plasalid 0.7.1 → 0.7.3

package/dist/scanner/engine/types.d.ts

@@ -0,0 +1,90 @@
+import type Database from "libsql";
+import type { ScannedFile } from "../walker.js";
+import type { ScanBuffer } from "../buffer/types.js";
+import type { Bus } from "../bus/types.js";
+export type MaybePromise<T> = T | Promise<T>;
+export interface Chunk {
+    readonly chunkId: string;
+    readonly fileId: string;
+    readonly fileName: string;
+    readonly relPath: string;
+    readonly pageNumber: number;
+    readonly totalPages: number;
+    readonly bytes: Buffer;
+    readonly mime: string;
+}
+export interface DecryptedFile {
+    readonly path: string;
+    readonly fileName: string;
+    readonly relPath: string;
+    readonly hash: string;
+    readonly mime: string;
+    readonly decryptedBytes: Buffer;
+    readonly replacesPriorScannedFileId?: string;
+}
+export interface SkippedFile {
+    readonly file: ScannedFile;
+    readonly existingScannedFileId: string;
+}
+export interface FailedFile {
+    readonly file: ScannedFile;
+    readonly error: string;
+}
+export interface PhaseError {
+    readonly phase: PhaseName;
+    readonly target?: string;
+    readonly error: unknown;
+}
+export type PhaseName = "decrypt" | "chunk" | "parse" | "review" | "commit";
+export type ReviewDecision = "commit" | "abort";
+export interface CommitOutcome {
+    readonly transactions: number;
+    readonly accounts: number;
+    readonly merchants: number;
+    readonly unknowns: number;
+    readonly scannedFileIds: readonly string[];
+}
+export interface RunScanOptions {
+    regex?: string;
+    force?: boolean;
+    interactive?: boolean;
+    /** Max FileWorkers running concurrently. Default 5, hard cap 8. */
+    maxFileWorkers?: number;
+    /** Max ChunkWorkers per FileWorker. Default 5, hard cap 8. */
+    maxChunkWorkersPerFile?: number;
+    review?: boolean;
+    autoCommit?: boolean;
+    /**
+     * Override the phase chain. Default = the five built-in phases. Extending
+     * tests / alternate flows (dry-run, OCR-only) inject their own without
+     * editing the engine. Open for extension; closed for modification.
+     */
+    phases?: ReadonlyArray<{
+        name: PhaseName;
+        phase: Phase;
+    }>;
+}
+/**
+ * The state object threaded through every phase. Phases mutate it in place;
+ * hooks read it. The buffer + bus are interfaces — the engine owns the
+ * factory instances and injects them here.
+ */
+export interface ScanState {
+    readonly scanId: string;
+    readonly startedAt: number;
+    readonly options: RunScanOptions;
+    readonly buffer: ScanBuffer;
+    readonly bus: Bus;
+    files: ScannedFile[];
+    decrypted: DecryptedFile[];
+    skipped: SkippedFile[];
+    failed: FailedFile[];
+    chunks: Chunk[];
+    review: ReviewDecision | null;
+    committed: CommitOutcome | null;
+    errors: PhaseError[];
+    auditApplied: Record<string, number>;
+}
+import type { ScanHooks } from "../hooks/types.js";
+export type Phase = (db: Database.Database, state: ScanState, hooks: ScanHooks) => Promise<void>;
+export type { ScanHooks } from "../hooks/types.js";

package/dist/scanner/engine/types.js

@@ -0,0 +1 @@
+export {};

