offgrid-ai 0.9.6 → 0.10.0

package/src/model-name.mjs

@@ -68,9 +68,8 @@ const TAG_TOKENS = [
 /**
  * Parse a raw model identifier into a structured display name.
  *
- * @param {string} rawId  The raw identifier: GGUF filename (no .gguf),
- *                        Ollama model name, or oMLX model id.
- * @param {"local-gguf"|"ollama"|"omlx"} source  Where this name came from.
+ * @param {string} rawId  The raw identifier: GGUF filename (no .gguf) or oMLX model id.
+ * @param {"local-gguf"|"omlx"} source  Where this name came from.
  * @returns {{ publisher: string|null, model: string, params: string|null,
  *             quant: string|null, tags: string[], display: string,
  *             sort: string, id: string }}
@@ -87,18 +86,7 @@ export function parseModelName(rawId, source) {
     name = rawId.slice(slashIdx + 1);
   }
 
-  // 2. For Ollama, split on : to separate model from tag (e.g. "gemma3:4b")
-  //    The tag after : is a model size/variant identifier — not a GGUF quant.
-  let ollamaTag = null;
-  if (source === "ollama") {
-    const colonIdx = name.lastIndexOf(":");
-    if (colonIdx !== -1) {
-      ollamaTag = name.slice(colonIdx + 1);
-      name = name.slice(0, colonIdx);
-    }
-  }
-
-  // 3. Extract quant (GGUF quantization suffix)
+  // 2. Extract quant (GGUF quantization suffix)
   let quant = null;
   for (const pattern of QUANT_PATTERNS) {
     const match = name.match(pattern);
@@ -125,13 +113,7 @@ export function parseModelName(rawId, source) {
   // Clean up leftover separators
   name = name.replace(/[-_]{2,}/g, "-").replace(/^[-_]+|[-_]+$/g, "");
 
-  // 5. For Ollama, re-attach the tag as part of the model name
-  //    (Ollama tags like "4b" or "30b-a3b" are size variants, not quants)
-  if (ollamaTag) {
-    name = name + "-" + ollamaTag;
-  }
-
-  // 6. Title-case the remaining model name
+  // 5. Title-case the remaining model name
   let model = titleCaseModel(name);
 
   // If nothing is left after parsing, fall back to the raw name
@@ -139,13 +121,13 @@ export function parseModelName(rawId, source) {
     model = rawId.includes("/") ? rawId : rawId.replace(/[-_]/g, " ");
   }
 
-  // 7. Extract params (size like 30B, 12B) for sort/filter convenience
+  // 6. Extract params (size like 30B, 12B) for sort/filter convenience
   const params = extractParams(model);
 
-  // 8. Build display string
+  // 7. Build display string
   const display = buildDisplay(publisher, model, tags, quant);
 
-  // 9. Build sort key (lowercase, no publisher, for alphabetical ordering)
+  // 8. Build sort key (lowercase, no publisher, for alphabetical ordering)
   const sort = model.toLowerCase().replace(/[-_]/g, " ");
 
   return { publisher, model, params, quant, tags, display, sort, id };

package/src/model-presenters.mjs

@@ -1,7 +1,8 @@
 import { existsSync, statSync } from "node:fs";
-import { BACKENDS, backendFor } from "./backends.mjs";
+import { basename, dirname } from "node:path";
+import { backendFor } from "./backends.mjs";
 import { readCommandArgv } from "./profiles.mjs";
-import { isProfileRunning, isProfileServerUp } from "./process.mjs";
+import { isProfileRunning } from "./process.mjs";
 import { buildPrettyCommand } from "./command.mjs";
 import { pc, formatBytes, renderRows, renderSection } from "./ui.mjs";
 import { capabilitySummary, ggufDetailParts, isProfileFileMissing, profileDetailParts } from "./model-summary.mjs";
@@ -11,6 +12,7 @@ import { findBenchmarkRepo } from "./benchmark.mjs";
 
 const OPTION_SEPARATOR = pc.dim("  │  ");
 const OPTION_STATUS_WIDTH = 10;
+const OPTION_BACKEND_WIDTH = 14;
 const OPTION_SOURCE_WIDTH = 14;
 const OPTION_CTX_WIDTH = 5;
 
