@oh-my-pi/pi-coding-agent 15.12.3 → 15.12.4

package/src/modes/components/mcp-add-wizard.ts

@@ -262,7 +262,7 @@ export class MCPAddWizard extends Container {
 		}
 
 		this.#contentContainer.addChild(
-			new Text(theme.fg("muted", "[Only letters, numbers, dash, underscore, dot]"), 0, 0),
+			new Text(theme.fg("muted", "[Only letters, numbers, dash, underscore, dot, colon]"), 0, 0),
 		);
 		this.#contentContainer.addChild(new Text(theme.fg("muted", "[Enter to continue, Esc to cancel]"), 0, 0));
 	}

package/src/modes/components/model-selector.ts

@@ -711,7 +711,7 @@ export class ModelSelectorComponent extends Container {
 		if (!this.#isModelOverContextLimit(model)) {
 			return "";
 		}
-		return ` ${theme.status.disabled} context>${formatNumber(model.contextWindow).toLowerCase()}`;
+		return ` ${theme.status.disabled} context>${formatNumber(model.contextWindow ?? 0).toLowerCase()}`;
 	}
 
 	#getVisibleItems(): ReadonlyArray<ModelItem | CanonicalModelItem> {
@@ -1016,7 +1016,7 @@ export class ModelSelectorComponent extends Container {
 			const limitWarning = this.#isItemDisabled(selected)
 				? theme.fg(
 						"dim",
-						` — current context ${formatNumber(this.#currentContextTokens).toLowerCase()} > ${formatNumber(selected.model.contextWindow).toLowerCase()} limit`,
+						` — current context ${formatNumber(this.#currentContextTokens).toLowerCase()} > ${formatNumber(selected.model.contextWindow ?? 0).toLowerCase()} limit`,
 					)
 				: "";
 			this.#listContainer.addChild(

package/src/modes/components/status-line/component.ts

@@ -1,7 +1,6 @@
 import * as fs from "node:fs";
 import * as path from "node:path";
 import type { AgentMessage } from "@oh-my-pi/pi-agent-core";
-import { estimateTokens } from "@oh-my-pi/pi-agent-core/compaction";
 import { type Component, truncateToWidth, visibleWidth } from "@oh-my-pi/pi-tui";
 import { getProjectDir } from "@oh-my-pi/pi-utils";
 import { $ } from "bun";
@@ -11,7 +10,6 @@ import * as git from "../../../utils/git";
 import { getSessionAccentAnsi, getSessionAccentHex } from "../../../utils/session-color";
 import { sanitizeStatusText } from "../../shared";
 import { theme } from "../../theme/theme";
-import { computeNonMessageTokens } from "../../utils/context-usage";
 import { canReuseCachedPr, createPrCacheContext, isSamePrCacheContext, type PrCacheContext } from "./git-utils";
 import { getPreset } from "./presets";
 import { renderSegment, type SegmentContext } from "./segments";
@@ -26,30 +24,15 @@ import type {
 } from "./types";
 
 // ═══════════════════════════════════════════════════════════════════════════
-// Per-message token cache
+// Context-usage memo
 // ═══════════════════════════════════════════════════════════════════════════
 
 /**
- * Symbol-keyed sidecar tagged onto each `AgentMessage` to memoize its
- * `estimateTokens` result. Keyed by message identity (the object itself);
- * a cheap content fingerprint detects in-place mutations (post-hoc error
- * attachment, retry-truncated branch rebuild, etc.) and forces recompute.
- *
- * Cache lives on the message — multiple `StatusLineComponent` instances
- * share it for free, and entries collect with the message itself when the
- * conversation is replaced or compacted.
- */
-const kTokenCache = Symbol("statusLine.tokenCache");
-interface TaggedMessage {
-	[kTokenCache]?: { fingerprint: string; tokens: number };
-}
-
-/**
- * Cheap structural fingerprint mirroring `estimateTokens`'s content walk.
- * O(blocks) — only reads string `.length` and primitives, never copies or
- * serializes content. Any in-place mutation that alters total tokenized
- * content also alters one of the byte-length sums or block counts captured
- * here, forcing the cached `estimateTokens` value to be recomputed.
+ * Cheap structural fingerprint of a message's tokenizable content. O(blocks) —
+ * only reads string `.length` and primitives, never copies or serializes.
+ * Detects in-place growth of the streaming tail (and other in-place mutations)
+ * so the cached `getContextUsage()` result is recomputed when — and only when —
+ * the numbers it depends on change.
  */
 function messageFingerprint(msg: AgentMessage): string {
 	const role = (msg as { role?: string }).role ?? "";
@@ -107,29 +90,17 @@ function messageFingerprint(msg: AgentMessage): string {
 	return `${role}:${ts}:${textLen}:${blocks}:${images}`;
 }
 
