jeopi-agent-core 16.2.13

package/dist/types/agent.d.ts

@@ -0,0 +1,427 @@
+import { type ApiKey, type AssistantMessage, type AssistantMessageEvent, type Context, type CursorExecHandlers, type CursorToolResultHandler, type Effort, type ImageContent, type Message, type Model, type ProviderSessionState, type ServiceTier, type SimpleStreamOptions, type ThinkingBudgets, type ToolChoice } from "jeopi-ai";
+import type { Dialect } from "jeopi-ai/dialect";
+import type { HarmonyAuditEvent } from "jeopi-ai/utils/harmony-leak";
+import type { AppendOnlyContextManager } from "./append-only-context";
+import type { AgentEvent, AgentLoopConfig, AgentMessage, AgentState, AgentTool, AgentToolContext, AgentTurnEndContext, AsideMessage, StreamFn, ToolCallContext, ToolChoiceDirective } from "./types";
+export declare class AgentBusyError extends Error {
+    constructor(message?: string);
+}
+export interface AgentOptions {
+    initialState?: Partial<AgentState>;
+    /**
+     * Converts AgentMessage[] to LLM-compatible Message[] before each LLM call.
+     * Default filters to user/assistant/toolResult and converts attachments.
+     */
+    convertToLlm?: (messages: AgentMessage[]) => Message[] | Promise<Message[]>;
+    /**
+     * Optional transform applied to context before convertToLlm.
+     * Use for context pruning, injecting external context, etc.
+     */
+    transformContext?: (messages: AgentMessage[], signal?: AbortSignal) => Promise<AgentMessage[]>;
+    /**
+     * Optional transform applied after provider context assembly and before
+     * telemetry capture/provider send.
+     */
+    transformProviderContext?: (context: Context, model: Model) => Context | Promise<Context>;
+    /**
+     * Steering mode: "all" = send all steering messages at once, "one-at-a-time" = one per turn
+     */
+    steeringMode?: "all" | "one-at-a-time";
+    /**
+     * Follow-up mode: "all" = send all follow-up messages at once, "one-at-a-time" = one per turn
+     */
+    followUpMode?: "all" | "one-at-a-time";
+    /**
+     * When to interrupt tool execution for steering messages.
+     * - "immediate": check after each tool call (default)
+     * - "wait": defer steering until the current turn completes
+     */
+    interruptMode?: "immediate" | "wait";
+    /**
+     * API format for Kimi Code provider: "openai" or "anthropic" (default: "anthropic")
+     */
+    kimiApiFormat?: "openai" | "anthropic";
+    /** Hint that websocket transport should be preferred when supported by the provider implementation. */
+    preferWebsockets?: boolean;
+    /**
+     * Custom stream function (for proxy backends, etc.). Default uses streamSimple.
+     */
+    streamFn?: StreamFn;
+    /** Absolute wall-clock deadline in Unix epoch milliseconds. */
+    deadline?: number;
+    /**
+     * Optional session identifier forwarded to LLM providers.
+     * Used by providers that support session-based caching (e.g., OpenAI Codex).
+     */
+    sessionId?: string;
+    /**
+     * Optional prompt cache key forwarded to LLM providers.
+     * When omitted, providers may fall back to sessionId.
+     */
+    promptCacheKey?: string;
+    /**
+     * Shared provider state map for session-scoped transport/session caches.
+     */
+    providerSessionState?: Map<string, ProviderSessionState>;
+    /**
+     * Resolves an API key or resolver dynamically for each LLM call.
+     * Useful for expiring tokens and model-scoped credential routing.
+     */
+    getApiKey?: (model: Model) => Promise<ApiKey | undefined> | ApiKey | undefined;
+    /**
+     * Inspect or replace provider payloads before they are sent.
+     */
+    onPayload?: SimpleStreamOptions["onPayload"];
+    /**
+     * Inspect provider response metadata after headers arrive and before streaming body consumption.
