@remit/imap-worker 0.0.71 → 0.0.73

package/src/filter-config.test.ts

@@ -1,171 +0,0 @@
-import assert from "node:assert/strict";
-import { describe, it } from "node:test";
-import type {
-	FilterItem,
-	IFilterAnchorRepository,
-	IFilterRepository,
-	IMessageLabelRepository,
-} from "@remit/data-ports";
-import { FilterMatchOperator, FilterState } from "@remit/domain-enums";
-import {
-	type FilterMessage,
-	FilterPipeline,
-	type MessageEmbedder,
-	NO_ACTION,
-	type PlacementMoveService,
-} from "@remit/mailbox-service";
-import { buildFilterConfig, type FilterConfigDeps } from "./filter-config.js";
-import { _resetMessageEmbedderForTest } from "./message-embedder.js";
-
-const anchorOnlyFilter = (destinationMailboxId: string): FilterItem =>
-	({
-		filterId: "flt-semantic",
-		accountConfigId: "cfg-1",
-		name: "receipts",
-		scope: "Standing",
-		state: FilterState.Active,
-		hasAnchor: true,
-		ruleChangedAt: 1,
-		matchOperator: FilterMatchOperator.Or,
-		literalClauses: [],
-		actionLabelId: NO_ACTION,
-		actionMailboxId: destinationMailboxId,
-		createdAt: 1,
-		updatedAt: 1,
-	}) as unknown as FilterItem;
-
-const CURRENT_MODEL = "test-model@3";
-
-const repos = (filter: FilterItem): FilterConfigDeps => ({
-	filterService: {
-		listByAccountAndState: async () => [filter],
-		refreshExpiry: async (f: FilterItem) => f,
-	} as unknown as IFilterRepository,
-	filterAnchorService: {
-		get: async () => ({
-			anchorEmbedding: [1, 0, 0],
-			anchorEmbeddingId: CURRENT_MODEL,
-		}),
-	} as unknown as IFilterAnchorRepository,
-	messageLabelService: {} as unknown as IMessageLabelRepository,
-	placementMoveService: {} as unknown as PlacementMoveService,
-});
-
-const message: FilterMessage = {
-	from: "billing@stripe.com",
-	fromName: "Stripe",
-	subject: "Your receipt",
-	text: "Thanks for your payment",
-	listId: "",
-};
-
-describe("buildFilterConfig", () => {
-	it("returns undefined without a placement mover — no move path, filters stay off", () => {
-		const deps = repos(anchorOnlyFilter("mbx-archive"));
-		deps.placementMoveService = undefined;
-
-		assert.equal(buildFilterConfig(deps), undefined);
-	});
-
-	it("wires an embedder into the config", () => {
-		const config = buildFilterConfig(repos(anchorOnlyFilter("mbx-archive")));
-
-		assert.ok(config);
-		assert.ok(config.embedder);
-	});
-
-	it("lights up a semantic anchor-only filter on the body-sync pass", async () => {
-		let embedCalls = 0;
-		const embedder: MessageEmbedder = {
-			embed: async () => {
-				embedCalls += 1;
-				return [1, 0, 0];
-			},
-			embeddingId: CURRENT_MODEL,
-		};
-		const config = buildFilterConfig(
-			repos(anchorOnlyFilter("mbx-archive")),
-			embedder,
-		);
-		assert.ok(config);
-
-		const decision = await new FilterPipeline(config, {
-			info: () => {},
-		}).evaluate("cfg-1", "m-1", message);
-
-		assert.deepEqual(decision.move, {
-			destinationMailboxId: "mbx-archive",
-			filterId: "flt-semantic",
-		});
-		assert.equal(embedCalls, 1);
-	});
-
-	it("does not match when the message embedding diverges from the anchor", async () => {
-		const embedder: MessageEmbedder = {
-			embed: async () => [0, 1, 0],
-			embeddingId: CURRENT_MODEL,
-		};
-		const config = buildFilterConfig(
-			repos(anchorOnlyFilter("mbx-archive")),
-			embedder,
-		);
-		assert.ok(config);
-
-		const decision = await new FilterPipeline(config, {
-			info: () => {},
-		}).evaluate("cfg-1", "m-2", message);
-
-		assert.equal(decision.move, undefined);
-	});
-});
-
-// An instance with semantic search off (issue #1068). `off` selects an embedder
-// that throws by design, and the throw lands in the pipeline's per-filter catch:
-// one error-level `filter_anchor_match_failed` for every semantic filter on
-// every message it syncs. The pipeline already has a designed skip for a
-// missing embedder, so the config passes none.
-describe("buildFilterConfig with semantic search off", () => {
-	const withProviderOff = async (body: () => Promise<void>): Promise<void> => {
-		const previous = process.env.SEARCH_EMBEDDING_PROVIDER;
-		process.env.SEARCH_EMBEDDING_PROVIDER = "off";
-		_resetMessageEmbedderForTest();
-		try {
-			await body();
-		} finally {
-			if (previous === undefined) {
-				delete process.env.SEARCH_EMBEDDING_PROVIDER;
-			} else {
-				process.env.SEARCH_EMBEDDING_PROVIDER = previous;
-			}
-			_resetMessageEmbedderForTest();
-		}
-	};
-
-	it("wires no embedder", async () => {
-		await withProviderOff(async () => {
-			const config = buildFilterConfig(repos(anchorOnlyFilter("mbx-archive")));
-
-			assert.ok(config);
-			assert.equal(config.embedder, undefined);
-		});
-	});
-
-	it("skips the semantic filter without logging an error", async () => {
-		await withProviderOff(async () => {
-			const config = buildFilterConfig(repos(anchorOnlyFilter("mbx-archive")));
-			assert.ok(config);
-
-			const errors: unknown[] = [];
-			const decision = await new FilterPipeline(config, {
-				info: () => {},
-				debug: () => {},
-				warn: () => {},
-				error: (fields) => errors.push(fields.alert),
-			}).evaluate("cfg-1", "m-off", message);
-
-			assert.deepEqual(decision.labels, []);
-			assert.equal(decision.move, undefined);
-			assert.deepEqual(errors, []);
-		});
-	});
-});

