@bastani/atomic 0.9.16-alpha.1 → 0.9.16-alpha.3

package/dist/builtin/workflows/src/durable/tool-primitive.ts

@@ -1,861 +0,0 @@
-/**
- * `ctx.tool` primitive — durable cached execution of arbitrary TypeScript code.
- *
- * Runs a user-supplied async function and caches the result durably via the
- * {@link DurableWorkflowBackend}. On resume, if the tool already completed
- * (matched by content hash of name + args), the cached result is returned
- * without re-executing the function — ensuring completed side effects are not
- * repeated.
- *
- * Only `ctx.*` blocks produce durable checkpoints. Anything outside `ctx.*`
- * (including bare `await someFunction()`) is never saved, matching the issue's
- * requirement: "checkpoints are effectively only `ctx.*` blocks."
- *
- * cross-ref: issue #1498 — "Introduce ctx.tool which allows you to run any
- * typescript code and cache the result for DBOS."
- */
-
-import { runCallback } from "@bastani/atomic";
-import { sleepOrAbort } from "../runs/shared/retry.js";
-import { flattenTruncatedString } from "../shared/flat-string.js";
-import type { ToolNodeSnapshot } from "../shared/store-types.js";
-import { boundedToolPayloadRecord, boundedToolText } from "../shared/tool-payload-bounds.js";
-import type {
-	WorkflowSerializableValue,
-	WorkflowToolContext,
-	WorkflowToolOptions,
-	WorkflowToolOutcome,
-	WorkflowToolPrimitive,
-} from "../shared/types.js";
-import { field, hasProcessFailureEvidence, normalizeCode } from "../shared/workflow-failures-signals.js";
-
-export { sleepOrAbort } from "../runs/shared/retry.js";
-
-import type { DurableWorkflowBackend } from "./backend.js";
-import { durableHash } from "./backend.js";
-import { recordThrowingToolFailure } from "./tool-failure-checkpoint.js";
-import {
-	replayedWorkflowToolOutcome,
-	workflowToolFailure,
-	workflowToolOutcomeFromValue,
-	workflowToolSuccess,
-} from "./tool-outcome.js";
-import {
-	DURABLE_TOOL_TOPOLOGY_VERSION,
-	type DurableCheckpoint,
-	type DurableStageRunTopology,
-	type DurableToolCheckpoint,
-} from "./types.js";
-
-export type { WorkflowToolContext, WorkflowToolOptions, WorkflowToolPrimitive } from "../shared/types.js";
-
-type WorkflowToolInvocationResult<TValue extends WorkflowSerializableValue> = TValue | WorkflowToolOutcome<TValue>;
-
-export type WorkflowToolExecutionAdmission =
-	| { readonly accepted?: true; bindNode(nodeId: string): void; noteCancelled?(): void }
-	| { readonly accepted: false; readonly error: Error; bindNode(nodeId: string): void; noteCancelled?(): void };
-
-/** Per-node abort control published while one logical tool call is in flight. */
-export interface ToolNodeControlRegistration {
-	readonly nodeId: string;
-	readonly name: string;
-	readonly controller: AbortController;
-	readonly settled: Promise<void>;
-}
-
-interface ToolInvocationAdmissionControl {
-	bindNode(nodeId: string): void;
-	noteCancelled(): void;
-	/**
-	 * End this call's admission registration section. Called as soon as the live
-	 * node controller is published, or as soon as the call resolves without a
-	 * live callback, so a quit closing admission waits microseconds.
-	 */
-	releaseAdmission(): void;
-}
-
-/**
- * Result of entering the root-shared admission boundary. Structural on purpose:
- * the boundary itself lives in the engine, and `durable/` must not depend on it.
- */
-export type ToolCallAdmission =
-	| { readonly accepted: true; readonly lease: { release(): void } }
-	| { readonly accepted: false; readonly error: Error };
-
-export interface CreateToolPrimitiveInput {
-	readonly workflowId: string;
-	readonly backend: DurableWorkflowBackend;
-	/** Monotonic checkpoint id counter source. */
-	readonly nextCheckpointId: () => string;
-	/** Abort check; throws if the workflow has been cancelled. */
-	readonly throwIfCancelled: () => void;
-	/** Optional run-level signal; combined with the per-node signal handed to `fn`. */
-	readonly signal?: AbortSignal;
-	/**
-	 * Publish the per-node abort control so `/workflow quit|interrupt` can abort
-	 * one in-flight node. The returned disposer runs when the node settles.