+     */
+    onResponse?: SimpleStreamOptions["onResponse"];
+    /**
+     * Inspect raw Server-Sent Events from HTTP streaming providers.
+     */
+    onSseEvent?: SimpleStreamOptions["onSseEvent"];
+    /**
+     * Inspect assistant streaming events before they are emitted to subscribers.
+     * Use this when abort decisions must happen before buffered events continue flowing.
+     */
+    onAssistantMessageEvent?: (message: AssistantMessage, event: AssistantMessageEvent) => void;
+    /**
+     * Called when GPT-5 Harmony protocol leakage is detected and mitigated.
+     */
+    onHarmonyLeak?: (event: HarmonyAuditEvent) => void | Promise<void>;
+    /**
+     * Custom token budgets for thinking levels (token-based providers only).
+     */
+    thinkingBudgets?: ThinkingBudgets;
+    /**
+     * Sampling temperature for LLM calls. `undefined` uses provider default.
+     */
+    temperature?: number;
+    /** Additional sampling controls for providers that support them. */
+    topP?: number;
+    topK?: number;
+    minP?: number;
+    presencePenalty?: number;
+    repetitionPenalty?: number;
+    serviceTier?: ServiceTier;
+    /**
+     * Per-call effective service-tier resolver. When set, it authoritatively
+     * supplies the request's tier (replacing the static `serviceTier` and its
+     * telemetry) per model — used to scope a provider/model into a priority
+     * serving path without mutating the shared session `serviceTier`.
+     */
+    serviceTierResolver?: (model: Model) => ServiceTier | undefined;
+    /**
+     * If true, request that the underlying provider omit reasoning/thinking summaries
+     * from the response. The model still reasons internally; only the human-readable
+     * summary stream is suppressed. Useful when the UI hides thinking blocks anyway.
+     */
+    hideThinkingSummary?: boolean;
+    /**
+     * Maximum delay in milliseconds to wait for a retry when the server requests a long wait.
+     * If the server's requested delay exceeds this value, the request fails immediately,
+     * allowing higher-level retry logic to handle it with user visibility.
+     * Default: 60000 (60 seconds). Set to 0 to disable the cap.
+     */
+    maxRetryDelayMs?: number;
+    /**
+     * Provides tool execution context, resolved per tool call.
+     * Use for late-bound UI or session state access.
+     */
+    getToolContext?: (toolCall?: ToolCallContext) => AgentToolContext | undefined;
+    /**
+     * Optional transform applied to tool call arguments before execution.
+     * Use for deobfuscating secrets or rewriting arguments.
+     */
+    transformToolCallArguments?: (args: Record<string, unknown>, toolName: string) => Record<string, unknown>;
+    /** Enable intent tracing schema injection/stripping in the harness. */
+    intentTracing?: boolean;
+    /**
+     * Strip tool descriptions from provider-bound tool specs (top-level + nested
+     * schema annotations). Use when the full catalog is rendered into the system
+     * prompt so descriptions are not duplicated on the wire. Native tool calling only.
+     */
+    pruneToolDescriptions?: boolean;
+    /** Owned tool-calling dialect. Undefined keeps provider-native tool calling. */
+    dialect?: Dialect;
+    /**
+     * When owned tool calling is active and the model fabricates a tool result
+     * mid-turn: `true` (default) aborts the provider request immediately; `false`
+     * drains the request and discards the fabricated continuation. Forwarded to
+     * the loop's {@link AgentLoopConfig.abortOnFabricatedToolResult}.
+     */
+    abortOnFabricatedToolResult?: boolean;
+    /** Dynamic tool-choice directive (hard {@link ToolChoice} or {@link SoftToolRequirement}), resolved once per turn. */
+    getToolChoice?: () => ToolChoiceDirective | undefined;
+    /**
+     * Cursor exec handlers for local tool execution.
