Add stackable toast notifications

Replaces the single-slot status-line flash with a top-right toast
stack over the focused pane. flashError, flashTransient, and
notifyAttention all push onto the same stack (cap 5, FIFO drop).
Ctrl-N dismisses the most recent toast; empty stack falls through to
the focused PTY so readline / nano / emacs / opencode bindings keep
working. A new "Clear notifications" palette item empties the stack.

.gitea/workflows/release.yml

@@ -11,14 +11,19 @@ jobs:
     steps:
       - uses: actions/checkout@v4
 
-      - uses: actions/setup-go@v5
+      - uses: jdx/mise-action@v2
         with:
-          go-version-file: go.mod
           cache: true
 
-      - uses: mlugg/setup-zig@v2
+      - name: Cache Go modules
+        uses: actions/cache@v4
         with:
-          version: 0.15.2
+          path: |
+            ~/.cache/go-build
+            ~/go/pkg/mod
+          key: ${{ runner.os }}-go-${{ hashFiles('**/go.sum') }}
+          restore-keys: |
+            ${{ runner.os }}-go-
 
       - name: Build libghostty-vt
         run: make deps

.mise.toml

@@ -3,6 +3,8 @@
 # libghostty-vt is built from a pinned upstream Ghostty commit; that
 # commit's build.zig.zon pins minimum_zig_version = 0.15.2. We match
 # it here so contributors don't have to puzzle out the version from
-# a deep upstream file.
+# a deep upstream file. The go pin matches go.mod so CI and local
+# builds use the same toolchain.
 [tools]
 zig = "0.15.2"
+go = "1.26.3"

CHANGELOG.md

@@ -6,6 +6,36 @@ loosely follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 
+### Changed
+- Replaced the single-slot status-line "flash" with a stackable toast
+  surface anchored at the top-right of the focused pane. `flashError`,
+  `flashTransient`, and MCP `request_human_attention` now push onto
+  the toast stack (cap 5, oldest drops). Toasts persist until
+  dismissed with `Ctrl-N`, or cleared via the new
+  "Clear notifications" palette command. The status line no longer
+  shows the `[!]` prefix.
+- `Ctrl-N` is consumed by the host only when there is a toast to
+  dismiss; an empty stack lets `Ctrl-N` pass through to the focused
+  child so readline / nano / emacs / opencode keep their bindings.
+
+## [0.0.4] - 2026-05-15
+
+### Changed
+- Release workflow (`.gitea/workflows/release.yml`) now provisions
+  Zig and Go through `jdx/mise-action@v2`, reading the versions from
+  `.mise.toml` (zig 0.15.2, go 1.26.3). Both toolchains were
+  previously installed via `mlugg/setup-zig` and `actions/setup-go`,
+  whose mirror chase / GitHub fetch combined for ~8 minutes per run
+  before any patterm code compiled. mise pulls each tool once and
+  caches the install dir, so subsequent runs hit the cache instead of
+  re-downloading. `make deps` still resolves zig via `mise which zig`
+  with a PATH fallback; `go.mod` already pinned `go 1.26.3`, so the
+  new `go` entry in `.mise.toml` just keeps CI and local builds on
+  the same toolchain.
+- A Go module/build cache step (`actions/cache@v4`, keyed on
+  `go.sum`) was added so `go build` doesn't re-download dependencies
+  on every tag push.
+
 ## [0.0.3] - 2026-05-15
 
 ### Added
@@ -70,6 +100,8 @@ loosely follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   the command field.
 
 ### Fixed
+- Error/status flashes now restore the currently focused pane instead
+  of drawing the empty-state hint over a running agent or process.
 - Release workflow (`.gitea/workflows/release.yml`) now uses
   `mlugg/setup-zig@v2` instead of the deprecated `@v1`. v1 hard-coded
   the pre-0.14 tarball name (`zig-linux-x86_64-<ver>.tar.xz`), so

