offgrid-ai 0.14.0 → 0.14.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "offgrid-ai",
-  "version": "0.14.0",
+  "version": "0.14.1",
   "description": "Privacy-first CLI for running local LLMs — discover, configure, run, benchmark",
   "author": "Eeshan Srivastava (https://eeshans.com)",
   "type": "module",

package/src/benchmark/flow.mjs

@@ -128,6 +128,28 @@ export async function runPreparedBenchmark(profile, runDirectory, options = {})
   return metadata;
 }
 
+// ── Shared benchmark selection ────────────────────────────────────────────
+
+async function selectBenchmark(prompt, repoPath) {
+  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");
+
+  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 null;
+  }
+  const benchmarkId = await prompt.choice("Prompt", benchmarks.map((b) => ({
+    value: b.id, label: b.title, hint: b.description || b.id,
+  })), benchmarks[0].id);
+  const benchmark = benchmarks.find((b) => b.id === benchmarkId);
+  if (!benchmark) return null;
+  return { kind, benchmark };
+}
+
 // ── Benchmark from a selected profile (from model picker) ────────────────
 
 export async function benchmarkForProfile(profile) {
@@ -137,22 +159,9 @@ export async function benchmarkForProfile(profile) {
     const repoPath = await linkBenchmarkRepo(prompt);
     if (!repoPath) return;
 
-    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");
-
-    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;
+    const selected = await selectBenchmark(prompt, repoPath);
+    if (!selected) return;
+    const { kind, benchmark: selectedBenchmark } = selected;
 
     const modelId = profile.modelAlias;
     const modelSource = benchmarkModelSource(profile);
@@ -177,28 +186,14 @@ export async function benchmarkForProfile(profile) {
 
 export async function benchmarkFlow() {
   await ensureDirs();
-
   const prompt = createPrompt();
   try {
     const repoPath = await linkBenchmarkRepo(prompt);
     if (!repoPath) return;
 
-    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");
-
-    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;
+    const selected = await selectBenchmark(prompt, repoPath);
+    if (!selected) return;
+    const { kind, benchmark: selectedBenchmark } = selected;
 
     const profiles = await loadProfiles();
     const source = await prompt.choice("Model source", [

package/src/benchmark/prepare.mjs

@@ -3,7 +3,7 @@
 import { mkdir, writeFile } from "node:fs/promises";
 import { join } from "node:path";
 import { pc, renderRows, renderSection } from "../ui.mjs";
-import { slugModelId, createRunId, buildToolPrompt } from "./shared.mjs";
+import { slugModelId, createRunId } from "./shared.mjs";
 import { parseModelName } from "../model-name.mjs";
 
 function harnessDisplayName(id) {
@@ -30,7 +30,7 @@ function printBenchmarkNextSteps({ repoPath, runDirectory, profile, modelId, run
 }
 
 export async function prepareBenchmarkRun({ repoPath, benchmark, kind, modelId, modelSource, backendLabel, profile, showNextSteps = true }) {
-  const toolPrompt = buildToolPrompt(benchmark);
+  const toolPrompt = benchmark.prompt;
   const now = new Date();
   const runId = createRunId(now);
   const modelSlug = slugModelId(modelId);

package/src/benchmark/sdk-runner.mjs

@@ -5,8 +5,8 @@ import { join, relative, basename } from "node:path";
 import { Agent } from "@earendil-works/pi-agent-core";
 import { streamSimple } from "@earendil-works/pi-ai/compat";
 import { createCodingTools } from "@earendil-works/pi-coding-agent";
-import { pc } from "../ui.mjs";
-import { piApiModelId } from "../harness-pi.mjs";
+import { pc, formatBytes } from "../ui.mjs";
+import { piApiModelId, modelReasoning, modelCompat } from "../harness-pi.mjs";
 
 const C = {
   thinking: pc.magenta,
@@ -257,16 +257,8 @@ export async function runBenchmarkInPi(profile, runDirectory, { signal } = {}) {
 // ── Model construction ──────────────────────────────────────────────────────
 
 function buildModel(profile) {
-  const text = [profile.id, profile.label, profile.modelAlias, profile.modelPath, profile.omlxModel]
-    .filter(Boolean).join(" ").toLowerCase();
-  const reasoning = profile.reasoning !== undefined
-    ? Boolean(profile.reasoning)
-    : /qwen|gemma-4|gemma 4/i.test(text);
-  const hasCompat = profile.compat
-    ? profile.compat
-    : /qwen|gemma-4|gemma 4/i.test(text)
-      ? { thinkingFormat: "qwen-chat-template" }
-      : null;
+  const reasoning = modelReasoning(profile) ?? false;
+  const compat = modelCompat(profile);
 
