kiri-mcp-server 0.10.0 → 0.12.0

package/dist/src/indexer/cli.js

@@ -1,21 +1,110 @@
 import { createHash } from "node:crypto";
 import { existsSync } from "node:fs";
-import { readFile, stat } from "node:fs/promises";
-import { join, resolve, extname } from "node:path";
+import { readFile, readdir, stat } from "node:fs/promises";
+import { join, resolve, extname, posix as pathPosix } from "node:path";
 import { pathToFileURL } from "node:url";
+import { parse as parseYAML } from "yaml";
 import { DuckDBClient } from "../shared/duckdb.js";
 import { generateEmbedding } from "../shared/embedding.js";
 import { acquireLock, releaseLock, LockfileError, getLockOwner } from "../shared/utils/lockfile.js";
-import { normalizeDbPath, ensureDbParentDir, getRepoPathCandidates } from "../shared/utils/path.js";
+import { normalizeDbPath, normalizeRepoPath, ensureDbParentDir, getRepoPathCandidates, } from "../shared/utils/path.js";
 import { analyzeSource, buildFallbackSnippet } from "./codeintel.js";
 import { getDefaultBranch, getHeadCommit, gitLsFiles, gitDiffNameOnly } from "./git.js";
 import { detectLanguage } from "./language.js";
+import { mergeRepoRecords } from "./migrations/repo-merger.js";
 import { getIndexerQueue } from "./queue.js";
-import { ensureBaseSchema, ensureRepoMetaColumns, rebuildFTSIfNeeded } from "./schema.js";
+import { ensureBaseSchema, ensureDocumentMetadataTables, ensureNormalizedRootColumn, ensureRepoMetaColumns, rebuildFTSIfNeeded, } from "./schema.js";
 import { IndexWatcher } from "./watch.js";
+function normalizePathForIndex(value) {
+    return value.replace(/\\/g, "/");
+}
+function ensurePairState(stateMap, path) {
+    const existing = stateMap.get(path);
+    if (existing) {
+        return existing;
+    }
+    const created = { count: 0, seen: new Set() };
+    stateMap.set(path, created);
+    return created;
+}
 const MAX_SAMPLE_BYTES = 32_768;
 const MAX_FILE_BYTES = 32 * 1024 * 1024; // 32MB limit to prevent memory exhaustion
 const SCAN_BATCH_SIZE = 100; // Process files in batches to limit memory usage
+const MARKDOWN_EXTENSIONS = new Set([".md", ".mdx", ".markdown"]);
+const DOCMETA_SNAPSHOT_DIR = "docmeta/";
+const DOCMETA_SNAPSHOT_TARGET_FIELD = "target_path";
+const DOCMETA_SNAPSHOT_DATA_FIELD = "front_matter";
+/**
+ * Metadata processing limits to prevent DoS attacks and memory exhaustion.
+ *
+ * These values balance security, performance, and real-world usage patterns.
+ * Adjust based on:
+ * - Performance testing with 10000+ file repositories
+ * - Memory profiling (Node.js heap size impact)
+ * - Analysis of 99th percentile values in production data
+ */
+/**
+ * Maximum length of a single metadata value (characters).
+ *
+ * Rationale: Typical YAML front matter fields (title, description) are 200-300 chars.
+ * Setting to 512 provides headroom while preventing abuse.
+ *
+ * Example use cases:
+ * - Document titles: ~100 chars
+ * - Descriptions: ~300 chars
+ * - Tags (as comma-separated string): ~200 chars
+ */
+const MAX_METADATA_VALUE_LENGTH = 512;
+/**
+ * Maximum nesting depth for metadata tree structures.
+ *
+ * Rationale: Normal YAML/JSON documents nest 3-5 levels deep.
+ * Setting to 8 accommodates complex configurations while preventing stack overflow.
+ *
+ * Defense: Prevents malicious deeply-nested documents from causing:
+ * - Stack overflow (recursive function calls)
+ * - Exponential memory growth
+ * - CPU exhaustion during traversal
+ */
+const MAX_METADATA_DEPTH = 8;
+/**
+ * Maximum number of elements in a metadata array.
+ *
+ * Rationale: Common use case is tags/categories arrays with ~10 items.
+ * Setting to 64 provides generous headroom for edge cases.
+ *
+ * Example arrays:
+ * - Tags: ["frontend", "react", "typescript"] (~3-10 items)
+ * - Authors: ["John Doe", "Jane Smith"] (~1-5 items)
+ * - Categories: ["guide", "tutorial", "api"] (~2-8 items)
+ */
+const MAX_METADATA_ARRAY_LENGTH = 64;
+/**
+ * Maximum number of key-value pairs extracted per file.
+ *
+ * Rationale: Memory footprint calculation:
+ * - 256 pairs × ~40 bytes/pair ≈ 10KB per file
+ * - For 10000 files: 10KB × 10000 = 100MB (acceptable overhead)
+ *
+ * Prevents DoS from files with thousands of metadata fields.
+ * Normal documents have 5-20 metadata fields.
+ */
+const MAX_METADATA_PAIRS_PER_FILE = 256;
+/**
+ * Maximum number of object keys processed in a metadata tree node.
+ *
+ * Rationale: Prevents memory exhaustion from maliciously crafted objects with excessive keys.
+ * Normal metadata objects have 5-20 keys. Setting to 256 provides generous headroom.
