@sayknow-cli/agent-core 0.3.16 → 0.4.0

package/dist/types/compaction/compaction.d.ts

@@ -1,298 +0,0 @@
-/**
- * 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 MessageAttribution, type Model, type ProviderSessionState, type Usage } from "@sayknow-cli/ai";
-import { type AgentTelemetry } from "../telemetry";
-import type { AgentMessage, AgentTool } from "../types";
-import type { SessionEntry } from "./entries";
-import { type ConvertToLlm } from "./messages";
-import { type FileOperations } from "./utils";
-/** Details stored in CompactionEntry.details for file tracking */
-export interface CompactionDetails {
-    readFiles: string[];
-    modifiedFiles: string[];
-}
-/** 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>;
-}
-export interface CompactionSettings {
-    enabled: boolean;
-    strategy?: "context-full" | "handoff" | "off";
-    thresholdPercent?: number;
-    thresholdTokens?: number;
-    reserveTokens: number;
-    keepRecentTokens: number;
-    autoContinue?: boolean;
-    remoteEnabled?: boolean;
-    remoteEndpoint?: string;
-}
-export declare const DEFAULT_COMPACTION_SETTINGS: CompactionSettings;
-/**
- * Calculate total context tokens from usage.
- * Uses the native totalTokens field when available, falls back to computing from components.
- */
-export declare function calculateContextTokens(usage: Usage): number;
-export declare function calculatePromptTokens(usage: Usage): number;
-/**
- * Find the last non-aborted assistant message usage from session entries.
- */
-export declare function getLastAssistantUsage(entries: SessionEntry[]): Usage | 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 declare function effectiveReserveTokens(contextWindow: number, settings: CompactionSettings, maxOutputTokens?: number): number;
-/**
- * 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 declare function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings, maxOutputTokens?: number): boolean;
-/** Reason a compaction was triggered. `token` is the normal user-configurable path; the rest are emergency floors. */
-export type CompactionTriggerReason = "token" | "heap" | "retainedMemory" | "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;
-}
-export interface EmergencyCompactionLimits {
-    heapUsedBytes: number;
-    providerBytes: number;
-    messageCount: number;
-    imageBytes: number;
-    retainedMemoryBytes?: number;
-    retainedMemoryDiagnosticBytes?: number;
-    tuiChatChildren?: number;
-    tuiChatChildrenDiagnostic?: number;
-}
-export declare function resetEmergencyRetainedMemoryDiagnosticsForTests(): void;
-export declare function resolveEmergencyCompactionLimits(totalMemoryBytes?: number): EmergencyCompactionLimits;
-/**
- * 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 declare const DEFAULT_EMERGENCY_COMPACTION_LIMITS: EmergencyCompactionLimits;
-/**
- * Returns the first emergency limit exceeded (heap > retainedMemory > 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 declare function emergencyCompactionReason(sample: EmergencyCompactionSample, limits?: EmergencyCompactionLimits): CompactionTriggerReason | null;
-export declare function resolveThresholdTokens(contextWindow: number, settings: CompactionSettings, maxOutputTokens?: number): number;
-/**
- * Image content has no tokenizer representation; charge a fixed estimate
- * matching what providers typically bill for inline images.
- */
-export declare const IMAGE_TOKEN_ESTIMATE = 1200;
-/**
- * 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 declare function estimateMessageTokensHeuristic(message: AgentMessage): number;
-/**
- * Script-aware native-free token estimate for plain string fragments.
- * Fragment-level counterpart of {@link estimateMessageTokensHeuristic}.
- */
-export declare function estimateTextTokensHeuristic(fragments: string | readonly string[]): number;
-export declare function estimateEntryTokens(entry: SessionEntry): number;
-export declare function estimateEntriesTokens(entries: SessionEntry[], startIndex: number, endIndex: number): number;
-/**
- * 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 declare function findTurnStartIndex(entries: SessionEntry[], entryIndex: number, startIndex: number): number;
-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 declare function findCutPoint(entries: SessionEntry[], startIndex: number, endIndex: number, keepRecentTokens: number): CutPointResult;
-export declare const AUTO_HANDOFF_THRESHOLD_FOCUS: string;
-/**
- * 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`, `compaction_short_summary`,
-     * or `compaction_turn_prefix`). `undefined` keeps the call paths zero-cost.
-     */
-    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;
-}
-/**
- * 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 `skc -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 declare function boundConversationTextForSummary(conversationText: string, model: Model, outputMaxTokens: number): string;
-export declare function generateSummary(currentMessages: AgentMessage[], model: Model, reserveTokens: number, apiKey: string, signal?: AbortSignal, customInstructions?: string, previousSummary?: string, options?: SummaryOptions): Promise<string>;
-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 `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 declare function renderHandoffPrompt(customInstructions?: string): string;
-export declare function generateHandoff(messages: AgentMessage[], model: Model, apiKey: string, options: HandoffOptions, signal?: AbortSignal): Promise<string>;
-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 declare const TOKEN_CORRECTION_MIN_RATIO = 0.5;
-export declare 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;
-}
-export declare function prepareCompaction(pathEntries: SessionEntry[], settings: CompactionSettings, options?: PrepareCompactionOptions): CompactionPreparation | undefined;
-/**
- * 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 declare function compact(preparation: CompactionPreparation, model: Model, apiKey: string, customInstructions?: string, signal?: AbortSignal, options?: SummaryOptions): Promise<CompactionResult>;

