@gajae-code/coding-agent 0.17.2 → 0.17.4

package/vendor/markit-ai/dist/converters/iwork.js

@@ -0,0 +1,391 @@
+import { mkdirSync, writeFileSync } from "node:fs";
+import { join } from "node:path";
+import JSZip from "jszip";
+const EXTENSIONS = [".pages", ".key", ".numbers"];
+const SF = "http://developer.apple.com/namespaces/sf";
+const SFA = "http://developer.apple.com/namespaces/sfa";
+const KEY = "http://developer.apple.com/namespaces/keynote2";
+/**
+ * Converts Apple iWork files (Pages, Keynote, Numbers) to markdown.
+ *
+ * All three formats are ZIP archives containing an XML file:
+ *   - Pages:   index.xml   (sf:p paragraphs with named styles)
+ *   - Keynote: index.apxl  (key:slide elements with sf:p text)
+ *   - Numbers: index.xml   (sf:t text cells + sf:n number cells)
+ */
+export class IWorkConverter {
+    name = "iwork";
+    accepts(streamInfo) {
+        if (streamInfo.extension && EXTENSIONS.includes(streamInfo.extension)) {
+            return true;
+        }
+        return false;
+    }
+    async convert(input, streamInfo) {
+        const zip = await JSZip.loadAsync(input);
+        const ext = streamInfo.extension;
+        if (ext === ".pages")
+            return this.convertPages(zip, streamInfo);
+        if (ext === ".key")
+            return this.convertKeynote(zip, streamInfo);
+        if (ext === ".numbers")
+            return this.convertNumbers(zip);
+        throw new Error(`Unsupported iWork format: ${ext}`);
+    }
+    // ---------------------------------------------------------------------------
+    // Pages
+    // ---------------------------------------------------------------------------
+    async convertPages(zip, streamInfo) {
+        const xml = await this.readIndex(zip, "index.xml");
+        const root = parseXml(xml);
+        const imageDir = streamInfo.imageDir;
+        if (imageDir)
+            mkdirSync(imageDir, { recursive: true });
+        let imageCount = 0;
+        const lines = [];
+        let title;
+        for (const p of iterAll(root, SF, "p")) {
+            const text = collectText(p).trim();
+            if (!text)
+                continue;
+            const style = p.getAttribute("sf:style") || "";
+            const prefix = paragraphPrefix(style);
+            if (!title && text.length > 0)
+                title = text;
+            lines.push(`${prefix}${text}`);
+        }
+        // Extract images
+        for (const name of Object.keys(zip.files)) {
+            if (!name.match(/\.(png|jpg|jpeg|gif|webp|tiff|bmp)$/i))
+                continue;
+            if (name.startsWith("QuickLook/"))
+                continue;
+            imageCount++;
+            const imgName = name.split("/").pop() || `image_${imageCount}`;
+            if (imageDir) {
+                const file = zip.file(name);
+                if (file) {
+                    const buf = await file.async("nodebuffer");
+                    const filepath = join(imageDir, imgName);
+                    writeFileSync(filepath, buf);
+                    lines.push(`![${imgName}](${filepath})`);
+                }
+            }
+            else {
+                lines.push(`<!-- image: ${imgName} -->`);
+            }
+        }
+        return { markdown: lines.join("\n\n"), title };
+    }
+    // ---------------------------------------------------------------------------
+    // Keynote
+    // ---------------------------------------------------------------------------
+    async convertKeynote(zip, streamInfo) {
+        const xml = await this.readIndex(zip, "index.apxl");
+        const root = parseXml(xml);
+        const imageDir = streamInfo.imageDir;
+        if (imageDir)
+            mkdirSync(imageDir, { recursive: true });
+        const sections = [];
+        let title;
+        const slides = [...iterAll(root, KEY, "slide")];
+        for (let i = 0; i < slides.length; i++) {
+            const slide = slides[i];
+            const slideLines = [`<!-- Slide ${i + 1} -->`];
+            const paragraphs = [...iterAll(slide, SF, "p")];
+            let isTitle = true;
+            for (const p of paragraphs) {
+                const text = collectText(p).trim();
+                if (!text)
+                    continue;
+                if (isTitle) {
+                    slideLines.push(`# ${text}`);
+                    if (!title)
+                        title = text;
+                    isTitle = false;
+                }
+                else {
+                    slideLines.push(text);
+                }
+            }
+            sections.push(slideLines.join("\n"));
+        }
+        // Extract media images
+        let imageCount = 0;
+        for (const name of Object.keys(zip.files)) {
+            if (!name.match(/\.(png|jpg|jpeg|gif|webp|tiff|bmp)$/i))