-/**
- * Token count for a single message, using the per-message sidecar cache.
- * The caller MUST skip caching for the last message during streaming —
- * it may still be growing and its tokens belong recomputed each refresh.
- */
-function tokensForMessage(msg: AgentMessage): number {
-	const fp = messageFingerprint(msg);
-	const tagged = msg as TaggedMessage;
-	const cached = tagged[kTokenCache];
-	if (cached && cached.fingerprint === fp) return cached.tokens;
-	const tokens = estimateTokens(msg);
-	tagged[kTokenCache] = { fingerprint: fp, tokens };
-	return tokens;
-}
-
-interface MessageTokenTotalsCache {
+interface ContextUsageMemo {
 	messagesRef: readonly AgentMessage[];
-	stableCount: number;
-	stableTokens: number;
-	lastStableMessage: AgentMessage | undefined;
-	lastStableFingerprint: string | undefined;
+	length: number;
+	lastFingerprint: string | undefined;
+	modelContextWindow: number;
+	usedTokens: number | null;
+	contextWindow: number;
 }
 
+const EMPTY_MESSAGES: readonly AgentMessage[] = [];
+
 function hasContextSegment(segments: readonly StatusLineSegmentId[]): boolean {
 	return segments.includes("context_pct") || segments.includes("context_total");
 }
@@ -176,19 +147,12 @@ export class StatusLineComponent implements Component {
 	} | null = null;
 	#usageFetchedAt = 0;
 	#usageInFlight = false;
-	// Context breakdown — incremental rolling cache. The status line refreshes
-	// on every agent event, so the hot path must not re-tokenize the full
-	// message list. Stable messages are accumulated once; normal streaming
-	// refreshes only recompute the current tail message and newly appended
-	// entries. History rewrites/compaction replace or shrink the message array
-	// and rebuild this cache. Stable messages are treated as immutable after
-	// promotion, matching the normal append-only session flow.
-	// Cached non-message total (system prompt + tools + skills). Invalidated
-	// when the inputs-identity fingerprint changes (model swap, skill toggle,
-	// tool registration).
-	#nonMessageTokensCache: number | undefined;
-	#nonMessageInputsKey: string | undefined;
-	#messageTokenTotalsCache: MessageTokenTotalsCache | undefined;
+	// Context-usage memo. The status line redraws on every agent event, so the
+	// hot path must not recompute context tokens unless an input changed.
+	// `getContextUsage()` anchors on the last assistant's real prompt-token
+	// count (matching the provider and the `/context` panel), so a stable
+	// message list + model window yields a stable result we can return verbatim.
+	#contextUsageCache: ContextUsageMemo | undefined;
 
 	constructor(private session: AgentSession) {
 		this.#settings = {
@@ -310,9 +274,7 @@ export class StatusLineComponent implements Component {
 		this.#cachedUsage = null;
 		this.#usageFetchedAt = 0;
 		this.#usageInFlight = false;
-		this.#nonMessageTokensCache = undefined;
-		this.#nonMessageInputsKey = undefined;
-		this.#messageTokenTotalsCache = undefined;
+		this.#contextUsageCache = undefined;
 		this.#lastTokensPerSecond = null;
 		this.#lastTokensPerSecondTimestamp = null;
 	}
@@ -544,109 +506,44 @@ export class StatusLineComponent implements Component {
 	}
 