package/dist/types/compaction/entries.d.ts

@@ -1,109 +0,0 @@
-import type { ImageContent, MessageAttribution, ServiceTier, TextContent } from "@sayknow-cli/ai";
-import type { AgentMessage } from "../types";
-export interface SessionEntryBase {
-    type: string;
-    id: string;
-    parentId: string | null;
-    timestamp: string;
-}
-export interface SessionMessageEntry extends SessionEntryBase {
-    type: "message";
-    message: AgentMessage;
-}
-export interface ThinkingLevelChangeEntry extends SessionEntryBase {
-    type: "thinking_level_change";
-    thinkingLevel?: string | null;
-}
-export interface ModelChangeEntry extends SessionEntryBase {
-    type: "model_change";
-    /** Model in "provider/modelId" format */
-    model: string;
-    /** Role: "default", "smol", "slow", etc. Undefined treated as "default" */
-    role?: string;
-    /** Requested model before a runtime substitution/fallback, in "provider/modelId" format. */
-    previousModel?: string;
-    /** Machine-readable reason for runtime model substitution/fallback. */
-    reason?: string;
-    /** Effective thinking level when the change was recorded. */
-    thinkingLevel?: string | null;
-}
-export interface ServiceTierChangeEntry extends SessionEntryBase {
-    type: "service_tier_change";
-    serviceTier: ServiceTier | null;
-}
-export interface CompactionEntry<T = unknown> extends SessionEntryBase {
-    type: "compaction";
-    summary: string;
-    shortSummary?: string;
-    firstKeptEntryId: string;
-    tokensBefore: number;
-    /** Extension-specific data (e.g., ArtifactIndex, version markers for structured compaction) */
-    details?: T;
-    /** Hook-provided data to persist across compaction */
-    preserveData?: Record<string, unknown>;
-    /** True if generated by an extension, undefined/false if pi-generated (backward compatible) */
-    fromExtension?: boolean;
-}
-export interface BranchSummaryEntry<T = unknown> extends SessionEntryBase {
-    type: "branch_summary";
-    fromId: string;
-    summary: string;
-    /** Extension-specific data (not sent to LLM) */
-    details?: T;
-    /** True if generated by an extension, false if pi-generated */
-    fromExtension?: boolean;
-}
-export interface CustomMessageEntry<T = unknown> extends SessionEntryBase {
-    type: "custom_message";
-    customType: string;
-    content: string | (TextContent | ImageContent)[];
-    details?: T;
-    display: boolean;
-    /** Who initiated this message for billing/attribution semantics. */
-    attribution?: MessageAttribution;
-}
-export interface CustomEntry<T = unknown> extends SessionEntryBase {
-    type: "custom";
-    customType: string;
-    data?: T;
-}
-export interface LabelEntry extends SessionEntryBase {
-    type: "label";
-    targetId: string;
-    label: string | undefined;
-}
-export interface TtsrInjectionEntry extends SessionEntryBase {
-    type: "ttsr_injection";
-    /** Names of rules that were injected */
-    injectedRules: string[];
-}
-export interface MCPToolSelectionEntry extends SessionEntryBase {
-    type: "mcp_tool_selection";
-    /** MCP tool names selected for visibility in discovery mode. */
-    selectedToolNames: string[];
-}
-export interface SessionInitEntry extends SessionEntryBase {
-    type: "session_init";
-    /** Full system prompt sent to the model */
-    systemPrompt: string;
-    /** Initial task/user message */
-    task: string;
-    /** Tools available to the agent */
-    tools: string[];
-    /** Output schema if structured output was requested */
-    outputSchema?: unknown;
-}
-export interface ModeChangeEntry extends SessionEntryBase {
-    type: "mode_change";
-    /** Current mode name, or "none" when exiting a mode */
-    mode: string;
-    /** Optional mode-specific data (e.g. plan file path) */
-    data?: Record<string, unknown>;
-}
-export interface CustomCompactionSessionEntries {
-}
-export type SessionEntry = SessionMessageEntry | ThinkingLevelChangeEntry | ModelChangeEntry | ServiceTierChangeEntry | CompactionEntry | BranchSummaryEntry | CustomEntry | CustomMessageEntry | LabelEntry | TtsrInjectionEntry | MCPToolSelectionEntry | SessionInitEntry | ModeChangeEntry | CustomCompactionSessionEntries[keyof CustomCompactionSessionEntries];
-export interface ReadonlySessionManager {
-    getBranch(leafId?: string | null): SessionEntry[];
-    getEntry(id: string): SessionEntry | undefined;
-}

