Cookbook UI: Ollama browser, advanced serve fold, API tokens form, diagnosis toolbar, polish

Surface a lot of accumulated cookbook + UI work as a single non-agent commit so the agent rework lands cleanly. Highlights: - Ollama as a first-class backend in the Cookbook: * Download input accepts ollama-style names (name:tag) → backend=ollama * /api/cookbook/ollama/library (cached scrape of ollama.com + curated fallback so classic models like qwen2.5 stay reachable) * "Browse Ollama library" toggle below Download with size chips * Engine=Ollama in hwfit toolbar merges the Ollama library into the main scan list as per-tag rows with the same Fit/Param/Quant/VRAM columns; click → fills Download input - API Tokens form added to Integrations panel (matching wired loadTokens()/initTokenForm() that had no HTML) - Serve panel polish: Advanced fold tightening (-8px nudges on vLLM checks, Extra args, Spec row), n_cpu_moe + Split Mode controls pulled up 8px to align with the row's checkboxes, GGUF File dropdown exposed for Ollama backend, GPU re-render on Edit serve restore, _forceBackend flag so saved serveState wins over backend detection, cookbook:servers-changed CustomEvent so panels don't need refresh - Models page redesign: Add Models row (URL + hidden API key reveal + Type select + Scan/Ollama/Key/Test/Add icon buttons), Probe All + Clear-offline buttons in Added Models toolbar, offline-pill removed (opacity already conveys state), Engine dropdown gains Ollama option - _ping_endpoint probes /v1/models then base, accepts 4xx as reachable (vLLM returns 404 on bare /v1, fully working endpoints were showing offline) - Diagnosis card: × dismiss + Copy bundle buttons restored on the serve error feedback card - Orphan tmux sweep re-enabled behind a 60s rate-limit + background Thread (off the main event loop) so dead serves get discovered - cookbook_routes auto-register watchdog: drops the endpoint if the serve session exits non-zero within the first ~3min - ollama-rocm sidecar awareness in download wrapper (`docker exec ollama-rocm ollama pull` when host ollama isn't installed) - Skill extractor sets initial_status="published" when auto_approve_skills pref is on (audit demotes later) - Skill list / model list / cookbook scan misc polish
2026-06-16 09:45:24 -04:00 · 2026-06-08 22:38:49 +09:00
parent 646f8bd2a9
commit fa8c93ec0a
28 changed files with 3033 additions and 1026 deletions
@@ -1149,6 +1149,144 @@ function initEndpointForm() {
    }
  }

+  // API Key reveal toggle. The key inputs are hidden by default so the Add
+  // form reads as a single action row; the Key button toggles the input row
+  // and flips aria-expanded for screen readers / CSS pseudo-classes.
+  const _wireKeyToggle = (btnId, rowId) => {
+    const btn = el(btnId);
+    const row = el(rowId);
+    if (!btn || !row) return;
+    btn.addEventListener('click', () => {
+      const showing = row.style.display !== 'none';
+      row.style.display = showing ? 'none' : '';
+      btn.setAttribute('aria-expanded', showing ? 'false' : 'true');
+      btn.style.opacity = showing ? '0.75' : '1';
+      if (!showing) {
+        const inp = row.querySelector('input');
+        if (inp) inp.focus();
+      }
+    });
+  };
+  _wireKeyToggle('adm-epLocalKeyBtn', 'adm-epLocalApiKey-row');
+  _wireKeyToggle('adm-epApiKeyBtn', 'adm-epApiKey-row');
+
+  // ── Added Models toolbar: Probe + Clear offline ────────────────────
+  // Both buttons act over the currently-rendered endpoint list. The
+  // online/offline marker is stamped on each row's [data-adm-ep-online]
+  // attribute by loadEndpoints(), so both buttons just iterate the DOM
+  // without re-fetching anything they don't already have.
+  const _refreshOfflineCount = () => {
+    const lbl = el('adm-epOfflineCount');
+    if (!lbl) return;
+    const n = document.querySelectorAll('[data-adm-ep-id] [data-adm-ep-online="0"]').length;
+    lbl.textContent = n > 0 ? `(${n})` : '';
+    // Keep the button enabled even when there are no offline rows — a
+    // click on the empty case fires a toast instead of feeling dead.
+    const btn = el('adm-epClearOfflineBtn');
+    if (btn) btn.style.opacity = n === 0 ? '0.55' : '0.85';
+  };
+  // Wire after every loadEndpoints() run by patching the render hook —
+  // simplest path: MutationObserver on the two list containers.
+  const _obsRoots = ['adm-epList-local', 'adm-epList-api']
+    .map(id => el(id)).filter(Boolean);
+  if (_obsRoots.length) {
+    const mo = new MutationObserver(_refreshOfflineCount);
+    _obsRoots.forEach(r => mo.observe(r, { childList: true, subtree: true }));
+    _refreshOfflineCount();
+  }
+
+  const probeAllBtn = el('adm-epProbeAllBtn');
+  if (probeAllBtn) {
+    probeAllBtn.addEventListener('click', async () => {
+      probeAllBtn.disabled = true;
+      const origHTML = probeAllBtn.innerHTML;
+      probeAllBtn.innerHTML = '<span style="opacity:0.7;">Probing…</span>';
+      try {
+        // Hit the bulk local probe (same one the model picker uses).
+        await fetch('/api/model-endpoints/probe-local', { credentials: 'same-origin' }).catch(() => {});
+        // Then per-endpoint /probe for the rest so API/cloud endpoints
+        // refresh too. Parallel — capped to 6 at a time so we don't
+        // hammer the backend on a big list.
+        const ids = Array.from(document.querySelectorAll('[data-adm-ep-id]')).map(r => r.getAttribute('data-adm-ep-id')).filter(Boolean);
+        const lane = async (id) => {
+          try { await fetch(`/api/model-endpoints/${id}/probe`, { credentials: 'same-origin' }); } catch (_) {}
+        };
+        const queue = [...ids];
+        const workers = Array.from({length: Math.min(6, queue.length)}, () => (async () => {
+          while (queue.length) {
+            const id = queue.shift();
+            if (id) await lane(id);
+          }
+        })());
+        await Promise.all(workers);
+        await loadEndpoints();
+        if (uiModule && uiModule.showToast) uiModule.showToast('Endpoint status refreshed', 1800);
+      } finally {
+        probeAllBtn.innerHTML = origHTML;
+        probeAllBtn.disabled = false;
+      }
+    });
+  }
+
+  const clearOfflineBtn = el('adm-epClearOfflineBtn');
+  if (clearOfflineBtn) {
+    clearOfflineBtn.addEventListener('click', async () => {
+      const offlineBtns = Array.from(document.querySelectorAll('[data-adm-del-ep][data-adm-ep-online="0"]'));
+      const ids = offlineBtns.map(b => b.getAttribute('data-adm-del-ep')).filter(Boolean);
+      if (!ids.length) {
+        if (uiModule && uiModule.showToast) {
+          uiModule.showToast('No offline endpoints — nothing to clear', 1800);
+        }
+        return;
+      }
+      const confirmMsg = ids.length === 1
+        ? 'Remove 1 offline endpoint?'
+        : `Remove ${ids.length} offline endpoints?`;
+      if (uiModule && uiModule.styledConfirm) {
+        const ok = await uiModule.styledConfirm(confirmMsg, { confirmText: 'Remove', danger: true });
+        if (!ok) return;
+      } else if (!confirm(confirmMsg)) {
+        return;
+      }
+      clearOfflineBtn.disabled = true;
+      // Optimistic UI: pull rows immediately, then fire the DELETEs.
+      offlineBtns.forEach(b => {
+        const row = b.closest('[data-adm-ep-id]');
+        if (row) row.remove();
+      });
+      await Promise.all(ids.map(id =>
+        fetch('/api/model-endpoints/' + id, { method: 'DELETE', credentials: 'same-origin' }).catch(() => {})
+      ));
+      try { await loadEndpoints(); } catch (_) {}
+      _refreshOfflineCount();
+      if (uiModule && uiModule.showToast) uiModule.showToast(`Removed ${ids.length} offline endpoint${ids.length === 1 ? '' : 's'}`, 1800);
+    });
+  }
+
+  // Clear-on-focus for the API key inputs. The fields are type=password so the
+  // value is masked; users can't see what's there to edit it in place, so the
+  // expected gesture is "click in, type new key". Wiping on focus removes the
+  // select-all-and-delete dance.
+  const _wireClearOnFocus = (id) => {
+    const inp = el(id);
+    if (!inp) return;
+    inp.addEventListener('focus', () => {
+      if (inp.value) inp.value = '';
+    });
+  };
+  _wireClearOnFocus('adm-epLocalApiKey');
+  _wireClearOnFocus('adm-epApiKey');
+
+  // Drop the Ollama provider logo into the Ollama Quickstart button. Reuses
+  // the same SVG the provider picker uses, so brand parity stays free.
+  try {
+    const _ollamaLogoSlot = document.querySelector('#adm-epOllamaBtn .adm-ollama-logo');
+    if (_ollamaLogoSlot) {
+      const svg = providerLogo('ollama') || '';
+      if (svg) _ollamaLogoSlot.innerHTML = svg;
+    }
+  } catch (_) {}
+
  // Local "Add" button — sibling form for self-hosted base URLs.
  const localAddBtn = el('adm-epLocalAddBtn');
  const localTestBtn = el('adm-epLocalTestBtn');
@@ -2073,17 +2211,28 @@ async function loadTokens() {
 }

 function initTokenForm() {
-  el('adm-tokenAddBtn').addEventListener('click', async () => {
+  const addBtn = el('adm-tokenAddBtn');
+  if (!addBtn || addBtn.dataset.bound) return;
+  addBtn.dataset.bound = '1';
+  addBtn.addEventListener('click', async () => {
    const msg = el('adm-tokenMsg');
    const reveal = el('adm-tokenReveal');
    msg.textContent = ''; msg.className = ''; reveal.style.display = 'none';
    const name = el('adm-tokenName').value.trim();
    if (!name) { msg.textContent = 'Token name is required'; msg.className = 'admin-error'; return; }
    const fd = new FormData(); fd.append('name', name);
+    const scopes = (el('adm-tokenScopes')?.value || '').trim();
+    if (scopes) fd.append('scopes', scopes);
    try {
      const res = await fetch('/api/tokens', { method: 'POST', body: fd, credentials: 'same-origin' });
      const data = await res.json();
-      if (res.ok) { el('adm-tokenValue').textContent = data.token; reveal.style.display = ''; el('adm-tokenName').value = ''; loadTokens(); }
+      if (res.ok) {
+        el('adm-tokenValue').textContent = data.token;
+        reveal.style.display = '';
+        el('adm-tokenName').value = '';
+        if (el('adm-tokenScopes')) el('adm-tokenScopes').value = '';
+        loadTokens();
+      }
      else { msg.textContent = data.detail || 'Failed'; msg.className = 'admin-error'; }
    } catch (e) { msg.textContent = 'Request failed'; msg.className = 'admin-error'; }
  });
@@ -2344,7 +2493,7 @@ function initDangerZone() {
   ═══════════════════════════════════════════ */
 function initAll() {
  modalEl = el('settings-modal');
-  const inits = [initSignupToggle, initAddUser, initEndpointForm, initMcpForm, initCalDAV, initBackup, initDangerZone, () => settingsModule.initIntegrations()];
+  const inits = [initSignupToggle, initAddUser, initEndpointForm, initMcpForm, initCalDAV, initBackup, initDangerZone, initTokenForm, () => settingsModule.initIntegrations()];
  for (const fn of inits) {
    try { fn(); } catch (e) { console.error('Admin init error in', fn.name || 'anonymous', e); }
  }
@@ -2357,6 +2506,7 @@ function refreshAll() {
  loadEndpoints();
  loadBuiltinTools();
  loadMcpServers();
+  loadTokens();
 }

 /* ═══════════════════════════════════════════
@@ -2118,6 +2118,28 @@ export function addMessage(role, content, modelName, metadata) {
      return lastWrap;
    }

+    // --- Wake-task / supervisor system check-in ---
+    // The self-wake mechanism injects "Did you finish?" as a user message
+    // (or persisted history shows a "[Task] Self-check: <id>" envelope)
+    // so the agent loop re-enters and re-checks status. Render as a
+    // normal user-style bubble — same chrome as a real user message,
+    // just with role "Supervisor" and a short summary body — instead of
+    // a slim system chip. Matches chat style and integrates cleanly
+    // into the conversation flow.
+    let _isWakeCheck = !!(metadata?.wake_check_in || metadata?.hidden_from_user_view);
+    if (!_isWakeCheck && typeof textRaw === 'string') {
+      // Also catch historical messages persisted as "[Task] Self-check: <sid>"
+      // (older wake tasks that didn't set wake_check_in metadata).
+      if (/^\s*\[Task\]\s+Self-check:/i.test(textRaw)) {
+        _isWakeCheck = true;
+      }
+    }
+    if (_isWakeCheck) {
+      // Supervisor self-check messages are an internal control signal —
+      // skip rendering entirely so they don't show up in the conversation.
+      return null;
+    }
+
    // --- Standard single-bubble message ---
    const wrap = document.createElement('div');
    wrap.className = 'msg ' + (role === 'user' ? 'msg-user' : 'msg-ai');
@@ -610,12 +610,47 @@ export function _showDiagnosis(panel, diagnosis, sourceText) {
    ? `Suggested action: ${fixes[0].label}.`
    : 'Suggested action: copy the error and adjust the serve settings.');

-  // Simplified diagnosis card: just the error message + suggestion + fix
-  // button(s). Removed the fold toggle, copy button, and × dismiss — they
-  // made the card noisy without earning their keep. _diagCollapsed is kept
-  // as a stub so callers don't have to change.
  panel._diagCollapsed = false;

+  // Top-right toolbar: Copy bundle + × dismiss. Restored after user feedback
+  // — without them there's no way to quietly close a stale diagnosis or grab
+  // the full error+context for a forum/discord paste.
+  const toolbar = document.createElement('div');
+  toolbar.className = 'cookbook-diag-toolbar';
+  toolbar.style.cssText = 'display:flex;justify-content:flex-end;align-items:center;gap:4px;margin-bottom:-2px;';
+
+  const copyBtn = document.createElement('button');
+  copyBtn.type = 'button';
+  copyBtn.className = 'cookbook-diag-copy';
+  copyBtn.title = 'Copy diagnosis details';
+  copyBtn.setAttribute('aria-label', 'Copy diagnosis');
+  copyBtn.innerHTML = '<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><rect x="9" y="9" width="13" height="13" rx="2" ry="2"/><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"/></svg>';
+  copyBtn.addEventListener('click', async (e) => {
+    e.stopPropagation();
+    const bundle = _diagnosisCopyBundle(task, diagnosis, sourceText, suggestionText);
+    try {
+      await navigator.clipboard.writeText(bundle);
+      copyBtn.classList.add('copied');
+      setTimeout(() => { if (copyBtn.isConnected) copyBtn.classList.remove('copied'); }, 1200);
+    } catch (_) {}
+  });
+
+  const dismissBtn = document.createElement('button');
+  dismissBtn.type = 'button';
+  dismissBtn.className = 'cookbook-diag-dismiss';
+  dismissBtn.title = 'Dismiss diagnosis';
+  dismissBtn.setAttribute('aria-label', 'Dismiss');
+  dismissBtn.textContent = '×';
+  dismissBtn.addEventListener('click', (e) => {
+    e.stopPropagation();
+    panel._diagDismissed = diagnosis.message;
+    _clearDiagnosis(panel);
+  });
+
+  toolbar.appendChild(copyBtn);
+  toolbar.appendChild(dismissBtn);
+  diag.appendChild(toolbar);
+
  const body = document.createElement('div');
  body.className = 'cookbook-diag-body';
  const msg = document.createElement('div');
@@ -416,9 +416,11 @@ function _hwfitShowError(list, host, detail) {
  if (rb) rb.addEventListener('click', () => { _resetGpuToggleState(); _hwfitFetch(true); });
 }

-// Client-side "Engine" filter (llama.cpp / vLLM / SGLang). Empty = show all.
-// Uses the same _detectBackend() the serve commands use, so what you filter to
-// is exactly what would be launched. Pure view filter — no refetch needed.
+// Client-side "Engine" filter (llama.cpp / vLLM / SGLang / Ollama). Empty =
+// show all. Uses the same _detectBackend() the serve commands use, so what you
+// filter to is exactly what would be launched. Pure view filter — no refetch
+// needed. Ollama rows are merged into the main list (see _ensureOllamaLib +
+// _ollamaToHwfitRows below) so the filter handles all engines uniformly.
 function _applyEngineFilter(models) {
  const want = document.getElementById('hwfit-engine')?.value || '';
  if (!want || !Array.isArray(models)) return models || [];
@@ -427,6 +429,86 @@ function _applyEngineFilter(models) {
  });
 }