+ *
+ * Memory impact: Each key entry requires ~50 bytes (key name + value reference).
+ * 256 keys × 50 bytes ≈ 12.8KB per object, which is acceptable.
+ */
+const MAX_METADATA_OBJECT_KEYS = 256;
+/**
+ * Key name used for root-level scalar values in metadata trees.
+ * Internal use only - not exposed in search results.
+ */
+const ROOT_METADATA_KEY = "__root";
 /**
  * Maximum number of SQL placeholders per INSERT statement.
  *
@@ -72,43 +161,17 @@ function isBinaryBuffer(buffer) {
  * @param defaultBranch - Default branch name (e.g., "main", "master"), or null if unknown
  * @returns The repository ID (auto-generated on first insert, reused thereafter)
  */
-async function mergeLegacyRepoRows(db, canonicalRepoId, legacyRepoIds) {
-    if (legacyRepoIds.length === 0) {
-        return;
-    }
-    const referencingTables = await db.all(`SELECT DISTINCT c.table_name
-       FROM duckdb_columns() AS c
-       JOIN duckdb_tables() AS t
-         ON c.database_name = t.database_name
-        AND c.schema_name = t.schema_name
-        AND c.table_name = t.table_name
-      WHERE c.column_name = 'repo_id'
-        AND c.table_name <> 'repo'
-        AND t.table_type = 'BASE TABLE'`);
-    const safeTables = referencingTables
-        .map((row) => row.table_name)
-        .filter((name) => /^[A-Za-z0-9_]+$/.test(name));
-    await db.transaction(async () => {
-        for (const legacyRepoId of legacyRepoIds) {
-            for (const tableName of safeTables) {
-                await db.run(`UPDATE ${tableName} SET repo_id = ? WHERE repo_id = ?`, [
-                    canonicalRepoId,
-                    legacyRepoId,
-                ]);
-            }
-            await db.run("DELETE FROM repo WHERE id = ?", [legacyRepoId]);
-        }
-    });
-}
 async function ensureRepo(db, repoRoot, defaultBranch, candidateRoots) {
     const searchRoots = Array.from(new Set([repoRoot, ...(candidateRoots ?? [])]));
     const placeholders = searchRoots.map(() => "?").join(", ");
     let rows = await db.all(`SELECT id, root FROM repo WHERE root IN (${placeholders})`, searchRoots);
     if (rows.length === 0) {
-        await db.run(`INSERT INTO repo (root, default_branch, indexed_at)
-       VALUES (?, ?, CURRENT_TIMESTAMP)
+        const normalized = normalizeRepoPath(repoRoot);
+        await db.run(`INSERT INTO repo (root, normalized_root, default_branch, indexed_at)
+       VALUES (?, ?, ?, CURRENT_TIMESTAMP)
        ON CONFLICT(root) DO UPDATE SET
-         default_branch = COALESCE(excluded.default_branch, repo.default_branch)`, [repoRoot, defaultBranch]);
+         normalized_root = excluded.normalized_root,
+         default_branch = COALESCE(excluded.default_branch, repo.default_branch)`, [repoRoot, normalized, defaultBranch]);
         rows = await db.all(`SELECT id, root FROM repo WHERE root IN (${placeholders})`, searchRoots);
     }
     if (rows.length === 0) {
@@ -123,7 +186,7 @@ async function ensureRepo(db, repoRoot, defaultBranch, candidateRoots) {
         canonicalRow = { ...canonicalRow, root: repoRoot };
     }
     const legacyIds = rows.filter((row) => row.id !== canonicalRow.id).map((row) => row.id);
-    await mergeLegacyRepoRows(db, canonicalRow.id, legacyIds);
+    await mergeRepoRecords(db, canonicalRow.id, legacyIds);
     return canonicalRow.id;
 }
 /**
@@ -302,6 +365,491 @@ async function persistEmbeddings(db, repoId, records) {
         ]),
     }));
 }
+async function persistDocumentMetadata(db, repoId, records) {
+    if (records.length === 0)
+        return;
+    const BATCH_SIZE = calculateBatchSize(4);
+    await persistInBatches(db, records, BATCH_SIZE, (batch) => ({
+        sql: `INSERT OR REPLACE INTO document_metadata (repo_id, path, source, data) VALUES ${batch.map(() => "(?, ?, ?, ?)").join(", ")}`,
+        params: batch.flatMap((record) => [
+            repoId,
+            record.path,
+            record.source,
+            JSON.stringify(record.data),
+        ]),
+    }));
+}
+async function persistMetadataPairs(db, repoId, records) {
+    if (records.length === 0)
+        return;
+    const BATCH_SIZE = calculateBatchSize(5);
+    await persistInBatches(db, records, BATCH_SIZE, (batch) => ({
+        sql: `INSERT OR REPLACE INTO document_metadata_kv (repo_id, path, source, key, value) VALUES ${batch.map(() => "(?, ?, ?, ?, ?)").join(", ")}`,
+        params: batch.flatMap((record) => [
+            repoId,
+            record.path,
+            record.source,
