mlx-bun 0.0.5 → 0.0.7

package/src/web/app.html

@@ -105,6 +105,39 @@
   #chat-think.think-on .dot{background:var(--blue);box-shadow:0 0 11px var(--blue)}
   .pill.bad .dot{background:var(--red);box-shadow:0 0 11px var(--red)}
   .pill.warn .dot{background:var(--orange);box-shadow:0 0 11px var(--orange)}
+  /* sampling disclosure: a small pill that opens a popover above the composer */
+  #chat-sampling-wrap{position:relative;display:inline-flex}
+  #chat-sampling{cursor:pointer;gap:6px;padding:5px 11px;font-size:12px;user-select:none;-webkit-user-select:none;
+    transition:color .22s var(--ease),border-color .22s var(--ease)}
+  #chat-sampling:hover{border-color:rgba(255,255,255,.28);color:var(--ink)}
+  #chat-sampling.on{color:var(--ink);border-color:rgba(255,255,255,.28)}
+  #chat-sampling.dirty .dot{background:var(--blue);box-shadow:0 0 11px var(--blue)}
+  #chat-sampling .dot{width:7px;height:7px}
+  .samplepop{position:absolute;bottom:calc(100% + 9px);right:0;z-index:40;width:300px;
+    background:rgba(30,30,34,.98);border:1px solid rgba(255,255,255,.12);border-radius:13px;padding:13px 14px;
+    backdrop-filter:blur(22px) saturate(1.3);-webkit-backdrop-filter:blur(22px) saturate(1.3);
+    box-shadow:0 18px 50px rgba(0,0,0,.7);display:none;
+    transform:translateY(4px);opacity:0;transition:opacity .16s var(--ease),transform .16s var(--ease)}
+  .samplepop.open{display:block;transform:translateY(0);opacity:1}
+  .samplepop h4{font-size:12px;font-weight:600;color:var(--dim);letter-spacing:.02em;
+    text-transform:uppercase;margin:0 0 10px}
+  .samprow{display:flex;align-items:center;gap:9px;margin:0 0 11px}
+  .samprow label{font-size:12.5px;color:var(--ink);width:74px;flex:0 0 auto}
+  .samprow input[type=range]{flex:1 1 auto;-webkit-appearance:none;appearance:none;height:4px;margin:0;
+    border-radius:3px;background:rgba(255,255,255,.15);outline:none;cursor:pointer}
+  .samprow input[type=range]::-webkit-slider-thumb{-webkit-appearance:none;appearance:none;width:15px;height:15px;
+    border-radius:50%;background:var(--ink);border:0;box-shadow:0 1px 4px rgba(0,0,0,.5);cursor:pointer;
+    transition:background .15s var(--ease)}
+  .samprow input[type=range]:hover::-webkit-slider-thumb,
+  .samprow input[type=range]:focus::-webkit-slider-thumb{background:var(--blue)}
+  .samp-val{width:40px;flex:0 0 auto;text-align:right;font-family:var(--mono);font-size:12px;color:var(--ink)}
+  .samp-val.auto{color:var(--dimmer)}
+  .samppop-foot{display:flex;justify-content:space-between;align-items:center;margin-top:3px}
+  .samppop-foot .note{font-size:10.5px;color:var(--dimmer);line-height:1.3;flex:1 1 auto;padding-right:8px}
+  .samppop-foot button{appearance:none;cursor:pointer;background:var(--card);border:1px solid var(--hairline);
+    color:var(--dim);border-radius:8px;padding:5px 11px;font-size:11.5px;font-weight:600;flex:0 0 auto;
+    transition:color .18s var(--ease),border-color .18s var(--ease)}
+  .samppop-foot button:hover{color:var(--ink);border-color:rgba(255,255,255,.28)}
   #nav-download{color:var(--ink);font-family:var(--mono);font-size:11.5px;font-weight:500}
   #nav-download .dot{background:var(--orange);box-shadow:0 0 11px var(--orange);animation:dlpulse 1.4s ease-in-out infinite}
   @keyframes dlpulse{0%,100%{opacity:.4}50%{opacity:1}}
@@ -485,6 +518,7 @@
     <a class="tab" href="#/dataset" data-tab="dataset">Build Dataset</a>
     <a class="tab" href="#/status" data-tab="status">Status</a>
     <a class="tab" href="/curves" data-tab="curves">Curves</a>
+    <a class="tab" href="#/routes" data-tab="routes">Routes</a>
   </div>
   <div class="right">
     <button class="hf-gear" id="nav-hf" title="Hugging Face token settings" aria-label="Hugging Face settings">
@@ -552,6 +586,32 @@
         <div class="chat-hint">
           <span id="chat-status-line">Enter to send · Shift+Enter for a new line · type while it streams to steer</span>
           <span class="hint-right">
+            <span id="chat-sampling-wrap">
+              <span id="chat-sampling" class="pill" role="button" tabindex="0" aria-haspopup="true" aria-expanded="false"
+                    title="Sampling controls (temperature · top_p · top_k)"><span class="dot"></span>Sampling</span>
+              <div id="chat-sampling-pop" class="samplepop" role="dialog" aria-label="Sampling controls">
+                <h4>Sampling</h4>
+                <div class="samprow">
+                  <label for="samp-temp">Temperature</label>
+                  <input id="samp-temp" type="range" min="0" max="2" step="0.05" aria-label="Temperature">
