diff --git a/frontend/src/lib/components/copilot/chat/AIChatDisplay.svelte b/frontend/src/lib/components/copilot/chat/AIChatDisplay.svelte
index ce192356c6..bbeb77b968 100644
--- a/frontend/src/lib/components/copilot/chat/AIChatDisplay.svelte
+++ b/frontend/src/lib/components/copilot/chat/AIChatDisplay.svelte
@@ -501,8 +501,8 @@
 	)
 	// Only when this tab could hold the posture itself: a mode its own selector
 	// does not offer would read as a mode the user could switch away from here.
-	const showsMirroredPlan = $derived(
-		aiChatManager.mirroredPlanMode && aiChatManager.planModeAvailable
+	const showsRemotePlan = $derived(
+		aiChatManager.remotePlanMode && aiChatManager.planModeAvailable
 	)
 	// Fall back to ask-permission when the persisted mode isn't applicable in the
 	// current AI mode (e.g. auto-accept edits while in a mode without edits).
@@ -510,7 +510,7 @@
 	// run this tab is showing, and this tab's own mode governs nothing until it
 	// drives a turn itself.
 	const effectiveAutonomyMode = $derived(
-		showsMirroredPlan
+		showsRemotePlan
 			? AIAutonomyMode.PLAN
 			: availableAutonomyModeOptions.some((option) => option.mode === aiChatManager.autonomyMode)
 				? aiChatManager.autonomyMode
@@ -535,16 +535,16 @@
 		return pending?.action === 'question' ? pending.toolCallId : undefined
 	})
 
-	// The composer is locked while another tab's run is on screen: this tab pairs
-	// a mirrored transcript with the history it held before that run, and a turn
-	// sent from that pair would reach the model as a conversation the driver has
-	// already moved past. It unlocks on its own once the re-read that follows the
-	// turn lands. A run parked on a question is the exception — answering is the
-	// one thing a watching tab is there to do, and the answer travels to the
-	// driver instead of starting a turn here.
-	const composerLocked = $derived(
-		aiChatManager.mirroringRemoteRun && pendingQuestionToolCallId === undefined
-	)
+	// The composer is locked for as long as another tab's run is in flight: this
+	// tab still shows the transcript from before that run, and a turn sent from
+	// it would reach the model as a conversation the driver has already moved
+	// past. It unlocks on its own once the re-read that follows the turn lands.
+	//
+	// No exception for a run parked on a question. Only the driving tab holds the
+	// question and the resolver waiting on it; a card that looks parked here is a
+	// restored one whose resolver left with the old page, and answering it would
+	// deliver to nobody.
+	const composerLocked = $derived(aiChatManager.runHeldElsewhere)
 
 	// Get app context for display when in APP mode
 	const appContext = $derived.by((): SelectedContext | undefined => {
@@ -639,10 +639,14 @@ the panel, or the Escape-to-stop focus check would wrongly reject them. -->
 									{#each pastChats as chat (chat.id)}
 										<button
 											class="text-left flex flex-row items-center gap-2 justify-between hover:bg-gray-100 dark:hover:bg-gray-700 rounded-md p-1 disabled:opacity-50 disabled:cursor-not-allowed disabled:hover:bg-transparent dark:disabled:hover:bg-transparent"
-											disabled={aiChatManager.loading || aiChatManager.sendInFlight}
-											title={aiChatManager.loading || aiChatManager.sendInFlight
-												? 'Stop the current answer to switch conversation'
-												: undefined}
+											disabled={aiChatManager.loading ||
+												aiChatManager.sendInFlight ||
+												aiChatManager.runHeldElsewhere}
+											title={aiChatManager.runHeldElsewhere
+												? 'This session is running in another tab'
+												: aiChatManager.loading || aiChatManager.sendInFlight
+													? 'Stop the current answer to switch conversation'
+													: undefined}
 											onclick={() => {
 												loadPastChat(chat.id)
 												close()
@@ -654,11 +658,15 @@ the panel, or the Escape-to-stop focus check would wrongly reject them. -->
 											>
 												{chat.title}
 											</div>
+											<!-- Gated separately: it sits inside the disabled row button,
+											     which does not disable it, and the catch-up re-read is
+											     keyed on a chat id this would delete out from under it. -->
 											<Button
 												iconOnly
 												size="xs2"
 												btnClasses="!p-1"
 												variant="default"
+												disabled={aiChatManager.runHeldElsewhere}
 												startIcon={{ icon: X }}
 												on:click={() => {
 													deletePastChat(chat.id)
@@ -672,10 +680,10 @@ the panel, or the Escape-to-stop focus check would wrongly reject them. -->
 					{/snippet}
 				</Popover>
 				<Button
-					title={aiChatManager.mirroringRemoteRun
+					title={aiChatManager.runHeldElsewhere
 						? 'This session is running in another tab'
 						: 'New chat'}
-					disabled={aiChatManager.mirroringRemoteRun}
+					disabled={aiChatManager.runHeldElsewhere}
 					on:click={() => {
 						saveAndClear()
 					}}
@@ -976,7 +984,7 @@ the panel, or the Escape-to-stop focus check would wrongly reject them. -->
 								onchange={onFolderInputChange}
 							/>
 						{/if}
-						{#if showsMirroredPlan}
+						{#if showsRemotePlan}
 							<!-- The posture belongs to the tab running the turn, so it reads here
 							     rather than offering a choice that would only move this tab's own
 							     preference while the label stayed put. -->
diff --git a/frontend/src/lib/components/copilot/chat/AIChatInput.svelte b/frontend/src/lib/components/copilot/chat/AIChatInput.svelte
index f21de0a209..341d14fd9b 100644
--- a/frontend/src/lib/components/copilot/chat/AIChatInput.svelte
+++ b/frontend/src/lib/components/copilot/chat/AIChatInput.svelte
@@ -133,10 +133,10 @@
 			return 'Answer the question above'
 		}
 
-		// Ahead of 'Ask followup': a mirrored run is almost always mid-conversation,
-		// so a check below that one would never be reached, and the locked composer
-		// would sit there inviting a followup it will not accept.
-		if (aiChatManager.mirroringRemoteRun) {
+		// Ahead of 'Ask followup': a run held elsewhere is almost always
+		// mid-conversation, so a check below that one would never be reached, and the
+		// locked composer would sit there inviting a followup it will not accept.
+		if (aiChatManager.runHeldElsewhere) {
 			return 'Running in another tab'
 		}
 
