bunnyquery 1.5.7 → 1.6.2

package/src/engine/session.ts

@@ -21,21 +21,26 @@ import {
 	callClaudeWithPublicMcp,
 	callOpenAIWithPublicMcp,
 	notifyAgentSaveAttachment,
+	notifyAgentContinueIndexing,
+	INDEXING_COMPLETE_MARKER,
+	MAX_INDEXING_RESUME_PASSES,
 	extractClaudeText,
 	extractOpenAIText,
 	getChatHistory,
 	POLL_INTERVAL,
 	BG_INDEXING_QUEUE_SUFFIX,
+	isBgIndexingQueue,
 	ANTHROPIC_MESSAGES_API_URL,
 	OPENAI_RESPONSES_API_URL,
 	type BgTaskEntry,
 } from './requests';
+import { isPagedReadFile, isImageVisionFile } from './office';
 import { isErrorResponseBody, isAuthExpiredError, isNonRetryableRequestError, getErrorMessage } from './errors';
 import { buildBoundedChatMessages } from './budget';
 import { createInlineLinkRegex } from './links';
 import { mapHistoryListToMessages, extractLastUserTextFromRequest } from './history';
 import { parseAttachmentContent } from './attachment_parsers';
-import type { ChatHost, ChatState, ChatMessage } from './host';
+import type { ChatHost, ChatState, ChatMessage, PinnedDispatchContext } from './host';
 
 function sleep(ms: number): Promise<void> {
 	return new Promise(function (r) { setTimeout(r, ms); });
@@ -53,6 +58,22 @@ function nextFrame(cb: (t: number) => void): void {
 	setTimeout(function () { cb(nowMs()); }, 16);
 }
 
+/** A live poll registered in ChatSession.historyItemPolls. */
+export type PollHandle = {
+	/** 'bg' = background indexing, pausable. 'fg' = a reply the user is waiting on. */
+	kind: 'fg' | 'bg';
+	/** Absent on an older skapi-js that cannot stop an attached poll. */
+	stop?: () => void;
+};
+
+/**
+ * True when a poll result came from stopPolling rather than the server. Duck-typed on
+ * purpose — see the note above about not importing skapi-js here.
+ */
+function isPollStopped(res: any): boolean {
+	return !!res && typeof res === 'object' && res.status === 'stopped';
+}
+
 export class ChatSession {
 	host: ChatHost;
 	state: ChatState;
@@ -60,7 +81,11 @@ export class ChatSession {
 	cancelledServerIds: Set<string>;
 	pendingAgentRequests: Record<string, Promise<any>>;
 	aiChatHistoryCache: Record<string, { messages: ChatMessage[]; endOfList: boolean; startKeyHistory: string[] }>;
-	historyItemPolls: Map<string, boolean>;
+	historyItemPolls: Map<string, PollHandle>;
+	/** Non-empty while polling is paused; keyed by reason so overlapping causes
+	 * (view detached AND tab hidden) do not resume each other prematurely. */
+	private _pauseReasons: Set<string>;
+	private _resuming: boolean;
 	private _lidSeq: number;
 
 	constructor(host: ChatHost) {
@@ -84,9 +109,91 @@ export class ChatSession {
 		this.pendingAgentRequests = {};
 		this.aiChatHistoryCache = {};
 		this.historyItemPolls = new Map();
+		this._pauseReasons = new Set();
+		this._resuming = false;
 		this._lidSeq = 0;
 	}
 
