offgrid-ai 0.5.4 → 0.6.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "offgrid-ai",
-  "version": "0.5.4",
+  "version": "0.6.0",
   "description": "Privacy-first CLI for running local LLMs — discover, configure, run, benchmark",
   "author": "Eeshan Srivastava (https://eeshans.com)",
   "type": "module",

package/src/benchmark.mjs

@@ -0,0 +1,277 @@
+import { mkdir, writeFile, readFile, readdir } from "node:fs/promises";
+import { existsSync } from "node:fs";
+import { createHash } from "node:crypto";
+import { join, resolve } from "node:path";
+import { homedir } from "node:os";
+import { execFile } from "node:child_process";
+import { promisify } from "node:util";
+import { ensureDirs, loadConfig, saveConfig } from "./config.mjs";
+import { pc, createPrompt, renderRows, renderSection } from "./ui.mjs";
+
+const execFileAsync = promisify(execFile);
+
+// ── Shared utilities (matches local-llm-visual-benchmark/scripts/local-llm/shared/run-utils.mjs) ──
+
+export function slugModelId(modelId, maxLength = 80) {
+  const hash = createHash("sha256").update(modelId).digest("hex").slice(0, 10);
+  const normalized = modelId.normalize("NFKD").replace(/[\u0300-\u036f]/gu, "").toLowerCase();
+  const slug = normalized.replace(/[^a-z0-9]+/gu, "-").replace(/^-+|-+$/gu, "").replace(/-{2,}/gu, "-");
+  if (slug.length > 0 && slug.length <= maxLength && slug === normalized) return slug;
+  const baseMaxLength = Math.max(1, maxLength - 11);
+  const base = slug.slice(0, baseMaxLength).replace(/^-+|-+$/gu, "") || "model";
+  return `${base}-${hash}`;
+}
+
+export function createRunId(date = new Date()) {
+  return date.toISOString().replace(/:/gu, "-").replace(/\./gu, "-");
+}
+
+export function buildToolPrompt(benchmark, kind) {
+  if (kind === "data-science") return benchmark.prompt;
+  return [
+    "Create a complete, self-contained HTML file for the request below.",
+    "Write the file as `index.html` in the current working directory.",
+    "Do not create any folders, do not infer a filesystem path, and do not print the HTML in chat.",
+    "",
+    "The HTML must include all CSS and JavaScript inline and must not depend on external network assets.",
+    "After building the page, run a visual QA pass with agent-browser or Playwright: open the saved index.html, inspect the rendered result, and fix any obvious layout, animation, console, or viewport issues before you finish.",
+    "",
+    benchmark.prompt,
+  ].join("\n");
+}
+
+export async function loadBenchmarks(benchDir) {
+  const entries = await readdir(benchDir);
+  const markdownFiles = entries.filter((f) => f.endsWith(".md")).sort();
+  const benchmarks = [];
+  for (const filename of markdownFiles) {
+    const raw = await readFile(join(benchDir, filename), "utf8");
+    // Simple frontmatter parser (no gray-matter dependency)
+    const match = raw.match(/^---\n([\s\S]*?)\n---\n([\s\S]*)$/);
+    const frontmatter = match ? match[1] : "";
+    const content = match ? match[2].trim() : raw.trim();
+    let id = filename.replace(/\.md$/u, "");
+    let title = id;
+    let description = "";
+    let kind = "visual";
+    for (const line of frontmatter.split("\n")) {
+      const kv = line.match(/^(\w+):\s*(.+)$/);
+      if (kv) {
+        const [, key, val] = kv;
+        if (key === "id") id = val.trim();
+        if (key === "title") title = val.trim();
+        if (key === "description") description = val.trim();
+        if (key === "kind") kind = val.trim();
+      }
+    }
+    if (id === "ab-test-analysis") kind = "data-science";
+    benchmarks.push({ id, title, description, prompt: content, kind });
+  }
+  return benchmarks;
+}
+
+// ── Benchmark repo linking ────────────────────────────────────────────────
+
+const BENCHMARK_REPO = "https://github.com/eeshansrivastava89/local-llm-visual-benchmark.git";
+
+export async function findBenchmarkRepo() {