@@ -704,6 +704,15 @@
 		// Read before `take()` empties the draft the id derives from, and only take
 		// once the answer is delivered — an undelivered one would leave the user
 		// with neither their text nor a resumed turn.
+		// Ahead of every `draft.take()` below, including the edit branch: a send
+		// refused after the draft is consumed loses the user's text, and
+		// restartGeneration has truncated the transcript by then as well. The
+		// composer is disabled while another tab drives, so this catches the ways
+		// in that bypass it — Enter, and submitting an edit.
+		if (aiChatManager.runHeldElsewhere) {
+			sendUserToast('This session is running in another tab. Continue it there.')
+			return
+		}
 		const answeredQuestionId = questionAnsweredBySend
 		if (
 			answeredQuestionId &&
@@ -716,16 +725,6 @@
 			return
 		}
 		if (aiChatManager.loading) {
-			// The composer is locked while another tab drives, with one exception:
-			// a run parked on a question stays answerable from here. An answer
-			// carrying an attachment misses the branch above (it only routes plain
-			// choices) and would land on the queue — which in this tab nothing
-			// drains, because the turn that would flush it belongs elsewhere. Keep
-			// the draft where the user put it and say what will send.
-			if (aiChatManager.mirroringRemoteRun) {
-				sendUserToast('This session is running in another tab. Answer with text to send it there.')
-				return
-			}
 			// Queue the message instead of silently discarding it — it is auto-sent
 			// when this tab's own streaming turn completes successfully.
 			// Editing-while-loading keeps the old discard behavior. Paste
diff --git a/frontend/src/lib/components/copilot/chat/AIChatManager.svelte.ts b/frontend/src/lib/components/copilot/chat/AIChatManager.svelte.ts
index c72afd16aa..ed6f8325d0 100644
--- a/frontend/src/lib/components/copilot/chat/AIChatManager.svelte.ts
+++ b/frontend/src/lib/components/copilot/chat/AIChatManager.svelte.ts
@@ -61,7 +61,7 @@ import { SvelteMap, SvelteSet } from 'svelte/reactivity'
 import { createLongHash } from '$lib/editorLangUtils'
 import type { AIProvider, UserDraftItemKind } from '$lib/gen'
 import { maskKey } from '$lib/components/sessions/modifiedItemsMask'
-import { isMirroring, runPosition } from '$lib/components/sessions/sessionRunOwner.svelte'
+import { runHeldElsewhere, runPosition } from '$lib/components/sessions/sessionRunOwner.svelte'
 import { getStringError } from './utils'
 import { type PasteAttachment } from './pasteTokens'
 import {
@@ -695,17 +695,16 @@ export class AIChatManager {
 	remoteToolConfirmation: ((toolId: string, confirmed: boolean) => boolean) | undefined = undefined
 	remoteQuestionAnswer: ((toolId: string, choices: string[]) => boolean) | undefined = undefined
 
-	/** True while this manager is showing a run another tab is driving, and on
-	 *  through the re-read that follows it. Its transcript then comes from mirror
-	 *  frames while `messages` still holds the pre-run history, so the pair is
-	 *  mismatched and must not be written to the shared record — the driving tab
-	 *  owns it until the re-read lands.
+	/** True while another tab is driving this session's run, and on through the
+	 *  re-read that follows it. What this manager holds is the conversation from
+	 *  before that turn, so it must not be written to the shared record — the
+	 *  driving tab owns it until the re-read lands.
 	 *
 	 *  Derived rather than stored: the position is what the runtime moves, and a
 	 *  second copy of it here could only ever be a copy that disagrees. A chat
-	 *  with no session (the docked copilot) is never mirroring. */
-	get mirroringRemoteRun(): boolean {
-		return isMirroring(this.sessionId)
+	 *  with no session (the docked copilot) never holds a run elsewhere. */
+	get runHeldElsewhere(): boolean {
+		return runHeldElsewhere(this.sessionId)
 	}
 
 	/** Whether the run on screen is one another tab is running in plan mode.
@@ -715,7 +714,7 @@ export class AIChatManager {
 	 *  only in the memory of the tab running the turn. It is read from the run's
 	 *  position for the same reason it is not stored there — the posture belongs
 	 *  to that run, and must not outlive it into a turn this tab drives. */
-	get mirroredPlanMode(): boolean {
+	get remotePlanMode(): boolean {
 		const position = runPosition(this.sessionId)
 		return position.state === 'watching' && position.planMode
 	}
@@ -804,7 +803,7 @@ export class AIChatManager {
 		// than deferred: the re-read when the run ends reseeds the mask from the
 		// driving tab's record, so a write held back here would be overwritten by
 		// it anyway — and writing now would put a mismatched pair in the record.
-		if (this.mirroringRemoteRun) return this.#maskPersistQueue
+		if (this.runHeldElsewhere) return this.#maskPersistQueue
 		this.#maskPersistQueue = this.#maskPersistQueue.then(() =>
 			this.historyManager
 				.saveChat(
@@ -1116,7 +1115,7 @@ export class AIChatManager {
 	#jobPersistQueue: Promise<void> = Promise.resolve()
 	#persistBackgroundJobs(): Promise<void> {
 		// Runs outside any turn, so the run guard never sees it.
-		if (this.mirroringRemoteRun) return this.#jobPersistQueue
+		if (this.runHeldElsewhere) return this.#jobPersistQueue
 		this.#jobPersistQueue = this.#jobPersistQueue.then(() =>
 			this.historyManager
 				.saveChat(
@@ -1540,9 +1539,9 @@ export class AIChatManager {
 	 * summary, leaving history untouched.
 	 */
 	compactManually = async (): Promise<void> => {
-		// `loading` is the mirrored driver's while another tab runs the session;
+		// `loading` reflects the driver's run while another tab has the session;
 		// compacting here would rewrite a conversation this tab does not own.
-		if (this.loading || this.mirroringRemoteRun) {
+		if (this.loading || this.runHeldElsewhere) {
 			return
 		}
 		// A summary round-trip only pays off once there's a prior exchange to fold
@@ -1803,7 +1802,7 @@ export class AIChatManager {
 		context?: ContextElement[],
 		files: AttachedTextFile[] = []
 	) {
-		if (this.mirroringRemoteRun) {
+		if (this.runHeldElsewhere) {
 			// Handed back, not dropped. These senders have no draft of their own to
 			// fall back on — the raw-app inline prompt would lose what the user
 			// typed outright. Not `restoreToInput`, whose no-composer fallback is
@@ -2926,12 +2925,12 @@ export class AIChatManager {
 			sendUserToast('This action needs the AI chat. Start an AI session to continue.', true)
 			return
 		}
-		// Still pairing a mirrored transcript with the pre-run history: sending now
-		// would put a conversation the driving tab has already moved past to the
-		// model, and persist it over what that tab saved. The composer is locked
-		// for exactly this window, so nothing the user typed reaches here — this
-		// catches the sends that start without one (tool handlers, auto-resume).
-		if (this.mirroringRemoteRun) {
+		// This tab holds the conversation from before the driver's turn: sending now
+		// would put one the driving tab has already moved past to the model, and
+		// persist it over what that tab saved. The composer is locked for exactly
+		// this window, so nothing the user typed reaches here — this catches the
+		// sends that start without one (tool handlers, auto-resume).
+		if (this.runHeldElsewhere) {
 			this.restoreRefusedSend(options)
 			return false
 		}
@@ -4080,7 +4079,7 @@ export class AIChatManager {
 	saveAndClear = async () => {
 		// The transcript on screen belongs to a run another tab owns; saving it
 		// here would write a history this manager does not have.
-		if (this.mirroringRemoteRun) return
+		if (this.runHeldElsewhere) return
 		this.cancel('saveAndClear')
 		// Drop any message queued in this conversation so it can't auto-send into
 		// the fresh chat or linger as a card across the switch.
diff --git a/frontend/src/lib/components/copilot/chat/AIChatManager.test.ts b/frontend/src/lib/components/copilot/chat/AIChatManager.test.ts
index c7840a6f62..f155bee011 100644
--- a/frontend/src/lib/components/copilot/chat/AIChatManager.test.ts
+++ b/frontend/src/lib/components/copilot/chat/AIChatManager.test.ts
@@ -294,10 +294,9 @@ describe('AIChatManager cross-tab run guard', () => {
 	})
 
 	// The save paths that run outside a turn have no guard over them, and a tab
-	// rendering someone else's run holds a transcript that does not match its own
-	// `messages` — writing that pair would clobber the record the driving tab is
-	// still appending to.
-	it('does not persist while rendering a run another tab owns', async () => {
+	// watching someone else's run holds the conversation from before it — writing
+	// that would clobber the record the driving tab is still appending to.
+	it('does not persist while another tab owns the run', async () => {
 		const manager = new AIChatManager()
 		manager.isSessionChat = true
 		manager.sessionId = 'session-watching'
@@ -319,16 +318,16 @@ describe('AIChatManager cross-tab run guard', () => {
 		manager.sessionId = 'session-planning'
 
 		noteDriverAlive('session-planning', true)
-		expect(manager.mirroredPlanMode).toBe(true)
+		expect(manager.remotePlanMode).toBe(true)
 
 		noteRemoteTurnEnded('session-planning')
-		expect(manager.mirroredPlanMode).toBe(false)
+		expect(manager.remotePlanMode).toBe(false)
 	})
 
