pi-usereq 0.7.0 → 0.11.0

package/pi.dev-src/pi-mono/coding-agent/src/core/agent-session.ts

@@ -0,0 +1,3096 @@
+/**
+ * AgentSession - Core abstraction for agent lifecycle and session management.
+ *
+ * This class is shared between all run modes (interactive, print, rpc).
+ * It encapsulates:
+ * - Agent state access
+ * - Event subscription with automatic session persistence
+ * - Model and thinking level management
+ * - Compaction (manual and auto)
+ * - Bash execution
+ * - Session switching and branching
+ *
+ * Modes use this class and add their own I/O layer on top.
+ */
+
+import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
+import { basename, dirname, join, resolve } from "node:path";
+import type {
+	Agent,
+	AgentEvent,
+	AgentMessage,
+	AgentState,
+	AgentTool,
+	ThinkingLevel,
+} from "@mariozechner/pi-agent-core";
+import type { AssistantMessage, ImageContent, Message, Model, TextContent } from "@mariozechner/pi-ai";
+import { isContextOverflow, modelsAreEqual, resetApiProviders, supportsXhigh } from "@mariozechner/pi-ai";
+import { getDocsPath } from "../config.js";
+import { theme } from "../modes/interactive/theme/theme.js";
+import { stripFrontmatter } from "../utils/frontmatter.js";
+import { sleep } from "../utils/sleep.js";
+import { type BashResult, executeBashWithOperations } from "./bash-executor.js";
+import {
+	type CompactionResult,
+	calculateContextTokens,
+	collectEntriesForBranchSummary,
+	compact,
+	estimateContextTokens,
+	generateBranchSummary,
+	prepareCompaction,
+	shouldCompact,
+} from "./compaction/index.js";
+import { DEFAULT_THINKING_LEVEL } from "./defaults.js";
+import { exportSessionToHtml, type ToolHtmlRenderer } from "./export-html/index.js";
+import { createToolHtmlRenderer } from "./export-html/tool-renderer.js";
+import {
+	type ContextUsage,
+	type ExtensionCommandContextActions,
+	type ExtensionErrorListener,
+	ExtensionRunner,
+	type ExtensionUIContext,
+	type InputSource,
+	type MessageEndEvent,
+	type MessageStartEvent,
+	type MessageUpdateEvent,
+	type ReplacedSessionContext,
+	type SessionBeforeCompactResult,
+	type SessionBeforeTreeResult,
+	type SessionStartEvent,
+	type ShutdownHandler,
+	type ToolDefinition,
+	type ToolExecutionEndEvent,
+	type ToolExecutionStartEvent,
+	type ToolExecutionUpdateEvent,
+	type ToolInfo,
+	type TreePreparation,
+	type TurnEndEvent,
+	type TurnStartEvent,
+	wrapRegisteredTools,
+} from "./extensions/index.js";
+import { emitSessionShutdownEvent } from "./extensions/runner.js";
+import type { BashExecutionMessage, CustomMessage } from "./messages.js";
+import type { ModelRegistry } from "./model-registry.js";
+import { expandPromptTemplate, type PromptTemplate } from "./prompt-templates.js";
+import type { ResourceExtensionPaths, ResourceLoader } from "./resource-loader.js";
+import type { BranchSummaryEntry, CompactionEntry, SessionManager } from "./session-manager.js";
+import { CURRENT_SESSION_VERSION, getLatestCompactionEntry, type SessionHeader } from "./session-manager.js";
+import type { SettingsManager } from "./settings-manager.js";
+import type { SlashCommandInfo } from "./slash-commands.js";
+import { createSyntheticSourceInfo, type SourceInfo } from "./source-info.js";
+import { type BuildSystemPromptOptions, buildSystemPrompt } from "./system-prompt.js";
+import { type BashOperations, createLocalBashOperations } from "./tools/bash.js";
+import { createAllToolDefinitions } from "./tools/index.js";
+import { createToolDefinitionFromAgentTool } from "./tools/tool-definition-wrapper.js";
+
+// ============================================================================
+// Skill Block Parsing
+// ============================================================================
+
+/** Parsed skill block from a user message */
+export interface ParsedSkillBlock {
+	name: string;
+	location: string;
+	content: string;
+	userMessage: string | undefined;
+}
+
+/**
+ * Parse a skill block from message text.
+ * Returns null if the text doesn't contain a skill block.
+ */
+export function parseSkillBlock(text: string): ParsedSkillBlock | null {
+	const match = text.match(/^<skill name="([^"]+)" location="([^"]+)">\n([\s\S]*?)\n<\/skill>(?:\n\n([\s\S]+))?$/);
+	if (!match) return null;
+	return {
+		name: match[1],
+		location: match[2],
+		content: match[3],
+		userMessage: match[4]?.trim() || undefined,
+	};
+}
+
+/** Session-specific events that extend the core AgentEvent */
+export type AgentSessionEvent =
+	| AgentEvent
+	| {
+			type: "queue_update";
+			steering: readonly string[];
+			followUp: readonly string[];
+	  }
+	| { type: "compaction_start"; reason: "manual" | "threshold" | "overflow" }
+	| {
+			type: "compaction_end";
+			reason: "manual" | "threshold" | "overflow";
+			result: CompactionResult | undefined;
+			aborted: boolean;
+			willRetry: boolean;
+			errorMessage?: string;
+	  }
+	| { type: "auto_retry_start"; attempt: number; maxAttempts: number; delayMs: number; errorMessage: string }
+	| { type: "auto_retry_end"; success: boolean; attempt: number; finalError?: string };
+
+/** Listener function for agent session events */
+export type AgentSessionEventListener = (event: AgentSessionEvent) => void;
+
+// ============================================================================
+// Types
+// ============================================================================
+
+export interface AgentSessionConfig {
+	agent: Agent;
+	sessionManager: SessionManager;
+	settingsManager: SettingsManager;
+	cwd: string;
+	/** Models to cycle through with Ctrl+P (from --models flag) */
+	scopedModels?: Array<{ model: Model<any>; thinkingLevel?: ThinkingLevel }>;
+	/** Resource loader for skills, prompts, themes, context files, system prompt */
+	resourceLoader: ResourceLoader;
+	/** SDK custom tools registered outside extensions */
+	customTools?: ToolDefinition[];
+	/** Model registry for API key resolution and model discovery */
+	modelRegistry: ModelRegistry;
+	/** Initial active built-in tool names. Default: [read, bash, edit, write] */
+	initialActiveToolNames?: string[];
+	/** Optional allowlist of tool names. When provided, only these tool names are exposed. */
+	allowedToolNames?: string[];
+	/**
+	 * Override base tools (useful for custom runtimes).
+	 *
+	 * These are synthesized into minimal ToolDefinitions internally so AgentSession can keep
+	 * a definition-first registry even when callers provide plain AgentTool instances.
+	 */
+	baseToolsOverride?: Record<string, AgentTool>;
+	/** Mutable ref used by Agent to access the current ExtensionRunner */
+	extensionRunnerRef?: { current?: ExtensionRunner };
+	/** Session start event metadata emitted when extensions bind to this runtime. */
+	sessionStartEvent?: SessionStartEvent;
+}
+
+export interface ExtensionBindings {
+	uiContext?: ExtensionUIContext;
+	commandContextActions?: ExtensionCommandContextActions;
+	shutdownHandler?: ShutdownHandler;
+	onError?: ExtensionErrorListener;
+}
+
+/** Options for AgentSession.prompt() */
+export interface PromptOptions {
+	/** Whether to expand file-based prompt templates (default: true) */
+	expandPromptTemplates?: boolean;
+	/** Image attachments */
+	images?: ImageContent[];
+	/** When streaming, how to queue the message: "steer" (interrupt) or "followUp" (wait). Required if streaming. */
+	streamingBehavior?: "steer" | "followUp";
+	/** Source of input for extension input event handlers. Defaults to "interactive". */
+	source?: InputSource;
+	/** Internal hook used by RPC mode to observe prompt preflight acceptance or rejection. */
+	preflightResult?: (success: boolean) => void;
+}
+
+/** Result from cycleModel() */
+export interface ModelCycleResult {
+	model: Model<any>;
+	thinkingLevel: ThinkingLevel;
+	/** Whether cycling through scoped models (--models flag) or all available */
+	isScoped: boolean;
+}
+
+/** Session statistics for /session command */
+export interface SessionStats {
+	sessionFile: string | undefined;
+	sessionId: string;
+	userMessages: number;
+	assistantMessages: number;
+	toolCalls: number;
+	toolResults: number;
+	totalMessages: number;
+	tokens: {
+		input: number;
+		output: number;
+		cacheRead: number;
+		cacheWrite: number;
+		total: number;
+	};
+	cost: number;
+	contextUsage?: ContextUsage;
+}
+
+interface ToolDefinitionEntry {
+	definition: ToolDefinition;
+	sourceInfo: SourceInfo;
+}
+
+// ============================================================================
+// Constants
+// ============================================================================
+
+/** Standard thinking levels */
+const THINKING_LEVELS: ThinkingLevel[] = ["off", "minimal", "low", "medium", "high"];
+
+/** Thinking levels including xhigh (for supported models) */
+const THINKING_LEVELS_WITH_XHIGH: ThinkingLevel[] = ["off", "minimal", "low", "medium", "high", "xhigh"];
+
+// ============================================================================
+// AgentSession Class
+// ============================================================================
+
+export class AgentSession {
+	readonly agent: Agent;
+	readonly sessionManager: SessionManager;
+	readonly settingsManager: SettingsManager;
+
+	private _scopedModels: Array<{ model: Model<any>; thinkingLevel?: ThinkingLevel }>;
+
+	// Event subscription state
+	private _unsubscribeAgent?: () => void;
+	private _eventListeners: AgentSessionEventListener[] = [];
+	private _agentEventQueue: Promise<void> = Promise.resolve();
+
+	/** Tracks pending steering messages for UI display. Removed when delivered. */
+	private _steeringMessages: string[] = [];
+	/** Tracks pending follow-up messages for UI display. Removed when delivered. */
+	private _followUpMessages: string[] = [];
+	/** Messages queued to be included with the next user prompt as context ("asides"). */
+	private _pendingNextTurnMessages: CustomMessage[] = [];
+
+	// Compaction state
+	private _compactionAbortController: AbortController | undefined = undefined;
+	private _autoCompactionAbortController: AbortController | undefined = undefined;
+	private _overflowRecoveryAttempted = false;
+
+	// Branch summarization state
+	private _branchSummaryAbortController: AbortController | undefined = undefined;
+
+	// Retry state
+	private _retryAbortController: AbortController | undefined = undefined;
+	private _retryAttempt = 0;
+	private _retryPromise: Promise<void> | undefined = undefined;
+	private _retryResolve: (() => void) | undefined = undefined;
+
+	// Bash execution state
+	private _bashAbortController: AbortController | undefined = undefined;
+	private _pendingBashMessages: BashExecutionMessage[] = [];
+
+	// Extension system
+	private _extensionRunner!: ExtensionRunner;
+	private _turnIndex = 0;
+
+	private _resourceLoader: ResourceLoader;
+	private _customTools: ToolDefinition[];
+	private _baseToolDefinitions: Map<string, ToolDefinition> = new Map();
+	private _cwd: string;
+	private _extensionRunnerRef?: { current?: ExtensionRunner };
+	private _initialActiveToolNames?: string[];
+	private _allowedToolNames?: Set<string>;
+	private _baseToolsOverride?: Record<string, AgentTool>;
+	private _sessionStartEvent: SessionStartEvent;
+	private _extensionUIContext?: ExtensionUIContext;
+	private _extensionCommandContextActions?: ExtensionCommandContextActions;
+	private _extensionShutdownHandler?: ShutdownHandler;
+	private _extensionErrorListener?: ExtensionErrorListener;
+	private _extensionErrorUnsubscriber?: () => void;
+
+	// Model registry for API key resolution
+	private _modelRegistry: ModelRegistry;
+
+	// Tool registry for extension getTools/setTools
+	private _toolRegistry: Map<string, AgentTool> = new Map();
+	private _toolDefinitions: Map<string, ToolDefinitionEntry> = new Map();
+	private _toolPromptSnippets: Map<string, string> = new Map();
+	private _toolPromptGuidelines: Map<string, string[]> = new Map();
+
+	// Base system prompt (without extension appends) - used to apply fresh appends each turn
+	private _baseSystemPrompt = "";
+	private _baseSystemPromptOptions!: BuildSystemPromptOptions;
+
+	constructor(config: AgentSessionConfig) {
+		this.agent = config.agent;
+		this.sessionManager = config.sessionManager;
+		this.settingsManager = config.settingsManager;
+		this._scopedModels = config.scopedModels ?? [];
+		this._resourceLoader = config.resourceLoader;
+		this._customTools = config.customTools ?? [];
+		this._cwd = config.cwd;
+		this._modelRegistry = config.modelRegistry;
+		this._extensionRunnerRef = config.extensionRunnerRef;
+		this._initialActiveToolNames = config.initialActiveToolNames;
+		this._allowedToolNames = config.allowedToolNames ? new Set(config.allowedToolNames) : undefined;
+		this._baseToolsOverride = config.baseToolsOverride;
+		this._sessionStartEvent = config.sessionStartEvent ?? { type: "session_start", reason: "startup" };
+
+		// Always subscribe to agent events for internal handling
+		// (session persistence, extensions, auto-compaction, retry logic)
+		this._unsubscribeAgent = this.agent.subscribe(this._handleAgentEvent);
+		this._installAgentToolHooks();
+
+		this._buildRuntime({
+			activeToolNames: this._initialActiveToolNames,
+			includeAllExtensionTools: true,
+		});
+	}
+
+	/** Model registry for API key resolution and model discovery */
+	get modelRegistry(): ModelRegistry {
+		return this._modelRegistry;
+	}
+
+	private async _getRequiredRequestAuth(model: Model<any>): Promise<{
+		apiKey: string;
+		headers?: Record<string, string>;
+	}> {
+		const result = await this._modelRegistry.getApiKeyAndHeaders(model);
+		if (!result.ok) {
+			throw new Error(result.error);
+		}
+		if (result.apiKey) {
+			return { apiKey: result.apiKey, headers: result.headers };
+		}
+
+		const isOAuth = this._modelRegistry.isUsingOAuth(model);
+		if (isOAuth) {
+			throw new Error(
+				`Authentication failed for "${model.provider}". ` +
+					`Credentials may have expired or network is unavailable. ` +
+					`Run '/login ${model.provider}' to re-authenticate.`,
+			);
+		}
+		throw new Error(
+			`No API key found for ${model.provider}.\n\n` +
+				`Use /login or set an API key environment variable. See ${join(getDocsPath(), "providers.md")}`,
+		);
+	}
+
+	/**
+	 * Install tool hooks once on the Agent instance.
