@prometheus-ai/agent-core 0.5.0

package/src/thinking.ts

@@ -0,0 +1,19 @@
+import { Effort } from "@prometheus-ai/ai";
+
+/**
+ * Agent-local thinking selector.
+ *
+ * `off` disables reasoning, while `inherit` defers to a higher-level selector.
+ */
+export const ThinkingLevel = {
+	Inherit: "inherit",
+	Off: "off",
+	Minimal: Effort.Minimal,
+	Low: Effort.Low,
+	Medium: Effort.Medium,
+	High: Effort.High,
+	XHigh: Effort.XHigh,
+} as const;
+
+export type ThinkingLevel = (typeof ThinkingLevel)[keyof typeof ThinkingLevel];
+export type ResolvedThinkingLevel = Exclude<ThinkingLevel, "inherit">;

package/src/types.ts

@@ -0,0 +1,505 @@
+import type {
+	AssistantMessage,
+	AssistantMessageEvent,
+	AssistantMessageEventStream,
+	Effort,
+	ImageContent,
+	Message,
+	Model,
+	SimpleStreamOptions,
+	Static,
+	streamSimple,
+	TextContent,
+	Tool,
+	ToolChoice,
+	ToolResultMessage,
+	TSchema,
+} from "@prometheus-ai/ai";
+import type { AppendOnlyContextManager } from "./append-only-context";
+import type { HarmonyAuditEvent } from "./harmony-leak";
+import type { AgentRunCoverage, AgentRunSummary } from "./run-collector";
+import type { AgentTelemetryConfig } from "./telemetry";
+
+/** Stream function - can return sync or Promise for async config lookup */
+export type StreamFn = (
+	...args: Parameters<typeof streamSimple>
+) => AssistantMessageEventStream | Promise<AssistantMessageEventStream>;
+
+/**
+ * Configuration for the agent loop.
+ */
+export interface AgentLoopConfig extends SimpleStreamOptions {
+	model: Model;
+
+	/**
+	 * 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";
+
+	/**
+	 * Maximum completed tool calls to accept from one streamed assistant turn before
+	 * cutting the provider stream and executing that batch. The cap is enforced on
+	 * `toolcall_end` so every executed call has complete arguments. Undefined disables
+	 * batching.
+	 */
+	maxToolCallsPerTurn?: number;
+
+	/**
+	 * Optional session identifier forwarded to LLM providers.
+	 * Used by providers that support session-based caching (e.g., OpenAI Codex).
+	 */
+	sessionId?: string;
+
+	/**
+	 * Optional resolver called per LLM request to produce request metadata.
+	 * When set, the agent loop evaluates it **after** `getApiKey` resolves the
+	 * session-sticky credential, ensuring the metadata's `account_uuid` reflects
+	 * the credential actually used for the request (not the credential that was
+	 * current when `AgentLoopConfig` was first constructed). Overrides the static
+	 * `metadata` field when present.
+	 */
+	metadataResolver?: (provider: string) => Record<string, unknown> | undefined;
+
+	/**
+	 * Converts AgentMessage[] to LLM-compatible Message[] before each LLM call.
+	 *
+	 * Each AgentMessage must be converted to a UserMessage, AssistantMessage, or ToolResultMessage
+	 * that the LLM can understand. AgentMessages that cannot be converted (e.g., UI-only notifications,
+	 * status messages) should be filtered out.
+	 *
+	 * @example
+	 * ```typescript
+	 * convertToLlm: (messages) => messages.flatMap(m => {
+	 *   if (m.role === "custom") {
+	 *     // Convert custom message to user message
+	 *     return [{ role: "user", content: m.content, timestamp: m.timestamp }];
+	 *   }
+	 *   if (m.role === "notification") {
+	 *     // Filter out UI-only messages
+	 *     return [];
+	 *   }
+	 *   // Pass through standard LLM messages
+	 *   return [m];
+	 * })
+	 * ```
+	 */
+	convertToLlm: (messages: AgentMessage[]) => Message[] | Promise<Message[]>;
+
+	/**
+	 * Optional transform applied to the context before `convertToLlm`.
+	 *
+	 * Use this for operations that work at the AgentMessage level:
+	 * - Context window management (pruning old messages)
+	 * - Injecting context from external sources
+	 *
+	 * @example
+	 * ```typescript
+	 * transformContext: async (messages) => {
+	 *   if (estimateTokens(messages) > MAX_TOKENS) {
+	 *     return pruneOldMessages(messages);
+	 *   }
+	 *   return messages;
+	 * }
+	 * ```
+	 */
+	transformContext?: (messages: AgentMessage[], signal?: AbortSignal) => Promise<AgentMessage[]>;
+
+	/**
+	 * Resolves an API key dynamically for each LLM call.
