@vib-rato/agent-core 0.16.0

package/src/compaction/compaction.ts

@@ -0,0 +1,1569 @@
+/**
+ * Context compaction for long sessions.
+ *
+ * Pure functions for compaction logic. The session manager handles I/O,
+ * and after compaction the session is reloaded.
+ */
+
+import * as os from "node:os";
+import {
+	type AssistantMessage,
+	Effort,
+	type Message,
+	type MessageAttribution,
+	type Model,
+	type ProviderSessionState,
+	type Usage,
+} from "@vib-rato/ai";
+import { logger, prompt } from "@vib-rato/utils";
+import { type AgentTelemetry, instrumentedCompleteSimple } from "../telemetry";
+import type { AgentMessage, AgentTool } from "../types";
+import type { AdaptiveCompactionDecisionState, AdaptiveCompactionOptions } from "./adaptive";
+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 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" | "off";
+	thresholdPercent?: number;
+	thresholdTokens?: number;
+	adaptive?: AdaptiveCompactionOptions;
+	adaptiveState?: AdaptiveCompactionDecisionState;
+	reserveTokens: number;
+	keepRecentTokens: number;
+	autoContinue?: boolean;
+	remoteEnabled?: boolean;
+	remoteEndpoint?: string;
+}
+
+export type RemoteCompactionFallbackHealthEvent =
+	| { kind: "success"; model: string; provider: string }
+	| { kind: "fallback"; model: string; provider: string; error: string };
+
+export interface RemoteCompactionFallbackHealthHooks {
+	recordRemoteCompactionFallback(event: RemoteCompactionFallbackHealthEvent): void;
+}
+
+function isAbortError(error: unknown): boolean {
+	return error instanceof Error && error.name === "AbortError";
+}
+
+export const DEFAULT_COMPACTION_SETTINGS: CompactionSettings = {
+	enabled: true,
+	strategy: "context-full",
+	thresholdPercent: -1,
+	thresholdTokens: -1,
+	reserveTokens: 16384,
+	keepRecentTokens: 20000,
+	autoContinue: true,
+	remoteEnabled: true,
+};
+
+export function computeAdaptiveThresholdPercent(
+	basePercent: number,
+	contextTokens: number,
+	contextWindow: number,
+	state: AdaptiveCompactionDecisionState | undefined,
+	options: AdaptiveCompactionOptions | undefined,
+): number {
+	const clampedBasePercent = Number.isFinite(basePercent) ? Math.min(99, Math.max(1, basePercent)) : 85;
+	if (!options?.enabled) return basePercent;
+	if (!state || !Number.isFinite(contextWindow) || contextWindow <= 0) return clampedBasePercent;
+	if (!Number.isFinite(options.turnWindow) || options.turnWindow <= 0) return clampedBasePercent;
+
+	const safeContextTokens = Number.isFinite(contextTokens) ? Math.max(0, contextTokens) : 0;
+	const fillRatio = safeContextTokens / contextWindow;
+	const baseRatio = clampedBasePercent / 100;
+	if (fillRatio < baseRatio * 0.7) return clampedBasePercent;
+
+	const turnsSinceCompact = Number.isFinite(state.turnsSinceCompact) ? Math.max(0, state.turnsSinceCompact) : 0;
+	const callsInWindow = Number.isFinite(state.callsInWindow) ? Math.max(0, state.callsInWindow) : 0;
+	if (turnsSinceCompact <= 3) return clampedBasePercent;
+	const windowTurns = Math.max(1, options.turnWindow * 4);
+	const intensity = Math.min(1, callsInWindow / windowTurns);
+	const aggression = Number.isFinite(options.aggression) ? Math.min(1, Math.max(0, options.aggression)) : 0;
+	const configuredMinThresholdPercent = options.minThresholdPercent;
+	const minThresholdPercent = Math.min(
+		clampedBasePercent,
+		typeof configuredMinThresholdPercent === "number" && Number.isFinite(configuredMinThresholdPercent)
+			? Math.max(1, configuredMinThresholdPercent)
+			: clampedBasePercent * 0.5,
+	);
+	const loweredPercent = clampedBasePercent - (clampedBasePercent - minThresholdPercent) * aggression * intensity;
+	return Math.max(1, Math.min(99, Math.round(loweredPercent)));
+}
+
+// ============================================================================
+// 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: the largest of 15% of the context window, the configured floor,
+ * and the model's reserved completion budget (`maxOutputTokens`).
+ *
+ * Reserving `maxOutputTokens` keeps the safe input/prompt-packing budget below the
+ * *total* context window for models whose completion reservation exceeds the 15%
+ * floor (e.g. a 400K-context model with 128K max output reserves 128K, not 60K, so
+ * input is capped near 272K instead of 340K).
+ */
+export function effectiveReserveTokens(
+	contextWindow: number,
+	settings: CompactionSettings,
+	maxOutputTokens = 0,
+): number {
+	return Math.max(Math.floor(contextWindow * 0.15), settings.reserveTokens, Math.max(0, maxOutputTokens));
+}
+
+/**
+ * Check if compaction should trigger based on context usage.