+                  <output id="samp-temp-val" class="samp-val auto">0.70</output>
+                </div>
+                <div class="samprow">
+                  <label for="samp-topp">top_p</label>
+                  <input id="samp-topp" type="range" min="0" max="1" step="0.01" aria-label="top_p">
+                  <output id="samp-topp-val" class="samp-val auto">0.95</output>
+                </div>
+                <div class="samprow">
+                  <label for="samp-topk">top_k</label>
+                  <input id="samp-topk" type="range" min="0" max="100" step="1" aria-label="top_k">
+                  <output id="samp-topk-val" class="samp-val auto">off</output>
+                </div>
+                <div class="samppop-foot">
+                  <span class="note">Sliders default to the recommended value — temperature follows the thinking toggle. Reset returns to defaults.</span>
+                  <button id="samp-reset" type="button">Reset</button>
+                </div>
+              </div>
+            </span>
             <select id="chat-adapter" class="pill" title="LoRA adapter for this chat (none = base model)">
               <option value="">no adapter</option>
             </select>
@@ -581,10 +641,14 @@
         <p class="lead" style="margin:6px 0 22px;font-size:14.5px">Point at a Hugging Face repo id or a local snapshot path. Inspect first to confirm support and size.</p>
         <div class="field">
           <label>Model id or path</label>
-          <input type="text" id="q-model" placeholder="mlx-community/Qwen3-8B-bf16" autocomplete="off">
-          <div class="hint">Anything on huggingface.co, or an absolute path to a local snapshot.</div>
+          <div id="q-drop">
+            <input type="text" id="q-model" placeholder="mlx-community/Qwen3-8B-bf16" autocomplete="off">
+          </div>
+          <input type="file" id="q-folder" webkitdirectory directory style="display:none">
+          <div class="hint">Anything on huggingface.co, an absolute path to a local snapshot, or <strong>choose / drop a folder</strong> below — we read its <code>config.json</code> and find the snapshot on disk.</div>
         </div>
         <div class="btnrow">
+          <button class="btn ghost" id="q-browse">Choose folder…</button>
           <button class="btn ghost" id="q-inspect">Inspect</button>
         </div>
         <div id="q-inspect-out"></div>
@@ -643,7 +707,7 @@
         <p class="lead" style="margin:6px 0 18px;font-size:14.5px">Your quantized model is written to disk.</p>
         <div class="field"><label>Output directory</label>
           <pre class="logbox" id="q-out" style="max-height:none;margin-top:0"></pre></div>
-        <div class="flash ok">Run <code>mlx-bun scan</code> (or it's auto-picked up on restart) to make this model serveable from the Library.</div>
+        <div class="flash ok">Written to your model cache and registered automatically — it's already serveable from the Library and the model picker.</div>
         <div class="grid g4" style="margin-top:14px">
           <div class="card"><h3>Push to Hugging Face</h3><div class="cap">Publish this quantized model to a Hugging Face repo.</div>
             <div class="btnrow" style="margin-top:14px"><button class="btn ghost sm" id="q-push">Push to HF</button></div>
@@ -671,10 +735,16 @@
         <p class="lead" style="margin:6px 0 22px;font-size:14.5px">The model you're adapting. Point at a local model directory or snapshot path.</p>
         <div class="field">
           <label>Model directory / path</label>
-          <input type="text" id="f-model" placeholder="/abs/path/to/model  or  mlx-community/Qwen3-8B-4bit" autocomplete="off">
-          <div class="hint">An OptiQ quant or any supported local model works.</div>
+          <div id="f-drop">
+            <input type="text" id="f-model" placeholder="/abs/path/to/model  or  mlx-community/Qwen3-8B-4bit" autocomplete="off">
+          </div>
+          <input type="file" id="f-folder" webkitdirectory directory style="display:none">
+          <div class="hint">An OptiQ quant or any supported local model works — type a path/repo id, or <strong>choose / drop a folder</strong> below.</div>
+        </div>
+        <div class="btnrow">
+          <button class="btn ghost" id="f-browse">Choose folder…</button>
+          <button class="btn primary" id="f-next0">Continue</button>
         </div>
-        <div class="btnrow"><button class="btn primary" id="f-next0">Continue</button></div>
       </div>
 
