codesift-mcp 0.1.0 → 0.2.0

package/dist/tools/search-tools.js

@@ -1,18 +1,36 @@
-import { readFile, readdir, stat } from "node:fs/promises";
-import { join, relative, extname } from "node:path";
+import { readFile } from "node:fs/promises";
+import { execFileSync } from "node:child_process";
+import { join } from "node:path";
 import { getBM25Index, getCodeIndex } from "./index-tools.js";
-import { searchBM25 } from "../search/bm25.js";
+import { searchBM25, applyCutoff } from "../search/bm25.js";
 import { loadConfig } from "../config.js";
-const DEFAULT_MAX_TEXT_MATCHES = 500;
-const MAX_FILE_SIZE = 1_000_000; // 1MB — skip giant files
+import { walkDirectory } from "../utils/walk.js";
+import { matchFilePattern } from "../utils/glob.js";
+const DEFAULT_MAX_TEXT_MATCHES = 200;
 const MAX_WALK_FILES = 50_000; // Safety limit — stop walking after this many files
-/** Directories to skip during text search file walk */
-const IGNORE_DIRS = new Set([
-    "node_modules", ".git", "dist", "build", "coverage",
-    ".codesift", ".next", "__pycache__", ".pytest_cache",
-    ".venv", "venv", ".tox", ".mypy_cache", ".turbo",
-    "generated", "audit-results", ".backup", "jscpd-report",
-]);
+const SEARCH_TIMEOUT_MS = 30_000; // Abort search after 30s to prevent 100s+ hangs
+const AUTO_GROUP_THRESHOLD = 50; // Auto-switch to group_by_file above this match count
+const MAX_RESPONSE_CHARS = 80_000; // ~20K tokens — force group_by_file above this
+const MAX_FIRST_MATCH_CHARS = 300; // Cap first_match preview in grouped output
+const MAX_LINE_CHARS = 500; // Truncate individual match lines (minified JS/JSON can be 100K+)
+const DEFAULT_TOP_K_WITH_SOURCE = 10; // Cap results when include_source=true without file_pattern
+const BM25_FILTER_MULTIPLIER = 5; // Widen BM25 candidate set when filters active
+const BM25_FILTER_MIN_K = 200; // Minimum candidate set size when filters active
+const DEFAULT_SOURCE_CHARS_NARROW = 200; // Source truncation without file_pattern (reduce waste)
+const DEFAULT_SOURCE_CHARS_WIDE = 500; // Source truncation with file_pattern
+const CHARS_PER_TOKEN = 3.5; // Approximate chars-per-token for budget calculation
+const DEFAULT_MAX_REGEX_RESULTS = 50; // Regex without file_pattern — tighter cap to limit timeout
+const JSON_OVERHEAD_PER_MATCH = 40; // Estimated JSON serialization overhead per TextMatch
+// SEC-003: Detect common catastrophic backtracking patterns (ReDoS)
+const REDOS_PATTERNS = [
+    /\(.*[+*].*\)[+*]/, // Nested quantifiers: (a+)+ or (a*)*
+    /\(.*\|.*\)[+*]/, // Alternation with quantifier: (a|b)+
+    /\(.*[+*].*\)\{/, // Nested quantifier with range: (a+){2,}
+    /\([^)]*\\[dDwWsS][+*].*\)[+*]/, // Character class with nested quantifier
+];
+function isSafeRegex(pattern) {
+    return !REDOS_PATTERNS.some((p) => p.test(pattern));
+}
 /** Binary/non-text extensions to skip during text search */
 const BINARY_EXTENSIONS = new Set([
     ".png", ".jpg", ".jpeg", ".gif", ".bmp", ".ico", ".svg",
@@ -25,128 +43,318 @@ const BINARY_EXTENSIONS = new Set([
     ".db", ".sqlite", ".sqlite3",
     ".lock",
 ]);
+// ── Private helpers ─────────────────────────────────────
+/** Check if a symbol matches the active kind and file_pattern filters. */
+function matchesSymbolFilters(symbol, options) {
+    if (options?.kind && symbol.kind !== options.kind)
+        return false;
+    if (options?.file_pattern && !matchFilePattern(symbol.file, options.file_pattern))
+        return false;
+    return true;
+}
 /**
- * Match a file path against a simple glob pattern.