+	/**
+	 * Register a live poll so (a) a remount dedupes against it instead of stacking a
+	 * SECOND poll on the same item, and (b) pausePolling can stop it.
+	 *
+	 * `stop` comes from the SDK and may be absent on an older skapi-js, in which case the
+	 * poll simply cannot be stopped and is left running — see pausePolling.
+	 */
+	private _trackPoll(id: string, kind: 'fg' | 'bg', p: any): any {
+		var stop = p && typeof p.stop === 'function' ? p.stop.bind(p) : undefined;
+		if (!stop) {
+			// The SDK could not give us a handle. Almost always means an older skapi-js
+			// is loaded (or a stale bundler dep cache) — the poll will be unstoppable.
+			console.debug('[chat-engine] poll has no stop handle', { id: id, kind: kind });
+		}
+		this.historyItemPolls.set(id, { kind: kind, stop: stop });
+		return p;
+	}
+
+	/** True while any pause reason is active. */
+	isPollingPaused(): boolean {
+		return this._pauseReasons.size > 0;
+	}
+
+	/**
+	 * Stop BACKGROUND polling until resumePolling. Foreground polls (a reply the user is
+	 * waiting on) keep running deliberately: their results must still land in the history
+	 * cache so resumePendingRequest can render them on return, otherwise a user who sends
+	 * a message then navigates away comes back to a permanently stuck "Thinking...".
+	 *
+	 * Server-side work is untouched; this only stops asking about it. That is safe for
+	 * document indexing because the worker drives that loop itself.
+	 */
+	pausePolling(reason: string): void {
+		this._pauseReasons.add(reason || 'paused');
+		var self = this;
+		var stopped: string[] = [];
+		this.historyItemPolls.forEach(function (handle, id) {
+			if (!handle || handle.kind !== 'bg') return;
+			// No stop available (older SDK): LEAVE the entry in place. Deleting it would
+			// let a later drain attach a second, uncancellable poll on the same item.
+			if (typeof handle.stop !== 'function') return;
+			try { handle.stop(); } catch (e) { /* best effort */ }
+			stopped.push(id);
+		});
+		// Delete only what we actually stopped, one at a time. NEVER clear() the map:
+		// wholesale clearing is what previously let loadHistory attach a duplicate poll
+		// on a live item, producing duplicate replies and stranded "Thinking" bubbles.
+		stopped.forEach(function (id) { self.historyItemPolls.delete(id); });
+
+	}
+
+	/**
+	 * Lift a pause reason WITHOUT running the reconcile. For a caller that is about to
+	 * reload history anyway (a view remounting), letting resumePolling also reconcile
+	 * would race that load and can double-attach.
+	 */
+	clearPauseReason(reason: string): void {
+		this._pauseReasons.delete(reason || 'paused');
+	}
+
+	/**
+	 * Clear a pause reason and, once none remain, re-attach polling and reconcile.
+	 * Deliberately does NOT touch gateRefreshToken: bumping it would silently discard
+	 * the results of anything still in flight across the pause.
+	 */
+	resumePolling(reason: string): Promise<void> {
+		this._pauseReasons.delete(reason || 'paused');
+		if (this._pauseReasons.size > 0 || this._resuming) return Promise.resolve();
+		if (!this.host.isViewMounted || !this.host.isViewMounted()) return Promise.resolve();
+		var self = this;
+		this._resuming = true;
+		return Promise.resolve()
+			.then(function () {
+				self.drainBgTaskQueue();
+				return self.loadHistory(false, self.state.gateRefreshToken);
+			})
+			.catch(function (e: any) { console.error('[chat-engine] resume polling failed', e); })
+			.then(function () { self._resuming = false; });
+	}
+
 	private _newLocalId(): string {
 		this._lidSeq += 1;
 		return 'lid_' + this._lidSeq;
@@ -101,18 +208,87 @@ export class ChatSession {
 	updateHistoryCache(): void {
 		var key = this.getHistoryCacheKey();
 		if (!key) return;
+		// Never persist another chat's in-flight bubbles under THIS key. The
+		// dashboard shares one ChatSession across projects, so a turn dispatched
+		// in project A can still be sitting in state.messages while the identity
+		// (and therefore `key`) already points at project B — without this filter
+		// A's user + "Thinking..." bubbles get written into B's cache entry and
+		// replay on every later visit to B. Bubbles with no _ownerKey (server
+		// history, bg tasks) are always kept. Single pass: this runs on the
+		// typewriter hot path.
 		this.aiChatHistoryCache[key] = {
-			messages: this.state.messages.slice(),
+			messages: this.state.messages.filter(function (m) {
+				return m._ownerKey === undefined || m._ownerKey === key;
+			}),
 			endOfList: this.state.historyEndOfList,
 			startKeyHistory: this.state.historyStartKeyHistory.slice(),
 		};
 	}
 
-	private _callProviderFor(platform: string, prompt: string, messages: any, system: string, model: string | undefined, userId: string, extractContent: any, fileUrls?: any) {
-		var id = this.host.getIdentity();
+	/**
+	 * Land a resolved reply in the history cache of a chat that is NOT currently
+	 * visible, without touching state.messages. Mirrors the cache-only path in
+	 * dispatchAgentRequest: REPLACE the trailing pending "Thinking..." bubble
+	 * (append only when there is none), and settle the matching pending user
+	 * bubble, so the cached copy never keeps a stuck "Thinking..." that a later
+	 * cache-first load would re-render forever.
+	 */
+	private _applyReplyToCache(key: string, reply: ChatMessage, serverId?: string): void {
+		if (!key) return;
+		var existing = this.aiChatHistoryCache[key] || { messages: [], endOfList: false, startKeyHistory: [] };
+		var msgs = existing.messages.slice();
+
+		var thIdx = -1;
+		for (var i = msgs.length - 1; i >= 0; i--) {
+			var m = msgs[i];
+			if (!m || !m.isPending || m.role !== 'assistant' || m.isBackgroundTask) continue;
+			if (serverId && m._serverItemId && m._serverItemId !== serverId) continue;
+			thIdx = i; break;
+		}
+		if (thIdx !== -1) {
+			if (reply._serverItemId === undefined && msgs[thIdx]._serverItemId !== undefined) reply._serverItemId = msgs[thIdx]._serverItemId;
+			msgs[thIdx] = reply;
+		} else {
+			msgs.push(reply);
+		}
+
+		// Settle the user bubble this turn belongs to (first pending one, or the
+		// one carrying serverId) so it stops rendering as still in flight.
+		for (var j = 0; j < msgs.length; j++) {
+			var u = msgs[j];
+			if (!u || u.role !== 'user' || u.isBackgroundTask) continue;
+			if (!(u.isPendingQueued || u.isPendingInProcess || u.isSendingToServer)) continue;
+			if (serverId && u._serverItemId && u._serverItemId !== serverId) continue;
+			var settled: ChatMessage = { role: 'user', content: u.content };
+			if (u._serverItemId !== undefined) settled._serverItemId = u._serverItemId;
+			if (u._ownerKey !== undefined) settled._ownerKey = u._ownerKey;
+			msgs[j] = settled;
+			break;
+		}
+
+		this.aiChatHistoryCache[key] = {
+			messages: msgs,
+			endOfList: existing.endOfList,
+			startKeyHistory: existing.startKeyHistory,
+		};
+	}
+
+	/**
+	 * serviceId/owner are passed explicitly by every caller: a request can be
+	 * dispatched after the user moved to another project, and re-reading the live
+	 * identity here would silently send the turn to THAT project instead of the
+	 * one it was composed for. Falls back to the live read only when a caller
+	 * omits them.
+	 */
+	private _callProviderFor(platform: string, prompt: string, messages: any, system: string, model: string | undefined, userId: string, extractContent: any, fileUrls?: any, serviceId?: string, owner?: string) {
+		if (serviceId === undefined || owner === undefined) {
+			var id = this.host.getIdentity();
+			if (serviceId === undefined) serviceId = id.serviceId;
+			if (owner === undefined) owner = id.owner;
+		}
 		return platform === 'openai'
-			? callOpenAIWithPublicMcp(prompt, id.serviceId, id.owner, messages, system, model, userId, extractContent, fileUrls)
-			: callClaudeWithPublicMcp(prompt, id.serviceId, id.owner, messages, system, model, userId, extractContent, fileUrls);
+			? callOpenAIWithPublicMcp(prompt, serviceId, owner, messages, system, model, userId, extractContent, fileUrls)
+			: callClaudeWithPublicMcp(prompt, serviceId, owner, messages, system, model, userId, extractContent, fileUrls);
 	}
 