       <!-- step 2: dataset -->
@@ -962,6 +1032,12 @@
     </div></div>
   </section>
 
+  <!-- ═══════════════════ ROUTES ═══════════════════ -->
+  <section data-route="routes" id="s-routes">
+    <iframe src="/dag" title="Training · inference route map"
+            style="width:100%;height:100%;border:0;display:block"></iframe>
+  </section>
+
 </div>
 
 <div id="toasts"></div>
@@ -1039,7 +1115,7 @@ function mdInline(t) {
   // Pull code spans out first (any run of backticks), then escape the rest.
   t = String(t).replace(/(`+)([\s\S]+?)\1/g, (_m, _ticks, c) => {
     codes.push(c.replace(/^ | $/g, ""));
-    return "C" + (codes.length - 1) + "";
+    return "\u0000C" + (codes.length - 1) + "\u0000";
   });
   let s = esc(t);
   // [text](url "title") — title optional and ignored.
@@ -1057,7 +1133,7 @@ function mdInline(t) {
   // Autolink bare URLs that markdown didn't already wrap.
   s = s.replace(/(^|[\s(])(https?:\/\/[^\s<)]+[^\s<).,;:!?'"])/g,
     (_m, pre, url) => pre + '<a href="' + url + '" target="_blank" rel="noopener noreferrer">' + url + "</a>");
-  return s.replace(/C(\d+)/g, (_m, i) => "<code>" + esc(codes[+i]) + "</code>");
+  return s.replace(/\u0000C(\d+)\u0000/g, (_m, i) => "<code>" + esc(codes[+i]) + "</code>");
 }
 
 function mdCodeBlock(lang, code) {
@@ -1288,7 +1364,7 @@ async function pushToHub(panel, { kind, job_id, source_path }) {
 /* ════════════════════════════════════════════════════════════════════
    ROUTER  — toggles section[data-route]; lazily inits each controller.
    ════════════════════════════════════════════════════════════════════ */
-const ROUTES = ["chat", "quantize", "finetune", "dataset", "status"];
+const ROUTES = ["chat", "quantize", "finetune", "dataset", "status", "routes"];
 const inited = {};
 const controllers = {}; // name -> { init, enter, leave }
 