- * Supports: "*.ts", "src/*.ts", "src/**\/*.ts", "**\/*.test.ts"
+ * Apply detail-level shaping, source truncation, and field cleanup.
+ * Compact: ~15 tok/result. Standard: signature + truncated source. Full: unlimited.
  */
-function matchFilePattern(filePath, pattern) {
-    // Exact match
-    if (filePath === pattern)
-        return true;
-    // "**\/" prefix — match anywhere in path
-    if (pattern.startsWith("**/")) {
-        const suffix = pattern.slice(3);
-        // Recursively match the suffix against every segment tail
-        return matchFilePattern(filePath, suffix) ||
-            filePath.includes("/" + suffix) ||
-            matchFileSuffix(filePath, suffix);
+function shapeSearchResults(results, detail, includeSource, options) {
+    if (detail === "compact") {
+        return results.map((r) => ({
+            symbol: {
+                id: r.symbol.id,
+                name: r.symbol.name,
+                kind: r.symbol.kind,
+                file: r.symbol.file,
+                start_line: r.symbol.start_line,
+            },
+            score: r.score,
+        }));
     }
-    // "*" at the start — match extension-style patterns like "*.ts"
-    if (pattern.startsWith("*") && !pattern.includes("/")) {
-        const suffix = pattern.slice(1);
-        return filePath.endsWith(suffix);
+    let shaped = results;
+    if (!includeSource) {
+        shaped = shaped.map((r) => {
+            const { source: _source, ...symbolWithoutSource } = r.symbol;
+            return { ...r, symbol: symbolWithoutSource };
+        });
     }
-    // "dir/**" — match everything under directory (e.g., "src/**")
-    if (pattern.endsWith("/**")) {
-        const prefix = pattern.slice(0, -3);
-        return filePath.startsWith(prefix + "/") || filePath === prefix;
+    const defaultSourceChars = detail === "full" ? undefined
+        : (includeSource && !options?.file_pattern) ? DEFAULT_SOURCE_CHARS_NARROW : DEFAULT_SOURCE_CHARS_WIDE;
+    const sourceChars = options?.source_chars ?? (includeSource ? defaultSourceChars : undefined);
+    if (includeSource && sourceChars !== undefined && sourceChars > 0) {
+        shaped = shaped.map((r) => {
+            const source = r.symbol.source;
+            if (source && source.length > sourceChars) {
+                return { ...r, symbol: { ...r.symbol, source: source.slice(0, sourceChars) + "..." } };
+            }
+            return r;
+        });
     }
-    // Pattern with "**" in the middle (e.g., "src/**/*.ts")
-    if (pattern.includes("/**/")) {
-        const [prefix, suffix] = splitFirst(pattern, "/**/");
-        if (!filePath.startsWith(prefix + "/") && filePath !== prefix)
-            return false;
-        const rest = filePath.slice(prefix.length + 1);
-        return matchFilePattern(rest, suffix) ||
-            matchFilePattern(rest, "**/" + suffix);
+    return shaped.map((r) => {
+        const { tokens: _tokens, repo: _repo, ...cleanSymbol } = r.symbol;
+        return { ...r, symbol: cleanSymbol };
+    });
+}
+/** Validate regex for ReDoS safety and compile without g/y flags, or throw descriptive error. */
+function compileSearchRegex(query) {
+    if (!isSafeRegex(query)) {
+        throw new Error("Regex pattern rejected: potential catastrophic backtracking (ReDoS)");
     }
-    // Simple directory prefix + filename pattern (e.g., "src/*.ts")
-    if (pattern.includes("/") && pattern.includes("*")) {
-        const lastSlash = pattern.lastIndexOf("/");
-        const dirPart = pattern.slice(0, lastSlash);
-        const filePart = pattern.slice(lastSlash + 1);
-        const fileLastSlash = filePath.lastIndexOf("/");
-        const fileDir = fileLastSlash >= 0 ? filePath.slice(0, fileLastSlash) : "";
-        const fileName = fileLastSlash >= 0 ? filePath.slice(fileLastSlash + 1) : filePath;
-        if (fileDir !== dirPart)
-            return false;
-        return matchFilePattern(fileName, filePart);