 	dispatchAgentRequest(params: any) {
@@ -126,15 +302,16 @@ export class ChatSession {
 		var dispatchItemId: string | undefined;
 		var sendAndPoll = function () {
 			return Promise.resolve(
-				self._callProviderFor(params.aiPlatform, params.text, params.boundedMessages, params.systemPrompt, params.aiModel, params.userId, params.extractContent, params.fileUrls)
+				self._callProviderFor(params.aiPlatform, params.text, params.boundedMessages, params.systemPrompt, params.aiModel, params.userId, params.extractContent, params.fileUrls, params.serviceId, params.owner)
 			).then(function (initial: any) {
 				if (initial && initial.poll && (initial.status === 'pending' || initial.status === 'running')) {
 					if (initial.id) {
 						if (dispatchItemId && dispatchItemId !== initial.id) self.historyItemPolls.delete(dispatchItemId);
 						dispatchItemId = initial.id;
-						self.historyItemPolls.set(initial.id, true);
 					}
-					return initial.poll({ latency: POLL_INTERVAL });
+					var dp = initial.poll({ latency: POLL_INTERVAL });
+					if (initial.id) self._trackPoll(initial.id, 'fg', dp);
+					return dp;
 				}
 				return initial;
 			});
@@ -208,24 +385,73 @@ export class ChatSession {
 	// composed = clean display text; composedForLlm carries office-extraction
 	// placeholders for the provider only. useBgQueue routes a post-attachment turn
 	// onto the "-bg" queue so it runs after indexing.
-	dispatchComposedMessage(composed: string, useBgQueue?: boolean, composedForLlm?: string, extractContent?: any, fileUrls?: any): void {
+	dispatchComposedMessage(composed: string, useBgQueue?: boolean, composedForLlm?: string, extractContent?: any, fileUrls?: any, pinned?: PinnedDispatchContext): void {
 		var self = this;
 		if (!composed) return;
-		var id = this.host.getIdentity();
+		// A send can be dispatched LONG after the user hit Send (attachment
+		// uploads are awaited first), by which time the live identity may have
+		// moved to another project. The caller pins the identity + system prompt
+		// it captured at Send time so the request still goes to the project the
+		// question was actually asked of. Falls back to the live read when the
+		// caller doesn't pin (the widget, which has only one project anyway).
+		var id = pinned ? pinned.identity : this.host.getIdentity();
 		if (id.platform === 'none') return;
 
 		var llmComposed = composedForLlm || composed;
 
-		var isQueuedSend = useBgQueue || this.state.sending || this.state.messages.some(function (m) {
+		// Cache key of the chat this turn belongs to. Every locally-created
+		// bubble is stamped with it so a project switch (which flips
+		// getIdentity()/getHistoryCacheKey() to the new project) can't
+		// misattribute this turn's bubbles to that project.
+		// (platform === 'none' already returned above, so serviceId is the only gate)
+		var key = !id.serviceId ? '' : id.serviceId + '#' + id.platform;
+		// True when the pinned chat is NOT the one currently on screen. Then
+		// state.messages belongs to a different project and MUST NOT be touched:
+		// the turn is staged in the pinned chat's cache instead and shows up when
+		// the user navigates back to it.
+		var offChat = !!key && key !== this.getHistoryCacheKey();
+
+		var isQueuedSend = !offChat && (useBgQueue || this.state.sending || this.state.messages.some(function (m) {
 			return (m.isPending || m.isPendingQueued) && !m.isBackgroundTask && !m._useBgQueue;
-		});
+		}));
 
 		var aiPlatform = id.platform;
 		var aiModel = id.model || undefined;
-		var systemPrompt = this.host.buildSystemPrompt();
+		var systemPrompt = pinned ? pinned.systemPrompt : this.host.buildSystemPrompt();
 		var userId = id.userId || id.serviceId;
 		var chatQueue = useBgQueue ? userId + BG_INDEXING_QUEUE_SUFFIX : userId;
 