package/dist/scanner/engine.d.ts

@@ -0,0 +1,90 @@
+import type Database from "libsql";
+import type { ScannedFile } from "./walker.js";
+import type { ScanHooks } from "./hooks.js";
+import type { ScanProgress } from "./progress.js";
+import type { ClarifySummary } from "./clarifier.js";
+export interface Chunk {
+    readonly chunkId: string;
+    readonly fileId: string;
+    readonly fileName: string;
+    readonly relPath: string;
+    readonly pageNumber: number;
+    readonly totalPages: number;
+    readonly bytes: Buffer;
+    readonly mime: string;
+}
+export interface DecryptedFile {
+    readonly path: string;
+    readonly fileName: string;
+    readonly relPath: string;
+    readonly hash: string;
+    readonly mime: string;
+    readonly decryptedBytes: Buffer;
+    readonly replacesPriorScannedFileId?: string;
+    /** scanned_files.id assigned in decryptPhase so scan-worker tools can stamp source_file_id. */
+    scannedFileId?: string;
+}
+export interface SkippedFile {
+    readonly file: ScannedFile;
+    readonly existingScannedFileId: string;
+}
+export interface FailedFile {
+    readonly file: ScannedFile;
+    readonly error: string;
+}
+export interface PhaseError {
+    readonly phase: PhaseName;
+    readonly target?: string;
+    readonly error: unknown;
+}
+export type PhaseName = "decrypt" | "chunk" | "parse" | "clarify";
+export interface RunScanOptions {
+    regex?: string;
+    force?: boolean;
+    interactive?: boolean;
+    /** Max files processed concurrently. Default 5, hard cap 8. */
+    maxFileWorkers?: number;
+    /** Max scan workers per file (one per chunk). Default 5, hard cap 8. */
+    maxScanWorkersPerFile?: number;
+    /**
+     * Override the phase chain. Default = the four built-ins. Tests and alternate
+     * flows (dry-run, OCR-only) inject their own without editing this file.
+     */
+    phases?: ReadonlyArray<{
+        name: PhaseName;
+        phase: Phase;
+    }>;
+}
+/**
+ * The state object threaded through every phase. Phases mutate it in place;
+ * hooks read it. `progress` is the single-consumer event sink scan-worker
+ * tools emit into; the CLI subscribes to drive the dashboard.
+ */
+export interface ScanState {
+    readonly scanId: string;
+    readonly startedAt: number;
+    readonly options: RunScanOptions;
+    readonly progress: ScanProgress;
+    files: ScannedFile[];
+    decrypted: DecryptedFile[];
+    skipped: SkippedFile[];
+    failed: FailedFile[];
+    chunks: Chunk[];
+    clarifySummary: ClarifySummary | null;
+    errors: PhaseError[];
+}
+export type Phase = (db: Database.Database, state: ScanState, hooks: ScanHooks) => Promise<void>;
+export interface ScanResult {
+    readonly scanId: string;
+    readonly state: ScanState;
+}
+export declare const DEFAULT_PHASES: readonly {
+    name: PhaseName;
+    phase: Phase;
+}[];
+/**
+ * Composition root. Builds the progress sink once per scan run, threads it
+ * through ScanState, then runs the phase chain. Nothing survives between
+ * scans.
+ */
+export declare function runScan(db: Database.Database, opts?: RunScanOptions, hooks?: ScanHooks): Promise<ScanResult>;