+    try {
+        // No g/y flags — regex is reused across files; stateful flags cause alternating matches
+        return new RegExp(query);
     }
-    // No wildcards: substring match on the full path
-    // "risk.service.ts" matches "src/lib/services/risk/risk.service.ts"
-    // "validators" matches "src/lib/validators/schema.ts"
-    if (!pattern.includes("*")) {
-        return filePath.includes(pattern);
+    catch (err) {
+        const message = err instanceof Error ? err.message : String(err);
+        throw new Error(`Invalid regex pattern: ${message}`);
     }
-    return false;
 }
-function matchFileSuffix(filePath, suffix) {
-    if (suffix.startsWith("*")) {
-        const ext = suffix.slice(1);
-        return filePath.endsWith(ext);
+// ── Ripgrep backend ────────────────────────────────────
+/** Directories always excluded from ripgrep search */
+const RG_EXCLUDE_DIRS = [
+    "node_modules", ".git", ".next", "dist", ".codesift", "coverage",
+    ".playwright-mcp", "__pycache__", ".mypy_cache", ".tox",
+];
+/** Detect whether `rg` (ripgrep) is available on this system. Cached at module level. */
+let rgAvailable = null;
+function hasRipgrep() {
+    if (rgAvailable !== null)
+        return rgAvailable;
+    try {
+        execFileSync("rg", ["--version"], { stdio: "pipe", timeout: 2000 });
+        rgAvailable = true;
     }
-    return filePath.endsWith("/" + suffix) || filePath === suffix;
-}
-function splitFirst(str, sep) {
-    const idx = str.indexOf(sep);
-    if (idx < 0)
-        return [str, ""];
-    return [str.slice(0, idx), str.slice(idx + sep.length)];
+    catch {
+        rgAvailable = false;
+    }
+    return rgAvailable;
 }
 /**
- * Walk a directory tree collecting all text files.
- * Returns relative paths from rootPath.
- * Unlike the index walk, this includes ALL text files (not just parseable ones).
+ * Search via ripgrep — fast C-based search, parses `rg -n` output.
+ * Falls back to Node.js search if rg is not available.
  */
-async function walkAllTextFiles(rootPath) {
-    const files = [];
-    let limitReached = false;
-    async function walk(dirPath) {
-        if (limitReached)
-            return;
-        let entries;
-        try {
-            entries = await readdir(dirPath, { withFileTypes: true });
+function searchWithRipgrep(root, query, options) {
+    const args = [
+        "-n", // line numbers
+        "--no-heading", // flat output
+        "--max-columns", String(MAX_LINE_CHARS),
+        "--max-columns-preview", // show truncated preview
+        "--max-count", String(Math.min(options.maxResults * 2, 5000)), // per-file cap (generous to hit global max)
+    ];
+    if (!options.regex) {
+        args.push("-F"); // fixed string (literal)
+    }
+    if (options.contextLines > 0) {
+        args.push("-C", String(options.contextLines));
+    }
+    // File pattern → rg glob
+    if (options.filePattern) {
+        // Handle patterns like "src/**" or "*.ts"
+        args.push("--glob", options.filePattern);
+    }
+    // Exclude dirs
+    for (const dir of RG_EXCLUDE_DIRS) {
+        args.push("--glob", `!${dir}`);
+    }
+    args.push("--", query, root);
+    let stdout;
+    try {
+        stdout = execFileSync("rg", args, {
+            encoding: "utf-8",
+            maxBuffer: 20 * 1024 * 1024, // 20MB
+            timeout: SEARCH_TIMEOUT_MS,
+        });
+    }
+    catch (err) {
+        // rg exits 1 = no matches, 2 = error
+        if (err && typeof err === "object" && "status" in err) {
+            const exitCode = err.status;
+            if (exitCode === 1)
+                return []; // no matches
+            if ("stdout" in err && typeof err.stdout === "string") {
+                stdout = err.stdout;
+                if (!stdout)
+                    return [];
+            }
+            else {
+                return [];
+            }
         }
-        catch {
-            return; // permission denied, etc.