+		if (offChat) {
+			// Stage the turn in the pinned chat's cache and dispatch. The
+			// client-side queue can't be consulted (its state is the other
+			// project's), but the SERVER serializes per queue name, so ordering
+			// within the pinned chat still holds. dispatchAgentRequest replaces
+			// the pending bubble in this same cache entry when the reply lands.
+			var offHistory = (this.aiChatHistoryCache[key] ? this.aiChatHistoryCache[key].messages : []).filter(function (m) {
+				return !m.isPending && !m.isPendingQueued && !m.isPendingInProcess && !m.isPendingOlder &&
+					!m.isCancelled && !m.isBackgroundTask;
+			});
+			var offBounded = buildBoundedChatMessages({
+				platform: aiPlatform, model: aiModel, systemPrompt: systemPrompt, serviceId: id.serviceId,
+				history: offHistory.concat([{ role: 'user', content: llmComposed }]),
+			});
+			var offExisting = this.aiChatHistoryCache[key] || { messages: [], endOfList: false, startKeyHistory: [] };
+			this.aiChatHistoryCache[key] = {
+				messages: offExisting.messages.concat([
+					{ role: 'user', content: composed, _ownerKey: key },
+					{ role: 'assistant', content: '', isPending: true, isPendingInProcess: true, _ownerKey: key },
+				]),
+				endOfList: offExisting.endOfList,
+				startKeyHistory: offExisting.startKeyHistory,
+			};
+			this.dispatchAgentRequest({
+				key: key, serviceId: id.serviceId, owner: id.owner, aiPlatform: aiPlatform, aiModel: aiModel,
+				systemPrompt: systemPrompt, text: composed, boundedMessages: offBounded.messages, userId: chatQueue,
+				extractContent: extractContent, fileUrls: fileUrls,
+			});
+			return;
+		}
+
 		if (isQueuedSend) {
 			var resolvedHistory = this.state.messages.filter(function (m) {
 				return !m.isPending && !m.isPendingQueued && !m.isPendingInProcess && !m.isPendingOlder &&
@@ -236,15 +462,20 @@ export class ChatSession {
 				history: resolvedHistory.concat([{ role: 'user', content: llmComposed }]),
 			});
 			var queuedBubble: ChatMessage = { role: 'user', content: composed, isPendingQueued: true, isSendingToServer: true };
+			if (key) queuedBubble._ownerKey = key;
 			if (useBgQueue) queuedBubble._useBgQueue = true;
 			this.state.messages.push(queuedBubble);
 			this.host.notify(); this.updateHistoryCache(); this.host.scrollToBottom(true);
 
-			var capturedComposed = composed, capturedPlatform = aiPlatform;
-			Promise.resolve(this._callProviderFor(aiPlatform, composed, boundedQ.messages, systemPrompt, aiModel, chatQueue, extractContent, fileUrls))
+			var capturedComposed = composed, capturedPlatform = aiPlatform, capturedKey = key;
+			Promise.resolve(this._callProviderFor(aiPlatform, composed, boundedQ.messages, systemPrompt, aiModel, chatQueue, extractContent, fileUrls, id.serviceId, id.owner))
 				.then(function (result: any) {
-					var sendingIdx = self.state.messages.findIndex(function (m) {
-						return m.isSendingToServer && (m.isPendingQueued || m.isPendingInProcess) && m.role === 'user';
+					// Only ack a bubble that belongs to THIS chat — the search is
+					// positional, so on another project it would stamp this turn's
+					// _serverItemId onto that project's unrelated in-flight bubble.
+					var sendingIdx = self.getHistoryCacheKey() !== capturedKey ? -1 : self.state.messages.findIndex(function (m) {
+						return m.isSendingToServer && (m.isPendingQueued || m.isPendingInProcess) && m.role === 'user' &&
+							(m._ownerKey === undefined || m._ownerKey === capturedKey);
 					});
 					var serverId = result && typeof result.id === 'string' ? result.id : undefined;
 					if (sendingIdx >= 0) {
@@ -255,14 +486,15 @@ export class ChatSession {
 					if (result && result.poll && (result.status === 'pending' || result.status === 'running')) {
 						// Track this queued item's poll so a remount/refetch dedups
 						// against it instead of attaching a duplicate history poll.
-						if (serverId) self.historyItemPolls.set(serverId, true);
-						return result.poll({ latency: POLL_INTERVAL })
-							.then(function (res: any) { return self.onQueuedSendResponse(capturedComposed, res, capturedPlatform, serverId); })
-							.catch(function (err: any) { return self.onQueuedSendError(capturedComposed, err, serverId); });
+						var qp = result.poll({ latency: POLL_INTERVAL });
+						if (serverId) self._trackPoll(serverId, 'fg', qp);
+						return qp
+							.then(function (res: any) { if (isPollStopped(res)) return; return self.onQueuedSendResponse(capturedComposed, res, capturedPlatform, serverId, capturedKey); })
+							.catch(function (err: any) { return self.onQueuedSendError(capturedComposed, err, serverId, capturedKey); });
 					}
-					return self.onQueuedSendResponse(capturedComposed, result, capturedPlatform, serverId);
+					return self.onQueuedSendResponse(capturedComposed, result, capturedPlatform, serverId, capturedKey);
 				})
-				.catch(function (err: any) { return self.onQueuedSendError(capturedComposed, err, undefined); });
+				.catch(function (err: any) { return self.onQueuedSendError(capturedComposed, err, undefined, capturedKey); });
 			return;
 		}
 