+ *
+ * `maxOutputTokens` is the model's reserved completion budget; it is excluded from
+ * the safe input budget so prompt + reserved output cannot exceed the total window.
+ */
+export function shouldCompact(
+	contextTokens: number,
+	contextWindow: number,
+	settings: CompactionSettings,
+	maxOutputTokens = 0,
+): boolean {
+	if (!settings.enabled || settings.strategy === "off" || contextWindow <= 0) return false;
+	const thresholdTokens = resolveThresholdTokens(contextWindow, settings, maxOutputTokens, contextTokens);
+	return contextTokens > thresholdTokens;
+}
+
+/** Reason a compaction was triggered. `token` is the normal user-configurable path; the rest are emergency floors. */
+export type CompactionTriggerReason =
+	| "token"
+	| "heap"
+	| "retainedMemory"
+	| "transcriptFile"
+	| "providerBytes"
+	| "messageCount"
+	| "imageBytes";
+
+/** A point-in-time resource sample. Supplied by an injectable sampler so tests never read real RSS. */
+export interface EmergencyCompactionSample {
+	/** Resident heap bytes (e.g. process.memoryUsage().heapUsed). */
+	heapUsedBytes: number;
+	/** Approximate serialized provider-context bytes. */
+	providerBytes: number;
+	/** Provider-visible message count. */
+	messageCount: number;
+	/** Approximate inline image bytes in the provider context. */
+	imageBytes: number;
+	/** Bytes retained by session resident image sentinels; separate from provider-visible bytes. */
+	sessionResidentImageBytes?: number;
+	/** Bytes retained by non-provider materialized/session-local caches. */
+	materializedResidentBytes?: number;
+	/** Number of live TUI chat-container children. */
+	tuiChatChildren?: number;
+	/** Bytes retained by TUI render caches. */
+	tuiCachedRenderBytes?: number;
+	/** On-disk JSONL transcript file size in bytes; 0/undefined when unknown. */
+	transcriptFileBytes?: number;
+}
+
+export interface EmergencyCompactionLimits {
+	heapUsedBytes: number;
+	providerBytes: number;
+	messageCount: number;
+	imageBytes: number;
+	retainedMemoryBytes?: number;
+	retainedMemoryDiagnosticBytes?: number;
+	tuiChatChildren?: number;
+	tuiChatChildrenDiagnostic?: number;
+	transcriptFileBytes?: number;
+}
+
+const MAX_EMERGENCY_HEAP_FLOOR_BYTES = 1_536 * 1024 * 1024; // 1.5 GiB resident heap
+const EMERGENCY_RETAINED_MEMORY_BYTES = 128 * 1024 * 1024;
+const DIAGNOSTIC_RETAINED_MEMORY_BYTES = 64 * 1024 * 1024;
+const EMERGENCY_TUI_CHAT_CHILDREN = 1000;
+const DIAGNOSTIC_TUI_CHAT_CHILDREN = 700;
+const EMERGENCY_TRANSCRIPT_FILE_BYTES = 48 * 1024 * 1024; // 48 MiB (75% of the 64 MiB managed cap)
+let retainedMemoryDiagnosticActive = false;
+let tuiChatChildrenDiagnosticActive = false;
+
+export function resetEmergencyRetainedMemoryDiagnosticsForTests(): void {
+	retainedMemoryDiagnosticActive = false;
+	tuiChatChildrenDiagnosticActive = false;
+}
+
+export function resolveEmergencyCompactionLimits(totalMemoryBytes: number = os.totalmem()): EmergencyCompactionLimits {
+	// Invalid or non-positive total memory (bad injection, exotic platform)
+	// must never disable the heap floor — fall back to the fixed 1.5 GiB cap.
+	const safeTotal =
+		Number.isFinite(totalMemoryBytes) && totalMemoryBytes > 0 ? totalMemoryBytes : Number.POSITIVE_INFINITY;
+	return {
+		heapUsedBytes: Math.min(MAX_EMERGENCY_HEAP_FLOOR_BYTES, Math.floor(0.5 * safeTotal)),
+		providerBytes: 24 * 1024 * 1024, // 24 MiB serialized provider context
+		messageCount: 4000,
+		imageBytes: 64 * 1024 * 1024, // 64 MiB inline image bytes
+		retainedMemoryBytes: EMERGENCY_RETAINED_MEMORY_BYTES,
+		retainedMemoryDiagnosticBytes: DIAGNOSTIC_RETAINED_MEMORY_BYTES,
+		tuiChatChildren: EMERGENCY_TUI_CHAT_CHILDREN,
+		tuiChatChildrenDiagnostic: DIAGNOSTIC_TUI_CHAT_CHILDREN,
+		transcriptFileBytes: EMERGENCY_TRANSCRIPT_FILE_BYTES,
+	};
+}
+
+/**
+ * Non-disableable emergency floors. These sit well above normal usage and exist so a
+ * long session on weak hardware compacts before OOM even when token-based compaction is
+ * disabled or its threshold is set too high. They are NOT user-tunable down to zero.