+        else {
+            return [];
         }
-        for (const entry of entries) {
-            if (limitReached)
-                return;
-            const fullPath = join(dirPath, entry.name);
-            if (entry.isDirectory()) {
-                if (IGNORE_DIRS.has(entry.name) || entry.name.startsWith(".")) {
-                    continue;
-                }
-                await walk(fullPath);
+    }
+    const matches = [];
+    const rootPrefix = root.endsWith("/") ? root : root + "/";
+    // Parse context blocks: lines separated by "--" separators
+    const blocks = options.contextLines > 0
+        ? stdout.split(/^--$/m)
+        : [stdout];
+    for (const block of blocks) {
+        if (matches.length >= options.maxResults)
+            break;
+        const lines = block.split("\n").filter(Boolean);
+        // In context mode, find the actual match line (has `:` separator) vs context (has `-` separator)
+        // In non-context mode, all lines are matches
+        for (const rawLine of lines) {
+            if (matches.length >= options.maxResults)
+                break;
+            // rg format: /abs/path/file.ts:42:content  (match)
+            // rg format: /abs/path/file.ts-40-content   (context, only with -C)
+            // We only want match lines (with `:` after line number)
+            const matchResult = rawLine.match(/^(.+?):(\d+):(.*)/);
+            if (!matchResult)
+                continue;
+            const [, absPath, lineNumStr, content] = matchResult;
+            if (!absPath || !lineNumStr || content === undefined)
+                continue;
+            const relPath = absPath.startsWith(rootPrefix)
+                ? absPath.slice(rootPrefix.length)
+                : absPath;
+            matches.push({
+                file: relPath,
+                line: parseInt(lineNumStr, 10),
+                content: content,
+            });
+        }
+    }
+    // For context mode, we need to re-parse to attach context_before/context_after
+    // But context_lines=0 is the default now, so this path is rarely hit
+    if (options.contextLines > 0 && blocks.length > 1) {
+        return parseRipgrepContextBlocks(stdout, rootPrefix, options.maxResults, options.contextLines);
+    }
+    return matches;
+}
+/**
+ * Parse rg output with context lines (-C N) into TextMatch[] with context_before/context_after.
+ */
+function parseRipgrepContextBlocks(stdout, rootPrefix, maxResults, contextLines) {
+    const matches = [];
+    const blocks = stdout.split(/^--$/m);
+    for (const block of blocks) {
+        if (matches.length >= maxResults)
+            break;
+        const lines = block.split("\n").filter(Boolean);
+        // Separate match lines from context lines
+        // Match: path:line:content  Context: path-line-content
+        const parsed = [];
+        for (const raw of lines) {
+            // Try match line first (colon after line number)
+            const matchLine = raw.match(/^(.+?):(\d+):(.*)/);
+            if (matchLine && matchLine[1] && matchLine[2] && matchLine[3] !== undefined) {
+                parsed.push({
+                    path: matchLine[1].startsWith(rootPrefix) ? matchLine[1].slice(rootPrefix.length) : matchLine[1],
+                    line: parseInt(matchLine[2], 10),
+                    content: matchLine[3],
+                    isMatch: true,
+                });
+                continue;
+            }
+            // Try context line (hyphen after line number)
+            const ctxLine = raw.match(/^(.+?)-(\d+)-(.*)/);
+            if (ctxLine && ctxLine[1] && ctxLine[2] && ctxLine[3] !== undefined) {
+                parsed.push({
+                    path: ctxLine[1].startsWith(rootPrefix) ? ctxLine[1].slice(rootPrefix.length) : ctxLine[1],
+                    line: parseInt(ctxLine[2], 10),
+                    content: ctxLine[3],
+                    isMatch: false,
+                });
+            }
+        }
+        // Build TextMatch for each match line with surrounding context
+        for (let i = 0; i < parsed.length; i++) {
+            const p = parsed[i];
+            if (!p.isMatch)