-	 */
-	readonly registerNodeControl?: (registration: ToolNodeControlRegistration) => (() => void) | undefined;
-	/**
-	 * Enter the root-shared tool-admission boundary. A refusal means graceful
-	 * quit already closed admission for this workflow tree; the call never runs.
-	 */
-	readonly admitToolCall?: () => ToolCallAdmission;
-	/** Deterministic boundary hooks before dispatch and after a tool node settles. */
-	readonly beforeToolCall?: () => void | Promise<void>;
-	readonly afterToolCall?: () => void | Promise<void>;
-	readonly trackExecution?: <T>(execution: Promise<T>) => WorkflowToolExecutionAdmission | undefined;
-	/** Observe a logical throwing-mode failure before graph publication or promise rejection. */
-	readonly onFailureObserved?: (error: unknown, nodeId: string) => void;
-	/** Admit/update a first-class graph node around the durable call. */
-	readonly onNodeStart?: (node: ToolNodeSnapshot) => void;
-	readonly onNodeRunning?: (nodeId: string, startedAt: number) => void;
-	readonly onNodeEnd?: (
-		nodeId: string,
-		update: Pick<ToolNodeSnapshot, "status"> &
-			Partial<Pick<ToolNodeSnapshot, "endedAt" | "durationMs" | "result" | "resultSummary" | "error">>,
-	) => void;
-	readonly onNodeSettle?: (nodeId: string) => void;
-	readonly runTopology?: DurableStageRunTopology;
-}
-
-function cloneToolArgs(
-	args: Readonly<Record<string, WorkflowSerializableValue>>,
-): Readonly<Record<string, WorkflowSerializableValue>> {
-	// Bounded, hostile-input safe, and always a plain object: author args may be
-	// cyclic, may throw on property access, may be arbitrarily large, and may
-	// carry a `toJSON` that would otherwise collapse the record to a scalar in a
-	// durable checkpoint. `argsHash` still hashes the raw args, so cache identity
-	// is unaffected by this inspection copy.
-	return boundedToolPayloadRecord(args);
-}
-
-/**
- * Capture the callback's own source for read-only inspection.
- *
- * `Function.prototype.toString` reads text the runtime already holds: it never
- * invokes the callback, never touches the filesystem, and never reaches any
- * data beyond the function passed to this call. A host that refuses the read
- * simply yields no source row.
- */
-function captureCallbackSource(fn: unknown): string | undefined {
-	if (typeof fn !== "function") return undefined;
-	try {
-		return boundedToolText(Function.prototype.toString.call(fn));
-	} catch {
-		return undefined;
-	}
-}
-
-/**
- * Create the `ctx.tool` primitive wired to a durable backend.
- */
-export function createToolPrimitive(input: CreateToolPrimitiveInput): WorkflowToolPrimitive {
-	const ordinals = new Map<string, number>();
-	return (<T extends WorkflowSerializableValue>(
-		name: string,
-		args: Readonly<Record<string, WorkflowSerializableValue>>,
-		fn: (toolCtx: WorkflowToolContext) => Promise<T>,
-		options?: WorkflowToolOptions,
-	): Promise<WorkflowToolInvocationResult<T>> => {
-		let resolveExecution!: (
-			value: WorkflowToolInvocationResult<T> | PromiseLike<WorkflowToolInvocationResult<T>>,
-		) => void;
-		let rejectExecution!: (reason?: unknown) => void;
-		const execution = new Promise<WorkflowToolInvocationResult<T>>((resolve, reject) => {
-			resolveExecution = resolve;
-			rejectExecution = reject;
-		});
-		// The admission boundary is checked before tracker admission so a refused
-		// post-quit call is a suspension, never an observed run failure.
-		const boundaryAdmission = input.admitToolCall?.();
-		if (boundaryAdmission?.accepted === false) {
-			void execution.catch(() => undefined);
-			rejectExecution(boundaryAdmission.error);
-			return execution;
-		}
-		const lease = boundaryAdmission?.accepted === true ? boundaryAdmission.lease : undefined;
-		const admission = input.trackExecution?.(execution);
-		if (admission?.accepted === false) {
-			lease?.release();
-			void execution.catch(() => undefined);
-			rejectExecution(admission.error);
-			return execution;
-		}
-		const control: ToolInvocationAdmissionControl = {
-			bindNode: (nodeId) => admission?.bindNode(nodeId),
-			noteCancelled: () => admission?.noteCancelled?.(),
-			releaseAdmission: () => lease?.release(),
-		};
-		const settleAfterBoundary = (settle: () => void): void => {
-			let afterToolCall: void | Promise<void>;
-			try {
-				afterToolCall = input.afterToolCall?.();
-			} catch (error) {
-				rejectExecution(error);
-				return;
-			}
-			if (afterToolCall === undefined) {
-				settle();
-				return;
-			}
-			void afterToolCall.then(settle, rejectExecution);
-		};
-		void executeToolInvocation(input, ordinals, name, args, fn, options, control, captureCallbackSource(fn))
-			// Backstop for a throw before either explicit release point; the lease
-			// release itself is idempotent.
