sweet-search 2.5.6 → 2.5.7

package/README.md

@@ -144,27 +144,29 @@ We measure sweet-search four ways — from how much it helps a real agent down t
 
 <table>
 <tr>
-<td width="25%" valign="top">
+<td width="50%" valign="top">
 
-**① Code-retrieval** *(agent-in-the-loop)*<br>
+🤖 **[① Code-retrieval](#bench-code-retrieval)** *(agent-in-the-loop)*<br>
 <sub>Does it make a real coding agent **cheaper and more useful** when it searches your repo? Paired against each model's own grep-and-read loop.</sub>
 
 </td>
-<td width="25%" valign="top">
+<td width="50%" valign="top">
 
-**② Task-completion** *(coming soon)*<br>
+🚧 **[② Task-completion](#bench-task-completion)** *(coming soon)*<br>
 <sub>Does cheaper, denser context **compound** into a higher resolve-rate on multi-step engineering tasks? Harness in progress.</sub>
 
 </td>
-<td width="25%" valign="top">
+</tr>
+<tr>
+<td width="50%" valign="top">
 
-**③ Paper-type IR** *(academic)*<br>
+📄 **[③ Paper-type IR](#bench-paper-type)** *(academic)*<br>
 <sub>The standard NL→code retrieval suites (GCSN, M2CRB, CoSQA…), full-corpus MRR@10.</sub>
 
 </td>
-<td width="25%" valign="top">
+<td width="50%" valign="top">
 
-**④ Engine speed**<br>
+⚡ **[④ Engine speed](#bench-engine-speed)**<br>
 <sub>Raw systems numbers — grep throughput, query latency, rerank kernels, HNSW.</sub>
 
 </td>
@@ -173,6 +175,7 @@ We measure sweet-search four ways — from how much it helps a real agent down t
 
 ---
 
+<a id="bench-code-retrieval"></a>
 ### 🤖 1. Code-retrieval benchmarks — *the agent-in-the-loop test*
 
 We install the evolved agent prompt (the [GEPA-evolved search discipline](#-an-agent-prompt-that-was-evolved-not-written)), point a coding agent at a real repo, and pair it **probe-for-probe against the same model running its own native grep-and-read loop**. Same model, same tasks, same judge — the only difference is whether sweet-search is wired in.
@@ -220,12 +223,14 @@ The win is **harness-adaptive**: where the native loop is disciplined (Claude Co
 
 ---
 
+<a id="bench-task-completion"></a>
 ### 🚧 2. Task-completion benchmarks — *coming soon*
 
 > Retrieval quality is necessary but not sufficient. Cheaper, denser context only matters if it **compounds across a real, multi-step engineering task** — finding the code, understanding it, changing it, and not breaking anything. The next suite measures exactly that: **resolve-rate on SWE-bench-style multi-file tasks**, sweet-search-wired vs. native, on the same paired, multiplicity-controlled bar as above. Harness and pilot are in progress — numbers land here when they clear that bar, and not before.
 
 ---
 
+<a id="bench-paper-type"></a>
 ### 📄 3. Paper-type retrieval benchmarks — *academic NL→code IR*
 
 > [!WARNING]
@@ -271,6 +276,7 @@ and French queries.
 
 ---
 
+<a id="bench-engine-speed"></a>
 ### ⚡ 4. Engine speed — *systems benchmarks, measured in-repo*
 
 <div align="center">
@@ -566,9 +572,36 @@ What it teaches:
 
 > **Chunk → enrich → embed → quantize** — every step on-device and in Rust. Batches are sized to *your CPU's actual cache*, two open code-models do the encoding, and two separate quantizations make the index both **faster to build** and **small enough to live in RAM**. Zero API keys; nothing ever leaves the machine.
 