@@ -25,7 +27,7 @@ function optionPad(text, color, width) {
 function optionStatusTag(kind) {
   const statuses = {
     running: ["RUNNING", pc.green],
-    serverup: ["SERVER UP", pc.yellow],
+    serverup: ["READY", pc.blue],
     ready: ["READY", pc.blue],
     missing: ["MISSING", pc.red],
     setup: ["SETUP", pc.yellow],
@@ -34,17 +36,70 @@ function optionStatusTag(kind) {
   return optionPad(text, color, OPTION_STATUS_WIDTH);
 }
 
-function optionSourceTag(sourceId, label) {
+function optionSourceTag(sourceId) {
+  const label = formatSourceLabel(sourceId);
   const colors = {
-    "llama-cpp": pc.cyan,
-    "llama-cpp-mtp": pc.blue,
-    ollama: pc.green,
+    huggingface: pc.cyan,
+    lmstudio: pc.blue,
     omlx: pc.magenta,
+    "llama.cpp": pc.cyan,
     gguf: pc.cyan,
+    mlx: pc.yellow,
+    "mlx-vlm": pc.yellow,
   };
   return optionPad(label, colors[sourceId] ?? pc.dim, OPTION_SOURCE_WIDTH);
 }
 
+function optionBackendTag(backendId) {
+  const backend = backendId ? backendFor(backendId) : null;
+  const label = backend?.label ?? backendId ?? "unknown";
+  const colors = {
+    "llama-cpp": pc.cyan,
+    "llama-cpp-mtp": pc.blue,
+    omlx: pc.magenta,
+    "mlx-vlm": pc.yellow,
+  };
+  return optionPad(label, colors[backendId] ?? pc.dim, OPTION_BACKEND_WIDTH);
+}
+
+function formatSourceLabel(sourceId) {
+  if (!sourceId) return "unknown";
+  const map = {
+    huggingface: "HuggingFace",
+    lmstudio: "LM Studio",
+    omlx: "oMLX",
+    "llama.cpp": "llama.cpp",
+    gguf: "GGUF file",
+    mlx: "MLX",
+    "mlx-vlm": "MLX",
+  };
+  return map[sourceId] ?? String(sourceId);
+}
+
+function inferSourceFromPath(modelPath) {
+  if (!modelPath) return null;
+  const normalized = modelPath.toLowerCase().replace(/\\/g, "/");
+  if (normalized.includes("/.omlx/models")) return "omlx";
+  if (normalized.includes("/.lmstudio/models")) return "lmstudio";
+  if (normalized.includes("/.cache/huggingface")) return "huggingface";
+  if (normalized.includes("/.cache/llama.cpp")) return "llama.cpp";
+  const parent = basename(dirname(modelPath));
+  if (parent && parent !== ".") return parent.replace(/^\./, "");
+  return null;
+}
+
+function discoverySourceForProfile(profile) {
+  const backend = backendFor(profile.backend);
+  if (backend.type === "managed-server") return backend.id;
+  if (profile.source && profile.source !== "local-gguf") return profile.source;
+  return inferSourceFromPath(profile.modelPath);
+}
+
+function discoverySourceForItem(item) {
+  if (item.type === "profile") return discoverySourceForProfile(item.profile);
+  return item.model?.source ?? null;
+}
+
 function optionCtxLabel(item) {
   if (item.type === "profile" && item.profile.flags?.ctxSize) {
     return optionPad(`${(item.profile.flags.ctxSize / 1000).toFixed(0)}k`, null, OPTION_CTX_WIDTH);
@@ -52,12 +107,18 @@ function optionCtxLabel(item) {
   return optionPad("—", null, OPTION_CTX_WIDTH);
 }
 
-function optionSizeLabel(item) {
+function optionSizeLabel(item, managedModels) {
   if (item.type === "profile") {
     if (item.fileMissing) return "—";
+    if (item.profile.modelSizeBytes) return formatBytes(item.profile.modelSizeBytes);
     if (item.profile.modelPath && existsSync(item.profile.modelPath)) {
-      return formatBytes(statSync(item.profile.modelPath).size);
+      const s = statSync(item.profile.modelPath);
+      // Only stat regular files — a modelPath that is a directory (MLX)
+      // reports the dir entry size, not the model size.
+      if (s.isFile()) return formatBytes(s.size);
     }
+    const managedSize = managedProfileSizeBytes(item.profile, managedModels);
+    if (managedSize) return formatBytes(managedSize);
     return "—";
   }
   if (item.type === "new") {
@@ -76,17 +137,18 @@ export function modelNameWidth(items) {
   return Math.max(20, maxName + 2);
 }
 