+ */
+export const DEFAULT_EMERGENCY_COMPACTION_LIMITS: EmergencyCompactionLimits = resolveEmergencyCompactionLimits();
+
+/**
+ * Returns the first emergency limit exceeded (heap > retainedMemory > transcriptFile > providerBytes > imageBytes > messageCount),
+ * or null when none is. Pure apart from retained-memory diagnostics; the caller routes the result through the
+ * normal pair-safe `compact()` cut logic so a tool_use/tool_result pair is never split.
+ */
+export function emergencyCompactionReason(
+	sample: EmergencyCompactionSample,
+	limits: EmergencyCompactionLimits = resolveEmergencyCompactionLimits(),
+): CompactionTriggerReason | null {
+	const retainedMemoryBytes = (sample.materializedResidentBytes ?? 0) + (sample.tuiCachedRenderBytes ?? 0);
+	const tuiChatChildren = sample.tuiChatChildren ?? 0;
+	const retainedDiagnostic =
+		retainedMemoryBytes >= (limits.retainedMemoryDiagnosticBytes ?? DIAGNOSTIC_RETAINED_MEMORY_BYTES);
+	const childDiagnostic = tuiChatChildren >= (limits.tuiChatChildrenDiagnostic ?? DIAGNOSTIC_TUI_CHAT_CHILDREN);
+	if (retainedDiagnostic && !retainedMemoryDiagnosticActive) {
+		logger.warn("Emergency compaction retained-memory diagnostic threshold crossed", {
+			retainedMemoryBytes,
+			limitBytes: limits.retainedMemoryDiagnosticBytes ?? DIAGNOSTIC_RETAINED_MEMORY_BYTES,
+		});
+	}
+	if (childDiagnostic && !tuiChatChildrenDiagnosticActive) {
+		logger.warn("Emergency compaction TUI chat-child diagnostic threshold crossed", {
+			tuiChatChildren,
+			limit: limits.tuiChatChildrenDiagnostic ?? DIAGNOSTIC_TUI_CHAT_CHILDREN,
+		});
+	}
+	retainedMemoryDiagnosticActive = retainedDiagnostic;
+	tuiChatChildrenDiagnosticActive = childDiagnostic;
+
+	if (sample.heapUsedBytes > limits.heapUsedBytes) return "heap";
+	if (
+		retainedMemoryBytes >= (limits.retainedMemoryBytes ?? EMERGENCY_RETAINED_MEMORY_BYTES) ||
+		tuiChatChildren >= (limits.tuiChatChildren ?? EMERGENCY_TUI_CHAT_CHILDREN)
+	)
+		return "retainedMemory";
+	if (
+		sample.transcriptFileBytes &&
+		sample.transcriptFileBytes > (limits.transcriptFileBytes ?? EMERGENCY_TRANSCRIPT_FILE_BYTES)
+	)
+		return "transcriptFile";
+	if (sample.providerBytes > limits.providerBytes) return "providerBytes";
+	if (sample.imageBytes > limits.imageBytes) return "imageBytes";
+	if (sample.messageCount > limits.messageCount) return "messageCount";
+	return null;
+}
+
+export function resolveThresholdTokens(
+	contextWindow: number,
+	settings: CompactionSettings,
+	maxOutputTokens = 0,
+	contextTokens?: number,
+): 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) {
+		if (!settings.adaptive?.enabled) {
+			return contextWindow - effectiveReserveTokens(contextWindow, settings, maxOutputTokens);
+		}
+		const adaptiveBasePercent = Number.isFinite(settings.adaptive.baseThresholdPercent)
+			? Math.min(99, Math.max(1, settings.adaptive.baseThresholdPercent))
+			: 85;
+		const adaptiveThresholdPercent = computeAdaptiveThresholdPercent(
+			adaptiveBasePercent,
+			adaptiveContextTokens(contextTokens, settings.adaptiveState?.lastContextTokens),
+			contextWindow,
+			settings.adaptiveState,
+			settings.adaptive,
+		);
+		return Math.floor(contextWindow * (adaptiveThresholdPercent / 100));
+	}
+	const clampedThresholdPercent = Math.min(99, Math.max(1, thresholdPercent));
+	const adaptiveThresholdPercent = computeAdaptiveThresholdPercent(
+		settings.adaptive?.baseThresholdPercent ?? clampedThresholdPercent,
+		adaptiveContextTokens(contextTokens, settings.adaptiveState?.lastContextTokens),
+		contextWindow,
+		settings.adaptiveState,
+		settings.adaptive,
+	);
+	const effectiveThresholdPercent = settings.adaptive?.enabled ? adaptiveThresholdPercent : clampedThresholdPercent;
+	return Math.floor(contextWindow * (effectiveThresholdPercent / 100));
+}
+
+function adaptiveContextTokens(contextTokens: number | undefined, lastContextTokens: number | undefined): number {
+	if (contextTokens !== undefined && Number.isFinite(contextTokens)) return Math.max(0, contextTokens);
+	if (typeof lastContextTokens === "number" && Number.isFinite(lastContextTokens))
+		return Math.max(0, lastContextTokens);
+	return 0;
+}
+
+// ============================================================================
+// Cut point detection
+// ============================================================================
+
+/**
+ * Image content has no tokenizer representation; charge a fixed estimate
+ * matching what providers typically bill for inline images.