-			.finally(() => lease?.release())
-			.then(
-				(value) => settleAfterBoundary(() => resolveExecution(value)),
-				(error) => settleAfterBoundary(() => rejectExecution(error)),
-			);
-		return execution;
-	}) as WorkflowToolPrimitive;
-}
-
-async function executeToolInvocation<T extends WorkflowSerializableValue>(
-	input: CreateToolPrimitiveInput,
-	ordinals: Map<string, number>,
-	name: string,
-	args: Readonly<Record<string, WorkflowSerializableValue>>,
-	fn: (toolCtx: WorkflowToolContext) => Promise<T>,
-	options: WorkflowToolOptions | undefined,
-	control: ToolInvocationAdmissionControl,
-	source: string | undefined,
-): Promise<WorkflowToolInvocationResult<T>> {
-	input.throwIfCancelled();
-	const beforeToolCall = input.beforeToolCall?.();
-	if (beforeToolCall !== undefined) await beforeToolCall;
-	input.throwIfCancelled();
-	if (
-		options?.retriesAllowed === true &&
-		options.maxAttempts !== undefined &&
-		(!Number.isInteger(options.maxAttempts) || options.maxAttempts < 1)
-	) {
-		throw new RangeError("atomic-workflows: ctx.tool maxAttempts must be a positive integer");
-	}
-	if (options?.timeoutMs !== undefined && (!Number.isFinite(options.timeoutMs) || options.timeoutMs <= 0)) {
-		throw new RangeError("atomic-workflows: ctx.tool timeoutMs must be a positive finite number");
-	}
-	const returnFailure = options?.failureMode === "return";
-	const identityMode = returnFailure ? { failureMode: "return" as const } : {};
-	const callKey = durableHash({ name, args, ...identityMode });
-	const ordinal = (ordinals.get(callKey) ?? 0) + 1;
-	ordinals.set(callKey, ordinal);
-	const argsHash = durableHash({ name, args, ordinal, ...identityMode });
-
-	const cached = input.backend.getToolCheckpoint(input.workflowId, argsHash);
-	const capturedSource = cached?.source ?? source;
-	const node: ToolNodeSnapshot = {
-		kind: "tool",
-		id: cached?.topology?.nodeId ?? `tool:${argsHash}`,
-		name,
-		args: cloneToolArgs(cached?.args ?? args),
-		...(capturedSource !== undefined ? { source: capturedSource } : {}),
-		argsHash,
-		ordinal: cached?.topology?.ordinal ?? ordinal,
-		parentIds: Object.freeze(cached?.topology?.parentIds ?? []),
-		status: "pending",
-		...(cached !== undefined && cached.topology === undefined ? { topologyState: "unavailable" as const } : {}),
-		...(cached !== undefined ? { replayed: true } : {}),
-		...(cached?.topology?.order !== undefined ? { executionOrder: cached.topology.order } : {}),
-		...(cached?.topology?.startedAt !== undefined ? { startedAt: cached.topology.startedAt } : {}),
-		attachable: false,
-	};
-	control.bindNode(node.id);
-	input.onNodeStart?.(node);
-	if (cached !== undefined) {
-		// A replayed call has no live callback to abort, so its registration
-		// section ends here rather than at node-controller publication.