+// Ollama library cache (per-page). Filled lazily on first _hwfitFetch; the raw
+// list is the same shape returned by /api/cookbook/ollama/library, then turned
+// into per-tag hwfit rows so they slot into the main list grid alongside HF
+// scan results.
+let _ollamaLibCache = null;
+async function _ensureOllamaLib() {
+  if (_ollamaLibCache) return _ollamaLibCache;
+  try {
+    const res = await fetch('/api/cookbook/ollama/library');
+    const data = await res.json();
+    _ollamaLibCache = Array.isArray(data?.models) ? data.models : [];
+  } catch { _ollamaLibCache = []; }
+  return _ollamaLibCache;
+}
+
+// Convert an Ollama library entry's sizes into per-tag hwfit rows. Shape
+// matches what _hwfitRenderList expects (fit_level, parameter_count,
+// required_gb, score, …) so the rows render identically to HF results.
+function _olParseSize(s) {
+  // "14b" → 14, "1.5b" → 1.5, "8x7b" → 56 (rough), "135m" → 0.135, "latest" → null
+  if (!s) return null;
+  const low = s.toLowerCase();
+  let m = low.match(/^(\d+(?:\.\d+)?)x(\d+(?:\.\d+)?)b$/);
+  if (m) return parseFloat(m[1]) * parseFloat(m[2]);
+  m = low.match(/^(\d+(?:\.\d+)?)b$/);
+  if (m) return parseFloat(m[1]);
+  m = low.match(/^(\d+(?:\.\d+)?)m$/);
+  if (m) return parseFloat(m[1]) / 1000;
+  return null;
+}
+function _ollamaToHwfitRows(libModels, vramAvail, ramAvail) {
+  const out = [];
+  if (!Array.isArray(libModels)) return out;
+  for (const m of libModels) {
+    const sizes = (Array.isArray(m.sizes) && m.sizes.length) ? m.sizes : ['latest'];
+    for (const sz of sizes) {
+      const params = _olParseSize(sz);
+      // Ollama default GGUF is ~Q4_K_M. Rough VRAM estimate: 0.6 GB / B.
+      const vramGb = params ? params * 0.6 : 0;
+      let fitLevel = 'no_fit';
+      if (vramGb && vramAvail) {
+        if (vramGb <= vramAvail * 0.6) fitLevel = 'perfect';
+        else if (vramGb <= vramAvail) fitLevel = 'good';
+        else if (ramAvail && vramGb <= ramAvail) fitLevel = 'marginal';
+        else fitLevel = 'too_tight';
+      } else if (vramGb && ramAvail && vramGb <= ramAvail) {
+        fitLevel = 'marginal';
+      }
+      const tag = `${m.name}:${sz}`;
+      const paramsLabel = params
+        ? (params >= 1 ? params.toFixed(params >= 10 ? 0 : 1) + 'B' : (params * 1000).toFixed(0) + 'M')
+        : '?';
+      // A modest score so Ollama rows still sort sensibly in the default
+      // score view — bigger models get a slightly higher base, but they
+      // always come in below well-scored HF results. Sort by Fit or VRAM
+      // to surface them more aggressively.
+      const score = params ? Math.min(30 + params * 0.3, 60) : 25;
+      out.push({
+        name: tag,
+        repo_id: tag,
+        quant: 'Q4_K_M',
+        parameter_count: paramsLabel,
+        params_b: params || 0,
+        required_gb: vramGb,
+        fit_level: fitLevel,
+        score,
+        speed_tps: 0,
+        context: 0,
+        is_gguf: true,
+        backend: 'ollama',
+        _isOllama: true,
+        _olName: m.name,
+        _olSize: sz,
+        _description: m.description || '',
+      });
+    }
+  }
+  return out;
+}
+
 export async function _hwfitFetch(fresh = false) {
  const _tk = ++_hwfitFetchToken;
  const useCase = document.getElementById('hwfit-usecase')?.value || '';
@@ -475,7 +557,12 @@ export async function _hwfitFetch(fresh = false) {
    _setLastCacheHost(remoteKey);
    const _cacheSrv = _serverByVal(_envState.remoteServerKey || remoteHost);
    const _cachePort = _cacheSrv?.port || '';
-    const _cacheParams = new URLSearchParams({ host: remoteHost }); if (_cachePort) _cacheParams.set('ssh_port', _cachePort); if (_cacheSrv?.platform) _cacheParams.set('platform', _cacheSrv.platform);
+    const _cacheParams = new URLSearchParams();
+    if (remoteHost) {
+      _cacheParams.set('host', remoteHost);
+      if (_cachePort) _cacheParams.set('ssh_port', _cachePort);
+      if (_cacheSrv?.platform) _cacheParams.set('platform', _cacheSrv.platform);
+    }
    fetch(`/api/model/cached?${_cacheParams}`, { credentials: 'same-origin' })
      .then(r => r.json())
      .then(d => {
@@ -543,7 +630,18 @@ export async function _hwfitFetch(fresh = false) {
    // A newer scan started while this one was in flight (user switched servers
    // mid-probe) — drop this stale response so it can't clobber the new one.
    if (_tk !== _hwfitFetchToken) { try { wp.destroy(); } catch {} return; }
-    if (!res.ok) throw new Error(res.statusText);
+    if (!res.ok) {
+      const body = await res.text().catch(() => '');
+      let msg = '';
+      try {
+        const payload = JSON.parse(body);
+        msg = payload && (payload.detail || payload.error || payload.message);
+      } catch {
+        msg = body;
+      }
+      msg = typeof msg === 'string' ? msg.trim() : '';
+      throw new Error(`HTTP ${res.status} ${res.statusText}${msg ? `: ${msg}` : ''}`);
+    }
    let data = await res.json();
    if (_tk !== _hwfitFetchToken) { try { wp.destroy(); } catch {} return; }
    if (!isImageMode && quantPref && !data.error && Array.isArray(data.models) && data.models.length === 0) {
@@ -583,6 +681,23 @@ export async function _hwfitFetch(fresh = false) {
      if (!_cached) { _hwfitShowError(list, remoteHost, data.error); if (hw) hw.innerHTML = ''; }
      return;
    }
+    // Merge Ollama library rows into the main list so they appear with the
+    // same Fit/Param/Quant/VRAM/Mode columns as HF results and respond to the
+    // Engine filter. Skipped in image-gen mode (Ollama doesn't serve diffusers).
+    if (!isImageMode) {
+      const _vramAvail = data.system?.gpu_vram_gb || 0;
+      const _ramAvail = data.system?.total_ram_gb || 0;
+      const _lib = await _ensureOllamaLib();
+      const _olRows = _ollamaToHwfitRows(_lib, _vramAvail, _ramAvail);
+      // Search filter on Ollama rows: HF API already filters by search; do the
+      // same client-side over Ollama name + description so the search box
+      // works consistently across both sources.
+      const _s = (search || '').trim().toLowerCase();
+      const _olFiltered = _s
+        ? _olRows.filter(r => r.name.toLowerCase().includes(_s) || (r._description || '').toLowerCase().includes(_s))
+        : _olRows;
+      data.models = (data.models || []).concat(_olFiltered);
+    }
    _hwfitCache = data;
    _hwfitRenderHw(hw, data.system);
    // Propagate local platform from hardware probe so _isWindows(task) works
@@ -964,14 +1079,36 @@ export function _hwfitRenderList(el, models) {
    html += `</div>`;
  }
  el.innerHTML = html;
-  // Click row → expand inline action panel
+  // Click row → expand inline action panel. Exception: Ollama rows skip the
+  // expand panel (no HF metadata to power it) and just fill the Download
+  // input with the `<name>:<size>` tag — one click → ready to pull.
  el.querySelectorAll('.hwfit-row:not(.hwfit-header)').forEach(row => {
    row.addEventListener('click', () => {
      const name = row.dataset.model;
      if (!name) return;
-      // Find model data from cache
      const modelData = (_hwfitCache?.models || []).find(m => m.name === name);
      if (!modelData) return;
+      if (modelData._isOllama) {
+        // Force-open the Download card if it's been collapsed — otherwise
+        // filling the (hidden) input silently swallows the click.
+        const dlBody = document.getElementById('cookbook-download-card-body');
+        const dlArrow = document.getElementById('cookbook-download-card-arrow');
+        if (dlBody && dlBody.style.display === 'none') {
+          dlBody.style.display = 'block';
+          if (dlArrow) dlArrow.style.transform = 'rotate(90deg)';
+        }
+        const dlInput = document.getElementById('cookbook-dl-repo');
+        if (dlInput) {
+          dlInput.value = modelData.name;
+          dlInput.focus();
+          // Briefly highlight so the user sees what got filled even when the
+          // download card sits far above the (long) hwfit list.
+          dlInput.classList.add('cookbook-dl-flash');
+          setTimeout(() => dlInput.classList.remove('cookbook-dl-flash'), 800);
+          dlInput.scrollIntoView({ behavior: 'smooth', block: 'center' });
+        }
+        return;
+      }
      _expandModelRow(row, modelData);
    });
  });
@@ -1297,7 +1434,7 @@ export function _hwfitInit() {
  if (sort) sort.addEventListener('change', () => _hwfitFetch());
  if (qpref) qpref.addEventListener('change', () => _hwfitFetch());
  // Engine filter is a pure client-side view filter over the already-fetched
-  // list, so just re-render from cache instead of re-probing hardware.
+  // list (HF + Ollama merged), so just re-render from cache.
  const engine = document.getElementById('hwfit-engine');
  if (engine) engine.addEventListener('change', () => {
    const list = document.getElementById('hwfit-list');
@@ -1694,6 +1831,15 @@ export function _hwfitInit() {
      saveBtn.addEventListener('click', () => {
        _syncServers();
        _rebuildServerSelect();
+        // Broadcast for anything outside the settings tab that depends on
+        // the server list (Serve dialog host picker, Running tasks, etc.).
+        // Without this the user had to hard-refresh to see the new entry
+        // in those other places.
+        try {
+          document.dispatchEvent(new CustomEvent('cookbook:servers-changed', {
+            detail: { servers: _envState.servers.slice() },
+          }));
+        } catch (_) {}
        saveBtn.classList.add('saved');
        saveBtn.innerHTML = '<svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="#50fa7b" stroke-width="2.6" stroke-linecap="round" stroke-linejoin="round" style="margin-right:4px;flex-shrink:0;"><polyline points="20 6 9 17 4 12"/></svg>Saved';
      });
@@ -1713,6 +1859,11 @@ export function _hwfitInit() {
      entry.remove();
      _syncServers();
      _rebuildServerSelect();
+      try {
+        document.dispatchEvent(new CustomEvent('cookbook:servers-changed', {
+          detail: { servers: _envState.servers.slice() },
+        }));
+      } catch (_) {}
      _hwfitCache = null;
      _hwfitFetch();
    });
@@ -72,7 +72,7 @@ function _platformIcon(platform) {
  return '';
 }

-export let _envState = { env: 'none', envPath: '', hfToken: '', hfTokenConfigured: false, hfTokenMasked: '', gpus: '', remoteHost: '', remoteServerKey: '', servers: [], modelPaths: [], platform: '', defaultServer: '' };
+export let _envState = { env: 'none', envPath: '', hfToken: '', hfTokenConfigured: false, hfTokenMasked: '', gpus: '', remoteHost: '', servers: [], modelPaths: [], platform: '', defaultServer: '' };
 let _lastCacheHostVal = null;
 let _cookbookOpeningSpinners = [];
 export function _lastCacheHost() { return _lastCacheHostVal; }
@@ -89,8 +89,8 @@ function _setCookbookOpening(on) {
  ].filter(Boolean);
  if (!on) {
    _cookbookOpeningSpinners.forEach(({ spinner, wrap, target }) => {
-      try { spinner?.stop?.(); } catch { }
-      try { wrap?.remove?.(); } catch { }
+      try { spinner?.stop?.(); } catch {}
+      try { wrap?.remove?.(); } catch {}
      target?.classList?.remove('cookbook-opening');
    });
    _cookbookOpeningSpinners = [];
@@ -114,44 +114,18 @@ function _setCookbookOpening(on) {
 // True for the local server entry (empty / "local" / "localhost" host).
 function _isLocalEntry(s) { return !s || !s.host || s.host === 'local' || s.host.toLowerCase() === 'localhost'; }

-// Resolve a dropdown option value to a server entry. New option values are
-// stable per-profile keys, so same-host SSH profiles stay distinguishable.
-// Host strings and numeric indices remain accepted for stale saved state.
-export function _serverKey(s) {
-  if (_isLocalEntry(s)) return 'local';
-  return 'srv:' + [
-    s?.name || '',
-    s?.host || '',
-    s?.port || '',
-    s?.envPath || '',
-    s?.platform || '',
-  ].map(v => encodeURIComponent(String(v).trim())).join('|');
-}
-
+// Resolve a dropdown option value to a server entry. Option values are the
+// stable HOST string ('local' for the local box) — NOT array indices — because
+// `_envState.servers` gets deduped/reordered, which made index-based selection
+// silently resolve to the wrong (or local) server. Accepts a numeric index too
+// for backwards-compat with any stale value.
 function _serverByVal(val) {
  if (val == null || val === 'local' || val === '') return null;
-  const raw = String(val);
-  let s = _envState.servers.find(x => _serverKey(x) === raw);
-  if (!s) s = _envState.servers.find(x => x.host === raw);
+  let s = _envState.servers.find(x => x.host === val);
  if (!s && /^\d+$/.test(String(val))) s = _envState.servers[parseInt(val)];
  return s || null;
 }

-export function _selectedServer() {
-  if (_envState.remoteServerKey) {
-    const keyed = _serverByVal(_envState.remoteServerKey);
-    if (keyed) return keyed;
-  }
-  if (_envState.remoteHost) return _envState.servers.find(s => s.host === _envState.remoteHost) || null;
-  return null;
-}
-
-export function _currentServerValue() {
-  const selected = _selectedServer();
-  if (selected) return _serverKey(selected);
-  return _envState.remoteHost || 'local';
-}
-
 function _buildServerOpts(excludeLocal = false) {
  // The local server is ALWAYS represented by the synthetic value="local" option
  // (showing its custom name from the "server name" feature). We must therefore
@@ -160,20 +134,13 @@ function _buildServerOpts(excludeLocal = false) {
  const _localSrv = _localIdx >= 0 ? _envState.servers[_localIdx] : null;
  const _localLabel = (_localSrv && _localSrv.name) ? _localSrv.name : 'Local';
  let html = `<option value="local"${!_envState.remoteHost ? ' selected' : ''}>${esc(_localLabel)}</option>`;
-  const selectedKey = _envState.remoteServerKey || '';
-  let legacyHostSelected = false;
  for (let i = 0; i < _envState.servers.length; i++) {
    const s = _envState.servers[i];
    if (i === _localIdx) continue;                 // already the synthetic "local" option
    if (excludeLocal && _isLocalEntry(s)) continue;
    const label = s.name || s.host || `Server ${i + 1}`;
-    const value = _serverKey(s);
-    let selected = selectedKey ? value === selectedKey : false;
-    if (!selectedKey && _envState.remoteHost === s.host && !legacyHostSelected) {
-      selected = true;
-      legacyHostSelected = true;
-    }
-    html += `<option value="${esc(value)}"${selected ? ' selected' : ''}>${esc(label)}</option>`;
+    const selected = _envState.remoteHost === s.host ? ' selected' : '';
+    html += `<option value="${esc(s.host)}"${selected}>${esc(label)}</option>`;
  }
  return html;
 }
@@ -187,41 +154,16 @@ export function _sshCmd(host, cmd, port) {
 /** Get SSH port for a given host (or task object) */
 function _getPort(hostOrTask) {
  if (!hostOrTask) return '';
-  if (typeof hostOrTask === 'object') return hostOrTask.sshPort || _getPort(hostOrTask.remoteServerKey || hostOrTask.remoteHost);
-  const selected = hostOrTask === _envState.remoteHost ? _selectedServer() : null;
-  const srv = selected || _serverByVal(hostOrTask);
+  if (typeof hostOrTask === 'object') return hostOrTask.sshPort || _getPort(hostOrTask.remoteHost);
+  const srv = _envState.servers.find(s => s.host === hostOrTask);
  return srv?.port || '';
 }

 /** Get platform for a given host (or task object). Returns 'windows', 'termux', 'linux', or '' */
 export function _getPlatform(hostOrTask) {
-  const isWinBrowser = (window.navigator.userAgent || window.navigator.platform || '').toLowerCase().includes('win');
-  // The browser's OS is NOT the server's OS when the UI is opened remotely —
-  // e.g. a Windows browser driving a Mac/Linux homeserver. Trusting the
-  // user-agent there makes the serve builder emit the Windows python-only
-  // shape (`python -m llama_cpp.server`, no `llama-server ||` fallback), which
-  // then fails on the actual Unix server. The local hardware probe is
-  // authoritative: it reports a backend (metal/cuda/rocm/cpu_*) for any Unix
-  // server and carries platform:"windows" for local Windows (which sets
-  // _envState.platform, short-circuiting below). So only fall back to the
-  // browser hint when we have no server-side signal at all.
-  const localPlatform = () => {
-    if (_envState.platform) return _envState.platform;
-    if (String(_hwfitCache?.system?.backend || '')) return '';
-    return isWinBrowser ? 'windows' : '';
-  };
-  if (!hostOrTask || hostOrTask === 'local') {
-    return localPlatform();
-  }
-  if (typeof hostOrTask === 'object') {
-    const h = hostOrTask.remoteHost;
-    if (!h || h === 'local') {
-      return hostOrTask.platform || localPlatform();
-    }
-    return hostOrTask.platform || _getPlatform(hostOrTask.remoteServerKey || h);
-  }
-  const selected = hostOrTask === _envState.remoteHost ? _selectedServer() : null;
-  const srv = selected || _serverByVal(hostOrTask);
+  if (!hostOrTask) return _envState.platform || '';
+  if (typeof hostOrTask === 'object') return hostOrTask.platform || _getPlatform(hostOrTask.remoteHost);
+  const srv = _envState.servers.find(s => s.host === hostOrTask);
  return srv?.platform || '';
 }

@@ -237,19 +179,6 @@ export function _isMetal() {
  return ['metal', 'mps', 'apple'].includes(String(_hwfitCache?.system?.backend || '').toLowerCase());
 }