package/dist/scanner/engine.js

@@ -0,0 +1,84 @@
+import { randomUUID } from "crypto";
+import { createProgress } from "./progress.js";
+import { decryptPhase } from "./decrypt.js";
+import { parsePhase } from "./parse.js";
+import { chunkPdf } from "./pdf/chunker.js";
+import { runClarify } from "./clarifier.js";
+import { errorMessage } from "./result.js";
+const chunkPhase = async (_db, state, hooks) => {
+    await hooks.beforeChunk?.(state);
+    for (const file of state.decrypted)
+        state.chunks.push(...await chunkPdf(file));
+    await hooks.afterChunk?.(state);
+};
+const clarifyPhase = async (db, state, hooks) => {
+    await hooks.beforeClarify?.(state);
+    const summary = await runClarify({
+        db,
+        scanId: state.scanId,
+        interactive: state.options.interactive ?? true,
+    });
+    state.clarifySummary = summary;
+    await hooks.afterClarify?.(state, summary);
+};
+export const DEFAULT_PHASES = [
+    { name: "decrypt", phase: decryptPhase },
+    { name: "chunk", phase: chunkPhase },
+    { name: "parse", phase: parsePhase },
+    { name: "clarify", phase: clarifyPhase },
+];
+/**
+ * Composition root. Builds the progress sink once per scan run, threads it
+ * through ScanState, then runs the phase chain. Nothing survives between
+ * scans.
+ */
+export async function runScan(db, opts = {}, hooks = {}) {
+    const scanId = `sc:${randomUUID()}`;
+    const progress = createProgress();
+    const state = {
+        scanId,
+        startedAt: Date.now(),
+        options: opts,
+        progress,
+        files: [],
+        decrypted: [],
+        skipped: [],
+        failed: [],
+        chunks: [],
+        clarifySummary: null,
+        errors: [],
+    };
+    await fire(hooks.onStart, state);
+    const phases = opts.phases ?? DEFAULT_PHASES;
+    await runPhaseChain(db, state, hooks, phases);
+    await fire(hooks.onFinish, state);
+    return { scanId, state };
+}
+async function runPhaseChain(db, state, hooks, phases) {
+    for (const { name, phase } of phases) {
+        const aborted = await tryPhase(db, state, hooks, name, phase);
+        if (aborted)
+            return;
+    }
+}
+async function tryPhase(db, state, hooks, name, phase) {
+    try {
+        await phase(db, state, hooks);
+        return false;
+    }
+    catch (err) {
+        state.errors.push({ phase: name, error: err });
+        await fire(hooks.onError, err, name, state);
+        return true;
+    }
+}
+async function fire(fn, ...args) {
+    if (!fn)
+        return;
+    try {
+        await fn(...args);
+    }
+    catch (err) {
+        console.error(`[scan-engine] ${errorMessage(err)}`);
+    }
+}

package/dist/scanner/file-worker.d.ts

@@ -0,0 +1,33 @@
+import type Database from "libsql";
+import { type ChunkWorkerDeps, type ChunkWorkerResult } from "./chunk-worker.js";
+import type { Chunk } from "./engine.js";
+import type { ScanHooks } from "./hooks.js";
+import type { ScanProgress } from "./progress.js";
+/**
+ * Pluggable chunk-parser strategy. Default is the LLM-driven runChunkWorker;
+ * tests and alternate flows (OCR-only, mock, dry-run) inject their own.
+ */
+export type ChunkWorkerFn = (deps: ChunkWorkerDeps, hooks: ScanHooks) => Promise<ChunkWorkerResult>;
+export interface FileWorkerDeps {
+    readonly db: Database.Database;
+    readonly scanId: string;
+    readonly scannedFileId: string | undefined;
+    readonly progress: ScanProgress;
+    readonly fileId: string;
+    readonly chunks: readonly Chunk[];
+    readonly maxChunkWorkers: number;
+    /** Optional override; defaults to the LLM-backed runChunkWorker. */
+    readonly chunkWorkerFn?: ChunkWorkerFn;
+}
+export interface FileWorkerResult {
+    readonly workerId: string;
+    readonly fileId: string;
+    readonly ok: number;
+    readonly failed: number;
+}
+/**
+ * Process every chunk of one file in parallel up to `maxChunkWorkers`.
+ * Chunk-worker tools write transactions + unknowns directly to the DB,
+ * scoped to `scanId`; per-row ticks fan out via the shared progress sink.
+ */
+export declare function runFileWorker(deps: FileWorkerDeps, hooks: ScanHooks): Promise<FileWorkerResult>;

package/dist/scanner/file-worker.js

@@ -0,0 +1,28 @@
+import { randomUUID } from "crypto";
+import { runWithConcurrency } from "./concurrency.js";
+import { runChunkWorker } from "./chunk-worker.js";
+/**
+ * Process every chunk of one file in parallel up to `maxChunkWorkers`.
+ * Chunk-worker tools write transactions + unknowns directly to the DB,
+ * scoped to `scanId`; per-row ticks fan out via the shared progress sink.
+ */
+export async function runFileWorker(deps, hooks) {
+    const workerId = `fw:${randomUUID()}`;
+    const chunkFn = deps.chunkWorkerFn ?? runChunkWorker;
+    const tasks = deps.chunks.map(chunk => () => chunkFn({
+        db: deps.db,
+        scanId: deps.scanId,
+        scannedFileId: deps.scannedFileId,
+        progress: deps.progress,
+        chunk,
+    }, hooks));
+    const settled = await runWithConcurrency(tasks, deps.maxChunkWorkers);
+    let ok = 0, failed = 0;
+    for (const r of settled) {
+        if (r.ok && r.value.ok)
+            ok++;
+        else
+            failed++;
+    }
+    return { workerId, fileId: deps.fileId, ok, failed };
+}

