local-echo overlay, no CJK IME textarea, no touch/mobile
- * handlers (a swipe on a touch screen pages nothing), no keyboard accessory
- * bar, and no SGR wheel forwarding to Claude's fullscreen renderer
- * (docs/tile-grid-plan.md follow-up 4). Built for wide screens; see
- * docs/split-pane-sessions-plan.md and docs/tile-grid-plan.md.
+ * handlers (a swipe on a touch screen pages nothing), and no keyboard
+ * accessory bar. Built for wide screens; see docs/split-pane-sessions-plan.md
+ * and docs/tile-grid-plan.md.
  *
  * What it does carry over from the primary pane, through the primary pane's
  * own code aimed at THIS pane (its terminal, its session, never the active
  * one):
+ *  - SGR wheel forwarding (_maybeForwardWheelToCli): Claude's fullscreen
+ *    renderer scrolls its own transcript on SGR wheel reports, while this
+ *    xterm holds only replayed repaint frames, so the wheel goes to the CLI
+ *    as reports at the pointer's cell in this pane, through the primary
+ *    pane's forwarding gate and its encoding (sgrWheelReports). Shift+wheel
+ *    scrolls the local scrollback itself (_maybeScrollLocalOnShift), as the
+ *    primary pane does, since xterm turns it into a horizontal no-op.
  *  - Hollow-buffer paging (#555): a CLI that draws in place (opencode on the
  *    alternate screen, Claude's repaint mode) leaves the xterm no scrollback,
  *    so the wheel pages the CLI's own transcript with PageUp/PageDown
@@ -43,7 +49,7 @@
  *
  * @dependency vendor/xterm.js, vendor/xterm-addon-fit.js
  * @dependency constants.js (window.CodemanTerminalFont, window.CodemanFetchDeadline, DEFAULT_SCROLLBACK, TERMINAL_TAIL_SIZE, TERMINAL_CHUNK_SIZE)
- * @dependency terminal-ui.js (codemanCurrentXtermTheme, codemanCurrentSkinIsLight, CodemanTerminalInput.shouldSuppressTerminalQueryResponse/isTerminalFocusOrMouseReport/wheelDeltaLines/pageKeysForTravel, app._shouldForwardWheelToApp/_localScrollbackIsHollow/_terminalViewportAtBottom/_handleDesktopTerminalClick)
+ * @dependency terminal-ui.js (codemanCurrentXtermTheme, codemanCurrentSkinIsLight, CodemanTerminalInput.shouldSuppressTerminalQueryResponse/isTerminalFocusOrMouseReport/wheelDeltaLines/wheelDeltaWholeLines/sgrWheelReports/pageKeysForTravel, app._shouldForwardWheelToApp/_localScrollbackIsHollow/_terminalViewportAtBottom/_clientPointToCell/_handleDesktopTerminalClick)
  * @dependency terminal-keycode229-recovery.js (window.CodemanKeyCode229Recovery, optional: absent, xterm's own textarea handling stands)
  * @loadorder 7.4 of 16, loaded after terminal-ui.js and before terminal-split.js
  */
@@ -201,6 +207,9 @@
       // Hollow-buffer paging (_maybePageCliTranscript): wheel travel short of a
       // whole page, carried to the next wheel event.
       this._pageKeyPending = 0;
+      // Shift+wheel travel short of a whole line, carried to the next wheel
+      // event (_maybeScrollLocalOnShift).
+      this._shiftScrollPending = 0;
       // Page keys waiting for the 40 ms flush, and its timer (_queueScrollBytes).
       this._scrollBytes = '';
       this._scrollFlushTimer = null;
@@ -1027,16 +1036,18 @@
 
     // Capture phase, because xterm's own wheel handler stopPropagation()s every
     // event it consumes, so a bubbling listener here would never see the wheel
-    // while the pane still has scrollback to scroll. Not passive: the one route
-    // this pane takes over, paging a hollow buffer's CLI transcript
-    // (_maybePageCliTranscript), is consumed right here (preventDefault plus
+    // while the pane still has scrollback to scroll. Not passive: the three
+    // routes this pane takes over, forwarding the wheel to Claude's fullscreen
+    // renderer (_maybeForwardWheelToCli), paging a hollow buffer's CLI
+    // transcript (_maybePageCliTranscript) and Shift+wheel's local scrollback
+    // (_maybeScrollLocalOnShift), are consumed right here (preventDefault plus
     // stopPropagation in the capture phase, the primary pane's technique), so
-    // xterm's viewport, a descendant, never sees it. Every other wheel is left
+    // xterm's viewport, a descendant, never sees them. Every other wheel is left
     // to xterm, which keeps doing the scrolling, and only observed for the
     // shell history pull.
