@prometheus-ai/agent-core 0.5.0

package/src/compaction/compaction.ts

@@ -0,0 +1,1155 @@
+/**
+ * Context compaction for long sessions.
+ *
+ * Pure functions for compaction logic. The session manager handles I/O,
+ * and after compaction the session is reloaded.
+ */
+
+import {
+	type AssistantMessage,
+	clampThinkingLevelForModel,
+	Effort,
+	type Message,
+	type MessageAttribution,
+	type Model,
+	type Usage,
+} from "@prometheus-ai/ai";
+import { countTokens } from "@prometheus-ai/natives";
+import { logger, prompt } from "@prometheus-ai/utils";
+import { type AgentTelemetry, instrumentedCompleteSimple } from "../telemetry";
+import { ThinkingLevel } from "../thinking";
+import type { AgentMessage, AgentTool } from "../types";
+import type { CompactionEntry, SessionEntry } from "./entries";
+import { type ConvertToLlm, convertToLlm, createBranchSummaryMessage, createCustomMessage } from "./messages";
+import {
+	buildOpenAiNativeHistory,
+	getPreservedOpenAiRemoteCompactionData,
+	requestOpenAiRemoteCompaction,
+	requestRemoteCompaction,
+	shouldUseOpenAiRemoteCompaction,
+	withOpenAiRemoteCompactionPreserveData,
+} from "./openai";
+import autoHandoffThresholdFocusPrompt from "./prompts/auto-handoff-threshold-focus.md" with { type: "text" };
+import compactionShortSummaryPrompt from "./prompts/compaction-short-summary.md" with { type: "text" };
+import compactionSummaryPrompt from "./prompts/compaction-summary.md" with { type: "text" };
+import compactionTurnPrefixPrompt from "./prompts/compaction-turn-prefix.md" with { type: "text" };
+import compactionUpdateSummaryPrompt from "./prompts/compaction-update-summary.md" with { type: "text" };
+import handoffDocumentPrompt from "./prompts/handoff-document.md" with { type: "text" };
+
+import {
+	computeFileLists,
+	createFileOps,
+	extractFileOpsFromMessage,
+	type FileOperations,
+	SUMMARIZATION_SYSTEM_PROMPT,
+	serializeConversation,
+	upsertFileOperations,
+} from "./utils";
+
+// ============================================================================
+// File Operation Tracking
+// ============================================================================
+
+/** Details stored in CompactionEntry.details for file tracking */
+export interface CompactionDetails {
+	readFiles: string[];
+	modifiedFiles: string[];
+}
+
+/**
+ * Extract file operations from messages and previous compaction entries.
+ */
+function extractFileOperations(
+	messages: AgentMessage[],
+	entries: SessionEntry[],
+	prevCompactionIndex: number,
+): FileOperations {
+	const fileOps = createFileOps();
+
+	// Collect from previous compaction's details (if pi-generated)
+	if (prevCompactionIndex >= 0) {
+		const prevCompaction = entries[prevCompactionIndex] as CompactionEntry;
+		if (!prevCompaction.fromExtension && prevCompaction.details) {
+			const details = prevCompaction.details as CompactionDetails;
+			if (Array.isArray(details.readFiles)) {
+				for (const f of details.readFiles) fileOps.read.add(f);
+			}
+			if (Array.isArray(details.modifiedFiles)) {
+				for (const f of details.modifiedFiles) fileOps.edited.add(f);
+			}
+		}
+	}
+
+	// Extract from tool calls in messages
+	for (const msg of messages) {
+		extractFileOpsFromMessage(msg, fileOps);
+	}
+
+	return fileOps;
+}
+
+// ============================================================================
+// Message Extraction
+// ============================================================================
+
+/**
+ * Extract AgentMessage from an entry if it produces one.
+ * Returns undefined for entries that don't contribute to LLM context.
+ */
+function getMessageFromEntry(entry: SessionEntry): AgentMessage | undefined {
+	if (entry.type === "message") {
+		return entry.message;
+	}
+	if (entry.type === "custom_message") {
+		return createCustomMessage(
+			entry.customType,
+			entry.content,
+			entry.display,
+			entry.details,
+			entry.timestamp,
+			entry.attribution,
+		);
+	}
+	if (entry.type === "branch_summary") {
+		return createBranchSummaryMessage(entry.summary, entry.fromId, entry.timestamp);
+	}
+	return undefined;
+}
+
+/** Result from compact() - SessionManager adds uuid/parentUuid when saving */
+export interface CompactionResult<T = unknown> {
+	summary: string;
+	/** Short PR-style summary for display purposes. */
+	shortSummary?: string;
+	firstKeptEntryId: string;
+	tokensBefore: number;
+	/** Hook-specific data (e.g., ArtifactIndex, version markers for structured compaction) */
+	details?: T;
+	/** Hook-provided data to persist alongside compaction entry. */
+	preserveData?: Record<string, unknown>;
+}
+
+// ============================================================================
+// Types
+// ============================================================================
+
+export interface CompactionSettings {
+	enabled: boolean;
+	strategy?: "context-full" | "handoff" | "shake" | "off";
+	thresholdPercent?: number;
+	thresholdTokens?: number;
+	reserveTokens: number;
+	keepRecentTokens: number;
+	autoContinue?: boolean;
+	remoteEnabled?: boolean;
+	remoteEndpoint?: string;
+}
+
+export const DEFAULT_COMPACTION_SETTINGS: CompactionSettings = {
+	enabled: true,
+	strategy: "context-full",
+	thresholdPercent: -1,
+	thresholdTokens: -1,
+	reserveTokens: 16384,
+	keepRecentTokens: 20000,
+	autoContinue: true,
+	remoteEnabled: true,
+};
+
+// ============================================================================
+// Token calculation
+// ============================================================================
+
+/**
+ * Calculate total context tokens from usage.