+                continue;
+            if (matches.length >= maxResults)
+                break;
+            const contextBefore = [];
+            const contextAfter = [];
+            // Collect context before
+            for (let j = Math.max(0, i - contextLines); j < i; j++) {
+                const ctx = parsed[j];
+                if (ctx && !ctx.isMatch)
+                    contextBefore.push(ctx.content);
             }
-            else if (entry.isFile()) {
-                const ext = extname(entry.name);
-                // Skip binary files
-                if (BINARY_EXTENSIONS.has(ext))
-                    continue;
-                // Skip files that are too large
-                try {
-                    const fileStat = await stat(fullPath);
-                    if (fileStat.size > MAX_FILE_SIZE)
-                        continue;
-                }
-                catch {
-                    continue;
-                }
-                files.push(relative(rootPath, fullPath));
-                if (files.length >= MAX_WALK_FILES) {
-                    console.warn(`[codesift] walkAllTextFiles: reached ${MAX_WALK_FILES} file limit, returning partial results`);
-                    limitReached = true;
-                    return;
-                }
+            // Collect context after
+            for (let j = i + 1; j <= Math.min(parsed.length - 1, i + contextLines); j++) {
+                const ctx = parsed[j];
+                if (ctx && !ctx.isMatch)
+                    contextAfter.push(ctx.content);
             }
+            const match = { file: p.path, line: p.line, content: p.content };
+            if (contextBefore.length > 0)
+                match.context_before = contextBefore;
+            if (contextAfter.length > 0)
+                match.context_after = contextAfter;
+            matches.push(match);
+        }
+    }
+    return matches;
+}
+// ── Node.js fallback search ───────────────────────────
+/** Search file content for line matches, collecting context lines around each hit. */
+function searchFileForMatches(content, filePath, query, regex, contextLines, maxMatches) {
+    const lines = content.split("\n");
+    const matches = [];
+    for (let i = 0; i < lines.length; i++) {
+        if (matches.length >= maxMatches)
+            break;
+        const line = lines[i];
+        if (line === undefined)
+            continue;
+        const isMatch = regex ? regex.test(line) : line.includes(query);
+        if (!isMatch)
+            continue;
+        const contextBefore = [];
+        for (let j = Math.max(0, i - contextLines); j < i; j++) {
+            const ctxLine = lines[j];
+            if (ctxLine !== undefined)
+                contextBefore.push(ctxLine);
+        }
+        const contextAfter = [];
+        for (let j = i + 1; j <= Math.min(lines.length - 1, i + contextLines); j++) {
+            const ctxLine = lines[j];
+            if (ctxLine !== undefined)
+                contextAfter.push(ctxLine);
+        }
+        const truncLine = line.length > MAX_LINE_CHARS
+            ? line.slice(0, MAX_LINE_CHARS) + "..."
+            : line;
+        const match = {
+            file: filePath,
+            line: i + 1,
+            content: truncLine,
+        };
+        if (contextBefore.length > 0)
+            match.context_before = contextBefore;
+        if (contextAfter.length > 0)
+            match.context_after = contextAfter;
+        matches.push(match);
+    }
+    return matches;
+}
+/** Aggregate flat TextMatch[] into per-file groups with counts and first_match preview. */
+function groupMatchesByFile(matches) {
+    const groups = new Map();
+    for (const m of matches) {
+        const existing = groups.get(m.file);
+        if (existing) {
+            existing.count++;
+            existing.lines.push(m.line);
+        }
+        else {
+            groups.set(m.file, {
+                file: m.file,
+                count: 1,
+                lines: [m.line],
+                first_match: m.content.length > MAX_FIRST_MATCH_CHARS
+                    ? m.content.slice(0, MAX_FIRST_MATCH_CHARS) + "..."