TODO.md

@@ -1,115 +1,3 @@
-# Perf Audit (reviewed 2026-05-15)
-Findings that survived the 2026-05-15 review pass. Low and marginal
-items from the original sweep were removed; remaining items have enough
-measured or workflow evidence to justify action.
-
-Baseline benchmark numbers (`go test -bench=. ./internal/app/`, AMD
-Ryzen 7 7800X3D, libghostty-vt **ReleaseFast** after the Makefile
-fix landed):
-
-```
-# Renderer alone
-ViewportRenderer_PlainASCII       229 MB/s     1.3 KB/op    6 allocs/op
-ViewportRenderer_StyledLines       89 MB/s    91   KB/op  4325 allocs/op
-ViewportRenderer_RatatuiBurst      40 MB/s   365   KB/op 17306 allocs/op
-RendererThroughput_ReuseInstance   90 MB/s   316   KB/op 17380 allocs/op
-ContainsOSC_NoOSC                3050 MB/s     0   B/op     0 allocs/op
-
-# ASCII-video stream (renderer only — 3 sec at the target fps)
-ASCIIVideo_Stream_8Color_120fps     260 µs/frame  3845 fps_ceiling   3.1% budget
-ASCIIVideo_Stream_TrueColor_120fps  576 µs/frame  1735 fps_ceiling   6.9% budget
-
-# Full pipeline (em.Write + renderer + io.Discard write)
-Pipeline_ASCIIVideo_8Color_120fps     493 µs/frame  2030 fps_ceiling   5.9% budget
-Pipeline_ASCIIVideo_TrueColor_120fps 1075 µs/frame   931 fps_ceiling  12.9% budget
-
-# Emulator alone (libghostty-vt CSI/SGR parser)
-Emulator_Write_Stream_8Color_120fps    257 µs/frame  3890 fps_ceiling
-Emulator_Write_Stream_TrueColor_120fps 488 µs/frame  2051 fps_ceiling
-```
-
-The current pipeline still has large 120 fps headroom. The remaining
-renderer concern is multi-MiB styled replay latency and allocation
-churn, not normal steady-state frame budget.
-
-
-- [ ] **viewport renderer allocates heavily on SGR/CSI-heavy chunks.** [MEDIUM]
-  - Review evidence: five benchmark reps confirmed
-    `ViewportRenderer_StyledLines` at about 4,325 allocs per 16 KiB
-    chunk (~91.5 KB/op, roughly 1 alloc per 3.8 input bytes), and
-    `ViewportRenderer_RatatuiBurst` at about 17,306 allocs per chunk
-    (~365 KB/op). A 5 MiB styled resume benchmark allocated about
-    31 MB across 1.38M objects.
-  - Likely hot paths: generic CSI/SGR output in
-    `internal/app/viewport_renderer.go` sends many sequences through
-    `vr.shifter.Shift(vr.buf)`, while `internal/app/cursorshift.go`
-    returns a fresh `[]byte` via `pending.String()` on every
-    `Shift` call and parses CSI params through `string(raw)` /
-    `strings.Split`. The mode-helper `string(params)` conversions
-    are real, but probably not the main SGR-heavy cost.
-  - Fix direction: make `cursorShifter` write into caller-owned
-    scratch output or directly into the viewport renderer's pending
-    builder; parse CSI params from byte slices; pre-grow/reuse
-    renderer and shifter buffers. Re-run styled-lines, ratatui, and
-    5 MiB resume benchmarks; use pprof when available to confirm the
-    top allocation sites.
-
-- [ ] **large styled resume/replay dumps spend visible time in viewport rendering.** [MEDIUM]
-  - Review evidence: `BenchmarkSessionResume_5MiBStyled` measured
-    about 58 ms median and 63 ms p95 over five reps. The plain 5 MiB
-    benchmark was about 23-24 ms with only 21 allocs. The live path
-    renders focused PTY chunks through `renderer.Render`, then still
-    pays emulator writes, ring writes, event dispatch, stdout writes,
-    and real terminal paint.
-  - Scope: this is not a Codex steady-state throughput limit. A
-    100 KB/s stream is far below the styled renderer's ~80-90 MB/s
-    ceiling. It matters for multi-MiB burst replay, resume/startup
-    dumps, and dense full-screen churn.
-  - Fix direction: do the allocation fix first, since it should also
-    improve throughput. After that, invest further only if styled
-    resume traces remain user-visible or the styled-lines benchmark
-    is still under roughly 300 MB/s.
-
-- [ ] **wait_for_pattern re-scans the entire stream/grid while waiting.** [MEDIUM]
-  - `internal/app/host.go:476-493` (the `check` closure). On
-    `scope="scrollback"` it calls `c.StreamRead(0)` followed by
-    `stripANSIBytes(nil, b)`, so each check can copy, strip, and
-    search the full 1 MiB ring. On `scope="grid"` it calls
-    `PlainText()` and runs the regex against the full grid string.
-  - Caveat from review: the current chunk notifier coalesces bursts
-    with a buffered channel and has a 500 ms fallback, so this is not
-    necessarily one full scan per PTY chunk. It is still meaningful
-    for active waits on chatty panes.
-  - Fix direction: for `scrollback`, track the last checked stream
-    offset and search only new output plus a bounded overlap/scratch
-    buffer so matches spanning chunks are not missed. For `grid`,
-    dedupe on `ScreenVersion()` and skip work when the version has
-    not changed.
-
-- [ ] **search_output rebuilds and searches whole scrollback on every call.** [MEDIUM]
-  - `internal/app/host.go:428-437` compiles a fresh regex, reads the
-    stream from offset 0, strips ANSI for `kind="rendered"`, converts
-    the full buffer to a string, and splits it into lines before
-    applying `limit`. This is meaningful when agents poll the same
-    pattern; it is low impact for ad hoc searches.
-  - Fix direction: cache compiled regexes by pattern; cache stripped
-    rendered output by child id and stream end offset; avoid
-    `strings.Split` over the whole ring when only the first `limit`
-    matches are needed. Prefer an incremental search shape if this
-    becomes the standard "watch for marker" path.
-
-# On Hold
-- [ ] There's a unicode <?> being displayed in opencode [ON HOLD]
-  - Investigated 2026-05-14: patterm passes ghostty grapheme codepoints
-    through unchanged (vt/ghostty.go:452-462), so the `<?>` glyph is
-    most likely the *host* terminal's font fallback for opencode's
-    Nerd Font private-use codepoints, not a patterm substitution.
-    Need a concrete reproduction (which codepoint, which host
-    terminal/font) before changing rendering.
-- [ ] After codex rips for like 15 minutes, the terminal becomes quite slow. [ON HOLD / VERIFYING]
-  - 2026-05-14: Perf plan P1-P11 landed (see CHANGELOG). Needs a real
-    long-running codex session to confirm whether the steady-state
-    slowdown is gone or some hotspot remains. Capture a pprof if it
-    still feels slow after ≥15 minutes — the structural drivers the
-    audit named are all addressed, so a remaining symptom is a new
-    one and probably wants fresh profiling.
+The close action in the command palette should just be "Close current agent" rather than "Close codex"
+Same with the other "focused" parts. It seems a bit clunky right now. "Close current agent"
+In general I think while the feature set has grown, the actual refinement of it isn't great, it feels a bit cluttered.