+            record.key,
+            record.value,
+        ]),
+    }));
+}
+async function persistMarkdownLinks(db, repoId, records) {
+    if (records.length === 0)
+        return;
+    const BATCH_SIZE = calculateBatchSize(6);
+    await persistInBatches(db, records, BATCH_SIZE, (batch) => ({
+        sql: `INSERT OR REPLACE INTO markdown_link (repo_id, src_path, target, resolved_path, anchor_text, kind) VALUES ${batch.map(() => "(?, ?, ?, ?, ?, ?)").join(", ")}`,
+        params: batch.flatMap((record) => [
+            repoId,
+            record.srcPath,
+            record.target,
+            record.resolvedPath,
+            record.anchorText,
+            record.kind,
+        ]),
+    }));
+}
+function sanitizeMetadataTree(value, depth = 0) {
+    // Depth check at the beginning to prevent stack overflow
+    if (depth > MAX_METADATA_DEPTH) {
+        console.warn(`Metadata depth limit (${MAX_METADATA_DEPTH}) exceeded, truncating nested value`);
+        return null;
+    }
+    if (value === null || value === undefined) {
+        return null;
+    }
+    if (value instanceof Date) {
+        return value.toISOString();
+    }
+    if (typeof value === "string") {
+        const trimmed = value.trim();
+        if (trimmed.length === 0) {
+            return null;
+        }
+        return trimmed.length > MAX_METADATA_VALUE_LENGTH
+            ? trimmed.slice(0, MAX_METADATA_VALUE_LENGTH)
+            : trimmed;
+    }
+    if (typeof value === "number") {
+        if (!Number.isFinite(value)) {
+            return null;
+        }
+        return value;
+    }
+    if (typeof value === "boolean") {
+        return value;
+    }
+    if (Array.isArray(value)) {
+        if (value.length === 0) {
+            return null;
+        }
+        // Warn if array is too large
+        if (value.length > MAX_METADATA_ARRAY_LENGTH) {
+            console.warn(`Metadata array has ${value.length} elements, limiting to ${MAX_METADATA_ARRAY_LENGTH}`);
+        }
+        const sanitized = [];
+        for (const item of value.slice(0, MAX_METADATA_ARRAY_LENGTH)) {
+            const child = sanitizeMetadataTree(item, depth + 1);
+            if (child !== null) {
+                sanitized.push(child);
+            }
+        }
+        return sanitized.length > 0 ? sanitized : null;
+    }
+    if (typeof value === "object") {
+        const result = {};
+        const entries = Object.entries(value);
+        // Limit number of object keys to prevent memory exhaustion
+        if (entries.length > MAX_METADATA_OBJECT_KEYS) {
+            console.warn(`Object has ${entries.length} keys, limiting to ${MAX_METADATA_OBJECT_KEYS} to prevent memory exhaustion`);
+        }
+        for (const [key, child] of entries.slice(0, MAX_METADATA_OBJECT_KEYS)) {
+            if (!key)
+                continue;
+            const sanitizedChild = sanitizeMetadataTree(child, depth + 1);
+            if (sanitizedChild !== null) {
+                result[key] = sanitizedChild;
+            }
+        }
+        return Object.keys(result).length > 0 ? result : null;
+    }
+    return null;
+}
+function metadataValueToString(value) {
+    if (typeof value === "string") {
+        return value;
+    }
+    if (typeof value === "number") {
+        return Number.isFinite(value) ? value.toString() : "";
+    }
+    return value ? "true" : "false";
+}
+function collectMetadataPairsFromValue(value, path, source, pairs, state, keyPrefix = "") {
+    if (state.count >= MAX_METADATA_PAIRS_PER_FILE) {
+        return;
+    }
+    if (typeof value === "string" || typeof value === "number" || typeof value === "boolean") {
+        const key = keyPrefix.length > 0 ? keyPrefix : ROOT_METADATA_KEY;
+        let normalized = metadataValueToString(value).trim();
+        if (normalized.length === 0) {
+            return;
+        }
+        if (normalized.length > MAX_METADATA_VALUE_LENGTH) {
+            normalized = normalized.slice(0, MAX_METADATA_VALUE_LENGTH);
+        }
+        const dedupeKey = `${source}:${key}:${normalized.toLowerCase()}`;
+        if (state.seen.has(dedupeKey)) {
+            return;
+        }
+        state.seen.add(dedupeKey);
+        pairs.push({ path, source, key, value: normalized });
+        state.count += 1;
+        return;
+    }
+    if (Array.isArray(value)) {
+        for (const item of value) {
+            collectMetadataPairsFromValue(item, path, source, pairs, state, keyPrefix);