+ */
+export const IMAGE_TOKEN_ESTIMATE = 1200;
+/**
+ * Estimate tokens for collected message fragments using the native-free
+ * heuristic. Provider usage is the authoritative anchor for context-changing
+ * decisions (see {@link calculatePromptTokens}); this chars/4 estimate covers
+ * only unsent/trailing deltas and per-entry budgeting, and callers add a
+ * conservative inflation factor where threshold safety requires it.
+ */
+function countCollectedMessageFragments(collected: { fragments: string[]; extra: number }): number {
+	return estimateTextTokensHeuristic(collected.fragments) + collected.extra;
+}
+
+/**
+ * Average bytes per token for the cheap heuristic. ~4 bytes/token is the
+ * conventional approximation for English/code text under modern BPE
+ * vocabularies; it intentionally errs slightly low-precision in exchange for
+ * never touching the native tokenizer (and its ~50MB BPE table).
+ */
+const HEURISTIC_BYTES_PER_TOKEN = 4;
+
+/**
+ * Token-dense character weight for the script-aware heuristic.
+ *
+ * Common-BMP CJK blocks (Hangul, unified/compat Han, Kana, CJK punctuation,
+ * full-width forms) tokenize at ~0.6–1.0 tokens per character under
+ * o200k-class BPE vocabularies (measured o200k_base: Hangul prose 0.604,
+ * spaceless Hangul 0.964, Han 0.793, Kana 0.740 tokens/char — versus the
+ * 0.25 the chars/4 heuristic assumes). Each such character is charged 1
+ * token: an upper bound for these measured blocks whose only failure mode is
+ * compacting slightly early, while undercounting risks overflowing the
+ * provider window.
+ *
+ * Surrogate code units are charged 0.5 each, i.e. 1 token per supplementary
+ * code point (supplementary Han extensions, emoji, and other astral chars).
+ * That is a floor rather than an upper bound — rare ideographs and emoji can
+ * cost several tokens — but it is strictly safer than the 0.5-per-pair the
+ * plain chars/4 rule produced.
+ */
+function tokenDenseCharWeight(text: string): { weight: number; units: number } {
+	let weight = 0;
+	let units = 0;
+	for (let i = 0; i < text.length; i++) {
+		const c = text.charCodeAt(i);
+		if (
+			(c >= 0x1100 && c <= 0x11ff) || // Hangul Jamo
+			(c >= 0x3000 && c <= 0x303f) || // CJK symbols & punctuation
+			(c >= 0x3040 && c <= 0x30ff) || // Hiragana & Katakana
+			(c >= 0x3130 && c <= 0x318f) || // Hangul compatibility Jamo
+			(c >= 0x3400 && c <= 0x4dbf) || // CJK ideographs extension A
+			(c >= 0x4e00 && c <= 0x9fff) || // CJK unified ideographs
+			(c >= 0xac00 && c <= 0xd7af) || // Hangul syllables
+			(c >= 0xf900 && c <= 0xfaff) || // CJK compatibility ideographs
+			(c >= 0xff00 && c <= 0xffef) // Half/full-width forms
+		) {
+			weight += 1;
+			units += 1;
+		} else if (c >= 0xd800 && c <= 0xdfff) {
+			// Surrogate half: a supplementary code point contributes two units.
+			weight += 0.5;
+			units += 1;
+		}
+	}
+	return { weight, units };
+}
+
+/**
+ * Script-aware native-free token estimate for a plain string fragment:
+ * token-dense characters cost ~1 token each, everything else chars/4.
+ */
+function estimateFragmentTokensHeuristic(fragment: string): { dense: number; otherChars: number } {
+	const { weight, units } = tokenDenseCharWeight(fragment);
+	return { dense: weight, otherChars: fragment.length - units };
+}
+
+/**
+ * Native-free token estimate for a message. This is the only message
+ * token estimator: provider usage (see {@link calculatePromptTokens}) anchors
+ * the already-sent context, and this covers unsent/trailing deltas, per-entry
+ * budgeting, and display surfaces. Callers add a conservative inflation factor
+ * where compaction-threshold safety requires it.
+ */
+export function estimateMessageTokensHeuristic(message: AgentMessage): number {
+	const { fragments, extra } = collectMessageFragments(message);
+	return extra + estimateTextTokensHeuristic(fragments);
+}
+
+/**
+ * Script-aware native-free token estimate for plain string fragments.
+ * Fragment-level counterpart of {@link estimateMessageTokensHeuristic}.
+ */
+export function estimateTextTokensHeuristic(fragments: string | readonly string[]): number {
+	const list = typeof fragments === "string" ? [fragments] : fragments;
+	let dense = 0;
+	let otherChars = 0;
+	for (const fragment of list) {
+		const counts = estimateFragmentTokensHeuristic(fragment);
+		dense += counts.dense;
+		otherChars += counts.otherChars;
+	}
+	return Math.ceil(dense + Math.max(0, otherChars) / HEURISTIC_BYTES_PER_TOKEN);
+}
+
+/** Shared content walk for both the native and heuristic estimators. */
+function collectMessageFragments(message: AgentMessage): { fragments: string[]; extra: 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, extra };
+	}
+
+	switch (message.role) {
+		case "user":
+		case "custom": {
+			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);
+					// `arguments` is typed non-null, but persisted history can carry a
+					// null/undefined payload from an aborted or malformed tool call;
+					// JSON.stringify returns undefined for those, and the token
+					// fingerprint below requires string fragments.