package/src/filter-config.ts

@@ -1,63 +0,0 @@
-import type {
-	IFilterAnchorRepository,
-	IFilterRepository,
-	IMessageLabelRepository,
-} from "@remit/data-ports";
-import type {
-	FilterConfig,
-	MessageEmbedder,
-	PlacementMoveService,
-} from "@remit/mailbox-service";
-import {
-	EMBEDDING_PROVIDER_OFF,
-	readEmbeddingProviderFromEnv,
-} from "@remit/search-service/from-env";
-import { getMessageEmbedder } from "./message-embedder.js";
-
-export interface FilterConfigDeps {
-	filterService: IFilterRepository;
-	filterAnchorService: IFilterAnchorRepository;
-	messageLabelService: IMessageLabelRepository;
-	placementMoveService?: PlacementMoveService;
-}
-
-/**
- * The env-selected embedder, or none on an instance with semantic search off.
- *
- * `off` embeds nothing by design, and its embedder says so by throwing
- * (EmbeddingDisabledError). The pipeline has a designed skip for a *missing*
- * embedder — semantic filters are passed over and logged at debug — but a
- * present embedder that throws lands in the per-filter catch instead, which is
- * an error-level `filter_anchor_match_failed` for every semantic filter on
- * every synced message. Same outcome, a log nobody can read.
- */
-const embedderFromEnv = (): MessageEmbedder | undefined =>
-	readEmbeddingProviderFromEnv() === EMBEDDING_PROVIDER_OFF
-		? undefined
-		: getMessageEmbedder();
-
-/**
- * Assemble the index-time filter config the body-sync pass runs (RFC 034). The
- * embedder is provisioned from env exactly as the backend read-path and
- * search-index worker provision theirs, so a semantic (anchor-only) filter is
- * evaluated on incoming mail instead of silently skipped.
- *
- * Absent the placement mover there is no move path, so filters stay off — a
- * matched filter's actions reuse the same enqueue plumbing the placement mover
- * owns.
- */
-export const buildFilterConfig = (
-	deps: FilterConfigDeps,
-	embedder?: MessageEmbedder,
-): FilterConfig | undefined => {
-	const { placementMoveService } = deps;
-	if (!placementMoveService) return undefined;
-	const resolved = embedder ?? embedderFromEnv();
-	return {
-		filterService: deps.filterService,
-		filterAnchorService: deps.filterAnchorService,
-		messageLabelService: deps.messageLabelService,
-		placementMoveService,
-		...(resolved ? { embedder: resolved } : {}),
-	};
-};