 	// The other tab's run has ended but this one is still reading what it left
-	// behind, so its transcript is mirrored while `messages` is the pre-turn
-	// history. A send in that window would put the stale history to the model and
-	// persist it over the completed turn.
+	// behind, so it holds the conversation from before that turn. A send in that
+	// window would put the stale history to the model and persist it over the
+	// completed turn.
 	it('refuses to send while still catching up on a finished remote turn', async () => {
 		const manager = new AIChatManager()
 		manager.isSessionChat = true
diff --git a/frontend/src/lib/components/sessions/sessionMirror.test.ts b/frontend/src/lib/components/sessions/sessionMirror.test.ts
deleted file mode 100644
index eb06eb01c2..0000000000
--- a/frontend/src/lib/components/sessions/sessionMirror.test.ts
+++ /dev/null
@@ -1,70 +0,0 @@
-import { describe, expect, it } from 'vitest'
-import { withoutHeavyPayloads } from './sessionMirrorPayload'
-import type { DisplayMessage } from '$lib/components/copilot/chat/shared'
-
-// Frames go out several times a second for the whole turn, so anything left in
-// them is re-cloned and re-broadcast on every tick. The caps allow megabytes per
-// image and per pasted file, and a structural typecheck cannot catch a field
-// name that no longer exists — the message union is cast on the way out.
-describe('withoutHeavyPayloads', () => {
-	it('strips the bytes and keeps what the transcript renders', () => {
-		const messages = [
-			{
-				role: 'user',
-				content: 'look at this',
-				images: [
-					{ dataUrl: 'data:image/png;base64,AAAA', mediaType: 'image/png', name: 'shot.png' }
-				],
-				files: [{ name: 'notes.md', content: 'a very long pasted file', id: 'f1' }],
-				pastes: [{ id: 1, lines: 400, content: 'a very long collapsed paste' }]
-			},
-			{ role: 'tool', content: 'took a screenshot', imageUrl: 'data:image/png;base64,BBBB' }
-		] as unknown as DisplayMessage[]
-
-		const stripped = withoutHeavyPayloads(messages)
-
-		expect(JSON.stringify(stripped)).not.toContain('data:image')
-		expect(JSON.stringify(stripped)).not.toContain('a very long pasted file')
-		expect(JSON.stringify(stripped)).not.toContain('a very long collapsed paste')
-		// The chip renders from the line count, so that has to survive.
-		expect((stripped[0] as any).pastes).toEqual([{ id: 1, lines: 400, content: '' }])
-		// The file chip is labelled from the name, so that has to survive.
-		expect((stripped[0] as any).files).toEqual([{ name: 'notes.md', content: '', id: 'f1' }])
-		// Images are dropped whole: the bubble renders one <img> per entry with no
-		// per-image guard, so an emptied url would show a broken image instead.
-		expect((stripped[0] as any).images).toBeUndefined()
-		expect((stripped[0] as any).content).toBe('look at this')
-	})
-
-	// A tool's output is the one payload with no ceiling at all — a query result
-	// or job logs ride the running turn's tail, which is re-cloned and re-sent
-	// several times a second until the turn ends.
-	it('drops a tool result and its logs, keeping the card that frames them', () => {
-		const messages = [
-			{
-				role: 'tool',
-				tool_call_id: 'tc-1',
-				content: 'Ran the query',
-				toolName: 'run_script',
-				parameters: { path: 'f/demo/q' },
-				result: { rows: Array.from({ length: 5000 }, (_, i) => ({ i, blob: 'x'.repeat(200) })) },
-				logs: 'y'.repeat(100_000)
-			}
-		] as unknown as DisplayMessage[]
-
-		const stripped = withoutHeavyPayloads(messages)
-
-		expect(JSON.stringify(stripped).length).toBeLessThan(500)
-		expect((stripped[0] as any).result).toBeUndefined()
-		expect((stripped[0] as any).logs).toBeUndefined()
-		// The card is still rendered from these while the output is in flight.
-		expect((stripped[0] as any).toolName).toBe('run_script')
-		expect((stripped[0] as any).parameters).toEqual({ path: 'f/demo/q' })
-		expect((stripped[0] as any).content).toBe('Ran the query')
-	})
-
-	it('passes through a message with nothing heavy in it', () => {
-		const messages = [{ role: 'assistant', content: 'plain reply' }] as unknown as DisplayMessage[]
-		expect(withoutHeavyPayloads(messages)[0]).toBe(messages[0])
-	})
-})
diff --git a/frontend/src/lib/components/sessions/sessionMirrorFrame.test.ts b/frontend/src/lib/components/sessions/sessionMirrorFrame.test.ts
deleted file mode 100644
index 927f0069ab..0000000000
--- a/frontend/src/lib/components/sessions/sessionMirrorFrame.test.ts
+++ /dev/null
@@ -1,47 +0,0 @@
-import { describe, expect, it } from 'vitest'
-import { canSpliceFrame, mirrorFrameStart } from './sessionMirrorPayload'
-
-describe('mirrorFrameStart', () => {
-	// A resync exists to rescue a receiver whose prefix does not fit. Answering it
-	// with a frame that still starts mid-transcript fails the same check again and
-	// asks for another one, so the two tabs trade messages and nothing renders.
-	it('sends the whole transcript when full, even mid-turn', () => {
-		expect(mirrorFrameStart({ total: 40, turnStart: 36, full: true })).toBe(0)
-	})
-
-	it('reaches no further back than the running turn', () => {
-		expect(mirrorFrameStart({ total: 40, turnStart: 36, full: false })).toBe(36)
-	})
-
-	it('caps the tail when the turn itself is long', () => {
-		expect(mirrorFrameStart({ total: 40, turnStart: 2, full: false })).toBe(30)
-	})
-
-	it('handles a turn that has produced nothing yet', () => {
-		expect(mirrorFrameStart({ total: 0, turnStart: 0, full: false })).toBe(0)
-	})
-})
-
-describe('canSpliceFrame', () => {
-	const base = { baseIndex: 30, total: 40, localLength: 35, onSameChat: true }
-
-	it('accepts a tail that lands on a matching prefix', () => {
-		expect(canSpliceFrame(base)).toBe(true)
-	})
-
-	it('always accepts a full frame', () => {
-		expect(canSpliceFrame({ ...base, baseIndex: 0, onSameChat: false })).toBe(true)
-	})
-
-	it('rejects a tail when the receiver joined mid-run and has a gap', () => {
-		expect(canSpliceFrame({ ...base, localLength: 12 })).toBe(false)
-	})
-
-	it('rejects a tail when the receiver holds more than the sender has', () => {
-		expect(canSpliceFrame({ ...base, localLength: 44 })).toBe(false)
-	})
-
-	it('rejects a tail from a different conversation', () => {
-		expect(canSpliceFrame({ ...base, onSameChat: false })).toBe(false)
-	})
-})
diff --git a/frontend/src/lib/components/sessions/sessionMirrorPayload.ts b/frontend/src/lib/components/sessions/sessionMirrorPayload.ts
deleted file mode 100644
index 1e9649b0ac..0000000000
--- a/frontend/src/lib/components/sessions/sessionMirrorPayload.ts
+++ /dev/null
@@ -1,119 +0,0 @@
-import type { DisplayMessage } from '$lib/components/copilot/chat/shared'
-
-/**
- * The attachment payloads a mirror frame does not carry, and how to put them
- * back. A frame goes out several times a second for the whole turn, and the
- * message holding an image or a pasted file sits in its tail unchanged
- * throughout — so shipping the payloads means re-cloning and re-broadcasting
- * the same megabytes on every tick (the caps allow roughly 16MB of base64 image
- * plus 5MB of file content on a single message, before pastes).
- *
- * Every heavy field is listed here rather than at its own call site: they are
- * spread across the message union, and one that goes unlisted silently
- * reintroduces the whole regression.
- */
-
-/** Emptied in place, keeping the sibling fields the chip renders from. */
-const BLANKED_ITEM_FIELDS = [
-	['files', 'content'],
-	['pastes', 'content']
-] as const
-
-/** Dropped whole: the bubble renders one <img> per entry with no per-image
- *  guard, so an emptied url would show a broken image where the real one is
- *  about to appear. */
-const DROPPED_LIST_FIELDS = ['images'] as const
-
-/** Dropped whole, and unbounded in a way the others are not: a tool's result or
- *  logs can be a whole query result or job output, and a frame re-clones and
- *  re-broadcasts the running turn's tail several times a second. A watching tab
- *  shows the tool card without its output until the turn-end re-read supplies
- *  it, which is the same trade the images above make. */
-const DROPPED_FIELDS = ['result', 'logs'] as const
-
-/** Emptied outright; guarded at the render site, so it renders nothing. */
-const BLANKED_FIELDS = ['imageUrl'] as const
-
-function isHeavy(message: DisplayMessage): boolean {
-	const m = message as Record<string, any>
-	return (
-		DROPPED_LIST_FIELDS.some((f) => m[f]?.length) ||
-		DROPPED_FIELDS.some((f) => m[f] !== undefined) ||
-		BLANKED_ITEM_FIELDS.some(([f]) => m[f]?.length) ||
-		BLANKED_FIELDS.some((f) => m[f])
-	)
-}
-
-/** Strip the bytes out of a transcript bound for a mirror frame. */
-export function withoutHeavyPayloads(messages: DisplayMessage[]): DisplayMessage[] {
-	return messages.map((message) => {
-		if (!isHeavy(message)) return message
-		const stripped: Record<string, any> = { ...message }
-		for (const f of DROPPED_LIST_FIELDS) delete stripped[f]
-		for (const f of DROPPED_FIELDS) delete stripped[f]
-		for (const [f, item] of BLANKED_ITEM_FIELDS) {
-			if (stripped[f]?.length) stripped[f] = stripped[f].map((v: any) => ({ ...v, [item]: '' }))
-		}
-		for (const f of BLANKED_FIELDS) if (stripped[f]) stripped[f] = ''
-		return stripped as DisplayMessage
-	})
-}
-
-/** How many of the newest messages a frame carries when it is not sending the
- *  whole transcript. In-place edits to already-rendered cards (a tool card
- *  settling) land within a few messages of the end, so a short tail carries them
- *  while keeping the frame bounded on a long conversation. */
-const MIRROR_TAIL = 10
-
-/**
- * Where the tail a frame carries starts.
- *
- * `full` sends the whole transcript and overrides everything else — it is what
- * answers a resync, and a resync that came back partial would fail the receiver's
- * prefix check again and ask for another one, forever.
- *
- * Otherwise the frame reaches no further back than `turnStart`, the index of the
- * running turn's first message: everything from there on is either new this turn
- * or a stripped copy the receiver already got from an earlier frame of the same
- * turn, so overwriting it can never destroy a message held complete from the
- * store.
- */
-export function mirrorFrameStart({
-	total,
-	turnStart,
-	full
-}: {
-	total: number
-	turnStart: number
-	full: boolean
-}): number {
-	if (full) return 0
-	return Math.max(turnStart, Math.max(0, total - MIRROR_TAIL))
-}
-
-/**
- * Whether a receiver can splice this frame onto what it already holds, or has to
- * ask for the whole transcript instead.
- *
- * A frame is positional, so it is only meaningful against the same conversation
- * and a prefix of the same shape. A receiver holding fewer messages than the
- * frame starts at has a gap; one holding more than the sender has messages the
- * sender no longer does (it compacted, or switched chats), and splicing would
- * render a conversation that never existed.
- */
-export function canSpliceFrame({
-	baseIndex,
-	total,
-	localLength,
-	onSameChat
-}: {
-	baseIndex: number
-	total: number
-	localLength: number
-	onSameChat: boolean
-}): boolean {
-	// A frame that starts at 0 replaces everything, so it needs no prefix to
-	// agree with and can be adopted even when it names a different conversation.
-	if (baseIndex === 0) return true
-	return onSameChat && localLength >= baseIndex && localLength <= total
-}
diff --git a/frontend/src/lib/components/sessions/sessionRunOwner.svelte.ts b/frontend/src/lib/components/sessions/sessionRunOwner.svelte.ts
index 35394a33e4..0d65e00778 100644
--- a/frontend/src/lib/components/sessions/sessionRunOwner.svelte.ts
+++ b/frontend/src/lib/components/sessions/sessionRunOwner.svelte.ts
@@ -8,9 +8,9 @@ import { SvelteMap } from 'svelte/reactivity'
 // have already run against the workspace. So one tab drives and the rest watch.
 //
 // Every gate in the cross-tab paths asks a variant of one question: may I send,
-// is this frame mine to adopt, is what I am showing still the other tab's. This
-// module answers all of them from a single position per session, so they cannot
-// disagree with each other.
+// may I save, is the run on screen still someone else's. This module answers all
+// of them from a single position per session, so they cannot disagree with each
+// other.
 