package/dist/scanner/fileWorker.d.ts

@@ -0,0 +1,33 @@
+import type Database from "libsql";
+import { type ChunkWorkerDeps, type ChunkWorkerResult } from "./chunkWorker.js";
+import type { ScanBuffer } from "./buffer.js";
+import type { Chunk } from "./engine.js";
+import type { ScanHooks } from "./hooks.js";
+/**
+ * Pluggable chunk-parser strategy. Default is the LLM-driven runChunkWorker;
+ * tests and alternate flows (OCR-only, mock, dry-run) inject their own
+ * without modifying the file worker. Open for extension; closed for
+ * modification.
+ */
+export type ChunkWorkerFn = (deps: ChunkWorkerDeps, hooks: ScanHooks) => Promise<ChunkWorkerResult>;
+export interface FileWorkerDeps {
+    readonly db: Database.Database;
+    readonly buffer: ScanBuffer;
+    readonly fileId: string;
+    readonly chunks: readonly Chunk[];
+    readonly maxChunkWorkers: number;
+    /** Optional override; defaults to the LLM-backed runChunkWorker. */
+    readonly chunkWorkerFn?: ChunkWorkerFn;
+}
+export interface FileWorkerResult {
+    readonly workerId: string;
+    readonly fileId: string;
+    readonly ok: number;
+    readonly failed: number;
+}
+/**
+ * Process every chunk of one file in parallel up to `maxChunkWorkers`. The
+ * shared buffer + bus are dependency-injected; this function has no global
+ * state and never reaches outside its args.
+ */
+export declare function runFileWorker(deps: FileWorkerDeps, hooks: ScanHooks): Promise<FileWorkerResult>;

package/dist/scanner/fileWorker.js

@@ -0,0 +1,22 @@
+import { randomUUID } from "crypto";
+import { runWithConcurrency } from "./concurrency.js";
+import { runChunkWorker } from "./chunkWorker.js";
+/**
+ * Process every chunk of one file in parallel up to `maxChunkWorkers`. The
+ * shared buffer + bus are dependency-injected; this function has no global
+ * state and never reaches outside its args.
+ */
+export async function runFileWorker(deps, hooks) {
+    const workerId = `fw:${randomUUID()}`;
+    const chunkFn = deps.chunkWorkerFn ?? runChunkWorker;
+    const tasks = deps.chunks.map(chunk => () => chunkFn({ db: deps.db, buffer: deps.buffer, chunk }, hooks));
+    const settled = await runWithConcurrency(tasks, deps.maxChunkWorkers);
+    let ok = 0, failed = 0;
+    for (const r of settled) {
+        if (r.ok && r.value.ok)
+            ok++;
+        else
+            failed++;
+    }
+    return { workerId, fileId: deps.fileId, ok, failed };
+}