+					fragments.push(JSON.stringify(block.arguments) ?? "null");
+				}
+			}
+			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:
+			break;
+	}
+
+	return { fragments, extra };
+}
+
+function entryTokenFingerprint(
+	entry: SessionEntry,
+	message: AgentMessage,
+	collected: { fragments: string[]; extra: number },
+): string {
+	const maybePruned = message as { prunedAt?: unknown };
+	let fingerprint = `${entry.type.length}:${entry.type}${(entry.id ?? "").length}:${entry.id ?? ""}${message.role.length}:${message.role}${String(collected.extra).length}:${String(collected.extra)}${collected.fragments.length}:`;
+	for (const fragment of collected.fragments) fingerprint += `${fragment.length}:${fragment}`;
+	if (maybePruned.prunedAt !== undefined) {
+		const prunedAt = String(maybePruned.prunedAt);
+		fingerprint += `prunedAt${prunedAt.length}:${prunedAt}`;
+	}
+	return fingerprint;
+}
+
+const entryTokenCache = new WeakMap<SessionEntry, { fingerprint: string; tokens: number }>();
+
+export function estimateEntryTokens(entry: SessionEntry): number {
+	const msg = getMessageFromEntry(entry);
+	if (!msg) return 0;
+	const collected = collectMessageFragments(msg);
+	const fingerprint = entryTokenFingerprint(entry, msg, collected);
+	const cached = entryTokenCache.get(entry);
+	if (cached?.fingerprint === fingerprint) return cached.tokens;
+	const tokens = countCollectedMessageFragments(collected);
+	entryTokenCache.set(entry, { fingerprint, tokens });
+	return tokens;
+}
+
+export function estimateEntriesTokens(entries: SessionEntry[], startIndex: number, endIndex: number): number {
+	let total = 0;
+	for (let i = startIndex; i < endIndex; i++) {
+		total += estimateEntryTokens(entries[i]);
+	}
+	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];
+		// Estimate this message's size
+		const messageTokens = estimateEntryTokens(entry);
+		accumulatedTokens += messageTokens;
+
+		// Check if we've exceeded the budget
+		if (accumulatedTokens >= keepRecentTokens) {
+			// Find the closest valid cut point at or after this entry
+			let foundCutPoint = false;
+			for (let c = 0; c < cutPoints.length; c++) {
+				if (cutPoints[c] >= i) {
+					cutIndex = cutPoints[c];
+					foundCutPoint = true;
+					break;
+				}
+			}
+			if (!foundCutPoint) {
+				cutIndex = cutPoints[cutPoints.length - 1];
+			}
+			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 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`;
+}
+
+/**
+ * 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
+	 * `pi.gen_ai.oneshot.kind` (`compaction_summary` or `compaction_turn_prefix`).
+	 */
+	telemetry?: AgentTelemetry;
+	authCredentialType?: "api_key" | "oauth";
+	/**
+	 * Provider session affinity id forwarded to the maintenance LLM call so it
+	 * reuses the live turn's provider/WebSocket session (matches the
+	 * `providerSessionId ?? sessionId` the agent loop sends for normal turns).
+	 */
+	sessionId?: string;
+	/** Shared provider state map so maintenance calls reuse session-scoped transport/session caches. */
+	providerSessionState?: Map<string, ProviderSessionState>;
+	/** Hint that websocket transport should be preferred when supported by the provider implementation. */
+	preferWebsockets?: boolean;
+	/** Session-owned health sink for remote-compaction fallback transition logging. */
+	remoteCompactionFallbackHealth?: RemoteCompactionFallbackHealthHooks;
+}
+
+/**
+ * Cap the serialized conversation fed to a summarization request so the request
+ * itself fits inside the model's context window.
+ *
+ * Without this, summarizing a near-full context serializes (nearly) the entire
+ * history back into a single summary request; on strict backends (e.g.
+ * OpenAI-code/Codex `context_length_exceeded`) that request itself overflows and
+ * throws, so context-overflow recovery cannot produce a summary and the agent
+ * fails to compact-and-continue — a non-interactive `vib -p` run then terminates
+ * on the very overflow the recovery was meant to absorb.
+ *
+ * The budget reserves the summary's own output tokens plus prompt/system/template
+ * overhead, and applies a conservative safety factor for estimator error on
+ * dense text (the reason the original overflow was missed).
+ * Truncation keeps the head (origin/goals) and the tail (most recent state) and
+ * elides the middle; it is a last resort that only triggers when the input would
+ * otherwise not fit.