-| ① Structure-aware chunk | ② Enrich from structure | ③ Embed — two models | ④ Quantize + persist |
-|:--|:--|:--|:--|
-| cAST over tree-sitter ASTs — whole functions, never sliced mid-body | deterministic preamble from the code graph — **no LLM call** | dense **CodeRankEmbed** + per-token **LateOn-Code** | INT8 weights → **2× faster build** · INT4 vectors → **fits in RAM** |
+<table>
+<tr>
+<td width="50%" valign="top">
+
+① 🧩 **[Structure-aware chunk](#idx-chunk)**<br>
+<sub>cAST over tree-sitter ASTs — whole functions, never sliced mid-body</sub>
+
+</td>
+<td width="50%" valign="top">
+
+② 🏷️ **[Enrich from structure](#idx-enrich)**<br>
+<sub>deterministic preamble from the code graph — **no LLM call**</sub>
+
+</td>
+</tr>
+<tr>
+<td width="50%" valign="top">
+
+③ 🤖 **[Embed — two models](#idx-embed)**<br>
+<sub>dense **CodeRankEmbed** + per-token **LateOn-Code**</sub>
+
+</td>
+<td width="50%" valign="top">
+
+④ 🗜️ **[Quantize + persist](#idx-quantize)**<br>
+<sub>INT8 weights → **2× faster build** · INT4 vectors → **fits in RAM**</sub>
+
+</td>
+</tr>
+</table>
 
 **The inference engine, picked for your silicon:**
 
@@ -579,10 +612,12 @@ What it teaches:
 | 🟩 NVIDIA GPU (SM 7.0+) | candle **CUDA**; **flash-attention** on Ampere+ |
 | 💻 No accelerator | **ONNX Runtime INT8** — tuned CPU path, 132 MB model, **zero GPU weights downloaded** |
 
+<a id="idx-chunk"></a>
 ### 🧩 Chunking — every chunk is whole code, never a fixed window
 - **[cAST](https://arxiv.org/abs/2506.15655)** structure-aware chunking over real **tree-sitter** ASTs: a recursive *split-then-merge* greedily packs sibling AST nodes up to the size cap and recurses *into* nodes too big to fit. So a chunk is always a **function, a class, or a contiguous run of declarations** — never a body cut in half, never a string split mid-literal.
 - **14 languages** get true AST grammars — `JS · TS · TSX · Python · Go · Rust · Java · C · C++ · Ruby · PHP · Kotlin · Swift · C#` — and a **39-config regex registry** carries structure-aware chunking to **70+ more extensions**.
 
+<a id="idx-enrich"></a>
 ### 🏷️ Metadata — context the encoder can actually see
 - Every chunk ships its **symbol name · entity type · signature · line span** — the metadata that powers the code graph, `ss-read` annotations, and the self-contained answers everywhere else.
 - **Contextual enrichment:** before embedding, each chunk is prefixed with a structured preamble assembled from the AST + code graph — *file path · enclosing-scope breadcrumb · name & type · merged siblings · the imports it actually uses*. **Both** encoders see it, so a bare `getId()` still retrieves on the class and module around it.
@@ -593,6 +628,7 @@ What it teaches:
 - **Uses every core the hardware really has** — full count on ARM/Apple Silicon; x86 SMT siblings discounted because they don't scale inference linearly.
 - **ORT drives the CPU path** (ONNX Runtime); GPU hosts swap in fused kernels (below). Either way inference runs off the event loop as a napi `AsyncTask`, so tokenization and SQLite writes overlap compute instead of stalling behind it.
 
+<a id="idx-quantize"></a>
 ### 🗜️ Two quantizations — one buys speed, one buys size
 | | **Model weights** · INT8 ORT | **Index vectors** · INT4 binary |
 |:--|:--|:--|
@@ -600,6 +636,7 @@ What it teaches:
 | **Win** | **~2× faster** indexing · 4× smaller model (**132 MB**) | LI index **1.34 GiB → ~396 MiB** · INT4 nibble-packing halves it again |
 | **Fidelity** | **≥ 0.96 cosine** vs FP32 | **no measurable retrieval loss** (A/B-tested vs INT8) |
 
+<a id="idx-embed"></a>
 ### 🤖 Two models — both open, both local, both code-specialized
 - **[CodeRankEmbed](https://huggingface.co/nomic-ai/CodeRankEmbed)** — 768-d dense bi-encoder (137M, Apache-2.0) for first-stage recall.
 - **[LateOn-Code](https://huggingface.co/lightonai/LateOn-Code)** — ModernBERT per-token **late interaction** (149M) for the rerank.

package/core/embedding/embedding-local-model.js

@@ -172,6 +172,7 @@ export function buildLocalSessionOptions(quantLabel = 'q8', coremlAvailable = fa
 
   const sessionOptions = {
     graphOptimizationLevel: 'all',
+    logSeverityLevel: 3, // ERROR — silence ORT's expected "optimized model is machine-specific" warning
     intraOpNumThreads: intraOpThreads,
     interOpNumThreads: interOpThreads,
     executionMode,

package/core/graph/relationship-resolver.js

@@ -160,7 +160,11 @@ export function resolveRelationshipTargets(db) {
 
   resolveAll();
 