     _installWheelListener() {
       this._onWheel = (ev) => {
-        if (this._maybePageCliTranscript(ev)) {
+        if (this._maybeForwardWheelToCli(ev) || this._maybePageCliTranscript(ev) || this._maybeScrollLocalOnShift(ev)) {
           ev.preventDefault();
           ev.stopPropagation();
           return;
@@ -1046,6 +1057,71 @@
       this.mountEl.addEventListener('wheel', this._onWheel, { capture: true, passive: false });
     }
 
+    // SGR wheel forwarding, the twin of the primary pane's capture-phase wheel
+    // handler and _forwardScrollToApp (terminal-ui.js; keep them in step).
+    // Claude's fullscreen renderer (claude 2.1.187+ while its mouse tracking is
+    // on, cliMouseTracking) scrolls its own transcript on SGR wheel reports,
+    // while this xterm holds only Codeman's replayed repaint frames (tmux keeps
+    // no history for such a pane). Left to xterm, the wheel dragged those stale
+    // frames, Claude's pinned input box with them, up the tile, or scrolled
+    // nothing at all. The gate is the primary pane's own, asked for THIS pane
+    // (its terminal, its session, never the active one), so the CLI rules stay
+    // in terminal-ui.js and this file names no CLI; the reports go to this
+    // pane's session through its own coalescer. Returns true when the wheel
+    // belongs to the CLI: a gesture with no whole line or no measurable cell is
+    // consumed too, as in the primary pane, so xterm never scrolls the stale
+    // frames under a forwarding session. Shift fails the gate, so Shift+wheel
+    // still scrolls the local scrollback (_maybeScrollLocalOnShift).
+    _maybeForwardWheelToCli(ev) {
+      if (this._destroyed || !this.terminal || !ev) return false;
+      const app = global.app;
+      const input = global.CodemanTerminalInput;
+      if (!app?._shouldForwardWheelToApp || !input?.sgrWheelReports || !input.wheelDeltaWholeLines) return false;
+      // xterm's own encoder forwards the wheel while the CLI's tracking reaches
+      // it, and its alt-scroll owns the alternate buffer, as in the primary pane.
+      const tracking = this.terminal.modes?.mouseTrackingMode;
+      if (tracking && tracking !== 'none') return false;
+      if (this.terminal.buffer?.active?.type === 'alternate') return false;
+      if (!app._shouldForwardWheelToApp(ev, { terminal: this.terminal, sessionId: this.sessionId })) return false;
+      // SGR coordinates address the live screen, so a report from a scrolled-up
+      // viewport would hit-test another row: snap home first (_forwardScrollToApp).
+      if (!app._terminalViewportAtBottom?.(this.terminal)) this.terminal.scrollToBottom?.();
+      const lines = input.wheelDeltaWholeLines(ev, this.terminal.rows);
+      const pos = app._clientPointToCell?.(ev.clientX, ev.clientY, this.terminal);
+      const bytes = input.sgrWheelReports(lines, pos);
+      if (bytes) this._queueScrollBytes(bytes);
+      return true;
+    }
+
+    // Shift+wheel scrolls this xterm's local scrollback, the explicit "local
+    // history" gesture, here as in the primary pane (whose capture-phase wheel
+    // handler scrolls with terminal.scrollLines() for the same reason). Left to
+    // xterm it was dead off macOS: Chrome on Windows sends Shift+wheel as a
+    // HORIZONTAL wheel (deltaX), and xterm's own scroller turns a Shift+vertical
+    // wheel into a horizontal one, so the viewport never moved. Reads the
+    // dominant axis under Shift (wheelDeltaLines), keeps the sub-line remainder
+    // for the next event (a trackpad's small deltas), and on the way up still
+    // asks a shell pane for more history. Returns true when the wheel was
+    // consumed here.
+    _maybeScrollLocalOnShift(ev) {
+      if (this._destroyed || !this.terminal || !ev?.shiftKey) return false;
+      const input = global.CodemanTerminalInput;
+      if (!input?.wheelDeltaLines) return false;
+      // xterm's own encoder forwards the wheel while the CLI's tracking reaches
+      // it, and its alt-scroll owns the alternate buffer, as in the primary pane.