+	 *
+	 * The callbacks read `this._extensionRunner` at execution time, so extension reload swaps in the
+	 * new runner without reinstalling hooks. Extension-specific tool wrappers are still used to adapt
+	 * registered tool execution to the extension context. Tool call and tool result interception now
+	 * happens here instead of in wrappers.
+	 */
+	private _installAgentToolHooks(): void {
+		this.agent.beforeToolCall = async ({ toolCall, args }) => {
+			const runner = this._extensionRunner;
+			if (!runner.hasHandlers("tool_call")) {
+				return undefined;
+			}
+
+			await this._agentEventQueue;
+
+			try {
+				return await runner.emitToolCall({
+					type: "tool_call",
+					toolName: toolCall.name,
+					toolCallId: toolCall.id,
+					input: args as Record<string, unknown>,
+				});
+			} catch (err) {
+				if (err instanceof Error) {
+					throw err;
+				}
+				throw new Error(`Extension failed, blocking execution: ${String(err)}`);
+			}
+		};
+
+		this.agent.afterToolCall = async ({ toolCall, args, result, isError }) => {
+			const runner = this._extensionRunner;
+			if (!runner.hasHandlers("tool_result")) {
+				return undefined;
+			}
+
+			const hookResult = await runner.emitToolResult({
+				type: "tool_result",
+				toolName: toolCall.name,
+				toolCallId: toolCall.id,
+				input: args as Record<string, unknown>,
+				content: result.content,
+				details: result.details,
+				isError,
+			});
+
+			if (!hookResult) {
+				return undefined;
+			}
+
+			return {
+				content: hookResult.content,
+				details: hookResult.details,
+				isError: hookResult.isError ?? isError,
+			};
+		};
+	}
+
+	// =========================================================================
+	// Event Subscription
+	// =========================================================================
+
+	/** Emit an event to all listeners */
+	private _emit(event: AgentSessionEvent): void {
+		for (const l of this._eventListeners) {
+			l(event);
+		}
+	}
+
+	private _emitQueueUpdate(): void {
+		this._emit({
+			type: "queue_update",
+			steering: [...this._steeringMessages],
+			followUp: [...this._followUpMessages],
+		});
+	}
+
+	// Track last assistant message for auto-compaction check
+	private _lastAssistantMessage: AssistantMessage | undefined = undefined;
+
+	/** Internal handler for agent events - shared by subscribe and reconnect */
+	private _handleAgentEvent = (event: AgentEvent): void => {
+		// Create retry promise synchronously before queueing async processing.
+		// Agent.emit() calls this handler synchronously, and prompt() calls waitForRetry()
+		// as soon as agent.prompt() resolves. If _retryPromise is created only inside
+		// _processAgentEvent, slow earlier queued events can delay agent_end processing
+		// and waitForRetry() can miss the in-flight retry.
+		this._createRetryPromiseForAgentEnd(event);
+
+		this._agentEventQueue = this._agentEventQueue.then(
+			() => this._processAgentEvent(event),
+			() => this._processAgentEvent(event),
+		);
+
+		// Keep queue alive if an event handler fails
+		this._agentEventQueue.catch(() => {});
+	};
+
+	private _createRetryPromiseForAgentEnd(event: AgentEvent): void {
+		if (event.type !== "agent_end" || this._retryPromise) {
+			return;
+		}
+
+		const settings = this.settingsManager.getRetrySettings();
+		if (!settings.enabled) {
+			return;
+		}
+
+		const lastAssistant = this._findLastAssistantInMessages(event.messages);
+		if (!lastAssistant || !this._isRetryableError(lastAssistant)) {
+			return;
+		}
+
+		this._retryPromise = new Promise((resolve) => {
+			this._retryResolve = resolve;
+		});
+	}
+
+	private _findLastAssistantInMessages(messages: AgentMessage[]): AssistantMessage | undefined {
+		for (let i = messages.length - 1; i >= 0; i--) {
+			const message = messages[i];
+			if (message.role === "assistant") {
+				return message as AssistantMessage;
+			}
+		}
+		return undefined;
+	}
+
+	private async _processAgentEvent(event: AgentEvent): Promise<void> {
+		// When a user message starts, check if it's from either queue and remove it BEFORE emitting
+		// This ensures the UI sees the updated queue state
+		if (event.type === "message_start" && event.message.role === "user") {
+			this._overflowRecoveryAttempted = false;
+			const messageText = this._getUserMessageText(event.message);
+			if (messageText) {
+				// Check steering queue first
+				const steeringIndex = this._steeringMessages.indexOf(messageText);
+				if (steeringIndex !== -1) {
+					this._steeringMessages.splice(steeringIndex, 1);
+					this._emitQueueUpdate();
+				} else {
+					// Check follow-up queue
+					const followUpIndex = this._followUpMessages.indexOf(messageText);
+					if (followUpIndex !== -1) {
+						this._followUpMessages.splice(followUpIndex, 1);
+						this._emitQueueUpdate();
+					}
+				}
+			}
+		}
+
+		// Emit to extensions first
+		await this._emitExtensionEvent(event);
+
+		// Notify all listeners
+		this._emit(event);
+
+		// Handle session persistence
+		if (event.type === "message_end") {
+			// Check if this is a custom message from extensions
+			if (event.message.role === "custom") {
+				// Persist as CustomMessageEntry
+				this.sessionManager.appendCustomMessageEntry(
+					event.message.customType,
+					event.message.content,
+					event.message.display,
+					event.message.details,
+				);
+			} else if (
+				event.message.role === "user" ||
+				event.message.role === "assistant" ||
+				event.message.role === "toolResult"
+			) {
+				// Regular LLM message - persist as SessionMessageEntry
+				this.sessionManager.appendMessage(event.message);
+			}
+			// Other message types (bashExecution, compactionSummary, branchSummary) are persisted elsewhere
+
+			// Track assistant message for auto-compaction (checked on agent_end)
+			if (event.message.role === "assistant") {
+				this._lastAssistantMessage = event.message;
+
+				const assistantMsg = event.message as AssistantMessage;
+				if (assistantMsg.stopReason !== "error") {
+					this._overflowRecoveryAttempted = false;
+				}
+
+				// Reset retry counter immediately on successful assistant response
+				// This prevents accumulation across multiple LLM calls within a turn
+				if (assistantMsg.stopReason !== "error" && this._retryAttempt > 0) {
+					this._emit({
+						type: "auto_retry_end",
+						success: true,
+						attempt: this._retryAttempt,
+					});
+					this._retryAttempt = 0;
+				}
+			}
+		}
+
+		// Check auto-retry and auto-compaction after agent completes
+		if (event.type === "agent_end" && this._lastAssistantMessage) {
+			const msg = this._lastAssistantMessage;
+			this._lastAssistantMessage = undefined;
+
+			// Check for retryable errors first (overloaded, rate limit, server errors)
+			if (this._isRetryableError(msg)) {
+				const didRetry = await this._handleRetryableError(msg);
+				if (didRetry) return; // Retry was initiated, don't proceed to compaction
+			}
+
+			this._resolveRetry();
+			await this._checkCompaction(msg);
+		}
+	}
+
+	/** Resolve the pending retry promise */
+	private _resolveRetry(): void {
+		if (this._retryResolve) {
+			this._retryResolve();
+			this._retryResolve = undefined;
+			this._retryPromise = undefined;
+		}
+	}
+
+	/** Extract text content from a message */
+	private _getUserMessageText(message: Message): string {
+		if (message.role !== "user") return "";
+		const content = message.content;
+		if (typeof content === "string") return content;
+		const textBlocks = content.filter((c) => c.type === "text");
+		return textBlocks.map((c) => (c as TextContent).text).join("");
+	}
+
+	/** Find the last assistant message in agent state (including aborted ones) */
+	private _findLastAssistantMessage(): AssistantMessage | undefined {
+		const messages = this.agent.state.messages;
+		for (let i = messages.length - 1; i >= 0; i--) {
+			const msg = messages[i];
+			if (msg.role === "assistant") {
+				return msg as AssistantMessage;
+			}
+		}
+		return undefined;
+	}
+
+	/** Emit extension events based on agent events */
+	private async _emitExtensionEvent(event: AgentEvent): Promise<void> {
+		if (event.type === "agent_start") {
+			this._turnIndex = 0;
+			await this._extensionRunner.emit({ type: "agent_start" });
+		} else if (event.type === "agent_end") {
+			await this._extensionRunner.emit({ type: "agent_end", messages: event.messages });
+		} else if (event.type === "turn_start") {
+			const extensionEvent: TurnStartEvent = {
+				type: "turn_start",
+				turnIndex: this._turnIndex,
+				timestamp: Date.now(),
+			};
+			await this._extensionRunner.emit(extensionEvent);
+		} else if (event.type === "turn_end") {
+			const extensionEvent: TurnEndEvent = {
+				type: "turn_end",
+				turnIndex: this._turnIndex,
+				message: event.message,
+				toolResults: event.toolResults,
+			};
+			await this._extensionRunner.emit(extensionEvent);
+			this._turnIndex++;
+		} else if (event.type === "message_start") {
+			const extensionEvent: MessageStartEvent = {
+				type: "message_start",
+				message: event.message,
+			};
+			await this._extensionRunner.emit(extensionEvent);
+		} else if (event.type === "message_update") {
+			const extensionEvent: MessageUpdateEvent = {
+				type: "message_update",
+				message: event.message,
+				assistantMessageEvent: event.assistantMessageEvent,
+			};
+			await this._extensionRunner.emit(extensionEvent);
+		} else if (event.type === "message_end") {
+			const extensionEvent: MessageEndEvent = {
+				type: "message_end",
+				message: event.message,
+			};
+			await this._extensionRunner.emit(extensionEvent);
+		} else if (event.type === "tool_execution_start") {
+			const extensionEvent: ToolExecutionStartEvent = {
+				type: "tool_execution_start",
+				toolCallId: event.toolCallId,
+				toolName: event.toolName,
+				args: event.args,
+			};
+			await this._extensionRunner.emit(extensionEvent);
+		} else if (event.type === "tool_execution_update") {
+			const extensionEvent: ToolExecutionUpdateEvent = {
+				type: "tool_execution_update",
+				toolCallId: event.toolCallId,
+				toolName: event.toolName,
+				args: event.args,
+				partialResult: event.partialResult,
+			};
+			await this._extensionRunner.emit(extensionEvent);
+		} else if (event.type === "tool_execution_end") {
+			const extensionEvent: ToolExecutionEndEvent = {
+				type: "tool_execution_end",
+				toolCallId: event.toolCallId,
+				toolName: event.toolName,
+				result: event.result,
+				isError: event.isError,
+			};
+			await this._extensionRunner.emit(extensionEvent);
+		}
+	}
+
+	/**
+	 * Subscribe to agent events.
+	 * Session persistence is handled internally (saves messages on message_end).
+	 * Multiple listeners can be added. Returns unsubscribe function for this listener.
+	 */
+	subscribe(listener: AgentSessionEventListener): () => void {
+		this._eventListeners.push(listener);
+
+		// Return unsubscribe function for this specific listener
+		return () => {
+			const index = this._eventListeners.indexOf(listener);
+			if (index !== -1) {
+				this._eventListeners.splice(index, 1);
+			}
+		};
+	}
+
+	/**
+	 * Temporarily disconnect from agent events.
+	 * User listeners are preserved and will receive events again after resubscribe().
+	 * Used internally during operations that need to pause event processing.
+	 */
+	private _disconnectFromAgent(): void {
+		if (this._unsubscribeAgent) {
+			this._unsubscribeAgent();
+			this._unsubscribeAgent = undefined;
+		}
+	}
+
+	/**
+	 * Reconnect to agent events after _disconnectFromAgent().
+	 * Preserves all existing listeners.
+	 */
+	private _reconnectToAgent(): void {
+		if (this._unsubscribeAgent) return; // Already connected
+		this._unsubscribeAgent = this.agent.subscribe(this._handleAgentEvent);
+	}
+
+	/**
+	 * Remove all listeners and disconnect from agent.
+	 * Call this when completely done with the session.