   return {
     id: piApiModelId(profile),
@@ -283,7 +275,7 @@ function buildModel(profile) {
       supportsDeveloperRole: false,
       supportsReasoningEffort: false,
       maxTokensField: "max_tokens",
-      ...(hasCompat ?? {}),
+      ...(compat ?? {}),
     },
   };
 }
@@ -364,15 +356,6 @@ function truncateOneLine(value, max = 80) {
   return text.length > max ? `${text.slice(0, Math.max(1, max - 1))}…` : text;
 }
 
-function formatBytes(bytes) {
-  if (!Number.isFinite(bytes)) return "unknown";
-  const units = ["B", "KB", "MB", "GB", "TB"];
-  let size = bytes;
-  let unit = 0;
-  while (size >= 1024 && unit < units.length - 1) { size /= 1024; unit += 1; }
-  return `${size.toFixed(unit === 0 ? 0 : 2)} ${units[unit]}`;
-}
-
 function formatTokens(n) {
   if (n >= 1_000_000) return `${(n / 1_000_000).toFixed(1)}M`;
   if (n >= 1_000) return `${Math.round(n / 1_000)}k`;

package/src/benchmark/shared.mjs

@@ -18,10 +18,6 @@ export function createRunId(date = new Date()) {
   return date.toISOString().replace(/:/gu, "-").replace(/\./gu, "-");
 }
 
-export function buildToolPrompt(benchmark) {
-  return benchmark.prompt;
-}
-
 export async function loadBenchmarks(benchDir) {
   const entries = await readdir(benchDir);
   const markdownFiles = entries.filter((f) => f.endsWith(".md")).sort();
@@ -47,8 +43,4 @@ export async function loadBenchmarks(benchDir) {
     benchmarks.push({ id, title, description, prompt: content, kind });
   }
   return benchmarks;
-}
-
-export function piModelString(profile) {
-  return profile.harnesses?.pi?.model ?? `${profile.providerId}/${profile.modelAlias}`;
 }

package/src/benchmark.mjs

@@ -1,7 +1,7 @@
 // ── Benchmark module (thin facade) ──────────────────────────────────────────
 // Submodules handle the actual logic. This file re-exports for backward compatibility.
 
-export { slugModelId, createRunId, buildToolPrompt, loadBenchmarks, piModelString } from "./benchmark/shared.mjs";
+export { slugModelId, createRunId, loadBenchmarks } from "./benchmark/shared.mjs";
 export { findBenchmarkRepo, linkBenchmarkRepo } from "./benchmark/repo.mjs";
 export { prepareBenchmarkRun } from "./benchmark/prepare.mjs";
 export { runBenchmarkInPi } from "./benchmark/sdk-runner.mjs";

package/src/commands/run.mjs

@@ -149,18 +149,6 @@ function textOnlyProfile(profile) {
     mmprojPath: null,
     disabledMmprojPath: profile.disabledMmprojPath ?? profile.mmprojPath,
     capabilities: { ...(profile.capabilities ?? {}), vision: false, visionDisabledReason: "unsupported-mmproj" },
-    commandArgv: removeCommandOption(profile.commandArgv ?? [], "--mmproj"),
   });
 }
 
-function removeCommandOption(argv, flag) {
-  const next = [];
-  for (let i = 0; i < argv.length; i++) {
-    if (argv[i] === flag) {
-      if (argv[i + 1] && !argv[i + 1].startsWith("--")) i += 1;
-      continue;
-    }
-    next.push(argv[i]);
-  }
-  return next;
-}

package/src/harness-pi.mjs

@@ -118,7 +118,7 @@ function modelInput(profile) {
   return ["text"];
 }
 
-function modelCompat(profile) {
+export function modelCompat(profile) {
   if (profile.compat) return profile.compat;
   const family = modelFamily(profile);
   if (family.includes("qwen") || family.includes("gemma-4") || family.includes("gemma 4")) {
@@ -127,14 +127,14 @@ function modelCompat(profile) {
   return null;
 }
 
-function modelReasoning(profile) {
+export function modelReasoning(profile) {
   if (profile.reasoning !== undefined) return Boolean(profile.reasoning);
   const family = modelFamily(profile);
   if (family.includes("qwen") || family.includes("gemma-4") || family.includes("gemma 4")) return true;
   return undefined;
 }
 