-  console.log(`  ✓ Resolved ${resolved}/${unresolved.length} relationships`);
+  if (resolved > 0) {
+    console.log(`  ✓ Linked ${resolved}/${unresolved.length} references to local definitions`);
+  } else {
+    console.log(`  ${unresolved.length} references resolve to external/library symbols (no local definition to link)`);
+  }
   if (ambiguous > 0) {
     console.log(`  ⚠ ${ambiguous} ambiguous targets (multiple matches)`);
   }

package/core/indexing/index-codebase-v21.js

@@ -140,10 +140,6 @@ async function main() {
     applyPersistedLiModel(process.env.SWEET_SEARCH_PROJECT_ROOT || process.cwd());
   }
 
-  log(`${colors.bright}╔═══════════════════════════════════════════════════╗${colors.reset}`, 'bright');
-  log(`${colors.bright}║   Sweet Search Codebase Indexer v2.3 (SOTA Dec'25) ║${colors.reset}`, 'bright');
-  log(`${colors.bright}╚═══════════════════════════════════════════════════╝${colors.reset}`, 'bright');
-
   if (vectorsOnly) {
     log('⚠ WARNING: --vectors-only skips code graph rebuild', 'yellow');
     log('  GraphRAG structural queries will use stale data', 'yellow');
@@ -338,13 +334,13 @@ Output:
     }
 
     // =========================================================================
-    // PHASE 3: Code Graph + HCGS Preparation (if not --vectors-only)
+    // PHASE 1: Code Graph (if not --vectors-only)
     // =========================================================================
     let graphStats = { entities: 0, relationships: 0 };
     let hcgsPromise = null;
 
     if (!vectorsOnly) {
-      const graphResult = await runPhase('Code Graph + HCGS Prep', buildCodeGraphWithHCGSPhase, {
+      const graphResult = await runPhase('Code Graph', buildCodeGraphWithHCGSPhase, {
         allFiles,
         filesToIndex,
         dryRun,

package/core/indexing/indexer-ann.js

@@ -396,7 +396,7 @@ function diversityFirstPermutationRowids(filePaths) {
 // =============================================================================
 
 export async function incrementalUpdateHNSW(dbPath, changedFiles, dryRun = false) {
-  log('\n━━━ Phase 3: HNSW Index (Incremental) ━━━', 'bright');
+  log('\n━━━ Phase 4: HNSW Index (Incremental) ━━━', 'bright');
 
   if (dryRun) {
     log('DRY RUN: Skipping HNSW incremental update', 'magenta');
@@ -510,7 +510,7 @@ export async function incrementalUpdateHNSW(dbPath, changedFiles, dryRun = false
 // =============================================================================
 
 export async function buildHNSWIndex(dbPath, dryRun = false) {
-  log('\n━━━ Phase 3: HNSW Index ━━━', 'bright');
+  log('\n━━━ Phase 4: HNSW Index ━━━', 'bright');
 
   if (dryRun) {
     log('DRY RUN: Skipping HNSW index', 'magenta');
@@ -705,7 +705,7 @@ export async function buildLateInteractionIndex(chunks, dryRun = false, filesToR
     segmentSize = null, // override SSLX-v3 segment threshold (default 10k)
     projectRoot,        // honored by LI skip policy for .sweet-search.config.json excludes
   } = options;
-  log('\n━━━ Phase 4: Late Interaction Index (LateOn-Code) ━━━', 'bright');
+  log('\n━━━ Phase 3: Late Interaction Index (LateOn-Code) ━━━', 'bright');
 
   if (dryRun) {
     log('DRY RUN: Skipping late interaction index', 'magenta');

package/core/indexing/indexer-build.js

@@ -643,7 +643,7 @@ export async function chunkFiles(files) {
     try {
       const enriched = await enrichChunksFromGraph(allChunks, ASTChunker);
       if (enriched > 0) {
-        log(`✓ Enriched ${enriched}/${allChunks.length} chunks with scope/import context`, 'green');
+        log(`✓ Added scope/import context to ${enriched} code chunks`, 'green');
       }
     } catch (err) {
       log(`⚠ Chunk enrichment skipped: ${err.message}`, 'yellow');