+                continue;
+            if (name.startsWith("QuickLook/"))
+                continue;
+            imageCount++;
+            const imgName = name.split("/").pop() || `image_${imageCount}`;
+            if (imageDir) {
+                const file = zip.file(name);
+                if (file) {
+                    const buf = await file.async("nodebuffer");
+                    const filepath = join(imageDir, imgName);
+                    writeFileSync(filepath, buf);
+                    sections.push(`![${imgName}](${filepath})`);
+                }
+            }
+            else {
+                sections.push(`<!-- image: ${imgName} -->`);
+            }
+        }
+        return { markdown: sections.join("\n\n"), title };
+    }
+    // ---------------------------------------------------------------------------
+    // Numbers
+    // ---------------------------------------------------------------------------
+    async convertNumbers(zip) {
+        const xml = await this.readIndex(zip, "index.xml");
+        const root = parseXml(xml);
+        // Find grid elements (tables)
+        const grids = [...iterAll(root, SF, "grid")];
+        if (grids.length === 0) {
+            // Fallback: extract all text and number cells
+            return this.convertNumbersFallback(root);
+        }
+        const sections = [];
+        for (const grid of grids) {
+            const rows = this.extractGrid(grid);
+            if (rows.length === 0)
+                continue;
+            const maxCols = Math.max(...rows.map((r) => r.length));
+            for (const row of rows) {
+                while (row.length < maxCols)
+                    row.push("");
+            }
+            const [header, ...body] = rows;
+            const lines = [];
+            lines.push(`| ${header.join(" | ")} |`);
+            lines.push(`| ${header.map(() => "---").join(" | ")} |`);
+            for (const row of body) {
+                lines.push(`| ${row.join(" | ")} |`);
+            }
+            sections.push(lines.join("\n"));
+        }
+        return { markdown: sections.join("\n\n") };
+    }
+    extractGrid(grid) {
+        const datasource = findFirst(grid, SF, "datasource");
+        if (!datasource)
+            return [];
+        const rows = [];
+        let currentRow = [];
+        let colCount = 0;
+        let totalCells = 0;
+        const allValues = [];
+        // Get column count from grid attributes (raw attribute names)
+        const numCols = Number.parseInt(grid.getAttribute("sf:numcols") || "0", 10);
+        for (const child of datasource.children) {
+            const tag = child.tagName;
+            let value = "";
+            if (tag === `${SF}:t`) {
+                // Text cell
+                const ct = findFirst(child, SF, "ct");
+                value = ct?.getAttribute("sfa:s") || collectText(child).trim();
+            }
+            else if (tag === `${SF}:n`) {
+                // Number cell
+                value = child.getAttribute("sf:v") || "";
+            }
+            else if (tag === `${SF}:b`) {
+                // Boolean cell
+                value = child.getAttribute("sf:v") === "1" ? "TRUE" : "FALSE";
+            }
+            else if (tag === `${SF}:d`) {
+                // Date cell
+                value = child.getAttribute("sf:v") || "";
+            }
+            else if (tag === `${SF}:du`) {
+                // Duration cell
+                value = child.getAttribute("sf:v") || "";
+            }
+            else if (tag === `${SF}:e`) {
+                // Empty cell
+                value = "";
+            }
+            else {
+                continue;
+            }
+            currentRow.push(value);
+            allValues.push(value);
+            colCount++;
+            totalCells++;
+            if (numCols > 0 && colCount >= numCols) {
+                rows.push(currentRow);
+                currentRow = [];
+                colCount = 0;
+            }
+        }
+        if (currentRow.length > 0)
+            rows.push(currentRow);
+        // If the grid used default dimensions and produced only one row,
+        // the data probably doesn't fill the full grid width. Fall back to
+        // treating the cells as a 2-column key/value list or single column.
+        if (rows.length <= 1 && totalCells > 0 && totalCells < numCols) {
+            // Re-layout: try 2 columns if even, otherwise single column
+            const cols = totalCells % 2 === 0 ? 2 : 1;
+            const relaid = [];
+            for (let i = 0; i < allValues.length; i += cols) {
+                relaid.push(allValues.slice(i, i + cols));
+            }
+            return relaid;
+        }
+        return rows;
+    }
+    convertNumbersFallback(root) {