+            if (state.count >= MAX_METADATA_PAIRS_PER_FILE) {
+                break;
+            }
+        }
+        return;
+    }
+    if (typeof value === "object" && value !== null) {
+        for (const [childKey, childValue] of Object.entries(value)) {
+            const normalizedKey = childKey.toLowerCase();
+            const nextPrefix = keyPrefix.length > 0 ? `${keyPrefix}.${normalizedKey}` : normalizedKey;
+            collectMetadataPairsFromValue(childValue, path, source, pairs, state, nextPrefix);
+            if (state.count >= MAX_METADATA_PAIRS_PER_FILE) {
+                break;
+            }
+        }
+    }
+}
+function parseFrontMatterBlock(content, path) {
+    const leading = content.startsWith("\uFEFF") ? content.slice(1) : content;
+    if (!leading.startsWith("---")) {
+        return null;
+    }
+    const match = leading.match(/^---\s*\r?\n([\s\S]*?)\r?\n---\s*(?:\r?\n|$)/);
+    if (!match) {
+        return null;
+    }
+    const rawBlock = match[1] ?? "";
+    const body = leading.slice(match[0].length);
+    try {
+        const data = parseYAML(rawBlock);
+        return { data: data ?? null, body };
+    }
+    catch (error) {
+        // Structured error logging for better debugging
+        const errorMessage = error instanceof Error ? error.message : String(error);
+        console.warn(JSON.stringify({
+            level: "warn",
+            message: "Failed to parse Markdown front matter",
+            file: path,
+            error: errorMessage,
+            context: "Front matter YAML parsing failed, metadata will be skipped for this file",
+        }));
+        return { data: null, body };
+    }
+}
+function stripLinkTitle(target) {
+    const trimmed = target.trim();
+    if (trimmed.length === 0) {
+        return trimmed;
+    }
+    const angleWrapped = trimmed.startsWith("<") && trimmed.endsWith(">");
+    const unwrapped = angleWrapped ? trimmed.slice(1, -1) : trimmed;
+    return unwrapped.replace(/\s+("[^"]*"|'[^']*')\s*$/, "").trim();
+}
+function extractMarkdownLinks(content, srcPath, repoFileSet) {
+    const links = [];
+    const pattern = /\[(?<text>[^\]]+)\]\((?<target>[^)]+)\)/g;
+    let match;
+    while ((match = pattern.exec(content)) !== null) {
+        if (match.index > 0 && content[match.index - 1] === "!") {
+            continue; // Skip images
+        }
+        const text = match.groups?.text?.trim() ?? "";
+        let target = match.groups?.target?.trim() ?? "";
+        if (!text || !target) {
+            continue;
+        }
+        target = stripLinkTitle(target);
+        if (!target) {
+            continue;
+        }
+        const kind = classifyMarkdownTarget(target);
+        const resolvedPath = resolveMarkdownLink(kind, target, srcPath, repoFileSet);
+        if (kind === "anchor" && resolvedPath === null) {
+            continue;
+        }
+        links.push({
+            srcPath,
+            target,
+            resolvedPath,
+            anchorText: text.slice(0, 160),
+            kind,
+        });
+    }
+    return links;
+}
+function classifyMarkdownTarget(target) {
+    const trimmed = target.trim();
+    if (!trimmed) {
+        return "external";
+    }
+    if (trimmed.startsWith("#")) {
+        return "anchor";
+    }
+    if (/^[a-z][a-z0-9+.-]*:/i.test(trimmed) || trimmed.startsWith("//")) {
+        return "external";
+    }
+    if (trimmed.startsWith("/")) {
+        return "absolute";
+    }
+    return "relative";
+}
+function resolveMarkdownLink(kind, target, srcPath, repoFileSet) {
+    if (kind === "external" || kind === "anchor") {
+        return null;
+    }
+    let cleanTarget = target.split("?")[0] ?? "";
+    const hashIndex = cleanTarget.indexOf("#");
+    if (hashIndex >= 0) {
+        cleanTarget = cleanTarget.slice(0, hashIndex);
+    }
+    cleanTarget = cleanTarget.trim().replace(/\\/g, "/");
+    if (!cleanTarget) {
+        return null;
+    }
+    let candidate;
+    if (kind === "absolute") {
+        candidate = cleanTarget.replace(/^\/+/, "");
+    }
+    else {
+        const dir = pathPosix.dirname(srcPath);
+        candidate = pathPosix.join(dir, cleanTarget);
+    }
+    candidate = pathPosix.normalize(candidate);
+    if (!candidate || candidate.startsWith("..")) {
+        return null;
+    }
+    // Security: Prevent directory traversal by checking for ".." segments
+    // Even after normalization, check that no path segment contains ".." or "."