package/core/indexing/indexer-utils.js

@@ -110,19 +110,22 @@ export function isVerboseMode() {
 }
 
 // ---------------------------------------------------------------------------
-// Progress rendering — an in-place "sticky" bar that animates as a phase runs.
+// Progress rendering — a live region of animated, in-place bars.
 //
-// On a TTY (verbose or not) the bar redraws on a single line via carriage return
-// + erase-to-EOL, with smooth 1/8-block fill. While a bar is active, log() pins it:
-// it clears the bar, prints the log line above, then redraws the bar below — so
-// interleaved diagnostics (e.g. the HNSW "checkpoint:" line) never split the bar.
-// Non-TTY (pipes / CI) falls back to throttled newlines so nothing is swallowed.
+// On a TTY (verbose included), each phase's bar animates in place via cursor
+// moves + erase-to-EOL, with smooth 1/8-block fill. Multiple bars can run at
+// once (e.g. Embedding + Late Interaction in parallel) — they share one pinned
+// region at the bottom and update independently. While bars are live, log()
+// prints its line above the region and redraws the bars below, so diagnostics
+// never split a bar. The region "commits" (stays on screen) once every bar in
+// it has reached 100%. Non-TTY (pipes / CI) falls back to throttled newlines.
 // ---------------------------------------------------------------------------
 const BAR_WIDTH = 30;
 const LABEL_COL = 17;           // pad "Label:" to this width so every bar's [ ] aligns
 const SUB_BLOCKS = ['', '▏', '▎', '▍', '▌', '▋', '▊', '▉']; // eighth-block partial fills
 const CLEAR_EOL = '\x1b[K';
-let activeBar = null;           // last-rendered bar string while a phase is in progress (TTY only)
+const liveBars = new Map();     // label -> { current, total }; insertion order = display order
+let regionLines = 0;            // bar lines currently pinned at the bottom (TTY)
 let lastLoggedPercent = {};
 
 function renderBar(current, total, label) {
@@ -137,12 +140,21 @@ function renderBar(current, total, label) {
   return `${colors.cyan}${head}[${bar}${empty}] ${pct}% (${current}/${total})${colors.reset}`;
 }
 
+function drawRegion() {
+  let out = regionLines > 0 ? `\x1b[${regionLines}A\r` : '\r';
+  for (const [label, b] of liveBars) out += renderBar(b.current, b.total, label) + CLEAR_EOL + '\n';
+  process.stdout.write(out);
+  regionLines = liveBars.size;
+}
+
 export function log(message, color = 'reset') {
   if (quietMode) return;
   const line = `${colors[color]}${message}${colors.reset}`;
-  if (activeBar && process.stdout.isTTY) {
-    // Pin the bar: clear it, print the log line above, redraw the bar below.
-    process.stdout.write(`\r${CLEAR_EOL}${line}\n${activeBar}${CLEAR_EOL}`);
+  if (regionLines > 0 && process.stdout.isTTY) {
+    // Print the line above the pinned bars, then redraw the bars below it.
+    let out = `\x1b[${regionLines}A\r${line}${CLEAR_EOL}\n`;
+    for (const [label, b] of liveBars) out += renderBar(b.current, b.total, label) + CLEAR_EOL + '\n';
+    process.stdout.write(out);
   } else {
     console.log(line);
   }
@@ -160,13 +172,16 @@ export function logProgress(current, total, label) {
     }
     return;
   }
-  // Interactive TTY: animate the bar in place.
-  activeBar = renderBar(current, total, label);
-  process.stdout.write(`\r${activeBar}${CLEAR_EOL}`);
-  if (current >= total) {
-    process.stdout.write('\n');
-    activeBar = null;
-    lastLoggedPercent[label] = 0;
+  // Interactive TTY: update this bar in the live region and redraw.
+  liveBars.set(label, { current, total });
+  drawRegion();
+  // Once every live bar is complete, commit the region (leave it on screen).
+  let allDone = true;
+  for (const b of liveBars.values()) if (b.current < b.total) { allDone = false; break; }
+  if (allDone) {
+    for (const k of liveBars.keys()) lastLoggedPercent[k] = 0;
+    liveBars.clear();
+    regionLines = 0;
   }
 }
 