+     */
+    cursorExecHandlers?: CursorExecHandlers;
+    /**
+     * Cursor tool result callback for exec tool responses.
+     */
+    cursorOnToolResult?: CursorToolResultHandler;
+    /** Current working directory used by local tool execution. */
+    cwd?: string;
+    /**
+     * Resolver for the live working directory, re-read on every turn. When set, it
+     * overrides the static {@link cwd} at config-build time so a session move
+     * (`/move`, which updates the host's cwd without reconstructing the Agent) is
+     * reflected in provider options — e.g. GitLab Duo Agent namespace/project
+     * discovery keys off this cwd's git remote. Falls back to `cwd` when it returns
+     * `undefined`.
+     */
+    cwdResolver?: () => string | undefined;
+    /**
+     * Called after a tool call has been validated and is about to execute.
+     * See {@link AgentLoopConfig.beforeToolCall} for full semantics.
+     */
+    beforeToolCall?: AgentLoopConfig["beforeToolCall"];
+    /**
+     * Called after a tool finishes executing, before `tool_execution_end` and the tool-result
+     * message are emitted. See {@link AgentLoopConfig.afterToolCall} for full semantics.
+     */
+    afterToolCall?: AgentLoopConfig["afterToolCall"];
+    /**
+     * Called once an assistant message is finalized, before it reaches the
+     * context, the UI, or tool dispatch. May mutate the message in place (text +
+     * tool-call arguments). See {@link AgentLoopConfig.transformAssistantMessage}.
+     */
+    transformAssistantMessage?: AgentLoopConfig["transformAssistantMessage"];
+    /**
+     * Opt-in OpenTelemetry instrumentation. Passing `{}` enables the loop's
+     * GenAI-semantic-convention spans using the global tracer provider. See
+     * {@link AgentLoopConfig.telemetry} for the full surface.
+     */
+    telemetry?: AgentLoopConfig["telemetry"];
+    /**
+     * Immutable context mode — stabilizes system prompt + tool spec bytes
+     * across turns so DeepSeek/Anthropic prefix caches hit at maximum rate.
+     */
+    appendOnlyContext?: AppendOnlyContextManager;
+}
+export interface AgentPromptOptions {
+    toolChoice?: ToolChoice;
+}
+export declare class Agent {
+    #private;
+    streamFn: StreamFn;
+    getApiKey?: (model: Model) => Promise<ApiKey | undefined> | ApiKey | undefined;
+    /**
+     * Hook invoked after tool arguments are validated and before execution.
+     * Reassign at any time to swap the implementation (e.g. on extension reload).
+     */
+    beforeToolCall?: AgentLoopConfig["beforeToolCall"];
+    /**
+     * Hook invoked after tool execution and before `tool_execution_end` / tool-result
+     * message emission. Reassign at any time to swap the implementation.
+     */
+    afterToolCall?: AgentLoopConfig["afterToolCall"];
+    /**
+     * Hook invoked once an assistant message is finalized, before context append,
+     * UI emission, and tool dispatch. Reassign at any time to swap the implementation.
+     */
+    transformAssistantMessage?: AgentLoopConfig["transformAssistantMessage"];
+    /**
+     * Hook that peeks whether interrupting IRC asides are queued for the next boundary.
+     */
+    hasIrcInterrupts?: AgentLoopConfig["hasIrcInterrupts"];
+    constructor(opts?: AgentOptions);
+    /**
+     * Get the current session ID used for provider caching.
+     */
+    get sessionId(): string | undefined;
+    /**
+     * Set the session ID for provider caching.
+     * Call this when switching sessions (new session, branch, resume).
+     */
+    set sessionId(value: string | undefined);
+    /**
+     * Get the prompt cache key forwarded to providers.
+     */
+    get promptCacheKey(): string | undefined;
+    /**
+     * Set the prompt cache key forwarded to providers.