internal/app/app.go

@@ -432,10 +432,11 @@ type uiState struct {
 	repaintNextPTY       string
 	repaintNextPTYBudget int
 
-	// attention is the latest request_human_attention surfaced via MCP;
-	// rendered in the status line until cleared.
-	attentionText string
-	attentionAt   string
+	// toasts is the stackable notification surface. flashError,
+	// flashTransient, and notifyAttention all push onto it; the user
+	// dismisses entries with Ctrl-N or the "Clear notifications"
+	// palette command.
+	toasts toastStack
 
 	// pendingTrust is the most recent trust prompt — surfaced in the
 	// status line until the user resolves it with Ctrl-K. v1 keeps the
@@ -722,20 +723,15 @@ func (st *uiState) updateActiveAgentLocked(c *Child) {
 }
 
 // notifyAttention is the request_human_attention sink (SPEC §7). We
-// surface a one-line toast in the status row and remember the most
-// recent ask so the status line keeps showing it. The sidebar-blink is
-// deferred until the §4 chrome lands.
+// push a toast onto the stack; the focused-pane render path picks it
+// up. The sidebar-blink is deferred until the §4 chrome lands.
 func (st *uiState) notifyAttention(childID, reason string) {
 	c := st.sess.FindChild(childID)
 	name := childID
 	if c != nil {
 		name = c.DisplayName()
 	}
-	st.mu.Lock()
-	st.attentionText = fmt.Sprintf("attention: %s — %s", name, reason)
-	st.attentionAt = childID
-	st.mu.Unlock()
-	st.drawStatusLine()
+	st.notifyToast(toastAttention, fmt.Sprintf("%s — %s", name, reason))
 }
 