-const GEMMA4_THINKING_CHAT_TEMPLATE = `{% for message in messages %}{% if message['role'] == 'system' %}<|turn>system\n<|think|>{{ message['content'] }}<turn|>\n{% elif message['role'] == 'user' %}<|turn>user\n{{ message['content'] }}<turn|>\n{% elif message['role'] == 'assistant' %}<|turn>model\n{{ message['content'] }}<turn|>\n{% endif %}{% endfor %}{% if add_generation_prompt %}<|turn>model\n<|channel>thought{% endif %}`;
-
-function _isGemma4ThinkingModel(modelName) {
-  const n = (modelName || '').toLowerCase();
-  return n.includes('gemma-4') || n.includes('gemma4');
-}
-
-function _gemma4ThinkingChatTemplateArg(modelName) {
-  return _isGemma4ThinkingModel(modelName)
-    ? _shellQuote(GEMMA4_THINKING_CHAT_TEMPLATE)
-    : '';
-}
-
 /** Detect model-specific vLLM optimizations */
 function _detectModelOptimizations(modelName) {
  const n = (modelName || '').toLowerCase();
@@ -326,7 +255,10 @@ export function _detectToolParser(modelName) {
 // ── Backend detection ──

 export function _detectBackend(model) {
-  if (model?.backend === 'ollama' || model?.is_ollama) {
+  const _ollamaName = String(model?.repo_id || model?.name || model?.id || '').trim();
+  const _ollamaMeta = `${model?.backend || ''} ${model?.endpoint_kind || ''} ${model?.provider || ''} ${model?.source || ''}`.toLowerCase();
+  const _looksLikeOllamaTag = /^[A-Za-z0-9][A-Za-z0-9._-]*(?::[A-Za-z0-9][A-Za-z0-9._-]*)$/.test(_ollamaName);
+  if (model?.backend === 'ollama' || model?.is_ollama || _ollamaMeta.includes('ollama') || _looksLikeOllamaTag) {
    return { backend: 'ollama', label: 'Ollama' };
  }
  const q = (model.quant || '').toUpperCase();
@@ -450,8 +382,6 @@ export function _buildServeCmd(f, modelName, backend) {
    const _extraEnv = (f.extra_env ?? '').toString().replace(/\s+/g, ' ').trim();
    if (_extraEnv) cmd += _extraEnv + ' ';
    cmd += `${_vllmBin} serve ${modelName} --host 0.0.0.0 --port ${f.port || '8000'}`;
-    const _gemma4ChatTemplate = _gemma4ThinkingChatTemplateArg(modelName);
-    if (_gemma4ChatTemplate) cmd += ` --chat-template ${_gemma4ChatTemplate}`;
    cmd += ` --tensor-parallel-size ${f.tp || '1'}`;
    cmd += ` --max-model-len ${f.ctx || '8192'}`;
    cmd += ` --gpu-memory-utilization ${f.gpu_mem || '0.90'}`;
@@ -482,8 +412,6 @@ export function _buildServeCmd(f, modelName, backend) {
    const _extraEnv = (f.extra_env ?? '').toString().replace(/\s+/g, ' ').trim();
    if (_extraEnv) cmd += _extraEnv + ' ';
    cmd += `${_py3Bin} -m sglang.launch_server --model-path ${modelName} --host 0.0.0.0 --port ${f.port || '30000'}`;
-    const _gemma4ChatTemplate = _gemma4ThinkingChatTemplateArg(modelName);
-    if (_gemma4ChatTemplate) cmd += ` --chat-template ${_gemma4ChatTemplate}`;
    if (f.tp && f.tp !== '1') cmd += ` --tp ${f.tp}`;
    if (f.ctx) cmd += ` --context-length ${f.ctx}`;
    if (f.gpu_mem && f.gpu_mem !== '0.90') cmd += ` --mem-fraction-static ${f.gpu_mem}`;
@@ -585,9 +513,34 @@ export function _buildServeCmd(f, modelName, backend) {
    }
  } else if (backend === 'ollama') {
    const ollamaPort = f.port || '11434';
-    const bindHost = _envState.remoteHost ? '0.0.0.0' : '127.0.0.1';
-    const hostEnv = ollamaPort !== '11434' ? `OLLAMA_HOST=${bindHost}:${ollamaPort} ` : '';
-    cmd = `${hostEnv}ollama serve`;
+    // GGUF + Ollama: delegate to the iGPU-bound ollama-test container via
+    // its /usr/local/bin/ollama-import helper. Plain `ollama serve` errors
+    // 127 on hosts where ollama isn't on PATH (and even when it is, it
+    // doesn't import the GGUF — it just starts the daemon). Args are all
+    // literal so the cookbook validator (which bans &&/||/;/$() ) is
+    // happy: `docker exec ollama-test ollama-import <repo> <name> <ctx>
+    // <file>`. The helper handles the find/Modelfile/preload dance.
+    if (modelName.includes('/') && (f.gguf_file || /-GGUF$/i.test(modelName))) {
+      // HF-GGUF repo → import + preload + tail
+      const _name = (modelName.split('/').pop() || modelName)
+        .replace(/-GGUF$/i, '')
+        .toLowerCase()
+        .replace(/[^a-z0-9._:-]+/g, '-')
+        .replace(/^-+|-+$/g, '');
+      const _ctx = f.ctx || '8192';
+      const _file = (f.gguf_file || '').split('/').pop() || '';
+      // Trailing GGUF_FILE is optional; helper picks the first match if empty.
+      cmd = `docker exec ollama-test ollama-import ${modelName} ${_name} ${_ctx}${_file ? ' ' + _file : ''}`;
+    } else if (!modelName.includes('/') && modelName) {
+      // Already-pulled Ollama tag (e.g. `qwen2.5:7b`). On kierkegaard the
+      // runtime is the ROCm Ollama sidecar; this quick command verifies the
+      // tag exists, then the backend auto-registers http://host.docker.internal:11434/v1.
+      cmd = `docker exec ollama-rocm ollama show ${modelName}`;
+    } else {
+      const bindHost = _envState.remoteHost ? '0.0.0.0' : '127.0.0.1';
+      const hostEnv = ollamaPort !== '11434' ? `OLLAMA_HOST=${bindHost}:${ollamaPort} ` : '';
+      cmd = `${hostEnv}ollama serve`;
+    }
  } else if (backend === 'diffusers') {
    const gpuStr = f.gpus?.trim();
    if (gpuStr) cmd += `CUDA_VISIBLE_DEVICES=${gpuStr} `;
@@ -630,7 +583,7 @@ function _fallbackCopy(text) {
  ta.style.cssText = 'position:fixed;left:-9999px;top:-9999px';
  document.body.appendChild(ta);
  ta.select();
-  try { document.execCommand('copy'); } catch (_) { }
+  try { document.execCommand('copy'); } catch (_) {}
  document.body.removeChild(ta);
  return Promise.resolve();
 }
@@ -663,7 +616,7 @@ function _readStoredEnvState() {

 export function _persistEnvState() {
  try { localStorage.setItem(LAST_STATE_KEY, JSON.stringify(_envStateForStorage())); }
-  catch (_) { }
+  catch (_) {}
  _saveTasks(_loadTasks());
 }

@@ -712,24 +665,22 @@ async function _fetchDependencies() {
    const data = await resp.json();
    const pkgs = data.packages || [];
    if (!pkgs.length) { list.innerHTML = '<div class="hwfit-loading">No packages found</div>'; return; }
-    const _winUnsupported = new Set(['vllm', 'rembg', 'gfpgan']);
+    const _winUnsupported = new Set(['diffusers', 'hf_transfer', 'vllm', 'rembg', 'gfpgan']);

    const _statusTag = (pkg, isLocal, isSystemDep, winBlocked) => {
      if (winBlocked) return `<span class="cookbook-dep-tag cookbook-dep-na">N/A</span>`;
-      const hasCustomInstall = !!pkg.install_cmd;
-      const hasCustomUpdate = !!pkg.update_cmd;
-      if (pkg.installed && isSystemDep && !hasCustomUpdate) return `<span class="cookbook-dep-tag cookbook-dep-installed" title="Found on selected server">Installed</span>`;
-      if (pkg.installed && pkg.pip_update_available === false && !hasCustomUpdate) {
+      if (pkg.installed && isSystemDep) return `<span class="cookbook-dep-tag cookbook-dep-installed" title="Found on selected server">Installed</span>`;
+      if (pkg.installed && pkg.pip_update_available === false) {
        const tip = esc(pkg.update_note || pkg.status_note || 'Found externally; update outside Odysseus.');
        return `<span class="cookbook-dep-tag cookbook-dep-installed" title="${tip}">Installed</span>`;
      }
      if (pkg.installed) return `<button class="cookbook-dep-tag cookbook-dep-installed cookbook-dep-installed-btn" title="Installed — click for actions"><span class="cookbook-dep-installed-label">Installed</span><span class="cookbook-dep-caret">&#9662;</span></button>`;
-      if (isSystemDep && !hasCustomInstall) {
+      if (isSystemDep) {
        const depTip = esc(pkg.install_hint || 'Install this OS package on the selected server.');
        const depLabel = pkg.applicable === false ? 'N/A ?' : 'Missing';
        return `<span class="cookbook-dep-tag cookbook-dep-na" title="${depTip}">${depLabel}</span>`;
      }
-      return `<button class="cookbook-dep-tag cookbook-dep-install" data-dep-pip="${esc(pkg.pip || '')}" data-dep-install-cmd="${esc(pkg.install_cmd || '')}" data-dep-update-cmd="${esc(pkg.update_cmd || '')}" data-dep-target="${isLocal ? 'local' : 'remote'}">Install</button>`;
+      return `<button class="cookbook-dep-tag cookbook-dep-install" data-dep-pip="${esc(pkg.pip)}" data-dep-target="${isLocal ? 'local' : 'remote'}">Install</button>`;
    };

    const _depRow = (pkg) => {
@@ -752,7 +703,7 @@ async function _fetchDependencies() {
      } else if (pkg.name === 'sglang' && pkg.installed) {
        _rebuildBtn = `<button type="button" class="cookbook-dep-tag cookbook-dep-rebuild cookbook-dep-reinstall" data-reinstall-pkg="sglang" title="Force-reinstall SGLang (pulls a matching torch). Runs as a tmux task in the Running tab.">Reinstall</button>`;
      }
-      return `<div class="cookbook-dep-row${winBlocked ? ' cookbook-dep-blocked' : ''}" data-pkg-name="${esc(pkg.name)}" data-dep-pip="${esc(pkg.pip || '')}" data-dep-install-cmd="${esc(pkg.install_cmd || '')}" data-dep-update-cmd="${esc(pkg.update_cmd || '')}" data-dep-target="${isLocal ? 'local' : 'remote'}" data-dep-kind="${esc(pkg.kind || 'python')}">`
+      return `<div class="cookbook-dep-row${winBlocked ? ' cookbook-dep-blocked' : ''}" data-pkg-name="${esc(pkg.name)}" data-dep-pip="${esc(pkg.pip || '')}" data-dep-target="${isLocal ? 'local' : 'remote'}" data-dep-kind="${esc(pkg.kind || 'python')}">`
        + `<div class="cookbook-dep-info">`
        + `<div class="memory-item-title">${esc(pkg.name)}</div>`
        + `<div class="memory-item-meta" style="font-size:10px;opacity:0.5;margin-top:2px;">${esc(pkg.desc)}</div>`
@@ -782,7 +733,7 @@ async function _fetchDependencies() {
    // Shared install/update routine — used by the Install button and the
    // "Update" item in an installed package's ⋮ menu. `upgrade` adds pip -U;
    // `statusEl`, when given, shows "Installing…/Updating…" and is disabled.
-    async function _installDep(pipName, pkgName, isLocalOnly, upgrade, statusEl, actionCmd = '') {
+    async function _installDep(pipName, pkgName, isLocalOnly, upgrade, statusEl) {
      if (isLocalOnly) {
        _envState.remoteHost = '';
        _envState.env = 'none';
@@ -827,43 +778,6 @@ async function _fetchDependencies() {
          envPrefix = 'eval "$(conda shell.bash hook)" && conda activate ' + _shellQuote(_envState.envPath);
        }
      }
-
-      if (actionCmd) {
-        const shellCmd = envPrefix ? `${envPrefix} ${actionCmd}` : actionCmd;
-        const fullCmd = (!isLocalOnly && _envState.remoteHost)
-          ? _sshCmd(_envState.remoteHost, shellCmd, _getPort(_envState.remoteHost))
-          : shellCmd;
-        try {
-          if (statusEl) { statusEl.textContent = upgrade ? 'Updating...' : 'Installing...'; statusEl.disabled = true; }
-          const res = await fetch('/api/shell/stream', {
-            method: 'POST', credentials: 'same-origin',
-            headers: { 'Content-Type': 'application/json' },
-            body: JSON.stringify({ command: fullCmd }),
-          });
-          uiModule.showToast(`${upgrade ? 'Updating' : 'Installing'} ${pkgName} on ${targetHost}...`);
-          const body = await res.text();
-          if (!res.ok) throw new Error(`HTTP ${res.status}`);
-          const exitMatches = [...body.matchAll(/"exit_code":\s*(-?\d+)/g)].map(m => Number(m[1]));
-          const exitCode = exitMatches.length ? exitMatches[exitMatches.length - 1] : 0;
-          if (exitCode !== 0) {
-            throw new Error((body.slice(-500).trim() || `${pkgName} command failed`) + ` (exit ${exitCode})`);
-          }
-
-          if (upgrade) { uiModule.showToast(`Successfully updated ${pkgName} on ${targetHost}.`); } else { uiModule.showToast(`Successfully installed ${pkgName} on ${targetHost}.`); }
-          await _fetchDependencies();
-          return;
-        } catch (err) {
-          if (statusEl) { statusEl.textContent = 'Install'; statusEl.disabled = false; }
-          uiModule.showToast(`${upgrade ? 'Update' : 'Install'} failed: ` + err.message);
-          return;
-        }
-      }
-
-      // Always go through `python -m pip` so the leading token is `python`
-      // — matches the /api/model/serve allow-list (bare `pip` is blocked).
-      // Inside a venv/conda env, `--user` is invalid (pip refuses), so we
-      // only add `--user --break-system-packages` when there's no env —
-      // for PEP-668-locked system pythons (Arch, newer Debian).
      try {
        const reqBody = {
          repo_id: pipName,
@@ -902,9 +816,8 @@ async function _fetchDependencies() {
      btn.addEventListener('click', async (e) => {
        e.stopPropagation();
        const pipName = btn.dataset.depPip;
-        const installCmd = btn.dataset.depInstallCmd || '';
        const pkgName = btn.closest('.cookbook-dep-row')?.querySelector('.memory-item-title')?.textContent || pipName;
-        await _installDep(pipName, pkgName, btn.dataset.depTarget === 'local', !!btn.dataset.upgrade, btn, installCmd);
+        await _installDep(pipName, pkgName, btn.dataset.depTarget === 'local', !!btn.dataset.upgrade, btn);
      });
    });