@@ -1431,7 +1507,7 @@ controllers.chat = (() => {
     el("div", "who", m).textContent = "local agent";
     const bubble = el("div", "bubble", m);
     const thinkBox = el("details", "thinkbox", bubble);
-    thinkBox.open = true;
+    thinkBox.open = false; // collapsed by default; the reasoning stays tucked behind "Thinking" — click to expand
     el("summary", "", thinkBox).textContent = "Thinking";
     const thinkBody = el("div", "thinkbody", thinkBox);
     thinkBox.style.display = "none";
@@ -1738,7 +1814,15 @@ controllers.chat = (() => {
   let attachments = [];       // { id, kind:"image"|"text", name, mimeType, data, text, truncated }
   let visionCapable = false;  // from the `ready` frame; gates image attachments
   let thinkingCapable = false; // from `ready`; gates the reasoning toggle
-  let thinkingOn = true;       // on/off state for the model's reasoning channel
+  let thinkingOn = true;       // default on: with tools (read/web_search) the reasoning step gives the small model far better tool judgment (answers math/writing itself, only searches when needed). The thinking box defaults collapsed. Users can toggle it off.
+  // Per-request sampling. Each field is either a user override (a number) or
+  // null = "use the recommended default" (resolved server-side). The recommended
+  // TEMPERATURE tracks the thinking toggle (0.7 off ↔ 0.9 on); top_p/top_k recs
+  // are fixed mirrors of the server defaults (display only). The placeholder in
+  // each input shows the live recommended value when the field isn't overridden.
+  const sampling = { temperature: null, top_p: null, top_k: null };
+  const SAMP_REC = { topP: 0.95, topK: 0 }; // fixed recommended top_p / top_k (server genDefaults)
+  function recTemp() { return thinkingOn ? 0.9 : 0.7; } // mirrors server toOptions defaultTemp
   let attachSeq = 0;
   const MAX_TEXT_BYTES = 256 * 1024;
   const TEXT_EXTS = /\.(txt|md|markdown|csv|tsv|json|jsonl|ya?ml|toml|ini|cfg|conf|log|xml|html?|css|scss|js|mjs|cjs|ts|tsx|jsx|py|rs|go|java|c|h|cpp|hpp|cc|cs|rb|php|sh|bash|zsh|fish|sql|swift|kt|lua|r|jl|tex|svg|env|gitignore|dockerfile|makefile)$/i;
@@ -1835,6 +1919,81 @@ controllers.chat = (() => {
     b.classList.toggle("think-on", thinkingOn);
     b.title = thinkingOn ? "Reasoning on — click to answer directly" : "Reasoning off — click to let the model think";
     b.setAttribute("aria-pressed", String(thinkingOn));
+    refreshSamplingRecs(); // recommended temperature follows the thinking state
+  }
+
+  /* — sampling controls — */
+  // Each row is a slider + a live value readout. A field is "auto" (null
+  // override → the server uses its mode-aware default; temperature's default
+  // follows the thinking toggle) until the user drags it, which records an
+  // explicit override. Reset returns every field to auto.
+  const SAMP_FIELDS = [
+    { id: "samp-temp", field: "temperature", valId: "samp-temp-val", rec: () => recTemp(), fmt: (v) => v.toFixed(2) },
+    { id: "samp-topp", field: "top_p", valId: "samp-topp-val", rec: () => SAMP_REC.topP, fmt: (v) => v.toFixed(2) },
+    { id: "samp-topk", field: "top_k", valId: "samp-topk-val", rec: () => SAMP_REC.topK, fmt: (v) => (v >= 1 ? String(Math.round(v)) : "off") },
+  ];
+
+  // Sync every slider + readout to current state: an overridden field shows its
+  // set value (bright readout); an auto field shows the recommended value
+  // (dimmed) and tracks it live — e.g. temperature when the thinking toggle
+  // flips. Called on thinking-toggle flips, popover open, and reset.
+  function refreshSamplingRecs() {
+    for (const f of SAMP_FIELDS) {
+      const slider = $(f.id), val = $(f.valId); if (!slider || !val) continue;
+      const over = sampling[f.field];
+      const v = over != null ? over : f.rec();
+      slider.value = String(v);
+      val.textContent = f.fmt(v);
+      val.classList.toggle("auto", over == null);
+    }
+  }
+
+  // User dragged a slider → record the explicit override, brighten the readout,
+  // and push the whole sampling state to the server.
+  function onSlide(f) {
+    const slider = $(f.id); if (!slider) return;
+    const v = Number(slider.value);
+    sampling[f.field] = v;
+    const val = $(f.valId); if (val) { val.textContent = f.fmt(v); val.classList.remove("auto"); }
+    updateSamplingUi();
+    send({ type: "set_sampling", temperature: sampling.temperature, top_p: sampling.top_p, top_k: sampling.top_k });
+  }
+
+  // The pill shows a lit dot whenever any field is overridden.
+  function updateSamplingUi() {
+    const dirty = sampling.temperature != null || sampling.top_p != null || sampling.top_k != null;
+    const pill = $("chat-sampling");
+    if (pill) { pill.classList.toggle("dirty", dirty); pill.title = dirty
+      ? "Sampling overridden — click to edit or reset" : "Sampling controls (temperature · top_p · top_k)"; }
+  }
+
+  function resetSampling() {
+    sampling.temperature = null; sampling.top_p = null; sampling.top_k = null;
+    refreshSamplingRecs(); updateSamplingUi();
+    send({ type: "set_sampling", temperature: null, top_p: null, top_k: null });
+  }
+  function initSampling() {
+    const pill = $("chat-sampling"), pop = $("chat-sampling-pop");
+    if (!pill || !pop) return;
+    const setOpen = (open) => {
+      pop.classList.toggle("open", open);
+      pill.classList.toggle("on", open);
+      pill.setAttribute("aria-expanded", String(open));
+      if (open) refreshSamplingRecs();
+    };
+    const toggle = () => setOpen(!pop.classList.contains("open"));
+    pill.onclick = (e) => { e.stopPropagation(); toggle(); };
+    pill.addEventListener("keydown", (e) => { if (e.key === "Enter" || e.key === " ") { e.preventDefault(); toggle(); } });
+    pop.addEventListener("click", (e) => e.stopPropagation());
+    // Click-away / Escape closes the popover.
+    document.addEventListener("click", () => setOpen(false));
+    document.addEventListener("keydown", (e) => { if (e.key === "Escape") setOpen(false); });
+    for (const f of SAMP_FIELDS) {
+      const e = $(f.id); if (!e) continue;
+      e.addEventListener("input", () => onSlide(f));
+    }
+    const reset = $("samp-reset"); if (reset) reset.onclick = (e) => { e.stopPropagation(); resetSampling(); };
+    updateSamplingUi(); refreshSamplingRecs();
   }
 