 func (st *uiState) scratchpadsChanged() {
@@ -1167,8 +1163,6 @@ func (st *uiState) drawStatusLine() {
 	palOpen := st.palette != nil
 	focusID := st.focusedID
 	focusName := st.focusedName
-	attention := st.attentionText
-	attentionAt := st.attentionAt
 	var trustMsg string
 	if st.pendingTrust != nil {
 		trustMsg = fmt.Sprintf("trust preset %q? [y]es / [n]o", st.pendingTrust.presetName)
@@ -1208,13 +1202,6 @@ func (st *uiState) drawStatusLine() {
 			left = owner
 		}
 	}
-	if attention != "" && attentionAt == focusID {
-		left = "[!] " + attention
-	}
-	if attention != "" && attentionAt == "" {
-		// Sticky attention/flash from somewhere outside the focused pane.
-		left = "[!] " + attention
-	}
 	if trustMsg != "" {
 		left = "[trust] " + trustMsg
 	}
@@ -1270,8 +1257,6 @@ func (st *uiState) drawStatusLine() {
 // child is focused.
 func (st *uiState) renderEmptyState() {
 	layout := st.layoutSnapshot()
-	st.outMu.Lock()
-	defer st.outMu.Unlock()
 	line := "Press Ctrl-K to spawn an agent or process"
 	row := int(layout.mainTop) + (int(layout.childRows()) / 2)
 	col := int(layout.mainLeft) + ((int(layout.childCols()) - len(line)) / 2)
@@ -1281,7 +1266,10 @@ func (st *uiState) renderEmptyState() {
 	if col < int(layout.mainLeft) {
 		col = int(layout.mainLeft)
 	}
+	st.outMu.Lock()
 	fmt.Fprintf(os.Stdout, "\x1b[?25l\x1b[H\x1b[2J\x1b[%d;%dH\x1b[2m%s\x1b[0m", row, col, line)
+	st.outMu.Unlock()
+	st.renderToasts()
 }
 
 func (st *uiState) hostSizeSnapshot() (uint16, uint16) {
@@ -1412,6 +1400,7 @@ func (st *uiState) processStdin(chunk []byte) {
 	var pendingViewportBottom bool
 	var pendingPadStep int
 	var pendingPadExit bool
+	var pendingDismissToast bool
 
 	flushForward := func() {
 		if len(forward) == 0 {
@@ -1598,6 +1587,11 @@ func (st *uiState) processStdin(chunk []byte) {
 			} else if hit, _ := matchCtrlChar(chunk, i, 'd'); hit {
 			} else if hit, _ := matchCtrlChar(chunk, i, 'w'); hit {
 			} else if hit, _ := matchCtrlChar(chunk, i, 's'); hit {
+			} else if hit, _ := matchCtrlChar(chunk, i, 'n'); hit {
+				// Ctrl-N is the toast dismiss key. In pad view we
+				// allow it through the chord block so the handler
+				// below can fire even though pads otherwise swallow
+				// bytes.
 			} else {
 				i++
 				continue
@@ -1696,6 +1690,22 @@ func (st *uiState) processStdin(chunk []byte) {
 				break
 			}
 		}
+		// Ctrl-N dismisses the most recent toast. We only consume the
+		// chord when there's actually a toast to dismiss; otherwise the
+		// bytes fall through to the focused PTY so readline /
+		// nano / emacs / opencode keep working in shells and editors.
+		if hit, adv := matchCtrlChar(chunk, i, 'n'); hit {
+			if st.toasts.length() > 0 {
+				flushForward()
+				pendingDismissToast = true
+				i += adv
+				continue
+			}
+			forward = append(forward, chunk[i:i+adv]...)
+			i += adv
+			continue
+		}
+
 		// Ctrl-B snaps the focused child's emulator viewport back to the
 		// active area. Use this as the escape hatch from a scrolled-up
 		// state — wheel scrolls move the viewport into the libghostty
@@ -1777,6 +1787,11 @@ func (st *uiState) processStdin(chunk []byte) {
 	if pendingPadExit {
 		st.exitPadView()
 	}
+	if pendingDismissToast {
+		if st.toasts.dismissTop() {
+			st.refreshToastSurface()
+		}
+	}
 }
 
 // scrollFocusedViewport scrolls the focused child's emulator viewport by
@@ -1985,6 +2000,11 @@ func (st *uiState) closePalette(action paletteAction) {
 	case "quit":
 		st.requestExit()
 
+	case "toasts-clear":
+		if st.toasts.clear() {
+			st.refreshToastSurface()
+		}
+
 	case "pad-delete":
 		st.handlePadDelete(action.padName)
 
@@ -2261,28 +2281,18 @@ func (st *uiState) handleProcRestart(childID string) {
 	st.drawStatusLine()
 }
 