@@ -271,11 +503,10 @@ export class ChatSession {
 		// view unmount), then rendered from the cache via typewriteLatestReply. A
 		// later resumePendingRequest() re-renders it if the view remounted while the
 		// request was still in flight.
-		this.state.messages.push({ role: 'user', content: composed });
-		this.state.messages.push({ role: 'assistant', content: '', isPending: true, isPendingInProcess: true });
+		this.state.messages.push({ role: 'user', content: composed, ...(key ? { _ownerKey: key } : {}) });
+		this.state.messages.push({ role: 'assistant', content: '', isPending: true, isPendingInProcess: true, ...(key ? { _ownerKey: key } : {}) });
 		this.host.notify(); this.updateHistoryCache(); this.state.sending = true; this.host.scrollToBottom(true);
 
-		var key = this.getHistoryCacheKey();
 		var historyForLlm = this.state.messages.filter(function (m) { return !m.isCancelled && !m.isBackgroundTask; });
 		if (llmComposed !== composed) {
 			for (var li = historyForLlm.length - 1; li >= 0; li--) {
@@ -305,8 +536,16 @@ export class ChatSession {
 		// already replaced the pending bubble with the answer IN THE CACHE, so a
 		// later loadChatHistory renders it.
 		Promise.resolve(run).catch(function () { }).then(function () {
-			if (!(self.host.isViewMounted() && self.getHistoryCacheKey() === key)) return;
+			// Clear `sending` UNCONDITIONALLY. It is session-global (it gates
+			// isQueuedSend above, plus the platform/model pickers in the view), so
+			// leaving it set when the user navigated to another project wedged
+			// every subsequent send in EVERY project onto the queued path and kept
+			// the view's user-bubble rescue arm armed forever. Only the RENDER of
+			// the reply stays gated on still showing this chat — when it isn't,
+			// dispatchAgentRequest has already written the answer into the cache
+			// under `key`, so a later loadChatHistory renders it.
 			self.state.sending = false;
+			if (!(self.host.isViewMounted() && self.getHistoryCacheKey() === key)) return;
 			return Promise.resolve(self.typewriteLatestReply(key)).then(function () { self.host.scrollToBottom(true); });
 		});
 	}
@@ -320,9 +559,11 @@ export class ChatSession {
 		var existing = this.state.messages[nextIdx];
 		var promoted: ChatMessage = { role: 'user', content: existing.content, isPendingInProcess: true, isBackgroundTask: true };
 		if (existing._serverItemId !== undefined) promoted._serverItemId = existing._serverItemId;
+		if (existing._ownerKey !== undefined) promoted._ownerKey = existing._ownerKey;
 		this.state.messages[nextIdx] = promoted;
 		var placeholder: ChatMessage = { role: 'assistant', content: '', isPending: true, isPendingInProcess: true, isBackgroundTask: true };
 		if (existing._serverItemId !== undefined) placeholder._serverItemId = existing._serverItemId;
+		if (existing._ownerKey !== undefined) placeholder._ownerKey = existing._ownerKey;
 		this.state.messages.splice(nextIdx + 1, 0, placeholder);
 		this.host.notify();
 	}
@@ -337,6 +578,7 @@ export class ChatSession {
 		var promoted: ChatMessage = { role: 'user', content: existing.content, isPendingInProcess: true };
 		if (existing.isBackgroundTask) promoted.isBackgroundTask = true;
 		if (existing._serverItemId !== undefined) promoted._serverItemId = existing._serverItemId;
+		if (existing._ownerKey !== undefined) promoted._ownerKey = existing._ownerKey;
 		if (existing.isSendingToServer) promoted.isSendingToServer = true;
 		this.state.messages[nextIdx] = promoted;
 		// Carry the promoted turn's _serverItemId onto the "Thinking..." placeholder
@@ -349,11 +591,16 @@ export class ChatSession {
 		// unaffected.
 		var placeholder: ChatMessage = { role: 'assistant', content: '', isPending: true };
 		if (existing._serverItemId !== undefined) placeholder._serverItemId = existing._serverItemId;
+		if (existing._ownerKey !== undefined) placeholder._ownerKey = existing._ownerKey;
 		this.state.messages.splice(nextIdx + 1, 0, placeholder);
 		this.host.notify();
 	}
 