 	/**
-	 * Compute the (cached) used-tokens / context-window totals for the
-	 * status-line context% segment. Exposed (non-private) so unit tests can
-	 * verify the incremental-cache invariants; not part of any external
-	 * API.
+	 * Used-tokens / context-window totals for the status-line context% segment,
+	 * memoized so the per-event redraw stays O(1) when nothing changed.
+	 *
+	 * The numerator comes from `session.getContextUsage()`, which anchors on the
+	 * last assistant's real prompt-token count — so the bar matches the provider
+	 * and the `/context` panel — and reports `null` while that count is unknown
+	 * (right after compaction, before the next response). Exposed (non-private)
+	 * for unit tests and the collab host's state broadcast.
 	 */
-	getCachedContextBreakdown(): { usedTokens: number; contextWindow: number } {
-		const messages = this.session.messages ?? [];
-		const contextWindow = this.session.model?.contextWindow ?? 0;
-
-		// 1) Non-message tokens (system prompt + tools + skills). Refresh only
-		//    when the inputs identity fingerprint changes — usually never
-		//    during a streaming turn. ~10-30 ms when it does refresh.
-		const inputsKey = this.#computeNonMessageInputsKey();
-		if (this.#nonMessageTokensCache === undefined || this.#nonMessageInputsKey !== inputsKey) {
-			this.#nonMessageTokensCache = computeNonMessageTokens(this.session);
-			this.#nonMessageInputsKey = inputsKey;
-		}
-
-		// 2) Message tokens — incremental rolling total. The sidecar cache lives
-		//    on each stable message object (all but the current tail). Normal
-		//    streaming turns only recompute the last message and newly appended
-		//    entries. Full rebuild only when the message array is replaced,
-		//    shrinks, or the recently-promoted stable tail mutates in place.
-		const messagesTokens = this.#getCachedMessageTokens(messages);
-
-		const usedTokens = this.#nonMessageTokensCache + messagesTokens;
-		return { usedTokens, contextWindow };
-	}
-
-	#getCachedMessageTokens(messages: readonly AgentMessage[]): number {
-		const cache = this.#messageTokenTotalsCache;
-		if (!cache || cache.messagesRef !== messages || messages.length <= cache.stableCount) {
-			return this.#rebuildMessageTokenTotals(messages);
-		}
-
-		let stableTokens = cache.stableTokens;
-		let stableCount = cache.stableCount;
-		const stableLimit = Math.max(0, messages.length - 1);
+	getCachedContextBreakdown(): { usedTokens: number | null; contextWindow: number } {
+		const messages = this.session.messages ?? EMPTY_MESSAGES;
+		const modelContextWindow = this.session.model?.contextWindow ?? 0;
+		const length = messages.length;
+		const lastFingerprint = length > 0 ? messageFingerprint(messages[length - 1]!) : undefined;
 
+		const cache = this.#contextUsageCache;
 		if (
-			cache.lastStableMessage &&
-			stableCount > 0 &&
-			messages[stableCount - 1] === cache.lastStableMessage &&
-			cache.lastStableFingerprint !== undefined &&
-			cache.lastStableFingerprint !== messageFingerprint(cache.lastStableMessage)
+			cache &&
+			cache.messagesRef === messages &&
+			cache.length === length &&
+			cache.lastFingerprint === lastFingerprint &&
+			cache.modelContextWindow === modelContextWindow
 		) {
-			return this.#rebuildMessageTokenTotals(messages);
-		}
-
-		while (stableCount < stableLimit) {
-			const promoted = messages[stableCount]!;
-			stableTokens += tokensForMessage(promoted);
-			stableCount++;
-		}
-
-		const lastStableMessage = stableCount > 0 ? messages[stableCount - 1] : undefined;
-		const lastStableFingerprint = lastStableMessage ? messageFingerprint(lastStableMessage) : undefined;
-		const lastMessage = messages.at(-1);
-		const lastTokens = lastMessage ? estimateTokens(lastMessage) : 0;
-		this.#messageTokenTotalsCache = {
-			messagesRef: messages,
-			stableCount,
-			stableTokens,
-			lastStableMessage,
-			lastStableFingerprint,
-		};
-		return stableTokens + lastTokens;
-	}
-
-	#rebuildMessageTokenTotals(messages: readonly AgentMessage[]): number {
-		let stableTokens = 0;
-		const stableLimit = Math.max(0, messages.length - 1);
-		for (let i = 0; i < stableLimit; i++) {
-			stableTokens += tokensForMessage(messages[i]!);
+			return { usedTokens: cache.usedTokens, contextWindow: cache.contextWindow };
 		}
 
