@sayknow-cli/agent-core 0.4.1 → 0.4.3

package/dist/types/agent-loop.d.ts

@@ -0,0 +1,108 @@
+/**
+ * Agent loop that works with AgentMessage throughout.
+ * Transforms to Message[] only at the LLM call boundary.
+ */
+import { type Context, EventStream } from "@sayknow-cli/ai";
+import { type AgentRunCoverage, type AgentRunSummary } from "./run-collector";
+import type { AgentContext, AgentEvent, AgentLoopConfig, AgentMessage, StreamFn } from "./types";
+/** Sentinel returned by the abort race in `streamAssistantResponse`. */
+/**
+ * Defensive caps for a provisional managed attempt. These are intentionally
+ * well above ordinary streamed responses; they only bound memory when an
+ * upstream emits an unbounded event stream before the attempt can commit.
+ */
+export declare const MANAGED_ATTEMPT_MAX_STAGED_EVENTS = 10000;
+export declare const MANAGED_ATTEMPT_MAX_STAGED_BYTES: number;
+/**
+ * Start an agent loop with a new prompt message.
+ * The prompt is added to the context and events are emitted for it.
+ */
+export declare function agentLoop(prompts: AgentMessage[], context: AgentContext, config: AgentLoopConfig, signal?: AbortSignal, streamFn?: StreamFn, emitManagedAgentStart?: boolean): EventStream<AgentEvent, AgentMessage[]>;
+/**
+ * Continue an agent loop from the current context without adding a new message.
+ * Used for retries - context already has user message or tool results.
+ *
+ * **Important:** The last message in context must convert to a `user` or `toolResult` message
+ * via `convertToLlm`. If it doesn't, the LLM provider will reject the request.
+ * This cannot be validated here since `convertToLlm` is only called once per turn.
+ */
+export declare function agentLoopContinue(context: AgentContext, config: AgentLoopConfig, signal?: AbortSignal, streamFn?: StreamFn, emitManagedAgentStart?: boolean): EventStream<AgentEvent, AgentMessage[]>;
+/**
+ * Hard work budget for one degraded snapshot: every visited node AND every
+ * enumerated own key is debited against this budget before it is processed
+ * (accessor keys and re-visits of shared objects included), and any remainder
+ * collapses to the deterministic `"[truncated]"` placeholder. Well above
+ * ordinary streamed events; it only bounds hostile graphs.
+ */
+export declare const MANAGED_SNAPSHOT_MAX_NODES = 100000;
+/**
+ * Cycle-aware deep clone that always returns a detached, JSON-serializable
+ * value. Used whenever a detached snapshot cannot be safely obtained or
+ * measured: after `structuredClone` fails, and again when a (successfully
+ * cloned) snapshot cannot be serialized for byte accounting.
+ *
+ * Totality rules — the walk must never dispatch through payload-controlled
+ * code, throw, or do unbounded work:
+ * - proxies (revoked or live) are collapsed to `"[unserializable]"` BEFORE
+ *   any reflective operation, so `ownKeys`/descriptor traps are never
+ *   dispatched (`util.types.isProxy` identifies proxies without touching
+ *   their handlers);
+ * - only intrinsics are used on the remaining ordinary objects (no
+ *   `input.map`, no `input.getTime()`, no `input.length` reads);
+ * - arrays are enumerated through their own present keys, never their
+ *   declared length, so a sparse array cannot force a dense allocation
+ *   proportional to `length`; sparse/exotic arrays degrade to a null-proto
+ *   record of their present indices, and the dense-shape decision verifies
+ *   every index against its ordinal;
+ * - the walk debits `maxNodes` budget per visited node and per enumerated
+ *   key before processing it; anything beyond the budget becomes
+ *   `"[truncated]"` (the one linear primitive per visited node is a single
+ *   `Object.keys` call on a non-proxy object the process already holds);
+ * - property values are read via own-property descriptors, so accessors are
+ *   never invoked (a snapshot must not cause observable side effects) and are
+ *   replaced with `"[accessor]"`;
+ * - functions/symbols and any property that cannot be read safely become
+ *   short placeholders, `bigint` becomes its decimal string, and references
+ *   back into the current path collapse to `"[Circular]"`;
+ * - records are built on a null prototype so a `__proto__` key cannot mutate
+ *   the clone's prototype chain.