+  const config = await loadConfig();
+  if (config.benchmarkRepoPath && existsSync(join(config.benchmarkRepoPath, "benchmarks"))) {
+    return config.benchmarkRepoPath;
+  }
+  return null;
+}
+
+export async function linkBenchmarkRepo(prompt) {
+  const existing = await findBenchmarkRepo();
+  if (existing) return existing;
+
+  // Check common locations
+  const candidates = [
+    join(homedir(), "dev", "local-llm-visual-benchmark"),
+    join(homedir(), "projects", "local-llm-visual-benchmark"),
+    join(homedir(), "local-llm-visual-benchmark"),
+  ];
+  for (const candidate of candidates) {
+    if (existsSync(join(candidate, "benchmarks"))) {
+      const config = await loadConfig();
+      config.benchmarkRepoPath = candidate;
+      await saveConfig(config);
+      return candidate;
+    }
+  }
+
+  console.log(pc.dim("\nThe benchmark gallery needs to be linked to offgrid-ai."));
+  console.log(pc.dim("This is the local-llm-visual-benchmark repo that stores prompts and run results.\n"));
+
+  const choice = await prompt.choice("Link benchmark gallery", [
+    { value: "clone", label: "Clone from GitHub", hint: "git clone into ~/dev" },
+    { value: "manual", label: "Enter path manually", hint: "If you already have it cloned" },
+  ], "clone");
+
+  if (choice === "clone") {
+    const targetDir = join(homedir(), "dev", "local-llm-visual-benchmark");
+    console.log(pc.dim(`\nCloning ${BENCHMARK_REPO}...`));
+    try {
+      await execFileAsync("git", ["clone", BENCHMARK_REPO, targetDir], { stdio: "pipe" });
+      const config = await loadConfig();
+      config.benchmarkRepoPath = targetDir;
+      await saveConfig(config);
+      console.log(pc.green(`✓ Cloned to ${targetDir}`));
+      return targetDir;
+    } catch (err) {
+      console.log(pc.red(`Clone failed: ${err.message}`));
+      return null;
+    }
+  }
+
+  // Manual path
+  const path = await prompt.text("Path to local-llm-visual-benchmark", "");
+  if (!path) return null;
+  const resolved = resolve(path.replace(/^~/, homedir()));
+  if (!existsSync(join(resolved, "benchmarks"))) {
+    console.log(pc.red(`No benchmarks/ directory found at ${resolved}`));
+    return null;
+  }
+  const config = await loadConfig();
+  config.benchmarkRepoPath = resolved;
+  await saveConfig(config);
+  console.log(pc.green(`✓ Linked to ${resolved}`));
+  return resolved;
+}
+
+// ── Prepare a benchmark run ────────────────────────────────────────────────
+
+export async function benchmarkFlow() {
+  await ensureDirs();
+
+  const prompt = createPrompt();
+  try {
+    const repoPath = await linkBenchmarkRepo(prompt);
+    if (!repoPath) return;
+
+    // 1. Pick category
+    const kind = await prompt.choice("Benchmark category", [
+      { value: "visual", label: "Visual Benchmark", hint: "HTML/CSS/JS animation benchmarks" },
+      { value: "data-science", label: "Data Science", hint: "Analysis and charting benchmarks" },
+    ], "visual");
+
+    // 2. Pick benchmark prompt
+    const benchDir = join(repoPath, "benchmarks");
+    const benchmarks = (await loadBenchmarks(benchDir)).filter((b) => b.kind === kind);
+    if (benchmarks.length === 0) {
+      console.log(pc.yellow(`No ${kind} benchmarks found in ${benchDir}`));
+      return;
+    }
+    const benchmarkId = await prompt.choice("Prompt", benchmarks.map((b) => ({
+      value: b.id,
+      label: b.title,
+      hint: b.description || b.id,
+    })), benchmarks[0].id);
+    const selectedBenchmark = benchmarks.find((b) => b.id === benchmarkId);
+    if (!selectedBenchmark) return;
+
+    // 3. Pick model source
+    const { loadProfiles } = await import("./profiles.mjs");
+    const { backendFor } = await import("./backends.mjs");
+
+    const profiles = await loadProfiles();
+    const source = await prompt.choice("Model source", [