@@ -927,12 +840,11 @@ async function _fetchDependencies() {
      const it = document.createElement('div');
      it.className = 'dropdown-item-compact';
      it.innerHTML = `<span class="dropdown-icon">${upIco}</span><span>Update</span>`;
-      it.title = row.dataset.depUpdateCmd ? `Update ${pkgName} using its custom command` : `Update ${pkgName} to the latest version (pip install -U)`;
+      it.title = `Update ${pkgName} to the latest version (pip install -U)`;
      it.addEventListener('click', async (e) => {
        e.stopPropagation();
        dropdown.remove();
-        const updateCmd = row.dataset.depUpdateCmd || '';
-        await _installDep(pipName, pkgName, isLocalOnly, true, null, updateCmd);
+        await _installDep(pipName, pkgName, isLocalOnly, true, null);
      });
      dropdown.appendChild(it);
      document.body.appendChild(dropdown);
@@ -964,7 +876,6 @@ async function _fetchDependencies() {
 function _applyServerSelection(val) {
  if (val === 'local') {
    _envState.remoteHost = '';
-    _envState.remoteServerKey = '';
    _envState.env = 'none';
    _envState.envPath = '';
    _envState.platform = '';
@@ -972,7 +883,6 @@ function _applyServerSelection(val) {
    const s = _serverByVal(val);
    if (s) {
      _envState.remoteHost = s.host;
-      _envState.remoteServerKey = _serverKey(s);
      _envState.env = s.env || 'none';
      _envState.envPath = s.envPath || '';
      _envState.platform = s.platform || '';
@@ -983,9 +893,10 @@ function _applyServerSelection(val) {
  // bug: the Download/Cache/Deps dropdowns set the host but never saved it, so
  // it silently reverted and downloads/scans hit the wrong server).
  _persistEnvState();
-  const _want = _currentServerValue();
+  const _want = _envState.remoteHost || 'local';
  document.querySelectorAll('#hwfit-server-select, #hwfit-dl-server, #hwfit-cache-server, #hwfit-deps-server').forEach(sel => {
    if (!sel || sel.tagName !== 'SELECT') return;
+    // Option values are host strings now ('local' for the local box).
    sel.value = _want;
    // If the host isn't among this select's current options (stale options after
    // the server list changed), the browser leaves the box BLANK/grey even though
@@ -993,7 +904,7 @@ function _applyServerSelection(val) {
    // re-apply; fall back to 'local' only if it's genuinely gone.
    if (sel.selectedIndex < 0) {
      sel.innerHTML = _buildServerOpts(sel.id === 'hwfit-dl-server');
-      sel.value = _currentServerValue();
+      sel.value = _want;
      if (sel.selectedIndex < 0) sel.value = 'local';
    }
  });
@@ -1031,7 +942,7 @@ function _wireTabEvents(body) {
      // Ignore swipes that start in a horizontally-scrollable tag row — those
      // should scroll the chips, not flip the tab.
      if (window.innerWidth > 768 || e.touches.length !== 1
-        || e.target.closest('input, textarea, select, .doclib-lang-chips')) { _sx = null; return; }
+          || e.target.closest('input, textarea, select, .doclib-lang-chips')) { _sx = null; return; }
      _sx = e.touches[0].clientX; _sy = e.touches[0].clientY;
    }, { passive: true });
    body.addEventListener('touchend', (e) => {
@@ -1081,13 +992,11 @@ function _wireTabEvents(body) {
      const remotes = servers.filter(s => !_isLocalEntry(s));
      if (remotes.length === 1) {
        _envState.remoteHost = remotes[0].host;
-        _envState.remoteServerKey = _serverKey(remotes[0]);
        _envState.env = remotes[0].env || 'none';
        _envState.envPath = remotes[0].envPath || '';
      }
    }
-    const activeSrv = _selectedServer();
-    if (activeSrv) _envState.remoteServerKey = _serverKey(activeSrv);
+    const activeSrv = servers.find(s => s.host === _envState.remoteHost);
    _envState.platform = activeSrv?.platform || '';
    localStorage.setItem('cookbook-last-state', JSON.stringify(_envStateForStorage()));
    _saveTasks(_loadTasks());
@@ -1095,7 +1004,7 @@ function _wireTabEvents(body) {
    // UI matches the resolved host. Done in a microtask so the dropdowns
    // exist by the time we set their .value.
    Promise.resolve().then(() => {
-      const _want = _currentServerValue();
+      const _want = _envState.remoteHost || 'local';
      document.querySelectorAll('#hwfit-server-select, #hwfit-dl-server, #hwfit-cache-server, #hwfit-deps-server').forEach(sel => {
        if (sel && sel.tagName === 'SELECT') sel.value = _want;
      });
@@ -1361,14 +1270,28 @@ function _wireTabEvents(body) {
      if (!m) return { repo: raw, include: null };
      return { repo: m[1], include: `*${m[2]}*` };
    }
+    // Ollama-library name. Matches `qwen2.5:14b`, `llama3:latest`, and the
+    // (rare) `library/<name>:<tag>` form which we normalize by stripping the
+    // namespace. The backend's _is_ollama_download check expects the same
+    // shape (no slash + has a colon).
+    function _ollamaName(raw) {
+      const stripped = raw.replace(/^library\//, '');
+      if (/^[A-Za-z0-9][A-Za-z0-9._-]{0,200}:[A-Za-z0-9][A-Za-z0-9._-]{0,200}$/.test(stripped)) {
+        return stripped;
+      }
+      return null;
+    }
    const triggerDownload = () => {
      const rawRepo = _stripHfUrl(dlInput.value);
      if (!rawRepo) return;
-      const { repo, include: autoInclude } = _splitRepoTag(rawRepo);
+      const ollamaName = _ollamaName(rawRepo);
+      const { repo, include: autoInclude } = ollamaName ? { repo: ollamaName, include: null } : _splitRepoTag(rawRepo);
      // HuggingFace repo IDs must be `org/model`. A bare model name would 404
      // at snapshot_download time with a raw traceback, so reject it up front.
-      if (!/^[^\s/]+\/[^\s/]+$/.test(repo)) {
-        uiModule.showToast('Enter a full HuggingFace repo ID like "org/model-name" (or paste the full HF URL).');
+      // Ollama names (single-segment with a tag) skip this check — they go
+      // through `ollama pull` server-side, not snapshot_download.
+      if (!ollamaName && !/^[^\s/]+\/[^\s/]+$/.test(repo)) {
+        uiModule.showToast('Enter a full HuggingFace repo ID like "org/model-name", or an Ollama name like "qwen2.5:14b".');
        dlInput.focus();
        return;
      }
@@ -1383,12 +1306,13 @@ function _wireTabEvents(body) {
      if (srvVal !== 'local') {
        host = _serverByVal(srvVal)?.host || '';
      }
-      const _hsrv = srvVal !== 'local' ? (_serverByVal(srvVal) || {}) : {};
+      const _hsrv = _envState.servers.find(sv => sv.host === host) || {};
      let env = host ? (_hsrv.env || 'none') : _envState.env;
      let envPath = host ? (_hsrv.envPath || '') : _envState.envPath;
      const payload = { repo_id: repo };
+      if (ollamaName) payload.backend = 'ollama';
      if (autoInclude) payload.include = autoInclude;
-      if (_envState.hfToken) payload.hf_token = _envState.hfToken;
+      if (_envState.hfToken && !ollamaName) payload.hf_token = _envState.hfToken;
      if (host) { payload.remote_host = host; const _sp3 = _getPort(host); if (_sp3) payload.ssh_port = _sp3; }
      const srvPlatform = _getPlatform(host);
      if (srvPlatform) payload.platform = srvPlatform;
@@ -1432,7 +1356,7 @@ function _wireTabEvents(body) {
      // the section is collapsed (the body's content normally provides
      // separation; with no body visible, the line gives the h2 definition).
      dlFold.classList.toggle('is-folded', !folded);
-      try { localStorage.setItem('cookbook_dl_tab_folded_v1', folded ? '0' : '1'); } catch { }
+      try { localStorage.setItem('cookbook_dl_tab_folded_v1', folded ? '0' : '1'); } catch {}
    });
  }
  const hfToggle = document.getElementById('cookbook-hf-latest-toggle');
@@ -1478,7 +1402,7 @@ function _wireTabEvents(body) {
          _hwCache[cacheKey] = hw;
          return hw;
        }
-      } catch { }
+      } catch {}
      _hwCache[cacheKey] = { vram: 0, backend: '' };
      return _hwCache[cacheKey];
    }
@@ -1591,6 +1515,84 @@ function _wireTabEvents(body) {
    document.getElementById('hwfit-server-select')?.addEventListener('change', _onServerChange);
  }

+  // Browse Ollama library — popular models from ollama.com via cached backend
+  // proxy. Click a row → fills the download input with `<name>:<size>` so the
+  // existing Download button kicks off `ollama pull`.
+  const olToggle = document.getElementById('cookbook-ollama-toggle');
+  const olArrow = document.getElementById('cookbook-ollama-arrow');
+  const olList = document.getElementById('cookbook-ollama-list');
+  const olRefresh = document.getElementById('cookbook-ollama-refresh');
+  if (olToggle && olList) {
+    let _olLoaded = false;
+    async function _loadOllama(refresh = false) {
+      olList.innerHTML = '<div class="hwfit-loading" style="opacity:0.5;font-size:11px;text-align:center;padding:12px;">Loading…</div>';
+      try {
+        const res = await fetch(`/api/cookbook/ollama/library${refresh ? '?refresh=1' : ''}`);
+        const data = await res.json();
+        const models = data.models || [];
+        if (!models.length) {
+          olList.innerHTML = '<div class="hwfit-loading">No models</div>';
+          return;
+        }
+        let html = '';
+        for (const m of models) {
+          const sizes = Array.isArray(m.sizes) && m.sizes.length ? m.sizes : ['latest'];
+          const sizeChips = sizes.map(s => `<button type="button" class="memory-toolbar-btn cookbook-ol-size" data-name="${esc(m.name)}" data-size="${esc(s)}" style="height:20px;padding:0 6px;font-size:10px;border-radius:3px;">${esc(s)}</button>`).join('');
+          html += `<div class="doclib-card memory-item cookbook-ollama-card" data-name="${esc(m.name)}">`;
+          html += `<div style="flex:1;min-width:0;">`;
+          html += `<div class="memory-item-title">${esc(m.name)} <a href="https://ollama.com/library/${esc(m.name)}" target="_blank" rel="noopener" class="cookbook-hf-link">ollama ↗</a></div>`;
+          if (m.description) html += `<div class="memory-item-meta" style="font-size:10px;opacity:0.55;margin-top:2px;">${esc(m.description)}</div>`;
+          html += `<div style="display:flex;flex-wrap:wrap;gap:3px;margin-top:4px;">${sizeChips}</div>`;
+          html += `</div></div>`;
+        }
+        olList.innerHTML = html;
+        olList.querySelectorAll('.cookbook-ol-size').forEach(btn => {
+          btn.addEventListener('click', (e) => {
+            e.stopPropagation();
+            const name = btn.dataset.name;
+            const size = btn.dataset.size;
+            if (dlInput) {
+              dlInput.value = `${name}:${size}`;
+              dlInput.focus();
+            }
+          });
+        });
+        // Clicking the card body (not a size chip / link) → default to first size
+        olList.querySelectorAll('.cookbook-ollama-card').forEach(card => {
+          card.addEventListener('click', (e) => {
+            if (e.target.closest('a') || e.target.closest('.cookbook-ol-size')) return;
+            const name = card.dataset.name;
+            const firstSize = card.querySelector('.cookbook-ol-size')?.dataset.size || 'latest';
+            if (dlInput) {
+              dlInput.value = `${name}:${firstSize}`;
+              dlInput.focus();
+            }
+          });
+        });
+      } catch (e) {
+        olList.innerHTML = '<div class="hwfit-loading">Failed to load</div>';
+      }
+    }
+    olToggle.addEventListener('click', () => {
+      const isOpen = olList.style.display !== 'none';
+      olList.style.display = isOpen ? 'none' : 'flex';
+      if (olArrow) olArrow.style.transform = isOpen ? 'rotate(0deg)' : 'rotate(90deg)';
+      if (!isOpen && !_olLoaded) {
+        _olLoaded = true;
+        _loadOllama(false);
+      }
+    });
+    if (olRefresh) olRefresh.addEventListener('click', (e) => {
+      e.stopPropagation();
+      _olLoaded = true;
+      _loadOllama(true);
+      if (olList.style.display === 'none') {
+        olList.style.display = 'flex';
+        if (olArrow) olArrow.style.transform = 'rotate(90deg)';
+      }
+    });
+  }
+
  // Server add button, row removal, model-dir add/remove, and per-row wiring
  // are ALL owned by cookbook-hwfit.js's _hwfitInit / _wireServerEntry.
  // A duplicate add handler used to live here and fired alongside the hwfit
@@ -1603,7 +1605,7 @@ function _wireTabEvents(body) {
    hfInput.addEventListener('change', async () => {
      const val = hfInput.value.trim();
      _envState.hfToken = val;
-      try { await _persistEnvState(); } catch { }
+      try { await _persistEnvState(); } catch {}
      if (val) {
        _envState.hfTokenConfigured = true;
        const masked = val.length > 6 ? val.slice(0, 3) + '…' + val.slice(-3) : '••••';
@@ -1643,9 +1645,8 @@ export function _serverEntryHtml(s, i, defaultServer, forceRemote, isNew) {
  let html = '';
  html += `<div class="cookbook-server-entry" data-idx="${i}" data-platform="${esc(s.platform || '')}">`;
  const _srvTitle = s.name || (isLocal ? 'Local' : (s.host || `Server ${i + 1}`));
-  const _srvKey = isLocal ? 'local' : _serverKey(s);
-  const _legacyDefault = !String(defaultServer || '').startsWith('srv:') && !isLocal && (defaultServer || '') === (s.host || '');
-  const _isDefaultSrv = (defaultServer || '') === _srvKey || _legacyDefault;
+  const _srvKey = isLocal ? 'local' : (s.host || '');
+  const _isDefaultSrv = (defaultServer || '') === _srvKey;
  const _pIco = _platformIcon(s.platform);
  const _keyBtn = `<button class="cookbook-server-key-btn" title="Set up SSH key for this server" style="height:22px;box-sizing:border-box;display:inline-flex;align-items:center;position:relative;top:-2px;"><svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="margin-right:4px;flex-shrink:0;"><circle cx="7.5" cy="15.5" r="5.5"/><path d="M12 11l8-8"/><path d="M17 6l3 3"/></svg>Key</button>`;
  const _checkBtn = `<button class="cookbook-server-check-btn" title="Check SSH connection" style="height:22px;box-sizing:border-box;display:inline-flex;align-items:center;position:relative;top:-2px;"><svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.2" stroke-linecap="round" stroke-linejoin="round" style="margin-right:4px;flex-shrink:0;"><polyline points="20 6 9 17 4 12"/></svg>Check</button>`;
@@ -1775,11 +1776,24 @@ function _renderRecipes() {
  html += `<button class="memory-toolbar-btn cookbook-dl-add-server" title="Add server in Settings" style="height:28px;">add server</button>`;
  html += `</div>`;
  html += `<div class="cookbook-dl-input" style="margin-top:0;">`;
-  html += `<input type="text" class="cookbook-dl-repo" id="cookbook-dl-repo" placeholder="org/model-name, HF URL, or org/model:QUANT_TAG" />`;
+  html += `<input type="text" class="cookbook-dl-repo" id="cookbook-dl-repo" placeholder="org/model-name, qwen2.5:14b, or HF URL" />`;
  html += `<button class="cookbook-btn cookbook-dl-btn" id="cookbook-dl-btn">Download</button>`;
  html += `</div>`;
+  // Browse Ollama library — fetches popular models from ollama.com via the
+  // /api/cookbook/ollama/library cached proxy, click → fills the input with
+  // `<name>:<size>` so the existing Download button kicks off `ollama pull`.
+  html += `<div style="margin-top:5px;position:relative;top:-3px;">`;
+  html += `<div style="display:flex;gap:4px;align-items:center;">`;
+  html += `<button type="button" class="memory-toolbar-btn" id="cookbook-ollama-toggle" style="flex:1;text-align:left;height:26px;display:flex;align-items:center;gap:6px;border-radius:4px;">`;
+  html += `<span id="cookbook-ollama-arrow" style="display:inline-block;transition:transform 0.15s;pointer-events:none;">▸</span>`;
+  html += `<span style="pointer-events:none;">Browse Ollama library</span>`;
+  html += `</button>`;
+  html += `<button type="button" class="memory-toolbar-btn" id="cookbook-ollama-refresh" title="Refresh" style="height:26px;width:26px;padding:0;border-radius:4px;">↻</button>`;
+  html += `</div>`;
+  html += `<div id="cookbook-ollama-list" style="display:none;margin-top:4px;max-height:320px;overflow-y:auto;flex-direction:column;gap:4px;"></div>`;
+  html += `</div>`;
  // Latest HF models that fit — collapsible card list
-  html += `<div style="margin-top:5px;position:relative;top:-7px;">`;
+  html += `<div style="margin-top:5px;position:relative;top:-3px;">`;
  html += `<div style="display:flex;gap:4px;align-items:center;">`;
  html += `<button type="button" class="memory-toolbar-btn" id="cookbook-hf-latest-toggle" style="flex:1;text-align:left;height:26px;display:flex;align-items:center;gap:6px;border-radius:4px;">`;
  html += `<span id="cookbook-hf-latest-arrow" style="display:inline-block;transition:transform 0.15s;pointer-events:none;">\u25B8</span>`;
@@ -1804,7 +1818,7 @@ function _renderRecipes() {
  html += '<option value="general" selected>Standard</option><option value="coding">Coding</option>';
  html += '<option value="reasoning">Reasoning</option><option value="chat">Chat</option>';
  // Image tab removed — text→image gen is gone from this build (only inpaint
-  // remains, which uses its own settings panel). Vision (multimodal) stays.
+   // remains, which uses its own settings panel). Vision (multimodal) stays.
  html += '<option value="multimodal">Vision</option></select>';
  // Engine sits next to the type filter so the "what category / which serving
  // path" filters live together; Quant + Context are storage-format and budget
@@ -1813,6 +1827,7 @@ function _renderRecipes() {
  html += '<select class="cookbook-field-input hwfit-engine" id="hwfit-engine" style="height:28px;" title="Filter by serving engine">';
  html += '<option value="">Engine</option>';
  html += '<option value="llamacpp">llama.cpp</option>';
+  html += '<option value="ollama">Ollama</option>';
  html += '<option value="vllm">vLLM</option>';
  html += '<option value="sglang">SGLang</option>';
  html += '</select>';
@@ -1869,13 +1884,13 @@ function _renderRecipes() {
  // Footer: link to the public discussion where users can request additions
  // to the curated model list. Sits below the list so it reads as a callout
  // after browsing, not a header.
-  html += '<div class="hwfit-list-footer" style="margin-top:8px;padding-top:6px;border-top:1px solid color-mix(in srgb, var(--border) 50%, transparent);font-size:9.5px;opacity:0.65;text-align:right;">'
-    + 'Don\'t see a model? '
-    + '<a href="https://github.com/pewdiepie-archdaemon/odysseus/discussions/1962" target="_blank" rel="noopener" style="color:var(--accent,var(--red));text-decoration:none;display:inline-flex;align-items:center;gap:4px;vertical-align:middle;">'
-    + 'Request it →'
-    + '<svg width="11" height="11" viewBox="0 0 16 16" fill="currentColor" aria-hidden="true" style="flex-shrink:0;"><path d="M8 0C3.58 0 0 3.58 0 8a8 8 0 0 0 5.47 7.59c.4.07.55-.17.55-.38 0-.19-.01-.82-.01-1.49-2.01.37-2.53-.49-2.69-.94-.09-.23-.48-.94-.82-1.13-.28-.15-.68-.52-.01-.53.63-.01 1.08.58 1.23.82.72 1.21 1.87.87 2.33.66.07-.52.28-.87.51-1.07-1.78-.2-3.64-.89-3.64-3.95 0-.87.31-1.59.82-2.15-.08-.2-.36-1.02.08-2.12 0 0 .67-.21 2.2.82.64-.18 1.32-.27 2-.27.68 0 1.36.09 2 .27 1.53-1.04 2.2-.82 2.2-.82.44 1.1.16 1.92.08 2.12.51.56.82 1.27.82 2.15 0 3.07-1.87 3.75-3.65 3.95.29.25.54.73.54 1.48 0 1.07-.01 1.93-.01 2.2 0 .21.15.46.55.38A8.013 8.013 0 0 0 16 8c0-4.42-3.58-8-8-8z"/></svg>'
-    + '</a>'
-    + '</div>';
+  html += '<div class="hwfit-list-footer" style="display:none;">'
+       + 'Don\'t see a model? '
+       + '<a href="https://github.com/pewdiepie-archdaemon/odysseus/discussions/1962" target="_blank" rel="noopener" style="color:var(--accent,var(--red));text-decoration:none;display:inline-flex;align-items:center;gap:4px;vertical-align:middle;position:relative;top:-1px;">'
+       + 'Request it →'
+       + '<svg width="11" height="11" viewBox="0 0 16 16" fill="currentColor" aria-hidden="true" style="flex-shrink:0;"><path d="M8 0C3.58 0 0 3.58 0 8a8 8 0 0 0 5.47 7.59c.4.07.55-.17.55-.38 0-.19-.01-.82-.01-1.49-2.01.37-2.53-.49-2.69-.94-.09-.23-.48-.94-.82-1.13-.28-.15-.68-.52-.01-.53.63-.01 1.08.58 1.23.82.72 1.21 1.87.87 2.33.66.07-.52.28-.87.51-1.07-1.78-.2-3.64-.89-3.64-3.95 0-.87.31-1.59.82-2.15-.08-.2-.36-1.02.08-2.12 0 0 .67-.21 2.2.82.64-.18 1.32-.27 2-.27.68 0 1.36.09 2 .27 1.53-1.04 2.2-.82 2.2-.82.44 1.1.16 1.92.08 2.12.51.56.82 1.27.82 2.15 0 3.07-1.87 3.75-3.65 3.95.29.25.54.73.54 1.48 0 1.07-.01 1.93-.01 2.2 0 .21.15.46.55.38A8.013 8.013 0 0 0 16 8c0-4.42-3.58-8-8-8z"/></svg>'
+       + '</a>'
+       + '</div>';

  html += '</div></div>';