-// flashError surfaces a spawn/etc. failure in the status line until the
-// next attention update overwrites it. stderr is hidden under the alt
-// screen so we can't rely on Fprintln(os.Stderr).
+// flashError surfaces a spawn/etc. failure as an error toast over the
+// focused pane. stderr is hidden under the alt screen so we can't rely
+// on Fprintln(os.Stderr).
 func (st *uiState) flashError(msg string) {
-	st.mu.Lock()
-	st.attentionText = msg
-	st.attentionAt = "" // shows on every focus until cleared
-	st.mu.Unlock()
-	st.renderEmptyState()
-	st.drawTabBar()
-	st.drawSidebar()
-	st.drawStatusLine()
+	st.notifyToast(toastError, msg)
 }
 
 // flashTransient is the softer cousin of flashError used for
-// trust-prompt resolutions. Same status-line surface; the prefix differs.
+// trust-prompt resolutions and other ack-style notices. Same
+// stackable surface, info styling.
 func (st *uiState) flashTransient(msg string) {
-	st.mu.Lock()
-	st.attentionText = msg
-	st.attentionAt = ""
-	st.mu.Unlock()
-	st.drawStatusLine()
+	st.notifyToast(toastInfo, msg)
 }
 
 // repaintFocused redraws the current focused child's screen snapshot.
@@ -2326,8 +2336,9 @@ func (st *uiState) repaintFocused() {
 	}
 	st.mu.Unlock()
 	st.outMu.Lock()
-	defer st.outMu.Unlock()
 	_, _ = os.Stdout.Write(out)
+	st.outMu.Unlock()
+	st.renderToasts()
 }
 
 // repaintFocusedPad paints the focused scratchpad's content into the
@@ -2351,8 +2362,9 @@ func (st *uiState) repaintFocusedPad() {
 		return
 	}
 	st.outMu.Lock()
-	defer st.outMu.Unlock()
 	_, _ = os.Stdout.Write(out)
+	st.outMu.Unlock()
+	st.renderToasts()
 }
 
 // renderPadView builds the bytes that paint a scratchpad's content

internal/app/palette.go

@@ -353,6 +353,12 @@ func (p *paletteState) buildItems(macro string) []paletteItem {
 		action: paletteAction{kind: "settings-open"},
 		group:  groupSettings,
 	})
+	out = append(out, paletteItem{
+		label:  "Clear notifications",
+		hint:   "dismiss all toasts in the top-right of the focused pane",
+		action: paletteAction{kind: "toasts-clear"},
+		group:  groupSettings,
+	})
 