+ * Uses the native totalTokens field when available, falls back to computing from components.
+ */
+export function calculateContextTokens(usage: Usage): number {
+	return usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;
+}
+
+export function calculatePromptTokens(usage: Usage): number {
+	const promptTokens = usage.input + usage.cacheRead + usage.cacheWrite;
+	if (promptTokens > 0) {
+		return promptTokens;
+	}
+	return calculateContextTokens(usage);
+}
+
+/**
+ * Get usage from an assistant message if available.
+ * Skips aborted and error messages as they don't have valid usage data.
+ */
+function getAssistantUsage(msg: AgentMessage): Usage | undefined {
+	if (msg.role === "assistant" && "usage" in msg) {
+		const assistantMsg = msg as AssistantMessage;
+		if (assistantMsg.stopReason !== "aborted" && assistantMsg.stopReason !== "error" && assistantMsg.usage) {
+			return assistantMsg.usage;
+		}
+	}
+	return undefined;
+}
+
+/**
+ * Find the last non-aborted assistant message usage from session entries.
+ */
+export function getLastAssistantUsage(entries: SessionEntry[]): Usage | undefined {
+	for (let i = entries.length - 1; i >= 0; i--) {
+		const entry = entries[i];
+		if (entry.type === "message") {
+			const usage = getAssistantUsage(entry.message);
+			if (usage) return usage;
+		}
+	}
+	return undefined;
+}
+
+/**
+ * Effective reserve: at least 15% of context window or the configured floor, whichever is larger.
+ */
+export function effectiveReserveTokens(contextWindow: number, settings: CompactionSettings): number {
+	return Math.max(Math.floor(contextWindow * 0.15), settings.reserveTokens);
+}
+
+/**
+ * Check if compaction should trigger based on context usage.
+ */
+export function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean {
+	if (!settings.enabled || settings.strategy === "off" || contextWindow <= 0) return false;
+	const thresholdTokens = resolveThresholdTokens(contextWindow, settings);
+	return contextTokens > thresholdTokens;
+}
+
+export function resolveThresholdTokens(contextWindow: number, settings: CompactionSettings): number {
+	// Fixed token limit takes priority over percentage
+	const thresholdTokens = settings.thresholdTokens;
+	if (typeof thresholdTokens === "number" && Number.isFinite(thresholdTokens) && thresholdTokens > 0) {
+		// Clamp to [1, contextWindow - 1] so there's always room
+		return Math.min(contextWindow - 1, Math.max(1, thresholdTokens));
+	}
+
+	// Percentage-based threshold
+	const thresholdPercent = settings.thresholdPercent;
+	if (typeof thresholdPercent !== "number" || !Number.isFinite(thresholdPercent) || thresholdPercent <= 0) {
+		return contextWindow - effectiveReserveTokens(contextWindow, settings);
+	}
+	const clampedThresholdPercent = Math.min(99, Math.max(1, thresholdPercent));
+	return Math.floor(contextWindow * (clampedThresholdPercent / 100));
+}
+
+// ============================================================================
+// Cut point detection
+// ============================================================================
+
+/**
+ * Image content has no tokenizer representation; charge a fixed estimate
+ * matching what providers typically bill for inline images.
+ */
+const IMAGE_TOKEN_ESTIMATE = 1200;
+
+/**
+ * Estimate token count for a message using cl100k_base via the native
+ * tokenizer. This is not Claude's first-party tokenizer (Anthropic doesn't
+ * publish one) but is within ~5–10% across English/code text.
+ */
+export function estimateTokens(message: AgentMessage): number {
+	const fragments: string[] = [];
+	let extra = 0;
+	if ((message as { role?: string }).role === "bashExecution") {
+		const bash = message as { command?: unknown; output?: unknown };
+		if (typeof bash.command === "string") fragments.push(bash.command);
+		if (typeof bash.output === "string") fragments.push(bash.output);
+		return fragments.length === 0 ? 0 : countTokens(fragments);
+	}
+
+	switch (message.role) {
+		case "user": {
+			const content = (message as { content: string | Array<{ type: string; text?: string }> }).content;
+			if (typeof content === "string") {
+				fragments.push(content);
+			} else if (Array.isArray(content)) {
+				for (const block of content) {
+					if (block.type === "text" && block.text) {
+						fragments.push(block.text);
+					}
+				}
+			}
+			break;
+		}
+		case "assistant": {
+			const assistant = message as AssistantMessage;
+			for (const block of assistant.content) {
+				if (block.type === "text") {
+					fragments.push(block.text);
+				} else if (block.type === "thinking") {
+					fragments.push(block.thinking);
+				} else if (block.type === "toolCall") {
+					fragments.push(block.name);
+					fragments.push(JSON.stringify(block.arguments));
+				}
+			}
+			break;
+		}
+		case "hookMessage":
+		case "toolResult": {
+			if (typeof message.content === "string") {
+				fragments.push(message.content);
+			} else {
+				for (const block of message.content) {
+					if (block.type === "text" && block.text) {
+						fragments.push(block.text);
+					} else if (block.type === "image") {
+						extra += IMAGE_TOKEN_ESTIMATE;
+					}
+				}
+			}
+			break;
+		}
+		case "branchSummary":
+		case "compactionSummary": {
+			fragments.push(message.summary);
+			break;
+		}
+		default:
+			return 0;
+	}
+
+	if (fragments.length === 0) return extra;
+	return extra + countTokens(fragments);
+}
+
+function estimateEntriesTokens(entries: SessionEntry[], startIndex: number, endIndex: number): number {
+	let total = 0;
+	for (let i = startIndex; i < endIndex; i++) {
+		const msg = getMessageFromEntry(entries[i]);
+		if (msg) {
+			total += estimateTokens(msg);
+		}
+	}
+	return total;
+}
+
+/**
+ * Find valid cut points: indices of user, assistant, custom, or bashExecution messages.