+ *
+ * Exported for direct regression coverage of the budget accounting; runtime
+ * callers use the default budget via {@link managedAttemptSnapshot}.
+ */
+export declare function sanitizedDetachedClone<T>(value: T, maxNodes?: number): T;
+/**
+ * Detailed-result handle returned by {@link agentLoopDetailed}. Adds the
+ * run-level telemetry/coverage rollup to the existing `AgentMessage[]`
+ * payload without changing the resolved type of `stream.result()`.
+ */
+export interface AgentLoopDetailedResult {
+    readonly messages: AgentMessage[];
+    readonly telemetry: AgentRunSummary | undefined;
+    readonly coverage: AgentRunCoverage | undefined;
+}
+/**
+ * Convenience wrapper over {@link agentLoop} that exposes the run-level
+ * summary + coverage alongside the messages. The returned `stream` is the
+ * same `EventStream` callers already consume; `detailed()` awaits the
+ * stream's `agent_end` event and returns the additive fields.
+ *
+ * Existing `stream.result()` semantics are preserved — it still resolves to
+ * `AgentMessage[]`. Use {@link agentLoopDetailed} when you need the rollup;
+ * use {@link agentLoop} when you do not.
+ */
+export declare function agentLoopDetailed(prompts: AgentMessage[], context: AgentContext, config: AgentLoopConfig, signal?: AbortSignal, streamFn?: StreamFn): {
+    readonly stream: EventStream<AgentEvent, AgentMessage[]>;
+    readonly detailed: () => Promise<AgentLoopDetailedResult>;
+};
+/**
+ * Like {@link agentLoopDetailed} but built on top of
+ * {@link agentLoopContinue}.
+ */
+export declare function agentLoopContinueDetailed(context: AgentContext, config: AgentLoopConfig, signal?: AbortSignal, streamFn?: StreamFn): {
+    readonly stream: EventStream<AgentEvent, AgentMessage[]>;
+    readonly detailed: () => Promise<AgentLoopDetailedResult>;
+};
+export declare function normalizeMessagesForProvider(messages: Context["messages"], model: AgentLoopConfig["model"]): Context["messages"];
+export declare const INTENT_FIELD = "_i";
+export declare function normalizeTools(tools: AgentContext["tools"], injectIntent: boolean): Context["tools"];

package/dist/types/agent.d.ts

@@ -0,0 +1,443 @@
+/** Agent class that uses the agent-loop directly.
+ * No transport abstraction - calls streamSimple via the loop.
+ */
+import { type AssistantMessage, type AssistantMessageEvent, type CursorExecHandlers, type CursorToolResultHandler, type Effort, type ImageContent, type Message, type Model, type ProviderSessionState, type ServiceTier, type SimpleStreamOptions, type ThinkingBudgets, type ToolChoice } from "@sayknow-cli/ai";
+import type { AppendOnlyContextManager } from "./append-only-context";
+import type { HarmonyAuditEvent } from "./harmony-leak";
+import type { AgentEvent, AgentLoopConfig, AgentMessage, AgentState, AgentTool, AgentToolContext, ManagedLogicalRunId, RunTerminalRequest, StreamFn, ToolCallContext } from "./types";
+/**
+ * Whether persisted history ends at a point where a new model turn can resume.
+ * Assistant-ended histories require an in-memory queued message and are handled
+ * separately by `Agent.continue()`.
+ */
+export declare function canContinuePersistedHistory(messages: readonly AgentMessage[]): boolean;
+export declare class ManagedCursorInvariantError extends Error {
+    constructor(message?: string);
+}
+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[]>;
+    /**
+     * 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";
+    /** Cooperative pause checkpoint passed through to AgentLoopConfig.shouldPause. */
+    shouldPause?: AgentLoopConfig["shouldPause"];
+    /**
+     * 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;
+    /**
+     * Optional session identifier forwarded to LLM providers.