@@ -1885,7 +1900,7 @@ function _renderRecipes() {
  html += '<div style="display:flex;align-items:baseline;gap:8px;margin-bottom:2px;">';
  html += '<h2 style="margin:0;padding:0;line-height:1;">Serve <span id="serve-stats" class="memory-count" style="font-size:0.6em;opacity:0.6;font-weight:normal"></span></h2>';
  html += '</div>';
-  const _selSrv = _selectedServer() || _es.servers[0] || {};
+  const _selSrv = _es.servers.find(s => s.host === _es.remoteHost) || _es.servers[0] || {};
  const _srvDirs = (Array.isArray(_selSrv.modelDirs) ? _selSrv.modelDirs : [_selSrv.modelDir || '~/.cache/huggingface/hub']).map(d => d.replaceAll('✕', '').replaceAll('✖', '').trim()).filter(Boolean);
  html += '<div class="cookbook-serve-dirs" style="margin-top:6px;">';
  html += _srvDirs.map(d => `<span class="cookbook-serve-dir-pill">${esc(d)}</span>`).join('');
@@ -1909,7 +1924,7 @@ function _renderRecipes() {
  html += '<label class="memory-bulk-check-all"><input type="checkbox" id="serve-select-all"> All</label>';
  html += '<span id="serve-bulk-count" style="font-size:10px;opacity:0.5;">0 selected</span>';
  html += '<button class="memory-toolbar-btn danger" id="serve-bulk-delete" style="position:relative;top:-3px;"><svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="vertical-align:-1px;margin-right:3px;"><polyline points="3 6 5 6 21 6"/><path d="M19 6l-1 14a2 2 0 0 1-2 2H8a2 2 0 0 1-2-2L5 6"/><path d="M10 11v6"/><path d="M14 11v6"/></svg>Delete</button>';
-  html += '<button class="memory-toolbar-btn" id="serve-bulk-cancel" title="Cancel (Esc)" style="margin-left:4px;padding:3px 6px;position:relative;top:-3px;"><svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round"><line x1="18" y1="6" x2="6" y2="18"/><line x1="6" y1="6" x2="18" y2="18"/></svg></button>';
+  html += '<button class="memory-toolbar-btn" id="serve-bulk-cancel" title="Cancel (Esc)" style="margin-left:4px;padding:3px 6px;position:relative;top:-7px;"><svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round"><line x1="18" y1="6" x2="6" y2="18"/><line x1="6" y1="6" x2="18" y2="18"/></svg></button>';
  html += '</div>';

  html += '<div class="doclib-grid hwfit-cached-list" id="hwfit-cached-list"></div>';
@@ -1963,7 +1978,7 @@ function _renderRecipes() {
  html += '<div style="display:flex;align-items:baseline;gap:8px;margin-bottom:2px;margin-top:-4px;">';
  html += '<h2 style="margin:0;padding:0;line-height:1;">Servers</h2>';
  // Reuse the calendar +New pill: spinning plus, label fades in idea uses
-  // the same `.cal-add-btn-text` rules, so styling stays consistent.
+   // the same `.cal-add-btn-text` rules, so styling stays consistent.
  html += '<button class="cal-add-btn cal-add-btn-text" id="cookbook-server-add" title="Add server" style="margin-left:auto;"><span class="cal-add-plus">+</span><span class="cal-add-label">Add</span></button>';
  html += '</div>';
  html += '<p class="memory-desc doclib-desc">Configure SSH servers, install Odysseus keys, choose model directories, and set the default server. Local is this machine.</p>';
@@ -2059,73 +2074,73 @@ export async function open(opts) {
  }
  _setCookbookOpening(true);
  try {
-    // Invalidate any pending close() animation handlers so they won't re-hide us
-    _closeGen++;
-    // Clear any leftover inline styles from a previous swipe-dismiss or close animation
-    const _content = modal.querySelector('.modal-content');
-    if (_content) {
-      _content.classList.remove('modal-closing', 'sheet-ready', 'cookbook-modal-entering');
-      _content.style.transform = '';
-      _content.style.transition = '';
-      _content.style.animation = '';
-      _content.style.opacity = '';
+  // Invalidate any pending close() animation handlers so they won't re-hide us
+  _closeGen++;
+  // Clear any leftover inline styles from a previous swipe-dismiss or close animation
+  const _content = modal.querySelector('.modal-content');
+  if (_content) {
+    _content.classList.remove('modal-closing', 'sheet-ready', 'cookbook-modal-entering');
+    _content.style.transform = '';
+    _content.style.transition = '';
+    _content.style.animation = '';
+    _content.style.opacity = '';
+  }
+  modal.style.display = '';
+  Modals.register('cookbook-modal', {
+    railBtnId: 'rail-cookbook',
+    sidebarBtnId: 'tool-cookbook-btn',
+    closeFn: () => _doClose(),
+    restoreFn: () => { _renderRunningTab(); },
+  });
+  _wireCookbookDrag(modal);
+  await _syncFromServer();
+  // `_syncFromServer` lives in cookbookRunning.js and populates *its* _envState
+  // (a different object reference than this module's), then mirrors the merged
+  // state to localStorage. So ALWAYS hydrate our _envState from that mirror —
+  // on a successful sync it holds the freshly-fetched servers; on failure it
+  // holds the last-known state. Gating this on `!synced` left the render's
+  // _envState empty whenever sync succeeded → "servers don't show".
+  try { Object.assign(_envState, _readStoredEnvState()); } catch {}
+  // Honour a user-set default server: always land on it when Cookbook opens, so
+  // every dropdown (scan/download/serve/cache/deps) starts on the same machine.
+  if (_envState.defaultServer) {
+    const _dk = _envState.defaultServer;
+    if (_dk === 'local') {
+      _envState.remoteHost = ''; _envState.env = 'none'; _envState.envPath = ''; _envState.platform = '';
+    } else {
+      const _ds = (_envState.servers || []).find(s => s.host === _dk);
+      if (_ds) { _envState.remoteHost = _ds.host; _envState.env = _ds.env || 'none'; _envState.envPath = _ds.envPath || ''; _envState.platform = _ds.platform || ''; }
    }
-    modal.style.display = '';
-    Modals.register('cookbook-modal', {
-      railBtnId: 'rail-cookbook',
-      sidebarBtnId: 'tool-cookbook-btn',
-      closeFn: () => _doClose(),
-      restoreFn: () => { _renderRunningTab(); },
-    });
-    _wireCookbookDrag(modal);
-    await _syncFromServer();
-    // `_syncFromServer` lives in cookbookRunning.js and populates *its* _envState
-    // (a different object reference than this module's), then mirrors the merged
-    // state to localStorage. So ALWAYS hydrate our _envState from that mirror —
-    // on a successful sync it holds the freshly-fetched servers; on failure it
-    // holds the last-known state. Gating this on `!synced` left the render's
-    // _envState empty whenever sync succeeded → "servers don't show".
-    try { Object.assign(_envState, _readStoredEnvState()); } catch { }
-    // Honour a user-set default server: always land on it when Cookbook opens, so
-    // every dropdown (scan/download/serve/cache/deps) starts on the same machine.
-    if (_envState.defaultServer) {
-      const _dk = _envState.defaultServer;
-      if (_dk === 'local') {
-        _envState.remoteHost = ''; _envState.remoteServerKey = ''; _envState.env = 'none'; _envState.envPath = ''; _envState.platform = '';
-      } else {
-        const _ds = _serverByVal(_dk);
-        if (_ds) { _envState.remoteHost = _ds.host; _envState.remoteServerKey = _serverKey(_ds); _envState.env = _ds.env || 'none'; _envState.envPath = _ds.envPath || ''; _envState.platform = _ds.platform || ''; }
-      }
-    }
-    // Re-render on every open AFTER sync so the freshly-fetched state (servers,
-    // HF token, presets) is always reflected. Gating this to once-per-page used
-    // to freeze a stale/empty servers list whenever the first sync raced or
-    // returned before hydration — and since close/reopen doesn't reset the page,
-    // only a full reload recovered it. Re-rendering is cheap and the in-progress
-    // Running tab is rendered separately just below.
-    _renderRecipes();
-    _rendered = true;
-    _clearCookbookNotif();
-    _renderRunningTab();
-    // Self-heal: revive any download tasks whose tmux session is still alive
-    // but were persisted as done/error (covers the "restarted server while a
-    // big multi-shard download was in flight" case — the task survived in
-    // tmux, the cookbook just lost track of it).
-    try { _selfHealStaleTasks({ oneShot: true }); } catch { }
-    if (_content) {
-      // Put the panel in its entering state before it becomes visible. On
-      // mobile, showing first and adding the class a frame later can paint the
-      // sheet at its final position, which makes the slide-up look like a snap.
-      _content.classList.add('cookbook-modal-entering');
-    }
-    modal.classList.remove('hidden');
-    if (_content) {
-      void _content.offsetWidth;
-      _content.addEventListener('animationend', () => {
-        _content.classList.remove('cookbook-modal-entering');
-      }, { once: true });
-    }
-    setTimeout(_applyIntent, 0);
+  }
+  // Re-render on every open AFTER sync so the freshly-fetched state (servers,
+  // HF token, presets) is always reflected. Gating this to once-per-page used
+  // to freeze a stale/empty servers list whenever the first sync raced or
+  // returned before hydration — and since close/reopen doesn't reset the page,
+  // only a full reload recovered it. Re-rendering is cheap and the in-progress
+  // Running tab is rendered separately just below.
+  _renderRecipes();
+  _rendered = true;
+  _clearCookbookNotif();
+  _renderRunningTab();
+  // Self-heal: revive any download tasks whose tmux session is still alive
+  // but were persisted as done/error (covers the "restarted server while a
+  // big multi-shard download was in flight" case — the task survived in
+  // tmux, the cookbook just lost track of it).
+  try { _selfHealStaleTasks({ oneShot: true }); } catch {}
+  if (_content) {
+    // Put the panel in its entering state before it becomes visible. On
+    // mobile, showing first and adding the class a frame later can paint the
+    // sheet at its final position, which makes the slide-up look like a snap.
+    _content.classList.add('cookbook-modal-entering');
+  }
+  modal.classList.remove('hidden');
+  if (_content) {
+    void _content.offsetWidth;
+    _content.addEventListener('animationend', () => {
+      _content.classList.remove('cookbook-modal-entering');
+    }, { once: true });
+  }
+  setTimeout(_applyIntent, 0);
  } finally {
    _setCookbookOpening(false);
  }
@@ -2217,9 +2232,6 @@ const shared = {
  _getPort,
  _sshPrefix,
  _getPlatform,
-  _serverByVal,
-  _selectedServer,
-  _currentServerValue,
  _isWindows,
  _isMetal,
  _buildEnvPrefix,
@@ -242,11 +242,7 @@ export function _wirePanelEvents(panel, model, backend) {
  const dlBtn = panel.querySelector('.hwfit-dl-btn');
  if (dlBtn) {
    dlBtn.addEventListener('click', () => {
-      if (backend === 'ollama') {
-        _runPanelCmd(panel, _buildDownloadCmd(model, backend), { timeout: 0 });
-      } else {
-        _runModelDownload(panel, model, backend);
-      }
+      _runModelDownload(panel, model, backend)
    });
  }

@@ -459,7 +455,9 @@ export async function _runModelDownload(panel, model, backend, hostOverride) {
    uiModule.showToast(_missingGgufMessage(model));
    return;
  }
-  const repo = ggufSource?.repo || model.quant_repo || model.name;
+  const repo = backend === 'ollama'
+    ? (model.ollama || model.ollama_name || model.name)
+    : (ggufSource?.repo || model.quant_repo || model.name);
  const include = backend === 'llamacpp' ? _ggufIncludePattern(model, ggufSource) : null;

  _syncEnvFromPanel(panel);
@@ -494,7 +492,7 @@ export async function _runModelDownload(panel, model, backend, hostOverride) {
  const platform = host ? (srv.platform || '') : (_envState.platform || '');
  const isWin = host ? (platform === 'windows') : _isWindows();

-  const payload = { repo_id: repo };
+  const payload = { repo_id: repo, backend };
  if (include) payload.include = include;
  // Large downloads are where hf_transfer most often dies near the end. Use the
  // plain HuggingFace downloader up front for big model files; it is slower, but
@@ -1564,6 +1564,10 @@ export async function _launchServeTask(shortName, repo, cmd, fields, hostOverrid
    const payload = { repo_id: repo, remote_host: _host || undefined, ssh_port: _sp || undefined, _cmd: cmd, _fields: fields || undefined, _env: _usedEnv, _envPath: _usedEnvPath, _gpus: _usedGpus };
    _addTask(data.session_id, shortName, 'serve', payload);
    uiModule.showToast(`Serving ${shortName}...`);
+    // Auto-register may have enabled an existing (offline) endpoint for this
+    // host:port. Refresh the picker so the row is no longer dimmed, and the
+    // user doesn't see "offline" on a serve they just started.
+    try { _refreshModelsAfterEndpointChange(); } catch (_) {}
  } catch (e) {
    uiModule.showToast('Failed: ' + e.message);
  }
@@ -3032,6 +3036,11 @@ async function _reconnectTask(el, task) {
            if (info.status === 'ready' && !task._serveReady) {
              task._serveReady = true;
              _updateTask(task.sessionId, { _serveReady: true });
+              // The auto-registered endpoint was marked offline while the
+              // server was coming up. Now that it's reachable, nudge the
+              // picker to re-probe so the offline pill clears without the
+              // user having to reopen Settings or refresh the page.
+              try { _refreshModelsAfterEndpointChange(); } catch (_) {}
            }
            if (info.phase) {
              badge.textContent = info.phase;
@@ -129,7 +129,7 @@ try { (function () {
          </label>
        </div>

-        <div class="hwfit-schedule-row">
+        <div class="hwfit-schedule-row hwfit-schedule-when-row">
          <label class="hwfit-schedule-field">
            <span>From</span>
            <input type="time" class="hwfit-sched-start cookbook-field-input" value="09:00" />
@@ -138,24 +138,24 @@ try { (function () {
            <span>Until</span>
            <input type="time" class="hwfit-sched-end cookbook-field-input" value="17:00" />
          </label>
-        </div>
-
-        <div class="hwfit-schedule-row hwfit-schedule-days-row">
-          <span class="hwfit-schedule-label">Days</span>
-          <div class="hwfit-sched-days">
-            ${DAYS.map(d => `
-              <button type="button" class="hwfit-sched-day-chip${WEEKDAYS.has(d.k) ? " is-on" : ""}" data-day="${d.k}">${d.l}</button>
-            `).join("")}
+          <label class="hwfit-schedule-field hwfit-schedule-days-field">
+            <span>Days</span>
+            <div class="hwfit-sched-days">
+              ${DAYS.map(d => `
+                <button type="button" class="hwfit-sched-day-chip${WEEKDAYS.has(d.k) ? " is-on" : ""}" data-day="${d.k}">${d.l}</button>
+              `).join("")}
+            </div>
+          </label>
+          <div class="hwfit-schedule-actions-inline">
+            <button type="button" class="cookbook-btn hwfit-sched-cancel" title="Cancel">
+              <svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round" style="vertical-align:-1px;margin-right:5px;flex-shrink:0;"><line x1="18" y1="6" x2="6" y2="18"/><line x1="6" y1="6" x2="18" y2="18"/></svg>
+              <span>Cancel</span>
+            </button>
+            <button type="button" class="cookbook-btn hwfit-sched-save" title="Save schedule" aria-label="Save schedule">
+              <svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="vertical-align:-1px;margin-right:5px;flex-shrink:0;"><rect x="3" y="4" width="18" height="18" rx="2"/><line x1="16" y1="2" x2="16" y2="6"/><line x1="8" y1="2" x2="8" y2="6"/><line x1="3" y1="10" x2="21" y2="10"/></svg>
+              <span>Save</span>
+            </button>
          </div>
-          <span class="hwfit-schedule-actions-spacer"></span>
-          <button type="button" class="cookbook-btn hwfit-sched-cancel" title="Cancel">
-            <svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round" style="vertical-align:-1px;margin-right:5px;flex-shrink:0;"><line x1="18" y1="6" x2="6" y2="18"/><line x1="6" y1="6" x2="18" y2="18"/></svg>
-            <span>Cancel</span>
-          </button>
-          <button type="button" class="cookbook-btn hwfit-sched-save" title="Save schedule" aria-label="Save schedule">
-            <svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="vertical-align:-1px;margin-right:5px;flex-shrink:0;"><rect x="3" y="4" width="18" height="18" rx="2"/><line x1="16" y1="2" x2="16" y2="6"/><line x1="8" y1="2" x2="8" y2="6"/><line x1="3" y1="10" x2="21" y2="10"/></svg>
-            <span>Save</span>
-          </button>
        </div>