   /* — composer — */
@@ -1854,8 +2013,10 @@ controllers.chat = (() => {
     addUserMsg(text, attachments);
     box.value = ""; box.style.height = "auto";
     const frame = imgs.length ? { text: combined, images: imgs } : { text: combined };
-    if (turnActive) send({ type: "steer", ...frame });   // mid-turn: steer/queue
-    else send({ type: "prompt", ...frame });
+    // A composer submission is always a new user turn. If the model is still
+    // streaming, the server passes it to pi with streamingBehavior:"followUp"
+    // so it becomes the next turn instead of steering/mutating the current one.
+    send({ type: "prompt", ...frame });
     clearAttachments();
   }
   const scroll = () => $("chat-scroll");
@@ -1872,6 +2033,7 @@ controllers.chat = (() => {
       $("chat-think").addEventListener("keydown", (e) => { if (e.key === "Enter" || e.key === " ") { e.preventDefault(); thinkToggle(); } });
       const adapterSel = $("chat-adapter");
       if (adapterSel) adapterSel.onchange = () => onSelectAdapter(adapterSel);
+      initSampling();
       refreshAdapters();
       const box = $("chat-box");
       box.addEventListener("keydown", (e) => { if (e.key === "Enter" && !e.shiftKey) { e.preventDefault(); submit(); } });
@@ -1934,6 +2096,33 @@ controllers.quantize = (() => {
       }));
   }
 
+  // --- OS folder picker / drag-drop → real on-disk snapshot path ---
+  // The browser can't reveal an absolute path, and the HF cache uses symlinked
+  // snapshots over a blobs/ store — so we DON'T read any files. We only take the
+  // picked folder's NAME (which encodes the repo id) and let the server
+  // reconstruct the snapshot path it owns. No upload, no "config.json" hunt.
+  async function resolveByName(name, relPath) {
+    if (!name && !relPath) return;
+    const out = $("q-inspect-out");
+    out.innerHTML = '<div class="flash"><span class="shimmer">locating on disk…</span></div>';
+    const d = await api("/api/quantize/resolve-folder", { method: "POST", body: { folder_name: name || "", rel_path: relPath || "" } })
+      .catch(() => ({ ok: false, error: "request failed" }));
+    if (!d.ok) { out.innerHTML = '<div class="flash err">' + esc(d.error || "could not locate this folder on the server") + "</div>"; return; }
+    $("q-model").value = d.path;
+    inspect();
+  }
+  async function pickFolder() {
+    // Native OS folder dialog with no "upload N files" prompt; we only read the
+    // folder's .name, never its contents.
+    if (window.showDirectoryPicker) {
+      try {
+        const handle = await window.showDirectoryPicker({ id: "mlx-models", mode: "read" });
+        return resolveByName(handle.name);
+      } catch (e) { return; } // user cancelled / denied
+    }
+    $("q-folder").click(); // fallback for browsers without the File System Access API
+  }
+
   async function inspect() {
     const id = $("q-model").value.trim();
     if (!id) { toast("Enter a model id or path first.", "err"); return; }
@@ -1990,6 +2179,26 @@ controllers.quantize = (() => {
       show(0);
       $("q-inspect").onclick = inspect;
       $("q-model").addEventListener("keydown", (e) => { if (e.key === "Enter") inspect(); });
+      // OS folder dialog (native; resolves by folder name, no upload)
+      $("q-browse").onclick = pickFolder;
+      $("q-folder").addEventListener("change", (e) => {
+        const f = e.target.files[0]; e.target.value = "";
+        if (f) resolveByName((f.webkitRelativePath || f.name).split("/")[0], f.webkitRelativePath || "");
+      });
+      // drag a folder onto the box — use the directory entry's NAME (no read)
+      const drop = $("q-drop");
+      const hi = (on) => { drop.style.outline = on ? "2px dashed var(--blue)" : ""; drop.style.outlineOffset = on ? "4px" : ""; };
+      ["dragenter", "dragover"].forEach((ev) => drop.addEventListener(ev, (e) => { e.preventDefault(); hi(true); }));
+      ["dragleave", "drop"].forEach((ev) => drop.addEventListener(ev, (e) => {
+        e.preventDefault(); hi(false);
+        if (ev !== "drop" || !e.dataTransfer) return;
+        for (const it of [...(e.dataTransfer.items || [])]) {
+          const en = it.webkitGetAsEntry && it.webkitGetAsEntry();
+          if (en && en.isDirectory) return resolveByName(en.name);
+        }
+        const f = e.dataTransfer.files && e.dataTransfer.files[0];
+        if (f) resolveByName((f.webkitRelativePath || f.name).split("/")[0]);
+      }));
       seg($("q-bits"), (v) => { bits = +v; $("q-bits-hint").textContent = v === "4"
         ? "4-bit · ~4× smaller than bf16 · the recommended default for local serving."
         : "8-bit · ~2× smaller than bf16 · highest fidelity quant."; });
@@ -2009,7 +2218,7 @@ controllers.quantize = (() => {
    ════════════════════════════════════════════════════════════════════ */
 controllers.finetune = (() => {
   const STEPS = ["Base", "Dataset", "Hyperparams", "Train", "Done"];
-  let step = 0, method = "sft", lrTouched = false, es = null, log = [];
+  let step = 0, method = "sft", lrTouched = false, rankTouched = false, es = null, log = [];
   let datasetOk = false, adapterPath = "";
   const trainPts = [], valPts = [];
 