package/dist/scanner/hooks/types.d.ts

@@ -0,0 +1,25 @@
+import type { Chunk, ScanState, CommitOutcome, PhaseName } from "../engine/types.js";
+import type { BufferSnapshot } from "../buffer/types.js";
+export type MaybePromise<T> = T | Promise<T>;
+/**
+ * Lifecycle hooks the scanner engine fires at phase edges. CLI registers
+ * spinner/Ink hooks; tests register assertions. Every hook is optional and
+ * best-effort — a hook that throws gets logged and the phase continues.
+ */
+export interface ScanHooks {
+    onStart?(s: Readonly<ScanState>): MaybePromise<void>;
+    beforeDecrypt?(s: Readonly<ScanState>): MaybePromise<void>;
+    afterDecrypt?(s: Readonly<ScanState>): MaybePromise<void>;
+    beforeChunk?(s: Readonly<ScanState>): MaybePromise<void>;
+    afterChunk?(s: Readonly<ScanState>): MaybePromise<void>;
+    beforeParse?(s: Readonly<ScanState>): MaybePromise<void>;
+    onWorkerStart?(workerId: string, chunk: Chunk): void;
+    onWorkerEnd?(workerId: string, chunk: Chunk, ok: boolean): void;
+    afterParse?(s: Readonly<ScanState>): MaybePromise<void>;
+    beforeReview?(s: Readonly<ScanState>, snapshot: BufferSnapshot): MaybePromise<void>;
+    afterReview?(s: Readonly<ScanState>): MaybePromise<void>;
+    beforeCommit?(s: Readonly<ScanState>): MaybePromise<void>;
+    afterCommit?(s: Readonly<ScanState>, outcome: CommitOutcome): MaybePromise<void>;
+    onError?(err: unknown, phase: PhaseName, s: Readonly<ScanState>): MaybePromise<void>;
+    onFinish?(s: Readonly<ScanState>): MaybePromise<void>;
+}

package/dist/scanner/hooks/types.js

@@ -0,0 +1 @@
+export {};

package/dist/scanner/hooks.d.ts

@@ -0,0 +1,23 @@
+import type { Chunk, ScanState, PhaseName } from "./engine.js";
+import type { ClarifySummary } from "./clarifier.js";
+export type MaybePromise<T> = T | Promise<T>;
+/**
+ * Lifecycle hooks the engine fires at phase edges. CLI registers spinner/Ink
+ * hooks; tests register assertions. Every hook is optional and best-effort —
+ * a hook that throws gets logged and the phase continues.
+ */
+export interface ScanHooks {
+    onStart?(s: Readonly<ScanState>): MaybePromise<void>;
+    beforeDecrypt?(s: Readonly<ScanState>): MaybePromise<void>;
+    afterDecrypt?(s: Readonly<ScanState>): MaybePromise<void>;
+    beforeChunk?(s: Readonly<ScanState>): MaybePromise<void>;
+    afterChunk?(s: Readonly<ScanState>): MaybePromise<void>;
+    beforeParse?(s: Readonly<ScanState>): MaybePromise<void>;
+    onWorkerStart?(workerId: string, chunk: Chunk): void;
+    onWorkerEnd?(workerId: string, chunk: Chunk, ok: boolean): void;
+    afterParse?(s: Readonly<ScanState>): MaybePromise<void>;
+    beforeClarify?(s: Readonly<ScanState>): MaybePromise<void>;
+    afterClarify?(s: Readonly<ScanState>, summary: ClarifySummary): MaybePromise<void>;
+    onError?(err: unknown, phase: PhaseName, s: Readonly<ScanState>): MaybePromise<void>;
+    onFinish?(s: Readonly<ScanState>): MaybePromise<void>;
+}

package/dist/scanner/hooks.js

@@ -0,0 +1 @@
+export {};

package/dist/scanner/parse.d.ts

@@ -0,0 +1,10 @@
+import type Database from "libsql";
+import type { ScanState } from "./engine.js";
+import type { ScanHooks } from "./hooks.js";
+/**
+ * Phase 3 — two-tier fan-out: up to `maxFile` files in parallel, each file
+ * processing up to `maxChunk` chunks in parallel. Chunk-worker tools write
+ * transactions and questions directly to the DB (scoped to `scanId`) and tick
+ * the shared progress sink.
+ */
+export declare function parsePhase(db: Database.Database, state: ScanState, hooks: ScanHooks): Promise<void>;