+	 *
+	 * Useful for short-lived OAuth tokens (e.g., GitHub Copilot) that may expire
+	 * during long-running tool execution phases.
+	 */
+	getApiKey?: (provider: string) => Promise<string | undefined> | string | undefined;
+
+	/**
+	 * Returns steering messages to inject into the conversation mid-run.
+	 *
+	 * Called after each tool execution to check for user interruptions unless interruptMode is "wait".
+	 * If messages are returned, remaining tool calls are skipped and
+	 * these messages are added to the context before the next LLM call.
+	 */
+	getSteeringMessages?: () => Promise<AgentMessage[]>;
+
+	/**
+	 * Returns follow-up messages to process after the agent would otherwise stop.
+	 *
+	 * Called when the agent has no more tool calls and no steering messages.
+	 * If messages are returned, they're added to the context and the agent
+	 * continues with another turn.
+	 */
+	getFollowUpMessages?: () => Promise<AgentMessage[]>;
+	/**
+	 * Hook fired right before the loop would exit.
+	 *
+	 * Called when the agent has no more tool calls and no steering messages,
+	 * immediately before polling follow-up messages.
+	 */
+	onBeforeYield?: () => Promise<void> | void;
+
+	/**
+	 * Provides tool execution context, resolved per tool call.
+	 * Use for late-bound UI or session state access.
+	 */
+	getToolContext?: (toolCall?: ToolCallContext) => AgentToolContext | undefined;
+
+	/**
+	 * Refreshes prompt/tool context from live session state before each model call.
+	 * Use this when tool availability or the system prompt can change mid-turn.
+	 */
+	syncContextBeforeModelCall?: (context: AgentContext) => void | Promise<void>;
+
+	/**
+	 * 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 for tool calls.
+	 * When enabled, the harness injects a `string` field into tool schemas sent to the model,
+	 * then strips from arguments before executing tools.
+	 */
+	intentTracing?: boolean;
+	/**
+	 * Append-only context mode — stabilizes system prompt + tool spec bytes
+	 * across turns so provider prefix caches hit at maximum rate.
+	 *
+	 * When set, the loop reads messages from the append-only log (stable
+	 * byte prefix) and caches system prompt + tools. Tools exclude per-turn
+	 * `_i` intent fields.
+	 */
+	appendOnlyContext?: AppendOnlyContextManager;
+
+	/**
+	 * Inspect assistant streaming events before they are published to the outer agent event stream.
+	 * Callers may abort synchronously to stop consuming buffered provider events.
+	 */
+	onAssistantMessageEvent?: (message: AssistantMessage, event: AssistantMessageEvent) => void;
+
+	/**
+	 * Called when GPT-5 Harmony protocol leakage is detected and mitigated.
+	 */
+	onHarmonyLeak?: (event: HarmonyAuditEvent) => void | Promise<void>;
+
+	/**
+	 * Dynamic tool choice override, resolved per LLM call.
+	 * When set and returns a value, overrides the static `toolChoice`.
+	 */
+	getToolChoice?: () => ToolChoice | undefined;
+
+	/**
+	 * Dynamic reasoning effort override, resolved per LLM call.
+	 * When set and returns a value, overrides the static `reasoning` captured
+	 * at run-loop start. Use this so mid-run thinking-level changes apply on
+	 * the next model call instead of waiting for the next prompt.
+	 */
+	getReasoning?: () => Effort | undefined;
+
+	/**
+	 * Called after a tool call has been validated and is about to execute.
+	 *
+	 * Return `{ block: true }` to prevent execution. The loop emits an error tool
+	 * result instead (using `reason` as the error text, or a default if omitted).
+	 *
+	 * Mutating `context.args` in place changes the arguments passed to `tool.execute`
+	 * — the loop does **not** re-validate after this hook runs.
+	 *
+	 * The hook receives the tool abort signal (`signal`) and is responsible for
+	 * honoring it. Throwing surfaces as a tool-error result and does not abort the
+	 * rest of the batch.
+	 */
+	beforeToolCall?: (
+		context: BeforeToolCallContext,
+		signal?: AbortSignal,
+	) => Promise<BeforeToolCallResult | undefined> | BeforeToolCallResult | undefined;
+
+	/**
+	 * Called after a tool finishes executing, before `tool_execution_end` and the
+	 * tool-result message are emitted.