+ * Never cut at tool results (they must follow their tool call).
+ * When we cut at an assistant message with tool calls, its tool results follow it
+ * and will be kept.
+ * BashExecutionMessage is treated like a user message (user-initiated context).
+ */
+function findValidCutPoints(entries: SessionEntry[], startIndex: number, endIndex: number): number[] {
+	const cutPoints: number[] = [];
+	for (let i = startIndex; i < endIndex; i++) {
+		const entry = entries[i];
+		switch (entry.type) {
+			case "message": {
+				const role = entry.message.role as string;
+				switch (role) {
+					case "bashExecution":
+					case "hookMessage":
+					case "branchSummary":
+					case "compactionSummary":
+					case "user":
+					case "assistant":
+						cutPoints.push(i);
+						break;
+					case "toolResult":
+						break;
+				}
+				break;
+			}
+			case "thinking_level_change":
+			case "model_change":
+			case "compaction":
+			case "branch_summary":
+			case "custom":
+			case "custom_message":
+			case "label":
+		}
+		// branch_summary and custom_message are user-role messages, valid cut points
+		if (entry.type === "branch_summary" || entry.type === "custom_message") {
+			cutPoints.push(i);
+		}
+	}
+	return cutPoints;
+}
+
+/**
+ * Find the user message (or bashExecution) that starts the turn containing the given entry index.
+ * Returns -1 if no turn start found before the index.
+ * BashExecutionMessage is treated like a user message for turn boundaries.
+ */
+export function findTurnStartIndex(entries: SessionEntry[], entryIndex: number, startIndex: number): number {
+	for (let i = entryIndex; i >= startIndex; i--) {
+		const entry = entries[i];
+		// branch_summary and custom_message are user-role messages, can start a turn
+		if (entry.type === "branch_summary" || entry.type === "custom_message") {
+			return i;
+		}
+		if (entry.type === "message") {
+			const role = entry.message.role as string;
+			if (role === "user" || role === "bashExecution") {
+				return i;
+			}
+		}
+	}
+	return -1;
+}
+
+export interface CutPointResult {
+	/** Index of first entry to keep */
+	firstKeptEntryIndex: number;
+	/** Index of user message that starts the turn being split, or -1 if not splitting */
+	turnStartIndex: number;
+	/** Whether this cut splits a turn (cut point is not a user message) */
+	isSplitTurn: boolean;
+}
+
+/**
+ * Find the cut point in session entries that keeps approximately `keepRecentTokens`.
+ *
+ * Algorithm: Walk backwards from newest, accumulating estimated message sizes.
+ * Stop when we've accumulated >= keepRecentTokens. Cut at that point.
+ *
+ * Can cut at user OR assistant messages (never tool results). When cutting at an
+ * assistant message with tool calls, its tool results come after and will be kept.
+ *
+ * Returns CutPointResult with:
+ * - firstKeptEntryIndex: the entry index to start keeping from
+ * - turnStartIndex: if cutting mid-turn, the user message that started that turn
+ * - isSplitTurn: whether we're cutting in the middle of a turn
+ *
+ * Only considers entries between `startIndex` and `endIndex` (exclusive).
+ */
+export function findCutPoint(
+	entries: SessionEntry[],
+	startIndex: number,
+	endIndex: number,
+	keepRecentTokens: number,
+): CutPointResult {
+	const cutPoints = findValidCutPoints(entries, startIndex, endIndex);
+
+	if (cutPoints.length === 0) {
+		return { firstKeptEntryIndex: startIndex, turnStartIndex: -1, isSplitTurn: false };
+	}
+
+	// Walk backwards from newest, accumulating estimated message sizes
+	let accumulatedTokens = 0;
+	let cutIndex = cutPoints[0]; // Default: keep from first message (not header)
+
+	for (let i = endIndex - 1; i >= startIndex; i--) {
+		const entry = entries[i];
+		if (entry.type !== "message") continue;
+
+		// Estimate this message's size
+		const messageTokens = estimateTokens(entry.message);
+		accumulatedTokens += messageTokens;
+
+		// Check if we've exceeded the budget
+		if (accumulatedTokens >= keepRecentTokens) {
+			// Find the closest valid cut point at or after this entry
+			for (let c = 0; c < cutPoints.length; c++) {
+				if (cutPoints[c] >= i) {
+					cutIndex = cutPoints[c];
+					break;
+				}
+			}
+			break;
+		}
+	}
+
+	// Scan backwards from cutIndex to include any non-message entries (bash, settings, etc.)