+      const tracking = this.terminal.modes?.mouseTrackingMode;
+      if (tracking && tracking !== 'none') return false;
+      if (this.terminal.buffer?.active?.type === 'alternate') return false;
+      const total = this._shiftScrollPending + input.wheelDeltaLines(ev, this.terminal.rows);
+      const lines = Math.trunc(total);
+      this._shiftScrollPending = total - lines;
+      if (lines) {
+        this.terminal.scrollLines(lines);
+        if (lines < 0) this._maybeLoadMoreHistory();
+      }
+      return true;
+    }
+
     // A plain left-click reported to the CLI, the primary pane's desktop click
     // (terminal-ui.js _handleDesktopTerminalClick) aimed at this pane. The
     // server strips the mouse DECSETs of some modes (opencode's since #555, so a
@@ -1087,9 +1163,9 @@
       const tracking = this.terminal.modes?.mouseTrackingMode;
       if (tracking && tracking !== 'none') return false;
       const target = { terminal: this.terminal, sessionId: this.sessionId };
-      // The primary pane would forward this wheel to Claude's fullscreen
-      // renderer as SGR reports. Tiles do not do that yet (docs/tile-grid-plan.md
-      // follow-up 4), so the wheel stays with xterm, as before.
+      // A wheel for Claude's fullscreen renderer was forwarded as SGR reports
+      // before this ran (_maybeForwardWheelToCli); the gate is repeated so a
+      // forwarding session is never paged.
       if (app._shouldForwardWheelToApp?.(ev, target)) return false;
       if (!app._localScrollbackIsHollow?.({ ...target, localRows: this._localRows() })) return false;
       // Only from the live screen. The one gate the primary pane never needs: a
@@ -1111,8 +1187,9 @@
       return true;
     }
 
-    // Coalesces the page keys into one send per 40 ms, bounded at 512 bytes so a
-    // fling cannot build a backlog that keeps paging after it stops. A narrow
+    // Coalesces the scroll bytes (SGR wheel reports and page keys alike) into
+    // one send per 40 ms, bounded at 512 bytes so a fling cannot build a backlog
+    // that keeps scrolling after it stops. A narrow
     // twin of the primary pane's _queueScrollBytes / _flushWheelSgrQueue
     // (terminal-ui.js; keep the two in step), which flushes to the active
     // session only. Sent ephemeral (no seq, never persisted) to THIS pane's
diff --git a/src/web/public/terminal-ui.js b/src/web/public/terminal-ui.js
index d95960b2..dda3d431 100644
--- a/src/web/public/terminal-ui.js
+++ b/src/web/public/terminal-ui.js
@@ -107,6 +107,32 @@
         : delta / 25; // DOM_DELTA_PIXEL (Chrome/WebKit, and every trackpad)
   }
 
+  // The same travel rounded to whole lines for the SGR wheel reports: a pure
+  // horizontal swipe is 0 (nothing to send), and anything else moves at least
+  // one line, so the small pixel deltas of a precision touchpad still scroll.
+  // The body of the primary pane's _wheelScrollLines.
+  function wheelDeltaWholeLines(ev, rows) {
+    const lines = wheelDeltaLines(ev, rows);
+    if (!lines) return 0;
+    return Math.round(lines) || (lines > 0 ? 1 : -1);
+  }
+
+  // Ticks one gesture batch may report: Claude applies its own scroll-speed
+  // multiplier and acceleration on top, so a bigger batch only overshoots.
+  const SGR_WHEEL_MAX_TICKS = 5;
+
+  // Whole wheel lines → SGR wheel reports at a 1-based cell `pos` ({ col, row },
+  // live-screen relative): button 64 per line up, 65 per line down, capped at
+  // SGR_WHEEL_MAX_TICKS. '' when there is nothing to send. The encoding of the
+  // primary pane's _sendSyntheticSgrWheel, pure so a TerminalTile forwards
+  // byte-identical reports to its own session.
+  function sgrWheelReports(lines, pos) {
+    if (!lines || !pos) return '';
+    const btn = lines < 0 ? 64 : 65;
+    const ticks = Math.min(Math.abs(lines), SGR_WHEEL_MAX_TICKS);
+    return `\x1b[<${btn};${pos.col};${pos.row}M`.repeat(ticks);
+  }
+
   // Gesture travel → PageUp/PageDown keys for a terminal `rows` tall: adds
   // `lines` to the sub-page travel already `pending`, and returns the travel
   // left over plus the keys to send ('' below one page). The arithmetic of the
@@ -260,6 +286,9 @@
     PAGE_KEY_SCREEN_FRACTION,
     PAGE_KEY_MAX_PER_BATCH,
     wheelDeltaLines,
+    wheelDeltaWholeLines,
+    SGR_WHEEL_MAX_TICKS,
+    sgrWheelReports,
     pageKeysForTravel,
     TUI_PROMPT_DEFAULT_ROWS_FROM_BOTTOM,
     MOBILE_KEYBOARD_DISMISS_EXEMPT_SELECTOR,
@@ -5563,9 +5592,8 @@ Object.assign(CodemanApp.prototype, {
   // the ±1 fallback — one line per notch, versus 4-5 for Chrome's ~110px. In
   // Claude mode the same value also capped the forwarded SGR report at one tick.