-function optionLabel({ status, source, name, ctx, size, nameWidth }) {
-  return [status, source, pc.bold(optionPad(name, null, nameWidth)), ctx, pc.dim(size)].join(OPTION_SEPARATOR);
+function optionLabel({ status, backend, source, name, ctx, size, nameWidth }) {
+  return [status, backend, source, pc.bold(optionPad(name, null, nameWidth)), ctx, pc.dim(size)].join(OPTION_SEPARATOR);
 }
 
-export function modelSelectOption(item, { runningProfilesNow, serverUpIds, modelMissingIds, nameWidth }) {
+export function modelSelectOption(item, { runningProfilesNow, modelMissingIds, nameWidth, managedModels }) {
+  const sourceId = discoverySourceForItem(item) ?? "unknown";
+  const backendId = inferBackendId(item);
   if (item.type === "profile") {
     const backend = backendFor(item.profile.backend);
     const running = runningProfilesNow.some((profile) => profile.id === item.profile.id);
-    const serverUp = !running && !item.fileMissing && serverUpIds?.has(item.profile.id);
     const modelMissing = !item.fileMissing && modelMissingIds?.has(item.profile.id);
-    const status = item.fileMissing || modelMissing ? "missing" : running ? "running" : serverUp ? "serverup" : "ready";
+    const status = item.fileMissing || modelMissing ? "missing" : running ? "running" : "ready";
     const drafterMissing = Boolean(item.profile.drafterPath) && !existsSync(item.profile.drafterPath);
     const hint = drafterMissing ? "MTP drafter missing — reconfigure"
       : modelMissing ? `${backend.label} model no longer available`
@@ -95,11 +157,12 @@ export function modelSelectOption(item, { runningProfilesNow, serverUpIds, model
       value: itemKey(item),
       label: optionLabel({
         status: optionStatusTag(status),
-        source: optionSourceTag(item.profile.backend, backend.label),
+        backend: optionBackendTag(backendId),
+        source: optionSourceTag(sourceId),
         name: item.profile.label,
         nameWidth,
         ctx: optionCtxLabel(item),
-        size: optionSizeLabel(item),
+        size: optionSizeLabel(item, managedModels),
       }),
       ...(hint ? { hint: pc.red(hint) } : {}),
     };
@@ -109,41 +172,60 @@ export function modelSelectOption(item, { runningProfilesNow, serverUpIds, model
       value: itemKey(item),
       label: optionLabel({
         status: optionStatusTag("setup"),
-        source: optionSourceTag("gguf", "GGUF file"),
+        backend: optionBackendTag(backendId),
+        source: optionSourceTag(sourceId),
         name: item.model.label,
         nameWidth,
         ctx: optionCtxLabel(item),
-        size: optionSizeLabel(item),
+        size: optionSizeLabel(item, managedModels),
       }),
     };
   }
-  const backend = BACKENDS[item.backendId];
   return {
     value: itemKey(item),
     label: optionLabel({
       status: optionStatusTag("setup"),
-      source: optionSourceTag(item.backendId, backend.label),
+      backend: optionBackendTag(backendId),
+      source: optionSourceTag(sourceId),
       name: item.model.label,
       nameWidth,
       ctx: optionCtxLabel(item),
-      size: optionSizeLabel(item),
+      size: optionSizeLabel(item, managedModels),
     }),
   };
 }
 