+                    : m.content,
+            });
         }
     }
-    await walk(rootPath);
-    return files;
+    return [...groups.values()];
 }
+// ── Public API ──────────────────────────────────────────
 /**
  * Search symbols by name/signature/docstring using BM25 ranking.
  * Supports filtering by symbol kind and file pattern.
@@ -161,149 +369,175 @@ export async function searchSymbols(repo, query, options) {
         throw new Error(`Repository "${repo}" not found. Run index_folder first.`);
     }
     const config = loadConfig();
-    const topK = options?.top_k ?? config.defaultTopK;
     const includeSource = options?.include_source ?? true;
-    const hasKindFilter = !!options?.kind;
-    const hasFileFilter = !!options?.file_pattern;
-    const hasFilters = hasKindFilter || hasFileFilter;
+    const defaultK = (includeSource && !options?.file_pattern) ? DEFAULT_TOP_K_WITH_SOURCE : config.defaultTopK;
+    const topK = options?.top_k ?? defaultK;
+    const hasFilters = !!options?.kind || !!options?.file_pattern;
     let results;
     if (!query.trim()) {
-        // Empty query: return all symbols matching filters (no BM25 scoring)
         const allSymbols = [...index.symbols.values()];
-        let filtered = allSymbols;
-        if (hasKindFilter) {
-            const kind = options.kind;
-            filtered = filtered.filter((s) => s.kind === kind);
-        }
-        if (hasFileFilter) {
-            const pattern = options.file_pattern;
-            filtered = filtered.filter((s) => matchFilePattern(s.file, pattern));
-        }
-        results = filtered.slice(0, topK).map((symbol) => ({
-            symbol,
-            score: 0,
-        }));
+        const filtered = allSymbols.filter((s) => matchesSymbolFilters(s, options));
+        results = filtered.slice(0, topK).map((symbol) => ({ symbol, score: 0 }));
     }
     else {
-        // When filters are active, search a wider candidate set from BM25
-        // so that post-filter truncation doesn't lose relevant results.
-        const searchTopK = hasFilters ? Math.max(topK * 5, 200) : topK;
+        const searchTopK = hasFilters ? Math.max(topK * BM25_FILTER_MULTIPLIER, BM25_FILTER_MIN_K) : topK;
         results = searchBM25(index, query, searchTopK, config.bm25FieldWeights);
-        // Filter by symbol kind
-        if (hasKindFilter) {
-            const kind = options.kind;
-            results = results.filter((r) => r.symbol.kind === kind);
-        }
-        // Filter by file pattern
-        if (hasFileFilter) {
-            const pattern = options.file_pattern;
-            results = results.filter((r) => matchFilePattern(r.symbol.file, pattern));
-        }
-        // Re-truncate to requested top_k after filtering
+        results = results.filter((r) => matchesSymbolFilters(r.symbol, options));
         results = results.slice(0, topK);
+        results = applyCutoff(results);
     }
-    // Strip source if not requested
-    if (!includeSource) {
-        results = results.map((r) => {
-            const { source: _source, ...symbolWithoutSource } = r.symbol;
-            return { ...r, symbol: symbolWithoutSource };
-        });
+    if (options?.rerank && results.length > 1) {
+        const { rerankResults } = await import("../search/reranker.js");
+        results = await rerankResults(query, results);
     }
-    // Truncate source to source_chars limit (default 500 when include_source=true)
-    const sourceChars = options?.source_chars ?? (includeSource ? 500 : undefined);
-    if (includeSource && sourceChars !== undefined && sourceChars > 0) {
-        results = results.map((r) => {
-            const source = r.symbol.source;
-            if (source && source.length > sourceChars) {
-                return {
-                    ...r,
-                    symbol: { ...r.symbol, source: source.slice(0, sourceChars) + "..." },
-                };
-            }
-            return r;
-        });
+    const detail = options?.detail_level ?? "standard";
+    const shaped = shapeSearchResults(results, detail, includeSource, options);
+    // Token budget: greedily pack results until budget exhausted
+    const budget = options?.token_budget;
+    if (budget && budget > 0) {
+        const packed = [];
+        let used = 0;
+        for (const r of shaped) {
+            const tok = Math.ceil(JSON.stringify(r).length / CHARS_PER_TOKEN);
+            if (used + tok > budget)
+                break;
+            packed.push(r);
+            used += tok;
+        }
+        return packed;
     }
-    return results;
+    return shaped;
 }
-/**
- * Full-text search across all files in a repository.