+ */
+export function boundConversationTextForSummary(
+	conversationText: string,
+	model: Model,
+	outputMaxTokens: number,
+): string {
+	const contextWindow = model.contextWindow;
+	if (!Number.isFinite(contextWindow) || contextWindow <= 0) return conversationText;
+
+	const OVERHEAD_TOKENS = 4096;
+	const SAFETY_FACTOR = 0.6;
+	const inputBudgetTokens = Math.floor(
+		(contextWindow - Math.max(0, outputMaxTokens) - OVERHEAD_TOKENS) * SAFETY_FACTOR,
+	);
+	const totalEstimatedTokens = estimateTextTokensHeuristic(conversationText);
+	const assemble = (head: string, tail: string): string => {
+		const elided = conversationText.length - head.length - tail.length;
+		return `${head}\n\n[... ${elided} characters of older conversation elided so this summarization request fits within the model context window ...]\n\n${tail}`;
+	};
+	const bareMarker = assemble("", "");
+	const fitsBudget = (candidate: string) => estimateTextTokensHeuristic(candidate) <= inputBudgetTokens;
+	if (inputBudgetTokens <= 0) {
+		// A window this small cannot fit any excerpt (not even the marker).
+		// Fail closed with an empty excerpt rather than submitting text into a
+		// request that is guaranteed to overflow.
+		return "";
+	}
+	if (totalEstimatedTokens <= inputBudgetTokens) return conversationText;
+
+	// Derive the character budget from the text's own measured token density
+	// instead of assuming 4 chars/token: a CJK-heavy conversation runs near
+	// 1 token/char, and a fixed 4-chars/token cut would overshoot the budget
+	// by up to ~4x — re-overflowing the very request this bound protects.
+	// Verify the complete assembled candidate (elision marker included)
+	// against the estimator and shrink until it fits.
+	const charsPerToken = conversationText.length / totalEstimatedTokens;
+	let budgetChars = Math.floor(inputBudgetTokens * charsPerToken);
+	for (let attempt = 0; attempt < 12 && budgetChars > 0; attempt++) {
+		const headChars = Math.floor(budgetChars * 0.35);
+		const tailChars = Math.max(0, budgetChars - headChars);
+		const head = conversationText.slice(0, headChars);
+		const tail = tailChars > 0 ? conversationText.slice(conversationText.length - tailChars) : "";
+		const assembled = assemble(head, tail);
+		if (fitsBudget(assembled)) return assembled;
+		budgetChars = Math.floor(budgetChars * 0.8);
+	}
+	// All attempts overshot (adversarially non-uniform density, or a budget
+	// smaller than the marker itself). Fail closed: return the bare marker
+	// only when it fits the budget, else an empty excerpt — never an
+	// over-budget result.
+	return fitsBudget(bareMarker) ? bareMarker : "";
+}
+
+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: ${customInstructions}`;
+	}
+
+	// Serialize conversation to text so model doesn't try to continue it
+	// Convert to LLM messages first (handles custom app messages when caller provides a transformer).
+	const llmMessages = (options?.convertToLlm ?? convertToLlm)(currentMessages);
+	const conversationText = boundConversationTextForSummary(serializeConversation(llmMessages), model, maxTokens);
+
+	// Build the prompt with conversation wrapped in tags
+	let promptText = `<conversation>\n${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: Effort.High,
+			initiatorOverride: options?.initiatorOverride,
+			metadata: options?.metadata,
+			sessionId: options?.sessionId,
+			providerSessionState: options?.providerSessionState,
+			preferWebsockets: options?.preferWebsockets,
+		},
+		{ telemetry: options?.telemetry, oneshotKind: "compaction_summary" },
+	);
+
+	if (response.stopReason === "error") {
+		throw new Error(`Summarization failed: ${response.errorMessage || "Unknown error"}`);
+	}
+
+	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;
+	/**
+	 * Optional user-configured extension appended to the base handoff prompt.
+	 * It SUPPLEMENTS the immutable base (safety/continuity structure); it never
+	 * replaces `HANDOFF_DOCUMENT_PROMPT`.
+	 */
+	promptExtension?: 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 `pi.gen_ai.oneshot.kind = "handoff"`.
+	 */
+	telemetry?: AgentTelemetry;
+	authCredentialType?: "api_key" | "oauth";
+	/**
+	 * Provider session affinity id forwarded to the handoff LLM call so it
+	 * reuses the live turn's provider/WebSocket session.