+     * Used by providers that support session-based caching (e.g., OpenAI code provider).
+     */
+    sessionId?: string;
+    /** Provider-facing cache/session affinity identifier. */
+    providerSessionId?: string;
+    /**
+     * Shared provider state map for session-scoped transport/session caches.
+     */
+    providerSessionState?: Map<string, ProviderSessionState>;
+    /**
+     * Resolves an API key dynamically for each LLM call.
+     * Useful for expiring tokens (e.g., GitHub Copilot OAuth).
+     */
+    getApiKey?: (provider: string) => Promise<string | undefined> | string | undefined;
+    getAuthCredentialType?: (provider: string) => "api_key" | "oauth" | 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 for non-content tool-choice incapability stream events. */
+    onToolChoiceIncapability?: AgentLoopConfig["onToolChoiceIncapability"];
+    /**
+     * 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;
+    /**
+     * 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;
+    /** Provider request retry budget. Counts retries, not the initial attempt. */
+    requestMaxRetries?: number;
+    /** Provider stream replay retry budget. Counts retries, not the initial attempt. */
+    streamMaxRetries?: 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;
+    /** Dynamic tool choice override, resolved per LLM call. */
+    getToolChoice?: () => ToolChoice | undefined;
+    /**
+     * Cursor exec handlers for local tool execution.
+     */
+    cursorExecHandlers?: CursorExecHandlers;
+    /**
+     * Cursor tool result callback for exec tool responses.
+     */
+    cursorOnToolResult?: CursorToolResultHandler;
+    /**
+     * 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"];
+    /**
+     * 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;
+    /** Disable transport replay; fallback accounting is owned by the caller. */
+    fallbackManaged?: boolean;
+    /** Called synchronously after this invocation claims the agent run, before asynchronous provider work. */
+    onRunAccepted?: () => void;
+    /** Called once immediately before every managed upstream request. */
+    nextFallbackAttempt?: AgentLoopConfig["nextFallbackAttempt"];
+    /** Called after a managed upstream request is accepted and committed. */
+    onManagedAttemptAccepted?: AgentLoopConfig["onManagedAttemptAccepted"];
+    /** Receives a discarded managed attempt without exposing assistant lifecycle events. */
+    onManagedAttemptOutcome?: AgentLoopConfig["onManagedAttemptOutcome"];
+}
+export type AgentQueueSnapshot = {
+    steering: AgentMessage[];
+    followUp: AgentMessage[];
+};
+export declare class Agent {
+    #private;
+    get intentTracing(): boolean;
+    streamFn: StreamFn;
+    getApiKey?: (provider: string) => Promise<string | undefined> | string | undefined;
+    getAuthCredentialType?: (provider: string) => "api_key" | "oauth" | 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"];
+    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 providerSessionId(): string | undefined;
+    set providerSessionId(value: string | undefined);
+    /**
+     * Whether websocket transport is preferred when the provider implementation
+     * supports it. Read by maintenance one-shot calls (compaction, handoff,
+     * branch summary) so they forward the same transport preference as live turns.
+     */
+    get preferWebsockets(): boolean | 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 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 requestMaxRetries(): number | undefined;
+    set requestMaxRetries(value: number | undefined);
+    get streamMaxRetries(): number | undefined;
+    set streamMaxRetries(value: number | undefined);
+    get state(): AgentState;
+    get contextRevision(): number;
+    get appendOnlyContext(): AppendOnlyContextManager | undefined;
+    setAppendOnlyContext(manager?: AppendOnlyContextManager): void;
+    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;
+    setShouldPause(fn: AgentLoopConfig["shouldPause"] | undefined): void;
+    setMaintainContext(fn: AgentLoopConfig["maintainContext"] | undefined): void;
+    emitExternalEvent(event: AgentEvent): void;
+    createExternalEventEmitterForCurrentRun(): ((event: AgentEvent) => void) | undefined;
+    setSystemPrompt(v: string[]): void;
+    setModel(m: Model | undefined): void;
+    setThinkingLevel(l: Effort | undefined): 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;
+    appendMessage(m: AgentMessage): void;
+    popMessage(): AgentMessage | undefined;
+    /**
+     * For callers that mutate committed messages or the system prompt in place
+     * outside Agent-owned mutators.