+        const values = [];
+        for (const t of iterAll(root, SF, "t")) {
+            const ct = findFirst(t, SF, "ct");
+            const val = ct?.getAttribute("sfa:s") || "";
+            if (val)
+                values.push(val);
+        }
+        for (const n of iterAll(root, SF, "n")) {
+            const val = n.getAttribute("sf:v") || "";
+            if (val)
+                values.push(val);
+        }
+        return { markdown: values.join("\n") };
+    }
+    // ---------------------------------------------------------------------------
+    // Helpers
+    // ---------------------------------------------------------------------------
+    async readIndex(zip, filename) {
+        const file = zip.file(filename);
+        if (!file) {
+            throw new Error(`Invalid iWork file: missing ${filename}`);
+        }
+        return file.async("string");
+    }
+}
+/**
+ * Minimal XML parser that preserves namespace prefixes in tag names
+ * and extracts text content and attributes.
+ */
+function parseXml(xml) {
+    // Use a simple recursive descent approach
+    const root = createElement("root");
+    const stack = [root];
+    // Match tags and text
+    const tagRe = /<(\/?)([a-zA-Z0-9_:.-]+)((?:\s+[a-zA-Z0-9_:.-]+\s*=\s*"[^"]*")*)\s*(\/?)>/g;
+    let lastIndex = 0;
+    let match = tagRe.exec(xml);
+    while (match !== null) {
+        const [fullMatch, isClose, tagName, attrs, isSelfClose] = match;
+        const textBefore = xml.slice(lastIndex, match.index);
+        lastIndex = match.index + fullMatch.length;
+        // Add text to current element
+        if (textBefore.trim()) {
+            const current = stack[stack.length - 1];
+            if (current.children.length > 0) {
+                current.children[current.children.length - 1].tail += textBefore.trim();
+            }
+            else {
+                current.text += textBefore.trim();
+            }
+        }
+        if (isClose) {
+            // Closing tag
+            stack.pop();
+        }
+        else {
+            // Opening tag
+            const el = createElement(expandTag(tagName, xml));
+            parseAttributes(attrs, el);
+            stack[stack.length - 1].children.push(el);
+            if (!isSelfClose) {
+                stack.push(el);
+            }
+        }
+        match = tagRe.exec(xml);
+    }
+    return root;
+}
+function createElement(tagName) {
+    return {
+        tagName,
+        children: [],
+        text: "",
+        tail: "",
+        attributes: {},
+        getAttribute(name) {
+            return this.attributes[name] ?? null;
+        },
+    };
+}
+function parseAttributes(attrStr, el) {
+    const attrRe = /([a-zA-Z0-9_:.-]+)\s*=\s*"([^"]*)"/g;
+    let m = attrRe.exec(attrStr);
+    while (m !== null) {
+        el.attributes[m[1]] = m[2];
+        m = attrRe.exec(attrStr);
+    }
+}
+/**
+ * Expand namespace prefix in tag name to full URI.
+ * e.g. "sf:p" with xmlns:sf="..." → "{uri}p"
+ * For simplicity, we use the known Apple namespaces.
+ */
+function expandTag(tag, _xml) {
+    const nsMap = {
+        sf: SF,
+        sfa: SFA,
+        sl: "http://developer.apple.com/namespaces/sl",
+        key: KEY,
+    };
+    const colon = tag.indexOf(":");
+    if (colon === -1)
+        return tag;
+    const prefix = tag.slice(0, colon);
+    const local = tag.slice(colon + 1);
+    const uri = nsMap[prefix];
+    return uri ? `${uri}:${local}` : tag;
+}
+function collectText(el) {
+    let result = el.text;
+    for (const child of el.children) {
+        result += collectText(child);
+        result += child.tail;
+    }
+    return result;
+}
+function* iterAll(el, ns, localName) {
+    const fullTag = `${ns}:${localName}`;
+    if (el.tagName === fullTag)
+        yield el;
+    for (const child of el.children) {
+        yield* iterAll(child, ns, localName);
+    }
+}
+function findFirst(el, ns, localName) {
+    for (const found of iterAll(el, ns, localName)) {
+        return found;
+    }
+    return null;
+}
+/**
+ * Map iWork paragraph style names to markdown heading prefixes.
+ */
+function paragraphPrefix(style) {
+    if (!style)
+        return "";
+    const lower = style.toLowerCase();
+    if (lower.includes("title"))
+        return "# ";
+    if (lower.includes("subtitle"))
+        return "## ";
+    if (lower.includes("heading-1") || lower.includes("heading 1"))
+        return "## ";
+    if (lower.includes("heading-2") || lower.includes("heading 2"))
+        return "### ";
+    if (lower.includes("heading-3") || lower.includes("heading 3"))
+        return "#### ";
+    if (lower.includes("heading-4") || lower.includes("heading 4"))
+        return "##### ";
+    if (lower.includes("caption"))
+        return "*";
+    return "";
+}