@@ -2019,6 +2228,23 @@ controllers.finetune = (() => {
     renderSteps($("f-steps"), STEPS, n);
   }
 
+  // OS folder picker for the base model — resolve by NAME server-side (no
+  // upload, works with the HF cache's symlinked layout). Mirrors Quantize.
+  async function resolveBaseByName(name, relPath) {
+    if (!name && !relPath) return;
+    const d = await api("/api/model/resolve-folder", { method: "POST", body: { folder_name: name || "", rel_path: relPath || "" } })
+      .catch(() => ({ ok: false, error: "request failed" }));
+    if (!d.ok) { toast(d.error || "could not locate this folder", "err"); return; }
+    $("f-model").value = d.path;
+  }
+  async function pickBaseFolder() {
+    if (window.showDirectoryPicker) {
+      try { const h = await window.showDirectoryPicker({ id: "mlx-models", mode: "read" }); return resolveBaseByName(h.name); }
+      catch (e) { return; }
+    }
+    $("f-folder").click();
+  }
+
   async function inspectDataset() {
     const path = $("f-data").value.trim();
     if (!path) { toast("Enter a dataset directory.", "err"); return; }
@@ -2043,6 +2269,8 @@ controllers.finetune = (() => {
         ? "Odds Ratio Preference Optimization (reference-free) — data is {prompt, chosen, rejected}."
         : "Supervised fine-tuning — data is messages or prompt+completion.";
     if (!lrTouched) $("f-lr").value = m === "dpo" ? 0.00005 : m === "orpo" ? 0.00001 : 0.0002;
+    // ORPO's validated LoRA recipe uses rank 16 (vs 8 for SFT); apply unless edited.
+    if (!rankTouched) $("f-rank").value = m === "orpo" ? 16 : 8;
   }
 
   function collectHP() {
@@ -2178,12 +2406,34 @@ controllers.finetune = (() => {
     init() {
       show(0);
       $("f-next0").onclick = () => { if (!$("f-model").value.trim()) { toast("Enter a base model path.", "err"); return; } show(1); };
+      // OS folder dialog for the base model (native; resolves by name, no upload)
+      $("f-browse").onclick = pickBaseFolder;
+      $("f-folder").addEventListener("change", (e) => {
+        const f = e.target.files[0]; e.target.value = "";
+        if (f) resolveBaseByName((f.webkitRelativePath || f.name).split("/")[0], f.webkitRelativePath || "");
+      });
+      {
+        const drop = $("f-drop");
+        const hi = (on) => { drop.style.outline = on ? "2px dashed var(--blue)" : ""; drop.style.outlineOffset = on ? "4px" : ""; };
+        ["dragenter", "dragover"].forEach((ev) => drop.addEventListener(ev, (e) => { e.preventDefault(); hi(true); }));
+        ["dragleave", "drop"].forEach((ev) => drop.addEventListener(ev, (e) => {
+          e.preventDefault(); hi(false);
+          if (ev !== "drop" || !e.dataTransfer) return;
+          for (const it of [...(e.dataTransfer.items || [])]) {
+            const en = it.webkitGetAsEntry && it.webkitGetAsEntry();
+            if (en && en.isDirectory) return resolveBaseByName(en.name);
+          }
+          const f = e.dataTransfer.files && e.dataTransfer.files[0];
+          if (f) resolveBaseByName((f.webkitRelativePath || f.name).split("/")[0]);
+        }));
+      }
       $("f-inspect").onclick = inspectDataset;
       $("f-data").addEventListener("keydown", (e) => { if (e.key === "Enter") inspectDataset(); });
       $("f-data").addEventListener("input", () => { datasetOk = false; $("f-next1").disabled = true; $("f-inspect-out").innerHTML = ""; });
       $("f-next1").onclick = () => { if (datasetOk) show(2); };
       $("f-method").querySelectorAll("button").forEach((b) => b.onclick = () => setMethod(b.dataset.v));
       $("f-lr").addEventListener("input", () => { lrTouched = true; });
+      $("f-rank").addEventListener("input", () => { rankTouched = true; });
       document.querySelectorAll("#s-finetune [data-fback]").forEach((b) => b.onclick = () => show(step - 1));
       $("f-submit").onclick = submit;
       $("f-merge-go").onclick = mergeAdapters;