-		control.releaseAdmission();
-		const endedAt = cached.topology?.endedAt ?? cached.completedAt;
-		try {
-			await recordReplayedToolTopology(input, node, cached, argsHash, endedAt);
-			const returnedOutcome =
-				cached.outcomeKind === undefined ? undefined : workflowToolOutcomeFromValue<T>(cached.output);
-			if (cached.outcomeKind !== undefined && returnedOutcome === undefined) {
-				throw new Error(`atomic-workflows: invalid durable return outcome for ctx.tool ${name}`);
-			}
-			const output =
-				returnedOutcome === undefined ? (cached.output as T) : replayedWorkflowToolOutcome(returnedOutcome);
-			const failed = cached.outcomeKind === "return_failure";
-			input.onNodeEnd?.(node.id, {
-				status: failed ? "failed" : "cached",
-				endedAt,
-				...(node.startedAt !== undefined ? { durationMs: Math.max(0, endedAt - node.startedAt) } : {}),
-				result: output,
-				...(failed && returnedOutcome?.ok === false
-					? { error: returnedOutcome.error.message }
-					: { resultSummary: summarizeToolResult(output) }),
-			});
-			input.onNodeSettle?.(node.id);
-			return output;
-		} catch (error) {
-			const cancelled = input.signal?.aborted === true;
-			if (cancelled) control.noteCancelled();
-			else input.onFailureObserved?.(error, node.id);
-			const endedAt = Date.now();
-			input.onNodeEnd?.(node.id, {
-				status: cancelled ? "cancelled" : "failed",
-				endedAt,
-				...(node.startedAt !== undefined ? { durationMs: Math.max(0, endedAt - node.startedAt) } : {}),
-				error: error instanceof Error ? error.message : String(error),
-			});
-			input.onNodeSettle?.(node.id);
-			throw error;
-		}
-	}
-	return executeLiveToolInvocation({
-		input,
-		node,
-		name,
-		args: node.args ?? args,
-		...(capturedSource !== undefined ? { source: capturedSource } : {}),
-		argsHash,
-		ordinal,
-		fn,
-		options,
-		control,
-		returnFailure,
-	});
-}
-
-interface LiveToolInvocation<T extends WorkflowSerializableValue> {
-	readonly input: CreateToolPrimitiveInput;
-	readonly node: ToolNodeSnapshot;
-	readonly name: string;
-	readonly args: Readonly<Record<string, WorkflowSerializableValue>>;
-	readonly source?: string;
-	readonly argsHash: string;
-	readonly ordinal: number;
-	readonly fn: (toolCtx: WorkflowToolContext) => Promise<T>;
-	readonly options: WorkflowToolOptions | undefined;
-	readonly control: ToolInvocationAdmissionControl;
-	readonly returnFailure: boolean;
-}
-class WorkflowToolTimeoutError extends Error {
-	constructor(toolName: string, timeoutMs: number) {
-		super(`atomic-workflows: ctx.tool ${toolName} timed out after ${timeoutMs}ms`);
-		this.name = "TimeoutError";
-	}
-}
-
-const MAX_HOST_TIMER_MS = 2_147_483_647;
-
-/** Schedule a finite deadline without triggering host timer overflow clamping. */
-function scheduleDeadline(callback: () => void, timeoutMs: number): () => void {
-	let remainingMs = timeoutMs;
-	let timer: ReturnType<typeof setTimeout> | undefined;
-	const scheduleNextChunk = () => {
-		const chunkMs = Math.min(remainingMs, MAX_HOST_TIMER_MS);
-		timer = setTimeout(() => {
-			remainingMs -= chunkMs;
-			if (remainingMs > 0) scheduleNextChunk();
-			else callback();
-		}, chunkMs);
-	};
-	scheduleNextChunk();
-	return () => {
-		if (timer !== undefined) clearTimeout(timer);
-	};
-}
-
-type ToolCallbackAttemptResult<T> =
-	| { readonly kind: "value"; readonly value: T }
-	| { readonly kind: "error"; readonly error: unknown };
-
-function abortReason(signal: AbortSignal): Error {
-	return signal.reason instanceof Error ? signal.reason : new Error("atomic-workflows: workflow cancelled");
-}
-
-/** Run one callback attempt with a fresh signal and deadline. */
-async function executeTimedToolAttempt<T extends WorkflowSerializableValue>(
-	toolName: string,
-	runId: string,
-	fn: (toolCtx: WorkflowToolContext) => Promise<T>,
-	baseSignal: AbortSignal,
-	timeoutMs: number,
-): Promise<T> {
-	const timeoutController = new AbortController();
-	const attemptSignal = AbortSignal.any([baseSignal, timeoutController.signal]);
-	return runCallback({ kind: "workflow.ctx_tool", name: toolName, runId }, async () => {
-		const callbackResult: Promise<ToolCallbackAttemptResult<T>> = Promise.resolve()
-			.then(() => fn({ signal: attemptSignal }))
-			.then(
-				(value) => ({ kind: "value" as const, value }),
-				(error: unknown) => ({ kind: "error" as const, error }),
-			);
-		let cancelDeadline: (() => void) | undefined;
-		let onAbort: (() => void) | undefined;
-		let timeoutError: WorkflowToolTimeoutError | undefined;
-		const timeoutResult = new Promise<never>((_, reject) => {
-			cancelDeadline = scheduleDeadline(() => {
-				if (baseSignal.aborted) return;
-				timeoutError = new WorkflowToolTimeoutError(toolName, timeoutMs);
-				timeoutController.abort(timeoutError);
-				reject(timeoutError);
-			}, timeoutMs);
-		});
-		const cancellationResult = new Promise<never>((_, reject) => {
-			onAbort = () => {
-				if (timeoutError === undefined) reject(abortReason(baseSignal));
-			};
-			if (baseSignal.aborted) onAbort();
-			else baseSignal.addEventListener("abort", onAbort, { once: true });
-		});
-		try {
-			const result = await Promise.race([callbackResult, timeoutResult, cancellationResult]);
-			if (result.kind === "value") return result.value;
-			throw result.error;
-		} finally {
-			cancelDeadline?.();
-			if (onAbort !== undefined) baseSignal.removeEventListener("abort", onAbort);
-		}
-	});
-}
-
-/**
- * Execute one uncached tool call under its own abort controller.