+    const segments = candidate.split("/");
+    if (segments.some((seg) => seg === ".." || seg === ".")) {
+        return null;
+    }
+    // Additional security: reject absolute paths that may have bypassed earlier checks
+    if (candidate.startsWith("/")) {
+        return null;
+    }
+    const candidates = buildLinkCandidatePaths(candidate);
+    for (const pathCandidate of candidates) {
+        if (repoFileSet.has(pathCandidate)) {
+            return pathCandidate;
+        }
+    }
+    return null;
+}
+function buildLinkCandidatePaths(basePath) {
+    const candidates = new Set();
+    candidates.add(basePath);
+    if (!pathPosix.extname(basePath)) {
+        candidates.add(`${basePath}.md`);
+        candidates.add(`${basePath}.mdx`);
+        candidates.add(`${basePath}/README.md`);
+        candidates.add(`${basePath}/readme.md`);
+        candidates.add(`${basePath}/index.md`);
+        candidates.add(`${basePath}/INDEX.md`);
+    }
+    return Array.from(candidates);
+}
+function parseJsonValue(content, path) {
+    try {
+        return JSON.parse(content);
+    }
+    catch (error) {
+        // Structured error logging for better debugging
+        const errorMessage = error instanceof Error ? error.message : String(error);
+        console.warn(JSON.stringify({
+            level: "warn",
+            message: "Failed to parse JSON metadata",
+            file: path,
+            error: errorMessage,
+            context: "JSON parsing failed, metadata will be skipped for this file",
+        }));
+        return null;
+    }
+}
+function parseYamlValue(content, path) {
+    try {
+        return parseYAML(content);
+    }
+    catch (error) {
+        // Structured error logging for better debugging
+        const errorMessage = error instanceof Error ? error.message : String(error);
+        console.warn(JSON.stringify({
+            level: "warn",
+            message: "Failed to parse YAML metadata",
+            file: path,
+            error: errorMessage,
+            context: "YAML parsing failed, metadata will be skipped for this file",
+        }));
+        return null;
+    }
+}
+function parseDocmetaSnapshot(content, path) {
+    const parsed = parseJsonValue(content, path);
+    if (!parsed || typeof parsed !== "object" || Array.isArray(parsed)) {
+        return null;
+    }
+    const candidate = parsed;
+    const targetPath = candidate[DOCMETA_SNAPSHOT_TARGET_FIELD];
+    const frontMatter = candidate[DOCMETA_SNAPSHOT_DATA_FIELD];
+    if (typeof targetPath !== "string") {
+        return null;
+    }
+    const sanitized = sanitizeMetadataTree(frontMatter);
+    if (!sanitized) {
+        return null;
+    }
+    return {
+        targetPath: normalizePathForIndex(targetPath),
+        data: sanitized,
+    };
+}
+async function collectPlainDocsPaths(repoRoot) {
+    const results = [];
+    async function walkRelative(relativeDir) {
+        const absDir = join(repoRoot, relativeDir);
+        let entries;
+        try {
+            entries = await readdir(absDir, { withFileTypes: true });
+        }
+        catch {
+            return;
+        }
+        for (const entry of entries) {
+            const relPath = pathPosix.join(relativeDir, entry.name);
+            if (entry.isDirectory()) {
+                await walkRelative(relPath);
+            }
+            else {
+                results.push(relPath);
+            }
+        }
+    }
+    await walkRelative("docs").catch(() => { });
+    await walkRelative("docmeta").catch(() => { });
+    return results;
+}
+function extractStructuredData(files, blobs, repoFileSet) {
+    const map = new Map();
+    const pairStates = new Map();
+    for (const file of files) {
+        if (file.isBinary)
+            continue;
+        const blob = blobs.get(file.blobHash);
+        if (!blob || blob.content === null) {
+            continue;
+        }
+        const ext = (file.ext ?? "").toLowerCase();
+        const normalizedPath = normalizePathForIndex(file.path);
+        if (normalizedPath.startsWith(DOCMETA_SNAPSHOT_DIR)) {
+            const snapshot = parseDocmetaSnapshot(blob.content, file.path);
+            if (snapshot) {
+                const existing = map.get(snapshot.targetPath);
+                const structured = existing ?? {
+                    metadataRecords: [],
+                    metadataPairs: [],
+                    links: [],
+                };
+                structured.metadataRecords.push({
+                    path: snapshot.targetPath,
+                    source: "front_matter",
+                    data: snapshot.data,
+                });
+                const pairState = ensurePairState(pairStates, snapshot.targetPath);
+                collectMetadataPairsFromValue(snapshot.data, snapshot.targetPath, "front_matter", structured.metadataPairs, pairState);
+                map.set(snapshot.targetPath, structured);
+            }
+            continue;
+        }
+        const existingEntry = map.get(file.path);
+        const structured = existingEntry ?? {
+            metadataRecords: [],
+            metadataPairs: [],
+            links: [],
+        };
+        let mutated = false;
+        if (ext === ".json") {
+            const parsed = parseJsonValue(blob.content, file.path);
+            const sanitized = sanitizeMetadataTree(parsed);
+            if (sanitized) {
+                structured.metadataRecords.push({ path: file.path, source: "json", data: sanitized });
+                const pairState = ensurePairState(pairStates, file.path);
+                collectMetadataPairsFromValue(sanitized, file.path, "json", structured.metadataPairs, pairState);
+                mutated = true;
+            }