 	// Group 4: Quit.
 	out = append(out, paletteItem{

internal/app/toast.go

@@ -0,0 +1,288 @@
+package app
+
+import (
+	"fmt"
+	"os"
+	"strings"
+	"sync"
+)
+
+// toastKind classifies a toast for styling and for migrating the
+// pre-existing flashError / flashTransient / notifyAttention call
+// sites onto the new stack.
+type toastKind int
+
+const (
+	toastInfo toastKind = iota
+	toastError
+	toastAttention
+)
+
+// toast is one entry in the host-level notification stack. Toasts
+// persist until the user dismisses them with Ctrl-N or the
+// "Clear notifications" palette command — there's no auto-expiry.
+type toast struct {
+	id   uint64
+	kind toastKind
+	text string
+}
+
+// toastStackCap caps how many toasts can be visible at once.
+// Older entries drop off the bottom when a new push would exceed it.
+const toastStackCap = 5
+
+// toastBoxMaxWidth bounds the rendered box width so a wide pane
+// doesn't produce huge toasts. Boxes shrink below this when the pane
+// is narrow.
+const toastBoxMaxWidth = 50
+
+// toastBoxMinWidth is the floor below which we refuse to render —
+// any narrower and there's not enough room for borders + content.
+const toastBoxMinWidth = 20
+
+// toastStack owns the ordered list of live toasts. Oldest at
+// index 0, newest (visually topmost) at the end. The stack's own
+// mutex is intentionally separate from uiState.mu so push / dismiss
+// can be called from any goroutine without participating in the
+// host's bigger lock-ordering rules.
+type toastStack struct {
+	mu    sync.Mutex
+	items []toast
+	next  uint64
+}
+
+func (s *toastStack) push(kind toastKind, text string) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.next++
+	s.items = append(s.items, toast{id: s.next, kind: kind, text: text})
+	if len(s.items) > toastStackCap {
+		s.items = s.items[len(s.items)-toastStackCap:]
+	}
+}
+
+// dismissTop pops the most recent toast (the one rendered at the
+// top of the stack). Returns true if something was removed so
+// callers can decide whether to repaint.
+func (s *toastStack) dismissTop() bool {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if len(s.items) == 0 {
+		return false
+	}
+	s.items = s.items[:len(s.items)-1]
+	return true
+}
+
+func (s *toastStack) clear() bool {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if len(s.items) == 0 {
+		return false
+	}
+	s.items = s.items[:0]
+	return true
+}
+
+func (s *toastStack) snapshot() []toast {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if len(s.items) == 0 {
+		return nil
+	}
+	out := make([]toast, len(s.items))
+	copy(out, s.items)
+	return out
+}
+
+func (s *toastStack) length() int {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return len(s.items)
+}
+
+// notifyToast is the single entry point that the former flash
+// helpers now delegate to. It pushes onto the stack and triggers a
+// repaint of the focused surface so the new toast appears
+// immediately; the repaint path also re-renders the stack on top.
+func (st *uiState) notifyToast(kind toastKind, text string) {
+	st.toasts.push(kind, text)
+	st.refreshToastSurface()
+}
+
+// refreshToastSurface re-renders whatever surface the toasts are
+// drawn over (focused child, focused pad, or the empty-state
+// canvas). Each of those paths calls renderToasts at the end, so
+// the toast layer is always reapplied on top of a freshly-drawn
+// pane. Centralised so push / dismiss / clear share one code path.
+func (st *uiState) refreshToastSurface() {
+	st.mu.Lock()
+	focusedPad := st.focusedPad
+	focusedID := st.focusedID
+	palOpen := st.palette != nil
+	st.mu.Unlock()
+	if palOpen {
+		// Palette owns the whole screen while it's open; toasts will
+		// repaint via closePalette's restore path.
+		return
+	}
+	switch {
+	case focusedPad != "":
+		st.repaintFocusedPad()
+	case focusedID != "":
+		st.repaintFocused()
+	default:
+		st.renderEmptyState()
+	}
+}
+
+// renderToasts draws the toast stack over the top-right of the
+// focused pane. Called from repaintFocused / repaintFocusedPad /
+// renderEmptyState after they finish so toasts always sit on top of
+// freshly-redrawn pane content. Safe to call when the stack is
+// empty (no-op).
+func (st *uiState) renderToasts() {
+	items := st.toasts.snapshot()
+	if len(items) == 0 {
+		return
+	}
+	st.mu.Lock()
+	palOpen := st.palette != nil
+	st.mu.Unlock()
+	if palOpen {
+		return
+	}
+	layout := st.layoutSnapshot()
+	paneCols := int(layout.childCols())
+	paneRows := int(layout.childRows())
+	if paneCols < toastBoxMinWidth+2 || paneRows < 3 {
+		return
+	}
+	boxWidth := toastBoxMaxWidth
+	if max := paneCols - 4; max < boxWidth {
+		boxWidth = max
+	}
+	if boxWidth < toastBoxMinWidth {
+		return
+	}
+	contentWidth := boxWidth - 4 // 2 border cells + 2 inner padding
+
+	var b strings.Builder
+	b.WriteString("\x1b7\x1b[?25l")
+
+	row := int(layout.mainTop) + 1
+	col := int(layout.mainLeft) + paneCols - boxWidth - 1
+	if col < int(layout.mainLeft) {
+		col = int(layout.mainLeft)
+	}
+
+	// Render newest first (visually on top), iterating items in
+	// reverse so the most recent push lands at the smallest row.
+	for idx := len(items) - 1; idx >= 0; idx-- {
+		t := items[idx]
+		isTopmost := idx == len(items)-1
+		hintLine := ""
+		if isTopmost && len(items) > 1 {
+			hintLine = fmt.Sprintf("Ctrl-N · %d more", len(items)-1)
+		}
+		height := 3
+		if hintLine != "" {
+			height++
+		}
+		// Stop if we'd run off the bottom of the pane.
+		if row+height > int(layout.mainTop)+paneRows {
+			break
+		}
+		border := toastBorderStyle(t.kind)
+
+		// Top border.
+		moveTo(&b, row, col)
+		b.WriteString(border)
+		b.WriteString("╭")
+		b.WriteString(strings.Repeat("─", boxWidth-2))
+		b.WriteString("╮")
+		b.WriteString(styleReset)
+		row++
+
+		// Content row.
+		moveTo(&b, row, col)
+		b.WriteString(border)
+		b.WriteString("│")
+		b.WriteString(styleReset)
+		b.WriteString(" ")
+		b.WriteString(toastIcon(t.kind))
+		body := t.text
+		bodyRoom := contentWidth - 2 // icon + space
+		if visibleLen(body) > bodyRoom {
+			body = clipRunes(body, bodyRoom-1) + "…"
+		}
+		b.WriteString(body)
+		b.WriteString(strings.Repeat(" ", max(0, bodyRoom-visibleLen(body))))
+		b.WriteString(" ")
+		b.WriteString(border)
+		b.WriteString("│")
+		b.WriteString(styleReset)
+		row++
+
+		// Hint row (topmost only, when stack has more than one).
+		if hintLine != "" {
+			if visibleLen(hintLine) > contentWidth {
+				hintLine = clipRunes(hintLine, contentWidth-1) + "…"
+			}
+			moveTo(&b, row, col)
+			b.WriteString(border)
+			b.WriteString("│")
+			b.WriteString(styleReset)
+			b.WriteString(" ")
+			b.WriteString(styleHint)
+			b.WriteString(hintLine)
+			b.WriteString(styleReset)
+			b.WriteString(strings.Repeat(" ", max(0, contentWidth-visibleLen(hintLine))))
+			b.WriteString(" ")
+			b.WriteString(border)
+			b.WriteString("│")
+			b.WriteString(styleReset)
+			row++
+		}
+
+		// Bottom border.
+		moveTo(&b, row, col)
+		b.WriteString(border)
+		b.WriteString("╰")
+		b.WriteString(strings.Repeat("─", boxWidth-2))
+		b.WriteString("╯")
+		b.WriteString(styleReset)
+		row++
+
+		// 1-row gap between stacked toasts.
+		row++
+	}
+
+	b.WriteString("\x1b[?25h\x1b8")
+
+	st.outMu.Lock()
+	defer st.outMu.Unlock()
+	_, _ = os.Stdout.WriteString(b.String())
+}
+
+func toastBorderStyle(kind toastKind) string {
+	switch kind {
+	case toastError:
+		return styleError
+	case toastAttention:
+		return styleAccent
+	default:
+		return styleBorder
+	}
+}
+
+func toastIcon(kind toastKind) string {
+	switch kind {
+	case toastError:
+		return styleError + "✗ " + styleReset
+	case toastAttention:
+		return styleAccent + "! " + styleReset
+	default:
+		return styleHint + "• " + styleReset
+	}
+}