- *
- * The callback signal combines the run signal with this node's controller, so a
- * run abort cascades to every node while `/workflow quit|interrupt` can abort
- * exactly one node without touching its siblings. A cancelled call never writes
- * a replayable checkpoint, so resume re-executes it at the same ordinal.
- */
-async function executeLiveToolInvocation<T extends WorkflowSerializableValue>(
-	live: LiveToolInvocation<T>,
-): Promise<WorkflowToolInvocationResult<T>> {
-	const { input, node, name, args, source, argsHash, ordinal, fn, options, control, returnFailure } = live;
-	const nodeController = new AbortController();
-	const toolSignal =
-		input.signal === undefined ? nodeController.signal : AbortSignal.any([input.signal, nodeController.signal]);
-	const settlement = Promise.withResolvers<void>();
-	const unregisterControl = input.registerNodeControl?.({
-		nodeId: node.id,
-		name,
-		controller: nodeController,
-		settled: settlement.promise,
-	});
-	// The node is now abortable by quit, so this call's admission registration
-	// section ends. Everything after this point is observable to `quitRun`.
-	control.releaseAdmission();
-	let callbackError: unknown;
-
-	const startedAt = Date.now();
-	let attempts = 0;
-	let callbackCompleted = false;
-	input.onNodeRunning?.(node.id, startedAt);
-	try {
-		const result = await executeWithRetries(
-			() => {
-				attempts += 1;
-				const callback =
-					options?.timeoutMs === undefined
-						? runCallback({ kind: "workflow.ctx_tool", name, runId: input.workflowId }, () =>
-								fn({ signal: toolSignal }),
-							)
-						: executeTimedToolAttempt(name, input.workflowId, fn, toolSignal, options.timeoutMs);
-				return callback.catch((error: unknown) => {
-					callbackError = error;
-					throw error;
-				});
-			},
-			options,
-			() => throwIfInvocationCancelled(input, toolSignal),
-			toolSignal,
-		);
-		callbackCompleted = true;
-		const output = returnFailure ? workflowToolSuccess(result, attempts, false) : result;
-
-		// Linearization policy: cancellation observed before persistence prevents
-		// a checkpoint. Once the durable write begins, a successful commit wins
-		// for this node; the root still observes its aborted signal and is killed.
-		// An abandoned callback that ignores its signal and returns late is
-		// caught here, so its value can never become a replayable checkpoint.
-		throwIfInvocationCancelled(input, toolSignal);
-		const completedAt = Date.now();
-		const checkpoint: DurableToolCheckpoint = {
-			kind: "tool",
-			workflowId: input.workflowId,
-			checkpointId: `tool:${argsHash}`,
-			name,
-			args: cloneToolArgs(args),
-			...(source !== undefined ? { source } : {}),
-			argsHash,
-			output,
-			...(returnFailure ? { outcomeKind: "return_success" as const } : {}),
-			completedAt,
-			topology: {
-				version: DURABLE_TOOL_TOPOLOGY_VERSION,
-				nodeId: node.id,
-				ordinal,
-				order: node.executionOrder ?? 0,
-				parentIds: [...node.parentIds],
-				startedAt,
-				endedAt: completedAt,
-				...(input.runTopology !== undefined ? { run: { ...input.runTopology } } : {}),
-			},
-		};
-		await recordCheckpointDurably(input.backend, checkpoint);
-		input.onNodeEnd?.(node.id, {
-			status: "completed",
-			endedAt: completedAt,
-			durationMs: Math.max(0, completedAt - startedAt),
-			result: output,
-			resultSummary: summarizeToolResult(output),
-		});
-		input.onNodeSettle?.(node.id);
-		return output;
-	} catch (error) {
-		const timeoutFailure = error instanceof WorkflowToolTimeoutError;
-		const cancellation = timeoutFailure ? undefined : invocationCancellation(input, toolSignal);
-		if (cancellation !== undefined) {
-			control.noteCancelled();
-			// A cancelled call never writes a replayable `tool:` checkpoint and never
-			// a `return_failure` outcome, so resume re-executes it at the same
-			// ordinal instead of replaying a cancellation as data. Return mode still
-			// keeps one inspection-only `tool-failure:` record, which the backend
-			// excludes from replay lookup. The record is written for every
-			// cancellation timing — while the callback awaits, when it throws, and
-			// when it fulfills after abort but before persistence — because this
-			// branch runs at most once per logical invocation.