package/dist/types/compaction/errors.d.ts

@@ -1,26 +0,0 @@
-/**
- * Compaction error types.
- *
- * `CompactionCancelledError` is the canonical signal raised when a compaction
- * is explicitly aborted — operator Esc, extension hook returning `cancel`,
- * programmatic `session.abortCompaction()` call, or any other deliberate
- * abort source. Downstream callers (e.g. `executeCompaction`) discriminate
- * cancellation from other failures via `instanceof CompactionCancelledError`
- * rather than introspecting error messages or `name` fields — the typed
- * sentinel makes classification source-agnostic and refactor-stable.
- */
-export declare class CompactionCancelledError extends Error {
-    readonly name: "CompactionCancelledError";
-    constructor(message?: string);
-}
-/**
- * Outcome of a compaction attempt, surfaced by `CommandController.executeCompaction`
- * so callers (e.g. the plan-mode approval flow) can distinguish a deliberate abort
- * from an unrelated failure.
- *
- *   "ok"        — compaction completed; transcript was summarized.
- *   "cancelled" — `CompactionCancelledError` was raised. Operator Esc, extension
- *                 hook, programmatic abort — all source-agnostic.
- *   "failed"    — any other rejection from `session.compact()`.
- */
-export type CompactionOutcome = "ok" | "cancelled" | "failed";

package/dist/types/compaction/index.d.ts

@@ -1,11 +0,0 @@
-/**
- * Compaction and summarization utilities.
- */
-export * from "./branch-summarization";
-export * from "./compaction";
-export * from "./entries";
-export * from "./errors";
-export * from "./messages";
-export * from "./openai";
-export * from "./pruning";
-export * from "./utils";

package/dist/types/compaction/messages.d.ts

@@ -1,61 +0,0 @@
-import type { ImageContent, Message, MessageAttribution, ProviderPayload, TextContent } from "@sayknow-cli/ai";
-import type { AgentMessage } from "../types";
-export interface CustomMessage<T = unknown> {
-    role: "custom";
-    customType: string;
-    content: string | (TextContent | ImageContent)[];
-    display: boolean;
-    details?: T;
-    /** Who initiated this message for billing/attribution semantics. */
-    attribution?: MessageAttribution;
-    timestamp: number;
-}
-/** Legacy hook message type (pre-extensions). Kept for session migration. */
-export interface HookMessage<T = unknown> {
-    role: "hookMessage";
-    customType: string;
-    content: string | (TextContent | ImageContent)[];
-    display: boolean;
-    details?: T;
-    /** Who initiated this message for billing/attribution semantics. */
-    attribution?: MessageAttribution;
-    timestamp: number;
-}
-export interface BranchSummaryMessage {
-    role: "branchSummary";
-    summary: string;
-    fromId: string;
-    timestamp: number;
-}
-export interface CompactionSummaryMessage {
-    role: "compactionSummary";
-    summary: string;
-    shortSummary?: string;
-    tokensBefore: number;
-    providerPayload?: ProviderPayload;
-    timestamp: number;
-}
-export type CoreCompactionMessage = CustomMessage | HookMessage | BranchSummaryMessage | CompactionSummaryMessage;
-declare module "../types" {
-    interface CustomAgentMessages {
-        custom: CustomMessage;
-        hookMessage: HookMessage;
-        branchSummary: BranchSummaryMessage;
-        compactionSummary: CompactionSummaryMessage;
-    }
-}
-export type ConvertToLlm = (messages: AgentMessage[]) => Message[];
-export declare function renderBranchSummaryContext(summary: string): string;
-export declare function renderCompactionSummaryContext(summary: string): string;
-export declare function createBranchSummaryMessage(summary: string, fromId: string, timestamp: string): BranchSummaryMessage;
-export declare function createCompactionSummaryMessage(summary: string, tokensBefore: number, timestamp: string, shortSummary?: string, providerPayload?: ProviderPayload): CompactionSummaryMessage;
-export declare function createCustomMessage(customType: string, content: string | (TextContent | ImageContent)[], display: boolean, details: unknown | undefined, timestamp: string, attribution?: MessageAttribution): CustomMessage;
-/**
- * Default compaction-domain transformer.
- *
- * Embedders with their own app messages should pass a richer transformer through
- * `SummaryOptions.convertToLlm`; this default intentionally preserves only the
- * core LLM roles and the compaction messages owned by this package.
- */
-export declare function defaultConvertToLlm(messages: AgentMessage[]): Message[];
-export declare const convertToLlm: typeof defaultConvertToLlm;