-export function printWorkspaceHeader(normalized, runningProfilesNow, serverUpIds = new Set(), modelMissingIds = new Set()) {
+function managedProfileSizeBytes(profile, managedModels) {
+  if (!managedModels || !Array.isArray(managedModels)) return null;
+  const backend = backendFor(profile.backend);
+  if (backend.type !== "managed-server") return null;
+  const backendModels = managedModels.find((m) => m.backendId === profile.backend)?.models ?? [];
+  const modelId = profile.omlxModel ?? null;
+  if (!modelId) return null;
+  const model = backendModels.find((m) => m.id === modelId);
+  return model?.sizeBytes || null;
+}
+
+function inferBackendId(item) {
+  if (item.type === "profile") return item.profile.backend;
+  if (item.type === "managed") return item.backendId;
+  // new model: derive from format
+  if (item.model?.format === "mlx") return "mlx-vlm";
+  if (item.model?.backend) return item.model.backend;
+  return "llama-cpp";
+}
+
+export function printWorkspaceHeader(normalized, runningProfilesNow, modelMissingIds = new Set()) {
   const profiles = normalized.profiles;
   const isRunning = (p) => runningProfilesNow.some((r) => r.id === p.id);
   const isMissing = (p) => isProfileFileMissing(p) || modelMissingIds.has(p.id);
-  const readyCount = profiles.filter((p) => !isMissing(p) && !isRunning(p) && !serverUpIds.has(p.id)).length;
+  const readyCount = profiles.filter((p) => !isMissing(p) && !isRunning(p)).length;
   const runningCount = runningProfilesNow.length;
-  const serverUpCount = profiles.filter((p) => !isMissing(p) && serverUpIds.has(p.id) && !isRunning(p)).length;
   const missingCount = profiles.filter(isMissing).length;
   const setupCount = normalized.newModels.length + normalized.managedItems.length;
 
   const countParts = [];
   if (runningCount > 0) countParts.push(pc.green(`${runningCount} running`));
-  if (serverUpCount > 0) countParts.push(pc.yellow(`${serverUpCount} server up, model not loaded`));
   if (readyCount > 0) countParts.push(pc.blue(`${readyCount} model${readyCount === 1 ? "" : "s"} ready`));
   if (missingCount > 0) countParts.push(pc.red(`${missingCount} model${missingCount === 1 ? "" : "s"} missing`));
   if (setupCount > 0) countParts.push(pc.yellow(`${setupCount} model${setupCount === 1 ? "" : "s"} need${setupCount === 1 ? "s" : ""} setup`));
@@ -166,11 +248,10 @@ export async function printProfileDetails(profile) {
   const backend = backendFor(profile.backend);
   const isManaged = backend.type === "managed-server";
   const running = await isProfileRunning(profile);
-  const serverUp = !running && isManaged && await isProfileServerUp(profile);
   const fileMissing = !isManaged && isProfileFileMissing(profile);
   console.log("\n" + renderSection("Model overview", renderRows([
     ["Name", pc.bold(profile.label)],
-    ["Status", fileMissing ? pc.red("File missing") : running ? pc.green("Running now") : serverUp ? pc.yellow("Server up, model not loaded") : pc.blue("Ready")],
+    ["Status", fileMissing ? pc.red("File missing") : running ? pc.green("Running now") : pc.blue("Ready")],
     ["Details", profileDetailParts(profile, { fileMissing }).join(pc.dim(" · "))],
     ["Server", fileMissing ? pc.red(profile.baseUrl) : profile.baseUrl],
   ])));
@@ -219,6 +300,29 @@ export function printGgufModelDetails(model, drafter) {
   console.log("\n" + renderSection("Model details", renderRows(detailRows), { columns: 110 }));
 }
 
+export async function printMlxModelDetails(model) {
+  const { detectMlxCapabilities } = await import("./mlx-discovery.mjs");
+  const caps = await detectMlxCapabilities(model.filePath ?? model.path);
+  const parts = [];
+  if (caps.architecture) parts.push(caps.architecture);
+  if (caps.thinking) parts.push("thinking");
+  if (caps.vision) parts.push("vision");
+  const summary = parts.length > 0 ? parts.join(pc.dim(" · ")) : "standard MLX";
+  console.log("\n" + renderSection("Downloaded model", renderRows([
+    ["Name", pc.bold(model.label)],
+    ["Status", pc.yellow("Needs one-time setup")],
+    ["Details", summary],
+  ])));
+  console.log("\n" + renderSection("Model details", renderRows([
+    ["Model dir", model.path],
+    ["Backend", "mlx-vlm"],
+    ["Source", formatSourceLabel(model.source)],
+    ["Detected", summary],
+    ["Size", formatBytes(model.sizeBytes)],
+    ["Context", caps.contextLength ? `${caps.contextLength.toLocaleString()} trained` : "unknown"],
+  ]), { columns: 110 }));
+}
+
 export function printManagedModelDetails(model, backend) {
   console.log("\n" + renderSection(`${backend.label} model`, renderRows([
     ["Name", pc.bold(model.label)],