 /** Whether this context can hold a real lock on a run. Web Locks is
  *  secure-context only, so a self-hosted instance served over plain HTTP has
@@ -18,17 +18,17 @@ import { SvelteMap } from 'svelte/reactivity'
  *  {@link withSessionRunLock} and {@link runLockHeld}. */
 const EXCLUSIVE_OWNERSHIP = BROWSER && !!globalThis.navigator?.locks?.query
 
-/** A driver with no frame for this long is presumed gone, pending the lock
- *  query that confirms it. Deliberately many times the frame cadence: a busy
- *  tab can be starved of ticks for a while without having gone anywhere. */
+/** A driver silent for this long is presumed gone, pending the lock query that
+ *  confirms it. Deliberately many times the status cadence: a busy tab can be
+ *  starved of ticks for a while without having gone anywhere. */
 const DRIVER_SILENCE_MS = 10_000
 
 /** Where this tab stands in one session's run.
  *
  *  `catchingUp` is its own position rather than a corner of `watching`: the
- *  driver's turn is over, but this tab still pairs a mirrored transcript with
- *  the history it held before that turn, and sending from that pair would put a
- *  conversation the driver has already moved past to the model. */
+ *  driver's turn is over, but this tab still holds the conversation from before
+ *  it, and sending from there would put one the driver has already moved past to
+ *  the model. */
 export type RunPosition =
 	| { state: 'idle' }
 	| { state: 'driving' }
@@ -60,24 +60,23 @@ export function isDriving(sessionId: string): boolean {
 	return runPosition(sessionId).state === 'driving'
 }
 
-/** Another tab is running the turn and this one is rendering its frames. */
+/** Another tab is running the turn and this one is showing that it is. */
 export function isWatching(sessionId: string): boolean {
 	return runPosition(sessionId).state === 'watching'
 }
 