+     */
+    set promptCacheKey(value: string | undefined);
+    /**
+     * Static metadata forwarded to every API request when no resolver is installed
+     * (e.g. `metadata.user_id` for Anthropic session attribution). Setting this
+     * clears any installed resolver.
+     *
+     * For live/provider-aware metadata (e.g. Anthropic OAuth `account_uuid` that
+     * must reflect the credential selected per-request), use
+     * {@link setMetadataResolver} and read via {@link metadataForProvider}.
+     */
+    get metadata(): Record<string, unknown> | undefined;
+    set metadata(value: Record<string, unknown> | undefined);
+    /**
+     * Resolve request metadata for the given provider at call time. When a
+     * resolver is installed via {@link setMetadataResolver}, it is invoked with
+     * the provider string so the result can be scoped (e.g. `account_uuid` is
+     * only included for `"anthropic"` requests). Falls back to the static
+     * {@link metadata} value when no resolver is set.
+     */
+    metadataForProvider(provider: string): Record<string, unknown> | undefined;
+    /**
+     * Install a function that resolves request metadata at call time. The
+     * resolver receives the target provider string and can gate provider-specific
+     * fields (e.g. `account_uuid` only for `"anthropic"`). Invoked per LLM
+     * request by `agent-loop` after `getApiKey` selects the session-sticky
+     * credential. Pass `undefined` to clear and revert to the static
+     * {@link metadata} value.
+     */
+    setMetadataResolver(resolver: ((provider: string) => Record<string, unknown> | undefined) | undefined): void;
+    /**
+     * Read the active OpenTelemetry configuration. Returns `undefined` when
+     * instrumentation is disabled. Callers spawning child runs (e.g. subagent
+     * dispatch) forward this to the child's loop so its spans appear under the
+     * parent's active context with the subagent's own identity stamped.
+     */
+    get telemetry(): AgentLoopConfig["telemetry"] | undefined;
+    /**
+     * Replace the active OpenTelemetry configuration. Pass `undefined` to
+     * disable instrumentation. Applies to the *next* `agentLoop` invocation —
+     * in-flight loops keep the configuration they started with.
+     */
+    setTelemetry(telemetry: AgentLoopConfig["telemetry"] | undefined): void;
+    /**
+     * Get provider-scoped mutable session state store.
+     */
+    get providerSessionState(): Map<string, ProviderSessionState> | undefined;
+    /**
+     * Set provider-scoped mutable session state store.
+     */
+    set providerSessionState(value: Map<string, ProviderSessionState> | undefined);
+    /**
+     * Get the current thinking budgets.
+     */
+    get thinkingBudgets(): ThinkingBudgets | undefined;
+    /**
+     * Set custom thinking budgets for token-based providers.
+     */
+    set thinkingBudgets(value: ThinkingBudgets | undefined);
+    /**
+     * Get the current sampling temperature.
+     */
+    get temperature(): number | undefined;
+    /**
+     * Set sampling temperature for LLM calls. `undefined` uses provider default.
+     */
+    set temperature(value: number | undefined);
+    get topP(): number | undefined;
+    set topP(value: number | undefined);
+    get topK(): number | undefined;
+    set topK(value: number | undefined);
+    get minP(): number | undefined;
+    set minP(value: number | undefined);
+    get presencePenalty(): number | undefined;
+    set presencePenalty(value: number | undefined);
+    get repetitionPenalty(): number | undefined;
+    set repetitionPenalty(value: number | undefined);
+    get serviceTier(): ServiceTier | undefined;
+    set serviceTier(value: ServiceTier | undefined);
+    get serviceTierResolver(): ((model: Model) => ServiceTier | undefined) | undefined;
+    set serviceTierResolver(value: ((model: Model) => ServiceTier | undefined) | undefined);
+    get hideThinkingSummary(): boolean | undefined;
+    set hideThinkingSummary(value: boolean | undefined);
+    /**
+     * Get the current max retry delay in milliseconds.