package/dist/scanner/parse.js

@@ -0,0 +1,47 @@
+import { runWithConcurrency } from "./concurrency.js";
+import { runScanWorker } from "./worker.js";
+import { errorMessage } from "./result.js";
+const DEFAULT_MAX_FILE_WORKERS = 5;
+const DEFAULT_MAX_SCAN_WORKERS_PER_FILE = 5;
+const HARD_CAP = 8;
+const clamp = (n, fallback) => Math.min(HARD_CAP, Math.max(1, n ?? fallback));
+/**
+ * Phase 3 — two-tier fan-out: up to `maxFile` files in parallel, each file
+ * processing up to `maxChunk` chunks in parallel. Chunk-worker tools write
+ * transactions and questions directly to the DB (scoped to `scanId`) and tick
+ * the shared progress sink.
+ */
+export async function parsePhase(db, state, hooks) {
+    await hooks.beforeParse?.(state);
+    const maxFile = clamp(state.options.maxFileWorkers, DEFAULT_MAX_FILE_WORKERS);
+    const maxChunk = clamp(state.options.maxScanWorkersPerFile, DEFAULT_MAX_SCAN_WORKERS_PER_FILE);
+    const fileGroups = state.decrypted
+        .map(file => ({
+        fileId: file.path,
+        scannedFileId: file.scannedFileId,
+        chunks: state.chunks.filter(c => c.fileId === file.path),
+    }))
+        .filter(g => g.chunks.length > 0);
+    const fileTasks = fileGroups.map(group => () => {
+        const chunkTasks = group.chunks.map(chunk => () => runScanWorker({
+            db,
+            scanId: state.scanId,
+            scannedFileId: group.scannedFileId,
+            progress: state.progress,
+            chunk,
+        }, hooks));
+        return runWithConcurrency(chunkTasks, maxChunk);
+    });
+    const settled = await runWithConcurrency(fileTasks, maxFile);
+    for (let i = 0; i < settled.length; i++) {
+        const r = settled[i];
+        if (!r.ok)
+            state.errors.push({ phase: "parse", target: fileGroups[i].fileId, error: errorMessage(r.error) });
+    }
+    for (const file of state.decrypted) {
+        if (!file.scannedFileId)
+            continue;
+        db.prepare(`UPDATE scanned_files SET status = 'scanned', scanned_at = datetime('now') WHERE id = ?`).run(file.scannedFileId);
+    }
+    await hooks.afterParse?.(state);
+}

package/dist/scanner/passes/index.d.ts

@@ -0,0 +1,8 @@
+import type { AuditPass } from "./types.js";
+/**
+ * Audit-pass registry. The audit engine indexes these by their declared
+ * `kinds` at startup and dispatches matching passes per buffer event.
+ * Append to the array to extend; do not edit the engine.
+ */
+export declare const AUDIT_PASSES: readonly AuditPass[];
+export type { AuditPass };

package/dist/scanner/passes/index.js

@@ -0,0 +1,6 @@
+/**
+ * Audit-pass registry. The audit engine indexes these by their declared
+ * `kinds` at startup and dispatches matching passes per buffer event.
+ * Append to the array to extend; do not edit the engine.
+ */
+export const AUDIT_PASSES = [];

package/dist/scanner/passes/types.d.ts

@@ -0,0 +1,22 @@
+import type Database from "libsql";
+import type { BufferEvent, EventKind } from "../bus.js";
+import type { ScanBuffer } from "../buffer.js";
+export interface AuditContext {
+    readonly db: Database.Database;
+    readonly buffer: ScanBuffer;
+    /** Optional tally bucket — passes increment their own count for the review summary. */
+    readonly tally: Record<string, number>;
+}
+/**
+ * Deterministic rule-based reaction to a BufferEvent. Passes mutate the
+ * ScanBuffer (apply memory rules, dedup, link to recurrences, etc.) so data
+ * lands clean before the review TUI sees it.
+ *
+ * Declare which event kinds you care about up front via `kinds` — the audit
+ * engine indexes once and dispatches only the matching passes per event.
+ */
+export interface AuditPass {
+    readonly name: string;
+    readonly kinds: readonly EventKind[];
+    apply(event: BufferEvent, ctx: AuditContext): Promise<void>;
+}

package/dist/scanner/passes/types.js

@@ -0,0 +1 @@
+export {};

package/dist/scanner/pdf/chunker.d.ts

@@ -0,0 +1,7 @@
+import type { Chunk, DecryptedFile } from "../engine.js";
+/**
+ * Split one decrypted PDF into N single-page Chunks. Each chunk is a
+ * standalone, valid PDF so the per-chunk LLM agent gets a clean document
+ * without siblings.
+ */
+export declare function chunkPdf(file: DecryptedFile): Promise<Chunk[]>;