   _wheelScrollLines(ev) {
-    const lines = this._wheelScrollLinesFloat(ev);
-    if (!lines) return 0; // pure horizontal swipe: don't fall through to -1
-    return Math.round(lines) || (lines > 0 ? 1 : -1);
+    // Pure horizontal swipe: 0, never the ±1 fallback (wheelDeltaWholeLines).
+    return window.CodemanTerminalInput.wheelDeltaWholeLines(ev, this.terminal?.rows);
   },
 
   /** Unrounded variant for the smooth local-scroll path, which accumulates
@@ -5639,13 +5667,13 @@ Object.assign(CodemanApp.prototype, {
   // scroll-speed multiplier and acceleration on top), and the queue is bounded
   // so a wild scroll can't build a backlog that keeps scrolling after the finger
   // stops. Flushed via _sendInputEphemeral — loss-tolerant, off the durable queue.
+  // The encoding is the pure CodemanTerminalInput.sgrWheelReports, which a
+  // TerminalTile calls with its own cell (TerminalTile._maybeForwardWheelToCli).
   _sendSyntheticSgrWheel(clientX, clientY, lines) {
     if (!this.activeSessionId || !lines) return;
     const pos = this._clientPointToCell(clientX, clientY);
     if (!pos) return;
-    const btn = lines < 0 ? 64 : 65;
-    const ticks = Math.min(Math.abs(lines), 5);
-    this._queueScrollBytes(`\x1b[<${btn};${pos.col};${pos.row}M`.repeat(ticks));
+    this._queueScrollBytes(window.CodemanTerminalInput.sgrWheelReports(lines, pos));
   },
 
   /**
diff --git a/test/mocks/terminal-tile-fakes.ts b/test/mocks/terminal-tile-fakes.ts
index 2b70614a..73e940f6 100644
--- a/test/mocks/terminal-tile-fakes.ts
+++ b/test/mocks/terminal-tile-fakes.ts
@@ -199,6 +199,19 @@ export class FakeTerminal {
   }
   scrollToLine() {}
   scrollToTop() {}
+  /** Every scrollLines() amount; the viewport moves within [0, baseY], as in xterm. */
+  scrolledLines: number[] = [];
+  scrollLines(amount: number) {
+    this.scrolledLines.push(amount);
+    const active = this.buffer.active;
+    active.viewportY = Math.max(0, Math.min(active.baseY, active.viewportY + amount));
+  }
+  /** Times scrollToBottom() was called; the viewport moves to the live screen, as in xterm. */
+  scrolledToBottom = 0;
+  scrollToBottom() {
+    this.scrolledToBottom++;
+    this.buffer.active.viewportY = this.buffer.active.baseY;
+  }
   dispose() {}
   type(data: string) {
     this.dataCb?.(data);
diff --git a/test/terminal-tile-scroll.browser.test.ts b/test/terminal-tile-scroll.browser.test.ts
index 55eebce4..cfeccbdb 100644
--- a/test/terminal-tile-scroll.browser.test.ts
+++ b/test/terminal-tile-scroll.browser.test.ts
@@ -14,6 +14,12 @@
  *    wheel-down scrolls xterm home instead of being paged, and paging resumes
  *    from the live screen.
  *
+ *  - a fullscreen Claude tile sends the wheel to the CLI as SGR wheel reports
+ *    at the pointer's cell in the TILE (Claude scrolls its own transcript on
+ *    them), and xterm's viewport stays on the live screen instead of
+ *    scrolling the stale replayed frames above it. Shift+wheel scrolls the
+ *    local scrollback, which xterm alone turns into a horizontal no-op.