-function modelFamily(profile) {
+export function modelFamily(profile) {
   return [profile.id, profile.label, profile.modelAlias, profile.modelPath, profile.omlxModel].filter(Boolean).join(" ").toLowerCase();
 }
 

package/src/model-presenters.mjs

@@ -159,21 +159,6 @@ export function modelSelectOption(item, { runningProfilesNow, modelMissingIds, n
       ...(hint ? { hint: pc.red(hint) } : {}),
     };
   }
-  if (item.type === "new") {
-    return {
-      value: itemKey(item),
-      label: optionLabel({
-        status: optionStatusTag("setup"),
-        backend: optionBackendTag(backendId),
-        source: optionSourceTag(sourceId),
-        name: item.label,
-        nameWidth,
-        quant: optionQuantLabel(item),
-        ctx: optionCtxLabel(item),
-        size: optionSizeLabel(item),
-      }),
-    };
-  }
   return {
     value: itemKey(item),
     label: optionLabel({

package/src/process.mjs

@@ -3,7 +3,6 @@ import { promisify } from "node:util";
 import { closeSync, openSync } from "node:fs";
 import { readFile, writeFile, chmod } from "node:fs/promises";
 import { basename, join } from "node:path";
-import { quoteShell } from "./command.mjs";
 import { LOG_DIR } from "./config.mjs";
 import { writeState, readState, profileDir } from "./profiles.mjs";
 import { backendFor, backendBinaryFor } from "./backends.mjs";
@@ -501,4 +500,9 @@ function timestampForFile() {
 
 function sleep(ms) {
   return new Promise((resolve) => setTimeout(resolve, ms));
+}
+
+function quoteShell(value) {
+  const text = String(value);
+  return /^[A-Za-z0-9_/@%+=:,.-]+$/u.test(text) ? text : `'${text.replace(/'/gu, ` '"'"'`)}'`;
 }

package/src/profile-setup.mjs

@@ -9,6 +9,7 @@ import { detectCapabilities } from "./autodetect.mjs";
 import { matchDrafter } from "./scan.mjs";
 import { scanGgufModels } from "./scan.mjs";
 import { estimateMemoryMb } from "./mlx-flags.mjs";
+import { capabilitySummary } from "./model-summary.mjs";
 
 const execFileAsync = promisify(execFile);
 
@@ -70,7 +71,7 @@ export async function configureLocalProfile(prompt, profile) {
   console.log("");
   console.log(renderSection("Model setup", renderRows([
     ["Model", pc.bold(profile.label)],
-    ["Detected", detectionSummary(caps)],
+    ["Detected", capabilitySummary(caps)],
     ["Context", `${profile.flags.ctxSize.toLocaleString()} tokens`],
     ["KV cache", `${profile.flags.cacheTypeK}/${profile.flags.cacheTypeV}`],
     ["Sampling", samplingSummary(profile.flags)],
@@ -151,16 +152,12 @@ export function applyRuntimeFlagOverrides(profile, overrides) {
 
 export function applyMtpDefaults(profile) {
   const flags = { ...profile.flags, port: LLAMA_CPP_MTP_PORT };
-  const edits = {
-    values: { "--spec-type": "draft-mtp", "--spec-draft-n-max": 4 },
-  };
-  if (profile.drafterPath) edits.values["--spec-draft-model"] = profile.drafterPath;
   return applyProfileFlags({
     ...profile,
     backend: "llama-cpp-mtp",
     providerId: "llama-cpp-mtp",
     capabilities: { ...(profile.capabilities ?? {}), mtp: true },
-  }, flags, edits);
+  }, flags);
 }
 
 export function removeMtpDefaults(profile) {
@@ -171,9 +168,7 @@ export function removeMtpDefaults(profile) {
     providerId: "llama-cpp",
     drafterPath: null,
     capabilities: { ...(profile.capabilities ?? {}), mtp: false },
-  }, flags, {
-    remove: ["--spec-type", "--spec-draft-n-max", "--spec-draft-model"],
-  });
+  }, flags);
 }
 
 function applyVisionDefaults(profile) {
@@ -201,10 +196,10 @@ function applyThinkingDefaults(profile) {
 function removeThinkingDefaults(profile) {
   const flags = { ...profile.flags, ...GENERAL_DEFAULTS };
   delete flags.chatTemplateKwargs;
-  return applyProfileFlags(profile, flags, { remove: ["--chat-template-kwargs"] });
+  return applyProfileFlags(profile, flags);
 }
 