+	 */
+	sessionId?: string;
+	/** Shared provider state map so the handoff call reuses session-scoped transport/session caches. */
+	providerSessionState?: Map<string, ProviderSessionState>;
+	/** Hint that websocket transport should be preferred when supported by the provider implementation. */
+	preferWebsockets?: boolean;
+}
+
+export function renderHandoffPrompt(customInstructions?: string, promptExtension?: string): string {
+	if (!customInstructions && !promptExtension) return HANDOFF_DOCUMENT_PROMPT;
+	return prompt.render(handoffDocumentPrompt, {
+		additionalFocus: customInstructions,
+		promptExtension,
+	});
+}
+
+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, options.promptExtension) }],
+			attribution: "agent",
+			timestamp: Date.now(),
+		},
+	];
+
+	const response = await instrumentedCompleteSimple(
+		model,
+		{
+			systemPrompt: options.systemPrompt,
+			messages: requestMessages,
+			tools: options.tools,
+		},
+		{
+			apiKey,
+			signal,
+			reasoning: Effort.High,
+			toolChoice: "none",
+			initiatorOverride: options.initiatorOverride,
+			metadata: options.metadata,
+			sessionId: options.sessionId,
+			providerSessionState: options.providerSessionState,
+			preferWebsockets: options.preferWebsockets,
+		},
+		{ telemetry: options.telemetry, oneshotKind: "handoff" },
+	);
+
+	if (response.stopReason === "error") {
+		throw new Error(`Handoff generation failed: ${response.errorMessage || "Unknown error"}`);
+	}
+
+	return response.content
+		.filter((c): c is { type: "text"; text: string } => c.type === "text")
+		.map(c => c.text)
+		.join("\n");
+}
+
+/** Derive a display summary locally to avoid a second compaction LLM request. */
+function deriveShortSummary(summary: string): string {
+	const firstParagraph = summary.trim().split(/\n\s*\n/, 1)[0] ?? "";
+	const maxLength = 2_000;
+	return firstParagraph.length <= maxLength ? firstParagraph : `${firstParagraph.slice(0, maxLength - 1)}…`;
+}
+
+// ============================================================================
+// 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;
+	/**
+	 * Diagnostics for the keep-window token correction (Finding 7). `ratio` is the
+	 * clamped heuristic→actual correction that was applied (1 when none supplied);
+	 * `keepRecentTokensCorrected` is the heuristic budget findCutPoint actually used.
+	 */
+	tokenCorrection: { ratio: number; keepRecentTokensCorrected: number };
+}
+
+/** Bounds for the keep-window token correction (Finding 7): never trust a ratio
+ * beyond 2x in either direction so a bad estimate cannot balloon or collapse the
+ * kept window. */
+export const TOKEN_CORRECTION_MIN_RATIO = 0.5;
+export const TOKEN_CORRECTION_MAX_RATIO = 2;
+
+export interface PrepareCompactionOptions {
+	/**
+	 * Observed heuristic→actual token correction for the post-boundary keep window
+	 * (actualTokens / chars-4-heuristicTokens), supplied by the caller from per-turn
+	 * Usage deltas or a stable-prefix-subtracted comparison. Clamped to
+	 * [0.5, 2] and applied bidirectionally. When omitted, no correction is applied
+	 * (the confounded raw promptTokens/estimatedTokens quotient is never used).
+	 */
+	tokenCorrectionRatio?: number;
+	/**
+	 * Model context-window size. Windows below 66k retain the legacy fixed
+	 * keepRecentTokens behavior; larger windows scale the keep window to 30%.
+	 */
+	contextWindow?: number;
+}
+
+export function prepareCompaction(
+	pathEntries: SessionEntry[],
+	settings: CompactionSettings,
+	options: PrepareCompactionOptions = {},
+): 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;
+
+	// Correct the keep-window budget for the chars/4 heuristic error using the
+	// caller-supplied observed ratio (actual/heuristic). The legacy raw
+	// promptTokens/estimatedTokens quotient is intentionally NOT used: promptTokens
+	// counts system+tools+full history while estimatedTokens counted only the
+	// post-boundary slice, so it was confounded and only ever shrank the window.
+	// Here the correction is bidirectional and clamped to [0.5, 2].
+	const configuredKeepRecentTokens = settings.keepRecentTokens;
+	const contextWindow = options.contextWindow;
+	const thresholdSafeKeepRecentTokens =
+		contextWindow !== undefined && Number.isFinite(contextWindow) && contextWindow > 1
+			? Math.max(
+					1,
+					resolveThresholdTokens(contextWindow, settings) - effectiveReserveTokens(contextWindow, settings, 0),
+				)
+			: configuredKeepRecentTokens;
+	const keepRecentTokens = Math.min(configuredKeepRecentTokens, thresholdSafeKeepRecentTokens);
+	// Preserve the legacy fixed window for smaller models. At 66k and above,
+	// retain up to 30% of the model context, but never enough to leave the
+	// post-compaction prompt immediately above its configured threshold.
+	const scaledKeepRecentTokens =
+		contextWindow !== undefined && Number.isFinite(contextWindow) && contextWindow >= 66_000
+			? Math.min(thresholdSafeKeepRecentTokens, Math.max(keepRecentTokens, Math.floor(contextWindow * 0.3)))
+			: keepRecentTokens;
+	const rawRatio = options.tokenCorrectionRatio;
+	const appliedRatio =
+		rawRatio !== undefined && Number.isFinite(rawRatio) && rawRatio > 0
+			? Math.min(TOKEN_CORRECTION_MAX_RATIO, Math.max(TOKEN_CORRECTION_MIN_RATIO, rawRatio))
+			: 1;
+	// Preserve an explicit keep floor that already covers the whole history: manual
+	// and emergency callers rely on prepareCompaction returning undefined rather
+	// than manufacturing a summary with no useful reduction. Otherwise, a scaled
+	// window that exceeds a short history falls back to the threshold-safe floor.