+	 *
+	 * Return an `AfterToolCallResult` to override individual fields of the executed
+	 * tool result. Omitted fields keep their original values; there is no deep merge.
+	 *
+	 * Throwing surfaces as a tool-error result and does not abort the rest of the batch.
+	 */
+	afterToolCall?: (
+		context: AfterToolCallContext,
+		signal?: AbortSignal,
+	) => Promise<AfterToolCallResult | undefined> | AfterToolCallResult | undefined;
+	/**
+	 * Opt-in OpenTelemetry instrumentation. Passing `{}` enables the loop's
+	 * GenAI-semantic-convention spans (`invoke_agent`, `chat`, `execute_tool`)
+	 * using the global tracer provider. Leaving this field undefined disables
+	 * the instrumentation entirely — the loop performs zero tracer lookups.
+	 *
+	 * See {@link AgentTelemetryConfig} for the full surface (hooks, content
+	 * capture, cost estimator, agent identity).
+	 */
+	telemetry?: AgentTelemetryConfig;
+}
+
+/**
+ * Batch/sequencing metadata for the tool call currently being processed.
+ */
+export interface ToolCallContext {
+	batchId: string;
+	index: number;
+	total: number;
+	toolCalls: Array<{ id: string; name: string }>;
+}
+
+/** A single tool-call content block emitted by an assistant message. */
+export type AgentToolCall = Extract<AssistantMessage["content"][number], { type: "toolCall" }>;
+
+/**
+ * Result returned from `beforeToolCall`.
+ *
+ * Set `block: true` to prevent the tool from executing. The loop emits an error tool
+ * result instead, using `reason` as the error text (or a default if omitted).
+ *
+ * Mutating the `args` reference passed in `BeforeToolCallContext` is supported and
+ * survives into execution — the loop does **not** re-validate after this hook runs.
+ */
+export interface BeforeToolCallResult {
+	block?: boolean;
+	reason?: string;
+}
+
+/**
+ * Partial override returned from `afterToolCall`.
+ *
+ * Merge semantics are field-by-field; omitted fields keep the executed values.
+ * No deep merge is performed.
+ */
+export interface AfterToolCallResult {
+	/** If provided, replaces the tool result content array in full. */
+	content?: (TextContent | ImageContent)[];
+	/** If provided, replaces the tool result details payload in full. */
+	details?: unknown;
+	/** If provided, replaces the error flag carried with the tool result. */
+	isError?: boolean;
+}
+
+/** Context passed to `beforeToolCall`. */
+export interface BeforeToolCallContext {
+	/** The assistant message that requested the tool call. */
+	assistantMessage: AssistantMessage;
+	/** The raw tool call block from `assistantMessage.content`. */
+	toolCall: AgentToolCall;
+	/**
+	 * Validated tool arguments. The same reference is forwarded to `tool.execute`
+	 * (after any `transformToolCallArguments` pass), so in-place mutations stick.
+	 */
+	args: Record<string, unknown>;
+	/** Current agent context at the time the tool call is prepared. */
+	context: AgentContext;
+}
+
+/** Context passed to `afterToolCall`. */
+export interface AfterToolCallContext {
+	/** The assistant message that requested the tool call. */
+	assistantMessage: AssistantMessage;
+	/** The raw tool call block from `assistantMessage.content`. */
+	toolCall: AgentToolCall;
+	/** Validated tool arguments used for execution (post `beforeToolCall` mutations). */
+	args: Record<string, unknown>;
+	/** The executed tool result before any `afterToolCall` overrides are applied. */
+	result: AgentToolResult<any>;
+	/** Whether the executed tool result is currently treated as an error. */
+	isError: boolean;
+	/** Current agent context at the time the tool call is finalized. */
+	context: AgentContext;
+}
+
+/**
+ * Extensible interface for custom app messages.
+ * Apps can extend via declaration merging:
+ *
+ * @example
+ * ```typescript
+ * declare module "@prometheus-ai/agent-core" {
+ *   interface CustomAgentMessages {
+ *     artifact: ArtifactMessage;
+ *     notification: NotificationMessage;
+ *   }
+ * }
+ * ```
+ */
+export interface CustomAgentMessages {
+	// Empty by default - apps extend via declaration merging
+}
+
+/**
+ * AgentMessage: Union of LLM messages + custom messages.
+ * This abstraction allows apps to add custom message types while maintaining
+ * type safety and compatibility with the base LLM messages.