+     */
+    get maxRetryDelayMs(): number | undefined;
+    /**
+     * Set the maximum delay to wait for server-requested retries.
+     * Set to 0 to disable the cap.
+     */
+    set maxRetryDelayMs(value: number | undefined);
+    get state(): AgentState;
+    get appendOnlyContext(): AppendOnlyContextManager | undefined;
+    setAppendOnlyContext(manager?: AppendOnlyContextManager): void;
+    /**
+     * Assemble the provider Context for a side-channel (no-loop) request, mirroring
+     * the main loop's prefix (system + normalized tools) so it shares the prompt
+     * cache. Never touches the append-only log or the tool-choice queue. Owned/
+     * in-band dialect sessions stay tools-less (matching their no-native-tools wire
+     * shape and avoiding tool-markup leakage). `llmMessages` is already converted
+     * (and, in production, obfuscated) by the caller.
+     *
+     * `systemPrompt` defaults to the live agent prompt so the side request hits the
+     * same cached prefix as the main loop. Callers that must pin a different prompt
+     * (e.g. handoff generation, which uses the base prompt rather than a per-turn
+     * `before_agent_start` hook override) pass it explicitly.
+     */
+    buildSideRequestContext(llmMessages: Message[], systemPrompt?: string[]): Promise<Context>;
+    subscribe(fn: (e: AgentEvent) => void): () => void;
+    setProviderResponseInterceptor(fn: SimpleStreamOptions["onResponse"] | undefined): void;
+    setRawSseEventInterceptor(fn: SimpleStreamOptions["onSseEvent"] | undefined): void;
+    setAssistantMessageEventInterceptor(fn: ((message: AssistantMessage, event: AssistantMessageEvent) => void) | undefined): void;
+    setOnBeforeYield(fn: (() => Promise<void> | void) | undefined): void;
+    setOnTurnEnd(fn: ((messages: AgentMessage[], signal?: AbortSignal, context?: AgentTurnEndContext) => Promise<void> | void) | undefined): void;
+    /**
+     * Provide a source of non-interrupting "aside" messages (e.g. background-job
+     * completions, late LSP diagnostics) drained at each step boundary. Never
+     * aborts in-flight tools. See `AgentLoopConfig.getAsideMessages`.
+     */
+    setAsideMessageProvider(fn: (() => AsideMessage[] | Promise<AsideMessage[]>) | undefined): void;
+    emitExternalEvent(event: AgentEvent): void;
+    setSystemPrompt(v: string[] | string): void;
+    setModel(m: Model): void;
+    setThinkingLevel(l: Effort | undefined): void;
+    setDisableReasoning(disabled: boolean): void;
+    setSteeringMode(mode: "all" | "one-at-a-time"): void;
+    getSteeringMode(): "all" | "one-at-a-time";
+    setFollowUpMode(mode: "all" | "one-at-a-time"): void;
+    getFollowUpMode(): "all" | "one-at-a-time";
+    setInterruptMode(mode: "immediate" | "wait"): void;
+    getInterruptMode(): "immediate" | "wait";
+    setTools(t: AgentTool<any>[]): void;
+    replaceMessages(ms: AgentMessage[]): void;
+    replaceQueues(steering: AgentMessage[], followUp: AgentMessage[]): void;
+    appendMessage(m: AgentMessage): void;
+    popMessage(): AgentMessage | undefined;
+    /**
+     * Queue a steering message to interrupt the agent mid-run.
+     * Delivered after current tool execution, skips remaining tools.
+     */
+    steer(m: AgentMessage): void;
+    /**
+     * Queue a follow-up message to be processed after the agent finishes.
+     * Delivered only when agent has no more tool calls or steering messages.