package/core/infrastructure/onnx-session-utils.js

@@ -192,6 +192,7 @@ export function buildSessionOptions(modelId, suffix, coremlAvailable = false, ru
     ?? parseInt(process.env.SWEET_SEARCH_ORT_INTER_OP_THREADS || '1', 10);
   const opts = {
     graphOptimizationLevel: 'all',
+    logSeverityLevel: 3, // ERROR — silence ORT's expected "optimized model is machine-specific" warning
     intraOpNumThreads: runtimeOptions.intraOpThreads ?? bestIntraOpThreads(runtimeOptions),
     interOpNumThreads: interOpThreads,
     executionMode,

package/core/ranking/late-interaction-model.js

@@ -193,6 +193,7 @@ async function loadModel() {
   const { getOptimizedGraphPath } = await import('../infrastructure/onnx-session-utils.js');
   const session = await ort.InferenceSession.create(onnxPath, {
     executionProviders: ['cpu'],
+    logSeverityLevel: 3, // ERROR — silence ORT's expected "optimized model is machine-specific" warning
     intraOpNumThreads: lateInteractionRuntimeConfig.intraOpThreads ?? bestIntraOpThreads(),
     interOpNumThreads: 1,
     optimizedModelFilePath: getOptimizedGraphPath(modelConfig.hfId, 'lateon'),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "sweet-search",
-  "version": "2.5.6",
+  "version": "2.5.7",
   "description": "Sweet Search - SOTA Hybrid Code Search Engine with WASM CatBoost Query Router, Semantic/Lexical/Structural Search, and Multilingual Support",
   "type": "module",
   "main": "core/search/sweet-search.js",
@@ -163,12 +163,12 @@
   },
   "optionalDependencies": {
     "usearch": "^2.21.4",
-    "@sweet-search/native-darwin-arm64": "2.5.6",
-    "@sweet-search/native-darwin-x64": "2.5.6",
-    "@sweet-search/native-linux-arm64-gnu": "2.5.6",
-    "@sweet-search/native-linux-arm64-gnu-cuda": "2.5.6",
-    "@sweet-search/native-linux-x64-gnu": "2.5.6",
-    "@sweet-search/native-linux-x64-gnu-cuda": "2.5.6"
+    "@sweet-search/native-darwin-arm64": "2.5.7",
+    "@sweet-search/native-darwin-x64": "2.5.7",
+    "@sweet-search/native-linux-arm64-gnu": "2.5.7",
+    "@sweet-search/native-linux-arm64-gnu-cuda": "2.5.7",
+    "@sweet-search/native-linux-x64-gnu": "2.5.7",
+    "@sweet-search/native-linux-x64-gnu-cuda": "2.5.7"
   },
   "engines": {
     "node": ">=18.0.0"

package/scripts/init.js

@@ -252,9 +252,26 @@ export function detectProjectRoot(cwd = process.cwd()) {
 export function ensureDataDir(projectRoot) {
   const dataDir = join(projectRoot, DATA_DIR_NAME);
   mkdirSync(dataDir, { recursive: true });
+  maybeIgnoreDataDir(projectRoot);
   return dataDir;
 }
 
+// Add `.sweet-search/` to the project's .gitignore so the local index isn't
+// committed — but ONLY if a .gitignore already exists. We never create one for
+// a project that doesn't already use it.
+function maybeIgnoreDataDir(projectRoot) {
+  try {
+    const gitignorePath = join(projectRoot, '.gitignore');
+    if (!existsSync(gitignorePath)) return;
+    const content = readFileSync(gitignorePath, 'utf8');
+    const already = content.split(/\r?\n/).map((l) => l.trim().replace(/^\//, '').replace(/\/$/, ''))
+      .some((l) => l === DATA_DIR_NAME);
+    if (already) return;
+    const sep = content.length === 0 || content.endsWith('\n') ? '' : '\n';
+    writeFileSync(gitignorePath, `${content}${sep}\n# Sweet Search local index\n${DATA_DIR_NAME}/\n`);
+  } catch { /* best-effort — never block init on .gitignore */ }
+}
+
 // ---------------------------------------------------------------------------
 // Init config read/write
 // ---------------------------------------------------------------------------
@@ -1576,11 +1593,10 @@ export async function runInit(args) {
   const skippedOptIns = getSkippedOptInModels(profile);
   let modelResults = new Map();
 