-		const lastStableMessage = stableLimit > 0 ? messages[stableLimit - 1] : undefined;
-		const lastStableFingerprint = lastStableMessage ? messageFingerprint(lastStableMessage) : undefined;
-		const lastMessage = messages.at(-1);
-		const lastTokens = lastMessage ? estimateTokens(lastMessage) : 0;
-
-		this.#messageTokenTotalsCache = {
+		const usage = this.session.getContextUsage();
+		const usedTokens = usage?.tokens ?? null;
+		const contextWindow = usage?.contextWindow ?? modelContextWindow;
+		this.#contextUsageCache = {
 			messagesRef: messages,
-			stableCount: stableLimit,
-			stableTokens,
-			lastStableMessage,
-			lastStableFingerprint,
+			length,
+			lastFingerprint,
+			modelContextWindow,
+			usedTokens,
+			contextWindow,
 		};
-		return stableTokens + lastTokens;
-	}
-
-	/**
-	 * Build an identity fingerprint for the non-message inputs (system prompt,
-	 * tools, skills). When this changes, the non-message token cache must be
-	 * recomputed. Cheap: just lengths + first-string-length. Doesn't need to
-	 * be cryptographically unique — only stable for the same inputs.
-	 */
-	#computeNonMessageInputsKey(): string {
-		const sp = this.session.systemPrompt ?? [];
-		const tools = this.session.agent?.state?.tools ?? [];
-		const skills = this.session.skills ?? [];
-		const modelId = this.session.model?.id ?? "";
-		return `${modelId}|${sp.length}:${sp[0]?.length ?? 0}|${tools.length}|${skills.length}`;
+		return { usedTokens, contextWindow };
 	}
 
 	#buildSegmentContext(
@@ -673,14 +570,14 @@ export class StatusLineComponent implements Component {
 			tokensPerSecond: this.#getTokensPerSecond(),
 		};
 
-		let contextTokens = 0;
 		let contextWindow = state.model?.contextWindow ?? this.session.model?.contextWindow ?? 0;
+		let contextPercent: number | null = 0;
 		if (includeContext) {
 			const breakdown = this.getCachedContextBreakdown();
-			contextTokens = breakdown.usedTokens;
 			contextWindow = breakdown.contextWindow || contextWindow;
+			contextPercent =
+				breakdown.usedTokens === null ? null : contextWindow > 0 ? (breakdown.usedTokens / contextWindow) * 100 : 0;
 		}
-		let contextPercent = contextWindow > 0 ? (contextTokens / contextWindow) * 100 : 0;
 
 		// Collab guest: context comes from the host's state frames — the local
 		// replica does no accounting of its own.

package/src/modes/components/status-line/segments.ts

@@ -364,7 +364,7 @@ const contextPctSegment: StatusLineSegment = {
 		const autoIcon = ctx.autoCompactEnabled && theme.icon.auto ? ` ${theme.icon.auto}` : "";
 		const text = `${formatContextUsage(pct, window)}${autoIcon}`;
 
-		const color = getContextUsageThemeColor(getContextUsageLevel(pct, window));
+		const color = getContextUsageThemeColor(getContextUsageLevel(pct ?? 0, window));
 		const content = withIcon(theme.icon.context, theme.fg(color, text));
 
 		return { content, visible: true };

package/src/modes/components/status-line/types.ts

@@ -71,7 +71,8 @@ export interface SegmentContext {
 		cost: number;
 		tokensPerSecond: number | null;
 	};
-	contextPercent: number;
+	/** Context usage percent, or null when unknown (e.g. right after compaction). */
+	contextPercent: number | null;
 	contextWindow: number;
 	autoCompactEnabled: boolean;
 	subagentCount: number;

package/src/modes/controllers/command-controller.ts

@@ -1026,11 +1026,6 @@ export class CommandController {
 		}
 		this.ctx.statusContainer.clear();
 