- * Walks the filesystem to search ALL text files, not just indexed ones.
- */
 export async function searchText(repo, query, options) {
     const index = await getCodeIndex(repo);
     if (!index) {
         throw new Error(`Repository "${repo}" not found. Run index_folder first.`);
     }
-    const contextLines = options?.context_lines ?? 2;
     const useRegex = options?.regex ?? false;
     const filePattern = options?.file_pattern;
-    const maxResults = options?.max_results ?? DEFAULT_MAX_TEXT_MATCHES;
-    let regex = null;
+    const maxResults = options?.max_results
+        ?? (useRegex && !filePattern ? DEFAULT_MAX_REGEX_RESULTS : DEFAULT_MAX_TEXT_MATCHES);
+    const contextLines = options?.context_lines ?? 0; // OPT-2: default 0 (was 2) — saves ~30 tokens/match
+    // Validate regex safety before passing to ripgrep
     if (useRegex) {
-        try {
-            regex = new RegExp(query);
-        }
-        catch (err) {
-            const message = err instanceof Error ? err.message : String(err);
-            throw new Error(`Invalid regex pattern: ${message}`);
-        }
+        compileSearchRegex(query); // throws on ReDoS patterns
     }
-    // Walk the filesystem to find ALL text files (not just indexed/parseable ones)
-    const allFiles = await walkAllTextFiles(index.root);
-    const matches = [];
-    for (const filePath of allFiles) {
-        if (matches.length >= maxResults)
-            break;
-        // Filter by file pattern
-        if (filePattern && !matchFilePattern(filePath, filePattern)) {
-            continue;
-        }
-        const fullPath = join(index.root, filePath);
-        let content;
-        try {
-            content = await readFile(fullPath, "utf-8");
+    let matches;
+    // OPT-1: Use ripgrep when available (10x faster)
+    if (hasRipgrep()) {
+        matches = searchWithRipgrep(index.root, query, {
+            regex: useRegex,
+            filePattern: filePattern,
+            maxResults: maxResults,
+            contextLines: contextLines,
+        });
+    }
+    else {
+        // Node.js fallback
+        const regex = useRegex ? compileSearchRegex(query) : null;
+        let allFiles;
+        if (filePattern) {
+            allFiles = index.files.map((f) => f.path);
         }
-        catch {
-            continue; // File may have been deleted or moved
+        else {
+            allFiles = await walkDirectory(index.root, {
+                fileFilter: (ext) => !BINARY_EXTENSIONS.has(ext),
+                maxFiles: MAX_WALK_FILES,
+                relative: true,
+            });
         }
-        const lines = content.split("\n");
-        for (let i = 0; i < lines.length; i++) {
+        matches = [];
+        const searchStart = Date.now();
+        for (const filePath of allFiles) {
             if (matches.length >= maxResults)
                 break;
-            const line = lines[i];
-            if (line === undefined)
-                continue;
-            const isMatch = regex ? regex.test(line) : line.includes(query);
-            if (!isMatch)
+            if (Date.now() - searchStart > SEARCH_TIMEOUT_MS)
+                break;
+            if (filePattern && !matchFilePattern(filePath, filePattern))
                 continue;
-            const contextBefore = [];
-            const contextAfter = [];
-            for (let j = Math.max(0, i - contextLines); j < i; j++) {
-                const ctxLine = lines[j];
-                if (ctxLine !== undefined) {
-                    contextBefore.push(ctxLine);
-                }
+            const fullPath = join(index.root, filePath);
+            let content;
+            try {
+                content = await readFile(fullPath, "utf-8");
             }
-            for (let j = i + 1; j <= Math.min(lines.length - 1, i + contextLines); j++) {
-                const ctxLine = lines[j];
-                if (ctxLine !== undefined) {
-                    contextAfter.push(ctxLine);
-                }
+            catch {
+                continue;
             }
-            const match = {
-                file: filePath,
-                line: i + 1, // 1-based
-                content: line,
-            };
-            if (contextBefore.length > 0) {
-                match.context_before = contextBefore;
+            const fileMatches = searchFileForMatches(content, filePath, query, regex, contextLines, maxResults - matches.length);
+            matches.push(...fileMatches);
+        }
+    }
+    // Ranked mode: classify hits with symbol context, deduplicate, and sort by centrality.