        <div class="hwfit-sched-err"></div>
@@ -14,7 +14,6 @@ import { bindMenuDismiss, dismissOrRemove } from './escMenuStack.js';
 let _envState;
 let _sshCmd;
 let _getPort;
-let _serverByVal;
 let _sshPrefix;
 let _getPlatform;
 let _isWindows;
@@ -98,14 +97,14 @@ function _selectedServeTarget(panel) {
  const select = document.getElementById('hwfit-server-select') || document.getElementById('hwfit-dl-server');
  const servers = Array.isArray(_envState.servers) ? _envState.servers : [];
  let host = _envState.remoteHost || '';
-  let server = host ? (_serverByVal?.(_envState.remoteServerKey || host) || servers.find(s => s.host === host)) : null;
+  let server = host ? servers.find(s => s.host === host) : null;
  if (select && select.value != null) {
    if (select.value === 'local') {
      host = '';
      server = servers.find(s => !s.host || s.host === 'local') || null;
    } else {
      const idx = /^\d+$/.test(String(select.value)) ? parseInt(select.value, 10) : -1;
-      server = _serverByVal?.(select.value) || (idx >= 0 ? servers[idx] : null) || null;
+      server = servers.find(s => s.host === select.value) || (idx >= 0 ? servers[idx] : null) || null;
      host = server?.host || '';
    }
  }
@@ -115,7 +114,7 @@ function _selectedServeTarget(panel) {
    : (server?.name || 'local server');
  return {
    host,
-    port: host ? (server?.port || _getPort(host) || '') : '',
+    port: host ? (_getPort(host) || server?.port || '') : '',
    venv,
    label,
  };
@@ -243,21 +242,6 @@ function _shellPathExpr(path) {
 function _selectedGgufExpr(model, repo, relPath) {
  const rel = String(relPath || '').replace(/^\/+/, '');
  if (!rel) return '';
-  if (_isWindows()) {
-    // PowerShell: plain path — no bash $() syntax (backend validator rejects
-    // $( ) in non-prelude commands, and PowerShell doesn't have printf).
-    const relW = rel.replace(/\//g, '\\');
-    if (model.is_local_dir && model.path) {
-      const base = String(model.path || '').replace(/\/+$/, '').replace(/\//g, '\\');
-      return `${base}\\${repo.replace(/\//g, '\\')}\\${relW}`;
-    }
-    if (model.path) {
-      const base = String(model.path || '').replace(/\/+$/, '').replace(/\//g, '\\');
-      return `${base}\\models--${repo.replace(/\//g, '--')}\\snapshots\\${relW}`;
-    }
-    const cacheRepo = repo.replace(/\//g, '--');
-    return `$env:USERPROFILE\\.cache\\huggingface\\hub\\models--${cacheRepo}\\snapshots\\${relW}`;
-  }
  if (model.is_local_dir && model.path) {
    const base = String(model.path || '').replace(/\/+$/, '');
    return `$(printf %s ${_shellPathExpr(`${base}/${repo}/${rel}`)})`;
@@ -271,15 +255,6 @@ function _selectedGgufExpr(model, repo, relPath) {
 }

 function _ggufSearchDirExpr(model, repo) {
-  if (_isWindows()) {
-    if (model.is_local_dir && model.path) {
-      return `${String(model.path || '').replace(/\/+$/, '').replace(/\//g, '\\')}\\${repo.replace(/\//g, '\\')}`;
-    }
-    if (model.path) {
-      return `${String(model.path || '').replace(/\/+$/, '').replace(/\//g, '\\')}\\models--${repo.replace(/\//g, '--')}\\snapshots`;
-    }
-    return `$env:USERPROFILE\\.cache\\huggingface\\hub\\models--${repo.replace(/\//g, '--')}\\snapshots`;
-  }
  if (model.is_local_dir && model.path) return _shellQuote(`${String(model.path || '').replace(/\/+$/, '')}/${repo}`);
  if (model.path) return _shellQuote(`${String(model.path || '').replace(/\/+$/, '')}/models--${repo.replace(/\//g, '--')}/snapshots`);
  return `"$HOME/.cache/huggingface/hub/models--${repo.replace(/\//g, '--')}/snapshots"`;
@@ -537,7 +512,7 @@ function _rerenderCachedModels() {
      // The venv set per-server in Settings (server.envPath). Used as the venv
      // field default when the global active env path isn't carrying it, so a
      // configured server venv shows up without re-typing it.
-      const _selSrv = _serverByVal?.(_es.remoteServerKey || _es.remoteHost || '') || {};
+      const _selSrv = (_es.servers || []).find(s => s.host === (_es.remoteHost || '')) || {};
      const _srvVenv = _selSrv.envPath || '';
      // Serve state schema: { _byRepo: { <repo>: {...} }, _lastUsed: {...} }.
      // Loading priority: this-repo's saved settings → last-used (from any
@@ -600,7 +575,7 @@ function _rerenderCachedModels() {
        + `<button type="button" class="cookbook-slot-btn cookbook-saved-arrow" title="${esc(_arrowTitle)}">${_arrowLabel}</button>`
        + `</div>`;

-      let panelHtml = `<div class="hwfit-serve-panel">${_slotsHtml}`;
+      let panelHtml = `<div class="hwfit-serve-panel">`;
      // Warn when serving a model whose download hasn't fully completed —
      // the user CAN still hit Launch (vLLM/llama-server will start, then
      // crash trying to read missing shards), but they should know.
@@ -633,26 +608,48 @@ function _rerenderCachedModels() {
        _gpuBtnsHtml += `<button type="button" class="cookbook-gpu-btn${on ? ' active' : ''}" data-gpu="${i}">${i}</button>`;
      }
      panelHtml += `<label>${_l('GPUs','Toggle which GPUs to use')}<div class="cookbook-gpu-group">${_gpuBtnsHtml}</div><input type="hidden" class="hwfit-sf" data-field="gpus" value="${esc(defaultGpus)}" /></label>`;
+      // Save / saved-configs split button — moved into Row 1 (next to GPUs)
+      // so it shares the same baseline as the rest of the top controls.
+      panelHtml += _slotsHtml;
      panelHtml += `</div>`;
      panelHtml += `<div class="hwfit-serve-runtime-note" style="display:none;font-size:11px;line-height:1.35;color:var(--fg-muted);margin-top:-4px;"></div>`;
      if (_ggufChoices.length > 1) {
-        panelHtml += `<div class="hwfit-serve-row hwfit-backend-llamacpp">`;
-        panelHtml += `<label class="hwfit-backend-llamacpp">${_l('GGUF File','Choose the exact GGUF artifact to serve from this cached model folder.')}<select class="hwfit-sf hwfit-sf-wide" data-field="gguf_file">${_ggufOptions}</select></label>`;
+        // Show the GGUF File dropdown for BOTH llama.cpp and Ollama — Ollama
+        // also needs to know which exact .gguf to import via the new
+        // `docker exec ollama-test ollama-import` auto-fill (otherwise the
+        // helper falls back to "first sorted gguf", which may not match what
+        // the user picked).
+        panelHtml += `<div class="hwfit-serve-row hwfit-backend-llamacpp hwfit-backend-ollama">`;
+        panelHtml += `<label class="hwfit-backend-llamacpp hwfit-backend-ollama">${_l('GGUF File','Choose the exact GGUF artifact to serve from this cached model folder.')}<select class="hwfit-sf hwfit-sf-wide" data-field="gguf_file">${_ggufOptions}</select></label>`;
        panelHtml += `</div>`;
      } else if (_defaultGguf) {
        panelHtml += `<input type="hidden" class="hwfit-sf" data-field="gguf_file" value="${esc(_defaultGguf)}" />`;
      }
-      // Row 2: Core settings
-      panelHtml += `<div class="hwfit-serve-row hwfit-backend-vllm hwfit-backend-sglang hwfit-backend-llamacpp">`;
+      // Row 2: Core settings — the handful you actually touch every launch.
+      // TP / Context / GPU / GPU Mem / Max Seqs / Dtype. Everything else
+      // (Swap, KV Cache, Attention backend, Env vars, llama.cpp batch/ubatch)
+      // moved to the Advanced fold below to keep this row scannable.
+      panelHtml += `<div class="hwfit-serve-row hwfit-backend-vllm hwfit-backend-sglang hwfit-backend-llamacpp hwfit-backend-ollama">`;
      panelHtml += `<label class="hwfit-backend-vllm hwfit-backend-sglang">${_l('TP','Tensor Parallelism — split model across N GPUs')}<select class="hwfit-sf" data-field="tp">${tpOpts}</select></label>`;
      // ctx resets to the model's max on every panel open (the real ctx slider
      // lives in the Scan/Download toolbar — see cookbook.js .hwfit-ctx-control).
      panelHtml += `<label>${_l('Context','Max tokens per request — resets to the model max on every open. Lower = less VRAM')}<input type="text" class="hwfit-sf" data-field="ctx" value="${esc(m.context_length || m.context || '20000')}" /></label>`;
      panelHtml += `<label>${_l('GPU','Which GPU to use. Leave empty for default')}<input type="text" class="hwfit-sf" data-field="gpu_id" value="${esc(sv('gpu_id', ''))}" placeholder="auto" style="width:50px;" /></label>`;
      panelHtml += `<label class="hwfit-backend-vllm hwfit-backend-sglang">${_l('GPU Mem','Fraction of GPU memory (0.0–1.0). Lower if OOM')}<input type="text" class="hwfit-sf" data-field="gpu_mem" value="${esc(sv('gpu_mem', '0.90'))}" /></label>`;
-      panelHtml += `<label class="hwfit-backend-vllm">${_l('Swap','CPU swap space in GB. Leave empty to omit (removed in newer vLLM)')}<input type="text" class="hwfit-sf" data-field="swap" value="${esc(sv('swap', ''))}" placeholder="off" /></label>`;
      panelHtml += `<label class="hwfit-backend-vllm hwfit-backend-sglang">${_l('Max Seqs','Maximum concurrent requests. Lower = less memory. Default 4 — prosumer GPUs often OOM on vLLM default 256 during CUDA graph capture.')}<input type="text" class="hwfit-sf" data-field="max_seqs" value="${esc(sv('max_seqs', '4'))}" placeholder="4" /></label>`;
      panelHtml += `<label>${_l('Dtype','Data type for weights. auto picks best for GPU')}<select class="hwfit-sf" data-field="dtype">${dtypeOpts}</select></label>`;
+      panelHtml += `</div>`;
+      // ── Advanced (collapsed by default) ──
+      // Everything below the fold is tuning users only touch occasionally:
+      // vLLM kernel/env knobs, llama.cpp fit/cache/split controls, the
+      // GGUF batch sizes, the speculative-decoding row, and the live VRAM
+      // monitor. Wrapped in a native <details> so toggle state survives
+      // re-renders cheaply and a closed fold doesn't trigger any layout
+      // work for the dozens of nested inputs.
+      panelHtml += `<details class="hwfit-serve-advanced">`;
+      panelHtml += `<summary class="hwfit-serve-advanced-summary">Advanced</summary>`;
+      // Advanced vLLM/SGLang row (KV Cache, Attention, Swap, Env)
+      panelHtml += `<div class="hwfit-serve-row hwfit-backend-vllm hwfit-backend-sglang">`;
      panelHtml += `<label class="hwfit-backend-vllm">${_l('KV Cache','vLLM --kv-cache-dtype. auto uses the model/runtime default; fp8 reduces KV memory for long context.')}<select class="hwfit-sf" data-field="vllm_kv_cache_dtype" style="height:32px;">${vllmKvCacheOpts}</select></label>`;
      // Attention backend selector — pin the kernel impl. Default `auto` lets
      // vLLM pick FlashInfer (which JITs on first use and breaks on older
@@ -662,6 +659,7 @@ function _rerenderCachedModels() {
      const vllmAttnBackendOpts = ['auto', 'FLASH_ATTN', 'XFORMERS', 'FLASHINFER', 'TORCH_SDPA']
        .map(b => `<option value="${b === 'auto' ? '' : b}"${(sv('vllm_attn_backend','') === (b === 'auto' ? '' : b)) ? ' selected' : ''}>${b}</option>`).join('');
      panelHtml += `<label class="hwfit-backend-vllm">${_l('Attention','vLLM VLLM_ATTENTION_BACKEND. auto = vLLM picks (often FLASHINFER, which JITs and can fail on old nvcc). FLASH_ATTN skips the JIT entirely.')}<select class="hwfit-sf" data-field="vllm_attn_backend" style="height:32px;">${vllmAttnBackendOpts}</select></label>`;
+      panelHtml += `<label class="hwfit-backend-vllm">${_l('Swap','CPU swap space in GB. Leave empty to omit (removed in newer vLLM)')}<input type="text" class="hwfit-sf" data-field="swap" value="${esc(sv('swap', ''))}" placeholder="off" /></label>`;
      // Free-text env-vars field. Anything pasted here is prepended to the
      // launch command verbatim. Use for CUDACXX, PATH overrides, NCCL_*
      // tuning, or any other KEY=VALUE pair that doesn't have a dedicated
@@ -669,6 +667,12 @@ function _rerenderCachedModels() {
      // already exported so they expand correctly here.
      panelHtml += `<label class="hwfit-backend-vllm hwfit-backend-sglang" style="flex:1 1 100%;">${_l('Env','Extra KEY=VALUE env-var pairs prepended to the launch (space-separated). Example: CUDACXX=$VIRTUAL_ENV/lib/python3.10/site-packages/nvidia/cuda_nvcc/bin/nvcc — points flashinfer at the venv-bundled nvcc when the system one is too old for your GPU.')}<input type="text" class="hwfit-sf" data-field="extra_env" value="${esc(sv('extra_env',''))}" placeholder="CUDACXX=/path/to/nvcc NCCL_P2P_DISABLE=1" style="width:100%;" /></label>`;
      panelHtml += `</div>`;
+      // Advanced llama.cpp row (Batch / UBatch — moved out of Core for the
+      // same "rarely touched" reason as the vLLM extras above).
+      panelHtml += `<div class="hwfit-serve-row hwfit-backend-llamacpp">`;
+      panelHtml += `<label class="hwfit-backend-llamacpp">${_l('Batch','llama.cpp prompt batch size. Leave blank for llama.cpp default.')}<input type="text" class="hwfit-sf" data-field="llama_batch_size" value="${esc(sv('llama_batch_size', ''))}" placeholder="2048" /></label>`;
+      panelHtml += `<label class="hwfit-backend-llamacpp">${_l('UBatch','llama.cpp physical micro-batch size. Leave blank for llama.cpp default.')}<input type="text" class="hwfit-sf" data-field="llama_ubatch_size" value="${esc(sv('llama_ubatch_size', ''))}" placeholder="512" /></label>`;
+      panelHtml += `</div>`;
      // Row 2b: Diffusers settings
      const diffDtypeOpts = ['bfloat16','float16','float32'].map(d => `<option value="${d}"${sv('diff_dtype','bfloat16')===d?' selected':''}>${d}</option>`).join('');
      const deviceMapOpts = ['balanced','auto','sequential'].map(d => `<option value="${d}"${sv('diff_device_map','balanced')===d?' selected':''}>${d}</option>`).join('');
@@ -691,7 +695,7 @@ function _rerenderCachedModels() {
      const llamaFitOpts = ['', 'off', 'on'].map(d => `<option value="${d}"${sv('llama_fit','')===d?' selected':''}>${d||'default'}</option>`).join('');
      const llamaSplitModeOpts = ['', 'layer', 'tensor', 'row', 'none'].map(d => `<option value="${d}"${sv('llama_split_mode','')===d?' selected':''}>${d||'default'}</option>`).join('');
      panelHtml += `<div class="hwfit-serve-row hwfit-backend-llamacpp">`;
-      panelHtml += `<label>${_l('CPU MoE','n-cpu-moe: number of MoE expert layers to run on CPU when the model is bigger than VRAM. 0 = all on GPU. Set automatically by the Auto profiles below.')}<input type="text" class="hwfit-sf" data-field="n_cpu_moe" value="${esc(sv('n_cpu_moe',''))}" placeholder="0" style="width:54px;" /></label>`;
+      panelHtml += `<label>${_l('CPU MoE','n-cpu-moe: number of MoE expert layers to run on CPU when the model is bigger than VRAM. 0 = all on GPU. Set automatically by the Auto profiles below.')}<input type="text" class="hwfit-sf" data-field="n_cpu_moe" value="${esc(sv('n_cpu_moe',''))}" placeholder="0" style="width:54px;position:relative;top:-8px;" /></label>`;
      panelHtml += `<label>${_l('KV Cache','cache-type-k/v: quantize the KV cache. q4_0 = smallest (more context), q8_0 = sharp long-context, f16 = full. Blank = llama.cpp default.')}<select class="hwfit-sf" data-field="cache_type">${_kvOpts}</select></label>`;
      panelHtml += `<label class="hwfit-sf-cb" style="align-self:end;"><input type="checkbox" class="hwfit-sf" data-field="flash_attn"${sv('flash_attn',false)?' checked':''} /> Flash Attn${_h('--flash-attn on: faster attention + needed for quantized KV cache.')}</label>`;
      panelHtml += `<label class="hwfit-sf-cb" style="align-self:end;"><input type="checkbox" class="hwfit-sf" data-field="vision"${sv('vision',false)?' checked':''} /> Vision${_h('Serve with the vision encoder so the model can read images. Auto-finds an mmproj-*.gguf next to the model (download one into the model folder). Adds ~1 GB VRAM + a small per-image cost.')}</label>`;
@@ -701,19 +705,16 @@ function _rerenderCachedModels() {
      // explicit overrides for known-good advanced presets; blank keeps
      // llama.cpp/profile defaults.
      panelHtml += `<div class="hwfit-serve-row hwfit-backend-llamacpp">`;
-      panelHtml += `<label>${_l('Split Mode','llama.cpp GPU placement. layer is the usual default; tensor splits weights and KV across GPUs.')}<select class="hwfit-sf" data-field="llama_split_mode">${llamaSplitModeOpts}</select></label>`;
+      panelHtml += `<label>${_l('Split Mode','llama.cpp GPU placement. layer is the usual default; tensor splits weights and KV across GPUs.')}<select class="hwfit-sf" data-field="llama_split_mode" style="position:relative;top:-8px;">${llamaSplitModeOpts}</select></label>`;
      panelHtml += `<label>${_l('Tensor Split','GPU proportions for llama.cpp, e.g. 50,50 across two visible GPUs. Leave blank for auto.')}<input type="text" class="hwfit-sf" data-field="llama_tensor_split" value="${esc(sv('llama_tensor_split', ''))}" placeholder="50,50" /></label>`;
      panelHtml += `<label>${_l('Main GPU','llama.cpp --main-gpu index inside the visible GPU set. Mostly useful for split mode none/row.')}<input type="text" class="hwfit-sf" data-field="llama_main_gpu" value="${esc(sv('llama_main_gpu', ''))}" placeholder="auto" /></label>`;
      panelHtml += `<label>${_l('Parallel','llama.cpp parallel slots. Leave blank for llama.cpp default; 1 matches single-lane presets.')}<input type="text" class="hwfit-sf" data-field="llama_parallel" value="${esc(sv('llama_parallel', ''))}" placeholder="1" /></label>`;
-      panelHtml += `<label>${_l('Batch','llama.cpp prompt batch size. Leave blank for llama.cpp default.')}<input type="text" class="hwfit-sf" data-field="llama_batch_size" value="${esc(sv('llama_batch_size', ''))}" placeholder="2048" /></label>`;
-      panelHtml += `<label>${_l('UBatch','llama.cpp physical micro-batch size. Leave blank for llama.cpp default.')}<input type="text" class="hwfit-sf" data-field="llama_ubatch_size" value="${esc(sv('llama_ubatch_size', ''))}" placeholder="512" /></label>`;
-      panelHtml += `</div>`;
-      // Row 2d: Auto profiles — computed from detected hardware (see profiles.py).
-      // Buttons are injected after the panel mounts (needs an async fetch).
-      panelHtml += `<div class="hwfit-serve-row hwfit-backend-llamacpp hwfit-serve-profiles" style="align-items:center;gap:8px;">`;
-      panelHtml += `<span style="opacity:0.7;font-size:11px;">Auto profiles:</span>`;
-      panelHtml += `<span class="hwfit-profile-btns" style="display:flex;gap:6px;flex-wrap:wrap;"><span style="opacity:0.5;font-size:11px;">computing…</span></span>`;
      panelHtml += `</div>`;
+      // Auto-profile chips row removed — visual fit with the rest of the
+      // serve panel was off, and the manual ctx/n_cpu_moe/cache controls
+      // above are already sufficient. The hwfit profile API
+      // (/api/hwfit/profiles) is still available for any caller that
+      // wants it.
      // Live VRAM / RAM-spillover monitor for the serve target's GPU. Polls
      // /api/cookbook/gpus while the panel is open so you can SEE whether the
      // config fits VRAM (fast) or spills to system RAM (slow). Populated after mount.
@@ -745,7 +746,7 @@ function _rerenderCachedModels() {
      // even for models the auto-detector doesn't recognize. Expert-parallel,
      // reasoning-parser and MoE-env still only appear when auto-detected.
      const _opts2 = _detectModelOptimizations(repo);
-      panelHtml += `<div class="hwfit-serve-checks hwfit-backend-vllm" style="margin-top:2px;">`;
+      panelHtml += `<div class="hwfit-serve-checks hwfit-backend-vllm">`;
      if (_opts2.flags.includes('--enable-expert-parallel')) panelHtml += `<label class="hwfit-sf-cb"><input type="checkbox" class="hwfit-sf" data-field="expert_parallel" /> Expert Parallel</label>`;
      if (_opts2.flags.some(f => f.includes('--reasoning-parser'))) { const rp = _opts2.flags.find(f => f.includes('--reasoning-parser')).split(' ')[1]; panelHtml += `<label class="hwfit-sf-cb"><input type="checkbox" class="hwfit-sf" data-field="reasoning_parser" data-parser="${rp}" /> Reasoning Parser <span class="hwfit-parser-tag">${rp}</span></label>`; }
      {
@@ -764,6 +765,8 @@ function _rerenderCachedModels() {
      }
      if (_opts2.envVars.length) panelHtml += `<label class="hwfit-sf-cb"><input type="checkbox" class="hwfit-sf" data-field="moe_env" /> MoE Env Vars</label>`;
      panelHtml += `</div>`;
+      // ── End Advanced fold ──
+      panelHtml += `</details>`;
      // Command preview + actions. Wrap the textarea so a floating Copy
      // button can sit at its top-right corner — same pattern as the chat
      // run-output panel.
@@ -825,27 +828,17 @@ function _rerenderCachedModels() {
          // model the file lives under "<path>/<repo>" — search there just like we
          // search the HF snapshots dir, so serving a GGUF from a custom dir works
          // instead of handing llama.cpp a directory (which fails).
-          const _ldir = m.path
-            ? (_isWindows() ? `${m.path.replace(/\//g, '\\')}\\${repo.replace(/\//g, '\\')}` : _shellQuote(`${m.path}/${repo}`))
-            : (_isWindows() ? '' : '""');
-          if (selectedGguf) {
-            f._gguf_path = _selectedGgufExpr(m, repo, selectedGguf.rel_path);
-          } else if (_isWindows()) {
-            // Windows fallback: no bash $() available; validator rejects it.
-            // Return empty so the serve fails with a clear message.
-            f._gguf_path = '';
-          } else if (m.is_local_dir && m.path) {
-            f._gguf_path = `$({ find ${_ldir} -name '*-00001-of-*.gguf' 2>/dev/null | sort; find ${_ldir} -name '*.gguf' 2>/dev/null | sort; } | head -1)`;
-          } else {
-            f._gguf_path = `$({ find ${dir} -name '*-00001-of-*.gguf' 2>/dev/null | sort; find ${dir} -name '*.gguf' 2>/dev/null | sort; } | head -1)`;
-          }
+          const _ldir = m.path ? _shellQuote(`${m.path}/${repo}`) : '""';
+          f._gguf_path = selectedGguf
+            ? _selectedGgufExpr(m, repo, selectedGguf.rel_path)
+            : m.is_local_dir && m.path
+            ? `$({ find ${_ldir} -name '*-00001-of-*.gguf' 2>/dev/null | sort; find ${_ldir} -name '*.gguf' 2>/dev/null | sort; } | head -1)`
+            : `$({ find ${dir} -name '*-00001-of-*.gguf' 2>/dev/null | sort; find ${dir} -name '*.gguf' 2>/dev/null | sort; } | head -1)`;
          // Vision: auto-find the mmproj (CLIP/projector) file in the same dir.
          // Resolved at runtime so the toggle just works if an mmproj-*.gguf is
          // present (downloaded alongside the model). Empty if none → cmd omits it.
          const _vsearchdir = (m.is_local_dir && m.path) ? _ldir : dir;
-          f._mmproj_path = _isWindows()
-            ? (_vsearchdir ? `${_vsearchdir}\\mmproj*.gguf` : '')
-            : `$(find ${_vsearchdir} -iname 'mmproj*.gguf' 2>/dev/null | sort | head -1)`;
+          f._mmproj_path = `$(find ${_vsearchdir} -iname 'mmproj*.gguf' 2>/dev/null | sort | head -1)`;
        }
        if (f.reasoning_parser) {
          const _rpEl2 = panel.querySelector('[data-field="reasoning_parser"]');
@@ -886,72 +879,29 @@ function _rerenderCachedModels() {
        _clampCtx(false);   // fix any stale/preset value already present
      }