-		const originalOnEscape = this.ctx.editor.onEscape;
-		this.ctx.editor.onEscape = () => {
-			this.ctx.session.abortCompaction();
-		};
-
 		this.ctx.chatContainer.addChild(new Spacer(1));
 		const label = isAuto ? "Auto-compacting context... (esc to cancel)" : "Compacting context... (esc to cancel)";
 		const compactingLoader = new Loader(
@@ -1068,7 +1063,6 @@ export class CommandController {
 		} finally {
 			compactingLoader.stop();
 			this.ctx.statusContainer.clear();
-			this.ctx.editor.onEscape = originalOnEscape;
 		}
 		await this.ctx.flushCompactionQueue({ willRetry: false });
 		return outcome;
@@ -1089,11 +1083,6 @@ export class CommandController {
 		}
 		this.ctx.statusContainer.clear();
 
-		const originalOnEscape = this.ctx.editor.onEscape;
-		this.ctx.editor.onEscape = () => {
-			this.ctx.session.abortHandoff();
-		};
-
 		const handoffLoader = new Loader(
 			this.ctx.ui,
 			spinner => theme.fg("accent", spinner),
@@ -1138,7 +1127,6 @@ export class CommandController {
 		} finally {
 			handoffLoader.stop();
 			this.ctx.statusContainer.clear();
-			this.ctx.editor.onEscape = originalOnEscape;
 		}
 		this.ctx.ui.requestRender();
 	}

package/src/modes/controllers/event-controller.ts

@@ -17,8 +17,9 @@ import { getSymbolTheme, theme } from "../../modes/theme/theme";
 import type { InteractiveModeContext, TodoPhase } from "../../modes/types";
 import type { PlanApprovalDetails } from "../../plan-mode/approved-plan";
 import type { AgentSessionEvent } from "../../session/agent-session";
-import { isSilentAbort, readPendingDisplayTag, resolveAbortLabel } from "../../session/messages";
+import { isSilentAbort, readQueueChipText, resolveAbortLabel } from "../../session/messages";
 import type { ResolveToolDetails } from "../../tools/resolve";
+import { hasVisibleThinking } from "../../utils/thinking-display";
 import { interruptHint } from "../shared";
 import { StreamingRevealController } from "./streaming-reveal";
 import { ToolArgsRevealController } from "./tool-args-reveal";
@@ -37,16 +38,6 @@ const IRC_MESSAGE_VISIBLE_TTL_MS = 10_000;
  */
 const MAX_LIVE_IRC_CARDS = 4;
 
-/**
- * Loader label shown the instant a user interrupt (Esc) is requested, kept until
- * the agent turn fully unwinds. Esc fires the abort synchronously, but the loop
- * only stops the spinner at `agent_end`, which it cannot reach until every
- * in-flight tool settles its abort in `executeToolCalls` (`Promise.allSettled`).
- * Swapping the steady "Working…" for this acknowledges the keypress instead of
- * reading as an ignored Esc for the seconds a slow tool takes to tear down.
- */
-export const INTERRUPTING_WORKING_MESSAGE = "Interrupting…";
-
 type AgentSessionEventHandlers = {
 	[E in AgentSessionEventKind]: (event: Extract<AgentSessionEvent, { type: E }>) => Promise<void>;
 };
@@ -63,8 +54,6 @@ export class EventController {
 	#renderedCustomMessages = new Set<string>();
 	#lastIntent: string | undefined = undefined;
 	#backgroundToolCallIds = new Set<string>();
-	#agentTurnActive = false;
-	#interrupting = false;
 	#readToolCallArgs = new Map<string, Record<string, unknown>>();
 	#readToolCallAssistantComponents = new Map<string, AssistantMessageComponent>();
 	#lastAssistantComponent: AssistantMessageComponent | undefined = undefined;
@@ -193,7 +182,7 @@ export class EventController {
 		return true;
 	}
 	#updateWorkingMessageFromIntent(intent: unknown): void {
-		if (this.#interrupting) return;
+		if (this.ctx.session.isAborting) return;
 		// Streamed JSON can deliver non-string `_i` (object, number, boolean) before
 		// schema validation; `?.` only guards null/undefined, so guard the type too.
 		if (typeof intent !== "string") return;
@@ -203,19 +192,6 @@ export class EventController {
 		this.ctx.setWorkingMessage(`${trimmed}${interruptHint()}`);
 	}
 