-/** True from the first frame adopted until the re-read that follows the
- *  driver's turn completes. The save and send paths gate on it: until it
- *  clears, this tab's transcript and its model history are not the same
- *  conversation. */
-export function isMirroring(sessionId: string | undefined): boolean {
+/** True from the driver's first status message until the re-read that follows
+ *  its turn completes. The save and send paths gate on it: until it clears, what
+ *  this tab holds is the conversation from before the driver's turn. */
+export function runHeldElsewhere(sessionId: string | undefined): boolean {
 	const state = runPosition(sessionId).state
 	return state === 'watching' || state === 'catchingUp'
 }
 
-/** A frame arrived, which is both the transcript and the sign of life. */
+/** The driver said how its run is going, which is also its sign of life. */
 export function noteDriverAlive(sessionId: string, planMode: boolean): void {
-	// A tab mid-turn is the authority on its own session; a frame reaching it can
-	// only be an echo of the run it is itself driving.
+	// A tab mid-turn is the authority on its own session; a status message
+	// reaching it can only be an echo of the run it is itself driving.
 	if (isDriving(sessionId)) return
 	positions.set(sessionId, { state: 'watching', lastHeardAt: Date.now(), planMode })
 	ensureReaper()
@@ -88,7 +87,12 @@ export function noteDriverAlive(sessionId: string, planMode: boolean): void {
  *  to perform it. The channel runs this in every tab that loads it, and one with
  *  no runtime would sit in a state nothing can leave. */
 export function noteRemoteTurnEnded(sessionId: string): void {
-	if (runPosition(sessionId).state !== 'watching') return
+	const state = runPosition(sessionId).state
+	// `idle` counts. A tab that heard the turn end without having heard it start
+	// — it opened between the driver's last status message and its turn-end, or
+	// its handler registered in that window — holds a transcript from before the
+	// turn, and idle is precisely the position that permits driving on it.
+	if (state === 'driving' || state === 'catchingUp') return
 	if (!canCompleteCatchUp()) {
 		positions.delete(sessionId)
 		return
@@ -131,6 +135,12 @@ export async function withSessionRunLock<T>(
 	sessionId: string,
 	body: () => Promise<T>
 ): Promise<T | 'busy'> {
+	// Ahead of the lock, because holding the lock is not the same as being
+	// entitled to the conversation: the driver releases it at turn-end while this
+	// tab still owes itself the re-read, and a turn driven in that window sends
+	// the pre-run history the re-read exists to replace. `runHeldElsewhere` and not
+	// `isWatching` for exactly that reason.
+	if (runHeldElsewhere(sessionId)) return 'busy'
 	if (!EXCLUSIVE_OWNERSHIP) return await bestEffort(sessionId, body)
 	// Set before the body runs, so a turn that throws is told apart from a lock
 	// that could not be taken. Without it a failing turn would look like a failed
@@ -164,8 +174,8 @@ export async function withSessionRunLock<T>(
  *  run is visibly on screen, and otherwise go.
  *
  *  This is deliberately not mutual exclusion, and cannot be made into it. A
- *  driver whose timers have been throttled in a hidden tab stops sending frames
- *  long before its turn ends, so it is reaped as dead and a send from here can
+ *  driver whose timers have been throttled in a hidden tab goes silent long
+ *  before its turn ends, so it is reaped as dead and a send from here can
  *  start a second turn against the same chat id. Nothing over the channel fixes
  *  that: a probe distinguishes a throttled tab from a closed one, but not a
  *  frozen tab from a closed one, and the browser freezes hidden tabs on much the
@@ -174,12 +184,8 @@ export async function withSessionRunLock<T>(
  *  Accepted rather than solved, because the only origins that land here are
  *  served over plain HTTP and are not localhost — a shape used for local testing
  *  rather than for running Windmill. Every HTTPS deployment, and localhost, is a
- *  secure context and takes the real lock above. `isMirroring` and not
- *  `isWatching`: a tab that has moved on to `catchingUp` is still owed its
- *  re-read, and driving from here would send the pre-run history that re-read
- *  exists to replace. */
+ *  secure context and takes the real lock above. */
 async function bestEffort<T>(sessionId: string, body: () => Promise<T>): Promise<T | 'busy'> {
-	if (isMirroring(sessionId)) return 'busy'
 	return await drive(sessionId, body)
 }
 
@@ -195,8 +201,8 @@ async function drive<T>(sessionId: string, body: () => Promise<T>): Promise<T> {
 }
 
 /** Whether any tab currently holds the run lock for this session. Used to
- *  settle a driver that stopped sending frames: a released lock proves the tab
- *  is gone, where silence alone only suggests it. */
+ *  settle a driver that went silent: a released lock proves the tab is gone,
+ *  where silence alone only suggests it. */
 async function runLockHeld(sessionId: string): Promise<boolean> {
 	// Nothing to consult without the lock API, so a silent driver is reaped on
 	// silence alone — and on that path reaching `idle` does entitle this tab to
@@ -250,8 +256,8 @@ async function reapDeadDrivers(): Promise<void> {
 		.filter(([, p]) => p.state === 'watching' && now - p.lastHeardAt > DRIVER_SILENCE_MS)
 		.map(([id]) => id)
 	for (const id of stale) {
-		// Re-checked after the await: a frame may have landed while the query was
-		// in flight, and reaping then would tear down a run that is visibly alive.
+		// Re-checked after the await: a status message may have landed while the
+		// query was in flight, and reaping then would tear down a live run.
 		if (await runLockHeld(id)) continue
 		if (!isWatching(id)) continue
 		noteRemoteTurnEnded(id)
diff --git a/frontend/src/lib/components/sessions/sessionRunOwner.test.ts b/frontend/src/lib/components/sessions/sessionRunOwner.test.ts
index 9e23f85733..6251882ff7 100644
--- a/frontend/src/lib/components/sessions/sessionRunOwner.test.ts
+++ b/frontend/src/lib/components/sessions/sessionRunOwner.test.ts
@@ -7,7 +7,7 @@ import {
 	withSessionRunLock
 } from './sessionRunOwner.svelte'
 
-const SESSIONS = ['session-watching', 'session-catching-up', 'session-idle']
+const SESSIONS = ['session-watching', 'session-catching-up', 'session-idle', 'session-unheard']
 
 // `positions` is module state and a watching entry keeps the reaper interval
 // armed, so leave nothing behind for the next suite.
@@ -44,6 +44,18 @@ describe('withSessionRunLock with no lock to take', () => {
 		expect(await withSessionRunLock('session-idle', body)).toBe('ran')
 		expect(body).toHaveBeenCalledTimes(1)
 	})
+
+	// A tab that opened between the driver's last status message and its turn-end
+	// never saw the run start, so it is idle — holding the conversation from
+	// before the turn, and idle is the position that permits driving on it.
+	it('refuses a turn-end it never saw start, rather than driving on stale history', async () => {
+		onDriverLost(() => {})
+		noteRemoteTurnEnded('session-unheard')
+		const body = vi.fn(async () => 'ran')
+
+		expect(await withSessionRunLock('session-unheard', body)).toBe('busy')
+		expect(body).not.toHaveBeenCalled()
+	})
 })
 