+	 */
+	dispose(): void {
+		this._extensionRunner.invalidate(
+			"This extension instance is stale after session replacement or reload. Use the provided replacement-session context instead.",
+		);
+		this._disconnectFromAgent();
+		this._eventListeners = [];
+	}
+
+	// =========================================================================
+	// Read-only State Access
+	// =========================================================================
+
+	/** Full agent state */
+	get state(): AgentState {
+		return this.agent.state;
+	}
+
+	/** Current model (may be undefined if not yet selected) */
+	get model(): Model<any> | undefined {
+		return this.agent.state.model;
+	}
+
+	/** Current thinking level */
+	get thinkingLevel(): ThinkingLevel {
+		return this.agent.state.thinkingLevel;
+	}
+
+	/** Whether agent is currently streaming a response */
+	get isStreaming(): boolean {
+		return this.agent.state.isStreaming;
+	}
+
+	/** Current effective system prompt (includes any per-turn extension modifications) */
+	get systemPrompt(): string {
+		return this.agent.state.systemPrompt;
+	}
+
+	/** Current retry attempt (0 if not retrying) */
+	get retryAttempt(): number {
+		return this._retryAttempt;
+	}
+
+	/**
+	 * Get the names of currently active tools.
+	 * Returns the names of tools currently set on the agent.
+	 */
+	getActiveToolNames(): string[] {
+		return this.agent.state.tools.map((t) => t.name);
+	}
+
+	/**
+	 * Get all configured tools with name, description, parameter schema, and source metadata.
+	 */
+	getAllTools(): ToolInfo[] {
+		return Array.from(this._toolDefinitions.values()).map(({ definition, sourceInfo }) => ({
+			name: definition.name,
+			description: definition.description,
+			parameters: definition.parameters,
+			sourceInfo,
+		}));
+	}
+
+	getToolDefinition(name: string): ToolDefinition | undefined {
+		return this._toolDefinitions.get(name)?.definition;
+	}
+
+	/**
+	 * Set active tools by name.
+	 * Only tools in the registry can be enabled. Unknown tool names are ignored.
+	 * Also rebuilds the system prompt to reflect the new tool set.
+	 * Changes take effect on the next agent turn.
+	 */
+	setActiveToolsByName(toolNames: string[]): void {
+		const tools: AgentTool[] = [];
+		const validToolNames: string[] = [];
+		for (const name of toolNames) {
+			const tool = this._toolRegistry.get(name);
+			if (tool) {
+				tools.push(tool);
+				validToolNames.push(name);
+			}
+		}
+		this.agent.state.tools = tools;
+
+		// Rebuild base system prompt with new tool set
+		this._baseSystemPrompt = this._rebuildSystemPrompt(validToolNames);
+		this.agent.state.systemPrompt = this._baseSystemPrompt;
+	}
+
+	/** Whether compaction or branch summarization is currently running */
+	get isCompacting(): boolean {
+		return (
+			this._autoCompactionAbortController !== undefined ||
+			this._compactionAbortController !== undefined ||
+			this._branchSummaryAbortController !== undefined
+		);
+	}
+
+	/** All messages including custom types like BashExecutionMessage */
+	get messages(): AgentMessage[] {
+		return this.agent.state.messages;
+	}
+
+	/** Current steering mode */
+	get steeringMode(): "all" | "one-at-a-time" {
+		return this.agent.steeringMode;
+	}
+
+	/** Current follow-up mode */
+	get followUpMode(): "all" | "one-at-a-time" {
+		return this.agent.followUpMode;
+	}
+
+	/** Current session file path, or undefined if sessions are disabled */
+	get sessionFile(): string | undefined {
+		return this.sessionManager.getSessionFile();
+	}
+
+	/** Current session ID */
+	get sessionId(): string {
+		return this.sessionManager.getSessionId();
+	}
+
+	/** Current session display name, if set */
+	get sessionName(): string | undefined {
+		return this.sessionManager.getSessionName();
+	}
+
+	/** Scoped models for cycling (from --models flag) */
+	get scopedModels(): ReadonlyArray<{ model: Model<any>; thinkingLevel?: ThinkingLevel }> {
+		return this._scopedModels;
+	}
+
+	/** Update scoped models for cycling */
+	setScopedModels(scopedModels: Array<{ model: Model<any>; thinkingLevel?: ThinkingLevel }>): void {
+		this._scopedModels = scopedModels;
+	}
+
+	/** File-based prompt templates */
+	get promptTemplates(): ReadonlyArray<PromptTemplate> {
+		return this._resourceLoader.getPrompts().prompts;
+	}
+
+	private _normalizePromptSnippet(text: string | undefined): string | undefined {
+		if (!text) return undefined;
+		const oneLine = text
+			.replace(/[\r\n]+/g, " ")
+			.replace(/\s+/g, " ")
+			.trim();
+		return oneLine.length > 0 ? oneLine : undefined;
+	}
+
+	private _normalizePromptGuidelines(guidelines: string[] | undefined): string[] {
+		if (!guidelines || guidelines.length === 0) {
+			return [];
+		}
+
+		const unique = new Set<string>();
+		for (const guideline of guidelines) {
+			const normalized = guideline.trim();
+			if (normalized.length > 0) {
+				unique.add(normalized);
+			}
+		}
+		return Array.from(unique);
+	}
+
+	private _rebuildSystemPrompt(toolNames: string[]): string {
+		const validToolNames = toolNames.filter((name) => this._toolRegistry.has(name));
+		const toolSnippets: Record<string, string> = {};
+		const promptGuidelines: string[] = [];
+		for (const name of validToolNames) {
+			const snippet = this._toolPromptSnippets.get(name);
+			if (snippet) {
+				toolSnippets[name] = snippet;
+			}
+
+			const toolGuidelines = this._toolPromptGuidelines.get(name);
+			if (toolGuidelines) {
+				promptGuidelines.push(...toolGuidelines);
+			}
+		}
+
+		const loaderSystemPrompt = this._resourceLoader.getSystemPrompt();
+		const loaderAppendSystemPrompt = this._resourceLoader.getAppendSystemPrompt();
+		const appendSystemPrompt =
+			loaderAppendSystemPrompt.length > 0 ? loaderAppendSystemPrompt.join("\n\n") : undefined;
+		const loadedSkills = this._resourceLoader.getSkills().skills;
+		const loadedContextFiles = this._resourceLoader.getAgentsFiles().agentsFiles;
+
+		this._baseSystemPromptOptions = {
+			cwd: this._cwd,
+			skills: loadedSkills,
+			contextFiles: loadedContextFiles,
+			customPrompt: loaderSystemPrompt,
+			appendSystemPrompt,
+			selectedTools: validToolNames,
+			toolSnippets,
+			promptGuidelines,
+		};
+		return buildSystemPrompt(this._baseSystemPromptOptions);
+	}
+
+	// =========================================================================
+	// Prompting
+	// =========================================================================
+
+	/**
+	 * Send a prompt to the agent.
+	 * - Handles extension commands (registered via pi.registerCommand) immediately, even during streaming
+	 * - Expands file-based prompt templates by default
+	 * - During streaming, queues via steer() or followUp() based on streamingBehavior option
+	 * - Validates model and API key before sending (when not streaming)
+	 * @throws Error if streaming and no streamingBehavior specified
+	 * @throws Error if no model selected or no API key available (when not streaming)
+	 */
+	async prompt(text: string, options?: PromptOptions): Promise<void> {
+		const expandPromptTemplates = options?.expandPromptTemplates ?? true;
+		const preflightResult = options?.preflightResult;
+		let messages: AgentMessage[] | undefined;
+
+		try {
+			// Handle extension commands first (execute immediately, even during streaming)
+			// Extension commands manage their own LLM interaction via pi.sendMessage()
+			if (expandPromptTemplates && text.startsWith("/")) {
+				const handled = await this._tryExecuteExtensionCommand(text);
+				if (handled) {
+					// Extension command executed, no prompt to send
+					preflightResult?.(true);
+					return;
+				}
+			}
+
+			// Emit input event for extension interception (before skill/template expansion)
+			let currentText = text;
+			let currentImages = options?.images;
+			if (this._extensionRunner.hasHandlers("input")) {
+				const inputResult = await this._extensionRunner.emitInput(
+					currentText,
+					currentImages,
+					options?.source ?? "interactive",
+				);
+				if (inputResult.action === "handled") {
+					preflightResult?.(true);
+					return;
+				}
+				if (inputResult.action === "transform") {
+					currentText = inputResult.text;
+					currentImages = inputResult.images ?? currentImages;
+				}
+			}
+
+			// Expand skill commands (/skill:name args) and prompt templates (/template args)
+			let expandedText = currentText;
+			if (expandPromptTemplates) {
+				expandedText = this._expandSkillCommand(expandedText);
+				expandedText = expandPromptTemplate(expandedText, [...this.promptTemplates]);
+			}
+
+			// If streaming, queue via steer() or followUp() based on option
+			if (this.isStreaming) {
+				if (!options?.streamingBehavior) {
+					throw new Error(
+						"Agent is already processing. Specify streamingBehavior ('steer' or 'followUp') to queue the message.",
+					);
+				}
+				if (options.streamingBehavior === "followUp") {
+					await this._queueFollowUp(expandedText, currentImages);
+				} else {
+					await this._queueSteer(expandedText, currentImages);
+				}
+				preflightResult?.(true);
+				return;
+			}
+
+			// Flush any pending bash messages before the new prompt
+			this._flushPendingBashMessages();
+
+			// Validate model
+			if (!this.model) {
+				throw new Error(
+					"No model selected.\n\n" +
+						`Use /login or set an API key environment variable. See ${join(getDocsPath(), "providers.md")}\n\n` +
+						"Then use /model to select a model.",
+				);
+			}
+
+			if (!this._modelRegistry.hasConfiguredAuth(this.model)) {
+				const isOAuth = this._modelRegistry.isUsingOAuth(this.model);
+				if (isOAuth) {
+					throw new Error(
+						`Authentication failed for "${this.model.provider}". ` +
+							`Credentials may have expired or network is unavailable. ` +
+							`Run '/login ${this.model.provider}' to re-authenticate.`,
+					);
+				}
+				throw new Error(
+					`No API key found for ${this.model.provider}.\n\n` +
+						`Use /login or set an API key environment variable. See ${join(getDocsPath(), "providers.md")}`,
+				);
+			}
+
+			// Check if we need to compact before sending (catches aborted responses)
+			const lastAssistant = this._findLastAssistantMessage();
+			if (lastAssistant) {
+				await this._checkCompaction(lastAssistant, false);
+			}
+
+			// Build messages array (custom message if any, then user message)
+			messages = [];
+
+			// Add user message
+			const userContent: (TextContent | ImageContent)[] = [{ type: "text", text: expandedText }];
+			if (currentImages) {
+				userContent.push(...currentImages);
+			}
+			messages.push({
+				role: "user",
+				content: userContent,
+				timestamp: Date.now(),
+			});
+
+			// Inject any pending "nextTurn" messages as context alongside the user message
+			for (const msg of this._pendingNextTurnMessages) {
+				messages.push(msg);
+			}
+			this._pendingNextTurnMessages = [];
+
+			// Emit before_agent_start extension event
+			const result = await this._extensionRunner.emitBeforeAgentStart(
+				expandedText,
+				currentImages,
+				this._baseSystemPrompt,
+				this._baseSystemPromptOptions,
+			);
+			// Add all custom messages from extensions
+			if (result?.messages) {
+				for (const msg of result.messages) {
+					messages.push({
+						role: "custom",
+						customType: msg.customType,
+						content: msg.content,
+						display: msg.display,
+						details: msg.details,
+						timestamp: Date.now(),
+					});
+				}
+			}
+			// Apply extension-modified system prompt, or reset to base
+			if (result?.systemPrompt) {
+				this.agent.state.systemPrompt = result.systemPrompt;
+			} else {
+				// Ensure we're using the base prompt (in case previous turn had modifications)
+				this.agent.state.systemPrompt = this._baseSystemPrompt;
+			}
+		} catch (error) {
+			preflightResult?.(false);
+			throw error;
+		}
+
+		if (!messages) {
+			return;
+		}
+
+		preflightResult?.(true);
+		await this.agent.prompt(messages);
+		await this.waitForRetry();
+	}
+
+	/**
+	 * Try to execute an extension command. Returns true if command was found and executed.
+	 */
+	private async _tryExecuteExtensionCommand(text: string): Promise<boolean> {
+		// Parse command name and args
+		const spaceIndex = text.indexOf(" ");
+		const commandName = spaceIndex === -1 ? text.slice(1) : text.slice(1, spaceIndex);
+		const args = spaceIndex === -1 ? "" : text.slice(spaceIndex + 1);
+
+		const command = this._extensionRunner.getCommand(commandName);
+		if (!command) return false;
+
+		// Get command context from extension runner (includes session control methods)
+		const ctx = this._extensionRunner.createCommandContext();
+
+		try {
+			await command.handler(args, ctx);
+			return true;
+		} catch (err) {
+			// Emit error via extension runner
+			this._extensionRunner.emitError({
+				extensionPath: `command:${commandName}`,
+				event: "command",
+				error: err instanceof Error ? err.message : String(err),
+			});
+			return true;
+		}
+	}
+
+	/**
+	 * Expand skill commands (/skill:name args) to their full content.
+	 * Returns the expanded text, or the original text if not a skill command or skill not found.
+	 * Emits errors via extension runner if file read fails.
+	 */
+	private _expandSkillCommand(text: string): string {
+		if (!text.startsWith("/skill:")) return text;
+
+		const spaceIndex = text.indexOf(" ");
+		const skillName = spaceIndex === -1 ? text.slice(7) : text.slice(7, spaceIndex);
+		const args = spaceIndex === -1 ? "" : text.slice(spaceIndex + 1).trim();
+
+		const skill = this.resourceLoader.getSkills().skills.find((s) => s.name === skillName);
+		if (!skill) return text; // Unknown skill, pass through
+
+		try {
+			const content = readFileSync(skill.filePath, "utf-8");
+			const body = stripFrontmatter(content).trim();
+			const skillBlock = `<skill name="${skill.name}" location="${skill.filePath}">\nReferences are relative to ${skill.baseDir}.\n\n${body}\n</skill>`;
+			return args ? `${skillBlock}\n\n${args}` : skillBlock;
+		} catch (err) {
+			// Emit error like extension commands do
+			this._extensionRunner.emitError({
+				extensionPath: skill.filePath,
+				event: "skill_expansion",
+				error: err instanceof Error ? err.message : String(err),
+			});
+			return text; // Return original on error
+		}
+	}
+
+	/**
+	 * Queue a steering message while the agent is running.