-      // Auto profiles — fetch hardware-computed llama.cpp profiles and render
-      // them as clickable chips. Clicking one fills the ctx/CPU-MoE/KV/flash
-      // fields and rebuilds the command. Computed from detected VRAM (see
-      // services/hwfit/profiles.py); rough on t/s, accurate on fit.
-      async function _loadServeProfiles() {
-        const wrap = panel.querySelector('.hwfit-profile-btns');
-        if (!wrap) return;
+      // Tighten the ctx slider's upper bound to the model's trained limit.
+      // Asking llama.cpp for ctx > n_ctx_train overflows and, with a quantized
+      // KV cache, can crash the GPU (radv ErrorDeviceLost). The auto-profile
+      // chip row that used to also live here was removed — visual fit with
+      // the rest of the serve panel was off — but this clamp is essential.
+      (async () => {
        try {
          const host = (_es.remoteHost || '').trim();
-          const selected = _serverByVal?.(_es.remoteServerKey || host);
          const params = new URLSearchParams({ model: repo });
          if (host) {
            params.set('host', host);
-            const _sp = selected?.port;
+            const _sp = (_es.servers || []).find(s => s.host === host)?.port;
            if (_sp) params.set('ssh_port', _sp);
          }
-          // SERVE mode: this is a specific GGUF file already on disk, so its quant
-          // is fixed — tell the profiler the file's real size + quant so it varies
-          // only the serving knobs (KV/ctx/offload), not the quant. Parse the size
-          // from m.size (e.g. "20.6 GB") and the quant from the file/repo name.
-          const _sizeMatch = String(m.size || '').match(/([\d.]+)\s*GB/i);
-          if (_sizeMatch) params.set('serve_weights_gb', _sizeMatch[1]);
-          const _qMatch = String(repo).match(/(Q\d[\w]*|IQ\d[\w]*|F16|BF16|FP8)/i);
-          if (_qMatch) params.set('serve_quant', _qMatch[1]);
          const res = await fetch(`/api/hwfit/profiles?${params}`);
          const data = await res.json();
-          // Remember the model's trained context limit and clamp the ctx field
-          // to it — asking llama.cpp for ctx > n_ctx_train overflows and, with a
-          // quantized KV cache, can crash the GPU (radv ErrorDeviceLost).
          const ctxMax = Number(data && data.model_ctx_max) || 0;
          if (ctxMax > 0) {
-            panel._modelCtxMax = ctxMax;   // tighten the clamp to the real limit
-            _clampCtx(false);              // re-apply now that we know the model's max
+            panel._modelCtxMax = ctxMax;
+            _clampCtx(false);
          }
-          const profs = (data && Array.isArray(data.profiles)) ? data.profiles : [];
-          if (!profs.length) { wrap.innerHTML = `<span style="opacity:0.5;font-size:11px;">no auto profile for this model</span>`; return; }
-          wrap.innerHTML = '';
-          for (const p of profs) {
-            const b = document.createElement('button');
-            b.type = 'button';
-            b.className = 'cookbook-btn hwfit-profile-chip';
-            b.style.cssText = 'height:24px;padding:0 9px;font-size:11px;';
-            const off = p.offloads ? `, ncm${p.n_cpu_moe}` : ', all-GPU';
-            b.textContent = `${p.label} · ${p.quant} · ${Math.round(p.ctx/1024)}k${off}`;
-            b.title = `${p.note}\nKV ${p.cache_type}, ~${p.est_vram_gb} GB VRAM`;
-            b.addEventListener('click', () => {
-              const set = (field, val) => {
-                const el = panel.querySelector(`[data-field="${field}"]`);
-                if (!el) return;
-                if (el.type === 'checkbox') el.checked = !!val; else el.value = val;
-              };
-              set('ctx', p.ctx);
-              set('n_cpu_moe', p.n_cpu_moe || '');
-              set('cache_type', p.cache_type || '');
-              set('flash_attn', true);   // required for a quantized KV cache
-              wrap.querySelectorAll('.hwfit-profile-chip').forEach(x => x.classList.remove('cookbook-btn-active'));
-              b.classList.add('cookbook-btn-active');
-              updateCmd();
-            });
-            wrap.appendChild(b);
-          }
-        } catch {
-          wrap.innerHTML = `<span style="opacity:0.5;font-size:11px;">profile compute failed</span>`;
-        }
-      }
-      _loadServeProfiles();
+        } catch { /* clamp falls back to the static default */ }
+      })();

      // Live GPU-memory monitor: poll /api/cookbook/gpus and show VRAM usage +
      // RAM-spillover, with a plain-language health/speed hint. Lets you tell at
@@ -962,11 +912,10 @@ function _rerenderCachedModels() {
        if (!el || !document.body.contains(el)) return false;  // panel closed → stop
        try {
          const host = (_es.remoteHost || '').trim();
-          const selected = _serverByVal?.(_es.remoteServerKey || host);
          const params = new URLSearchParams();
          if (host) {
            params.set('host', host);
-            const _sp = selected?.port;
+            const _sp = (_es.servers || []).find(s => s.host === host)?.port;
            if (_sp) params.set('ssh_port', _sp);
          }
          const res = await fetch('/api/cookbook/gpus' + (params.toString() ? '?' + params : ''));
@@ -1535,6 +1484,38 @@ function _rerenderCachedModels() {
          }
          panel._gpuProbe.byIdx = new Map(data.gpus.map(g => [g.index, g]));
          panel._gpuProbe.host = remoteHost;
+          // If the probe found more GPUs than the panel originally
+          // rendered (e.g. host switched from a 1-iGPU local box to an
+          // 8-GPU remote), append buttons for the missing indexes so the
+          // user can actually toggle them. Reuse the parent <div> from
+          // the first existing button as the insertion target.
+          try {
+            const _existing = Array.from(panel.querySelectorAll('.cookbook-gpu-btn'));
+            const _grp = _existing[0] && _existing[0].parentElement;
+            if (_grp) {
+              const _have = new Set(_existing.map(b => parseInt(b.dataset.gpu, 10)));
+              const _activeStr = (panel.querySelector('[data-field="gpus"]')?.value || '').split(',').map(s => s.trim());
+              data.gpus.forEach(g => {
+                if (_have.has(g.index)) return;
+                const _b = document.createElement('button');
+                _b.type = 'button';
+                _b.className = 'cookbook-gpu-btn' + (_activeStr.includes(String(g.index)) ? ' active' : '');
+                _b.dataset.gpu = String(g.index);
+                _b.textContent = String(g.index);
+                _grp.appendChild(_b);
+                // Re-wire the click handler the same way the panel did
+                // on first render. Toggles active + rewrites the hidden
+                // gpus input from the live set of active buttons.
+                _b.addEventListener('click', () => {
+                  _b.classList.toggle('active');
+                  const activeBtns = [...panel.querySelectorAll('.cookbook-gpu-btn.active')];
+                  const ids = activeBtns.map(x => x.dataset.gpu).sort((a, b) => +a - +b).join(',');
+                  const hidden = panel.querySelector('[data-field="gpus"]');
+                  if (hidden) { hidden.value = ids; hidden.dispatchEvent(new Event('change', { bubbles: true })); }
+                });
+              });
+            }
+          } catch (_) {}
          panel.querySelectorAll('.cookbook-gpu-btn').forEach(b => {
            const idx = parseInt(b.dataset.gpu);
            const g = panel._gpuProbe.byIdx.get(idx);
@@ -1790,7 +1771,7 @@ function _rerenderCachedModels() {
            const _probeParams = new URLSearchParams();
            if (_probeHost) {
              _probeParams.set('host', _probeHost);
-              const _sp = (_serverByVal?.(_envState.remoteServerKey || _probeHost) || {}).port;
+              const _sp = (_envState.servers || []).find(s => s.host === _probeHost)?.port;
              if (_sp) _probeParams.set('ssh_port', _sp);
            }
            const _probeRes = await fetch('/api/cookbook/gpus' + (_probeParams.toString() ? '?' + _probeParams : ''), { credentials: 'same-origin' });
@@ -1861,12 +1842,20 @@ function _rerenderCachedModels() {
        }
        // Save in the { _byRepo, _lastUsed } schema — no legacy flat keys at
        // the root so per-model state doesn't leak between models.
+        // Stamp `_forceBackend: true` so the next open of this model defaults
+        // to the launched configuration end-to-end, even when the detector
+        // would have picked a different backend. Without this flag, the
+        // `savedMatchesBackend` gate inside sv() throws away every saved
+        // value when the detected backend doesn't match — the user opens
+        // Serve again and the panel looks like a fresh form despite a
+        // known-good prior launch.
        try {
          let cur = {};
          try { cur = JSON.parse(localStorage.getItem(SERVE_STATE_KEY)) || {}; } catch {}
          const byRepo = (cur && cur._byRepo && typeof cur._byRepo === 'object') ? cur._byRepo : {};
-          byRepo[repo] = serveState;
-          localStorage.setItem(SERVE_STATE_KEY, JSON.stringify({ _byRepo: byRepo, _lastUsed: serveState }));
+          const _saved = { ...serveState, _forceBackend: true };
+          byRepo[repo] = _saved;
+          localStorage.setItem(SERVE_STATE_KEY, JSON.stringify({ _byRepo: byRepo, _lastUsed: _saved }));
        } catch {}
        const origEnv = _envState.env;
        const origEnvPath = _envState.envPath;
@@ -1882,7 +1871,8 @@ function _rerenderCachedModels() {
        if (_ssEl && _ssEl.value != null) {
          if (_ssEl.value === 'local') serveHost = '';
          else {
-            const _srv = _serverByVal?.(_ssEl.value) || _envState.servers[parseInt(_ssEl.value)];
+            // Values are host strings now; resolve by host (numeric fallback).
+            const _srv = _envState.servers.find(s => s.host === _ssEl.value) || _envState.servers[parseInt(_ssEl.value)];
            if (_srv) {
              serveHost = _srv.host;
              _srvEnv = _srv.env || '';
@@ -1938,10 +1928,24 @@ function _rerenderCachedModels() {
 function _resolveCacheHost() {
  let host = _envState.remoteHost || '';
  const cacheSrv = document.getElementById('hwfit-cache-server');
+
+  function _serverByCacheValue(val) {
+    if (val === 'local') return null;
+    const found = _envState.servers.find(x => x.host === val)
+      || (/^\d+$/.test(String(val)) ? _envState.servers[parseInt(val)] : null)
+      || _envState.servers.find(x => x.name === val)
+      || null;
+    return found || null;
+  }
+
  if (cacheSrv) {
    const val = cacheSrv.value;
-    if (val === 'local') host = '';
-    else { const s = _serverByVal?.(val) || _envState.servers[parseInt(val)]; if (s) host = s.host; }
+    if (val === 'local') {
+      host = '';
+    } else {
+      const s = _serverByCacheValue(val);
+      if (s) host = s.host;
+    }
  }
  return host;
 }
@@ -2071,8 +2075,12 @@ export async function openServePanelForRepo(repo, fields) {
      let cur = {};
      try { cur = JSON.parse(localStorage.getItem(SERVE_STATE_KEY)) || {}; } catch {}
      const byRepo = (cur && cur._byRepo && typeof cur._byRepo === 'object') ? cur._byRepo : {};
-      byRepo[repo] = fields;
-      localStorage.setItem(SERVE_STATE_KEY, JSON.stringify({ _byRepo: byRepo, _lastUsed: fields }));
+      // Mirror the launch-time save: stamp _forceBackend so the panel's
+      // sv() helper treats these seeded fields as authoritative, not as
+      // overridable defaults.
+      const _seeded = { ...fields, _forceBackend: true };
+      byRepo[repo] = _seeded;
+      localStorage.setItem(SERVE_STATE_KEY, JSON.stringify({ _byRepo: byRepo, _lastUsed: _seeded }));
    } catch {}
  }
  // Switch to the Serve tab (its click handler triggers _fetchCachedModels).
@@ -2099,7 +2107,18 @@ export async function openServePanelForRepo(repo, fields) {
             .find(el => (el.dataset.repo || '').split('/').pop() === _short);
    }
    if (card) {
-      if (!card.classList.contains('doclib-card-expanded')) card.click();
+      // If we were given fields to restore, force a fresh render of the
+      // serve panel so it reads the just-written _byRepo[repo] values
+      // from localStorage. Without this, an already-expanded card kept
+      // its stale form and the "Edit serve" → previous settings round-
+      // trip looked broken from the user's side.
+      if (fields && card.classList.contains('doclib-card-expanded')) {
+        card.click();
+        await new Promise(r => setTimeout(r, 40));
+        card.click();
+      } else if (!card.classList.contains('doclib-card-expanded')) {
+        card.click();
+      }
      try { card.scrollIntoView({ behavior: 'smooth', block: 'center' }); } catch {}
      return true;
    }
@@ -2130,6 +2149,14 @@ export async function _fetchCachedModels() {
  try {
    let host = _envState.remoteHost || '';
    let selectedServer = null;
+    const _serverByCacheValue = (val) => {
+      if (val === 'local') return null;
+      return _envState.servers.find(x => x.host === val)
+        || (/^\d+$/.test(String(val)) ? _envState.servers[parseInt(val)] : null)
+        || _envState.servers.find(x => x.name === val)
+        || null;
+    };
+
    const cacheSrv = document.getElementById('hwfit-cache-server');
    if (cacheSrv) {
      const val = cacheSrv.value;
@@ -2137,11 +2164,11 @@ export async function _fetchCachedModels() {
        host = '';
        selectedServer = _envState.servers.find(s => !s.host || s.host === 'local') || _envState.servers[0];
      } else {
-        const s = _serverByVal?.(val) || _envState.servers[parseInt(val)];
+        const s = _serverByCacheValue(val);
        if (s) { host = s.host; selectedServer = s; }
      }
    } else {
-      selectedServer = _serverByVal?.(_envState.remoteServerKey || host) || _envState.servers[0];
+      selectedServer = _envState.servers.find(s => s.host === host) || _envState.servers[0];
    }
    // Read extra model dirs from the SELECTED server's modelDirs (canonical source)
    const modelDirs = [];
@@ -2171,7 +2198,18 @@ export async function _fetchCachedModels() {
    if (modelDirs.length) qp.set('model_dir', modelDirs.join(','));
    const params = qp.toString() ? `?${qp}` : '';
    const res = await fetch(`/api/model/cached${params}`);
-    if (!res.ok) throw new Error(res.statusText);
+    if (!res.ok) {
+      const body = await res.text().catch(() => '');
+      let msg = '';
+      try {
+        const payload = JSON.parse(body);
+        msg = payload && (payload.detail || payload.error || payload.message);
+      } catch {
+        msg = body;
+      }
+      msg = typeof msg === 'string' ? msg.trim() : '';
+      throw new Error(`HTTP ${res.status} ${res.statusText}${msg ? `: ${msg}` : ''}`);
+    }
    const data = await res.json();
    _dlWp.destroy();