+	while (cutIndex > startIndex) {
+		const prevEntry = entries[cutIndex - 1];
+		// Stop at session header or compaction boundaries
+		if (prevEntry.type === "compaction") {
+			break;
+		}
+		if (prevEntry.type === "message") {
+			// Stop if we hit any message
+			break;
+		}
+		// Include this non-message entry (bash, settings change, etc.)
+		cutIndex--;
+	}
+
+	// Determine if this is a split turn
+	const cutEntry = entries[cutIndex];
+	const isUserMessage = cutEntry.type === "message" && cutEntry.message.role === "user";
+	const turnStartIndex = isUserMessage ? -1 : findTurnStartIndex(entries, cutIndex, startIndex);
+
+	return {
+		firstKeptEntryIndex: cutIndex,
+		turnStartIndex,
+		isSplitTurn: !isUserMessage && turnStartIndex !== -1,
+	};
+}
+
+// ============================================================================
+// Summarization
+// ============================================================================
+
+const SUMMARIZATION_PROMPT = prompt.render(compactionSummaryPrompt);
+
+const UPDATE_SUMMARIZATION_PROMPT = prompt.render(compactionUpdateSummaryPrompt);
+
+const SHORT_SUMMARY_PROMPT = prompt.render(compactionShortSummaryPrompt);
+
+const HANDOFF_DOCUMENT_PROMPT = prompt.render(handoffDocumentPrompt);
+
+export const AUTO_HANDOFF_THRESHOLD_FOCUS = prompt.render(autoHandoffThresholdFocusPrompt);
+
+function formatAdditionalContext(context: string[] | undefined): string {
+	if (!context || context.length === 0) return "";
+	const lines = context.map(line => `- ${line}`).join("\n");
+	return `<additional-context>\n${lines}\n</additional-context>\n\n`;
+}
+
+/**
+ * Maps the non-special `ThinkingLevel` values to their `Effort` counterparts.
+ * Exhaustive over the union; throws for `Off`/`Inherit` to surface logic
+ * errors in callers that forgot to filter those out. Never use a TS cast for
+ * this — `ThinkingLevel` is a string-union over distinct concepts (Off /
+ * Inherit are not Efforts), and a cast hides the contract.
+ */
+function effortFromThinkingLevel(level: ThinkingLevel): Effort {
+	switch (level) {
+		case ThinkingLevel.Minimal:
+			return Effort.Minimal;
+		case ThinkingLevel.Low:
+			return Effort.Low;
+		case ThinkingLevel.Medium:
+			return Effort.Medium;
+		case ThinkingLevel.High:
+			return Effort.High;
+		case ThinkingLevel.XHigh:
+			return Effort.XHigh;
+		case ThinkingLevel.Off:
+		case ThinkingLevel.Inherit:
+			throw new Error(`effortFromThinkingLevel: ${level} must be handled by caller`);
+	}
+}
+
+/**
+ * Resolves the reasoning effort to send on a compaction LLM call.
+ *
+ * - Explicit `Off` → `undefined` (omit reasoning entirely; the user said no thinking).
+ * - `undefined` / `Inherit` → historical `Effort.High` default → clamped per model
+ *   (preserves current behavior for users who never touched the dial).
+ * - Explicit effort → respect user choice → clamped per model.
+ *
+ * The clamp routes through `clampThinkingLevelForModel`, which returns
+ * `undefined` for models with `compat.supportsReasoningEffort: false`
+ * (e.g. `xai-oauth/grok-build`). That `undefined` then flows through to the
+ * openai-responses mapper where `modelOmitsReasoningEffort` short-circuits
+ * the wire param — no `requireSupportedEffort` throw.
+ */
+function resolveCompactionEffort(model: Model, level: ThinkingLevel | undefined): Effort | undefined {
+	if (level === ThinkingLevel.Off) return undefined;
+	const requested: Effort =
+		level === undefined || level === ThinkingLevel.Inherit ? Effort.High : effortFromThinkingLevel(level);
+	return clampThinkingLevelForModel(model, requested);
+}
+
+/**
+ * Build the error thrown when an LLM summarization call ends with
+ * `stopReason === "error"`. Carries the provider's HTTP `errorStatus`
+ * onto a top-level `.status` field so callers (notably
+ * `AgentSession.#isCompactionAuthFailure`) can branch on 401/403 without
+ * regex-scraping `error.message`. The `auth_unavailable` synthetic
+ * (Prometheus native gateway) does not populate `errorStatus`, hence the legacy
+ * message-based check is still required upstream — see issue #986.
+ */
+function createSummarizationError(prefix: string, response: AssistantMessage): Error {
+	const error: Error & { status?: number } = new Error(`${prefix}: ${response.errorMessage || "Unknown error"}`);
+	if (response.errorStatus !== undefined) {
+		error.status = response.errorStatus;
+	}
+	return error;
+}
+
+/**
+ * Generate a summary of the conversation using the LLM.
+ * If previousSummary is provided, uses the update prompt to merge.
+ */
+export interface SummaryOptions {
+	promptOverride?: string;
+	extraContext?: string[];
+	remoteEndpoint?: string;
+	remoteInstructions?: string;
+	initiatorOverride?: MessageAttribution;
+	metadata?: Record<string, unknown>;
+	convertToLlm?: ConvertToLlm;
+	/**
+	 * Optional telemetry handle. When provided, every LLM call emitted during
+	 * compaction is wrapped in an OTEL chat span tagged with
+	 * `prometheus.gen_ai.oneshot.kind` (`compaction_summary`, `compaction_short_summary`,
+	 * or `compaction_turn_prefix`). `undefined` keeps the call paths zero-cost.