internal/app/toast_test.go

@@ -0,0 +1,100 @@
+package app
+
+import "testing"
+
+func TestToastStackPushAndOrder(t *testing.T) {
+	var s toastStack
+	s.push(toastInfo, "one")
+	s.push(toastError, "two")
+	s.push(toastAttention, "three")
+
+	snap := s.snapshot()
+	if len(snap) != 3 {
+		t.Fatalf("snapshot len = %d, want 3", len(snap))
+	}
+	if snap[0].text != "one" || snap[1].text != "two" || snap[2].text != "three" {
+		t.Fatalf("snapshot order wrong: %#v", snap)
+	}
+	if snap[0].kind != toastInfo || snap[1].kind != toastError || snap[2].kind != toastAttention {
+		t.Fatalf("snapshot kinds wrong: %#v", snap)
+	}
+	// IDs strictly increase.
+	if !(snap[0].id < snap[1].id && snap[1].id < snap[2].id) {
+		t.Fatalf("ids not increasing: %#v", snap)
+	}
+}
+
+func TestToastStackCapDropsOldest(t *testing.T) {
+	var s toastStack
+	for i := 0; i < toastStackCap+3; i++ {
+		s.push(toastInfo, "msg")
+	}
+	snap := s.snapshot()
+	if len(snap) != toastStackCap {
+		t.Fatalf("len = %d, want %d", len(snap), toastStackCap)
+	}
+	// The earliest IDs should have been dropped, leaving the highest
+	// toastStackCap IDs.
+	for i := 1; i < len(snap); i++ {
+		if snap[i].id <= snap[i-1].id {
+			t.Fatalf("ordering broken after cap: %#v", snap)
+		}
+	}
+	// First retained id should be 4 (1,2,3 dropped; cap=5 leaves 4..8).
+	want := uint64(toastStackCap + 3 - toastStackCap + 1)
+	if snap[0].id != want {
+		t.Fatalf("first retained id = %d, want %d", snap[0].id, want)
+	}
+}
+
+func TestToastStackDismissTop(t *testing.T) {
+	var s toastStack
+	if s.dismissTop() {
+		t.Fatalf("dismissTop on empty stack returned true")
+	}
+	s.push(toastInfo, "a")
+	s.push(toastError, "b")
+	if !s.dismissTop() {
+		t.Fatalf("dismissTop returned false with items present")
+	}
+	snap := s.snapshot()
+	if len(snap) != 1 || snap[0].text != "a" {
+		t.Fatalf("after dismissTop: %#v", snap)
+	}
+	if !s.dismissTop() {
+		t.Fatalf("dismissTop on last item returned false")
+	}
+	if s.length() != 0 {
+		t.Fatalf("length after final dismiss = %d, want 0", s.length())
+	}
+}
+
+func TestToastStackClear(t *testing.T) {
+	var s toastStack
+	if s.clear() {
+		t.Fatalf("clear on empty returned true")
+	}
+	s.push(toastInfo, "a")
+	s.push(toastError, "b")
+	s.push(toastAttention, "c")
+	if !s.clear() {
+		t.Fatalf("clear returned false with items present")
+	}
+	if s.length() != 0 {
+		t.Fatalf("length after clear = %d, want 0", s.length())
+	}
+	if snap := s.snapshot(); snap != nil {
+		t.Fatalf("snapshot after clear = %#v, want nil", snap)
+	}
+}
+
+func TestToastStackSnapshotIsCopy(t *testing.T) {
+	var s toastStack
+	s.push(toastInfo, "a")
+	snap := s.snapshot()
+	snap[0].text = "mutated"
+	again := s.snapshot()
+	if again[0].text != "a" {
+		t.Fatalf("snapshot is not an independent copy: %#v", again)
+	}
+}