@@ -2268,7 +2306,6 @@ export function initServe(shared) {
  _envState = shared._envState;
  _sshCmd = shared._sshCmd;
  _getPort = shared._getPort;
-  _serverByVal = shared._serverByVal;
  _sshPrefix = shared._sshPrefix;
  _getPlatform = shared._getPlatform;
  _isWindows = shared._isWindows;
@@ -578,13 +578,12 @@ let _libraryArchivedView = false;   // Documents tab showing archived docs?
    const pieces = [];
    if (doc.session_name) pieces.push(`<span>${_esc(doc.session_name)}</span>`);
    if (doc.language && doc.language !== 'text') {
-      const ic = langIcon(doc.language, 11, { style: 'vertical-align:-2px;flex-shrink:0;opacity:0.65;color:currentColor;' });
-      pieces.push(`<span style="display:inline-flex;align-items:center;gap:3px;">${ic}${_esc(doc.language)}</span>`);
+      // Per-language icon lives in the title row above; just the language
+      // name here keeps the meta line scannable without duplicating the icon.
+      pieces.push(`<span>${_esc(doc.language)}</span>`);
    }
    pieces.push(`<span>${_esc(libraryRelativeTime(doc.updated_at))}</span>`);
    meta.innerHTML = pieces.join('<span style="opacity:0.5;">\u00b7</span>');
-    // Strip the per-language icon from the meta line \u2014 it now sits next to the
-    // title above, so duplicating it here was redundant.
    content.appendChild(meta);
    card.appendChild(content);

@@ -788,7 +788,7 @@ export function openEmailLibrary(opts = {}) {
        <div class="admin-card" style="flex:1;flex-direction:column;display:flex;overflow:hidden;">
          <p class="memory-desc doclib-desc">All emails. Click to open as a document.</p>
          <div class="email-accounts-row">
-            <div id="email-lib-accounts" style="display:flex;gap:4px;flex-wrap:wrap;flex:1;"></div>
+            <div id="email-lib-accounts" style="display:flex;gap:4px;flex:1;min-width:0;"></div>
            <button class="memory-toolbar-btn email-compose-jiggle" id="email-lib-compose-btn">
              <svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" style="vertical-align:-2px;margin-right:3px;"><rect x="2" y="4" width="20" height="16" rx="2"/><path d="m22 7-8.97 5.7a1.94 1.94 0 0 1-2.06 0L2 7"/></svg>
              New
@@ -36,6 +36,17 @@ function linkHtml(text, url) {
  return `<a href="${escapeHtml(safeUrl)}" target="_blank" rel="noopener noreferrer">${safeText}</a>`;
 }

+function _isModelEndpointUrl(rawUrl) {
+  try {
+    const parsed = new URL(String(rawUrl || ''), window.location.origin);
+    if (parsed.protocol !== 'http:' && parsed.protocol !== 'https:') return false;
+    const path = parsed.pathname.replace(/\/+$/, '');
+    return path === '/v1';
+  } catch (_) {
+    return false;
+  }
+}
+
 /**
 * Sanitize the raw-HTML fragments that mdToHtml deliberately preserves from
 * the source text — <details> blocks (collapsible agent output) and <a> tags
@@ -327,6 +338,17 @@ function createThinkingSection(thinkingContent, index = 0, thinkingTime = null)
  `;
 }

+function createTaskCompletedMarker() {
+  return `
+    <div class="task-completed-marker" role="status" aria-label="Task completed">
+      <span class="task-completed-icon" aria-hidden="true">
+        <svg viewBox="0 0 24 24" width="14" height="14" fill="none" stroke="currentColor" stroke-width="2.6" stroke-linecap="round" stroke-linejoin="round"><polyline points="20 6 9 17 4 12"/></svg>
+      </span>
+      <span>Task completed</span>
+    </div>
+  `;
+}
+
 /**
 * Process text and render with thinking sections
 */
@@ -422,6 +444,9 @@ export function processWithThinking(text) {
  const { thinkingBlocks, content, thinkingTime } = extractThinkingBlocks(text);

  let html = '';
+  let visibleContent = content || '';
+  const doneOnly = /^\s*\[DONE\]\s*$/i.test(visibleContent);
+  const hadTrailingDone = !doneOnly && /(?:^|\n)\s*\[DONE\]\s*$/i.test(visibleContent);

  // Add thinking sections (collapsed by default)
  thinkingBlocks.forEach((block, index) => {
@@ -429,8 +454,12 @@ export function processWithThinking(text) {
  });

  // Add the actual content
-  if (content) {
-    html += mdToHtml(content);
+  if (doneOnly) {
+    html += createTaskCompletedMarker();
+  } else {
+    if (hadTrailingDone) visibleContent = visibleContent.replace(/\n?\s*\[DONE\]\s*$/i, '').trimEnd();
+    if (visibleContent) html += mdToHtml(visibleContent);
+    if (hadTrailingDone) html += createTaskCompletedMarker();
  }

  return _useSvgEmoji() ? svgifyEmoji(html) : html;
@@ -885,3 +914,121 @@ document.addEventListener('click', function(e) {
    start();
  }
 })();
+
+function _endpointNameFromUrl(url) {
+  try {
+    const parsed = new URL(url, window.location.origin);
+    return parsed.host || parsed.hostname || 'Model endpoint';
+  } catch (_) {
+    return 'Model endpoint';
+  }
+}
+
+function _appendEndpointAddButtons(root) {
+  if (!root || !root.querySelectorAll) return;
+  const anchors = root.matches?.('a[href]')
+    ? [root]
+    : [...root.querySelectorAll('a[href]')];
+  for (const anchor of anchors) {
+    if (anchor.dataset.endpointAddChecked === '1') continue;
+    anchor.dataset.endpointAddChecked = '1';
+    const href = anchor.getAttribute('href') || '';
+    if (!_isModelEndpointUrl(href)) continue;
+    if (anchor.nextElementSibling?.classList?.contains('model-endpoint-add-btn')) continue;
+
+    const btn = document.createElement('button');
+    btn.type = 'button';
+    btn.className = 'model-endpoint-add-btn';
+    btn.dataset.endpointUrl = new URL(href, window.location.origin).href.replace(/\/+$/, '');
+    btn.title = 'Add this OpenAI-compatible endpoint to the model picker';
+    btn.innerHTML = '<span aria-hidden="true">+</span><span>Add to model picker</span>';
+    anchor.insertAdjacentElement('afterend', btn);
+  }
+}
+
+async function _registerEndpointFromButton(btn) {
+  const baseUrl = String(btn?.dataset?.endpointUrl || '').trim();
+  if (!baseUrl || !_isModelEndpointUrl(baseUrl)) return;
+  const original = btn.innerHTML;
+  btn.disabled = true;
+  btn.innerHTML = '<span aria-hidden="true">...</span><span>Adding</span>';
+  try {
+    const existingRes = await fetch('/api/model-endpoints', { credentials: 'same-origin' });
+    if (existingRes.ok) {
+      const endpoints = await existingRes.json();
+      const existing = Array.isArray(endpoints)
+        ? endpoints.find((ep) => String(ep.base_url || '').replace(/\/+$/, '') === baseUrl)
+        : null;
+      if (existing) {
+        btn.classList.add('added');
+        btn.innerHTML = '<span aria-hidden="true">✓</span><span>Already added</span>';
+        window.dispatchEvent(new CustomEvent('ge:model-endpoints-updated', { detail: { baseUrl } }));
+        if (window.modelsModule?.refreshModels) window.modelsModule.refreshModels(true);
+        if (window.sessionModule?.updateModelPicker) window.sessionModule.updateModelPicker();
+        uiModule.showToast?.(`Already in model picker: ${existing.name || _endpointNameFromUrl(baseUrl)}`);
+        return;
+      }
+    }
+
+    const parsed = new URL(baseUrl, window.location.origin);
+    const fd = new FormData();
+    fd.append('base_url', baseUrl);
+    fd.append('name', _endpointNameFromUrl(baseUrl));
+    fd.append('model_type', 'llm');
+    fd.append('endpoint_kind', 'auto');
+    fd.append('skip_probe', 'true');
+    if (/^(localhost|127\.0\.0\.1|0\.0\.0\.0)$/i.test(parsed.hostname)) {
+      fd.append('container_local', 'true');
+    }
+    const res = await fetch('/api/model-endpoints', {
+      method: 'POST',
+      credentials: 'same-origin',
+      body: fd,
+    });
+    if (!res.ok) {
+      const body = await res.text().catch(() => '');
+      throw new Error(`HTTP ${res.status}${body ? ': ' + body.slice(0, 160) : ''}`);
+    }
+    btn.classList.add('added');
+    btn.innerHTML = '<span aria-hidden="true">✓</span><span>Added</span>';
+    window.dispatchEvent(new CustomEvent('ge:model-endpoints-updated', { detail: { baseUrl } }));
+    if (window.modelsModule?.refreshModels) await window.modelsModule.refreshModels(true);
+    if (window.sessionModule?.updateModelPicker) window.sessionModule.updateModelPicker();
+    uiModule.showToast?.(`Model endpoint added: ${_endpointNameFromUrl(baseUrl)}`);
+  } catch (err) {
+    btn.disabled = false;
+    btn.innerHTML = original;
+    uiModule.showError?.(`Add endpoint failed: ${err.message || err}`);
+  }
+}
+
+(function _watchModelEndpointLinks() {
+  if (window._modelEndpointLinkWatcherWired) return;
+  window._modelEndpointLinkWatcherWired = true;
+
+  document.addEventListener('click', (e) => {
+    const btn = e.target.closest?.('.model-endpoint-add-btn');
+    if (!btn) return;
+    e.preventDefault();
+    e.stopPropagation();
+    _registerEndpointFromButton(btn);
+  });
+
+  const start = () => {
+    const root = document.body;
+    if (!root) return;
+    _appendEndpointAddButtons(root);
+    new MutationObserver((mutations) => {
+      for (const m of mutations) {
+        for (const node of m.addedNodes) {
+          if (node.nodeType === 1) _appendEndpointAddButtons(node);
+        }
+      }
+    }).observe(root, { childList: true, subtree: true });
+  };
+  if (document.readyState === 'loading') {
+    document.addEventListener('DOMContentLoaded', start, { once: true });
+  } else {
+    start();
+  }
+})();
@@ -327,13 +327,10 @@ function _initModelPickerDropdown() {
      // hover so the suffix/variant tag is still discoverable (#1982).
      nameSpan.title = m.display;
      row.appendChild(nameSpan);
-      if (m.stale) {
-        const badge = document.createElement('span');
-        badge.className = 'model-switch-stale-badge';
-        badge.textContent = 'offline';
-        badge.style.cssText = 'font-size:10px;opacity:0.7;padding:1px 6px;border:1px solid var(--border);border-radius:8px;margin-left:6px;';
-        row.appendChild(badge);
-      }
+      // Offline state is already conveyed by the row's reduced opacity —
+      // a redundant "offline" pill on top of that just added clutter.
+      // (Class kept on `row` so the opacity rule still applies; the text
+      // badge is gone.)
      const epSpan = document.createElement('span');
      epSpan.className = 'model-switch-ep';
      // Don't show endpoint name if it matches the model name (local self-hosted)
@@ -178,7 +178,14 @@ export async function refreshModels(force = false) {
    _loadingSpinner.start();
    try {
      if (!_fetchInflight) {
-        _fetchInflight = fetch(`${API_BASE}/api/models`, { credentials: 'same-origin' })
+        // Pass ?refresh=true on forced refreshes so the BACKEND's 30s
+        // per-user cache also gets bypassed. Without this, `force=true`
+        // only clears the frontend cache and the same stale list comes
+        // back — newly-served endpoints don't appear until the cache
+        // ages out. (Bug repro: serve a model, picker is empty for ~30s
+        // even though the endpoint is in the DB and online.)
+        const _url = `${API_BASE}/api/models` + (force ? '?refresh=true' : '');
+        _fetchInflight = fetch(_url, { credentials: 'same-origin' })
          .then(async (res) => {
            if (!res.ok) throw new Error(`HTTP ${res.status}`);
            return res.json();
@@ -1559,6 +1559,7 @@ async function initResearchSearchSettings() {
 async function initAgentSettings() {
  var toolsInput = el('set-agentMaxTools');
  var roundsInput = el('set-agentMaxRounds');
+  var supInput = el('set-agentSupervisorLadder');
  var msg = el('set-agentMsg');
  if (!toolsInput) return;

@@ -1567,6 +1568,7 @@ async function initAgentSettings() {
    var settings = await res.json();
    if (settings.agent_max_tool_calls) toolsInput.value = settings.agent_max_tool_calls;
    if (roundsInput && settings.agent_max_rounds) roundsInput.value = settings.agent_max_rounds;
+    if (supInput) supInput.checked = !!settings.agent_supervisor_ladder;
  } catch (e) {}

  // Clamp + coerce a raw input to an int in [lo, hi]; falls back to `dflt`
@@ -1584,23 +1586,27 @@ async function initAgentSettings() {
    if (roundsInput) roundsInput.value = rounds;
    var payload = { agent_max_tool_calls: tools };
    if (rounds != null) payload.agent_max_rounds = rounds;
+    if (supInput) payload.agent_supervisor_ladder = !!supInput.checked;
    try {
      await fetch('/api/auth/settings', { method: 'POST', credentials: 'same-origin',
        headers: { 'Content-Type': 'application/json' },
        body: JSON.stringify(payload)
      });
      msg.textContent = (tools > 0 ? 'Limit: ' + tools + ' tool calls' : 'Unlimited tool calls') +
-        (rounds != null ? ' · ' + rounds + ' steps/message' : '');
+        (rounds != null ? ' · ' + rounds + ' steps/message' : '') +
+        (supInput && supInput.checked ? ' · supervisor on' : '');
      msg.style.color = 'var(--fg)';
    } catch (e) { msg.textContent = 'Failed to save'; msg.style.color = 'var(--red)'; }
  }

  toolsInput.addEventListener('change', save);
  if (roundsInput) roundsInput.addEventListener('change', save);
+  if (supInput) supInput.addEventListener('change', save);
  var cur = parseInt(toolsInput.value, 10) || 0;
  var curR = roundsInput ? (parseInt(roundsInput.value, 10) || 20) : null;
  msg.textContent = (cur > 0 ? 'Limit: ' + cur + ' tool calls' : 'Unlimited tool calls') +
-    (curR != null ? ' · ' + curR + ' steps/message' : '');
+    (curR != null ? ' · ' + curR + ' steps/message' : '') +
+    (supInput && supInput.checked ? ' · supervisor on' : '');
 }

 /* ═══════════════════════════════════════════
@@ -890,10 +890,10 @@ function renderSkillsList() {
    });
  }

-  // Background-load the visible skills' SKILL.md so expanding any of them is
-  // instant (no first-time async fetch → no jump). Deferred so it never
-  // competes with the render/cascade paint.
-  setTimeout(_preloadVisibleMarkdown, 0);
+  // Do not eager-load every visible SKILL.md. On large skill libraries this
+  // creates dozens of simultaneous /api/skills/<name>/markdown requests during
+  // app startup and can peg uvicorn. Markdown is fetched lazily when a card is
+  // expanded.
 }

 // ---- Card expand / edit / actions ----