+	 * Delivered after the current assistant turn finishes executing its tool calls,
+	 * before the next LLM call.
+	 * Expands skill commands and prompt templates. Errors on extension commands.
+	 * @param images Optional image attachments to include with the message
+	 * @throws Error if text is an extension command
+	 */
+	async steer(text: string, images?: ImageContent[]): Promise<void> {
+		// Check for extension commands (cannot be queued)
+		if (text.startsWith("/")) {
+			this._throwIfExtensionCommand(text);
+		}
+
+		// Expand skill commands and prompt templates
+		let expandedText = this._expandSkillCommand(text);
+		expandedText = expandPromptTemplate(expandedText, [...this.promptTemplates]);
+
+		await this._queueSteer(expandedText, images);
+	}
+
+	/**
+	 * Queue a follow-up message to be processed after the agent finishes.
+	 * Delivered only when agent has no more tool calls or steering messages.
+	 * Expands skill commands and prompt templates. Errors on extension commands.
+	 * @param images Optional image attachments to include with the message
+	 * @throws Error if text is an extension command
+	 */
+	async followUp(text: string, images?: ImageContent[]): Promise<void> {
+		// Check for extension commands (cannot be queued)
+		if (text.startsWith("/")) {
+			this._throwIfExtensionCommand(text);
+		}
+
+		// Expand skill commands and prompt templates
+		let expandedText = this._expandSkillCommand(text);
+		expandedText = expandPromptTemplate(expandedText, [...this.promptTemplates]);
+
+		await this._queueFollowUp(expandedText, images);
+	}
+
+	/**
+	 * Internal: Queue a steering message (already expanded, no extension command check).
+	 */
+	private async _queueSteer(text: string, images?: ImageContent[]): Promise<void> {
+		this._steeringMessages.push(text);
+		this._emitQueueUpdate();
+		const content: (TextContent | ImageContent)[] = [{ type: "text", text }];
+		if (images) {
+			content.push(...images);
+		}
+		this.agent.steer({
+			role: "user",
+			content,
+			timestamp: Date.now(),
+		});
+	}
+
+	/**
+	 * Internal: Queue a follow-up message (already expanded, no extension command check).
+	 */
+	private async _queueFollowUp(text: string, images?: ImageContent[]): Promise<void> {
+		this._followUpMessages.push(text);
+		this._emitQueueUpdate();
+		const content: (TextContent | ImageContent)[] = [{ type: "text", text }];
+		if (images) {
+			content.push(...images);
+		}
+		this.agent.followUp({
+			role: "user",
+			content,
+			timestamp: Date.now(),
+		});
+	}
+
+	/**
+	 * Throw an error if the text is an extension command.
+	 */
+	private _throwIfExtensionCommand(text: string): void {
+		const spaceIndex = text.indexOf(" ");
+		const commandName = spaceIndex === -1 ? text.slice(1) : text.slice(1, spaceIndex);
+		const command = this._extensionRunner.getCommand(commandName);
+
+		if (command) {
+			throw new Error(
+				`Extension command "/${commandName}" cannot be queued. Use prompt() or execute the command when not streaming.`,
+			);
+		}
+	}
+
+	/**
+	 * Send a custom message to the session. Creates a CustomMessageEntry.
+	 *
+	 * Handles three cases:
+	 * - Streaming: queues message, processed when loop pulls from queue
+	 * - Not streaming + triggerTurn: appends to state/session, starts new turn
+	 * - Not streaming + no trigger: appends to state/session, no turn
+	 *
+	 * @param message Custom message with customType, content, display, details
+	 * @param options.triggerTurn If true and not streaming, triggers a new LLM turn
+	 * @param options.deliverAs Delivery mode: "steer", "followUp", or "nextTurn"
+	 */
+	async sendCustomMessage<T = unknown>(
+		message: Pick<CustomMessage<T>, "customType" | "content" | "display" | "details">,
+		options?: { triggerTurn?: boolean; deliverAs?: "steer" | "followUp" | "nextTurn" },
+	): Promise<void> {
+		const appMessage = {
+			role: "custom" as const,
+			customType: message.customType,
+			content: message.content,
+			display: message.display,
+			details: message.details,
+			timestamp: Date.now(),
+		} satisfies CustomMessage<T>;
+		if (options?.deliverAs === "nextTurn") {
+			this._pendingNextTurnMessages.push(appMessage);
+		} else if (this.isStreaming) {
+			if (options?.deliverAs === "followUp") {
+				this.agent.followUp(appMessage);
+			} else {
+				this.agent.steer(appMessage);
+			}
+		} else if (options?.triggerTurn) {
+			await this.agent.prompt(appMessage);
+		} else {
+			this.agent.state.messages.push(appMessage);
+			this.sessionManager.appendCustomMessageEntry(
+				message.customType,
+				message.content,
+				message.display,
+				message.details,
+			);
+			this._emit({ type: "message_start", message: appMessage });
+			this._emit({ type: "message_end", message: appMessage });
+		}
+	}
+
+	/**
+	 * Send a user message to the agent. Always triggers a turn.
+	 * When the agent is streaming, use deliverAs to specify how to queue the message.
+	 *
+	 * @param content User message content (string or content array)
+	 * @param options.deliverAs Delivery mode when streaming: "steer" or "followUp"
+	 */
+	async sendUserMessage(
+		content: string | (TextContent | ImageContent)[],
+		options?: { deliverAs?: "steer" | "followUp" },
+	): Promise<void> {
+		// Normalize content to text string + optional images
+		let text: string;
+		let images: ImageContent[] | undefined;
+
+		if (typeof content === "string") {
+			text = content;
+		} else {
+			const textParts: string[] = [];
+			images = [];
+			for (const part of content) {
+				if (part.type === "text") {
+					textParts.push(part.text);
+				} else {
+					images.push(part);
+				}
+			}
+			text = textParts.join("\n");
+			if (images.length === 0) images = undefined;
+		}
+
+		// Use prompt() with expandPromptTemplates: false to skip command handling and template expansion
+		await this.prompt(text, {
+			expandPromptTemplates: false,
+			streamingBehavior: options?.deliverAs,
+			images,
+			source: "extension",
+		});
+	}
+
+	/**
+	 * Clear all queued messages and return them.
+	 * Useful for restoring to editor when user aborts.
+	 * @returns Object with steering and followUp arrays
+	 */
+	clearQueue(): { steering: string[]; followUp: string[] } {
+		const steering = [...this._steeringMessages];
+		const followUp = [...this._followUpMessages];
+		this._steeringMessages = [];
+		this._followUpMessages = [];
+		this.agent.clearAllQueues();
+		this._emitQueueUpdate();
+		return { steering, followUp };
+	}
+
+	/** Number of pending messages (includes both steering and follow-up) */
+	get pendingMessageCount(): number {
+		return this._steeringMessages.length + this._followUpMessages.length;
+	}
+
+	/** Get pending steering messages (read-only) */
+	getSteeringMessages(): readonly string[] {
+		return this._steeringMessages;
+	}
+
+	/** Get pending follow-up messages (read-only) */
+	getFollowUpMessages(): readonly string[] {
+		return this._followUpMessages;
+	}
+
+	get resourceLoader(): ResourceLoader {
+		return this._resourceLoader;
+	}
+
+	/**
+	 * Abort current operation and wait for agent to become idle.
+	 */
+	async abort(): Promise<void> {
+		this.abortRetry();
+		this.agent.abort();
+		await this.agent.waitForIdle();
+	}
+
+	// =========================================================================
+	// Model Management
+	// =========================================================================
+
+	private async _emitModelSelect(
+		nextModel: Model<any>,
+		previousModel: Model<any> | undefined,
+		source: "set" | "cycle" | "restore",
+	): Promise<void> {
+		if (modelsAreEqual(previousModel, nextModel)) return;
+		await this._extensionRunner.emit({
+			type: "model_select",
+			model: nextModel,
+			previousModel,
+			source,
+		});
+	}
+
+	/**
+	 * Set model directly.
+	 * Validates that auth is configured, saves to session and settings.
+	 * @throws Error if no auth is configured for the model
+	 */
+	async setModel(model: Model<any>): Promise<void> {
+		if (!this._modelRegistry.hasConfiguredAuth(model)) {
+			throw new Error(`No API key for ${model.provider}/${model.id}`);
+		}
+
+		const previousModel = this.model;
+		const thinkingLevel = this._getThinkingLevelForModelSwitch();
+		this.agent.state.model = model;
+		this.sessionManager.appendModelChange(model.provider, model.id);
+		this.settingsManager.setDefaultModelAndProvider(model.provider, model.id);
+
+		// Re-clamp thinking level for new model's capabilities
+		this.setThinkingLevel(thinkingLevel);
+
+		await this._emitModelSelect(model, previousModel, "set");
+	}
+
+	/**
+	 * Cycle to next/previous model.
+	 * Uses scoped models (from --models flag) if available, otherwise all available models.
+	 * @param direction - "forward" (default) or "backward"
+	 * @returns The new model info, or undefined if only one model available
+	 */
+	async cycleModel(direction: "forward" | "backward" = "forward"): Promise<ModelCycleResult | undefined> {
+		if (this._scopedModels.length > 0) {
+			return this._cycleScopedModel(direction);
+		}
+		return this._cycleAvailableModel(direction);
+	}
+
+	private async _cycleScopedModel(direction: "forward" | "backward"): Promise<ModelCycleResult | undefined> {
+		const scopedModels = this._scopedModels.filter((scoped) => this._modelRegistry.hasConfiguredAuth(scoped.model));
+		if (scopedModels.length <= 1) return undefined;
+
+		const currentModel = this.model;
+		let currentIndex = scopedModels.findIndex((sm) => modelsAreEqual(sm.model, currentModel));
+
+		if (currentIndex === -1) currentIndex = 0;
+		const len = scopedModels.length;
+		const nextIndex = direction === "forward" ? (currentIndex + 1) % len : (currentIndex - 1 + len) % len;
+		const next = scopedModels[nextIndex];
+		const thinkingLevel = this._getThinkingLevelForModelSwitch(next.thinkingLevel);
+
+		// Apply model
+		this.agent.state.model = next.model;
+		this.sessionManager.appendModelChange(next.model.provider, next.model.id);
+		this.settingsManager.setDefaultModelAndProvider(next.model.provider, next.model.id);
+
+		// Apply thinking level.
+		// - Explicit scoped model thinking level overrides current session level
+		// - Undefined scoped model thinking level inherits the current session preference
+		// setThinkingLevel clamps to model capabilities.
+		this.setThinkingLevel(thinkingLevel);
+
+		await this._emitModelSelect(next.model, currentModel, "cycle");
+
+		return { model: next.model, thinkingLevel: this.thinkingLevel, isScoped: true };
+	}
+
+	private async _cycleAvailableModel(direction: "forward" | "backward"): Promise<ModelCycleResult | undefined> {
+		const availableModels = await this._modelRegistry.getAvailable();
+		if (availableModels.length <= 1) return undefined;
+
+		const currentModel = this.model;
+		let currentIndex = availableModels.findIndex((m) => modelsAreEqual(m, currentModel));
+
+		if (currentIndex === -1) currentIndex = 0;
+		const len = availableModels.length;
+		const nextIndex = direction === "forward" ? (currentIndex + 1) % len : (currentIndex - 1 + len) % len;
+		const nextModel = availableModels[nextIndex];
+
+		const thinkingLevel = this._getThinkingLevelForModelSwitch();
+		this.agent.state.model = nextModel;
+		this.sessionManager.appendModelChange(nextModel.provider, nextModel.id);
+		this.settingsManager.setDefaultModelAndProvider(nextModel.provider, nextModel.id);
+
+		// Re-clamp thinking level for new model's capabilities
+		this.setThinkingLevel(thinkingLevel);
+
+		await this._emitModelSelect(nextModel, currentModel, "cycle");
+
+		return { model: nextModel, thinkingLevel: this.thinkingLevel, isScoped: false };
+	}
+
+	// =========================================================================
+	// Thinking Level Management
+	// =========================================================================
+
+	/**
+	 * Set thinking level.
+	 * Clamps to model capabilities based on available thinking levels.
+	 * Saves to session and settings only if the level actually changes.
+	 */
+	setThinkingLevel(level: ThinkingLevel): void {
+		const availableLevels = this.getAvailableThinkingLevels();
+		const effectiveLevel = availableLevels.includes(level) ? level : this._clampThinkingLevel(level, availableLevels);
+
+		// Only persist if actually changing
+		const isChanging = effectiveLevel !== this.agent.state.thinkingLevel;
+
+		this.agent.state.thinkingLevel = effectiveLevel;
+
+		if (isChanging) {
+			this.sessionManager.appendThinkingLevelChange(effectiveLevel);
+			if (this.supportsThinking() || effectiveLevel !== "off") {
+				this.settingsManager.setDefaultThinkingLevel(effectiveLevel);
+			}
+		}
+	}
+
+	/**
+	 * Cycle to next thinking level.