package/src/handlers/restore-source-placement.ts

@@ -1,135 +0,0 @@
-import type {
-	IMessageRepository,
-	IThreadMessageRepository,
-	MessageItem,
-	ThreadMessageItem,
-} from "@remit/data-ports";
-import { MessageStatus } from "@remit/domain-enums";
-import { isNotFoundError } from "../is-not-found.js";
-import { buildThreadMessageMoveRevert } from "./thread-message-rows.js";
-
-export interface RestoreSourcePlacementDeps {
-	messageService: Pick<IMessageRepository, "updateForMove">;
-	threadMessageService: Pick<
-		IThreadMessageRepository,
-		"findAllByMessageId" | "update"
-	>;
-}
-
-export interface RestoreSourcePlacementInput {
-	accountConfigId: string;
-	messageId: string;
-	sourceMailboxId: string;
-	uid: number;
-	/**
-	 * What the row records about the mutation that will now never happen.
-	 * `synced` where the row is a faithful projection of the source again — the
-	 * server never received the command, or has just said the message is still
-	 * there. `failed` where the product refused an operation the user asked for,
-	 * which is a failure the row has to carry.
-	 */
-	syncStatus: MessageItem["syncStatus"];
-	/**
-	 * The listing rows, where the caller has already read them to pick this
-	 * outcome. Omitted, they are read here.
-	 */
-	threadMessages?: ThreadMessageItem[];
-}
-
-/**
- * Put a row back on the pre-mutation source pair and hand its listing rows back
- * with it.
- *
- * `status: moving` has to be cleared here or the row is stuck: a give-up that
- * writes `syncStatus: failed` and returns leaves the row naming the destination
- * with the source's uid — the pair `bindsForeignUid` calls a lie and
- * `MessagePlacementUnsettledError` refuses to act on. Nothing routine repairs it
- * (sync does not touch `status`, and the cursor rebuild's `updateUid` runs only
- * on a UIDVALIDITY change), so the message is undeletable and unmovable for good
- * (issue #1005).
- *
- * `updateForMove`, not `updateUid`: both write the same settled pair, but
- * `updateUid` also appends a `message.moved` outbox row so the search index can
- * refresh the mailbox it stores per message. Nothing here moved. The optimistic
- * `updateForMove` that pointed the row at the destination never enqueued one
- * either, so the index still holds the source and a revert has nothing to
- * correct — the re-index would be a forced pass over an unchanged document.
- *
- * The caller must have evidence, not an inference: this writes a placement, and
- * a placement written on a guess binds live mail to a uid that names somebody
- * else's message. Two kinds of caller qualify, and nothing else does.
- *
- * A terminal resolver reaches here from {@link isMessageGoneFromOpenMailbox}
- * answering that the message is STILL at `sourceMailboxId`/`uid`, which makes
- * this the server's own answer rather than a revert on ambiguity (what PR #652
- * was pulled for).
- *
- * A paused-cursor caller reaches here from a `MailboxCursorPausedError` the
- * openBox guard threw before the outbound command was issued (issue #1203), and
- * restores the pre-mutation pair off its own event. On a FIRST delivery that is
- * not an inference but an undo: the command provably never left, so the row is
- * put back exactly as it was found. A redelivery cannot claim that — the
- * earlier attempt's tagged OK can be lost with the connection — so it asks
- * `probePausedPlacement` which folder holds the message and only reaches here
- * once the answer rules the destination out. Either way the source mailbox is
- * by definition awaiting a cursor rebuild, which matches its rows by Message-ID
- * and adjudicates the uid: handing the row back is what puts it in the set the
- * rebuild walks, since a row still naming the destination is in neither
- * folder's.
- *
- * A row another path deleted while the probe was in flight has nothing left to
- * restore, and a thread row whose composites have moved on is being rewritten
- * by that other path anyway. Both surface as `NotFoundError` — ElectroDB wraps
- * the conditional-check miss as one — and both are settled states here, not
- * faults. Throwing on them would escape the caller's `.catch()` and cost the
- * record its ack and its resync, which is the contract the terminal resolvers
- * state: never re-thrown.
- */
-export const restoreSourcePlacement = async (
-	deps: RestoreSourcePlacementDeps,
-	input: RestoreSourcePlacementInput,
-): Promise<void> => {
-	const { accountConfigId, messageId, sourceMailboxId, uid, syncStatus } =
-		input;
-
-	const skipNotFound = (error: unknown): void => {
-		if (isNotFoundError(error)) return;
-		throw error;
-	};
-
-	const messageRestored = await deps.messageService
-		.updateForMove(messageId, {
-			mailboxId: sourceMailboxId,
-			uid,
-			status: MessageStatus.active,
-			syncStatus,
-		})
-		.then(() => true)
-		.catch((error: unknown) => {
-			skipNotFound(error);
-			return false;
-		});
-	if (!messageRestored) return;
-
-	const threadMessages =
-		input.threadMessages ??
-		(await deps.threadMessageService.findAllByMessageId(
-			accountConfigId,
-			messageId,
-		));
-	for (const threadMessage of threadMessages) {
-		const args = buildThreadMessageMoveRevert(
-			threadMessage,
-			uid,
-			sourceMailboxId,
-		);
-		await deps.threadMessageService
-			.update(
-				threadMessage.accountConfigId,
-				threadMessage.threadMessageId,
-				args.set,
-				{ composites: args.composites },
-			)
-			.catch(skipNotFound);
-	}
-};