+ */
+export type AgentMessage = Message | CustomAgentMessages[keyof CustomAgentMessages];
+
+/**
+ * Agent state containing all configuration and conversation data.
+ */
+export interface AgentState {
+	systemPrompt: string[];
+	model: Model;
+	thinkingLevel?: Effort;
+	tools: AgentTool<any>[];
+	messages: AgentMessage[]; // Can include attachments + custom message types
+	isStreaming: boolean;
+	streamMessage: AgentMessage | null;
+	pendingToolCalls: Set<string>;
+	error?: string;
+}
+
+export interface AgentToolResult<T = any, _TInput = unknown> {
+	// Content blocks supporting text and images
+	content: (TextContent | ImageContent)[];
+	// Details to be displayed in a UI or logged
+	details?: T;
+	// Marks a non-throwing failure (e.g. an aggregator catching per-entry errors).
+	// agent-loop honors this and surfaces it as a tool error on the wire.
+	isError?: boolean;
+}
+
+// Callback for streaming tool execution updates
+export type AgentToolUpdateCallback<T = any, TInput = unknown> = (partialResult: AgentToolResult<T, TInput>) => void;
+
+/** Options passed to renderResult */
+export interface RenderResultOptions {
+	/** Whether the result view is expanded */
+	expanded: boolean;
+	/** Whether this is a partial/streaming result */
+	isPartial: boolean;
+	/** Current spinner frame index for animated elements (optional) */
+	spinnerFrame?: number;
+}
+
+/** Capability tier a tool exercises. Determines which approval modes auto-approve it. */
+export type ToolTier = "read" | "write" | "exec";
+
+/**
+ * Per-tool approval declaration.
+ * - bare tier ("read" / "write" / "exec") — static classification.
+ * - object form — adds a `reason` (shown in the prompt) and/or `override: true`
+ *   (force-prompt even in modes that would otherwise auto-approve this tier).
+ * - function — dynamic, given parsed args. Returns either form above.
+ *
+ * Omitted approvals are treated as "exec" by callers that enforce approvals.
+ */
+export type ToolApprovalDecision = ToolTier | { tier: ToolTier; reason?: string; override?: boolean };
+export type ToolApproval = ToolApprovalDecision | ((args: unknown) => ToolApprovalDecision);
+
+/**
+ * Context passed to tool execution.
+ * Apps can extend via declaration merging.
+ */
+export interface AgentToolContext {
+	// Empty by default - apps extend via declaration merging
+}
+
+export type AgentToolExecFn<TParameters extends TSchema = TSchema, TDetails = any, TTheme = unknown> = (
+	this: AgentTool<TParameters, TDetails, TTheme>,
+	toolCallId: string,
+	params: Static<TParameters>,
+	signal?: AbortSignal,
+	onUpdate?: AgentToolUpdateCallback<TDetails, TParameters>,
+	context?: AgentToolContext,
+) => Promise<AgentToolResult<TDetails, TParameters>>;
+
+// AgentTool extends Tool but adds the execute function
+export interface AgentTool<TParameters extends TSchema = TSchema, TDetails = any, TTheme = unknown>
+	extends Tool<TParameters> {
+	// A human-readable label for the tool to be displayed in UI
+	label: string;
+	/** If true, tool is excluded unless explicitly listed in --tools or agent's tools field */
+	hidden?: boolean;
+	/** If true, tool can stage a pending action that requires explicit resolution via the resolve tool. */
+	deferrable?: boolean;
+	/** Built-in tool loading behavior. "essential" loads initially; "discoverable" can be activated by tool search. */
+	loadMode?: "essential" | "discoverable";
+	/** Short one-line summary used for tool discovery indexes. */
+	summary?: string;
+	/** If true, tool execution ignores abort signals (runs to completion) */
+	nonAbortable?: boolean;
+	/**
+	 * Concurrency mode for tool scheduling when multiple calls are in one turn.
+	 * - "shared": can run alongside other shared tools (default)
+	 * - "exclusive": runs alone; other tools wait until it finishes
+	 */
+	concurrency?: "shared" | "exclusive";
+	/** If true, argument validation errors are non-fatal: raw args are passed to execute() instead of returning an error to the LLM. */
+	lenientArgValidation?: boolean;
+	/**
+	 * Controls how the INTENT_FIELD (`_i`) is handled for this tool.
+	 * - `"require"` (default): `_i` is injected and required in the parameter schema.
+	 * - `"optional"`: `_i` is injected as an optional/nullable field.