+        }
+        else if (ext === ".yaml" || ext === ".yml") {
+            const parsed = parseYamlValue(blob.content, file.path);
+            const sanitized = sanitizeMetadataTree(parsed);
+            if (sanitized) {
+                structured.metadataRecords.push({ path: file.path, source: "yaml", data: sanitized });
+                const pairState = ensurePairState(pairStates, file.path);
+                collectMetadataPairsFromValue(sanitized, file.path, "yaml", structured.metadataPairs, pairState);
+                mutated = true;
+            }
+        }
+        if (MARKDOWN_EXTENSIONS.has(ext)) {
+            const frontMatter = parseFrontMatterBlock(blob.content, file.path);
+            let markdownBody = blob.content;
+            if (frontMatter) {
+                if (frontMatter.data) {
+                    const sanitized = sanitizeMetadataTree(frontMatter.data);
+                    if (sanitized) {
+                        structured.metadataRecords.push({
+                            path: file.path,
+                            source: "front_matter",
+                            data: sanitized,
+                        });
+                        const pairState = ensurePairState(pairStates, file.path);
+                        collectMetadataPairsFromValue(sanitized, file.path, "front_matter", structured.metadataPairs, pairState);
+                        mutated = true;
+                    }
+                }
+                markdownBody = frontMatter.body;
+            }
+            const links = extractMarkdownLinks(markdownBody, file.path, repoFileSet);
+            if (links.length > 0) {
+                structured.links.push(...links);
+                mutated = true;
+            }
+        }
+        if (mutated || existingEntry) {
+            map.set(file.path, structured);
+        }
+    }
+    return map;
+}
+function aggregateStructuredData(map) {
+    const aggregated = {
+        metadataRecords: [],
+        metadataPairs: [],
+        links: [],
+    };
+    for (const entry of map.values()) {
+        aggregated.metadataRecords.push(...entry.metadataRecords);
+        aggregated.metadataPairs.push(...entry.metadataPairs);
+        aggregated.links.push(...entry.links);
+    }
+    return aggregated;
+}
 async function buildCodeIntel(files, blobs, workspaceRoot) {
     const fileSet = new Set(files.map((file) => file.path));
     const symbols = [];
@@ -498,16 +1046,20 @@ async function reconcileDeletedFiles(db, repoId, repoRoot) {
         }
     }
     // Delete all records for removed files in a single transaction
+    // Batched DELETE operations to avoid N+1 query problem
     if (deletedPaths.length > 0) {
         await db.transaction(async () => {
-            for (const path of deletedPaths) {
-                await db.run("DELETE FROM symbol WHERE repo_id = ? AND path = ?", [repoId, path]);
-                await db.run("DELETE FROM snippet WHERE repo_id = ? AND path = ?", [repoId, path]);
-                await db.run("DELETE FROM dependency WHERE repo_id = ? AND src_path = ?", [repoId, path]);
-                await db.run("DELETE FROM file_embedding WHERE repo_id = ? AND path = ?", [repoId, path]);
-                await db.run("DELETE FROM tree WHERE repo_id = ? AND path = ?", [repoId, path]);
-                await db.run("DELETE FROM file WHERE repo_id = ? AND path = ?", [repoId, path]);
-            }
+            const placeholders = deletedPaths.map(() => "?").join(", ");
+            const params = [repoId, ...deletedPaths];
+            await db.run(`DELETE FROM symbol WHERE repo_id = ? AND path IN (${placeholders})`, params);
+            await db.run(`DELETE FROM snippet WHERE repo_id = ? AND path IN (${placeholders})`, params);
+            await db.run(`DELETE FROM dependency WHERE repo_id = ? AND src_path IN (${placeholders})`, params);
+            await db.run(`DELETE FROM file_embedding WHERE repo_id = ? AND path IN (${placeholders})`, params);
+            await db.run(`DELETE FROM document_metadata WHERE repo_id = ? AND path IN (${placeholders})`, params);
+            await db.run(`DELETE FROM document_metadata_kv WHERE repo_id = ? AND path IN (${placeholders})`, params);
+            await db.run(`DELETE FROM markdown_link WHERE repo_id = ? AND src_path IN (${placeholders})`, params);
+            await db.run(`DELETE FROM tree WHERE repo_id = ? AND path IN (${placeholders})`, params);
+            await db.run(`DELETE FROM file WHERE repo_id = ? AND path IN (${placeholders})`, params);
         });
     }
     return deletedPaths;
@@ -526,6 +1078,9 @@ async function deleteFileRecords(db, repoId, headCommit, path) {
     await db.run("DELETE FROM snippet WHERE repo_id = ? AND path = ?", [repoId, path]);
     await db.run("DELETE FROM dependency WHERE repo_id = ? AND src_path = ?", [repoId, path]);
     await db.run("DELETE FROM file_embedding WHERE repo_id = ? AND path = ?", [repoId, path]);
+    await db.run("DELETE FROM document_metadata WHERE repo_id = ? AND path = ?", [repoId, path]);
+    await db.run("DELETE FROM document_metadata_kv WHERE repo_id = ? AND path = ?", [repoId, path]);
+    await db.run("DELETE FROM markdown_link WHERE repo_id = ? AND src_path = ?", [repoId, path]);
     await db.run("DELETE FROM tree WHERE repo_id = ? AND commit_hash = ? AND path = ?", [
         repoId,
         headCommit,
@@ -533,6 +1088,25 @@ async function deleteFileRecords(db, repoId, headCommit, path) {
     ]);
     await db.run("DELETE FROM file WHERE repo_id = ? AND path = ?", [repoId, path]);
 }
+/**
+ * Remove blob records that are no longer referenced by any file.
+ * This garbage collection should be run after full re-indexing or periodically as maintenance.