package/src/handlers/thread-message-rows.ts

@@ -1,77 +0,0 @@
-import type { ThreadMessageItem } from "@remit/data-ports";
-
-/**
- * Build the `set` and `composites` payload for the ThreadMessage update on a
- * MESSAGE_DELETE move-to-trash.
- *
- * The CURRENT row state goes in `composites`; the NEW values go in `set`.
- * ElectroDB uses `composites` to run the conditional check on the existing row
- * AND to compute the previous sort-key values needed to recompute the new ones.
- * Passing the NEW values in `composites` makes the conditional check fail with
- * ConditionalCheckFailedException, which ElectroDB wraps as NotFoundError, and
- * the caller silently drops the update. Same root cause as PR #186 fixed for
- * `flag-queue.ts`.
- */
-export type ThreadMessageRowState = Pick<
-	ThreadMessageItem,
-	| "sentDate"
-	| "mailboxId"
-	| "isRead"
-	| "isDeleted"
-	| "hasStars"
-	| "hasAttachment"
->;
-
-const currentComposites = (threadMessage: ThreadMessageRowState) => ({
-	sentDate: threadMessage.sentDate,
-	mailboxId: threadMessage.mailboxId,
-	isRead: threadMessage.isRead,
-	isDeleted: threadMessage.isDeleted,
-	hasStars: threadMessage.hasStars,
-	hasAttachment: threadMessage.hasAttachment,
-});
-
-export const buildThreadMessageTrashUpdate = (
-	threadMessage: ThreadMessageRowState,
-	newUid: number,
-	destinationMailboxId: string,
-) => ({
-	set: {
-		uid: newUid,
-		mailboxId: destinationMailboxId,
-		isDeleted: true,
-	},
-	composites: currentComposites(threadMessage),
-});
-
-/**
- * The inverse payload: put the thread row back where the message still is on
- * the server. The delete was recorded optimistically, so a delete that gives up
- * with the message still at its source has to hand the row back rather than
- * leave it claiming Trash — an invisible `failed` on a row the user cannot see
- * is the shape of the incident this whole change is about.
- */
-export const buildThreadMessageMoveRevert = (
-	threadMessage: ThreadMessageRowState,
-	sourceUid: number,
-	sourceMailboxId: string,
-) => ({
-	set: {
-		uid: sourceUid,
-		mailboxId: sourceMailboxId,
-		isDeleted: false,
-	},
-	composites: currentComposites(threadMessage),
-});
-
-/**
- * Hand a row back after an abandoned expunge. The mail never left Trash, so
- * only the deletion mark reverts — the uid and mailbox on the row are still
- * where the server has it.
- */
-export const buildThreadMessageUndelete = (
-	threadMessage: ThreadMessageRowState,
-) => ({
-	set: { isDeleted: false },
-	composites: currentComposites(threadMessage),
-});