+ *
  * The wheel is a real one (`page.mouse.wheel()`), and the last step proves it
  * reaches xterm: a wheel the tile does not page scrolls xterm's viewport, so
  * "xterm did not scroll" on a paged wheel means something. What keeps xterm
@@ -200,4 +206,95 @@ describe('TerminalTile wheel paging in a real browser', () => {
       }, PROBE_ID);
     }
   });
+
+  it("forwards a fullscreen Claude tile's wheel as SGR reports from its own cells, and keeps xterm on the live screen", async () => {
+    await page.evaluate(async (id) => {
+      const w = window as any;
+      const app = w.app;
+      const CAPTURE_ROWS = 40;
+      const capture = Array.from({ length: CAPTURE_ROWS }, (_, i) => `frame ${i}`).join('\r\n');
+
+      // The tile's session: Claude in its fullscreen renderer (mouse tracking on).
+      app.sessions.set(id, { id, mode: 'claude', cliVersion: '2.1.295', cliMouseTracking: true });
+      const sent: Array<[string, string]> = [];
+      const realEphemeral = app._sendInputEphemeral;
+      app._sendInputEphemeral = (sessionId: string, data: string) => sent.push([sessionId, data]);
+      const realFetch = w.fetch;
+      w.fetch = async (url: string, init?: unknown) =>
+        String(url).includes(`/api/sessions/${id}/terminal`)
+          ? new Response(JSON.stringify({ data: { terminalBuffer: capture, captureRows: CAPTURE_ROWS } }))
+          : realFetch(url, init);
+      const RealWebSocket = w.WebSocket;
+      w.WebSocket = class {
+        static OPEN = 1;
+        readyState = 0;
+        send() {}
+        close() {}
+      };
+
+      // Offset from the page corner, so a cell computed against the wrong
+      // element (the primary pane's) would come out different.
+      const mount = document.createElement('div');
+      mount.style.cssText =
+        'position:fixed;left:100px;top:120px;width:640px;height:300px;z-index:99999;background:#000';
+      document.body.appendChild(mount);
+      const tile = new w.TerminalTile(id, mount, { mode: 'claude' });
+      try {
+        await tile.connect();
+      } finally {
+        w.fetch = realFetch;
+        w.WebSocket = RealWebSocket;
+      }
+      w.__tileProbe = { tile, mount, sent, realEphemeral };
+    }, PROBE_ID);
+
+    try {
+      // Where the pointer will be, as the tile's own screen maps it (1-based).
+      const expected = await page.evaluate(() => {
+        const term = (window as any).__tileProbe.tile.terminal;
+        const rect = term.element.querySelector('.xterm-screen').getBoundingClientRect();
+        const cell = term._core._renderService.dimensions.css.cell;
+        const x = rect.left + cell.width * 7.5;
+        const y = rect.top + cell.height * 3.5;
+        return { x, y, col: 8, row: 4 };
+      });
+      // The stale frames sit above the live screen: something xterm COULD scroll.
+      const before = await snap();
+      expect(before.baseY).toBeGreaterThan(0);
+      expect(before.viewportY).toBe(before.baseY);
+
+      await page.mouse.move(expected.x, expected.y);
+      await page.mouse.wheel(0, -100); // one Chrome-on-Windows notch up
+      await page.waitForTimeout(150);
+      await page.mouse.wheel(0, 50); // half a notch down
+      await page.waitForTimeout(150);
+
+      const after = await snap();
+      expect(after.viewportY).toBe(before.baseY); // xterm never scrolled the stale frames
+      expect(after.sent.every(([sid]) => sid === PROBE_ID)).toBe(true);
+      const up = `\x1b[<64;${expected.col};${expected.row}M`;
+      const down = `\x1b[<65;${expected.col};${expected.row}M`;
+      expect(after.sent.map(([, data]) => data)).toEqual([up.repeat(4), down.repeat(2)]);
+
+      // Shift+wheel is the explicit "local scrollback" gesture: scrolled locally
+      // by the tile (xterm alone turns it into a horizontal no-op), unforwarded.
+      await page.keyboard.down('Shift');
+      await page.mouse.wheel(0, -100);
+      await page.keyboard.up('Shift');
+      await page.waitForTimeout(150);
+      const shifted = await snap();
+      expect(shifted.sent.length).toBe(after.sent.length);
+      expect(shifted.viewportY).toBeLessThan(before.baseY);
+    } finally {
+      await page.evaluate((id) => {
+        const w = window as any;
+        const probe = w.__tileProbe;
+        probe.tile.destroy();
+        probe.mount.remove();
+        w.app.sessions.delete(id);
+        w.app._sendInputEphemeral = probe.realEphemeral;
+        delete w.__tileProbe;
+      }, PROBE_ID);
+    }
+  });
 });
diff --git a/test/terminal-tile-scroll.test.ts b/test/terminal-tile-scroll.test.ts
index 1ad5459e..a75640cd 100644
--- a/test/terminal-tile-scroll.test.ts
+++ b/test/terminal-tile-scroll.test.ts
@@ -18,6 +18,13 @@
  * history, so the tile discounts them (`_localRows`); the overflow cases below
  * fail without that discount.