package/src/process.mjs

@@ -35,11 +35,13 @@ async function startLocalServer(profile) {
 
   // Build argv: binary + command.json args
   const argv = [...commandArgv];
+  // mlx-vlm requires APC_ENABLED=1 (86x TTFT improvement; fixes Metal cache clearing).
+  const env = profile.backend === "mlx-vlm" ? { ...process.env, APC_ENABLED: "1" } : process.env;
 
   const rawFd = openSync(rawLogPath, "a");
   let child;
   try {
-    child = spawn(binary, argv, { detached: true, stdio: ["ignore", rawFd, rawFd] });
+    child = spawn(binary, argv, { detached: true, stdio: ["ignore", rawFd, rawFd], env });
   } finally {
     closeSync(rawFd);
   }
@@ -96,16 +98,134 @@ export async function stopProfile(profile) {
     await writeState(profile.id, { ...state, pid: null, stoppedAt: new Date().toISOString(), stopReason: "pid-not-running" });
     return { stopped: false, message: `${profile.id} pid ${state.pid} is no longer running.` };
   }
+  const pid = state.pid;
   try {
-    try {
-      process.kill(-state.pid, "SIGTERM");
-    } catch {
-      process.kill(state.pid, "SIGTERM");
-    }
-    await writeState(profile.id, { ...state, pid: null, stoppedAt: new Date().toISOString(), stopSignal: "SIGTERM" });
-    return { stopped: true, message: `Stopped ${profile.id} pid ${state.pid}` };
+    const signal = await terminateProcess(pid);
+    await writeState(profile.id, { ...state, pid: null, stoppedAt: new Date().toISOString(), stopSignal: signal });
+    return { stopped: true, message: `Stopped ${profile.id} pid ${pid}` };
   } catch (error) {
-    return { stopped: false, message: `Could not stop pid ${state.pid}: ${error.message}` };
+    return { stopped: false, message: `Could not stop pid ${pid}: ${error.message}` };
+  }
+}
+
+// Reliably terminate a detached local-server process group: SIGTERM with a
+// grace period for graceful shutdown (lets mlx-vlm/llama-server release the
+// model), then SIGKILL if still alive. Guarantees the model is unloaded when a
+// profile stops — consistent across backends (llama-server exits on SIGTERM;
+// mlx-vlm/uvicorn often does not, hence the SIGKILL fallback).
+async function terminateProcess(pid) {
+  const signalGroup = (sig) => {
+    try { process.kill(-pid, sig); }
+    catch { process.kill(pid, sig); } // not a group leader — kill the proc itself
+  };
+  signalGroup("SIGTERM");
+  for (let i = 0; i < 50; i++) { // 5s grace for graceful shutdown
+    if (await processGone(pid)) return "SIGTERM";
+    await sleep(100);
+  }
+  signalGroup("SIGKILL");
+  for (let i = 0; i < 30; i++) { // 3s for SIGKILL to take effect
+    if (await processGone(pid)) return "SIGKILL";
+    await sleep(100);
+  }
+  throw new Error(`pid ${pid} did not exit after SIGKILL`);
+}
+
+// True if the process is dead (or a zombie about to be reaped).
+async function processGone(pid) {
+  try { process.kill(pid, 0); }
+  catch { return true; } // no such process
+  // Alive to signal(0) — but a detached setsid child can briefly appear as a
+  // zombie before launchd reaps it. Treat zombie as gone.
+  try {
+    const { stdout } = await execFileAsync("ps", ["-o", "stat=", "-p", String(pid)]);
+    return /^Z/.test(stdout.trim());
+  } catch {
+    return false;
+  }
+}
+
+// ── Unload model from a managed server (oMLX) ─────────────────────────────
+// Counterpart to stopProfile for local-server backends: stopProfile kills the
+// server process (which unloads the model); unloadModelFromServer asks a
+// managed server to release the model from memory via its HTTP API, leaving the
+// server itself running. Together they give a consistent UX: quitting Pi
+// unloads the model regardless of backend type.
+
+export async function unloadModelFromServer(profile) {
+  const backend = backendFor(profile.backend);
+
+  if (backend.id === "llama-cpp" || backend.id === "llama-cpp-mtp") {
+    // llama.cpp unloads when the server process exits; no HTTP unload API exists.
+    // If offgrid-ai started the server, stopProfile already handled it.
+    return { unloaded: false, backend: backend.id, reason: "stop server to unload" };
+  }
+
+  if (backend.id === "omlx") {
+    return await unloadOmlxModel(profile);
+  }
+
+  if (backend.id === "mlx-vlm") {
+    // mlx-vlm is a local-server backend — stopProfile handles unload by killing
+    // the process. No HTTP unload API.
+    return { unloaded: false, backend: backend.id, reason: "stop server to unload" };
+  }
+
+  return { unloaded: false, backend: backend.id, reason: "unsupported backend" };
+}
+
+async function unloadOmlxModel(profile) {
+  const baseUrl = profile.baseUrl?.replace(/\/v1\/?$/u, "") || "";
+  const adminUrl = `${baseUrl}/admin/api/models`;
+  const modelId = profile.modelAlias || profile.omlxModel || profile.id;
+
+  try {
+    const ids = await serverModelIds(profile.baseUrl);
+    const match = ids.find((id) => id.toLowerCase() === modelId.toLowerCase());
+    const targetId = match ?? modelId;
+
+    const response = await fetch(`${adminUrl}/${encodeURIComponent(targetId)}/unload`, {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      signal: AbortSignal.timeout(30000),
+    });
+
+    if (response.ok) {
+      return { unloaded: true, backend: "omlx", modelId: targetId };
+    }
+
+    const detail = await responseErrorDetail(response);
+
+    if (response.status === 400 && /not loaded/i.test(detail)) {
+      return { unloaded: true, backend: "omlx", modelId: targetId, reason: "model was not loaded" };
+    }
+
+    if (response.status === 401 || response.status === 403) {
+      return {
+        unloaded: false,
+        backend: "omlx",
+        modelId: targetId,
+        error: "oMLX admin authentication required. Enable skip_api_key_verification in oMLX settings, or unload manually from the admin panel.",
+      };
+    }
+
+    return { unloaded: false, backend: "omlx", modelId: targetId, error: `HTTP ${response.status}: ${detail}` };
+  } catch (err) {
+    if (err?.name === "AbortError" || err?.name === "TimeoutError") {
+      return { unloaded: false, backend: "omlx", modelId, error: "Unload request timed out. The model may still be unloading in the background." };
+    }
+    return { unloaded: false, backend: "omlx", modelId, error: err.message };
+  }
+}
+
+async function responseErrorDetail(response) {
+  const text = await response.text().catch(() => "");
+  if (!text) return "";
+  try {
+    const body = JSON.parse(text);
+    return body?.detail ?? body?.message ?? text;
+  } catch {
+    return text;
   }
 }
 