+	 * - `"omit"`: `_i` is NOT injected. Use for tools where intent is obvious (yield, resolve, todo, …).
+	 * - function: `_i` is NOT injected; intent is derived dynamically from (potentially partial / streaming) args.
+	 */
+	intent?: "omit" | "optional" | "require" | ((args: Partial<Static<TParameters>>) => string | undefined);
+
+	/**
+	 * Normalize (potentially partial) streamed arguments into the plain text that
+	 * stream-content matchers (e.g. TTSR rules) should inspect — the real content
+	 * the call introduces, without wire grammar such as patch prefixes or JSON
+	 * string escaping. Return `undefined` to fall back to raw argument-delta
+	 * matching.
+	 */
+	matcherDigest?: (args: unknown) => string | undefined;
+
+	/** Capability tier declaration used by approval gates. Omitted means "exec". */
+	approval?: ToolApproval;
+
+	/** Lines appended after the standard approval prompt header. */
+	formatApprovalDetails?: (args: unknown) => string | string[] | undefined;
+
+	/** The main execution callback for this tool. */
+	execute: AgentToolExecFn<TParameters, TDetails, TTheme>;
+
+	/** Optional custom rendering for tool call display (returns UI component) */
+	renderCall?: (args: Static<TParameters>, options: RenderResultOptions, theme: TTheme) => unknown;
+
+	/** Optional custom rendering for tool result display (returns UI component) */
+	renderResult?: (
+		result: AgentToolResult<TDetails, TParameters>,
+		options: RenderResultOptions,
+		theme: TTheme,
+	) => unknown;
+}
+
+// AgentContext is like Context but uses AgentTool
+export interface AgentContext {
+	systemPrompt: string[];
+	messages: AgentMessage[];
+	tools?: AgentTool<any>[];
+}
+
+/**
+ * Events emitted by the Agent for UI updates.
+ * These events provide fine-grained lifecycle information for messages, turns, and tool executions.
+ */
+export type AgentEvent =
+	// Agent lifecycle
+	| { type: "agent_start" }
+	| {
+			type: "agent_end";
+			messages: AgentMessage[];
+			/** Present iff `AgentTelemetryConfig` was supplied on this run. */
+			telemetry?: AgentRunSummary;
+			coverage?: AgentRunCoverage;
+	  }
+	// Turn lifecycle - a turn is one assistant response + any tool calls/results
+	| { type: "turn_start" }
+	| { type: "turn_end"; message: AgentMessage; toolResults: ToolResultMessage[] }
+	// Message lifecycle - emitted for user, assistant, and toolResult messages
+	| { type: "message_start"; message: AgentMessage }
+	// Only emitted for assistant messages during streaming
+	| { type: "message_update"; message: AgentMessage; assistantMessageEvent: AssistantMessageEvent }
+	| { type: "message_end"; message: AgentMessage }
+	// Tool execution lifecycle
+	| { type: "tool_execution_start"; toolCallId: string; toolName: string; args: any; intent?: string }
+	| { type: "tool_execution_update"; toolCallId: string; toolName: string; args: any; partialResult: any }
+	| { type: "tool_execution_end"; toolCallId: string; toolName: string; result: any; isError?: boolean };

package/src/utils/yield.ts

@@ -0,0 +1,146 @@
+/**
+ * Cooperative yield utility for preventing Bun event-loop busy-wait.
+ *
+ * ## Root Cause
+ *
+ * Bun 1.3.x (JavaScriptCore) event loop busy-waits (spins in userspace)
+ * when the only pending work is an unresolved Promise — even if there are
+ * active I/O watchers (stdin, child process pipes, etc.).  The event loop
+ * continuously polls for microtask resolution instead of blocking in
+ * `epoll_wait`, consuming ~100% of a CPU core.
+ *
+ * This affects any `await` on a never-resolved Promise, including:
+ * - `Promise.withResolvers()` used for user input callbacks
+ * - `await proc.exited` for long-running child processes
+ * - Agent loop iterations waiting for the next tool call
+ *
+ * ## Fix
+ *
+ * A recurring `setInterval` keeps the event loop sleeping in `epoll_wait`.
+ * The `EventLoopKeepalive` class and `keepaliveWhile()` wrapper provide a
+ * clean way to install and clean up this keepalive timer.
+ *
+ * The older `yieldIfDue()` and `ExponentialYield` approaches (compensated
+ * sleep loops) are retained for the agent-loop hot-path where Promises
+ * resolve frequently and the keepalive alone is insufficient.