+ *
+ * @param db - Database client
+ */
+async function garbageCollectBlobs(db) {
+    console.info("Running garbage collection on blob table...");
+    try {
+        await db.run(`
+      DELETE FROM blob
+      WHERE hash NOT IN (SELECT DISTINCT blob_hash FROM file)
+    `);
+        console.info("Blob garbage collection complete.");
+    }
+    catch (error) {
+        console.warn("Failed to garbage collect blobs:", error instanceof Error ? error.message : String(error));
+    }
+}
 export async function runIndexer(options) {
     const repoPathCandidates = getRepoPathCandidates(options.repoRoot);
     const repoRoot = repoPathCandidates[0];
@@ -571,6 +1145,10 @@ export async function runIndexer(options) {
             const dbClient = await DuckDBClient.connect({ databasePath, ensureDirectory: true });
             db = dbClient;
             await ensureBaseSchema(dbClient);
+            // Migration: Ensure document_metadata tables exist for existing DBs
+            await ensureDocumentMetadataTables(dbClient);
+            // Phase 1: Ensure normalized_root column exists (Critical #1)
+            await ensureNormalizedRootColumn(dbClient);
             // Phase 3: Ensure FTS metadata columns exist for existing DBs (migration)
             await ensureRepoMetaColumns(dbClient);
             const [headCommit, defaultBranch] = await Promise.all([
@@ -626,6 +1204,12 @@ export async function runIndexer(options) {
                     }
                     return;
                 }
+                const existingFileRows = await dbClient.all("SELECT path FROM file WHERE repo_id = ?", [repoId]);
+                const repoFileSet = new Set(existingFileRows.map((row) => row.path));
+                for (const file of files) {
+                    repoFileSet.add(file.path);
+                }
+                const structuredByFile = extractStructuredData(changedFiles, changedBlobs, repoFileSet);
                 // Process all changed files in a single transaction for atomicity
                 const fileSet = new Set(files.map((f) => f.path));
                 const embeddingMap = new Map();
@@ -648,67 +1232,79 @@ export async function runIndexer(options) {
                         const blob = changedBlobs.get(file.blobHash);
                         if (!blob)
                             continue;
-                        // Build code intelligence for this file
-                        const fileSymbols = [];
-                        const fileSnippets = [];
-                        const fileDependencies = [];
-                        if (!file.isBinary && blob.content) {
-                            const analysis = await analyzeSource(file.path, file.lang, blob.content, fileSet, repoRoot);
-                            for (const symbol of analysis.symbols) {
-                                fileSymbols.push({
-                                    path: file.path,
-                                    symbolId: symbol.symbolId,
-                                    name: symbol.name,
-                                    kind: symbol.kind,
-                                    rangeStartLine: symbol.rangeStartLine,
-                                    rangeEndLine: symbol.rangeEndLine,
-                                    signature: symbol.signature,
-                                    doc: symbol.doc,
-                                });
+                        try {
+                            // Build code intelligence for this file
+                            const fileSymbols = [];
+                            const fileSnippets = [];
+                            const fileDependencies = [];
+                            if (!file.isBinary && blob.content) {
+                                const analysis = await analyzeSource(file.path, file.lang, blob.content, fileSet, repoRoot);
+                                for (const symbol of analysis.symbols) {
+                                    fileSymbols.push({
+                                        path: file.path,
+                                        symbolId: symbol.symbolId,
+                                        name: symbol.name,
+                                        kind: symbol.kind,
+                                        rangeStartLine: symbol.rangeStartLine,
+                                        rangeEndLine: symbol.rangeEndLine,
+                                        signature: symbol.signature,
+                                        doc: symbol.doc,
+                                    });
+                                }
+                                for (const snippet of analysis.snippets) {
+                                    fileSnippets.push({
+                                        path: file.path,
+                                        snippetId: snippet.startLine,
+                                        startLine: snippet.startLine,
+                                        endLine: snippet.endLine,
+                                        symbolId: snippet.symbolId,
+                                    });
+                                }
+                                for (const dep of analysis.dependencies) {
+                                    fileDependencies.push({
+                                        srcPath: file.path,
+                                        dstKind: dep.dstKind,
+                                        dst: dep.dst,
+                                        rel: dep.rel,
+                                    });
+                                }
                             }
-                            for (const snippet of analysis.snippets) {
+                            else {
+                                // Binary or no content: add fallback snippet
+                                const fallback = buildFallbackSnippet(blob.lineCount ?? 1);
                                 fileSnippets.push({
                                     path: file.path,
-                                    snippetId: snippet.startLine,
-                                    startLine: snippet.startLine,
-                                    endLine: snippet.endLine,
-                                    symbolId: snippet.symbolId,
+                                    snippetId: fallback.startLine,
+                                    startLine: fallback.startLine,
+                                    endLine: fallback.endLine,
+                                    symbolId: fallback.symbolId,
                                 });
                             }
-                            for (const dep of analysis.dependencies) {
-                                fileDependencies.push({
-                                    srcPath: file.path,
-                                    dstKind: dep.dstKind,
-                                    dst: dep.dst,
-                                    rel: dep.rel,
-                                });
+                            const fileEmbedding = embeddingMap.get(file.path) ?? null;
+                            // Delete old records for this file (within main transaction)
+                            await deleteFileRecords(dbClient, repoId, headCommit, file.path);
+                            // Insert new records (within main transaction)