 // A fresh module instance is the only honest way to test this: the runtime's
@@ -57,7 +69,7 @@ describe('a turn ending in a tab with no session runtime', () => {
 		owner.noteRemoteTurnEnded('session-no-runtime')
 
 		expect(owner.isCatchingUp('session-no-runtime')).toBe(false)
-		expect(owner.isMirroring('session-no-runtime')).toBe(false)
+		expect(owner.runHeldElsewhere('session-no-runtime')).toBe(false)
 		owner.clearRunPosition('session-no-runtime')
 	})
 })
diff --git a/frontend/src/lib/components/sessions/sessionRuntime.svelte.ts b/frontend/src/lib/components/sessions/sessionRuntime.svelte.ts
index ffc492a5af..84b4365b7b 100644
--- a/frontend/src/lib/components/sessions/sessionRuntime.svelte.ts
+++ b/frontend/src/lib/components/sessions/sessionRuntime.svelte.ts
@@ -93,18 +93,15 @@ import { getNonStreamingMetadataCompletion } from '$lib/components/copilot/lib'
 import { sendUserToast } from '$lib/toast'
 import { pendingUserAction, type DisplayMessage } from '$lib/components/copilot/chat/shared'
 import type { ChatCompletionMessageParam } from 'openai/resources/index.mjs'
-import { canSpliceFrame, mirrorFrameStart, withoutHeavyPayloads } from './sessionMirrorPayload'
 import {
-	broadcastMirror,
+	broadcastRunStatus,
 	broadcastTurnEnd,
-	MIRROR_THROTTLE_MS,
 	registerSyncHandlers,
 	requestCancel,
-	requestResync,
+	RUN_STATUS_INTERVAL_MS,
 	sendQuestionAnswer,
 	sendToolConfirmation,
-	type MirrorMsg,
-	type MirrorSnapshot
+	type RunStatusMsg
 } from './sessionSync.svelte'
 import {
 	clearRunPosition,
@@ -938,17 +935,15 @@ async function initRuntime(runtime: SessionRuntime, session: Session) {
 	manager.onChatRotated = (chatId) => setSessionChatId(session.id, chatId)
 
 	// One tab drives a session at a time. The lock brackets the turn so the other
-	// tabs can mirror it, and refuses a second driver rather than letting two
-	// turns interleave into one chat id.
+	// tabs can show that it is running, and refuses a second driver rather than
+	// letting two turns interleave into one chat id.
 	manager.runGuard = async (body) => {
 		const outcome = await withSessionRunLock(session.id, async () => {
-			// The first frame doubles as the "a run started here" signal: it is
-			// posted immediately and carries the chat id the watchers need.
-			startMirroring(session.id)
+			startRunStatus(session.id)
 			try {
 				return await body()
 			} finally {
-				stopMirroring(session.id)
+				stopRunStatus(session.id)
 				// The chat id is re-read here, not reused from above: the turn may
 				// have rotated it, and the listeners key their IndexedDB re-read on it.
 				broadcastTurnEnd(session.id, manager.historyManager.getCurrentChatId())
@@ -1007,154 +1002,67 @@ async function initRuntime(runtime: SessionRuntime, session: Session) {
 }
 
 // ---------------------------------------------------------------------------
-// Cross-tab mirroring
+// Cross-tab run status
 // ---------------------------------------------------------------------------
 
-// Driving tabs post a frame on this cadence for as long as their turn runs.
-// A plain interval rather than an $effect on the manager's fields: it also
-// serves as the liveness heartbeat (a run that stalls in a long tool call
-// still ticks), and it picks up in-place edits to already-rendered tool cards
-// that reactive tracking of the array root would miss.
-const mirrorTimers = new Map<string, ReturnType<typeof setInterval>>()
+// Driving tabs post their status on this cadence for as long as their turn
+// runs. A plain interval rather than an $effect on the manager's fields: it
+// doubles as the liveness heartbeat, so a run stalled in a long tool call still
+// ticks and is not reaped as a closed tab.
+const statusTimers = new Map<string, ReturnType<typeof setInterval>>()
 
-// Where the running turn's own messages start, per session. A frame never
-// reaches below it, which is what makes overwriting safe: everything at or after
-// it either is new this turn or is a stripped copy the watcher got from an
-// earlier frame of this same turn, so a frame can never replace a message the
-// watcher holds complete from the store. Reset to 0 by a rewrite (compaction),
-// after which no index is stable and whole transcripts travel instead.
-const turnStarts = new Map<string, number>()
-// The driver's transcript length at its last frame, so a shrink is detectable.
-const lastSentTotals = new Map<string, number>()
-
-function mirrorSnapshotOf(sessionId: string, full: boolean): MirrorSnapshot | undefined {
+function postRunStatus(sessionId: string): void {
 	const runtime = runtimes.get(sessionId)
-	if (!runtime) return undefined
+	if (!runtime) return
 	const m = runtime.manager
-	const total = m.displayMessages.length
-	// A shrink means the transcript was rewritten, not appended to: compaction
-	// folds the head into a summary mid-turn. A watcher's prefix is then a
-	// conversation that no longer exists, and no index check can tell — the
-	// lengths still line up. Send the whole thing instead.
-	const rewritten = total < (lastSentTotals.get(sessionId) ?? 0)
-	lastSentTotals.set(sessionId, total)
-	if (rewritten) turnStarts.set(sessionId, 0)
-	// Slice before cloning: `$state.snapshot` walks whatever it is handed, so
-	// snapshotting the whole transcript to send ten messages would traverse the
-	// entire conversation on every tick.
-	const baseIndex = mirrorFrameStart({
-		total,
-		turnStart: turnStarts.get(sessionId) ?? 0,
-		full: full || rewritten
-	})
-	return {
+	broadcastRunStatus({
 		sessionId,
 		chatId: m.historyManager.getCurrentChatId(),
-		baseIndex,
-		// Stripped before the clone, not after: `$state.snapshot` deep-copies
-		// whatever it is handed, so stripping second still copies every image data
-		// URL and pasted file body first — megabytes per tick, several times a
-		// second, for bytes about to be thrown away.
-		tail: $state.snapshot(
-			withoutHeavyPayloads(m.displayMessages.slice(baseIndex))
-		) as DisplayMessage[],
-		total,
 		loading: m.loading,
-		currentReply: m.currentReply,
-		currentReasoning: m.currentReasoning,
-		currentReasoningActive: m.currentReasoningActive,
-		loadingLabel: m.loadingLabel,
 		compacting: m.compacting,
+		blockedOnUser: pendingUserAction(m.displayMessages) !== undefined,
+		loadingLabel: m.loadingLabel,
 		planModeActive: m.planModeActive
-	}
+	})
 }
 
-function postMirror(sessionId: string, { full = false } = {}): void {
-	const snap = mirrorSnapshotOf(sessionId, full)
-	if (snap) broadcastMirror(snap)
-}
-
-function startMirroring(sessionId: string): void {
-	stopMirroring(sessionId)
-	// Captured before the turn pushes its first message, so it is the index of
-	// the oldest message this turn owns.
-	turnStarts.set(sessionId, runtimes.get(sessionId)?.manager.displayMessages.length ?? 0)
-	postMirror(sessionId)
-	mirrorTimers.set(
+function startRunStatus(sessionId: string): void {
+	stopRunStatus(sessionId)
+	// Posted immediately: this first message doubles as the "a run started here"
+	// signal, and carries the chat id the watching tabs key their re-read on.
+	postRunStatus(sessionId)
+	statusTimers.set(
 		sessionId,
-		setInterval(() => postMirror(sessionId), MIRROR_THROTTLE_MS)
+		setInterval(() => postRunStatus(sessionId), RUN_STATUS_INTERVAL_MS)
 	)
 }
 
-function stopMirroring(sessionId: string): void {
-	const timer = mirrorTimers.get(sessionId)
+function stopRunStatus(sessionId: string): void {
+	const timer = statusTimers.get(sessionId)
 	if (!timer) return
 	clearInterval(timer)
-	mirrorTimers.delete(sessionId)
-	// One last frame: the closing tokens of a turn usually land between ticks,
-	// and this is what the passive tabs render until their re-read completes.
-	// Posted while the turn's bookkeeping still stands — it is a frame like any
-	// other, and one sent without `turnStart` reaches below the turn and replaces
-	// complete messages with payload-stripped copies, while one sent without
-	// `lastSentTotals` cannot notice a compaction landing on this very tick.
-	postMirror(sessionId)
-	lastSentTotals.delete(sessionId)
-	turnStarts.delete(sessionId)
+	statusTimers.delete(sessionId)
 }
 