+     */
+    touchContext(): void;
+    /**
+     * 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.
+     *
+     * `forceOneAtATime` lets UI composer queues preserve prompt-by-prompt
+     * delivery even when the session-wide follow-up mode is set to `all` for
+     * other integration paths.
+     */
+    followUp(m: AgentMessage, options?: {
+        forceOneAtATime?: boolean;
+    }): void;
+    clearSteeringQueue(): void;
+    clearFollowUpQueue(): void;
+    clearAllQueues(): void;
+    hasQueuedMessages(): boolean;
+    hasQueuedSteering(): boolean;
+    /**
+     * Snapshot the steering queue without mutating it. Used to preserve queued
+     * steering across maintenance ops (compaction/handoff) that call reset().
+     */
+    snapshotSteering(): AgentMessage[];
+    /**
+     * Restore previously snapshotted steering messages ahead of any newly
+     * queued ones. No-op for an empty snapshot.
+     */
+    restoreSteering(messages: AgentMessage[]): void;
+    /** Snapshot the follow-up queue without mutating it. */
+    snapshotFollowUp(): AgentMessage[];
+    /** Restore previously snapshotted follow-up messages ahead of any newly queued ones. */
+    restoreFollowUp(messages: AgentMessage[]): void;
+    /** Snapshot both executable queues as one atomic session-level view. */
+    snapshotQueues(): AgentQueueSnapshot;
+    /** Replace both executable queues with a prior snapshot. */
+    restoreQueues(snapshot: AgentQueueSnapshot): void;
+    /**
+     * Remove and return the last steering message from the queue (LIFO).
+     * Used by dequeue keybinding.
+     */
+    popLastSteer(): AgentMessage | undefined;
+    removeSteerAt(index: number): AgentMessage | undefined;
+    moveSteer(fromIndex: number, toIndex: number): boolean;
+    /**
+     * Remove and return the last follow-up message from the queue (LIFO).
+     * Used by dequeue keybinding.
+     */
+    popLastFollowUp(): AgentMessage | undefined;
+    removeFollowUpAt(index: number): AgentMessage | undefined;
+    moveFollowUp(fromIndex: number, toIndex: number): boolean;
+    /** Remove queued steering+follow-up messages matching `predicate`, preserving order of the rest. */
+    removeQueuedMessages(predicate: (message: AgentMessage) => boolean): {
+        steering: number;
+        followUp: number;
+        total: number;
+    };
+    clearMessages(): void;
+    abort(): void;
+    /**
+     * Force the current run out of the busy/streaming state when cooperative abort
+     * did not drain. The abandoned provider/tool stream may still settle later, so
+     * #runLoop guards every state mutation with a run id.
+     */
+    forceAbort(reason?: string): boolean;
+    waitForIdle(): Promise<void>;
+    /** The active per-attempt run identifier. */
+    get activeRunId(): number | undefined;
+    /**
+     * Stable identifier for the active managed logical run, shared by every retry
+     * attempt. Pass this value to requestRunTerminal(); never retain activeRunId
+     * for managed terminal completion.
+     */
+    get currentManagedLogicalRunId(): ManagedLogicalRunId | undefined;
+    /**
+     * Request terminal completion through the single logical-run keyed finalizer.
+     *
+     * For managed runs, logicalRunId must be currentManagedLogicalRunId from any
+     * attempt in the retry chain. Non-managed runs use their activeRunId. Terminal
+     * requests with messages emit a committed message_start/message_end lifecycle
+     * for each diagnostic before agent_end. Requests without messages (such as
+     * cancellation) emit only agent_end.
+     */
+    requestRunTerminal(logicalRunId: ManagedLogicalRunId, request: RunTerminalRequest): boolean;
+    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(options?: AgentPromptOptions): Promise<void>;
+}