-  // Tell the user once which optional models are being skipped. This is
-  // NOT an error — these are opt-in features (e.g. cross-encoder
-  // rerankers disabled by default since commit 43a61eb). Without this
-  // line, init silently omitting them looks like a missing-model bug.
-  if (skippedOptIns.length > 0) {
+  // Opt-in models (e.g. cross-encoder rerankers, disabled by default since
+  // commit 43a61eb) are skipped silently — they're optional features, not
+  // missing models. Set DEBUG=1 to see which were skipped and how to enable.
+  if (process.env.DEBUG && skippedOptIns.length > 0) {
     for (const skipped of skippedOptIns) {
       process.stderr.write(
         `[init] Skipping opt-in model "${skipped.key}" — ` +

package/scripts/postinstall-banner.js

@@ -1,46 +1,34 @@
 #!/usr/bin/env node
 /**
- * postinstall — play the animated banner once after install.
+ * postinstall — print a short "what next" message after install.
  *
- * npm pipes lifecycle-script stdout (it's not a TTY), so we render to the
- * controlling terminal directly via /dev/tty when possible. This is Unix-only;
- * on Windows (no /dev/tty) or when there is no controlling terminal (CI, detached,
- * sandboxed installs) we simply skip.
- *
- * Defensive by design: renders only to a real terminal, honours CI / NO_BANNER /
- * SWEET_SEARCH_NO_BANNER, swallows every error, and always exits 0 so it can never
- * fail `npm install`.
+ * Deliberately plain text: during `npm install` npm writes its own progress
+ * spinner to the terminal CONCURRENTLY with this script, which would interrupt
+ * any graphics/animation escape sequence mid-stream and leak its payload as
+ * garbage text. So the rich animated banner is reserved for `sweet-search init`
+ * and `sweet-search index` (where we own the TTY); install just prints a clean,
+ * escape-light pointer. Best-effort; always exits 0 so it can't fail an install.
  */
 import process from 'node:process';
-import tty from 'node:tty';
-import { openSync, closeSync } from 'node:fs';
-import { dirname, join } from 'node:path';
-import { fileURLToPath } from 'node:url';
 
-async function run() {
+function run() {
   const env = process.env;
-  if (env.CI || env.NO_BANNER || env.SWEET_SEARCH_NO_BANNER) return;
-
-  // Pick an output stream that is a real terminal.
-  let stream = process.stdout.isTTY ? process.stdout : null;
-  let ownedFd = -1;
-  if (!stream && process.platform !== 'win32') {
-    try {
-      ownedFd = openSync('/dev/tty', 'r+');     // throws if no controlling terminal
-      const s = new tty.WriteStream(ownedFd);
-      if (s.isTTY) stream = s;
-    } catch { /* no controlling terminal — skip */ }
-  }
-  if (!stream) return;
-
+  if (env.NO_BANNER || env.SWEET_SEARCH_NO_BANNER) return;
   try {
-    const here = dirname(fileURLToPath(import.meta.url));
-    const { showBanner } = await import(join(here, '..', 'core', 'banner', 'render-banner.js'));
-    // query:false — we have no matching stdin for this tty stream; rely on env-based detection.
-    const shown = await showBanner({ stream, env, query: false, maxMs: 2200 });
-    if (shown) stream.write('  sweet-search installed — run `sweet-search init` to get started.\n');
+    const c = (n, s) => (process.stdout.isTTY ? `\x1b[${n}m${s}\x1b[0m` : s);
+    const lines = [
+      '',
+      `  ${c('1;38;5;213', 'sweet-search')} installed ${c('2', '— SOTA hybrid code search')}`,
+      '',
+      `  ${c('1', 'Get started:')}`,
+      `    ${c('36', 'sweet-search init')}        set up the current project`,
+      `    ${c('36', 'sweet-search index')}       build the search index`,
+      `    ${c('36', 'sweet-search "query"')}     search your code`,
+      `  ${c('2', '(installed locally? prefix with')} ${c('2;36', 'npx')}${c('2', ', e.g. `npx sweet-search init`)')}`,
+      '',
+    ];
+    process.stdout.write(lines.join('\n') + '\n');
   } catch { /* never break an install */ }
-  finally { if (ownedFd >= 0) { try { closeSync(ownedFd); } catch { /* noop */ } } }
 }
 
-run().finally(() => process.exit(0));
+run();