-			if (returnFailure) {
-				await recordCancelledToolInspection(live, startedAt, cancellation, attempts);
-			}
-			const endedAt = Date.now();
-			input.onNodeEnd?.(node.id, {
-				status: "cancelled",
-				endedAt,
-				durationMs: Math.max(0, endedAt - startedAt),
-				error: cancellation.message,
-			});
-			input.onNodeSettle?.(node.id);
-			throw cancellation;
-		}
-		const callbackFailure = callbackError ?? error;
-		if (returnFailure && !callbackCompleted && isExplicitCallbackCancellation(callbackFailure)) {
-			input.onFailureObserved?.(callbackFailure, node.id);
-			const failure = await recordThrowingToolFailure(
-				input,
-				node,
-				{ name, args, ...(source !== undefined ? { source } : {}), argsHash, ordinal, startedAt },
-				callbackFailure,
-				attempts,
-			);
-			input.onNodeEnd?.(node.id, {
-				status: "failed",
-				endedAt: failure.failedAt,
-				durationMs: Math.max(0, failure.failedAt - startedAt),
-				error: failure.message,
-			});
-			input.onNodeSettle?.(node.id);
-			throw callbackFailure;
-		}
-		if (returnFailure && !callbackCompleted) {
-			const outcome = workflowToolFailure(callbackFailure, attempts, false);
-			const completedAt = Date.now();
-			const checkpoint: DurableToolCheckpoint = {
-				kind: "tool",
-				workflowId: input.workflowId,
-				checkpointId: `tool:${argsHash}`,
-				name,
-				args: cloneToolArgs(args),
-				...(source !== undefined ? { source } : {}),
-				argsHash,
-				output: outcome,
-				outcomeKind: "return_failure",
-				completedAt,
-				topology: {
-					version: DURABLE_TOOL_TOPOLOGY_VERSION,
-					nodeId: node.id,
-					ordinal,
-					order: node.executionOrder ?? 0,
-					parentIds: [...node.parentIds],
-					startedAt,
-					endedAt: completedAt,
-					...(input.runTopology !== undefined ? { run: { ...input.runTopology } } : {}),
-				},
-			};
-			try {
-				await recordCheckpointDurably(input.backend, checkpoint);
-			} catch (persistenceError) {
-				input.onFailureObserved?.(persistenceError, node.id);
-				const endedAt = Date.now();
-				input.onNodeEnd?.(node.id, {
-					status: "failed",
-					endedAt,
-					durationMs: Math.max(0, endedAt - startedAt),
-					error: persistenceError instanceof Error ? persistenceError.message : String(persistenceError),
-				});
-				input.onNodeSettle?.(node.id);
-				throw persistenceError;
-			}
-			input.onNodeEnd?.(node.id, {
-				status: "failed",
-				endedAt: completedAt,
-				durationMs: Math.max(0, completedAt - startedAt),
-				result: outcome,
-				error: outcome.error.message,
-			});
-			input.onNodeSettle?.(node.id);
-			if (!timeoutFailure) throwIfInvocationCancelled(input, toolSignal);
-			return outcome;
-		}
-		const reportedFailure = timeoutFailure ? callbackError : error;
-		input.onFailureObserved?.(reportedFailure, node.id);
-		const failure = await recordThrowingToolFailure(
-			input,
-			node,
-			{ name, args, ...(source !== undefined ? { source } : {}), argsHash, ordinal, startedAt },
-			reportedFailure,
-			attempts,
-		);
-		input.onNodeEnd?.(node.id, {
-			status: "failed",
-			endedAt: failure.failedAt,
-			durationMs: Math.max(0, failure.failedAt - startedAt),
-			error: failure.message,
-		});
-		input.onNodeSettle?.(node.id);
-		throw reportedFailure;
-	} finally {
-		settlement.resolve();
-		unregisterControl?.();
-	}
-}
-
-/**
- * Persist exactly one inspection-only cancellation record for a
- * `failureMode: "return"` call.