package/dist/types/compaction/openai.d.ts

@@ -1,63 +0,0 @@
-/**
- * Remote compaction utilities.
- *
- * Provider-side conversation summarization endpoints. Two flavors:
- *
- * - **OpenAI remote compaction** (`/responses/compact`): preserves encrypted
- *   reasoning across compactions by submitting the full responses-API native
- *   history and storing the returned `compaction` / `compaction_summary`
- *   item in `preserveData` so future turns can replay the encrypted state.
- * - **Generic remote compaction**: a thin POST helper for self-hosted
- *   summarization endpoints that accept `{ systemPrompt, prompt }` and reply
- *   with `{ summary, shortSummary? }`.
- */
-import type { Message, Model } from "@sayknow-cli/ai/types";
-export declare const OPENAI_REMOTE_COMPACTION_PRESERVE_KEY = "openaiRemoteCompaction";
-export type OpenAiRemoteCompactionItem = {
-    type: "compaction" | "compaction_summary";
-    encrypted_content?: string;
-    summary?: string;
-};
-export interface OpenAiRemoteCompactionPreserveData {
-    provider?: string;
-    replacementHistory: Array<Record<string, unknown>>;
-    compactionItem: OpenAiRemoteCompactionItem;
-}
-export interface OpenAiRemoteCompactionRequest {
-    model: string;
-    input: Array<Record<string, unknown>>;
-    instructions: string;
-}
-export interface OpenAiRemoteCompactionResponse extends OpenAiRemoteCompactionPreserveData {
-}
-export interface RemoteCompactionRequest {
-    systemPrompt: string;
-    prompt: string;
-}
-export interface RemoteCompactionResponse {
-    summary: string;
-    shortSummary?: string;
-}
-export declare function shouldUseOpenAiRemoteCompaction(model: Model): boolean;
-export declare function getPreservedOpenAiRemoteCompactionData(preserveData: Record<string, unknown> | undefined): OpenAiRemoteCompactionPreserveData | undefined;
-export declare function withOpenAiRemoteCompactionPreserveData(preserveData: Record<string, unknown> | undefined, remoteCompaction: OpenAiRemoteCompactionPreserveData | undefined): Record<string, unknown> | undefined;
-export declare function estimateOpenAiCompactInputTokens(input: Array<Record<string, unknown>>, instructions: string): number;
-export declare function trimOpenAiCompactInput(input: Array<Record<string, unknown>>, contextWindow: number, instructions: string): Array<Record<string, unknown>>;
-export declare function resolveOpenAiCompactInputBudget(contextWindow: number, maxOutputTokens?: number): number;
-/**
- * Build the OpenAI Responses-API native history array from LLM messages.
- *
- * Caller is responsible for converting any custom message types to
- * `Message[]` first (e.g. via the agent's `convertToLlm`); this function
- * operates purely on the LLM-domain shape.
- *
- * @param messages - LLM messages to encode.
- * @param model - Target model (used for provider gating + tool-call id rules).
- * @param previousReplacementHistory - History from a prior compaction whose
- *   encrypted reasoning we want to preserve.
- */
-export declare function buildOpenAiNativeHistory(messages: Message[], model: Model, previousReplacementHistory?: Array<Record<string, unknown>>): Array<Record<string, unknown>>;
-export declare function requestOpenAiRemoteCompaction(model: Model, apiKey: string, compactInput: Array<Record<string, unknown>>, instructions: string, signal?: AbortSignal, options?: {
-    authCredentialType?: "api_key" | "oauth";
-}): Promise<OpenAiRemoteCompactionResponse>;
-export declare function requestRemoteCompaction(endpoint: string, request: RemoteCompactionRequest, signal?: AbortSignal): Promise<RemoteCompactionResponse>;