  *
+ * Claude's fullscreen renderer is the other half: it scrolls its own
+ * transcript on SGR wheel reports, which the primary pane sends it, while a
+ * tile's xterm holds only replayed repaint frames. A tile left that wheel to
+ * xterm too, so in a grid of fullscreen Claude sessions the wheel dragged
+ * stale frames around or did nothing; it now forwards the reports to its own
+ * session, from its own geometry (`_maybeForwardWheelToCli`).
+ *
  * Real code under test: constants.js + app.js + terminal-ui.js +
  * terminal-tile.js in one `vm` context, as in terminal-tile-input.test.ts.
  * xterm, the fit addon and WebSocket are fakes (test/mocks/terminal-tile-fakes.ts),
@@ -260,7 +267,8 @@ describe('a tile pages a hollow buffer through the primary pane gates', () => {
       expect(ev.stopPropagation).not.toHaveBeenCalled();
     };
 
-    fireAndCheck(wheelLines(-12, { shiftKey: true })); // the explicit "local scrollback" gesture
+    // The explicit "local scrollback" gesture: scrolled locally, never paged.
+    mount.fire('wheel', wheelLines(-12, { shiftKey: true }));
 
     term.modes.mouseTrackingMode = 'any'; // xterm's own encoder forwards the wheel
     fireAndCheck(wheelLines(-12));
@@ -280,18 +288,207 @@ describe('a tile pages a hollow buffer through the primary pane gates', () => {
   });
 });
 
-describe("the forwarding gate behaves as in the primary pane, for the tile's session", () => {
-  it('leaves a fullscreen Claude tile to xterm (tiles do not forward SGR wheel yet)', async () => {
-    const app = makeApp({ 's-tile': { mode: 'claude', cliVersion: '2.1.223', cliMouseTracking: true } });
-    const { ws, mount } = await connectTile(app, { mode: 'claude' });
+describe('a fullscreen Claude tile forwards the wheel to the CLI as SGR reports', () => {
+  const FULLSCREEN: Session = { mode: 'claude', cliVersion: '2.1.295', cliMouseTracking: true };
+  // 10 + 8 * 20 + 1, 20 + 16 * 5 + 1 inside the tile's own screen: column 21, row 6.
+  const AT = { clientX: 171, clientY: 101 };
+  const UP = '\x1b[<64;21;6M';
+  const DOWN = '\x1b[<65;21;6M';
 
-    const ev = wheelLines(-12);
-    mount.fire('wheel', ev);
+  it("sends ephemeral reports on the tile's socket, from the tile's own geometry, and keeps xterm out", async () => {
+    const { ws, mount } = await connectTile(makeApp({ 's-tile': FULLSCREEN }), { mode: 'claude' });
+
+    const up = wheelLines(-3, AT);
+    mount.fire('wheel', up);
+
+    expect(up.preventDefault).toHaveBeenCalled();
+    expect(up.stopPropagation).toHaveBeenCalled();
+    expect(ws.inputFrames()).toEqual([]); // coalesced, not sent per event
+    expect(flushed(ws)).toEqual([{ t: 'i', d: UP.repeat(3) }]); // no seq: never persisted
+
+    mount.fire('wheel', wheelLines(2, AT));
+    expect(flushed(ws).at(-1)).toEqual({ t: 'i', d: DOWN.repeat(2) });
+  });
+
+  it('forwards a Chrome-on-Windows notch (100 px) as four ticks and caps a fling at five', async () => {
+    const { ws, mount } = await connectTile(makeApp({ 's-tile': FULLSCREEN }), { mode: 'claude' });
+
+    mount.fire('wheel', wheel(-100, AT));
+    expect(flushed(ws)).toEqual([{ t: 'i', d: UP.repeat(4) }]);
+
+    mount.fire('wheel', wheelLines(-1000, AT));
+    expect(flushed(ws).at(-1)).toEqual({ t: 'i', d: UP.repeat(5) });
+  });
+
+  it('moves a precision touchpad delta one line, and consumes a pure horizontal swipe without sending', async () => {
+    const { ws, mount } = await connectTile(makeApp({ 's-tile': FULLSCREEN }), { mode: 'claude' });
+
+    mount.fire('wheel', wheel(-3, AT)); // 0.12 of a line
+    expect(flushed(ws)).toEqual([{ t: 'i', d: UP }]);
+
+    const sideways = wheel(0, { ...AT, deltaX: 120 });