+ */
+
+import { scheduler } from "node:timers/promises";
+
+// ---------------------------------------------------------------------------
+// EventLoopKeepalive — the primary fix for idle-state busy-wait
+// ---------------------------------------------------------------------------
+
+export class EventLoopKeepalive {
+	#tmr = setInterval(() => {}, 86_400_000).unref();
+	[Symbol.dispose](): void {
+		clearInterval(this.#tmr);
+	}
+}
+
+// ---------------------------------------------------------------------------
+// yieldIfDue — retained for agent-loop hot-path
+// ---------------------------------------------------------------------------
+
+const YIELD_SLEEP_MS = 20;
+const YIELD_INTERVAL_MS = 50;
+
+/**
+ * Wall-clock timestamp of the last completed yield. Module-level so that
+ * tight loops sharing this helper collectively respect the gate, not just
+ * one caller at a time.
+ */
+let lastYieldAt = 0;
+
+/**
+ * Sleep for at least `ms` milliseconds of wall-clock time.
+ * Retries the wait if it returns prematurely (which can happen when napi
+ * callbacks wake the event loop via `uv_async_send`). When `signal` is
+ * provided, the wait is cancellable and silently returns on abort instead
+ * of throwing — callers race against another promise that decides what to
+ * do next.
+ */
+async function sleepAtLeast(ms: number, signal?: AbortSignal): Promise<void> {
+	const start = performance.now();
+	let remaining = ms;
+	while (remaining > 0) {
+		if (signal?.aborted) return;
+		try {
+			await scheduler.wait(remaining, { signal });
+		} catch (err) {
+			if ((err as { name?: string })?.name === "AbortError") return;
+			throw err;
+		}
+		remaining = ms - (performance.now() - start);
+	}
+}
+
+/**
+ * Yield to the Bun event loop, sleeping for at least 20 ms — but at most
+ * once every {@link YIELD_INTERVAL_MS}. Callers in hot paths can invoke
+ * this freely; only the slow path actually sleeps.
+ */
+export async function yieldIfDue(): Promise<void> {
+	const now = Date.now();
+	if (now - lastYieldAt < YIELD_INTERVAL_MS) return;
+	await sleepAtLeast(YIELD_SLEEP_MS);
+	lastYieldAt = Date.now();
+}
+
+// ---------------------------------------------------------------------------
+// ExponentialYield — retained for bash-executor long waits
+// ---------------------------------------------------------------------------
+
+const EXP_DEFAULT_MIN_MS = 20;
+const EXP_DEFAULT_MAX_MS = 10_000;
+const EXP_DEFAULT_MULTIPLIER = 2;
+
+export class ExponentialYield {
+	#currentMs: number;
+	readonly #minMs: number;
+	readonly #maxMs: number;
+	readonly #multiplier: number;
+
+	constructor(opts?: { minMs?: number; maxMs?: number; multiplier?: number }) {
+		this.#minMs = opts?.minMs ?? EXP_DEFAULT_MIN_MS;
+		this.#maxMs = opts?.maxMs ?? EXP_DEFAULT_MAX_MS;
+		this.#multiplier = opts?.multiplier ?? EXP_DEFAULT_MULTIPLIER;
+		this.#currentMs = this.#minMs;
+	}
+
+	notifyActivity(): void {
+		this.#currentMs = this.#minMs;
+	}
+
+	async sleep(signal?: AbortSignal): Promise<number> {
+		const ms = this.#currentMs;
+		await sleepAtLeast(ms, signal);
+		this.#currentMs = Math.min(this.#currentMs * this.#multiplier, this.#maxMs);
+		return ms;
+	}
+
+	/**
+	 * Race `racers` against an exponentially-backed-off cooperative yield.
+	 * The losing sleep is cancelled as soon as a racer settles, so no stray
+	 * timers keep the event loop alive past the racer's resolution.
+	 */
+	async race<T>(racers: Array<Promise<T>>): Promise<T> {
+		const racer = Promise.race(racers);
+		const controller = new AbortController();
+		try {
+			const yieldMarker = Symbol("exp-yield");
+			for (;;) {
+				const result = await Promise.race<T | typeof yieldMarker>([
+					racer,
+					this.sleep(controller.signal).then(() => yieldMarker as T | typeof yieldMarker),
+				]);
+				if (result !== yieldMarker) {
+					this.notifyActivity();
+					return result;
+				}
+			}
+		} finally {
+			controller.abort();
+		}
+	}
+}