+	 */
+	telemetry?: AgentTelemetry;
+	/**
+	 * Active session thinking level. Threaded from `agent-session.ts` so
+	 * compaction honors the user's `/model` thinking selection instead of
+	 * silently overriding it with `Effort.High` (the historical default).
+	 * `undefined` / `ThinkingLevel.Inherit` falls back to that historical
+	 * default; `ThinkingLevel.Off` omits reasoning entirely. See
+	 * `resolveCompactionEffort` for the conversion contract.
+	 */
+	thinkingLevel?: ThinkingLevel;
+}
+
+export async function generateSummary(
+	currentMessages: AgentMessage[],
+	model: Model,
+	reserveTokens: number,
+	apiKey: string,
+	signal?: AbortSignal,
+	customInstructions?: string,
+	previousSummary?: string,
+	options?: SummaryOptions,
+): Promise<string> {
+	const maxTokens = Math.floor(0.8 * reserveTokens);
+
+	// Use update prompt if we have a previous summary, otherwise initial prompt
+	let basePrompt = previousSummary ? UPDATE_SUMMARIZATION_PROMPT : SUMMARIZATION_PROMPT;
+	if (options?.promptOverride) {
+		basePrompt = options.promptOverride;
+	}
+	if (customInstructions) {
+		basePrompt = `${basePrompt}\n\nAdditional focus: ${customInstructions}`;
+	}
+
+	// Serialize conversation to text so model doesn't try to continue it
+	// Convert to LLM messages first (handles custom app messages when caller provides a transformer).
+	const llmMessages = (options?.convertToLlm ?? convertToLlm)(currentMessages);
+	const conversationText = serializeConversation(llmMessages);
+
+	// Build the prompt with conversation wrapped in tags
+	let promptText = `<conversation>\n${conversationText}\n</conversation>\n\n`;
+	if (previousSummary) {
+		promptText += `<previous-summary>\n${previousSummary}\n</previous-summary>\n\n`;
+	}
+	promptText += formatAdditionalContext(options?.extraContext);
+	promptText += basePrompt;
+
+	const summarizationMessages = [
+		{
+			role: "user" as const,
+			content: [{ type: "text" as const, text: promptText }],
+			timestamp: Date.now(),
+		},
+	];
+
+	if (options?.remoteEndpoint) {
+		const remote = await requestRemoteCompaction(
+			options.remoteEndpoint,
+			{
+				systemPrompt: SUMMARIZATION_SYSTEM_PROMPT,
+				prompt: promptText,
+			},
+			signal,
+		);
+		return remote.summary;
+	}
+
+	const response = await instrumentedCompleteSimple(
+		model,
+		{ systemPrompt: [SUMMARIZATION_SYSTEM_PROMPT], messages: summarizationMessages },
+		{
+			maxTokens,
+			signal,
+			apiKey,
+			reasoning: resolveCompactionEffort(model, options?.thinkingLevel),
+			initiatorOverride: options?.initiatorOverride,
+			metadata: options?.metadata,
+		},
+		{ telemetry: options?.telemetry, oneshotKind: "compaction_summary" },
+	);
+
+	if (response.stopReason === "error") {
+		throw createSummarizationError("Summarization failed", response);
+	}
+
+	const textContent = response.content
+		.filter((c): c is { type: "text"; text: string } => c.type === "text")
+		.map(c => c.text)
+		.join("\n");
+
+	return textContent;
+}
+
+// ============================================================================
+// Handoff generation
+// ============================================================================
+
+export interface HandoffOptions {
+	/** Live agent system prompt — passed verbatim so providers hit the cached prefix. */
+	systemPrompt: string[];
+	/** Live agent tool list — same purpose. Forced to `toolChoice: "none"`. */
+	tools?: AgentTool<any>[];
+	customInstructions?: string;
+	convertToLlm?: ConvertToLlm;
+	initiatorOverride?: MessageAttribution;
+	metadata?: Record<string, unknown>;
+	/**
+	 * Optional telemetry handle. When provided, the handoff LLM call is
+	 * wrapped in an OTEL chat span tagged with `prometheus.gen_ai.oneshot.kind = "handoff"`.
+	 */
+	telemetry?: AgentTelemetry;
+	/**
+	 * Active session thinking level. Threaded from `agent-session.ts` so
+	 * handoff generation honors the user's `/model` thinking selection
+	 * instead of silently overriding it with `Effort.High`. See
+	 * `resolveCompactionEffort` for the conversion contract.