 	resolveQueuedUserBubble(serverId?: string): number | undefined {
+		// The two fallbacks below match by POSITION, not identity, so they must
+		// never consider a bubble stamped for a different chat.
+		var liveKey = this.getHistoryCacheKey();
+		var isLocal = function (m: ChatMessage) { return m._ownerKey === undefined || m._ownerKey === liveKey; };
 		var userIdx = -1;
 		if (serverId) {
 			userIdx = this.state.messages.findIndex(function (m) {
@@ -363,19 +610,19 @@ export class ChatSession {
 		}
 		if (userIdx === -1) {
 			userIdx = this.state.messages.findIndex(function (m) {
-				return m.isPendingInProcess && m.role === 'user' && !m.isBackgroundTask && !m._useBgQueue;
+				return m.isPendingInProcess && m.role === 'user' && !m.isBackgroundTask && !m._useBgQueue && isLocal(m);
 			});
 		}
 		if (userIdx === -1) {
 			userIdx = this.state.messages.findIndex(function (m) {
-				return m.isPendingQueued && m.role === 'user' && !m.isBackgroundTask && !m._useBgQueue;
+				return m.isPendingQueued && m.role === 'user' && !m.isBackgroundTask && !m._useBgQueue && isLocal(m);
 			});
 		}
 		if (serverId && this.cancelledServerIds.has(serverId)) {
 			this.cancelledServerIds.delete(serverId);
 			if (userIdx >= 0) {
 				var ex = this.state.messages[userIdx];
-				this.state.messages[userIdx] = { role: 'user', content: ex.content, isCancelled: true, _serverItemId: ex._serverItemId };
+				this.state.messages[userIdx] = { role: 'user', content: ex.content, isCancelled: true, _serverItemId: ex._serverItemId, ...(ex._ownerKey !== undefined ? { _ownerKey: ex._ownerKey } : {}) };
 				var thIdx = this.state.messages.findIndex(function (m, i) {
 					return i > userIdx && m.isPending && m.role === 'assistant' && !m.isBackgroundTask;
 				});
@@ -388,6 +635,7 @@ export class ChatSession {
 			var exist = this.state.messages[userIdx];
 			var repl: ChatMessage = { role: 'user', content: exist.content };
 			if (exist._serverItemId !== undefined) repl._serverItemId = exist._serverItemId;
+			if (exist._ownerKey !== undefined) repl._ownerKey = exist._ownerKey;
 			this.state.messages[userIdx] = repl;
 		}
 		var thinkingIdx = userIdx >= 0
@@ -402,8 +650,21 @@ export class ChatSession {
 		else this.state.messages.push(msg);
 	}
 
-	onQueuedSendResponse(_composed: string, response: any, platform: string, serverId?: string): void {
+	onQueuedSendResponse(_composed: string, response: any, platform: string, serverId?: string, ownerKey?: string): void {
 		if (serverId) this.historyItemPolls.delete(serverId);
+		// This turn resolved while a DIFFERENT chat is on screen (the user moved to
+		// another project mid-flight). state.messages now belongs to that chat, and
+		// resolveQueuedUserBubble's positional fallbacks would happily hijack ITS
+		// pending bubble — or, finding nothing, push this answer onto its list.
+		// Settle in the owning chat's cache instead and leave the view untouched.
+		if (ownerKey && this.getHistoryCacheKey() !== ownerKey) {
+			var offReply: ChatMessage = isErrorResponseBody(response)
+				? { role: 'assistant', content: getErrorMessage(response), isError: true }
+				: { role: 'assistant', content: ((platform === 'openai' ? extractOpenAIText(response) : extractClaudeText(response)) || '').trim() || 'No text response received from AI provider.' };
+			this._applyReplyToCache(ownerKey, offReply, serverId);
+			if (serverId) this.cancelledServerIds.delete(serverId);
+			return;
+		}
 		var targetIdx = this.resolveQueuedUserBubble(serverId);
 		if (targetIdx === undefined) { this.host.notify(); this.updateHistoryCache(); return; }
 		if (isErrorResponseBody(response)) {
@@ -431,9 +692,19 @@ export class ChatSession {
 		this.host.scrollToBottom(true);
 	}
 