- *
- * The record uses the non-replayable `tool-failure:` id, so
- * `getToolCheckpoint()` still returns undefined and resume re-runs the call. It
- * is best effort: a diagnostic write failure must not turn cancellation into a
- * storage failure, and no `return_failure` outcome is ever written.
- */
-async function recordCancelledToolInspection<T extends WorkflowSerializableValue>(
-	live: LiveToolInvocation<T>,
-	startedAt: number,
-	cancellation: Error,
-	attempts: number,
-): Promise<void> {
-	try {
-		await recordThrowingToolFailure(
-			live.input,
-			live.node,
-			{
-				name: live.name,
-				args: live.args,
-				...(live.source !== undefined ? { source: live.source } : {}),
-				argsHash: live.argsHash,
-				ordinal: live.ordinal,
-				startedAt,
-			},
-			cancellation,
-			attempts,
-		);
-	} catch {
-		// Inspection metadata is optional; the cancellation itself is authoritative.
-	}
-}
-
-/** Resolve the cancellation reason for this invocation, if it was cancelled. */
-function invocationCancellation(input: CreateToolPrimitiveInput, signal: AbortSignal): Error | undefined {
-	try {
-		input.throwIfCancelled();
-	} catch (error) {
-		return error instanceof Error ? error : new Error(String(error));
-	}
-	if (!signal.aborted) return undefined;
-	return signal.reason instanceof Error ? signal.reason : new Error("atomic-workflows: workflow cancelled");
-}
-
-function throwIfInvocationCancelled(input: CreateToolPrimitiveInput, signal: AbortSignal): void {
-	const cancellation = invocationCancellation(input, signal);
-	if (cancellation !== undefined) throw cancellation;
-}
-
-const CALLBACK_CANCELLATION_NAMES = new Set([
-	"aborterror",
-	"aborted",
-	"cancelederror",
-	"cancellederror",
-	"canceled",
-	"cancelled",
-]);
-const CALLBACK_CANCELLATION_CODES = new Set([
-	"aborterror",
-	"abort_err",
-	"aborted",
-	"canceled",
-	"cancelled",
-	"ecanceled",
-	"ecancelled",
-	"err_canceled",
-	"err_cancelled",
-]);
-const CALLBACK_CANCELLATION_STOP_REASONS = new Set(["aborted", "canceled", "cancelled"]);
-
-function safeCancellationField(error: unknown, key: string): unknown {
-	try {
-		return field(error, key);
-	} catch {
-		return undefined;
-	}
-}
-
-function hasCancellationMarker(markers: ReadonlySet<string>, value: unknown): boolean {
-	return (typeof value === "string" || typeof value === "number") && markers.has(normalizeCode(value) ?? "");
-}
-
-function isExplicitCallbackCancellation(error: unknown): boolean {
-	const seen = new Set<object>();
-	const pending: Array<{ readonly value: unknown; readonly depth: number }> = [{ value: error, depth: 0 }];
-	while (pending.length > 0) {
-		const current = pending.pop()!;
-		if (current.value === undefined || current.value === null || current.depth >= 8) continue;
-		if (typeof current.value === "object") {
-			if (seen.has(current.value)) continue;
-			seen.add(current.value);
-		}
-		const hasProcessFailure = hasProcessFailureEvidence(current.value);
-		if (
-			!hasProcessFailure &&
-			hasCancellationMarker(CALLBACK_CANCELLATION_NAMES, safeCancellationField(current.value, "name"))
-		)
-			return true;
-		if (
-			!hasProcessFailure &&
-			hasCancellationMarker(CALLBACK_CANCELLATION_STOP_REASONS, safeCancellationField(current.value, "stopReason"))
-		)
-			return true;
-		if (
-			!hasProcessFailure &&
-			hasCancellationMarker(CALLBACK_CANCELLATION_CODES, safeCancellationField(current.value, "code"))
-		)
-			return true;
-
-		for (const key of ["cause", "error", "response", "body"] as const) {
-			pending.push({ value: safeCancellationField(current.value, key), depth: current.depth + 1 });
-		}
-		for (const key of ["diagnostics", "errors"] as const) {
-			const entries = safeCancellationField(current.value, key);
-			if (!Array.isArray(entries)) continue;
-			for (const entry of entries) {
-				pending.push({