@@ -126,7 +246,6 @@ export async function isProfileServerUp(profile) {
 
 export async function modelLoadedOnServer(profile) {
   const backend = backendFor(profile.backend);
-  if (backend.id === "ollama") return modelIdsMatch(await ollamaLoadedModelIds(profile), expectedModelIds(profile));
   if (backend.id === "omlx") return modelIdsMatch(await omlxLoadedModelIds(profile), expectedModelIds(profile));
   const { matches } = await serverMatchesProfile(profile);
   return matches;
@@ -134,9 +253,6 @@ export async function modelLoadedOnServer(profile) {
 
 export async function modelAvailableOnServer(profile) {
   const backend = backendFor(profile.backend);
-  if (backend.id === "ollama") {
-    return modelIdsMatch(await ollamaAvailableModelIds(profile), expectedModelIds(profile));
-  }
   if (backend.id === "omlx") {
     // /v1/models lists discovered models; an ID must exist there to be usable.
     return modelIdsMatch(await serverModelIds(profile.baseUrl), expectedModelIds(profile));
@@ -217,24 +333,6 @@ export async function serverModelIds(baseUrl) {
     .filter(Boolean);
 }
 
-async function ollamaLoadedModelIds(profile) {
-  const result = await fetchJson(`${apiRootUrl(profile.baseUrl)}/api/ps`);
-  if (!result.ok) return [];
-  return (Array.isArray(result.data?.models) ? result.data.models : [])
-    .flatMap((model) => [model?.name, model?.model])
-    .map((id) => String(id ?? "").trim())
-    .filter(Boolean);
-}
-
-async function ollamaAvailableModelIds(profile) {
-  const result = await fetchJson(`${apiRootUrl(profile.baseUrl)}/api/tags`);
-  if (!result.ok) return [];
-  return (Array.isArray(result.data?.models) ? result.data.models : [])
-    .flatMap((model) => [model?.name, model?.model])
-    .map((id) => String(id ?? "").trim())
-    .filter(Boolean);
-}
-
 async function omlxLoadedModelIds(profile) {
   const statusResult = await fetchJson(`${profile.baseUrl.replace(/\/+$/u, "")}/models/status`);
   const fromStatus = statusResult.ok
@@ -305,7 +403,6 @@ function expectedModelIds(profile) {
   return [
     profile.modelAlias,
     profile.label,
-    profile.ollamaModel,
     profile.omlxModel,
     profile.modelPath,
     fileName,