-	onQueuedSendError(_composed: string, err: any, serverId?: string): void {
+	onQueuedSendError(_composed: string, err: any, serverId?: string, ownerKey?: string): void {
 		var self = this;
 		if (serverId) this.historyItemPolls.delete(serverId);
+		// Off-chat resolution — see onQueuedSendResponse. Settle in the owning
+		// chat's cache rather than mutating whatever chat is on screen now.
+		if (ownerKey && this.getHistoryCacheKey() !== ownerKey) {
+			var isGone = err && (err.code === 'NOT_EXISTS' || (err.body && err.body.code === 'NOT_EXISTS'));
+			this._applyReplyToCache(ownerKey, isGone
+				? { role: 'assistant', content: 'Request was cancelled.', isError: true }
+				: { role: 'assistant', content: getErrorMessage(err), isError: true }, serverId);
+			if (serverId) this.cancelledServerIds.delete(serverId);
+			return;
+		}
 		var isNotExists = err && (err.code === 'NOT_EXISTS' || (err.body && err.body.code === 'NOT_EXISTS'));
 		if (isNotExists) {
 			var userIdx = serverId
@@ -736,6 +1007,8 @@ export class ChatSession {
 		var isErr = isErrorResponseBody(response);
 		var answer = isErr ? getErrorMessage(response)
 			: ((platform === 'openai' ? extractOpenAIText(response) : extractClaudeText(response)) || '').trim();
+		// Hide the internal indexing-completion marker from the displayed summary.
+		if (!isErr && answer) answer = answer.split(INDEXING_COMPLETE_MARKER).join('').trim();
 		var idx = this.state.messages.findIndex(function (m) { return m.isPending && m._serverItemId === itemId; });
 		if (idx !== -1) {
 			// A bg "Indexing:" turn pushes a user bubble (isPendingInProcess) ALONGSIDE
@@ -793,7 +1066,11 @@ export class ChatSession {
 		}
 		this.bgTaskQueue.forEach(function (entry) {
 			if (entry.serviceId !== svcId || entry.platform !== plat) return;
-			if (presentIds[entry.id]) return;
+			// Bubble injection and poll attachment are INDEPENDENT. An entry whose bubble
+			// already exists may still need a poll — that is exactly the state a paused
+			// drain leaves behind, and returning early here stranded it as a permanent
+			// "Thinking..." once polling resumed.
+			if (!presentIds[entry.id]) {
 			var isRunning = entry.status === 'running';
 			var userBubble: ChatMessage = { role: 'user', content: self.host.formatIndexingLabel(entry.filename, entry.mime, entry.size, entry.storagePath, entry.isReindex), isBackgroundTask: true, _serverItemId: entry.id };
 			if (isRunning) userBubble.isPendingInProcess = true; else userBubble.isPendingQueued = true;
@@ -803,11 +1080,19 @@ export class ChatSession {
 			}
 			presentIds[entry.id] = true; // keep the index consistent with the pushed bubbles
 			self.host.notify(); self.updateHistoryCache(); self.host.scrollToBottom(false);
-			if (!self.historyItemPolls.has(entry.id) && typeof entry.poll === 'function') {
-				self.historyItemPolls.set(entry.id, true);
+			}
+			if (!self.isPollingPaused() && !self.historyItemPolls.has(entry.id) && typeof entry.poll === 'function') {
 				var capturedId = entry.id, capturedPlat = plat;
-				entry.poll({ latency: POLL_INTERVAL }).then(function (response: any) {
+				var capturedEntry = entry;
+				var wasStopped = false;
+				var bp = entry.poll({ latency: POLL_INTERVAL });
+				self._trackPoll(entry.id, 'bg', bp);
+				bp.then(function (response: any) {
+					// A stopped poll is not a result: leave the bubble and the queue entry
+					// exactly as they were so resumePolling can re-attach.
+					if (isPollStopped(response)) { wasStopped = true; return; }
 					self.handleHistoryItemResolution(capturedId, response, capturedPlat);
+					self.maybeResumeIndexing(capturedEntry, response, capturedPlat);
 				}).catch(function (err: any) {
 					self.historyItemPolls.delete(capturedId);
 					var isNotExists = err && (err.code === 'NOT_EXISTS' || (err.body && err.body.code === 'NOT_EXISTS'));
@@ -818,6 +1103,9 @@ export class ChatSession {
 						self.host.notify(); self.updateHistoryCache();
 					}
 				}).then(function () {
+					// Keep the queue entry when the poll was merely stopped, or resuming
+					// would have nothing left to re-attach to.
+					if (wasStopped) return;
 					var qi = self.bgTaskQueue.findIndex(function (q) { return q.id === capturedId; });
 					if (qi !== -1) self.bgTaskQueue.splice(qi, 1);
 				});
@@ -826,6 +1114,61 @@ export class ChatSession {
 		this.promoteNextBgQueuedToRunning();
 	}
 