+	 * @returns New level, or undefined if model doesn't support thinking
+	 */
+	cycleThinkingLevel(): ThinkingLevel | undefined {
+		if (!this.supportsThinking()) return undefined;
+
+		const levels = this.getAvailableThinkingLevels();
+		const currentIndex = levels.indexOf(this.thinkingLevel);
+		const nextIndex = (currentIndex + 1) % levels.length;
+		const nextLevel = levels[nextIndex];
+
+		this.setThinkingLevel(nextLevel);
+		return nextLevel;
+	}
+
+	/**
+	 * Get available thinking levels for current model.
+	 * The provider will clamp to what the specific model supports internally.
+	 */
+	getAvailableThinkingLevels(): ThinkingLevel[] {
+		if (!this.supportsThinking()) return ["off"];
+		return this.supportsXhighThinking() ? THINKING_LEVELS_WITH_XHIGH : THINKING_LEVELS;
+	}
+
+	/**
+	 * Check if current model supports xhigh thinking level.
+	 */
+	supportsXhighThinking(): boolean {
+		return this.model ? supportsXhigh(this.model) : false;
+	}
+
+	/**
+	 * Check if current model supports thinking/reasoning.
+	 */
+	supportsThinking(): boolean {
+		return !!this.model?.reasoning;
+	}
+
+	private _getThinkingLevelForModelSwitch(explicitLevel?: ThinkingLevel): ThinkingLevel {
+		if (explicitLevel !== undefined) {
+			return explicitLevel;
+		}
+		if (!this.supportsThinking()) {
+			return this.settingsManager.getDefaultThinkingLevel() ?? DEFAULT_THINKING_LEVEL;
+		}
+		return this.thinkingLevel;
+	}
+
+	private _clampThinkingLevel(level: ThinkingLevel, availableLevels: ThinkingLevel[]): ThinkingLevel {
+		const ordered = THINKING_LEVELS_WITH_XHIGH;
+		const available = new Set(availableLevels);
+		const requestedIndex = ordered.indexOf(level);
+		if (requestedIndex === -1) {
+			return availableLevels[0] ?? "off";
+		}
+		for (let i = requestedIndex; i < ordered.length; i++) {
+			const candidate = ordered[i];
+			if (available.has(candidate)) return candidate;
+		}
+		for (let i = requestedIndex - 1; i >= 0; i--) {
+			const candidate = ordered[i];
+			if (available.has(candidate)) return candidate;
+		}
+		return availableLevels[0] ?? "off";
+	}
+
+	// =========================================================================
+	// Queue Mode Management
+	// =========================================================================
+
+	/**
+	 * Set steering message mode.
+	 * Saves to settings.
+	 */
+	setSteeringMode(mode: "all" | "one-at-a-time"): void {
+		this.agent.steeringMode = mode;
+		this.settingsManager.setSteeringMode(mode);
+	}
+
+	/**
+	 * Set follow-up message mode.
+	 * Saves to settings.
+	 */
+	setFollowUpMode(mode: "all" | "one-at-a-time"): void {
+		this.agent.followUpMode = mode;
+		this.settingsManager.setFollowUpMode(mode);
+	}
+
+	// =========================================================================
+	// Compaction
+	// =========================================================================
+
+	/**
+	 * Manually compact the session context.
+	 * Aborts current agent operation first.
+	 * @param customInstructions Optional instructions for the compaction summary
+	 */
+	async compact(customInstructions?: string): Promise<CompactionResult> {
+		this._disconnectFromAgent();
+		await this.abort();
+		this._compactionAbortController = new AbortController();
+		this._emit({ type: "compaction_start", reason: "manual" });
+
+		try {
+			if (!this.model) {
+				throw new Error("No model selected");
+			}
+
+			const { apiKey, headers } = await this._getRequiredRequestAuth(this.model);
+
+			const pathEntries = this.sessionManager.getBranch();
+			const settings = this.settingsManager.getCompactionSettings();
+
+			const preparation = prepareCompaction(pathEntries, settings);
+			if (!preparation) {
+				// Check why we can't compact
+				const lastEntry = pathEntries[pathEntries.length - 1];
+				if (lastEntry?.type === "compaction") {
+					throw new Error("Already compacted");
+				}
+				throw new Error("Nothing to compact (session too small)");
+			}
+
+			let extensionCompaction: CompactionResult | undefined;
+			let fromExtension = false;
+
+			if (this._extensionRunner.hasHandlers("session_before_compact")) {
+				const result = (await this._extensionRunner.emit({
+					type: "session_before_compact",
+					preparation,
+					branchEntries: pathEntries,
+					customInstructions,
+					signal: this._compactionAbortController.signal,
+				})) as SessionBeforeCompactResult | undefined;
+
+				if (result?.cancel) {
+					throw new Error("Compaction cancelled");
+				}
+
+				if (result?.compaction) {
+					extensionCompaction = result.compaction;
+					fromExtension = true;
+				}
+			}
+
+			let summary: string;
+			let firstKeptEntryId: string;
+			let tokensBefore: number;
+			let details: unknown;
+
+			if (extensionCompaction) {
+				// Extension provided compaction content
+				summary = extensionCompaction.summary;
+				firstKeptEntryId = extensionCompaction.firstKeptEntryId;
+				tokensBefore = extensionCompaction.tokensBefore;
+				details = extensionCompaction.details;
+			} else {
+				// Generate compaction result
+				const result = await compact(
+					preparation,
+					this.model,
+					apiKey,
+					headers,
+					customInstructions,
+					this._compactionAbortController.signal,
+					this.thinkingLevel,
+				);
+				summary = result.summary;
+				firstKeptEntryId = result.firstKeptEntryId;
+				tokensBefore = result.tokensBefore;
+				details = result.details;
+			}
+
+			if (this._compactionAbortController.signal.aborted) {
+				throw new Error("Compaction cancelled");
+			}
+
+			this.sessionManager.appendCompaction(summary, firstKeptEntryId, tokensBefore, details, fromExtension);
+			const newEntries = this.sessionManager.getEntries();
+			const sessionContext = this.sessionManager.buildSessionContext();
+			this.agent.state.messages = sessionContext.messages;
+
+			// Get the saved compaction entry for the extension event
+			const savedCompactionEntry = newEntries.find((e) => e.type === "compaction" && e.summary === summary) as
+				| CompactionEntry
+				| undefined;
+
+			if (this._extensionRunner && savedCompactionEntry) {
+				await this._extensionRunner.emit({
+					type: "session_compact",
+					compactionEntry: savedCompactionEntry,
+					fromExtension,
+				});
+			}
+
+			const compactionResult = {
+				summary,
+				firstKeptEntryId,
+				tokensBefore,
+				details,
+			};
+			this._emit({
+				type: "compaction_end",
+				reason: "manual",
+				result: compactionResult,
+				aborted: false,
+				willRetry: false,
+			});
+			return compactionResult;
+		} catch (error) {
+			const message = error instanceof Error ? error.message : String(error);
+			const aborted = message === "Compaction cancelled" || (error instanceof Error && error.name === "AbortError");
+			this._emit({
+				type: "compaction_end",
+				reason: "manual",
+				result: undefined,
+				aborted,
+				willRetry: false,
+				errorMessage: aborted ? undefined : `Compaction failed: ${message}`,
+			});
+			throw error;
+		} finally {
+			this._compactionAbortController = undefined;
+			this._reconnectToAgent();
+		}
+	}
+
+	/**
+	 * Cancel in-progress compaction (manual or auto).
+	 */
+	abortCompaction(): void {
+		this._compactionAbortController?.abort();
+		this._autoCompactionAbortController?.abort();
+	}
+
+	/**
+	 * Cancel in-progress branch summarization.
+	 */
+	abortBranchSummary(): void {
+		this._branchSummaryAbortController?.abort();
+	}
+
+	/**
+	 * Check if compaction is needed and run it.
+	 * Called after agent_end and before prompt submission.
+	 *
+	 * Two cases:
+	 * 1. Overflow: LLM returned context overflow error, remove error message from agent state, compact, auto-retry
+	 * 2. Threshold: Context over threshold, compact, NO auto-retry (user continues manually)
+	 *
+	 * @param assistantMessage The assistant message to check
+	 * @param skipAbortedCheck If false, include aborted messages (for pre-prompt check). Default: true
+	 */
+	private async _checkCompaction(assistantMessage: AssistantMessage, skipAbortedCheck = true): Promise<void> {
+		const settings = this.settingsManager.getCompactionSettings();
+		if (!settings.enabled) return;
+
+		// Skip if message was aborted (user cancelled) - unless skipAbortedCheck is false
+		if (skipAbortedCheck && assistantMessage.stopReason === "aborted") return;
+
+		const contextWindow = this.model?.contextWindow ?? 0;
+
+		// Skip overflow check if the message came from a different model.
+		// This handles the case where user switched from a smaller-context model (e.g. opus)
+		// to a larger-context model (e.g. codex) - the overflow error from the old model
+		// shouldn't trigger compaction for the new model.
+		const sameModel =
+			this.model && assistantMessage.provider === this.model.provider && assistantMessage.model === this.model.id;
+
+		// Skip compaction checks if this assistant message is older than the latest
+		// compaction boundary. This prevents a stale pre-compaction usage/error
+		// from retriggering compaction on the first prompt after compaction.
+		const compactionEntry = getLatestCompactionEntry(this.sessionManager.getBranch());
+		const assistantIsFromBeforeCompaction =
+			compactionEntry !== null && assistantMessage.timestamp <= new Date(compactionEntry.timestamp).getTime();
+		if (assistantIsFromBeforeCompaction) {
+			return;
+		}
+
+		// Case 1: Overflow - LLM returned context overflow error
+		if (sameModel && isContextOverflow(assistantMessage, contextWindow)) {
+			if (this._overflowRecoveryAttempted) {
+				this._emit({
+					type: "compaction_end",
+					reason: "overflow",
+					result: undefined,
+					aborted: false,
+					willRetry: false,
+					errorMessage:
+						"Context overflow recovery failed after one compact-and-retry attempt. Try reducing context or switching to a larger-context model.",
+				});
+				return;
+			}
+
+			this._overflowRecoveryAttempted = true;
+			// Remove the error message from agent state (it IS saved to session for history,
+			// but we don't want it in context for the retry)
+			const messages = this.agent.state.messages;
+			if (messages.length > 0 && messages[messages.length - 1].role === "assistant") {
+				this.agent.state.messages = messages.slice(0, -1);
+			}
+			await this._runAutoCompaction("overflow", true);
+			return;
+		}
+
+		// Case 2: Threshold - context is getting large
+		// For error messages (no usage data), estimate from last successful response.
+		// This ensures sessions that hit persistent API errors (e.g. 529) can still compact.
+		let contextTokens: number;
+		if (assistantMessage.stopReason === "error") {
+			const messages = this.agent.state.messages;
+			const estimate = estimateContextTokens(messages);
+			if (estimate.lastUsageIndex === null) return; // No usage data at all
+			// Verify the usage source is post-compaction. Kept pre-compaction messages
+			// have stale usage reflecting the old (larger) context and would falsely
+			// trigger compaction right after one just finished.
+			const usageMsg = messages[estimate.lastUsageIndex];
+			if (
+				compactionEntry &&
+				usageMsg.role === "assistant" &&
+				(usageMsg as AssistantMessage).timestamp <= new Date(compactionEntry.timestamp).getTime()
+			) {
+				return;
+			}
+			contextTokens = estimate.tokens;
+		} else {
+			contextTokens = calculateContextTokens(assistantMessage.usage);
+		}
+		if (shouldCompact(contextTokens, contextWindow, settings)) {
+			await this._runAutoCompaction("threshold", false);
+		}
+	}
+
+	/**
+	 * Internal: Run auto-compaction with events.