-function applyProfileFlags(profile, flags, edits = {}) {
+function applyProfileFlags(profile, flags) {
   const next = {
     ...profile,
     flags,
@@ -214,43 +209,9 @@ function applyProfileFlags(profile, flags, edits = {}) {
       pi: { ...(profile.harnesses?.pi ?? {}), enabled: true, model: `${profile.providerId ?? profile.backend}/${profile.modelAlias ?? profile.id}` },
     },
   };
-  next.commandArgv = updateArgv(profile.commandArgv ?? [], {
-    "--host": flags.host,
-    "--port": flags.port,
-    "--ctx-size": flags.ctxSize,
-    "--cache-type-k": flags.cacheTypeK,
-    "--cache-type-v": flags.cacheTypeV,
-    "--top-k": flags.topK,
-    "--presence-penalty": flags.presencePenalty,
-    "--repeat-penalty": flags.repeatPenalty,
-    ...(flags.chatTemplateKwargs ? { "--chat-template-kwargs": JSON.stringify(flags.chatTemplateKwargs) } : {}),
-  }, edits);
-  return next;
-}
-
-function updateArgv(argv, values, edits = {}) {
-  let next = [...argv];
-  for (const flag of edits.remove ?? []) next = removeOption(next, flag);
-  for (const [flag, value] of Object.entries({ ...values, ...(edits.values ?? {}) })) {
-    if (value === undefined) continue;
-    const index = next.indexOf(flag);
-    if (index === -1) next.push(flag, String(value));
-    else next[index + 1] = String(value);
-  }
   return next;
 }
 
-function removeOption(argv, flag) {
-  const next = [];
-  for (let i = 0; i < argv.length; i++) {
-    if (argv[i] === flag) {
-      if (argv[i + 1] && !argv[i + 1].startsWith("--")) i += 1;
-      continue;
-    }
-    next.push(argv[i]);
-  }
-  return next;
-}
 
 function renderMemoryEstimate(profile) {
   try {
@@ -283,18 +244,6 @@ async function runtimeSupportsGemma4Unified() {
   }
 }
 
-function detectionSummary(caps) {
-  const parts = [];
-  if (caps.architecture) parts.push(caps.architecture);
-  if (caps.quant) parts.push(caps.quant);
-  if (caps.mtp) parts.push("MTP");
-  if (caps.qat) parts.push("QAT");
-
-  if (caps.thinking) parts.push("thinking");
-  if (caps.vision) parts.push("vision");
-  return parts.length > 0 ? parts.join(" · ") : "standard GGUF";
-}
-
 function samplingSummary(flags) {
   return `temp ${flags.temperature}, top-p ${flags.topP}, top-k ${flags.topK}`;
 }
@@ -342,7 +291,7 @@ export async function configureMlxProfile(prompt, profile) {
     ["Backend", configured.backend],
     ["Endpoint", configured.baseUrl],
     ["Context", String(configured.flags.ctxSize) + " tokens"],
-    ["Thinking", configured.capabilities.thinking && configured.commandArgv.includes("--enable-thinking") ? "on" : "off"],
+    ["Thinking", configured.capabilities?.thinking ? "on" : "off"],
     ["Vision", configured.capabilities.vision ? "yes" : "no"],
   ])));
 
@@ -352,33 +301,19 @@ export async function configureMlxProfile(prompt, profile) {
 
 async function applyMlxThinkingToggle(profile, enabled) {
   if (!profile.capabilities.thinking) return profile;
-  const { computeMlxVlmFlags } = await import("./mlx-flags.mjs");
-  const { args } = computeMlxVlmFlags(profile.modelPath, {
-    port: profile.flags.port,
-    ctxSize: profile.flags.ctxSize,
-    thinkingEnabled: enabled,
-  });
   return {
     ...profile,
-    commandArgv: args,
     capabilities: { ...profile.capabilities, thinkingEnabled: enabled },
   };
 }
 
 function applyMlxContextSize(profile, ctxSize) {
   const flags = { ...profile.flags, ctxSize };
-  const next = {
+  return {
     ...profile,
     flags,
     baseUrl: baseUrlForFlags(flags),
   };
-  const idx = next.commandArgv.indexOf("--max-kv-size");
-  if (idx !== -1 && next.commandArgv[idx + 1] != null) {
-    next.commandArgv[idx + 1] = String(ctxSize);
-  } else if (ctxSize && ctxSize > 0) {
-    next.commandArgv.push("--max-kv-size", String(ctxSize));
-  }
-  return next;
 }
 
 function renderMlxMemoryEstimate(profile) {

package/src/profiles.mjs

@@ -18,10 +18,6 @@ export function profileJsonPath(id) {
   return join(profileDir(id), "profile.json");
 }
 