+	 */
+	thinkingLevel?: ThinkingLevel;
+}
+
+export function renderHandoffPrompt(customInstructions?: string): string {
+	if (!customInstructions) return HANDOFF_DOCUMENT_PROMPT;
+	return prompt.render(handoffDocumentPrompt, {
+		additionalFocus: customInstructions,
+	});
+}
+
+export async function generateHandoff(
+	messages: AgentMessage[],
+	model: Model,
+	apiKey: string,
+	options: HandoffOptions,
+	signal?: AbortSignal,
+): Promise<string> {
+	const llmMessages = (options.convertToLlm ?? convertToLlm)(messages);
+	const requestMessages: Message[] = [
+		...llmMessages,
+		{
+			role: "user",
+			content: [{ type: "text", text: renderHandoffPrompt(options.customInstructions) }],
+			attribution: "agent",
+			timestamp: Date.now(),
+		},
+	];
+
+	const response = await instrumentedCompleteSimple(
+		model,
+		{
+			systemPrompt: options.systemPrompt,
+			messages: requestMessages,
+			tools: options.tools,
+		},
+		{
+			apiKey,
+			signal,
+			reasoning: resolveCompactionEffort(model, options.thinkingLevel),
+			toolChoice: "none",
+			initiatorOverride: options.initiatorOverride,
+			metadata: options.metadata,
+		},
+		{ telemetry: options.telemetry, oneshotKind: "handoff" },
+	);
+
+	if (response.stopReason === "error") {
+		throw createSummarizationError("Handoff generation failed", response);
+	}
+
+	return response.content
+		.filter((c): c is { type: "text"; text: string } => c.type === "text")
+		.map(c => c.text)
+		.join("\n");
+}
+
+async function generateShortSummary(
+	recentMessages: AgentMessage[],
+	historySummary: string | undefined,
+	model: Model,
+	reserveTokens: number,
+	apiKey: string,
+	signal?: AbortSignal,
+	options?: SummaryOptions,
+): Promise<string> {
+	const maxTokens = Math.min(512, Math.floor(0.2 * reserveTokens));
+	const llmMessages = (options?.convertToLlm ?? convertToLlm)(recentMessages);
+	const conversationText = serializeConversation(llmMessages);
+
+	let promptText = `<conversation>\n${conversationText}\n</conversation>\n\n`;
+	if (historySummary) {
+		promptText += `<previous-summary>\n${historySummary}\n</previous-summary>\n\n`;
+	}
+	promptText += formatAdditionalContext(options?.extraContext);
+	promptText += SHORT_SUMMARY_PROMPT;
+
+	if (options?.remoteEndpoint) {
+		const remote = await requestRemoteCompaction(
+			options.remoteEndpoint,
+			{
+				systemPrompt: SUMMARIZATION_SYSTEM_PROMPT,
+				prompt: promptText,
+			},
+			signal,
+		);
+		return remote.summary;
+	}
+
+	const response = await instrumentedCompleteSimple(
+		model,
+		{
+			systemPrompt: [SUMMARIZATION_SYSTEM_PROMPT],
+			messages: [{ role: "user", content: [{ type: "text", text: promptText }], timestamp: Date.now() }],
+		},
+		{
+			maxTokens,
+			signal,
+			apiKey,
+			reasoning: resolveCompactionEffort(model, options?.thinkingLevel),
+			initiatorOverride: options?.initiatorOverride,
+			metadata: options?.metadata,
+		},
+		{ telemetry: options?.telemetry, oneshotKind: "compaction_short_summary" },
+	);
+
+	if (response.stopReason === "error") {
+		throw createSummarizationError("Short summary failed", response);
+	}
+
+	return response.content
+		.filter((c): c is { type: "text"; text: string } => c.type === "text")
+		.map(c => c.text)
+		.join("\n");
+}
+
+// ============================================================================
+// Compaction Preparation (for hooks)
+// ============================================================================
+
+export interface CompactionPreparation {
+	/** UUID of first entry to keep */
+	firstKeptEntryId: string;
+	/** Messages that will be summarized and discarded */
+	messagesToSummarize: AgentMessage[];
+	/** Messages that will be turned into turn prefix summary (if splitting) */
+	turnPrefixMessages: AgentMessage[];
+	/** Messages kept in full after compaction (recent history) */
+	recentMessages: AgentMessage[];
+	/** Whether this is a split turn (cut point in middle of turn) */
+	isSplitTurn: boolean;
+	tokensBefore: number;
+	/** Summary from previous compaction, for iterative update */
+	previousSummary?: string;
+	/** Preserved opaque compaction payload from the previous compaction, if any. */
+	previousPreserveData?: Record<string, unknown>;
+	/** File operations extracted from messagesToSummarize */
+	fileOps: FileOperations;
+	/** Compaction settions from settings.jsonl	*/
+	settings: CompactionSettings;
+}
+
+export function prepareCompaction(
+	pathEntries: SessionEntry[],
+	settings: CompactionSettings,
+): CompactionPreparation | undefined {
+	if (pathEntries.length > 0 && pathEntries[pathEntries.length - 1].type === "compaction") {
+		return undefined;
+	}
+
+	let prevCompactionIndex = -1;
+	for (let i = pathEntries.length - 1; i >= 0; i--) {
+		if (pathEntries[i].type === "compaction") {
+			prevCompactionIndex = i;
+			break;
+		}
+	}
+	const boundaryStart = prevCompactionIndex + 1;