+	 */
+	private async _runAutoCompaction(reason: "overflow" | "threshold", willRetry: boolean): Promise<void> {
+		const settings = this.settingsManager.getCompactionSettings();
+
+		this._emit({ type: "compaction_start", reason });
+		this._autoCompactionAbortController = new AbortController();
+
+		try {
+			if (!this.model) {
+				this._emit({
+					type: "compaction_end",
+					reason,
+					result: undefined,
+					aborted: false,
+					willRetry: false,
+				});
+				return;
+			}
+
+			const authResult = await this._modelRegistry.getApiKeyAndHeaders(this.model);
+			if (!authResult.ok || !authResult.apiKey) {
+				this._emit({
+					type: "compaction_end",
+					reason,
+					result: undefined,
+					aborted: false,
+					willRetry: false,
+				});
+				return;
+			}
+			const { apiKey, headers } = authResult;
+
+			const pathEntries = this.sessionManager.getBranch();
+
+			const preparation = prepareCompaction(pathEntries, settings);
+			if (!preparation) {
+				this._emit({
+					type: "compaction_end",
+					reason,
+					result: undefined,
+					aborted: false,
+					willRetry: false,
+				});
+				return;
+			}
+
+			let extensionCompaction: CompactionResult | undefined;
+			let fromExtension = false;
+
+			if (this._extensionRunner.hasHandlers("session_before_compact")) {
+				const extensionResult = (await this._extensionRunner.emit({
+					type: "session_before_compact",
+					preparation,
+					branchEntries: pathEntries,
+					customInstructions: undefined,
+					signal: this._autoCompactionAbortController.signal,
+				})) as SessionBeforeCompactResult | undefined;
+
+				if (extensionResult?.cancel) {
+					this._emit({
+						type: "compaction_end",
+						reason,
+						result: undefined,
+						aborted: true,
+						willRetry: false,
+					});
+					return;
+				}
+
+				if (extensionResult?.compaction) {
+					extensionCompaction = extensionResult.compaction;
+					fromExtension = true;
+				}
+			}
+
+			let summary: string;
+			let firstKeptEntryId: string;
+			let tokensBefore: number;
+			let details: unknown;
+
+			if (extensionCompaction) {
+				// Extension provided compaction content
+				summary = extensionCompaction.summary;
+				firstKeptEntryId = extensionCompaction.firstKeptEntryId;
+				tokensBefore = extensionCompaction.tokensBefore;
+				details = extensionCompaction.details;
+			} else {
+				// Generate compaction result
+				const compactResult = await compact(
+					preparation,
+					this.model,
+					apiKey,
+					headers,
+					undefined,
+					this._autoCompactionAbortController.signal,
+					this.thinkingLevel,
+				);
+				summary = compactResult.summary;
+				firstKeptEntryId = compactResult.firstKeptEntryId;
+				tokensBefore = compactResult.tokensBefore;
+				details = compactResult.details;
+			}
+
+			if (this._autoCompactionAbortController.signal.aborted) {
+				this._emit({
+					type: "compaction_end",
+					reason,
+					result: undefined,
+					aborted: true,
+					willRetry: false,
+				});
+				return;
+			}
+
+			this.sessionManager.appendCompaction(summary, firstKeptEntryId, tokensBefore, details, fromExtension);
+			const newEntries = this.sessionManager.getEntries();
+			const sessionContext = this.sessionManager.buildSessionContext();
+			this.agent.state.messages = sessionContext.messages;
+
+			// Get the saved compaction entry for the extension event
+			const savedCompactionEntry = newEntries.find((e) => e.type === "compaction" && e.summary === summary) as
+				| CompactionEntry
+				| undefined;
+
+			if (this._extensionRunner && savedCompactionEntry) {
+				await this._extensionRunner.emit({
+					type: "session_compact",
+					compactionEntry: savedCompactionEntry,
+					fromExtension,
+				});
+			}
+
+			const result: CompactionResult = {
+				summary,
+				firstKeptEntryId,
+				tokensBefore,
+				details,
+			};
+			this._emit({ type: "compaction_end", reason, result, aborted: false, willRetry });
+
+			if (willRetry) {
+				const messages = this.agent.state.messages;
+				const lastMsg = messages[messages.length - 1];
+				if (lastMsg?.role === "assistant" && (lastMsg as AssistantMessage).stopReason === "error") {
+					this.agent.state.messages = messages.slice(0, -1);
+				}
+
+				setTimeout(() => {
+					this.agent.continue().catch(() => {});
+				}, 100);
+			} else if (this.agent.hasQueuedMessages()) {
+				// Auto-compaction can complete while follow-up/steering/custom messages are waiting.
+				// Kick the loop so queued messages are actually delivered.
+				setTimeout(() => {
+					this.agent.continue().catch(() => {});
+				}, 100);
+			}
+		} catch (error) {
+			const errorMessage = error instanceof Error ? error.message : "compaction failed";
+			this._emit({
+				type: "compaction_end",
+				reason,
+				result: undefined,
+				aborted: false,
+				willRetry: false,
+				errorMessage:
+					reason === "overflow"
+						? `Context overflow recovery failed: ${errorMessage}`
+						: `Auto-compaction failed: ${errorMessage}`,
+			});
+		} finally {
+			this._autoCompactionAbortController = undefined;
+		}
+	}
+
+	/**
+	 * Toggle auto-compaction setting.
+	 */
+	setAutoCompactionEnabled(enabled: boolean): void {
+		this.settingsManager.setCompactionEnabled(enabled);
+	}
+
+	/** Whether auto-compaction is enabled */
+	get autoCompactionEnabled(): boolean {
+		return this.settingsManager.getCompactionEnabled();
+	}
+
+	async bindExtensions(bindings: ExtensionBindings): Promise<void> {
+		if (bindings.uiContext !== undefined) {
+			this._extensionUIContext = bindings.uiContext;
+		}
+		if (bindings.commandContextActions !== undefined) {
+			this._extensionCommandContextActions = bindings.commandContextActions;
+		}
+		if (bindings.shutdownHandler !== undefined) {
+			this._extensionShutdownHandler = bindings.shutdownHandler;
+		}
+		if (bindings.onError !== undefined) {
+			this._extensionErrorListener = bindings.onError;
+		}
+
+		this._applyExtensionBindings(this._extensionRunner);
+		await this._extensionRunner.emit(this._sessionStartEvent);
+		await this.extendResourcesFromExtensions(this._sessionStartEvent.reason === "reload" ? "reload" : "startup");
+	}
+
+	private async extendResourcesFromExtensions(reason: "startup" | "reload"): Promise<void> {
+		if (!this._extensionRunner.hasHandlers("resources_discover")) {
+			return;
+		}
+
+		const { skillPaths, promptPaths, themePaths } = await this._extensionRunner.emitResourcesDiscover(
+			this._cwd,
+			reason,
+		);
+
+		if (skillPaths.length === 0 && promptPaths.length === 0 && themePaths.length === 0) {
+			return;
+		}
+
+		const extensionPaths: ResourceExtensionPaths = {
+			skillPaths: this.buildExtensionResourcePaths(skillPaths),
+			promptPaths: this.buildExtensionResourcePaths(promptPaths),
+			themePaths: this.buildExtensionResourcePaths(themePaths),
+		};
+
+		this._resourceLoader.extendResources(extensionPaths);
+		this._baseSystemPrompt = this._rebuildSystemPrompt(this.getActiveToolNames());
+		this.agent.state.systemPrompt = this._baseSystemPrompt;
+	}
+
+	private buildExtensionResourcePaths(entries: Array<{ path: string; extensionPath: string }>): Array<{
+		path: string;
+		metadata: { source: string; scope: "temporary"; origin: "top-level"; baseDir?: string };
+	}> {
+		return entries.map((entry) => {
+			const source = this.getExtensionSourceLabel(entry.extensionPath);
+			const baseDir = entry.extensionPath.startsWith("<") ? undefined : dirname(entry.extensionPath);
+			return {
+				path: entry.path,
+				metadata: {
+					source,
+					scope: "temporary",
+					origin: "top-level",
+					baseDir,
+				},
+			};
+		});
+	}
+
+	private getExtensionSourceLabel(extensionPath: string): string {
+		if (extensionPath.startsWith("<")) {
+			return `extension:${extensionPath.replace(/[<>]/g, "")}`;
+		}
+		const base = basename(extensionPath);
+		const name = base.replace(/\.(ts|js)$/, "");
+		return `extension:${name}`;
+	}
+
+	private _applyExtensionBindings(runner: ExtensionRunner): void {
+		runner.setUIContext(this._extensionUIContext);
+		runner.bindCommandContext(this._extensionCommandContextActions);
+
+		this._extensionErrorUnsubscriber?.();
+		this._extensionErrorUnsubscriber = this._extensionErrorListener
+			? runner.onError(this._extensionErrorListener)
+			: undefined;
+	}
+
+	private _refreshCurrentModelFromRegistry(): void {
+		const currentModel = this.model;
+		if (!currentModel) {
+			return;
+		}
+
+		const refreshedModel = this._modelRegistry.find(currentModel.provider, currentModel.id);
+		if (!refreshedModel || refreshedModel === currentModel) {
+			return;
+		}
+
+		this.agent.state.model = refreshedModel;
+	}
+
+	private _bindExtensionCore(runner: ExtensionRunner): void {
+		const getCommands = (): SlashCommandInfo[] => {
+			const extensionCommands: SlashCommandInfo[] = runner.getRegisteredCommands().map((command) => ({
+				name: command.invocationName,
+				description: command.description,
+				source: "extension",
+				sourceInfo: command.sourceInfo,
+			}));
+
+			const templates: SlashCommandInfo[] = this.promptTemplates.map((template) => ({
+				name: template.name,
+				description: template.description,
+				source: "prompt",
+				sourceInfo: template.sourceInfo,
+			}));
+
+			const skills: SlashCommandInfo[] = this._resourceLoader.getSkills().skills.map((skill) => ({
+				name: `skill:${skill.name}`,
+				description: skill.description,
+				source: "skill",
+				sourceInfo: skill.sourceInfo,
+			}));
+
+			return [...extensionCommands, ...templates, ...skills];
+		};
+
+		runner.bindCore(
+			{
+				sendMessage: (message, options) => {
+					this.sendCustomMessage(message, options).catch((err) => {
+						runner.emitError({
+							extensionPath: "<runtime>",
+							event: "send_message",
+							error: err instanceof Error ? err.message : String(err),
+						});
+					});
+				},
+				sendUserMessage: (content, options) => {
+					this.sendUserMessage(content, options).catch((err) => {
+						runner.emitError({
+							extensionPath: "<runtime>",
+							event: "send_user_message",
+							error: err instanceof Error ? err.message : String(err),
+						});
+					});
+				},
+				appendEntry: (customType, data) => {
+					this.sessionManager.appendCustomEntry(customType, data);
+				},
+				setSessionName: (name) => {
+					this.sessionManager.appendSessionInfo(name);
+				},
+				getSessionName: () => {
+					return this.sessionManager.getSessionName();
+				},
+				setLabel: (entryId, label) => {
+					this.sessionManager.appendLabelChange(entryId, label);
+				},
+				getActiveTools: () => this.getActiveToolNames(),
+				getAllTools: () => this.getAllTools(),
+				setActiveTools: (toolNames) => this.setActiveToolsByName(toolNames),
+				refreshTools: () => this._refreshToolRegistry(),
+				getCommands,
+				setModel: async (model) => {
+					if (!this.modelRegistry.hasConfiguredAuth(model)) return false;
+					await this.setModel(model);
+					return true;
+				},
+				getThinkingLevel: () => this.thinkingLevel,
+				setThinkingLevel: (level) => this.setThinkingLevel(level),
+			},
+			{
+				getModel: () => this.model,
+				isIdle: () => !this.isStreaming,
+				getSignal: () => this.agent.signal,
+				abort: () => this.abort(),
+				hasPendingMessages: () => this.pendingMessageCount > 0,
+				shutdown: () => {
+					this._extensionShutdownHandler?.();
+				},
+				getContextUsage: () => this.getContextUsage(),
+				compact: (options) => {
+					void (async () => {
+						try {
+							const result = await this.compact(options?.customInstructions);
+							options?.onComplete?.(result);
+						} catch (error) {
+							const err = error instanceof Error ? error : new Error(String(error));
+							options?.onError?.(err);
+						}
+					})();
+				},
+				getSystemPrompt: () => this.systemPrompt,
+			},
+			{
+				registerProvider: (name, config) => {
+					this._modelRegistry.registerProvider(name, config);
+					this._refreshCurrentModelFromRegistry();
+				},
+				unregisterProvider: (name) => {
+					this._modelRegistry.unregisterProvider(name);
+					this._refreshCurrentModelFromRegistry();
+				},
+			},
+		);
+	}
+
+	private _refreshToolRegistry(options?: { activeToolNames?: string[]; includeAllExtensionTools?: boolean }): void {
+		const previousRegistryNames = new Set(this._toolRegistry.keys());
+		const previousActiveToolNames = this.getActiveToolNames();
+		const allowedToolNames = this._allowedToolNames;
+		const isAllowedTool = (name: string): boolean => !allowedToolNames || allowedToolNames.has(name);
+
+		const registeredTools = this._extensionRunner.getAllRegisteredTools();
+		const allCustomTools = [
+			...registeredTools,
+			...this._customTools.map((definition) => ({
+				definition,
+				sourceInfo: createSyntheticSourceInfo(`<sdk:${definition.name}>`, { source: "sdk" }),
+			})),
+		].filter((tool) => isAllowedTool(tool.definition.name));
+		const definitionRegistry = new Map<string, ToolDefinitionEntry>(
+			Array.from(this._baseToolDefinitions.entries())
+				.filter(([name]) => isAllowedTool(name))
+				.map(([name, definition]) => [
+					name,
+					{
+						definition,
+						sourceInfo: createSyntheticSourceInfo(`<builtin:${name}>`, { source: "builtin" }),
+					},
+				]),
+		);
+		for (const tool of allCustomTools) {
+			definitionRegistry.set(tool.definition.name, {
+				definition: tool.definition,
+				sourceInfo: tool.sourceInfo,
+			});
+		}
+		this._toolDefinitions = definitionRegistry;
+		this._toolPromptSnippets = new Map(
+			Array.from(definitionRegistry.values())
+				.map(({ definition }) => {