+      { value: "profile", label: "Use existing profile", hint: "Pick a saved offgrid-ai profile" },
+      { value: "cloud", label: "Custom / cloud", hint: "Free-form model label for cloud runs" },
+    ], "profile");
+
+    let modelId, modelSource, backendLabel;
+
+    if (source === "profile") {
+      if (profiles.length === 0) {
+        console.log(pc.yellow("No profiles yet. Run: offgrid-ai models"));
+        return;
+      }
+      const profileId = await prompt.choice("Profile", profiles.map((p) => ({
+        value: p.id,
+        label: p.label,
+        hint: `${backendFor(p.backend).label} · ${p.modelAlias}`,
+      })), profiles[0].id);
+      const profile = profiles.find((p) => p.id === profileId);
+      if (!profile) return;
+      modelId = profile.modelAlias;
+      modelSource = profile.providerId === "llama-cpp-mtp" ? "llama-cpp-mtp" : profile.backend === "ollama" ? "ollama" : profile.backend === "omlx" ? "omlx" : "llama-cpp";
+      backendLabel = backendFor(profile.backend).label;
+    } else {
+      backendLabel = await prompt.text("Backend label", "cloud");
+      modelId = await prompt.text("Model name", "");
+      if (!modelId) { console.log(pc.yellow("Model name is required.")); return; }
+      modelSource = "cloud";
+    }
+
+    // 4. Create the run directory
+    const toolPrompt = buildToolPrompt(selectedBenchmark, kind);
+    const now = new Date();
+    const runId = createRunId(now);
+    const modelSlug = slugModelId(modelId);
+    const runsDir = join(repoPath, "runs");
+    const benchmarkDirectory = join(runsDir, selectedBenchmark.id);
+    const modelDirectory = join(benchmarkDirectory, modelSlug);
+    const runDirectory = join(modelDirectory, runId);
+
+    await mkdir(runDirectory, { recursive: true });
+
+    const isDs = kind === "data-science";
+    const metadata = {
+      schemaVersion: 1,
+      kind,
+      runId,
+      benchmark: { id: selectedBenchmark.id, title: selectedBenchmark.title, description: selectedBenchmark.description, prompt: selectedBenchmark.prompt },
+      model: { id: modelId, slug: modelSlug },
+      status: "prepared",
+      createdAt: now.toISOString(),
+      updatedAt: now.toISOString(),
+      preparedAt: now.toISOString(),
+      runDirectory,
+      assets: isDs
+        ? { metadata: "metadata.json", prompt: "prompt.md", rawResponse: "response.raw.txt", ds: { notebook: "analysis.ipynb", summary: "summary.json", chartDistribution: "chart-distribution.png", chartTreatmentEffect: "chart-treatment-effect.png", chartCompletionRates: "chart-completion-rates.png" } }
+        : { metadata: "metadata.json", prompt: "prompt.md", html: "index.html", preview: "preview.png", video: "preview.webm", rawResponse: "response.raw.txt" },
+      runner: {
+        mode: source === "profile" ? "external" : "manual",
+        ...(modelSource ? { modelSource } : {}),
+        ...(backendLabel ? { backendLabel } : {}),
+        model: modelId,
+        retries: 0,
+        tokenMetrics: { reported: false },
+      },
+    };
+
+    // Clean up the baseUrl field — only include if we have a real one
+    if (!metadata.runner.baseUrl) delete metadata.runner.baseUrl;
+
+    await writeFile(join(runDirectory, "metadata.json"), JSON.stringify(metadata, null, 2) + "\n", "utf8");
+    await writeFile(join(runDirectory, "prompt.md"), toolPrompt + "\n", "utf8");
+
+    // 5. Print result
+    console.log("");
+    console.log(pc.green("✓ Run slot prepared"));
+    console.log(renderSection("Run", renderRows([
+      ["Directory", pc.cyan(runDirectory)],
+      ["Benchmark", selectedBenchmark.title],
+      ["Kind", kind],
+      ["Model", pc.bold(modelId)],
+      ["Source", backendLabel || modelSource],
+    ])));
+
+    console.log("");
+    console.log(pc.bold("Next steps"));
+    if (source === "profile") {
+      console.log(`  1. ${pc.cyan(`cd ${runDirectory}`)}`);