+	// Resume-across-passes: if a background INDEXING task for a paged file (spreadsheet or
+	// text) finished WITHOUT the completion marker, the agent ran out of room before reading
+	// the whole file - dispatch a CONTINUE pass (up to a cap) that resumes from where the
+	// already-saved records leave off. Additive + guarded so it never loops forever and
+	// never breaks the resolution path.
+	//
+	// VISION files (PDFs rendered to page images) are NOT resumed here: the proxy worker
+	// advances their page window itself, off the renderer's true page count. Driving them
+	// from the browser is what used to lose pages on long documents - the chain lived in tab
+	// memory (a reload or a closed tab ended it), and it stopped whenever the model claimed
+	// completion, which on an 88-page file happened at page 15. Continuing to dispatch here
+	// as well would now double-index every window.
+	maybeResumeIndexing(entry: BgTaskEntry, response: any, platform: string): void {
+		var self = this;
+		try {
+			if (!entry || !entry.storagePath) return;
+			if (!isPagedReadFile(entry.filename, entry.mime)) return;
+			if (isImageVisionFile(entry.filename, entry.mime)) return; // worker owns this loop
+			if (isErrorResponseBody(response)) return; // a failed pass is not "incomplete"
+			var answer = (platform === 'openai' ? extractOpenAIText(response) : extractClaudeText(response)) || '';
+			if (answer.indexOf(INDEXING_COMPLETE_MARKER) !== -1) return; // fully indexed
+			var pass = (entry.resumePass || 0) + 1;
+			if (pass > MAX_INDEXING_RESUME_PASSES) return; // give up after the cap
+			var id = this.host.getIdentity();
+			if (!id || id.platform === 'none' || id.serviceId !== entry.serviceId) return;
+			notifyAgentContinueIndexing({
+				platform: id.platform as 'claude' | 'openai',
+				model: id.model,
+				service: id.serviceId,
+				owner: id.owner,
+				userId: id.userId || id.serviceId,
+				serviceName: id.serviceName,
+				serviceDescription: id.serviceDescription,
+				attachment: {
+					name: entry.filename,
+					storagePath: entry.storagePath,
+					mime: entry.mime,
+					size: entry.size,
+					url: '',
+				},
+			}).then(function (ack: any) {
+				if (ack && typeof ack.id === 'string') {
+					self.bgTaskQueue.push({
+						serviceId: id.serviceId, platform: id.platform as 'claude' | 'openai', id: ack.id,
+						filename: entry.filename, storagePath: entry.storagePath,
+						isReindex: entry.isReindex, mime: entry.mime, size: entry.size,
+						status: ack.status === 'running' ? 'running' : 'pending',
+						poll: ack.poll, resumePass: pass,
+					});
+					self.drainBgTaskQueue();
+				}
+			}, function (e: any) { console.error('[chat-engine] resume-indexing dispatch failed', e); });
+		} catch (e) { /* best-effort: resume must never break bg-task resolution */ }
+	}
+
 	// --- history fetch + pagination --------------------------------------
 	// Initial load (fetchMore=false) replaces the list (with in-flight rescue +
 	// cancelled-merge) and attaches polls to running/pending items; pagination
@@ -835,6 +1178,11 @@ export class ChatSession {
 	loadHistory(fetchMore?: boolean, token?: number): Promise<void> {
 		var self = this;
 		var id = this.host.getIdentity();
+		// Key this fetch is FOR, snapshotted from the same identity read the
+		// request is built from. The rescue below compares against this rather
+		// than a live getHistoryCacheKey(), so a project switch mid-fetch can't
+		// make another chat's in-flight bubbles look local.
+		var loadKey = (!id.serviceId || id.platform === 'none') ? '' : id.serviceId + '#' + id.platform;
 		if (token === undefined) token = this.state.gateRefreshToken;
 		if ((this.state.loadingHistory && this.state.historyRequestToken === token) || id.platform === 'none' || !id.serviceId) {
 			return Promise.resolve();
@@ -857,9 +1205,9 @@ export class ChatSession {
 			if (token !== self.state.gateRefreshToken) return;
 			var chatList = history && Array.isArray(history.list) ? history.list : [];
 			chatList.forEach(function (item: any) {
-				if (typeof item.queue_name === 'string' && item.queue_name.slice(-BG_INDEXING_QUEUE_SUFFIX.length) === BG_INDEXING_QUEUE_SUFFIX) {
+				if (isBgIndexingQueue(item.queue_name)) {
 					var userText = extractLastUserTextFromRequest(item.request_body);
-					if (typeof userText === 'string' && userText.indexOf('A new file has just been uploaded') === 0) item._isBgTask = true;
+					if (typeof userText === 'string' && (userText.indexOf('A new file has just been uploaded') === 0 || userText.indexOf('CONTINUE indexing') === 0)) item._isBgTask = true;
 					else item._isOnBgQueue = true;
 				}
 			});
@@ -895,6 +1243,10 @@ export class ChatSession {
 				for (var ri = 0; ri < self.state.messages.length; ri++) {
 					var mm = self.state.messages[ri];
 					if (mm.isBackgroundTask) continue;
+					// Belongs to a different chat (another project, or another
+					// platform on this one) — it must not be carried onto THIS
+					// chat's freshly-fetched history.
+					if (mm._ownerKey !== undefined && mm._ownerKey !== loadKey) continue;
 					if (mm._serverItemId && serverIds[mm._serverItemId]) continue;
 					if (!mm._serverItemId) {
 						if (mappedHasPendingAssistant) continue;
@@ -957,11 +1309,13 @@ export class ChatSession {
 					// dispatch never polls, so they MUST still get their poll — skipping
 					// them was the stranded-"Thinking" bug this guard must not revive.
 					if (self.pendingAgentRequests[self.getHistoryCacheKey()] && !item._isBgTask && !item._isOnBgQueue) return;
-					self.historyItemPolls.set(item.id, true);
+					// Background indexing polls are suppressed while paused; foreground
+					// replies the user is waiting on are not.
+					if ((item._isBgTask || item._isOnBgQueue) && self.isPollingPaused()) return;
 					var capturedId = item.id;
 					var pp = item.poll({
 						latency: POLL_INTERVAL,
-						onResponse: function (response: any) { self.handleHistoryItemResolution(capturedId, response, platform); },
+						onResponse: function (response: any) { if (isPollStopped(response)) return; self.handleHistoryItemResolution(capturedId, response, platform); },
 						onError: function (err: any) {
 							self.historyItemPolls.delete(capturedId);
 							var isNotExists = err && (err.code === 'NOT_EXISTS' || (err.body && err.body.code === 'NOT_EXISTS'));
@@ -984,6 +1338,12 @@ export class ChatSession {
 							}
 						},
 					});
+					// Anything on the BACKGROUND queue is pausable, not just items whose
+					// prompt text we recognise as an indexing task. _isBgTask vs
+					// _isOnBgQueue is a DISPLAY distinction ("Indexing: file" vs a normal
+					// chat bubble); keying the pause off _isBgTask alone left every
+					// bg-queue item we could not text-match polling forever.
+					self._trackPoll(capturedId, (item._isBgTask || item._isOnBgQueue) ? 'bg' : 'fg', pp);
 					if (pp && pp.catch) pp.catch(function () { });
 				});
 				self.drainBgTaskQueue();