+					const snippet = this._normalizePromptSnippet(definition.promptSnippet);
+					return snippet ? ([definition.name, snippet] as const) : undefined;
+				})
+				.filter((entry): entry is readonly [string, string] => entry !== undefined),
+		);
+		this._toolPromptGuidelines = new Map(
+			Array.from(definitionRegistry.values())
+				.map(({ definition }) => {
+					const guidelines = this._normalizePromptGuidelines(definition.promptGuidelines);
+					return guidelines.length > 0 ? ([definition.name, guidelines] as const) : undefined;
+				})
+				.filter((entry): entry is readonly [string, string[]] => entry !== undefined),
+		);
+		const runner = this._extensionRunner;
+		const wrappedExtensionTools = wrapRegisteredTools(allCustomTools, runner);
+		const wrappedBuiltInTools = wrapRegisteredTools(
+			Array.from(this._baseToolDefinitions.values())
+				.filter((definition) => isAllowedTool(definition.name))
+				.map((definition) => ({
+					definition,
+					sourceInfo: createSyntheticSourceInfo(`<builtin:${definition.name}>`, { source: "builtin" }),
+				})),
+			runner,
+		);
+
+		const toolRegistry = new Map(wrappedBuiltInTools.map((tool) => [tool.name, tool]));
+		for (const tool of wrappedExtensionTools as AgentTool[]) {
+			toolRegistry.set(tool.name, tool);
+		}
+		this._toolRegistry = toolRegistry;
+
+		const nextActiveToolNames = (
+			options?.activeToolNames ? [...options.activeToolNames] : [...previousActiveToolNames]
+		).filter((name) => isAllowedTool(name));
+
+		if (allowedToolNames) {
+			for (const toolName of this._toolRegistry.keys()) {
+				if (allowedToolNames.has(toolName)) {
+					nextActiveToolNames.push(toolName);
+				}
+			}
+		} else if (options?.includeAllExtensionTools) {
+			for (const tool of wrappedExtensionTools) {
+				nextActiveToolNames.push(tool.name);
+			}
+		} else if (!options?.activeToolNames) {
+			for (const toolName of this._toolRegistry.keys()) {
+				if (!previousRegistryNames.has(toolName)) {
+					nextActiveToolNames.push(toolName);
+				}
+			}
+		}
+
+		this.setActiveToolsByName([...new Set(nextActiveToolNames)]);
+	}
+
+	private _buildRuntime(options: {
+		activeToolNames?: string[];
+		flagValues?: Map<string, boolean | string>;
+		includeAllExtensionTools?: boolean;
+	}): void {
+		const autoResizeImages = this.settingsManager.getImageAutoResize();
+		const shellCommandPrefix = this.settingsManager.getShellCommandPrefix();
+		const shellPath = this.settingsManager.getShellPath();
+		const baseToolDefinitions = this._baseToolsOverride
+			? Object.fromEntries(
+					Object.entries(this._baseToolsOverride).map(([name, tool]) => [
+						name,
+						createToolDefinitionFromAgentTool(tool),
+					]),
+				)
+			: createAllToolDefinitions(this._cwd, {
+					read: { autoResizeImages },
+					bash: { commandPrefix: shellCommandPrefix, shellPath },
+				});
+
+		this._baseToolDefinitions = new Map(
+			Object.entries(baseToolDefinitions).map(([name, tool]) => [name, tool as ToolDefinition]),
+		);
+
+		const extensionsResult = this._resourceLoader.getExtensions();
+		if (options.flagValues) {
+			for (const [name, value] of options.flagValues) {
+				extensionsResult.runtime.flagValues.set(name, value);
+			}
+		}
+
+		this._extensionRunner = new ExtensionRunner(
+			extensionsResult.extensions,
+			extensionsResult.runtime,
+			this._cwd,
+			this.sessionManager,
+			this._modelRegistry,
+		);
+		if (this._extensionRunnerRef) {
+			this._extensionRunnerRef.current = this._extensionRunner;
+		}
+		this._bindExtensionCore(this._extensionRunner);
+		this._applyExtensionBindings(this._extensionRunner);
+
+		const defaultActiveToolNames = this._baseToolsOverride
+			? Object.keys(this._baseToolsOverride)
+			: ["read", "bash", "edit", "write"];
+		const baseActiveToolNames = options.activeToolNames ?? defaultActiveToolNames;
+		this._refreshToolRegistry({
+			activeToolNames: baseActiveToolNames,
+			includeAllExtensionTools: options.includeAllExtensionTools,
+		});
+	}
+
+	async reload(): Promise<void> {
+		const previousFlagValues = this._extensionRunner.getFlagValues();
+		await emitSessionShutdownEvent(this._extensionRunner, { type: "session_shutdown", reason: "reload" });
+		await this.settingsManager.reload();
+		resetApiProviders();
+		await this._resourceLoader.reload();
+		this._buildRuntime({
+			activeToolNames: this.getActiveToolNames(),
+			flagValues: previousFlagValues,
+			includeAllExtensionTools: true,
+		});
+
+		const hasBindings =
+			this._extensionUIContext ||
+			this._extensionCommandContextActions ||
+			this._extensionShutdownHandler ||
+			this._extensionErrorListener;
+		if (hasBindings) {
+			await this._extensionRunner.emit({ type: "session_start", reason: "reload" });
+			await this.extendResourcesFromExtensions("reload");
+		}
+	}
+
+	// =========================================================================
+	// Auto-Retry
+	// =========================================================================
+
+	/**
+	 * Check if an error is retryable (overloaded, rate limit, server errors).
+	 * Context overflow errors are NOT retryable (handled by compaction instead).
+	 */
+	private _isRetryableError(message: AssistantMessage): boolean {
+		if (message.stopReason !== "error" || !message.errorMessage) return false;
+
+		// Context overflow is handled by compaction, not retry
+		const contextWindow = this.model?.contextWindow ?? 0;
+		if (isContextOverflow(message, contextWindow)) return false;
+
+		const err = message.errorMessage;
+		// Match: overloaded_error, provider returned error, rate limit, 429, 500, 502, 503, 504, service unavailable, network/connection errors (including connection lost), fetch failed, request ended without sending chunks, terminated, retry delay exceeded
+		return /overloaded|provider.?returned.?error|rate.?limit|too many requests|429|500|502|503|504|service.?unavailable|server.?error|internal.?error|network.?error|connection.?error|connection.?refused|connection.?lost|other side closed|fetch failed|upstream.?connect|reset before headers|socket hang up|ended without|timed? out|timeout|terminated|retry delay/i.test(
+			err,
+		);
+	}
+
+	/**
+	 * Handle retryable errors with exponential backoff.
+	 * @returns true if retry was initiated, false if max retries exceeded or disabled
+	 */
+	private async _handleRetryableError(message: AssistantMessage): Promise<boolean> {
+		const settings = this.settingsManager.getRetrySettings();
+		if (!settings.enabled) {
+			this._resolveRetry();
+			return false;
+		}
+
+		// Retry promise is created synchronously in _handleAgentEvent for agent_end.
+		// Keep a defensive fallback here in case a future refactor bypasses that path.
+		if (!this._retryPromise) {
+			this._retryPromise = new Promise((resolve) => {
+				this._retryResolve = resolve;
+			});
+		}
+
+		this._retryAttempt++;
+
+		if (this._retryAttempt > settings.maxRetries) {
+			// Max retries exceeded, emit final failure and reset
+			this._emit({
+				type: "auto_retry_end",
+				success: false,
+				attempt: this._retryAttempt - 1,
+				finalError: message.errorMessage,
+			});
+			this._retryAttempt = 0;
+			this._resolveRetry(); // Resolve so waitForRetry() completes
+			return false;
+		}
+
+		const delayMs = settings.baseDelayMs * 2 ** (this._retryAttempt - 1);
+
+		this._emit({
+			type: "auto_retry_start",
+			attempt: this._retryAttempt,
+			maxAttempts: settings.maxRetries,
+			delayMs,
+			errorMessage: message.errorMessage || "Unknown error",
+		});
+
+		// Remove error message from agent state (keep in session for history)
+		const messages = this.agent.state.messages;
+		if (messages.length > 0 && messages[messages.length - 1].role === "assistant") {
+			this.agent.state.messages = messages.slice(0, -1);
+		}
+
+		// Wait with exponential backoff (abortable)
+		this._retryAbortController = new AbortController();
+		try {
+			await sleep(delayMs, this._retryAbortController.signal);
+		} catch {
+			// Aborted during sleep - emit end event so UI can clean up
+			const attempt = this._retryAttempt;
+			this._retryAttempt = 0;
+			this._retryAbortController = undefined;
+			this._emit({
+				type: "auto_retry_end",
+				success: false,
+				attempt,
+				finalError: "Retry cancelled",
+			});
+			this._resolveRetry();
+			return false;
+		}
+		this._retryAbortController = undefined;
+
+		// Retry via continue() - use setTimeout to break out of event handler chain
+		setTimeout(() => {
+			this.agent.continue().catch(() => {
+				// Retry failed - will be caught by next agent_end
+			});
+		}, 0);
+
+		return true;
+	}
+
+	/**
+	 * Cancel in-progress retry.
+	 */
+	abortRetry(): void {
+		this._retryAbortController?.abort();
+		// Note: _retryAttempt is reset in the catch block of _autoRetry
+		this._resolveRetry();
+	}
+
+	/**
+	 * Wait for any in-progress retry to complete.
+	 * Returns immediately if no retry is in progress.
+	 */
+	private async waitForRetry(): Promise<void> {
+		if (!this._retryPromise) {
+			return;
+		}
+
+		await this._retryPromise;
+		await this.agent.waitForIdle();
+	}
+
+	/** Whether auto-retry is currently in progress */
+	get isRetrying(): boolean {
+		return this._retryPromise !== undefined;
+	}
+
+	/** Whether auto-retry is enabled */
+	get autoRetryEnabled(): boolean {
+		return this.settingsManager.getRetryEnabled();
+	}
+
+	/**
+	 * Toggle auto-retry setting.
+	 */
+	setAutoRetryEnabled(enabled: boolean): void {
+		this.settingsManager.setRetryEnabled(enabled);
+	}
+
+	// =========================================================================
+	// Bash Execution
+	// =========================================================================
+
+	/**
+	 * Execute a bash command.
+	 * Adds result to agent context and session.
+	 * @param command The bash command to execute
+	 * @param onChunk Optional streaming callback for output
+	 * @param options.excludeFromContext If true, command output won't be sent to LLM (!! prefix)
+	 * @param options.operations Custom BashOperations for remote execution
+	 */
+	async executeBash(
+		command: string,
+		onChunk?: (chunk: string) => void,
+		options?: { excludeFromContext?: boolean; operations?: BashOperations },
+	): Promise<BashResult> {
+		this._bashAbortController = new AbortController();
+
+		// Apply command prefix if configured (e.g., "shopt -s expand_aliases" for alias support)
+		const prefix = this.settingsManager.getShellCommandPrefix();
+		const shellPath = this.settingsManager.getShellPath();
+		const resolvedCommand = prefix ? `${prefix}\n${command}` : command;
+
+		try {
+			const result = await executeBashWithOperations(
+				resolvedCommand,
+				this.sessionManager.getCwd(),
+				options?.operations ?? createLocalBashOperations({ shellPath }),
+				{
+					onChunk,
+					signal: this._bashAbortController.signal,
+				},
+			);
+
+			this.recordBashResult(command, result, options);
+			return result;
+		} finally {
+			this._bashAbortController = undefined;
+		}
+	}
+
+	/**
+	 * Record a bash execution result in session history.
+	 * Used by executeBash and by extensions that handle bash execution themselves.
+	 */
+	recordBashResult(command: string, result: BashResult, options?: { excludeFromContext?: boolean }): void {
+		const bashMessage: BashExecutionMessage = {
+			role: "bashExecution",
+			command,
+			output: result.output,
+			exitCode: result.exitCode,
+			cancelled: result.cancelled,
+			truncated: result.truncated,
+			fullOutputPath: result.fullOutputPath,
+			timestamp: Date.now(),
+			excludeFromContext: options?.excludeFromContext,
+		};
+
+		// If agent is streaming, defer adding to avoid breaking tool_use/tool_result ordering
+		if (this.isStreaming) {
+			// Queue for later - will be flushed on agent_end
+			this._pendingBashMessages.push(bashMessage);
+		} else {
+			// Add to agent state immediately
+			this.agent.state.messages.push(bashMessage);
+
+			// Save to session
+			this.sessionManager.appendMessage(bashMessage);
+		}
+	}
+
+	/**
+	 * Cancel running bash command.
+	 */
+	abortBash(): void {
+		this._bashAbortController?.abort();
+	}
+
+	/** Whether a bash command is currently running */
+	get isBashRunning(): boolean {
+		return this._bashAbortController !== undefined;
+	}
+
+	/** Whether there are pending bash messages waiting to be flushed */
+	get hasPendingBashMessages(): boolean {
+		return this._pendingBashMessages.length > 0;
+	}
+
+	/**
+	 * Flush pending bash messages to agent state and session.
+	 * Called after agent turn completes to maintain proper message ordering.
+	 */
+	private _flushPendingBashMessages(): void {
+		if (this._pendingBashMessages.length === 0) return;
+
+		for (const bashMessage of this._pendingBashMessages) {
+			// Add to agent state
+			this.agent.state.messages.push(bashMessage);
+
+			// Save to session
+			this.sessionManager.appendMessage(bashMessage);
+		}
+
+		this._pendingBashMessages = [];
+	}
+
+	// =========================================================================
+	// Session Management
+	// =========================================================================
+
+	/**
+	 * Set a display name for the current session.
+	 */
+	setSessionName(name: string): void {
+		this.sessionManager.appendSessionInfo(name);
+	}
+
+	// =========================================================================
+	// Tree Navigation
+	// =========================================================================
+
+	/**
+	 * Navigate to a different node in the session tree.
+	 * Unlike fork() which creates a new session file, this stays in the same file.