+     */
+    followUp(m: AgentMessage): void;
+    clearSteeringQueue(): void;
+    clearFollowUpQueue(): void;
+    clearAllQueues(): void;
+    hasQueuedMessages(): boolean;
+    /** Non-consuming view of the pending steering queue (insertion order, newest
+     *  last). The session layer derives its queued-message display/count from
+     *  this live view instead of a mirror, so the agent-core queue stays the
+     *  single source of truth. */
+    peekSteeringQueue(): readonly AgentMessage[];
+    /** Non-consuming view of the pending follow-up queue. See
+     *  {@link peekSteeringQueue}. */
+    peekFollowUpQueue(): readonly AgentMessage[];
+    get isAborting(): boolean;
+    /**
+     * Remove and return the last steering message from the queue (LIFO).
+     * Used by dequeue keybinding.
+     */
+    popLastSteer(): AgentMessage | undefined;
+    /**
+     * Remove and return the last follow-up message from the queue (LIFO).
+     * Used by dequeue keybinding.
+     */
+    popLastFollowUp(): AgentMessage | undefined;
+    clearMessages(): void;
+    abort(reason?: unknown): void;
+    waitForIdle(): Promise<void>;
+    reset(): void;
+    /** Send a prompt with an AgentMessage */
+    prompt(message: AgentMessage | AgentMessage[], options?: AgentPromptOptions): Promise<void>;
+    prompt(input: string, options?: AgentPromptOptions): Promise<void>;
+    prompt(input: string, images?: ImageContent[], options?: AgentPromptOptions): Promise<void>;
+    /**
+     * Continue from current context (used for retries and resuming queued messages).
+     */
+    continue(): Promise<void>;
+}

package/dist/types/append-only-context.d.ts

@@ -0,0 +1,133 @@
+/**
+ * Append-only context mode — stabilizes the byte prefix sent to the LLM
+ * across turns so provider prefix caches (DeepSeek, Anthropic, etc.)
+ * hit at the maximum possible rate.
+ *
+ * Two mechanisms:
+ *
+ * 1. **StablePrefix** — system prompt + tool specs are computed once
+ *    and frozen. Subsequent turns reuse the exact same byte sequence
+ *    unless `invalidate()` is called (e.g. after MCP reconnect).
+ *
+ * 2. **AppendOnlyLog** — messages only grow; prior turns are never
+ *    re-serialized. Combined with a stable prefix, only the user's new
+ *    message delta is a cache miss each turn.
+ */
+import type { Context, Message, Tool } from "jeopi-ai";
+import type { Dialect } from "jeopi-ai/dialect";
+import type { AgentContext } from "./types";
+/** Frozen system prompt + tool spec snapshot. */
+export interface StablePrefixSnapshot {
+    systemPrompt: string[];
+    tools: Tool[];
+    fingerprint: string;
+}
+/** Options threaded through `build()` so the snapshot reflects loop-time settings. */
+export interface BuildOptions {
+    /** Inject the `i` intent field into tool schemas (must match agent-loop's normalizeTools). */
+    intentTracing: boolean;
+    exampleDialect?: Dialect;
+    /** Strip tool descriptions from the provider-bound specs (must match normalizeTools). */
+    pruneToolDescriptions?: boolean;
+}
+/**
+ * A frozen prefix (system prompt + tools) that produces stable byte
+ * sequences across `build()` calls.
+ *
+ * The first `build()` snapshots the live state. Subsequent calls reuse
+ * the cached copy until `invalidate()` is called or the live state's
+ * fingerprint changes.
+ */
+export declare class StablePrefix {
+    #private;
+    get fingerprint(): string;
+    get version(): number;
+    get built(): boolean;
+    /**
+     * Build or rebuild from live context.
+     * Returns `true` if the prefix actually changed (cache miss imminent).
+     */
+    build(context: AgentContext, options: BuildOptions): boolean;
+    /** Force rebuild on the next `build()` call. */
+    invalidate(): void;
+    /**
+     * Returns the cached prefix.
+     * @throws if `build()` was never called.