package/src/is-not-found.ts

@@ -1,17 +0,0 @@
-/**
- * A repository lookup or write for a row that no longer exists rejects with a
- * `NotFoundError` (name-matched, since the class crosses the data-ports adapter
- * boundary — see `@remit/data-ports/errors`).
- *
- * A worker event that references a deliberately-deleted mailbox is completed or
- * moot work, never a transient fault: every redelivery re-throws the same
- * `NotFoundError`, and because the sync queues carry `MessageGroupId=accountId`
- * that permanently-failing head message stalls the whole account's per-group
- * FIFO. Handlers use this predicate to resolve such an event terminally (ack
- * with a WARN) instead of retrying forever (issues #287, #289, #290).
- *
- * The guard is narrow on purpose: only a genuine not-found terminates. Real
- * IMAP/infra failures carry other errors and must still propagate to be retried.
- */
-export const isNotFoundError = (error: unknown): boolean =>
-	(error as { name?: string })?.name === "NotFoundError";

package/src/message-embedder.test.ts

@@ -1,39 +0,0 @@
-import assert from "node:assert/strict";
-import { afterEach, describe, it } from "node:test";
-import {
-	_resetMessageEmbedderForTest,
-	getMessageEmbedder,
-} from "./message-embedder.js";
-
-describe("getMessageEmbedder", () => {
-	afterEach(() => {
-		_resetMessageEmbedderForTest();
-	});
-
-	it("adapts the batch embedding service into a single-text message embedder", async () => {
-		const vector = await getMessageEmbedder().embed("invoice from stripe");
-
-		assert.ok(Array.isArray(vector));
-		assert.ok(vector.length > 0);
-		assert.ok(vector.every((value) => typeof value === "number"));
-	});
-
-	it("embeds the same text deterministically", async () => {
-		const embedder = getMessageEmbedder();
-		const a = await embedder.embed("your monthly receipt");
-		const b = await embedder.embed("your monthly receipt");
-
-		assert.deepEqual(a, b);
-	});
-
-	it("memoizes the embedder across calls", () => {
-		assert.strictEqual(getMessageEmbedder(), getMessageEmbedder());
-	});
-
-	it("rebuilds after a reset", () => {
-		const first = getMessageEmbedder();
-		_resetMessageEmbedderForTest();
-
-		assert.notStrictEqual(getMessageEmbedder(), first);
-	});
-});

package/src/message-embedder.ts

@@ -1,33 +0,0 @@
-import type { MessageEmbedder } from "@remit/mailbox-service";
-import { buildEmbeddingServiceFromEnv } from "@remit/search-service/from-env";
-
-/**
- * Adapt the batch {@link EmbeddingService} the rest of the stack composes from
- * env into the single-text {@link MessageEmbedder} the filter pipeline needs: one
- * incoming message becomes one message-level vector to compare against a filter's
- * persisted anchor (RFC 034 Decision 2.1/2.3).
- *
- * The embedder is selected by the same `SEARCH_EMBEDDING_*` env the backend and
- * search-index worker read, so the worker embeds under the identical model the
- * anchors were built with. The instance is memoized across warm invocations so a
- * local Transformers.js model loads once per container rather than per event.
- */
-let cached: MessageEmbedder | undefined;
-
-export const getMessageEmbedder = (): MessageEmbedder => {
-	if (cached) return cached;
-	const service = buildEmbeddingServiceFromEnv();
-	cached = {
-		embed: async (text: string): Promise<number[]> => {
-			const [vector] = await service.embed([text]);
-			return vector;
-		},
-		embeddingId: service.embeddingId,
-	};
-	return cached;
-};
-
-/** Reset the singleton — test use only. */
-export const _resetMessageEmbedderForTest = (): void => {
-	cached = undefined;
-};