package/src/web/skills/memory/SKILL.md

@@ -0,0 +1,97 @@
+---
+name: memory
+description: Use when the answer depends on the user's prior conversations or personal context: explicit memory/recall requests, named personal projects, people, preferences, decisions, or history not explained in the current chat. Do not use for weather, current public facts, generic web research, or ordinary coding/file tasks when current files are sufficient. Also use when the user wants to set up, inspect, or manage memory.
+---
+
+# Memory
+
+The user has a personal **memory** — a wiki of Markdown articles synthesized
+from their past conversations, stored locally at `~/.mlx-bun/wiki/`. User memory
+articles live under `articles/`. Read-only mlx-bun reference docs live under
+`Reference/` as symlinks and are also searchable/readable, but synthesis never
+writes them. Each user article is the consolidated, finished understanding of one
+topic. Your job is to *use it* so the user never has to re-explain what they've
+already told you.
+
+It is entirely local: the articles, the synthesis, and these tools all run on
+the user's machine. Nothing about their memory leaves the device.
+
+## Consulting memory (the common case)
+
+When the answer depends on personal continuity — an explicit memory/recall
+request, or a named project, person, decision, preference, or history that is
+not explained in the current chat — consult memory before answering or asking
+the user to re-explain.
+
+Do **not** use memory for weather, current public facts, generic web research,
+or ordinary coding/file tasks when the current files or the web tools answer the
+question. Memory is for user-specific context, not for performing familiarity.
+
+1. **Search.** Call `memory_search` with the topic (plain phrasing is fine —
+   a name or key noun). It returns ranked articles and sample lines.
+2. **Read enough.** For short or central articles, call `memory_read`. For long
+   articles, use `memory_toc` then `memory_section` to read the relevant part.
+   The article is the answer — read it rather than reconstructing the user's
+   position from scattered fragments. Treat it as their current view.
+3. **Follow nearby context when useful.** Use `memory_links` if backlinks or
+   related concepts matter. `memory_list` gives an overview when search comes up
+   short. mlx-bun reference docs are separate from personal memory; use
+   `reference_search`, `reference_read`, and `reference_list` for those.
+   `memory_status` answers setup/schedule/location questions.
+4. **Use it silently, as a prior.** Let what you find shape your response like
+   a long-time colleague would — don't announce "I searched your memory."
+
+Use memory as an invisible prior: it should narrow relevance and improve the
+answer, not become the topic of the answer. Avoid performative recall — do not
+name-drop unrelated remembered facts to sound familiar.
+
+## Opening memory for the user
+
+The wiki is a normal Markdown/Obsidian vault. If the user asks to browse, open,
+inspect, or edit their memory directly, you may open it for them after a brief
+confirmation:
+
+```bash
+mlx-bun memory open
+mlx-bun memory open <article>
+```
+
+The first form opens `~/.mlx-bun/wiki/` in Obsidian when available, falling back
+to the folder in Finder. The second form opens a specific article stem such as
+`Archie_Project` directly in Obsidian, falling back to the default Markdown app.
+Opening the vault or article is allowed; editing articles during chat is not. If
+the user wants a change to become durable memory, keep the conversation natural
+and let synthesis incorporate it later, unless they explicitly ask to edit files
+by hand in Obsidian.
+
+## Setting up memory (if it doesn't exist yet)
+
+If memory isn't set up (the memory tools aren't available, or the user asks to
+turn it on), explain it plainly and set it up on agreement:
+
+> Your memory is a folder of Markdown articles at `~/.mlx-bun/wiki/`,
+> synthesized from your conversations so I can remember your context across
+> sessions. It's a normal git-tracked wiki you fully own and can edit or
+> delete — and it never leaves your machine.
+
+On a yes, run the setup via the bash tool (the user approves it):
+
+```
+mlx-bun memory init
+```
+
+This creates the vault and walks through optional steps (importing an existing
+wiki, scheduling nightly synthesis). Confirm before each persistent action.
+
+## What not to do
+
+- Don't claim a fact "from memory" that you didn't read in an article — search
+  and read first, or say you don't have it.
+- Don't write to or edit the wiki during a normal conversation; the read tools
+  are read-only by design. Synthesis is a separate, deliberate step.
+- Do open the wiki or a specific article in Obsidian/Finder if the user asks to
+  browse or edit their memory themselves (`mlx-bun memory open [article]`).
+- Don't dump article contents at the user verbatim — use them to inform a
+  direct, relevant answer.
+- Don't perform familiarity. If the user asks for weather, current facts, or a
+  normal coding task, use the task-appropriate tool instead of memory.

package/src/web/skills/web-research/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: web-research
-description: Use when the user asks you to research something, do a deep dive, fact-check a claim, compare options, or answer a question that needs current, accurate, real-world information ("what's the latest on…", prices, releases, news, specs). Gives a rigorous workflow for searching the web, reading primary sources, cross-checking claims, and answering with citations.
+description: Use only when the user asks you to research something, do a deep dive, fact-check a claim, compare options, or answer a question that needs current, accurate, real-world information ("what's the latest on…", prices, releases, news, specs). Do not use for ordinary conversation, ordinary coding tasks, or questions answerable from the current conversation/files.
 ---
 
 # Web Research
@@ -16,6 +16,8 @@ trustworthy, well-sourced answer — not a guess dressed up as one.
 - The user wants a thorough or authoritative answer, or asks for sources.
 - A specific claim needs verifying, or you would otherwise be speculating.
 