package/vendor/markit-ai/dist/converters/json.d.ts

@@ -0,0 +1,6 @@
+import type { ConversionResult, Converter, StreamInfo } from "../types.js";
+export declare class JsonConverter implements Converter {
+    name: string;
+    accepts(streamInfo: StreamInfo): boolean;
+    convert(input: Buffer, _streamInfo: StreamInfo): Promise<ConversionResult>;
+}

package/vendor/markit-ai/dist/converters/json.js

@@ -0,0 +1,21 @@
+const EXTENSIONS = [".json"];
+const MIMETYPES = ["application/json"];
+export class JsonConverter {
+    name = "json";
+    accepts(streamInfo) {
+        if (streamInfo.extension && EXTENSIONS.includes(streamInfo.extension)) {
+            return true;
+        }
+        if (streamInfo.mimetype &&
+            MIMETYPES.some((m) => streamInfo.mimetype?.startsWith(m))) {
+            return true;
+        }
+        return false;
+    }
+    async convert(input, _streamInfo) {
+        const text = new TextDecoder("utf-8").decode(input);
+        const parsed = JSON.parse(text);
+        const pretty = JSON.stringify(parsed, null, 2);
+        return { markdown: `\`\`\`json\n${pretty}\n\`\`\`` };
+    }
+}

package/vendor/markit-ai/dist/converters/pdf/columns.d.ts

@@ -0,0 +1,35 @@
+/**
+ * Multi-column layout detection and text box reordering.
+ *
+ * Many PDFs (legal documents, datasheets, academic papers) use two-column
+ * layouts. Without column detection, text boxes are ordered by Y position
+ * only, interleaving left and right column content.
+ *
+ * Algorithm:
+ *   1. Collect left edges of all text boxes on the page
+ *   2. Find the largest horizontal gap between consecutive left edges
+ *   3. If gap > MIN_GAP_RATIO of the text width and both sides have
+ *      enough boxes → multi-column detected
+ *   4. Assign each text box to a column based on its center X
+ *   5. Return columns in reading order (left-to-right, top-to-bottom)
+ *
+ * This only detects the column structure. The caller is responsible for
+ * processing each column's text boxes independently (table detection,
+ * rendering, etc.).
+ */
+import type { TextBox } from "./types.js";
+export interface ColumnLayout {
+    /** Number of columns detected (1 = single column, 2+ = multi-column). */
+    columnCount: number;
+    /** Text boxes grouped by column, in reading order (left to right). */
+    columns: TextBox[][];
+    /** X positions of column boundaries (between columns). */
+    boundaries: number[];
+}
+/**
+ * Detect column layout and return text boxes grouped by column.
+ *
+ * For single-column pages, returns all boxes in one group.
+ * For multi-column pages, returns boxes split by column in reading order.
+ */
+export declare function detectColumns(textBoxes: TextBox[]): ColumnLayout;