-/** Adopt a frame from the tab driving this session. */
-function applyMirror(msg: MirrorMsg): void {
-	// A tab mid-turn is the authority on its own session; a frame can only be
-	// an echo of a run this tab is itself driving.
+/** Adopt the driving tab's status. The transcript is left alone: this tab keeps
+ *  showing its own until `turn-end` triggers the re-read. */
+function applyRunStatus(msg: RunStatusMsg): void {
+	// A tab mid-turn is the authority on its own session; a status message can
+	// only be an echo of a run this tab is itself driving.
 	if (isDriving(msg.sessionId)) return
 	const runtime = runtimes.get(msg.sessionId)
 	if (!runtime) return
 	const m = runtime.manager
-	const onSameChat = m.historyManager.getCurrentChatId() === msg.chatId
-	if (
-		!canSpliceFrame({
-			baseIndex: msg.baseIndex,
-			total: msg.total,
-			localLength: m.displayMessages.length,
-			onSameChat
-		})
-	) {
-		// Nothing here can host this tail: this tab joined mid-run, or the driver
-		// rotated to a chat it isn't on. Ask for the whole transcript instead of
-		// splicing onto a prefix that isn't the same conversation.
-		requestResync(msg.sessionId)
-		return
-	}
-	if (!onSameChat) {
-		// A full snapshot names the conversation it came from, so follow the
-		// driver onto it rather than rendering it under the wrong chat id.
-		m.historyManager.setCurrentChatId(msg.chatId)
-		setSessionChatId(msg.sessionId, msg.chatId)
-	}
-	// Safe to overwrite wholesale: a frame reaches no further back than the
-	// running turn's first message, so everything it replaces is either new or a
-	// stripped copy of its own from an earlier frame — never a message this tab
-	// holds complete from the store.
-	m.displayMessages =
-		msg.baseIndex === 0 ? msg.tail : [...m.displayMessages.slice(0, msg.baseIndex), ...msg.tail]
 	m.loading = msg.loading
-	m.currentReply = msg.currentReply
-	m.currentReasoning = msg.currentReasoning
-	m.currentReasoningActive = msg.currentReasoningActive
-	m.loadingLabel = msg.loadingLabel
 	m.compacting = msg.compacting
+	m.loadingLabel = msg.blockedOnUser ? 'Waiting for your answer in the other tab' : msg.loadingLabel
 }
 
-/** The driver answers a resync with its whole transcript. */
-function answerResync(sessionId: string): void {
-	if (!isDriving(sessionId)) return
-	postMirror(sessionId, { full: true })
-}
-
-/** A run finished in another tab. The mirror carried the rendered transcript
- *  only, so re-read the record the driver just saved for everything else — the
- *  API-format history, context usage, the edits mask, background jobs — leaving
- *  this tab able to take the conversation over. */
+/** A run finished in another tab. Nothing of it crossed the channel but its
+ *  status, so re-read the record the driver just saved for everything else —
+ *  the rendered transcript, the API-format history, context usage, the edits
+ *  mask, background jobs — leaving this tab able to take the conversation
+ *  over. */
 async function applyTurnEnd(sessionId: string, chatId: string, attempt = 0): Promise<void> {
 	if (isDriving(sessionId)) return
 	let caughtUp = false
@@ -1166,11 +1074,9 @@ async function applyTurnEnd(sessionId: string, chatId: string, attempt = 0): Pro
 		}
 		const m = runtime.manager
 		// `loading` has to go first: loadPastChat refuses to run while the manager
-		// looks busy, and this one is the mirrored driver's, not a turn of our own.
+		// looks busy, and this one is the driver's, not a turn of our own. These
+		// three are exactly what applyRunStatus sets.
 		m.loading = false
-		m.currentReply = ''
-		m.currentReasoning = ''
-		m.currentReasoningActive = false
 		m.loadingLabel = undefined
 		m.compacting = false
 		const id = chatId || m.historyManager.getCurrentChatId()
@@ -1181,10 +1087,10 @@ async function applyTurnEnd(sessionId: string, chatId: string, attempt = 0): Pro
 		const found = await m.historyManager.reloadChat(id)
 		if (found === 'unavailable') {
 			// The store is unreadable *right now*, which says nothing about the
-			// conversation — so this tab still pairs the mirrored transcript with
-			// the history it held before the run. Freeing it here would let a turn
-			// go out under that pair and save it over what the driver wrote. Stay
-			// gated and keep asking; the read is the only thing that can settle it.
+			// conversation — so this tab still holds the one from before the run.
+			// Freeing it here would let a turn go out under that and save it over
+			// what the driver wrote. Stay gated and keep asking; the read is the only
+			// thing that can settle it.
 			scheduleCatchUpRetry(sessionId, chatId, attempt)
 			return
 		}
@@ -1218,9 +1124,9 @@ async function applyTurnEnd(sessionId: string, chatId: string, attempt = 0): Pro
  *
  *  It retries for as long as the runtime lives, including against a store that
  *  will never open. Deliberate: the alternative is giving up and releasing the
- *  gate, and this tab would then send a mirrored transcript paired with pre-run
- *  history — the driver's completed turn missing from what reaches the model,
- *  and its record overwritten. A read every few seconds is the cheaper half of
+ *  gate, and this tab would then send the pre-run conversation — the driver's
+ *  completed turn missing from what reaches the model, and its record
+ *  overwritten. A read every few seconds is the cheaper half of
  *  that trade, and a browser whose IndexedDB never opens has no session history,
  *  artifacts or records either, so a locked composer is not what is broken. */
 const CATCH_UP_BACKOFF_MS = [300, 700, 1500, 3000, 5000]
@@ -1268,8 +1174,7 @@ function applyQuestionAnswer(sessionId: string, toolId: string, choices: string[
 }
 
 registerSyncHandlers({
-	onMirror: applyMirror,
-	onResyncRequest: answerResync,
+	onRunStatus: applyRunStatus,
 	onCancelRequest: applyCancelRequest,
 	onToolConfirmation: applyToolConfirmation,
 	onQuestionAnswer: applyQuestionAnswer,
@@ -1279,7 +1184,7 @@ registerSyncHandlers({
 	onSessionDelete: (id) => disposeRuntime(id),
 	// Artifacts persist to a store both tabs share, but each keeps its own
 	// reactive copy, so the tab that did not write one never hears of it. That is
-	// what leaves a mirrored plan awaiting approval with no document behind it.
+	// what leaves a plan awaiting approval with no document behind it.
 	onSessionArtifact: (sessionId, artifactId) =>
 		void runtimes.get(sessionId)?.manager.artifacts.applyRemoteArtifact(artifactId)
 })
@@ -1306,10 +1211,10 @@ export function disposeRuntime(sessionId: string) {
 	runtime.manager.historyManager.close()
 	runtimes.delete(sessionId)
 	// The per-run bookkeeping outlives the manager it describes otherwise: a
-	// frame timer with no runtime to read keeps ticking, a pending catch-up would
-	// re-read for a manager that is gone, and a leftover position would answer
-	// for a session this tab no longer holds.
-	stopMirroring(sessionId)
+	// status timer with no runtime to read keeps ticking, a pending catch-up
+	// would re-read for a manager that is gone, and a leftover position would
+	// answer for a session this tab no longer holds.
+	stopRunStatus(sessionId)
 	cancelCatchUpRetry(sessionId)
 	clearRunPosition(sessionId)
 }
diff --git a/frontend/src/lib/components/sessions/sessionSync.svelte.ts b/frontend/src/lib/components/sessions/sessionSync.svelte.ts
index 057a97b0f6..fe95cf0836 100644
--- a/frontend/src/lib/components/sessions/sessionSync.svelte.ts
+++ b/frontend/src/lib/components/sessions/sessionSync.svelte.ts
@@ -1,23 +1,27 @@
 import { BROWSER } from 'esm-env'
 import { onUserChange, scopedKey } from '$lib/userScopedStorage'
-import type { DisplayMessage } from '$lib/components/copilot/chat/shared'
 import { noteDriverAlive, noteRemoteTurnEnded } from './sessionRunOwner.svelte'
 