+     */
+    toContext(): {
+        systemPrompt: string[];
+        tools: Tool[];
+    };
+}
+/**
+ * Append-only message log at the `Message[]` (provider-level) layer.
+ *
+ * The only mutation path is `replaceTail()`, reserved for compaction.
+ * Every other operation is append-only.
+ */
+export declare class AppendOnlyLog {
+    #private;
+    get length(): number;
+    append(message: any): void;
+    extend(messages: any[]): void;
+    /** Replace the last entry — only legal for compaction. */
+    replaceTail(replacement: any): void;
+    /** Returns a shallow copy of all entries. */
+    toMessages(): Message[];
+    /** Direct readonly access for in-place inspection. */
+    entries(): readonly Message[];
+    /** Drop entries past index `count`, keeping the first `count` byte-stable.
+     * Used by {@link AppendOnlyContextManager.syncMessages} to preserve the
+     * already-on-the-wire prefix when a later message diverges. */
+    truncate(count: number): void;
+    clear(): void;
+}
+/**
+ * Manages a stable prefix + append-only log for the agent loop.
+ *
+ * Call `build(context)` each turn to get a `Context` with stable
+ * `systemPrompt` and `tools` and append-only messages. Call
+ * `syncMessages(normalizedMessages)` after `convertToLlm` each
+ * turn to keep the log in sync.
+ *
+ * Example:
+ * ```
+ * const mgr = new AppendOnlyContextManager();
+ * const ctx = mgr.build(context);  // first call snapshots prefix
+ * mgr.syncMessages(normalized);    // grow the log
+ * ctx = mgr.build(context);        // subsequent calls use cache
+ * ```
+ */
+export declare class AppendOnlyContextManager {
+    #private;
+    readonly prefix: StablePrefix;
+    readonly log: AppendOnlyLog;
+    build(context: AgentContext, options: BuildOptions): Context;
+    /**
+     * Sync normalized (provider-level) messages into the append-only log.
+     *
+     * Three cases:
+     *
+     * 1. **Append**: same prefix, new tail → push the new entries.
+     * 2. **Compaction**: shorter array → clear the log and replay.
+     * 3. **In-place rewrite** (per-turn pruning, transformContext re-render,
+     *    image strip, etc.): find the longest byte-stable prefix between
+     *    the previously-synced messages and the new ones, drop the log
+     *    down to that prefix, then append the diverged tail. Earlier
+     *    revisions cleared the whole log on any digest change, which on
+     *    llama.cpp / local backends forced a full ~40k-token re-prefill
+     *    every turn that an extension, prune pass, or steering re-wrap
+     *    rewrote a single message (#3406). Preserving the stable prefix
+     *    lets the provider's KV cache stay warm up to the divergence
+     *    point — the model only re-prefills from the changed message on.
+     */
+    syncMessages(normalizedMessages: any[]): void;
+    /** Reset prefix + log for a model/provider switch while mode stays active. */
+    invalidateForModelChange(): void;
+    /** Reset the sync cursor AND clear the log. */
+    resetSyncCursor(): void;
+    appendMessage(message: any): void;
+    replaceTailMessage(message: any): void;
+    invalidate(): void;
+    reset(context: AgentContext, options: BuildOptions): void;
+}

package/dist/types/compaction/branch-summarization.d.ts

@@ -0,0 +1,101 @@
+/**
+ * Branch summarization for tree navigation.
+ *
+ * When navigating to a different point in the session tree, this generates
+ * a summary of the branch being left so context isn't lost.