+    mount.fire('wheel', sideways);
+    expect(sideways.preventDefault).toHaveBeenCalled(); // never xterm's stale frames
+    expect(flushed(ws)).toHaveLength(1);
+  });
+
+  it('coalesces the reports of wheels within 40 ms into one frame', async () => {
+    const { ws, mount } = await connectTile(makeApp({ 's-tile': FULLSCREEN }), { mode: 'claude' });
+
+    mount.fire('wheel', wheelLines(-2, AT));
+    vi.advanceTimersByTime(20);
+    mount.fire('wheel', wheelLines(-2, AT));
+
+    expect(flushed(ws)).toEqual([{ t: 'i', d: UP.repeat(4) }]);
+  });
+
+  it('snaps a scrolled-up viewport home before reporting, since SGR rows address the live screen', async () => {
+    const { ws, term, mount } = await connectTile(makeApp({ 's-tile': FULLSCREEN }), { mode: 'claude' });
+    term.buffer.active.baseY = 10;
+    term.buffer.active.viewportY = 0;
+
+    mount.fire('wheel', wheelLines(-1, AT));
+
+    expect(term.scrolledToBottom).toBe(1);
+    expect(term.buffer.active.viewportY).toBe(10);
+    expect(flushed(ws)).toEqual([{ t: 'i', d: UP }]);
+  });
+
+  it("follows the tile session's gate, never the active session's", async () => {
+    // Active session is a shell; the tile shows fullscreen Claude: the tile forwards.
+    const forwarded = await connectTile(makeApp({ other: { mode: 'shell' }, 's-tile': FULLSCREEN }, 'other'), {
+      mode: 'claude',
+    });
+    forwarded.mount.fire('wheel', wheelLines(-1, AT));
+    expect(flushed(forwarded.ws)).toEqual([{ t: 'i', d: UP }]);
+
+    // Active session is fullscreen Claude; the tile shows inline Claude (no
+    // tracking): no reports, the hollow buffer is paged instead.
+    const inline = await connectTile(
+      makeApp({ other: FULLSCREEN, 's-tile': { ...FULLSCREEN, cliMouseTracking: false } }, 'other'),
+      { mode: 'claude' }
+    );
+    inline.mount.fire('wheel', wheelLines(-12, AT));
+    expect(flushed(inline.ws)).toEqual([{ t: 'i', d: PAGE_UP }]);
+  });
+
+  it('scrolls Shift locally, and leaves a tracking xterm and the alternate buffer to xterm', async () => {
+    const { ws, term, mount } = await connectTile(makeApp({ 's-tile': FULLSCREEN }), { mode: 'claude' });
+    const fireAndCheck = (ev: ReturnType<typeof wheel>) => {
+      mount.fire('wheel', ev);
+      expect(ev.preventDefault).not.toHaveBeenCalled();
+      expect(ev.stopPropagation).not.toHaveBeenCalled();
+    };
+
+    const shifted = wheelLines(-3, { ...AT, shiftKey: true }); // the explicit "local scrollback" gesture
+    mount.fire('wheel', shifted);
+    expect(shifted.preventDefault).toHaveBeenCalled();
+    expect(term.scrolledLines).toEqual([-3]);
+
+    term.modes.mouseTrackingMode = 'any'; // xterm's own encoder forwards the wheel
+    fireAndCheck(wheelLines(-3, AT));
+    term.modes.mouseTrackingMode = 'none';
+
+    term.buffer.active.type = 'alternate'; // xterm's alt-scroll owns it
+    fireAndCheck(wheelLines(-3, AT));
 
-    expect(ev.preventDefault).not.toHaveBeenCalled();
     expect(flushed(ws)).toEqual([]);
   });
 
+  it('pages, never reports, a Claude older than 2.1.187', async () => {
+    const app = makeApp({ 's-tile': { ...FULLSCREEN, cliVersion: '2.1.100' } });
+    const { ws, mount } = await connectTile(app, { mode: 'claude' });
+
+    mount.fire('wheel', wheelLines(-12, AT));
+
+    expect(flushed(ws)).toEqual([{ t: 'i', d: PAGE_UP }]);
+  });
+
+  it('sends exactly the bytes the primary pane sends for the same wheel sequence', async () => {
+    const sequence = [-1, -3, 2, -100, 40, -0.2, 7].map((n) => wheelLines(n, AT));
+    const { ws, mount } = await connectTile(makeApp({ 's-tile': FULLSCREEN }), { mode: 'claude' });
+    const tileBytes: string[] = [];