+      console.log(`  2. ${pc.cyan("offgrid-ai models")} → select the model → Run`);
+      console.log(`  3. In Pi, paste the prompt from ${pc.cyan("prompt.md")}`);
+      console.log(`  4. View results: ${pc.cyan(`cd ${repoPath} && npm run dev`)}`);
+    } else {
+      console.log(`  1. ${pc.cyan(`cd ${runDirectory}`)}`);
+      console.log(`  2. Copy the prompt from ${pc.cyan("prompt.md")} into your tool of choice`);
+      console.log(`  3. View results: ${pc.cyan(`cd ${repoPath} && npm run dev`)}`);
+    }
+  } finally {
+    prompt.close();
+  }
+}

package/src/cli.mjs

@@ -59,6 +59,7 @@ export async function run(argv) {
   if (command === "run") return runCommand(argv.slice(1));
   if (command === "status") return statusCommand();
   if (command === "stop") return stopCommand(argv.slice(1));
+  if (command === "benchmark") return benchmarkCommand();
   if (command === "uninstall" || command === "--uninstall") return uninstallCommand(argv.slice(1));
   if (command === "--verbose") return mainFlow(); // verbose flag handled inside onboardFlow
 
@@ -181,41 +182,37 @@ async function modelCommandCenter(initialCatalog) {
     return;
   }
 
+  const catalog = initialCatalog.newModels ? initialCatalog : await loadModelCatalog();
+  const normalized = normalizeCatalog(catalog);
+  const allItems = buildCatalogItems(normalized);
+  if (allItems.length === 0) { console.log(pc.dim("No models found.")); return; }
+  const runningProfilesNow = [];
+  for (const profile of normalized.profiles) {
+    if (await isProfileRunning(profile)) runningProfilesNow.push(profile);
+  }
+  const fileMissingCount = normalized.profiles.filter((p) => isProfileFileMissing(p)).length;
+  const summaryBorder = fileMissingCount > 0 ? pc.red : normalized.newModels.length > 0 ? pc.yellow : pc.dim;
+  console.log("\n" + renderCard("Your local AI workspace", renderRows([
+    ["Setups", `${normalized.profiles.length} saved`],
+    ["Need setup", normalized.newModels.length > 0 ? pc.yellow(`${normalized.newModels.length} model${normalized.newModels.length === 1 ? "" : "s"}`) : pc.dim("none")],
+    ["Running", runningProfilesNow.length > 0 ? pc.green(String(runningProfilesNow.length)) : pc.dim("none")],
+    ["File missing", fileMissingCount > 0 ? pc.red(`${fileMissingCount} setup${fileMissingCount === 1 ? "" : "s"}`) : pc.dim("none")],
+  ]), { formatBorder: summaryBorder }));
+
+  printModelCards(allItems, { runningProfilesNow, drafters: normalized.drafters });
+
   const prompt = createPrompt();
   try {
-    while (true) {
-      const catalog = await loadModelCatalog();
-      const normalized = normalizeCatalog(catalog);
-      const allItems = buildCatalogItems(normalized);
-      if (allItems.length === 0) { console.log(pc.dim("No models found.")); break; }
-      const runningProfilesNow = [];
-      for (const profile of normalized.profiles) {
-        if (await isProfileRunning(profile)) runningProfilesNow.push(profile);
-      }
-      const fileMissingCount = normalized.profiles.filter((p) => isProfileFileMissing(p)).length;
-      const summaryBorder = fileMissingCount > 0 ? pc.red : normalized.newModels.length > 0 ? pc.yellow : pc.dim;
-      console.log("\n" + renderCard("Your local AI workspace", renderRows([
-        ["Setups", `${normalized.profiles.length} saved`],
-        ["Need setup", normalized.newModels.length > 0 ? pc.yellow(`${normalized.newModels.length} model${normalized.newModels.length === 1 ? "" : "s"}`) : pc.dim("none")],
-        ["Running", runningProfilesNow.length > 0 ? pc.green(String(runningProfilesNow.length)) : pc.dim("none")],
-        ["File missing", fileMissingCount > 0 ? pc.red(`${fileMissingCount} setup${fileMissingCount === 1 ? "" : "s"}`) : pc.dim("none")],
-      ]), { formatBorder: summaryBorder }));
-
-      // Show model cards for all items
-      printModelCards(allItems, { runningProfilesNow, drafters: normalized.drafters });