package/dist/scanner/pdf/chunker.js

@@ -0,0 +1,60 @@
+let mupdfPromise = null;
+function getMupdf() {
+    if (!mupdfPromise)
+        mupdfPromise = import("mupdf");
+    return mupdfPromise;
+}
+/**
+ * Build one Chunk holding exactly page `pageIndex` of `file`. mupdf has no
+ * native page-range extract, so we clone the source doc and delete every
+ * other page, back-to-front so indices stay stable as we splice. Resource
+ * lifetime is contained in the try/finally so a saveToBuffer failure can't
+ * leak the cloned doc.
+ */
+async function extractPage(file, pageIndex, pageCount) {
+    const mupdf = await getMupdf();
+    const clone = mupdf.Document.openDocument(file.decryptedBytes, file.mime);
+    try {
+        for (let j = pageCount - 1; j >= 0; j--) {
+            if (j !== pageIndex)
+                clone.deletePage(j);
+        }
+        const out = clone.saveToBuffer("decrypt");
+        return {
+            chunkId: `${file.path}#p${pageIndex + 1}`,
+            fileId: file.path,
+            fileName: file.fileName,
+            relPath: file.relPath,
+            pageNumber: pageIndex + 1,
+            totalPages: pageCount,
+            bytes: Buffer.from(out.asUint8Array()),
+            mime: file.mime,
+        };
+    }
+    finally {
+        clone.destroy();
+    }
+}
+/**
+ * Split one decrypted PDF into N single-page Chunks. Each chunk is a
+ * standalone, valid PDF so the per-chunk LLM agent gets a clean document
+ * without siblings.
+ */
+export async function chunkPdf(file) {
+    const mupdf = await getMupdf();
+    const probe = mupdf.Document.openDocument(file.decryptedBytes, file.mime);
+    let pageCount;
+    try {
+        pageCount = probe.countPages();
+    }
+    finally {
+        probe.destroy();
+    }
+    if (pageCount <= 0)
+        return [];
+    const chunks = [];
+    for (let i = 0; i < pageCount; i++) {
+        chunks.push(await extractPage(file, i, pageCount));
+    }
+    return chunks;
+}

package/dist/scanner/pdf/password-store.d.ts

@@ -0,0 +1,34 @@
+import type Database from "libsql";
+export interface StoredPassword {
+    id: string;
+    pattern: string;
+    password: string;
+    useCount: number;
+    lastUsedAt: string | null;
+}
+/**
+ * Derive a regex from a filename. Strategy: take the leading alphabetic-ish
+ * prefix (up to the first separator: underscore, hyphen, space, or dot) and
+ * wildcard everything after it. Looser than a literal match — `AcctSt_May26.pdf`
+ * and `AcctSt_Jun26.pdf` share the same pattern.
+ *
+ * Falls back to the older digit-collapse strategy when the prefix is too short
+ * (<3 chars) or doesn't start with a letter, so we don't end up with overly
+ * generic patterns like `^a.*` or `^\d+.*`.
+ *
+ * Examples:
+ *   `AcctSt_May26.pdf`          → `^acctst.*`
+ *   `KBank-Savings-2026-01.pdf` → `^kbank.*`
+ *   `statement.pdf`             → `^statement.*`
+ *   `1234567890.pdf`            → `^\d+\.pdf$`           (fallback)
+ *   `e-statement.pdf`           → `^e\-statement\.pdf$`  (fallback — prefix too short)
+ */
+export declare function suggestPattern(filename: string): string;
+/** Stored passwords whose pattern matches the basename of `filePath`. */
+export declare function findCandidates(db: Database.Database, filePath: string, dbKey: string): StoredPassword[];
+/**
+ * Upsert by pattern. If the pattern already exists the row is replaced — useful
+ * when the bank rotates the password for a recurring statement series.
+ */
+export declare function savePassword(db: Database.Database, pattern: string, password: string, dbKey: string): string;
+export declare function recordUse(db: Database.Database, id: string): void;