-					value: key === "diagnostics" ? (safeCancellationField(entry, "error") ?? entry) : entry,
-					depth: current.depth + 1,
-				});
-			}
-		}
-	}
-	return false;
-}
-
-async function recordReplayedToolTopology(
-	input: CreateToolPrimitiveInput,
-	node: ToolNodeSnapshot,
-	cached: DurableToolCheckpoint,
-	argsHash: string,
-	endedAt: number,
-): Promise<void> {
-	const runTopology = input.runTopology;
-	if (runTopology === undefined) return;
-	if (cached.topology === undefined && runTopology.parentRunId === undefined) return;
-	if (cached.topology?.run?.runId === runTopology.runId) return;
-	const topology =
-		cached.topology === undefined
-			? {
-					version: DURABLE_TOOL_TOPOLOGY_VERSION,
-					nodeId: node.id,
-					ordinal: node.ordinal,
-					order: node.executionOrder ?? 0,
-					parentIds: [...node.parentIds],
-					endedAt,
-					run: { ...runTopology },
-				}
-			: {
-					...cached.topology,
-					parentIds: [...node.parentIds],
-					order: node.executionOrder ?? cached.topology.order,
-					endedAt,
-					run: { ...runTopology },
-				};
-	const unchanged =
-		cached.topology !== undefined &&
-		JSON.stringify(cached.topology.parentIds) === JSON.stringify(topology.parentIds) &&
-		JSON.stringify(cached.topology.run) === JSON.stringify(topology.run) &&
-		cached.topology.endedAt === topology.endedAt;
-	if (unchanged) return;
-	await input.backend.recordAdditiveCheckpointBestEffort({
-		kind: "tool",
-		workflowId: input.workflowId,
-		checkpointId: `tool-replay-meta:${durableHash({ argsHash, topology })}`,
-		name: node.name,
-		...(cached.args !== undefined ? { args: cached.args } : {}),
-		...(cached.source !== undefined ? { source: cached.source } : {}),
-		argsHash,
-		output: cached.output,
-		...(cached.outcomeKind !== undefined ? { outcomeKind: cached.outcomeKind } : {}),
-		completedAt: Date.now(),
-		topology,
-	});
-	input.throwIfCancelled();
-}
-/**
- * Summaries are written into durable checkpoints and read back by status and
- * graph inspection, so they outlive the value they came from. Flatten the
- * truncation: a bare `slice` leaves a SlicedString pointing at the whole
- * serialized payload, and the 240-character bound would cap what you can read
- * without capping what is retained.
- */
-function summarizeToolResult(value: WorkflowSerializableValue): string {
-	const serialized = JSON.stringify(value);
-	if (serialized === undefined) return flattenTruncatedString(String(value).slice(0, 240));
-	return serialized.length <= 240 ? serialized : `${flattenTruncatedString(serialized.slice(0, 237))}...`;
-}
-
-export async function recordCheckpointDurably(
-	backend: DurableWorkflowBackend,
-	checkpoint: DurableCheckpoint,
-): Promise<void> {
-	await backend.recordCheckpointAsync(checkpoint);
-}
-
-async function executeWithRetries<T>(
-	fn: () => Promise<T>,
-	options: WorkflowToolOptions | undefined,
-	throwIfCancelled: () => void,
-	signal?: AbortSignal,
-): Promise<T> {
-	throwIfCancelled();
-	if (!options?.retriesAllowed) return fn();
-	const maxAttempts = options.maxAttempts ?? 3;
-	const intervalMs = options.intervalMs ?? 1000;
-	const backoffRate = options.backoffRate ?? 2;
-	let lastError: Error | undefined;
-	let delay = intervalMs;
-	for (let attempt = 1; attempt <= maxAttempts; attempt++) {
-		throwIfCancelled();
-		try {
-			return await fn();
-		} catch (err) {
-			if (isExplicitCallbackCancellation(err)) throw err;
-			lastError = err instanceof Error ? err : new Error(String(err));
-			if (attempt < maxAttempts) {
-				await sleepOrAbort(delay, signal);
-				throwIfCancelled();
-				delay = Math.min(delay * backoffRate, 3600_000);
-			}
-		}
-	}
-	throw lastError ?? new Error("ctx.tool: retries exhausted");
-}
-
-/**
- * Create a monotonic checkpoint id generator for a workflow.
- */
-export function createCheckpointIdGenerator(): () => string {
-	let counter = 0;
-	return () => `cp-${++counter}`;
-}