+    for (const ev of sequence) {
+      mount.fire('wheel', ev);
+      tileBytes.push(
+        ...flushed(ws)
+          .splice(tileBytes.length)
+          .map((f) => f.d as string)
+      );
+    }
+
+    const primary = makeApp({ other: FULLSCREEN }, 'other') as App & {
+      terminal: unknown;
+      _sendSyntheticSgrWheel(x: number, y: number, lines: number): void;
+      _wheelScrollLines(ev: unknown): number;
+      _flushWheelSgrQueue(): void;
+    };
+    const sent: string[] = [];
+    primary._sendInputEphemeral = (_id: string, data: string) => sent.push(data);
+    primary.terminal = new FakeTerminal({});
+    for (const ev of sequence) {
+      primary._sendSyntheticSgrWheel(ev.clientX, ev.clientY, primary._wheelScrollLines(ev));
+      primary._flushWheelSgrQueue();
+    }
+
+    expect(tileBytes.join('')).not.toBe('');
+    expect(tileBytes).toEqual(sent);
+  });
+});
+
+describe("Shift+wheel scrolls the tile's local scrollback itself", () => {
+  // xterm's own scroller turns a Shift+vertical wheel into a horizontal one off
+  // macOS, and Chrome on Windows already delivers Shift+wheel as deltaX, so
+  // left to xterm the gesture moved nothing.
+  it.each(['claude', 'opencode', 'shell'])('scrolls a %s tile with real history, and sends nothing', async (mode) => {
+    const { ws, term, mount } = await connectTile(makeApp({ 's-tile': { mode } }), { mode });
+    term.write(lines(40));
+    const bottom = term.buffer.active.baseY;
+    expect(bottom).toBeGreaterThan(5);
+
+    const up = wheelLines(-3, { shiftKey: true });
+    mount.fire('wheel', up);
+
+    expect(up.preventDefault).toHaveBeenCalled();
+    expect(up.stopPropagation).toHaveBeenCalled();
+    expect(term.buffer.active.viewportY).toBe(bottom - 3);
+    mount.fire('wheel', wheelLines(3, { shiftKey: true }));
+    expect(term.buffer.active.viewportY).toBe(bottom);
+    expect(flushed(ws)).toEqual([]);
+  });
+
+  it("reads Chrome-on-Windows' horizontal Shift+wheel (deltaX) as vertical travel", async () => {
+    const { term, mount } = await connectTile(makeApp({ 's-tile': { mode: 'claude' } }), { mode: 'claude' });
+
+    mount.fire('wheel', wheel(0, { deltaX: -100, shiftKey: true }));
+
+    expect(term.scrolledLines).toEqual([-4]);
+  });
+
+  it('carries sub-line travel to the next event, so a trackpad still scrolls', async () => {
+    const { term, mount } = await connectTile(makeApp({ 's-tile': { mode: 'claude' } }), { mode: 'claude' });
+
+    mount.fire('wheel', wheel(-10, { shiftKey: true })); // 0.4 of a line
+    mount.fire('wheel', wheel(-10, { shiftKey: true }));
+    expect(term.scrolledLines).toEqual([]);
+    mount.fire('wheel', wheel(-10, { shiftKey: true }));
+
+    expect(term.scrolledLines).toEqual([-1]);
+  });
+
+  it('still pulls history on a Shift+wheel-up at the top of a shell tile', async () => {
+    const { mount } = await connectTile(makeApp({ 's-tile': { mode: 'shell' } }), { mode: 'shell' });
+    expect(fetchMock).toHaveBeenCalledTimes(1); // the initial load
+
+    mount.fire('wheel', wheelLines(-3, { shiftKey: true }));
+    await vi.advanceTimersByTimeAsync(0);
+
+    expect(fetchMock).toHaveBeenCalledTimes(2);
+    expect(String(fetchMock.mock.calls[1][0])).toContain('full=1&tail=');
+  });
+});
+
+describe("the forwarding gate behaves as in the primary pane, for the tile's session", () => {
   it('pages that same tile under the "Wheel scrolls local history" opt-out (the footgun rescue)', async () => {
     const app = makeApp({ 's-tile': { mode: 'claude', cliVersion: '2.1.223', cliMouseTracking: true } });
     app.loadAppSettingsFromStorage = () => ({ terminalWheelLocalScrollback: true });