+	const boundaryEnd = pathEntries.length;
+
+	const lastUsage = getLastAssistantUsage(pathEntries);
+	const tokensBefore = lastUsage ? calculateContextTokens(lastUsage) : 0;
+	let keepRecentTokens = settings.keepRecentTokens;
+	if (lastUsage) {
+		const estimatedTokens = estimateEntriesTokens(pathEntries, boundaryStart, boundaryEnd);
+		const promptTokens = calculatePromptTokens(lastUsage);
+		const ratio = estimatedTokens > 0 ? promptTokens / estimatedTokens : 0;
+		if (Number.isFinite(ratio) && ratio > 1) {
+			keepRecentTokens = Math.max(1, Math.floor(keepRecentTokens / ratio));
+		}
+	}
+
+	const cutPoint = findCutPoint(pathEntries, boundaryStart, boundaryEnd, keepRecentTokens);
+
+	// Get ID of first kept entry
+	const firstKeptEntry = pathEntries[cutPoint.firstKeptEntryIndex];
+	if (!firstKeptEntry?.id) {
+		return undefined; // Session needs migration
+	}
+	const firstKeptEntryId = firstKeptEntry.id;
+
+	const historyEnd = cutPoint.isSplitTurn ? cutPoint.turnStartIndex : cutPoint.firstKeptEntryIndex;
+
+	// Messages to summarize (will be discarded after summary)
+	const messagesToSummarize: AgentMessage[] = [];
+	for (let i = boundaryStart; i < historyEnd; i++) {
+		const msg = getMessageFromEntry(pathEntries[i]);
+		if (msg) messagesToSummarize.push(msg);
+	}
+
+	// Messages for turn prefix summary (if splitting a turn)
+	const turnPrefixMessages: AgentMessage[] = [];
+	if (cutPoint.isSplitTurn) {
+		for (let i = cutPoint.turnStartIndex; i < cutPoint.firstKeptEntryIndex; i++) {
+			const msg = getMessageFromEntry(pathEntries[i]);
+			if (msg) turnPrefixMessages.push(msg);
+		}
+	}
+
+	// Messages kept after compaction (recent history)
+	const recentMessages: AgentMessage[] = [];
+	for (let i = cutPoint.firstKeptEntryIndex; i < boundaryEnd; i++) {
+		const msg = getMessageFromEntry(pathEntries[i]);
+		if (msg) recentMessages.push(msg);
+	}
+	// Nothing to summarize means compaction would be a no-op.
+	if (messagesToSummarize.length === 0 && turnPrefixMessages.length === 0) {
+		return undefined;
+	}
+
+	// Get previous summary and preserved data for iterative updates
+	let previousSummary: string | undefined;
+	let previousPreserveData: Record<string, unknown> | undefined;
+	if (prevCompactionIndex >= 0) {
+		const prevCompaction = pathEntries[prevCompactionIndex] as CompactionEntry;
+		previousSummary = prevCompaction.summary;
+		previousPreserveData = prevCompaction.preserveData;
+	}
+
+	// Extract file operations from messages and previous compaction
+	const fileOps = extractFileOperations(messagesToSummarize, pathEntries, prevCompactionIndex);
+
+	// Also extract file ops from turn prefix if splitting
+	if (cutPoint.isSplitTurn) {
+		for (const msg of turnPrefixMessages) {
+			extractFileOpsFromMessage(msg, fileOps);
+		}
+	}
+
+	return {
+		firstKeptEntryId,
+		messagesToSummarize,
+		turnPrefixMessages,
+		recentMessages,
+		isSplitTurn: cutPoint.isSplitTurn,
+		tokensBefore,
+		previousSummary,
+		previousPreserveData,
+		fileOps,
+		settings,
+	};
+}
+
+// ============================================================================
+// Main compaction function
+// ============================================================================
+
+const TURN_PREFIX_SUMMARIZATION_PROMPT = prompt.render(compactionTurnPrefixPrompt);
+
+/**
+ * Generate summaries for compaction using prepared data.
+ * Returns CompactionResult - SessionManager adds id/parentId when saving.
+ *
+ * @param preparation - Pre-calculated preparation from prepareCompaction()
+ * @param customInstructions - Optional custom focus for the summary
+ */
+export async function compact(
+	preparation: CompactionPreparation,
+	model: Model,
+	apiKey: string,
+	customInstructions?: string,
+	signal?: AbortSignal,
+	options?: SummaryOptions,
+): Promise<CompactionResult> {
+	const {
+		firstKeptEntryId,
+		messagesToSummarize,
+		turnPrefixMessages,
+		recentMessages,
+		isSplitTurn,
+		tokensBefore,
+		previousSummary,
+		previousPreserveData,
+		fileOps,
+		settings,
+	} = preparation;
+
+	const summaryOptions: SummaryOptions = {
+		promptOverride: options?.promptOverride,
+		extraContext: options?.extraContext,
+		remoteEndpoint: settings.remoteEnabled === false ? undefined : settings.remoteEndpoint,
+		remoteInstructions: options?.remoteInstructions,
+		initiatorOverride: options?.initiatorOverride,
+		metadata: options?.metadata,
+		convertToLlm: options?.convertToLlm,
+		telemetry: options?.telemetry,
+		// Honor /model thinking selection on every fan-out summarizer.
+		// Without this propagation, generateSummary / generateTurnPrefixSummary
+		// see options?.thinkingLevel === undefined and resolveCompactionEffort
+		// silently falls back to Effort.High — the same defect e07b47ee4 fixed
+		// at the call sites, leaked back in here. See resolveCompactionEffort.