package/dist/types/compaction/pruning.d.ts

@@ -1,69 +0,0 @@
-/**
- * Tool output pruning utilities for compaction.
- *
- * Candidate selection is staleness-aware: tool results that have been
- * superseded by a later result for the same target (same file read again,
- * same search re-run) or invalidated by a later successful edit/write to a
- * covered file are pruned in preference to merely-old results. Protect-window
- * and minimum-savings hysteresis semantics are unchanged.
- */
-import type { SessionEntry, SessionMessageEntry } from "./entries";
-export interface PruneConfig {
-    /** Keep the most recent tool output tokens intact. */
-    protectTokens: number;
-    /** Only prune if total savings meets this threshold. */
-    minimumSavings: number;
-    /** Tool names that should never be pruned. */
-    protectedTools: string[];
-    /**
-     * Tools in `protectedTools` whose protection is waived once the result is
-     * superseded (a later result for the same target, or a later successful
-     * edit/write to the covered file). The most recent result per target is
-     * never considered superseded. Optional; defaults to none.
-     */
-    staleOverridableTools?: string[];
-}
-export declare const DEFAULT_PRUNE_CONFIG: PruneConfig;
-export interface PruneResult {
-    prunedCount: number;
-    tokensSaved: number;
-    /**
-     * The mutated message entries. Callers whose entry source returns
-     * materialized copies (not live references) must write these back into
-     * their canonical store by id.
-     */
-    prunedEntries: SessionMessageEntry[];
-}
-export interface AssistantArgumentPruneResult {
-    argumentPrunedCount: number;
-    argumentTokensSaved: number;
-    /**
-     * The mutated assistant message entries. Callers whose entry source returns
-     * materialized copies must write these back into their canonical store by id.
-     */
-    prunedEntries: SessionMessageEntry[];
-}
-export declare function pruneAssistantToolArguments(entries: SessionEntry[], config?: PruneConfig): AssistantArgumentPruneResult;
-/**
- * Estimate the token savings {@link pruneToolOutputs} would achieve, without
- * mutating any entry. Returns 0 savings when below the configured minimum so the
- * caller sees the same gate the real prune enforces.
- */
-export declare function estimateToolOutputPruneSavings(entries: SessionEntry[], config?: PruneConfig): {
-    prunableCount: number;
-    tokensSaved: number;
-};
-/**
- * Evidence gate for below-threshold maintenance pruning (Finding 13). Pruning
- * forces a prompt-cache-epoch reset, so it only runs when opted in AND the
- * estimated stale savings clear a high minimum AND exceed the one-time reset
- * cost (so the reclaim pays the reset back). Default-off/blocked until live
- * evidence justifies enabling.
- */
-export declare function shouldRunMaintenancePrune(args: {
-    enabled: boolean;
-    estimatedSavings: number;
-    minSavings: number;
-    cacheEpochResetCost: number;
-}): boolean;
-export declare function pruneToolOutputs(entries: SessionEntry[], config?: PruneConfig): PruneResult;

package/dist/types/compaction/utils.d.ts

@@ -1,32 +0,0 @@
-/**
- * Shared utilities for compaction and branch summarization.
- */
-import type { Message } from "@sayknow-cli/ai";
-import type { AgentMessage } from "../types";
-export interface FileOperations {
-    read: Set<string>;
-    written: Set<string>;
-    edited: Set<string>;
-}
-export declare function createFileOps(): FileOperations;
-/**
- * Extract file operations from tool calls in an assistant message.
- */
-export declare function extractFileOpsFromMessage(message: AgentMessage, fileOps: FileOperations): void;
-/**
- * Compute final file lists from file operations.
- * Returns readFiles (files only read, not modified) and modifiedFiles.
- */
-export declare function computeFileLists(fileOps: FileOperations): {
-    readFiles: string[];
-    modifiedFiles: string[];
-};
-export declare function formatFileOperations(readFiles: string[], modifiedFiles: string[]): string;
-export declare function upsertFileOperations(summary: string, readFiles: string[], modifiedFiles: string[]): string;
-/**
- * Serialize LLM messages to text for summarization.
- * This prevents the model from treating it as a conversation to continue.
- * Call convertToLlm() first to handle custom message types.
- */
-export declare function serializeConversation(messages: Message[]): string;
-export declare const SUMMARIZATION_SYSTEM_PROMPT: string;