-
-      const options = allItems.map((item) => modelSelectOption(item, { runningProfilesNow, drafters: normalized.drafters }));
-      options.push({ value: "__exit__", label: pc.dim("Exit"), hint: "Close offgrid-ai" });
-      const selected = await prompt.choice("Select a model", options);
-      if (!selected || selected === "__exit__") break;
-      const item = allItems.find((i) => itemKey(i) === selected);
-      if (!item) break;
-
-      const actions = actionsForItem(item);
-      const action = await prompt.choice(item.label, actions, actions[0].value);
-      if (!action) continue;
-      await performAction(prompt, action, item);
-    }
+    const options = allItems.map((item) => modelSelectOption(item, { runningProfilesNow, drafters: normalized.drafters }));
+    const selected = await prompt.choice("Select a model", options);
+    if (!selected) return;
+    const item = allItems.find((i) => itemKey(i) === selected);
+    if (!item) return;
+
+    const actions = actionsForItem(item);
+    const action = await prompt.choice(item.label, actions, actions[0].value);
+    if (!action) return;
+    await performAction(prompt, action, item);
   } finally {
     prompt.close();
   }
@@ -315,6 +312,11 @@ function actionsForItem(item) {
       { value: "reconfigure", label: "Reconfigure", hint: "Change context, MTP, settings" },
       { value: "inspect", label: "Details", hint: "Paths, ports, flags" },
     ];
+    // Benchmark is offered if repo is linked and model is local
+    const backend = backendFor(item.profile.backend);
+    if (backend.type === "local-server" || backend.type === "managed-server") {
+      actions.push({ value: "benchmark", label: "Benchmark", hint: "Prepare a benchmark run" });
+    }
     if (!item.fileMissing) {
       actions.push({ value: "remove", label: "Remove", hint: "Delete this setup" });
     }
@@ -416,6 +418,10 @@ async function performAction(prompt, action, item) {
     if (item.type === "managed") return printManagedModelDetails(item.model, BACKENDS[item.backendId]);
     return printGgufModelDetails(item.model, item.drafter);
   }
+  if (action === "benchmark") {
+    const { benchmarkFlow } = await import("./benchmark.mjs");
+    return await benchmarkFlow();
+  }
   if (action === "run") {
     if (item.type === "profile") return await runProfile(await readProfile(item.profile.id));
     // Shouldn't reach here for new/managed, but handle gracefully
@@ -714,6 +720,11 @@ async function removeProfileInteractive(id) {
 
 // ── Status ──────────────────────────────────────────────────────────────────
 
+async function benchmarkCommand() {
+  const { benchmarkFlow } = await import("./benchmark.mjs");
+  return await benchmarkFlow();
+}
+
 async function statusCommand() {
   await ensureDirs();
   const profiles = await loadProfiles();
@@ -1227,9 +1238,10 @@ function printHelp() {
     ["Start", pc.bold("offgrid-ai")],
     ["Status", "offgrid-ai status"],
     ["Stop", "offgrid-ai stop"],
+    ["Benchmark", "offgrid-ai benchmark"],
     ["Uninstall", "offgrid-ai uninstall"],
     ["Version", "offgrid-ai version"],
   ]), { formatBorder: pc.cyan }));
-  console.log("\n" + renderCard("How it works", "Run offgrid-ai, choose a local model, and start chatting in Pi.\n\nFirst run walks you through missing tools. After that, offgrid-ai remembers your model setup.", { formatBorder: pc.magenta }));
+  console.log("\n" + renderCard("How it works", "Run offgrid-ai, choose a local model, and start chatting in Pi.\n\nFirst run walks you through missing tools. After that, offgrid-ai remembers your model setup.\n\nFor benchmarks, run offgrid-ai benchmark to prepare a visual or data-science benchmark run.", { formatBorder: pc.magenta }));
   console.log("\n" + pc.dim("Tip: use --verbose only when you want detailed install output."));
 }