+		thinkingLevel: options?.thinkingLevel,
+	};
+
+	let preserveData = withOpenAiRemoteCompactionPreserveData(previousPreserveData, undefined);
+	if (settings.remoteEnabled !== false && shouldUseOpenAiRemoteCompaction(model)) {
+		const previousRemoteCompaction = getPreservedOpenAiRemoteCompactionData(previousPreserveData);
+		const remoteMessages = [...messagesToSummarize, ...turnPrefixMessages, ...recentMessages];
+		const previousReplacementHistory =
+			previousRemoteCompaction?.provider === model.provider
+				? previousRemoteCompaction.replacementHistory
+				: undefined;
+		const remoteHistory = buildOpenAiNativeHistory(
+			(summaryOptions.convertToLlm ?? convertToLlm)(remoteMessages),
+			model,
+			previousReplacementHistory,
+		);
+		if (remoteHistory.length > 0) {
+			try {
+				const remote = await requestOpenAiRemoteCompaction(
+					model,
+					apiKey,
+					remoteHistory,
+					summaryOptions.remoteInstructions ?? SUMMARIZATION_SYSTEM_PROMPT,
+					signal,
+				);
+				preserveData = withOpenAiRemoteCompactionPreserveData(previousPreserveData, remote);
+			} catch (err) {
+				logger.warn("OpenAI remote compaction failed, falling back to local summarization", {
+					error: err instanceof Error ? err.message : String(err),
+					model: model.id,
+					provider: model.provider,
+				});
+			}
+		}
+	}
+
+	// Generate summaries (can be parallel if both needed) and merge into one
+	let summary: string;
+
+	if (isSplitTurn && turnPrefixMessages.length > 0) {
+		// Generate both summaries in parallel
+		const [historyResult, turnPrefixResult] = await Promise.all([
+			messagesToSummarize.length > 0
+				? generateSummary(
+						messagesToSummarize,
+						model,
+						settings.reserveTokens,
+						apiKey,
+						signal,
+						customInstructions,
+						previousSummary,
+						summaryOptions,
+					)
+				: Promise.resolve("No prior history."),
+			generateTurnPrefixSummary(turnPrefixMessages, model, settings.reserveTokens, apiKey, signal, summaryOptions),
+		]);
+		// Merge into single summary
+		summary = `${historyResult}\n\n---\n\n**Turn Context (split turn):**\n\n${turnPrefixResult}`;
+	} else if (messagesToSummarize.length > 0) {
+		// Generate history summary from messages to summarize
+		summary = await generateSummary(
+			messagesToSummarize,
+			model,
+			settings.reserveTokens,
+			apiKey,
+			signal,
+			customInstructions,
+			previousSummary,
+			summaryOptions,
+		);
+	} else if (previousSummary) {
+		// No new messages to summarize, preserve previous summary
+		summary = previousSummary;
+	} else {
+		// No messages and no previous summary
+		summary = "No prior history.";
+	}
+
+	const shortSummary = await generateShortSummary(
+		recentMessages,
+		summary,
+		model,
+		settings.reserveTokens,
+		apiKey,
+		signal,
+		{
+			extraContext: options?.extraContext,
+			remoteEndpoint: summaryOptions.remoteEndpoint,
+			initiatorOverride: summaryOptions.initiatorOverride,
+			metadata: summaryOptions.metadata,
+			telemetry: summaryOptions.telemetry,
+			// Same propagation as summaryOptions above — generateShortSummary
+			// resolves its own reasoning via resolveCompactionEffort.
+			thinkingLevel: options?.thinkingLevel,
+		},
+	);
+
+	// Compute file lists and append to summary
+	const { readFiles, modifiedFiles } = computeFileLists(fileOps);
+	summary = upsertFileOperations(summary, readFiles, modifiedFiles);
+
+	if (!firstKeptEntryId) {
+		throw new Error("First kept entry has no ID - session may need migration");
+	}
+
+	return {
+		summary,
+		shortSummary,
+		firstKeptEntryId,
+		tokensBefore,
+		details: { readFiles, modifiedFiles } as CompactionDetails,
+		preserveData,
+	};
+}
+
+/**
+ * Generate a summary for a turn prefix (when splitting a turn).
+ */
+async function generateTurnPrefixSummary(
+	messages: AgentMessage[],
+	model: Model,
+	reserveTokens: number,
+	apiKey: string,
+	signal?: AbortSignal,
+	options?: SummaryOptions,
+): Promise<string> {
+	const maxTokens = Math.floor(0.5 * reserveTokens); // Smaller budget for turn prefix
+
+	const llmMessages = (options?.convertToLlm ?? convertToLlm)(messages);
+	const conversationText = serializeConversation(llmMessages);
+	const promptText = `<conversation>\n${conversationText}\n</conversation>\n\n${TURN_PREFIX_SUMMARIZATION_PROMPT}`;
+	const summarizationMessages = [
+		{
+			role: "user" as const,
+			content: [{ type: "text" as const, text: promptText }],
+			timestamp: Date.now(),
+		},
+	];
+
+	const response = await instrumentedCompleteSimple(
+		model,
+		{ systemPrompt: [SUMMARIZATION_SYSTEM_PROMPT], messages: summarizationMessages },
+		{
+			maxTokens,
+			signal,
+			apiKey,
+			reasoning: resolveCompactionEffort(model, options?.thinkingLevel),
+			initiatorOverride: options?.initiatorOverride,
+			metadata: options?.metadata,
+		},
+		{ telemetry: options?.telemetry, oneshotKind: "compaction_turn_prefix" },
+	);
+
+	if (response.stopReason === "error") {
+		throw createSummarizationError("Turn prefix summarization failed", response);
+	}
+
+	return response.content
+		.filter((c): c is { type: "text"; text: string } => c.type === "text")
+		.map(c => c.text)
+		.join("\n");
+}