internal/harness/scenarios/error_flash_preserves_focused_pane.json

@@ -0,0 +1,37 @@
+{
+  "name": "error_flash_preserves_focused_pane",
+  "presets": {
+    "processes": [
+      {
+        "name": "steady",
+        "argv": ["sh", "-lc", "printf 'STEADY READY\\n'; sleep 5"]
+      }
+    ]
+  },
+  "trust": ["steady"],
+  "steps": [
+    {
+      "type": "mcp_call",
+      "method": "spawn_process",
+      "params": {"kind": "command", "preset": "steady", "name": "steady"},
+      "save_as": "proc"
+    },
+    { "type": "wait_text", "contains": "STEADY READY", "timeout_ms": 5000 },
+    { "type": "send_chord", "chord": "ctrl-k" },
+    { "type": "send_text", "text": "Open Settings" },
+    { "type": "send_chord", "chord": "enter" },
+    { "type": "send_chord", "chord": "enter" },
+    { "type": "send_chord", "chord": "ctrl-n" },
+    { "type": "send_chord", "chord": "ctrl-n" },
+    { "type": "send_chord", "chord": "ctrl-n" },
+    { "type": "send_chord", "chord": "ctrl-n" },
+    { "type": "send_chord", "chord": "ctrl-n" },
+    { "type": "send_chord", "chord": "ctrl-n" },
+    { "type": "send_chord", "chord": "ctrl-n" },
+    { "type": "send_chord", "chord": "enter" },
+    { "type": "wait_text", "contains": "no active top-level agent to summarize", "timeout_ms": 5000 },
+    { "type": "wait_text", "contains": "STEADY READY", "timeout_ms": 5000 },
+    { "type": "assert_contains", "contains": "STEADY READY" },
+    { "type": "assert_not_contains", "contains": "Press Ctrl-K to spawn an agent or process" }
+  ]
+}

internal/harness/scenarios/toast_dismiss.json

@@ -0,0 +1,32 @@
+{
+  "name": "toast_dismiss",
+  "presets": {
+    "processes": [
+      {
+        "name": "steady",
+        "argv": ["sh", "-lc", "printf 'STEADY READY\\n'; sleep 30"]
+      }
+    ]
+  },
+  "trust": ["steady"],
+  "steps": [
+    {
+      "type": "mcp_call",
+      "method": "spawn_process",
+      "params": {"kind": "command", "preset": "steady", "name": "steady"},
+      "save_as": "proc"
+    },
+    { "type": "wait_text", "contains": "STEADY READY", "timeout_ms": 5000 },
+    {
+      "type": "mcp_call",
+      "method": "request_human_attention",
+      "params": {"process_id": "{{proc.process_id}}", "reason": "needs eyes on the deploy"}
+    },
+    { "type": "wait_text", "contains": "needs eyes on the deploy", "timeout_ms": 5000 },
+    { "type": "assert_contains", "contains": "STEADY READY" },
+    { "type": "send_chord", "chord": "ctrl-n" },
+    { "type": "wait_stable", "timeout_ms": 2000 },
+    { "type": "assert_contains", "contains": "STEADY READY" },
+    { "type": "assert_not_contains", "contains": "needs eyes on the deploy" }
+  ]
+}