+                            await persistBlobs(dbClient, new Map([[blob.hash, blob]]));
+                            await persistTrees(dbClient, repoId, headCommit, [file]);
+                            await persistFiles(dbClient, repoId, [file]);
+                            await persistSymbols(dbClient, repoId, fileSymbols);
+                            await persistSnippets(dbClient, repoId, fileSnippets);
+                            await persistDependencies(dbClient, repoId, fileDependencies);
+                            const structured = structuredByFile.get(file.path);
+                            if (structured) {
+                                await persistDocumentMetadata(dbClient, repoId, structured.metadataRecords);
+                                await persistMetadataPairs(dbClient, repoId, structured.metadataPairs);
+                                await persistMarkdownLinks(dbClient, repoId, structured.links);
                             }
+                            if (fileEmbedding) {
+                                await persistEmbeddings(dbClient, repoId, [fileEmbedding]);
+                            }
+                            processedCount++;
                         }
-                        else {
-                            // Binary or no content: add fallback snippet
-                            const fallback = buildFallbackSnippet(blob.lineCount ?? 1);
-                            fileSnippets.push({
-                                path: file.path,
-                                snippetId: fallback.startLine,
-                                startLine: fallback.startLine,
-                                endLine: fallback.endLine,
-                                symbolId: fallback.symbolId,
-                            });
-                        }
-                        const fileEmbedding = embeddingMap.get(file.path) ?? null;
-                        // Delete old records for this file (within main transaction)
-                        await deleteFileRecords(dbClient, repoId, headCommit, file.path);
-                        // Insert new records (within main transaction)
-                        await persistBlobs(dbClient, new Map([[blob.hash, blob]]));
-                        await persistTrees(dbClient, repoId, headCommit, [file]);
-                        await persistFiles(dbClient, repoId, [file]);
-                        await persistSymbols(dbClient, repoId, fileSymbols);
-                        await persistSnippets(dbClient, repoId, fileSnippets);
-                        await persistDependencies(dbClient, repoId, fileDependencies);
-                        if (fileEmbedding) {
-                            await persistEmbeddings(dbClient, repoId, [fileEmbedding]);
+                        catch (error) {
+                            console.error(`Failed to process file ${file.path}, transaction will rollback:`, error instanceof Error ? error.message : String(error));
+                            throw error; // Re-throw to rollback the transaction
                         }
-                        processedCount++;
                     }
                     // Update timestamp and mark FTS dirty inside transaction for atomicity
                     // Fix: Increment fts_generation to prevent lost updates during concurrent rebuilds
@@ -725,7 +1321,14 @@ export async function runIndexer(options) {
                 return;
             }
             // Full mode: reindex entire repository
-            const paths = await gitLsFiles(repoRoot);
+            let paths = await gitLsFiles(repoRoot);
+            if (paths.length === 0) {
+                const fallbackPaths = await collectPlainDocsPaths(repoRoot);
+                if (fallbackPaths.length > 0) {
+                    console.warn(`git ls-files returned 0 paths for ${repoRoot}. Falling back to filesystem scan (${fallbackPaths.length} files).`);
+                    paths = fallbackPaths;
+                }
+            }
             const { blobs, files, embeddings, missingPaths } = await scanFilesInBatches(repoRoot, paths);
             // In full mode, missingPaths should be rare (git ls-files returns existing files)
             // But log them if they occur (race condition: file deleted between ls-files and scan)
@@ -733,6 +1336,9 @@ export async function runIndexer(options) {
                 console.warn(`${missingPaths.length} file(s) disappeared during full reindex (race condition)`);
             }
             const codeIntel = await buildCodeIntel(files, blobs, repoRoot);
+            const repoFileSetFull = new Set(files.map((file) => file.path));
+            const structuredMap = extractStructuredData(files, blobs, repoFileSetFull);
+            const aggregatedStructured = aggregateStructuredData(structuredMap);
             await dbClient.transaction(async () => {
                 await dbClient.run("DELETE FROM tree WHERE repo_id = ?", [repoId]);
                 await dbClient.run("DELETE FROM file WHERE repo_id = ?", [repoId]);
@@ -740,6 +1346,9 @@ export async function runIndexer(options) {
                 await dbClient.run("DELETE FROM snippet WHERE repo_id = ?", [repoId]);
                 await dbClient.run("DELETE FROM dependency WHERE repo_id = ?", [repoId]);
                 await dbClient.run("DELETE FROM file_embedding WHERE repo_id = ?", [repoId]);
+                await dbClient.run("DELETE FROM document_metadata WHERE repo_id = ?", [repoId]);
+                await dbClient.run("DELETE FROM document_metadata_kv WHERE repo_id = ?", [repoId]);
+                await dbClient.run("DELETE FROM markdown_link WHERE repo_id = ?", [repoId]);
                 await persistBlobs(dbClient, blobs);
                 await persistTrees(dbClient, repoId, headCommit, files);
                 await persistFiles(dbClient, repoId, files);
@@ -747,6 +1356,9 @@ export async function runIndexer(options) {
                 await persistSnippets(dbClient, repoId, codeIntel.snippets);
                 await persistDependencies(dbClient, repoId, codeIntel.dependencies);
                 await persistEmbeddings(dbClient, repoId, embeddings);
+                await persistDocumentMetadata(dbClient, repoId, aggregatedStructured.metadataRecords);
+                await persistMetadataPairs(dbClient, repoId, aggregatedStructured.metadataPairs);
+                await persistMarkdownLinks(dbClient, repoId, aggregatedStructured.links);
                 // Update timestamp and mark FTS dirty inside transaction to ensure atomicity
                 // Fix: Increment fts_generation to prevent lost updates during concurrent rebuilds
                 if (defaultBranch) {
@@ -759,6 +1371,8 @@ export async function runIndexer(options) {
             console.info(`Indexed ${files.length} files for repo ${repoRoot} at ${databasePath} (commit ${headCommit.slice(0, 12)})`);
             // Phase 2+3: Force rebuild FTS index after full reindex
             await rebuildFTSIfNeeded(dbClient, repoId, true);
+            // Garbage collect orphaned blobs after full reindex
+            await garbageCollectBlobs(dbClient);
         }
         finally {
             // Fix #2: Ensure lock is released even if DB connection fails