package/vendor/markit-ai/dist/converters/pdf/columns.js

@@ -0,0 +1,93 @@
+/**
+ * Multi-column layout detection and text box reordering.
+ *
+ * Many PDFs (legal documents, datasheets, academic papers) use two-column
+ * layouts. Without column detection, text boxes are ordered by Y position
+ * only, interleaving left and right column content.
+ *
+ * Algorithm:
+ *   1. Collect left edges of all text boxes on the page
+ *   2. Find the largest horizontal gap between consecutive left edges
+ *   3. If gap > MIN_GAP_RATIO of the text width and both sides have
+ *      enough boxes → multi-column detected
+ *   4. Assign each text box to a column based on its center X
+ *   5. Return columns in reading order (left-to-right, top-to-bottom)
+ *
+ * This only detects the column structure. The caller is responsible for
+ * processing each column's text boxes independently (table detection,
+ * rendering, etc.).
+ */
+/**
+ * Minimum gap as a fraction of the total text width to consider a column
+ * boundary. A two-column layout typically has ~50% gap; we use a lower
+ * threshold to catch asymmetric columns.
+ */
+const MIN_GAP_RATIO = 0.15;
+/** Minimum number of text boxes on each side of the gap. */
+const MIN_BOXES_PER_COLUMN = 4;
+/** Minimum gap in absolute points to avoid splitting on small whitespace. */
+const MIN_GAP_PTS = 40;
+/**
+ * Detect column layout and return text boxes grouped by column.
+ *
+ * For single-column pages, returns all boxes in one group.
+ * For multi-column pages, returns boxes split by column in reading order.
+ */
+export function detectColumns(textBoxes) {
+    if (textBoxes.length < MIN_BOXES_PER_COLUMN * 2) {
+        return { columnCount: 1, columns: [textBoxes], boundaries: [] };
+    }
+    // Collect unique left edges (rounded to avoid float noise)
+    const lefts = [
+        ...new Set(textBoxes.map((tb) => Math.round(tb.bounds.left))),
+    ].sort((a, b) => a - b);
+    if (lefts.length < 2) {
+        return { columnCount: 1, columns: [textBoxes], boundaries: [] };
+    }
+    const textXMin = lefts[0];
+    const textXMax = Math.max(...textBoxes.map((tb) => Math.round(tb.bounds.right)));
+    const textWidth = textXMax - textXMin;
+    if (textWidth <= 0) {
+        return { columnCount: 1, columns: [textBoxes], boundaries: [] };
+    }
+    // Find the largest gap between consecutive left-edge positions
+    let maxGap = 0;
+    let gapLeft = 0;
+    let gapRight = 0;
+    for (let i = 1; i < lefts.length; i++) {
+        const gap = lefts[i] - lefts[i - 1];
+        if (gap > maxGap) {
+            maxGap = gap;
+            gapLeft = lefts[i - 1];
+            gapRight = lefts[i];
+        }
+    }
+    const gapRatio = maxGap / textWidth;
+    if (gapRatio < MIN_GAP_RATIO || maxGap < MIN_GAP_PTS) {
+        return { columnCount: 1, columns: [textBoxes], boundaries: [] };
+    }
+    // Split point is the midpoint of the gap
+    const splitX = (gapLeft + gapRight) / 2;
+    // Assign boxes to columns based on center X
+    const leftCol = [];
+    const rightCol = [];
+    for (const tb of textBoxes) {
+        const cx = (tb.bounds.left + tb.bounds.right) / 2;
+        if (cx < splitX) {
+            leftCol.push(tb);
+        }
+        else {
+            rightCol.push(tb);
+        }
+    }
+    // Validate both columns have enough content
+    if (leftCol.length < MIN_BOXES_PER_COLUMN ||
+        rightCol.length < MIN_BOXES_PER_COLUMN) {
+        return { columnCount: 1, columns: [textBoxes], boundaries: [] };
+    }
+    return {
+        columnCount: 2,
+        columns: [leftCol, rightCol],
+        boundaries: [splitX],
+    };
+}

package/vendor/markit-ai/dist/converters/pdf/extract.d.ts

@@ -0,0 +1,19 @@
+/**
+ * PDF content extraction using mupdf.
+ *
+ * Extracts text boxes (with position, font size, bold) and vector line
+ * segments (table borders) from each page. Uses mupdf's native WASM
+ * engine for fast parsing, and reads raw content streams for vector graphics.
+ *
+ * Coordinate system: PDF native (origin = bottom-left, Y increases upward).
+ */
+import type { ImageRegion, PageContent } from "./types.js";
+/**
+ * Render an image region from a PDF page as a PNG buffer.
+ * Uses mupdf's DrawDevice to render just the cropped area at 2x resolution.
+ */
+export declare function renderImageRegion(input: Uint8Array, region: ImageRegion): Uint8Array;
+/**
+ * Extract text boxes and vector segments from all pages of a PDF buffer.
+ */
+export declare function extractPages(input: Uint8Array): Promise<PageContent[]>;