 // The channel between tabs open on the same AI session. Everything a session is
 // made of — the record list, the chat transcript, the run itself — lives in the
 // tab, so two tabs on one session are two independent copies of it. This module
-// carries messages between those copies: record writes, and the driving tab's
-// live transcript. Who is entitled to drive is sessionRunOwner's question; this
-// module only tells it what arrived.
+// carries messages between those copies: record writes, and how the driving
+// tab's run is going. Who is entitled to drive is sessionRunOwner's question;
+// this module only tells it what arrived.
+//
+// No message carries a transcript. Tabs converge by re-reading the shared
+// IndexedDB record once a run ends, so what crosses the channel stays small and
+// bounded whatever a turn produced.
 //
 // The channel is per-user (same email scoping as the IndexedDB stores), so a
 // browser shared by two accounts never crosses them.
 
 const CHANNEL_BASE = 'windmill-sessions-sync'
 
-/** Mirror ticks are throttled to this while a turn streams. Also the heartbeat
- *  interval: an unchanged run still ticks, so silence means the driver is gone. */
-export const MIRROR_THROTTLE_MS = 250
+/** How often a driving tab posts its status. An unchanged run still ticks, so
+ *  silence is what tells the other tabs the driver is gone; the reaper's
+ *  threshold is several times this. */
+export const RUN_STATUS_INTERVAL_MS = 1000
 
 /** Carries only the id on purpose. Broadcast delivery order and IndexedDB
  *  commit order are independent, so shipping a copy of the record lets an older
@@ -36,34 +40,33 @@ type SessionArtifactMsg = { kind: 'session-artifact'; sessionId: string; artifac
  *  watcher does the same thing either way — re-read the record and stop
  *  showing the run. */
 type TurnEndMsg = { kind: 'turn-end'; sessionId: string; chatId: string }
-type MirrorMsg = {
-	kind: 'mirror'
+/**
+ * How the driving tab's run is going. Deliberately carries no transcript: a
+ * watching tab keeps showing its own, and drives the same loading indicator a
+ * local turn shows from these fields, then re-reads the record once `turn-end`
+ * says the run is over.
+ *
+ * Every field here is a scalar of fixed size. That is the property to preserve:
+ * the transcript this replaced carried tool results and job logs, which are
+ * bounded by nothing, and re-broadcasting them several times a second was a
+ * responsiveness bug that recurred once per field anyone forgot to strip.
+ */
+type RunStatusMsg = {
+	kind: 'run-status'
 	sessionId: string
 	chatId: string
-	/** Index the tail starts at; the receiver keeps its own prefix below it, and
-	 *  asks for a full snapshot when it has no prefix that reaches this far. */
-	baseIndex: number
-	tail: DisplayMessage[]
-	/** The driver's whole transcript length. A watcher holding more than this has
-	 *  a prefix the driver no longer has, so its splice would invent a transcript
-	 *  that never existed; it resyncs instead. */
-	total: number
 	loading: boolean
-	currentReply: string
-	currentReasoning: string
-	currentReasoningActive: boolean
-	loadingLabel: string | undefined
 	compacting: boolean
+	/** Parked on a question that only the driving tab can render. The watcher
+	 *  says where to answer it rather than implying work is in progress. */
+	blockedOnUser: boolean
+	loadingLabel: string | undefined
 	/** The driver's plan-mode posture. The only autonomy state worth carrying:
 	 *  every other one is a stored preference each tab keeps its own copy of,
 	 *  while plan mode is never persisted and so exists nowhere but the driving
 	 *  tab's memory. */
 	planModeActive: boolean
 }
-/** Sent by a tab whose local prefix can't host the tail it just received (it
- *  joined mid-run, or is on a different chat). The driver answers with a full
- *  snapshot. */
-type ResyncRequestMsg = { kind: 'resync-request'; sessionId: string }
 /** A Stop pressed in a tab that is only watching the run. */
 type CancelRequestMsg = { kind: 'cancel-request'; sessionId: string }
 /** A run blocked on the user is unblocked from whichever tab the user is in;
@@ -86,8 +89,7 @@ type SyncMsg =
 	| SessionDeleteMsg
 	| SessionArtifactMsg
 	| TurnEndMsg
-	| MirrorMsg
-	| ResyncRequestMsg
+	| RunStatusMsg
 	| CancelRequestMsg
 	| ToolConfirmationMsg
 	| QuestionAnswerMsg
@@ -97,8 +99,7 @@ type Handlers = {
 	onSessionDelete: (id: string) => void
 	onSessionArtifact: (sessionId: string, artifactId: string) => void
 	onTurnEnd: (sessionId: string, chatId: string) => void
-	onMirror: (msg: MirrorMsg) => void
-	onResyncRequest: (sessionId: string) => void
+	onRunStatus: (msg: RunStatusMsg) => void
 	onCancelRequest: (sessionId: string) => void
 	onToolConfirmation: (sessionId: string, toolId: string, confirmed: boolean) => void
 	onQuestionAnswer: (sessionId: string, toolId: string, choices: string[]) => void
@@ -172,12 +173,9 @@ function receive(msg: SyncMsg): void {
 			noteRemoteTurnEnded(msg.sessionId)
 			emit('onTurnEnd', msg.sessionId, msg.chatId)
 			break
-		case 'mirror':
+		case 'run-status':
 			noteDriverAlive(msg.sessionId, msg.planModeActive)
-			emit('onMirror', msg)
-			break
-		case 'resync-request':
-			emit('onResyncRequest', msg.sessionId)
+			emit('onRunStatus', msg)
 			break
 		case 'cancel-request':
 			emit('onCancelRequest', msg.sessionId)
@@ -218,17 +216,13 @@ export function broadcastSessionArtifact(sessionId: string, artifactId: string):
 }
 
 // ---------------------------------------------------------------------------
-// Live mirroring
+// Run status
 // ---------------------------------------------------------------------------
 
 export function broadcastTurnEnd(sessionId: string, chatId: string): void {
 	post({ kind: 'turn-end', sessionId, chatId })
 }
 
-export function requestResync(sessionId: string): void {
-	post({ kind: 'resync-request', sessionId })
-}
-
 export function requestCancel(sessionId: string): void {
 	post({ kind: 'cancel-request', sessionId })
 }
@@ -241,11 +235,11 @@ export function sendQuestionAnswer(sessionId: string, toolId: string, choices: s
 	post({ kind: 'question-answer', sessionId, toolId, choices })
 }
 
-export type MirrorSnapshot = Omit<MirrorMsg, 'kind'>
+export type RunStatus = Omit<RunStatusMsg, 'kind'>
 
-/** Send the driver's current view of a run. */
-export function broadcastMirror(snap: MirrorSnapshot): void {
-	post({ kind: 'mirror', ...snap })
+/** Say how the run this tab is driving is going. */
+export function broadcastRunStatus(status: RunStatus): void {
+	post({ kind: 'run-status', ...status })
 }
 
-export type { MirrorMsg }
+export type { RunStatusMsg }