-export function commandJsonPath(id) {
-  return join(profileDir(id), "command.json");
-}
-
 export function notesPath(id) {
   return join(profileDir(id), "notes.md");
 }
@@ -69,9 +65,8 @@ export async function saveProfile(profile, options = {}) {
   };
   await writeJson(profileJsonPath(id), saved);
 
-  // Note: command.json is no longer written — the server command is computed
-  // fresh from the profile config at launch time (see computeServerCommand in
-  // process.mjs).  commandArgv is kept in the profile for backwards compat.
+  // Note: commandArgv is no longer stored — computeServerCommand computes
+  // the full command fresh from profile config at launch time.
 
   if (!existsSync(notesPath(id))) {
     await writeFile(notesPath(id), `# ${saved.label}\n\nNotes for this model profile.\n`, "utf8");
@@ -134,7 +129,7 @@ export async function createProfileFromModel(model, backendId, drafterPath) {
   // If a drafter is provided, this model supports MTP regardless of filename
   const hasMtp = caps.mtp || Boolean(drafterPath);
   const backend = backendId ?? (hasMtp ? "llama-cpp-mtp" : "llama-cpp");
-  const { flags, argv } = computeFlags(
+  const { flags } = computeFlags(
     { ...caps, mtp: hasMtp },
     model.path,
     model.mmprojPath,
@@ -154,21 +149,15 @@ export async function createProfileFromModel(model, backendId, drafterPath) {
     capabilities: summarizeCapabilities({ ...caps, mtp: hasMtp }),
     preset: null, // no presets — auto-detected
     flags,
-    commandArgv: argv,
   });
 }
 
 // ── Auto-create profile from a discovered MLX model ────────────────────────
 
 export async function createProfileFromMlxModel(model) {
-  const { computeMlxVlmFlags, DEFAULT_PORT } = await import("./mlx-flags.mjs");
+  const { DEFAULT_PORT } = await import("./mlx-flags.mjs");
   const caps = await detectMlxCapabilities(model.filePath);
   const ctxSize = defaultMlxContextLength(caps.contextLength, detectHardware().totalRamBytes / (1024 ** 3));
-  const { args } = computeMlxVlmFlags(model.filePath, {
-    port: DEFAULT_PORT,
-    ctxSize,
-    thinkingEnabled: caps.thinking,
-  });
   return normalizeProfile({
     id: slugFromLabel(model.label),
     label: model.label,
@@ -182,7 +171,6 @@ export async function createProfileFromMlxModel(model) {
     modelSizeBytes: model.sizeBytes,
     capabilities: caps,
     flags: { host: "127.0.0.1", port: DEFAULT_PORT, ctxSize },
-    commandArgv: args,
   });
 }
 
@@ -203,11 +191,6 @@ function summarizeCapabilities(caps) {
 
 // ── State files (for running servers) ──────────────────────────────────────
 
-export async function readCommandArgv(profile) {
-  const command = await readJson(commandJsonPath(profile.id), null);
-  return Array.isArray(command?.argv) ? command.argv.map(String) : (profile.commandArgv ?? []).map(String);
-}
-
 export async function readState(id) {
   return readJson(statePath(id), null);
 }

package/src/command.mjs

@@ -1,21 +0,0 @@
-export function buildPrettyCommand(profile, binary = "llama-server") {
-  const argv = profile.commandArgv ?? [];
-  const lines = [`${quoteShell(binary)} \\`];
-  for (let i = 0; i < argv.length; i++) {
-    const arg = argv[i];
-    const next = argv[i + 1];
-    const hasValue = arg.startsWith("--") && next && !next.startsWith("--");
-    if (hasValue) {
-      lines.push(`  ${arg} ${quoteShell(next)}${i + 2 < argv.length ? " \\" : ""}`);
-      i += 1;
-    } else {
-      lines.push(`  ${arg}${i + 1 < argv.length ? " \\" : ""}`);
-    }
-  }
-  return lines.join("\n");
-}
-
-export function quoteShell(value) {
-  const text = String(value);
-  return /^[A-Za-z0-9_/@%+=:,.-]+$/u.test(text) ? text : `'${text.replace(/'/gu, `'"'"'`)}'`;
-}