-	/**
-	 * Acknowledge a user interrupt (Esc) immediately: switch the loader to
-	 * `INTERRUPTING_WORKING_MESSAGE` and freeze intent-driven working-message
-	 * updates for the rest of the turn so a late `tool_execution_start` intent
-	 * cannot repaint a "Working…/<intent>" line over the acknowledgment. Reset at
-	 * the next `agent_start`. No-op outside an active turn or if already set.
-	 */
-	notifyInterrupting(): void {
-		if (!this.#agentTurnActive || this.#interrupting) return;
-		this.#interrupting = true;
-		this.ctx.setWorkingMessage(INTERRUPTING_WORKING_MESSAGE);
-	}
-
 	subscribeToAgent(): void {
 		this.ctx.unsubscribe = this.ctx.session.subscribe(async (event: AgentSessionEvent) => {
 			await this.handleEvent(event);
@@ -233,8 +209,6 @@ export class EventController {
 		this.#renderedCustomMessages.clear();
 		this.#lastIntent = undefined;
 		this.#backgroundToolCallIds.clear();
-		this.#agentTurnActive = false;
-		this.#interrupting = false;
 		this.#readToolCallArgs.clear();
 		this.#readToolCallAssistantComponents.clear();
 		this.#lastAssistantComponent = undefined;
@@ -264,8 +238,6 @@ export class EventController {
 	}
 