+	const historyTokens = pathEntries
+		.slice(boundaryStart, boundaryEnd)
+		.reduce((tokens, entry) => tokens + estimateEntryTokens(entry), 0);
+	const effectiveKeepRecentTokens =
+		configuredKeepRecentTokens > historyTokens
+			? configuredKeepRecentTokens
+			: scaledKeepRecentTokens > keepRecentTokens && scaledKeepRecentTokens > historyTokens
+				? keepRecentTokens
+				: scaledKeepRecentTokens;
+	const keepRecentTokensCorrected = Math.max(1, Math.round(effectiveKeepRecentTokens / appliedRatio));
+
+	const cutPoint = findCutPoint(pathEntries, boundaryStart, boundaryEnd, keepRecentTokensCorrected);
+
+	// 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,
+		tokenCorrection: { ratio: appliedRatio, keepRecentTokensCorrected },
+	};
+}
+
+// ============================================================================
+// 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,
+		sessionId: options?.sessionId,
+		providerSessionState: options?.providerSessionState,
+		preferWebsockets: options?.preferWebsockets,
+		remoteCompactionFallbackHealth: options?.remoteCompactionFallbackHealth,
+	};
+
+	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,
+					{ authCredentialType: options?.authCredentialType },
+				);
+				preserveData = withOpenAiRemoteCompactionPreserveData(previousPreserveData, remote);
+				summaryOptions.remoteCompactionFallbackHealth?.recordRemoteCompactionFallback({
+					kind: "success",
+					model: model.id,
+					provider: model.provider,
+				});
+			} catch (err) {
+				if (signal?.aborted || isAbortError(err)) throw err;
+				const error = err instanceof Error ? err.message : String(err);
+				if (summaryOptions.remoteCompactionFallbackHealth) {
+					summaryOptions.remoteCompactionFallbackHealth.recordRemoteCompactionFallback({
+						kind: "fallback",
+						error,
+						model: model.id,
+						provider: model.provider,
+					});
+				} else {
+					logger.warn("OpenAI remote compaction failed, falling back to local summarization", {
+						error,
+						model: model.id,
+						provider: model.provider,
+					});
+				}
+			}
+		}
+	}
+
+	// Generate summaries (can be parallel if both needed) and merge into one
+	let summary: string;
+
+	// A single active Codex WebSocket session cannot service two concurrent
+	// requests ("websocket request already in progress"). When the maintenance
+	// calls use the Codex Responses provider, share one provider session, and
+	// websocket transport is not explicitly disabled, run the split-turn history
+	// and turn-prefix summaries sequentially. This covers websocket activation
+	// from config/env/model defaults too: the provider can select websockets even
+	// when `preferWebsockets` is undefined, while non-Codex providers keep the
+	// previous parallel behavior.
+	const summariesMayShareWebSocketSession = Boolean(
+		model.api === "openai-codex-responses" &&
+			summaryOptions.providerSessionState &&
+			summaryOptions.preferWebsockets !== false,
+	);
+
+	if (isSplitTurn && turnPrefixMessages.length > 0) {
+		const runHistorySummary = () =>
+			messagesToSummarize.length > 0
+				? generateSummary(
+						messagesToSummarize,
+						model,
+						settings.reserveTokens,
+						apiKey,
+						signal,
+						customInstructions,
+						previousSummary,
+						summaryOptions,
+					)
+				: Promise.resolve("No prior history.");
+		const runTurnPrefixSummary = () =>
+			generateTurnPrefixSummary(turnPrefixMessages, model, settings.reserveTokens, apiKey, signal, summaryOptions);
+
+		let historyResult: string;
+		let turnPrefixResult: string;
+		if (summariesMayShareWebSocketSession) {
+			// Sequential: avoids concurrent requests on the same provider session.
+			historyResult = await runHistorySummary();
+			turnPrefixResult = await runTurnPrefixSummary();
+		} else {
+			[historyResult, turnPrefixResult] = await Promise.all([runHistorySummary(), runTurnPrefixSummary()]);
+		}
+		// 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.";
+	}
+
+	// Compute file lists and append to summary
+	const { readFiles, modifiedFiles } = computeFileLists(fileOps);
+	summary = upsertFileOperations(summary, readFiles, modifiedFiles);
+	const shortSummary = deriveShortSummary(summary);
+
+	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 = boundConversationTextForSummary(serializeConversation(llmMessages), model, maxTokens);
+	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: Effort.High,
+			initiatorOverride: options?.initiatorOverride,
+			metadata: options?.metadata,
+			sessionId: options?.sessionId,
+			providerSessionState: options?.providerSessionState,
+			preferWebsockets: options?.preferWebsockets,
+		},
+		{ telemetry: options?.telemetry, oneshotKind: "compaction_turn_prefix" },
+	);
+
+	if (response.stopReason === "error") {
+		throw new Error(`Turn prefix summarization failed: ${response.errorMessage || "Unknown error"}`);
+	}
+
+	return response.content
+		.filter((c): c is { type: "text"; text: string } => c.type === "text")
+		.map(c => c.text)
+		.join("\n");
+}