+Do **not** use this for ordinary conversation that does not require external information. In that case, reply naturally instead of explaining the research workflow.
+
 ## Workflow
 
 1. **Plan.** Restate the question in your own words and list the 2–4 concrete

package/src/web/skills.ts

@@ -13,6 +13,10 @@
 // the user's personal ~/.pi and project .pi skills, so the chat stays
 // isolated while still getting our curated set (resource-loader.js:277).
 //
+// Memory is deliberately gated by vault existence. If memory is not enabled,
+// we do not expose its skill at all; otherwise small local models can fixate on
+// the skill description even when the user asks unrelated things like weather.
+//
 // To add a skill: drop a folder under ./skills/<name>/SKILL.md, add a text
 // import for it, and append it to BUNDLED_SKILLS.
 
@@ -20,6 +24,7 @@ import { mkdirSync, writeFileSync } from "node:fs";
 import { homedir } from "node:os";
 import { join } from "node:path";
 import webResearchMd from "./skills/web-research/SKILL.md" with { type: "text" };
+import memoryMd from "./skills/memory/SKILL.md" with { type: "text" };
 
 interface BundledSkill {
   /** Directory name under the skills root; should match the SKILL.md `name`. */
@@ -30,15 +35,23 @@ interface BundledSkill {
 
 const BUNDLED_SKILLS: readonly BundledSkill[] = [
   { name: "web-research", md: webResearchMd },
+  { name: "memory", md: memoryMd },
 ];
 
+export interface BundledSkillOptions {
+  /** Include the web-research workflow skill. Default false: tool schemas are enough, and small local models can over-apply workflow skills. */
+  includeWebResearch?: boolean;
+  /** Include the personal-memory skill. Default true for backward compatibility. */
+  includeMemory?: boolean;
+}
+
 /** Where the bundled skills are written; also pi's per-session scratch root. */
 export function bundledSkillsRoot(): string {
   return join(homedir(), ".mlx-bun", "skills");
 }
 
 /** Memoize: the embedded content can't change within a process run. */
-let materialized = false;
+const materialized = new Set<string>();
 
 /**
  * Write the embedded skills to disk (overwriting, so updates propagate) and
@@ -46,14 +59,39 @@ let materialized = false;
  * pi scans it for per-skill SKILL.md files. Runs its disk writes once per
  * process even when called per browser connection.
  */
-export function materializeBundledSkills(): string {
+function selectedSkills(opts: BundledSkillOptions = {}): readonly BundledSkill[] {
+  const includeMemory = opts.includeMemory ?? true;
+  const includeWebResearch = opts.includeWebResearch ?? false;
+  return BUNDLED_SKILLS.filter((skill) => {
+    if (!includeMemory && skill.name === "memory") return false;
+    if (!includeWebResearch && skill.name === "web-research") return false;
+    return true;
+  });
+}
+
+/**
+ * Materialize selected bundled skills and return exact skill directories. Use
+ * these paths in DefaultResourceLoader.additionalSkillPaths. Returning exact
+ * directories (rather than the shared root) prevents an old on-disk memory
+ * skill from being discovered when memory is disabled for the current session.
+ */
+export function materializeBundledSkillPaths(opts: BundledSkillOptions = {}): string[] {
   const root = bundledSkillsRoot();
-  if (materialized) return root;
-  for (const skill of BUNDLED_SKILLS) {
+  const paths: string[] = [];
+  for (const skill of selectedSkills(opts)) {
     const dir = join(root, skill.name);
-    mkdirSync(dir, { recursive: true });
-    writeFileSync(join(dir, "SKILL.md"), skill.md);
+    if (!materialized.has(skill.name)) {
+      mkdirSync(dir, { recursive: true });
+      writeFileSync(join(dir, "SKILL.md"), skill.md);
+      materialized.add(skill.name);
+    }
+    paths.push(dir);
   }
-  materialized = true;
-  return root;
+  return paths;
+}
+
+/** Backward-compatible root materializer for older call sites. Prefer materializeBundledSkillPaths(). */
+export function materializeBundledSkills(): string {
+  materializeBundledSkillPaths({ includeMemory: true, includeWebResearch: true });
+  return bundledSkillsRoot();
 }