+    // Takes precedence over auto_group/compact — returns TextMatch[] with containing_symbol.
+    if (options?.ranked && matches.length > 0) {
+        try {
+            const { classifyHitsWithSymbols } = await import("./search-ranker.js");
+            const bm25Idx = await getBM25Index(repo);
+            if (bm25Idx) {
+                matches = await classifyHitsWithSymbols(matches, index, { centrality: bm25Idx.centrality });
             }
-            if (contextAfter.length > 0) {
-                match.context_after = contextAfter;
+        }
+        catch {
+            // Graceful fallback — return unranked matches if pipeline fails
+        }
+        return matches;
+    }
+    // OPT-3: Compact format — grep-like `file:line: content` output, ~50% less tokens than JSON
+    // Auto-enable when auto_group is set (caller is optimization-aware) and results are small
+    const useCompact = options?.compact
+        ?? (options?.auto_group && contextLines === 0 && matches.length > 0 && matches.length <= AUTO_GROUP_THRESHOLD);
+    if (useCompact && !options?.group_by_file) {
+        // Group by file to avoid repeating long paths (saves ~30% on multi-match files)
+        const groups = new Map();
+        for (const m of matches) {
+            let g = groups.get(m.file);
+            if (!g) {
+                g = [];
+                groups.set(m.file, g);
             }
-            matches.push(match);
+            g.push(`  ${m.line}: ${m.content}`);
+        }
+        if (groups.size === matches.length) {
+            // Each file has 1 match — flat format is fine
+            return matches.map((m) => `${m.file}:${m.line}: ${m.content}`).join("\n");
         }
+        // Grouped: file header + indented matches
+        const parts = [];
+        for (const [file, lines] of groups) {
+            parts.push(`${file}\n${lines.join("\n")}`);
+        }
+        return parts.join("\n");
+    }
+    // Estimate response size; force grouping when output would be enormous
+    const estimatedChars = matches.reduce((sum, m) => {
+        let chars = m.file.length + m.content.length + JSON_OVERHEAD_PER_MATCH;
+        if (m.context_before)
+            chars += m.context_before.reduce((s, l) => s + l.length, 0);
+        if (m.context_after)
+            chars += m.context_after.reduce((s, l) => s + l.length, 0);
+        return sum + chars;
+    }, 0);
+    const shouldGroup = options?.group_by_file
+        || (options?.auto_group && matches.length > AUTO_GROUP_THRESHOLD)
+        || estimatedChars > MAX_RESPONSE_CHARS;
+    if (shouldGroup) {
+        return groupMatchesByFile(matches);
     }
     return matches;
 }
+// ---------------------------------------------------------------------------
+// Semantic search — standalone wrapper around retrieval infrastructure
+// ---------------------------------------------------------------------------
+export async function semanticSearch(repo, query, options) {
+    const { handleSemanticQuery } = await import("../retrieval/semantic-handlers.js");
+    const result = await handleSemanticQuery(repo, {
+        type: "semantic",
+        query,
+        top_k: options?.top_k,
+        file_filter: options?.file_pattern,
+        exclude_tests: options?.exclude_tests,
+        rerank: options?.rerank,
+    });
+    return typeof result.data === "string" ? result.data : JSON.stringify(result.data);
+}
 //# sourceMappingURL=search-tools.js.map