+	 *
+	 * @param targetId The entry ID to navigate to
+	 * @param options.summarize Whether user wants to summarize abandoned branch
+	 * @param options.customInstructions Custom instructions for summarizer
+	 * @param options.replaceInstructions If true, customInstructions replaces the default prompt
+	 * @param options.label Label to attach to the branch summary entry
+	 * @returns Result with editorText (if user message) and cancelled status
+	 */
+	async navigateTree(
+		targetId: string,
+		options: { summarize?: boolean; customInstructions?: string; replaceInstructions?: boolean; label?: string } = {},
+	): Promise<{ editorText?: string; cancelled: boolean; aborted?: boolean; summaryEntry?: BranchSummaryEntry }> {
+		const oldLeafId = this.sessionManager.getLeafId();
+
+		// No-op if already at target
+		if (targetId === oldLeafId) {
+			return { cancelled: false };
+		}
+
+		// Model required for summarization
+		if (options.summarize && !this.model) {
+			throw new Error("No model available for summarization");
+		}
+
+		const targetEntry = this.sessionManager.getEntry(targetId);
+		if (!targetEntry) {
+			throw new Error(`Entry ${targetId} not found`);
+		}
+
+		// Collect entries to summarize (from old leaf to common ancestor)
+		const { entries: entriesToSummarize, commonAncestorId } = collectEntriesForBranchSummary(
+			this.sessionManager,
+			oldLeafId,
+			targetId,
+		);
+
+		// Prepare event data - mutable so extensions can override
+		let customInstructions = options.customInstructions;
+		let replaceInstructions = options.replaceInstructions;
+		let label = options.label;
+
+		const preparation: TreePreparation = {
+			targetId,
+			oldLeafId,
+			commonAncestorId,
+			entriesToSummarize,
+			userWantsSummary: options.summarize ?? false,
+			customInstructions,
+			replaceInstructions,
+			label,
+		};
+
+		// Set up abort controller for summarization
+		this._branchSummaryAbortController = new AbortController();
+		let extensionSummary: { summary: string; details?: unknown } | undefined;
+		let fromExtension = false;
+
+		// Emit session_before_tree event
+		if (this._extensionRunner.hasHandlers("session_before_tree")) {
+			const result = (await this._extensionRunner.emit({
+				type: "session_before_tree",
+				preparation,
+				signal: this._branchSummaryAbortController.signal,
+			})) as SessionBeforeTreeResult | undefined;
+
+			if (result?.cancel) {
+				return { cancelled: true };
+			}
+
+			if (result?.summary && options.summarize) {
+				extensionSummary = result.summary;
+				fromExtension = true;
+			}
+
+			// Allow extensions to override instructions and label
+			if (result?.customInstructions !== undefined) {
+				customInstructions = result.customInstructions;
+			}
+			if (result?.replaceInstructions !== undefined) {
+				replaceInstructions = result.replaceInstructions;
+			}
+			if (result?.label !== undefined) {
+				label = result.label;
+			}
+		}
+
+		// Run default summarizer if needed
+		let summaryText: string | undefined;
+		let summaryDetails: unknown;
+		if (options.summarize && entriesToSummarize.length > 0 && !extensionSummary) {
+			const model = this.model!;
+			const { apiKey, headers } = await this._getRequiredRequestAuth(model);
+			const branchSummarySettings = this.settingsManager.getBranchSummarySettings();
+			const result = await generateBranchSummary(entriesToSummarize, {
+				model,
+				apiKey,
+				headers,
+				signal: this._branchSummaryAbortController.signal,
+				customInstructions,
+				replaceInstructions,
+				reserveTokens: branchSummarySettings.reserveTokens,
+			});
+			this._branchSummaryAbortController = undefined;
+			if (result.aborted) {
+				return { cancelled: true, aborted: true };
+			}
+			if (result.error) {
+				throw new Error(result.error);
+			}
+			summaryText = result.summary;
+			summaryDetails = {
+				readFiles: result.readFiles || [],
+				modifiedFiles: result.modifiedFiles || [],
+			};
+		} else if (extensionSummary) {
+			summaryText = extensionSummary.summary;
+			summaryDetails = extensionSummary.details;
+		}
+
+		// Determine the new leaf position based on target type
+		let newLeafId: string | null;
+		let editorText: string | undefined;
+
+		if (targetEntry.type === "message" && targetEntry.message.role === "user") {
+			// User message: leaf = parent (null if root), text goes to editor
+			newLeafId = targetEntry.parentId;
+			editorText = this._extractUserMessageText(targetEntry.message.content);
+		} else if (targetEntry.type === "custom_message") {
+			// Custom message: leaf = parent (null if root), text goes to editor
+			newLeafId = targetEntry.parentId;
+			editorText =
+				typeof targetEntry.content === "string"
+					? targetEntry.content
+					: targetEntry.content
+							.filter((c): c is { type: "text"; text: string } => c.type === "text")
+							.map((c) => c.text)
+							.join("");
+		} else {
+			// Non-user message: leaf = selected node
+			newLeafId = targetId;
+		}
+
+		// Switch leaf (with or without summary)
+		// Summary is attached at the navigation target position (newLeafId), not the old branch
+		let summaryEntry: BranchSummaryEntry | undefined;
+		if (summaryText) {
+			// Create summary at target position (can be null for root)
+			const summaryId = this.sessionManager.branchWithSummary(newLeafId, summaryText, summaryDetails, fromExtension);
+			summaryEntry = this.sessionManager.getEntry(summaryId) as BranchSummaryEntry;
+
+			// Attach label to the summary entry
+			if (label) {
+				this.sessionManager.appendLabelChange(summaryId, label);
+			}
+		} else if (newLeafId === null) {
+			// No summary, navigating to root - reset leaf
+			this.sessionManager.resetLeaf();
+		} else {
+			// No summary, navigating to non-root
+			this.sessionManager.branch(newLeafId);
+		}
+
+		// Attach label to target entry when not summarizing (no summary entry to label)
+		if (label && !summaryText) {
+			this.sessionManager.appendLabelChange(targetId, label);
+		}
+
+		// Update agent state
+		const sessionContext = this.sessionManager.buildSessionContext();
+		this.agent.state.messages = sessionContext.messages;
+
+		// Emit session_tree event
+		await this._extensionRunner.emit({
+			type: "session_tree",
+			newLeafId: this.sessionManager.getLeafId(),
+			oldLeafId,
+			summaryEntry,
+			fromExtension: summaryText ? fromExtension : undefined,
+		});
+
+		// Emit to custom tools
+
+		this._branchSummaryAbortController = undefined;
+		return { editorText, cancelled: false, summaryEntry };
+	}
+
+	/**
+	 * Get all user messages from session for fork selector.
+	 */
+	getUserMessagesForForking(): Array<{ entryId: string; text: string }> {
+		const entries = this.sessionManager.getEntries();
+		const result: Array<{ entryId: string; text: string }> = [];
+
+		for (const entry of entries) {
+			if (entry.type !== "message") continue;
+			if (entry.message.role !== "user") continue;
+
+			const text = this._extractUserMessageText(entry.message.content);
+			if (text) {
+				result.push({ entryId: entry.id, text });
+			}
+		}
+
+		return result;
+	}
+
+	private _extractUserMessageText(content: string | Array<{ type: string; text?: string }>): string {
+		if (typeof content === "string") return content;
+		if (Array.isArray(content)) {
+			return content
+				.filter((c): c is { type: "text"; text: string } => c.type === "text")
+				.map((c) => c.text)
+				.join("");
+		}
+		return "";
+	}
+
+	/**
+	 * Get session statistics.
+	 */
+	getSessionStats(): SessionStats {
+		const state = this.state;
+		const userMessages = state.messages.filter((m) => m.role === "user").length;
+		const assistantMessages = state.messages.filter((m) => m.role === "assistant").length;
+		const toolResults = state.messages.filter((m) => m.role === "toolResult").length;
+
+		let toolCalls = 0;
+		let totalInput = 0;
+		let totalOutput = 0;
+		let totalCacheRead = 0;
+		let totalCacheWrite = 0;
+		let totalCost = 0;
+
+		for (const message of state.messages) {
+			if (message.role === "assistant") {
+				const assistantMsg = message as AssistantMessage;
+				toolCalls += assistantMsg.content.filter((c) => c.type === "toolCall").length;
+				totalInput += assistantMsg.usage.input;
+				totalOutput += assistantMsg.usage.output;
+				totalCacheRead += assistantMsg.usage.cacheRead;
+				totalCacheWrite += assistantMsg.usage.cacheWrite;
+				totalCost += assistantMsg.usage.cost.total;
+			}
+		}
+
+		return {
+			sessionFile: this.sessionFile,
+			sessionId: this.sessionId,
+			userMessages,
+			assistantMessages,
+			toolCalls,
+			toolResults,
+			totalMessages: state.messages.length,
+			tokens: {
+				input: totalInput,
+				output: totalOutput,
+				cacheRead: totalCacheRead,
+				cacheWrite: totalCacheWrite,
+				total: totalInput + totalOutput + totalCacheRead + totalCacheWrite,
+			},
+			cost: totalCost,
+			contextUsage: this.getContextUsage(),
+		};
+	}
+
+	getContextUsage(): ContextUsage | undefined {
+		const model = this.model;
+		if (!model) return undefined;
+
+		const contextWindow = model.contextWindow ?? 0;
+		if (contextWindow <= 0) return undefined;
+
+		// After compaction, the last assistant usage reflects pre-compaction context size.
+		// We can only trust usage from an assistant that responded after the latest compaction.
+		// If no such assistant exists, context token count is unknown until the next LLM response.
+		const branchEntries = this.sessionManager.getBranch();
+		const latestCompaction = getLatestCompactionEntry(branchEntries);
+
+		if (latestCompaction) {
+			// Check if there's a valid assistant usage after the compaction boundary
+			const compactionIndex = branchEntries.lastIndexOf(latestCompaction);
+			let hasPostCompactionUsage = false;
+			for (let i = branchEntries.length - 1; i > compactionIndex; i--) {
+				const entry = branchEntries[i];
+				if (entry.type === "message" && entry.message.role === "assistant") {
+					const assistant = entry.message;
+					if (assistant.stopReason !== "aborted" && assistant.stopReason !== "error") {
+						const contextTokens = calculateContextTokens(assistant.usage);
+						if (contextTokens > 0) {
+							hasPostCompactionUsage = true;
+						}
+						break;
+					}
+				}
+			}
+
+			if (!hasPostCompactionUsage) {
+				return { tokens: null, contextWindow, percent: null };
+			}
+		}
+
+		const estimate = estimateContextTokens(this.messages);
+		const percent = (estimate.tokens / contextWindow) * 100;
+
+		return {
+			tokens: estimate.tokens,
+			contextWindow,
+			percent,
+		};
+	}
+
+	/**
+	 * Export session to HTML.
+	 * @param outputPath Optional output path (defaults to session directory)
+	 * @returns Path to exported file
+	 */
+	async exportToHtml(outputPath?: string): Promise<string> {
+		const themeName = this.settingsManager.getTheme();
+
+		// Create tool renderer if we have an extension runner (for custom tool HTML rendering)
+		const toolRenderer: ToolHtmlRenderer = createToolHtmlRenderer({
+			getToolDefinition: (name) => this.getToolDefinition(name),
+			theme,
+			cwd: this.sessionManager.getCwd(),
+		});
+
+		return await exportSessionToHtml(this.sessionManager, this.state, {
+			outputPath,
+			themeName,
+			toolRenderer,
+		});
+	}
+
+	/**
+	 * Export the current session branch to a JSONL file.
+	 * Writes the session header followed by all entries on the current branch path.
+	 * @param outputPath Target file path. If omitted, generates a timestamped file in cwd.
+	 * @returns The resolved output file path.
+	 */
+	exportToJsonl(outputPath?: string): string {
+		const filePath = resolve(outputPath ?? `session-${new Date().toISOString().replace(/[:.]/g, "-")}.jsonl`);
+		const dir = dirname(filePath);
+		if (!existsSync(dir)) {
+			mkdirSync(dir, { recursive: true });
+		}
+
+		const header: SessionHeader = {
+			type: "session",
+			version: CURRENT_SESSION_VERSION,
+			id: this.sessionManager.getSessionId(),
+			timestamp: new Date().toISOString(),
+			cwd: this.sessionManager.getCwd(),
+		};
+
+		const branchEntries = this.sessionManager.getBranch();
+		const lines = [JSON.stringify(header)];
+
+		// Re-chain parentIds to form a linear sequence
+		let prevId: string | null = null;
+		for (const entry of branchEntries) {
+			const linear = { ...entry, parentId: prevId };
+			lines.push(JSON.stringify(linear));
+			prevId = entry.id;
+		}
+
+		writeFileSync(filePath, `${lines.join("\n")}\n`);
+		return filePath;
+	}
+
+	// =========================================================================
+	// Utilities
+	// =========================================================================
+
+	/**
+	 * Get text content of last assistant message.
+	 * Useful for /copy command.
+	 * @returns Text content, or undefined if no assistant message exists
+	 */
+	getLastAssistantText(): string | undefined {
+		const lastAssistant = this.messages
+			.slice()
+			.reverse()
+			.find((m) => {
+				if (m.role !== "assistant") return false;
+				const msg = m as AssistantMessage;
+				// Skip aborted messages with no content
+				if (msg.stopReason === "aborted" && msg.content.length === 0) return false;
+				return true;
+			});
+
+		if (!lastAssistant) return undefined;
+
+		let text = "";
+		for (const content of (lastAssistant as AssistantMessage).content) {
+			if (content.type === "text") {
+				text += content.text;
+			}
+		}
+
+		return text.trim() || undefined;
+	}
+
+	// =========================================================================
+	// Extension System
+	// =========================================================================
+
+	createReplacedSessionContext(): ReplacedSessionContext {
+		const context = Object.defineProperties(
+			{},
+			Object.getOwnPropertyDescriptors(this._extensionRunner.createCommandContext()),
+		) as ReplacedSessionContext;
+		context.sendMessage = (message, options) => this.sendCustomMessage(message, options);
+		context.sendUserMessage = (content, options) => this.sendUserMessage(content, options);
+		return context;
+	}
+
+	/**
+	 * Check if extensions have handlers for a specific event type.
+	 */
+	hasExtensionHandlers(eventType: string): boolean {
+		return this._extensionRunner.hasHandlers(eventType);
+	}
+
+	/**
+	 * Get the extension runner (for setting UI context and error handlers).
+	 */
+	get extensionRunner(): ExtensionRunner {
+		return this._extensionRunner;
+	}
+}