+ */
+import type { Api, ApiKey, AssistantMessage, Context, Model, SimpleStreamOptions } from "jeopi-ai";
+import { type AgentTelemetry } from "../telemetry";
+import type { AgentMessage } from "../types";
+import type { ReadonlySessionManager, SessionEntry } from "./entries";
+import { type ConvertToLlm } from "./messages";
+import { type FileOperations } from "./utils";
+export interface BranchSummaryResult {
+    summary?: string;
+    readFiles?: string[];
+    modifiedFiles?: string[];
+    aborted?: boolean;
+    error?: string;
+}
+/** Details stored in BranchSummaryEntry.details for file tracking */
+export interface BranchSummaryDetails {
+    readFiles: string[];
+    modifiedFiles: string[];
+}
+export type { FileOperations } from "./utils";
+export interface BranchPreparation {
+    /** Messages extracted for summarization, in chronological order */
+    messages: AgentMessage[];
+    /** File operations extracted from tool calls */
+    fileOps: FileOperations;
+    /** Total estimated tokens in messages */
+    totalTokens: number;
+}
+export interface CollectEntriesResult {
+    /** Entries to summarize, in chronological order */
+    entries: SessionEntry[];
+    /** Common ancestor between old and new position, if any */
+    commonAncestorId: string | null;
+}
+export interface GenerateBranchSummaryOptions {
+    /** Model to use for summarization */
+    model: Model;
+    /** API key for the model */
+    apiKey: ApiKey;
+    /** Abort signal for cancellation */
+    signal: AbortSignal;
+    /** Optional custom instructions for summarization */
+    customInstructions?: string;
+    /** Tokens reserved for prompt + LLM response (default 16384) */
+    reserveTokens?: number;
+    /** Optional metadata forwarded to the underlying API request (e.g. user_id for session attribution). */
+    metadata?: Record<string, unknown>;
+    /** Convert app-specific messages before serializing the branch summary prompt. */
+    convertToLlm?: ConvertToLlm;
+    /**
+     * Optional telemetry handle. When provided, the branch summary LLM call is
+     * wrapped in an OTEL chat span tagged with `pi.gen_ai.oneshot.kind = "branch_summary"`.
+     */
+    telemetry?: AgentTelemetry;
+    /**
+     * Optional completion transport override (same contract as
+     * {@link SummaryOptions.completeImpl}). Lets the host route the branch
+     * summary HTTP request through its provider-concurrency limiter instead
+     * of the default `completeSimple` transport.
+     */
+    completeImpl?: <TApi extends Api>(model: Model<TApi>, ctx: Context, options: SimpleStreamOptions) => Promise<AssistantMessage>;
+}
+/**
+ * Collect entries that should be summarized when navigating from one position to another.
+ *
+ * Walks from oldLeafId back to the common ancestor with targetId, collecting entries
+ * along the way. Does NOT stop at compaction boundaries - those are included and their
+ * summaries become context.
+ *
+ * @param session - Session manager (read-only access)
+ * @param oldLeafId - Current position (where we're navigating from)
+ * @param targetId - Target position (where we're navigating to)
+ * @returns Entries to summarize and the common ancestor
+ */
+export declare function collectEntriesForBranchSummary(session: ReadonlySessionManager, oldLeafId: string | null, targetId: string): CollectEntriesResult;
+/**
+ * Prepare entries for summarization with token budget.
+ *
+ * Walks entries from NEWEST to OLDEST, adding messages until we hit the token budget.
+ * This ensures we keep the most recent context when the branch is too long.
+ *
+ * Also collects file operations from:
+ * - Tool calls in assistant messages
+ * - Existing branch_summary entries' details (for cumulative tracking)
+ *
+ * @param entries - Entries in chronological order
+ * @param tokenBudget - Maximum tokens to include (0 = no limit)
+ */
+export declare function prepareBranchEntries(entries: SessionEntry[], tokenBudget?: number): BranchPreparation;
+/**
+ * Generate a summary of abandoned branch entries.
+ *
+ * @param entries - Session entries to summarize (chronological order)
+ * @param options - Generation options
+ */
+export declare function generateBranchSummary(entries: SessionEntry[], options: GenerateBranchSummaryOptions): Promise<BranchSummaryResult>;