package/src/profile-setup.mjs

@@ -8,6 +8,7 @@ import { pc, formatBytes, renderRows, renderSection } from "./ui.mjs";
 import { detectCapabilities } from "./autodetect.mjs";
 import { matchDrafter } from "./scan.mjs";
 import { scanGgufModels } from "./scan.mjs";
+import { estimateMemoryMb } from "./mlx-flags.mjs";
 
 const execFileAsync = promisify(execFile);
 
@@ -297,3 +298,107 @@ function detectionSummary(caps) {
 function samplingSummary(flags) {
   return `temp ${flags.temperature}, top-p ${flags.topP}, top-k ${flags.topK}`;
 }
+
+// ── MLX profile configuration ─────────────────────────────────────────────
+
+/**
+ * Interactive configuration for an mlx-vlm profile.
+ */
+export async function configureMlxProfile(prompt, profile) {
+  let configured = profile;
+
+  console.log("");
+  console.log(renderSection("Model setup", renderRows([
+    ["Model", pc.bold(profile.label)],
+    ["Detected", mlxDetectionSummary(configured.capabilities)],
+    ["Context", String(configured.flags.ctxSize) + " tokens"],
+  ])));
+  console.log(pc.dim("Larger context windows use more memory. You can edit the profile later if needed.\n"));
+
+  if (configured.capabilities.vision) {
+    console.log(renderSection("Vision detected", renderRows([
+      ["Capability", "image / multimodal input"],
+      ["Note", "mlx-vlm loads vision from the model directory automatically."],
+    ])));
+  }
+
+  if (configured.capabilities.thinking) {
+    console.log("");
+    console.log(renderSection("Thinking mode", renderRows([
+      ["Flag", "--enable-thinking"],
+      ["Default", "on for Qwen 3 / Gemma 4 / DeepSeek-R class models"],
+    ])));
+    const useThinking = await prompt.yesNo("Enable thinking mode?", true);
+    configured = await applyMlxThinkingToggle(configured, useThinking);
+  }
+
+  const ctxSize = await prompt.number("Context window tokens", configured.flags.ctxSize, 1024, 1048576);
+  configured = applyMlxContextSize(configured, ctxSize);
+
+  console.log("\n" + renderMlxMemoryEstimate(configured));
+
+  console.log("");
+  console.log(renderSection("Defaults", renderRows([
+    ["Backend", configured.backend],
+    ["Endpoint", configured.baseUrl],
+    ["Context", String(configured.flags.ctxSize) + " tokens"],
+    ["Thinking", configured.capabilities.thinking && configured.commandArgv.includes("--enable-thinking") ? "on" : "off"],
+    ["Vision", configured.capabilities.vision ? "yes" : "no"],
+  ])));
+
+  if (!(await prompt.yesNo("Save profile with these settings?", true))) return null;
+  return configured;
+}
+
+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 = {
+    ...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) {
+  const modelBytes = profile.modelSizeBytes || 0;
+  if (!modelBytes) {
+    return renderSection("Memory estimate", pc.dim("Model size unknown — save the profile to estimate."));
+  }
+  const totalMb = estimateMemoryMb(modelBytes);
+  const overheadBytes = Math.max(0, totalMb * 1024 * 1024 - modelBytes);
+  return renderSection("Memory estimate", renderRows([
+    ["Estimated total", pc.bold(`~${formatBytes(totalMb * 1024 * 1024)}`)],
+    ["Model", formatBytes(modelBytes)],
+    ["Overhead", `~${formatBytes(overheadBytes)} (KV cache, APC, runtime)`],
+  ]));
+}
+
+function mlxDetectionSummary(caps) {
+  const parts = [];
+  if (caps.architecture) parts.push(caps.architecture);
+  if (caps.thinking) parts.push("thinking");
+  if (caps.vision) parts.push("vision");
+  return parts.length > 0 ? parts.join(" · ") : "standard MLX";
+}