 	async #handleAgentStart(_event: Extract<AgentSessionEvent, { type: "agent_start" }>): Promise<void> {
-		this.#agentTurnActive = true;
-		this.#interrupting = false;
 		this.#lastIntent = undefined;
 		this.#readToolCallArgs.clear();
 		this.#readToolCallAssistantComponents.clear();
@@ -276,10 +248,6 @@ export class EventController {
 		this.#pinnedErrorComponent?.setErrorPinned(false);
 		this.#pinnedErrorComponent = undefined;
 		this.ctx.clearPinnedError();
-		if (this.ctx.retryEscapeHandler) {
-			this.ctx.editor.onEscape = this.ctx.retryEscapeHandler;
-			this.ctx.retryEscapeHandler = undefined;
-		}
 		if (this.ctx.retryLoader) {
 			this.ctx.retryLoader.stop();
 			this.ctx.retryLoader = undefined;
@@ -299,15 +267,10 @@ export class EventController {
 			this.#renderedCustomMessages.add(signature);
 			this.#resetReadGroup();
 			this.ctx.addMessageToChat(event.message);
-			// Tag-keyed pending-bar refresh: when AgentSession.#handleAgentEvent
-			// spliced this dequeued custom message out of #steeringMessages /
-			// #followUpMessages (it ran before this emit), the array state is
-			// already correct — pendingMessagesContainer just needs to be
-			// re-rendered to match. Gated on tag presence so non-queued customs
-			// (ttsr-injection, irc:*, async-result, hookMessage) skip the
-			// rebuild; their dispatch path never registered a pending chip.
-			// Mirrors the user-role refresh at the bottom of this function.
-			if (event.message.role === "custom" && readPendingDisplayTag(event.message.details)) {
+			// Queued custom-message chips are derived from the agent queue; refresh the
+			// pending bar when the queued custom is consumed so the chip disappears
+			// immediately.
+			if (event.message.role === "custom" && readQueueChipText(event.message.details)) {
 				this.ctx.updatePendingMessagesDisplay();
 			}
 			this.ctx.ui.requestRender();
@@ -475,7 +438,7 @@ export class EventController {
 			const visibleBlockCount = this.ctx.streamingMessage.content.filter(
 				content =>
 					(content.type === "text" && content.text.trim().length > 0) ||
-					(content.type === "thinking" && content.thinking.trim().length > 0),
+					(content.type === "thinking" && hasVisibleThinking(content)),
 			).length;
 			if (visibleBlockCount > this.#lastVisibleBlockCount) {
 				this.#resetReadGroup();
@@ -823,7 +786,21 @@ export class EventController {
 		}
 	}
 	async #handleAgentEnd(_event: Extract<AgentSessionEvent, { type: "agent_end" }>): Promise<void> {
-		this.#agentTurnActive = false;
+		// A superseded agent_end: the agent is already streaming a fresh turn, so
+		// this event belongs to a turn that has already been replaced. The session
+		// dispatches to listeners fire-and-forget across an async extension-emit hop
+		// (#emitSessionEvent), so an interrupted turn's agent_end can land AFTER the
+		// resumed turn's agent_start (e.g. any post-turn agent.continue()). Running
+		// the turn-end teardown now would stop the loader the live turn just created,
+		// leaving "Working…" gone while the agent keeps running. The live turn owns
+		// the loader and finalizes it at its own agent_end (isStreaming === false by
+		// then). Mirrors the collab guest's !isStreaming loader reconciler.
+		if (this.ctx.session.isStreaming) return;
+
+		await this.#finishAgentEnd();
+	}
+
+	async #finishAgentEnd(): Promise<void> {
 		this.#streamingReveal.stop();
 		this.#toolArgsReveal.flushAll();
 		if (this.ctx.loadingAnimation) {
@@ -871,10 +848,6 @@ export class EventController {
 		event: Extract<AgentSessionEvent, { type: "auto_compaction_start" }>,
 	): Promise<void> {
 		this.#cancelIdleCompaction();
-		this.ctx.autoCompactionEscapeHandler = this.ctx.editor.onEscape;
-		this.ctx.editor.onEscape = () => {
-			this.ctx.viewSession.abortCompaction();
-		};
 		this.ctx.statusContainer.clear();
 		const reasonText =
 			event.reason === "overflow"
@@ -903,10 +876,6 @@ export class EventController {
 
 	async #handleAutoCompactionEnd(event: Extract<AgentSessionEvent, { type: "auto_compaction_end" }>): Promise<void> {
 		this.#cancelIdleCompaction();
-		if (this.ctx.autoCompactionEscapeHandler) {
-			this.ctx.editor.onEscape = this.ctx.autoCompactionEscapeHandler;
-			this.ctx.autoCompactionEscapeHandler = undefined;
-		}
 		if (this.ctx.autoCompactionLoader) {
 			this.ctx.autoCompactionLoader.stop();
 			this.ctx.autoCompactionLoader = undefined;
@@ -964,10 +933,6 @@ export class EventController {
 	}
 
 	async #handleAutoRetryStart(event: Extract<AgentSessionEvent, { type: "auto_retry_start" }>): Promise<void> {
-		this.ctx.retryEscapeHandler = this.ctx.editor.onEscape;
-		this.ctx.editor.onEscape = () => {
-			this.ctx.viewSession.abortRetry();
-		};
 		this.ctx.statusContainer.clear();
 		const delaySeconds = Math.round(event.delayMs / 1000);
 		this.ctx.retryLoader = new Loader(
@@ -982,10 +947,6 @@ export class EventController {
 	}
 
 	async #handleAutoRetryEnd(event: Extract<AgentSessionEvent, { type: "auto_retry_end" }>): Promise<void> {
-		if (this.ctx.retryEscapeHandler) {
-			this.ctx.editor.onEscape = this.ctx.retryEscapeHandler;
-			this.ctx.retryEscapeHandler = undefined;
-		}
 		if (this.ctx.retryLoader) {
 			this.ctx.retryLoader.stop();
 			this.ctx.retryLoader = undefined;