dsh-local-models 0.3.1 → 0.4.0

package/README-es.md

@@ -67,7 +67,7 @@ dsh plugin --profile web remove dsh-local-models
 
 ## Usage
 
-1. **Choose GGUF…** — elige un archivo de modelo (atajos Home / Models, navegación con Up).
+1. **Choose GGUF…** — elige un archivo de modelo (botones Home / Root, atajos Models, navegación con Up). Up sube hasta `/`, así que un modelo en un disco montado (`/mnt/…`, `/media/…`) se abre sin registrar antes su carpeta.
 2. Ajusta **context**, **KV cache K / V**, **Max MTP head** (borrador fijo, 0-7; 3 es el punto óptimo ajustado — más profundidad colapsa con ctx grandes), **thinking level** + casilla **preserve thinking**, y los ajustes opcionales de **mmproj** y **MoE**.
 3. **Load model**, observa la tarjeta de estado, inspecciona la salida con **Open terminal**.
 4. **Register in dsh** — la ruta (por defecto `local-<alias>`) aparece en el selector de modelos.

package/README-hi.md

@@ -53,7 +53,7 @@ dsh plugin --profile web remove dsh-local-models
 
 ## Usage
 
-1. **Choose GGUF…** — कोई model फ़ाइल चुनें (Home / Models shortcuts, Up navigation)।
+1. **Choose GGUF…** — कोई model फ़ाइल चुनें (Home / Root buttons, Models shortcuts, Up navigation)। Up सीधे `/` तक जाता है, इसलिए mounted disk पर रखा model (`/mnt/…`, `/media/…`) बिना folder दर्ज किए खुल जाता है।
 2. **context**, **KV cache K / V**, **Max MTP head** (fixed draft, 0-7; 3 ही ट्यून किया गया sweet spot है — इससे गहरा draft बड़े ctx पर collapse हो जाता है), **thinking level** + **preserve thinking** checkbox, वैकल्पिक **mmproj** और **MoE** settings ट्यून करें।
 3. **Load model**, status card देखें, **Open terminal** से output जाँचें।
 4. **Register in dsh** — route (default `local-<alias>`) Models picker में दिखने लगता है।

package/README-pt.md

@@ -67,7 +67,7 @@ dsh plugin --profile web remove dsh-local-models
 
 ## Usage
 
-1. **Choose GGUF…** — escolha um arquivo de modelo (atalhos Home / Models, navegação Up).
+1. **Choose GGUF…** — escolha um arquivo de modelo (botões Home / Root, atalhos Models, navegação Up). O Up sobe até `/`, então um modelo em um disco montado (`/mnt/…`, `/media/…`) abre sem cadastrar a pasta antes.
 2. Ajuste **context**, **KV cache K / V**, **Max MTP head** (draft fixo, 0-7; 3 é o ponto ideal ajustado — profundidades maiores colapsam em ctx grande), **thinking level** + checkbox **preserve thinking**, e as configurações opcionais de **mmproj** e **MoE**.
 3. **Load model**, acompanhe o card de status e inspecione a saída via **Open terminal**.
 4. **Register in dsh** — a rota (padrão `local-<alias>`) aparece no seletor de modelos.

package/README-zh.md

@@ -67,7 +67,7 @@ dsh plugin --profile web remove dsh-local-models
 
 ## Usage
 
-1. **Choose GGUF…** —— 选择模型文件（Home / Models 快捷入口，Up 向上导航）。
+1. **Choose GGUF…** —— 选择模型文件（Home / Root 按钮，Models 快捷入口，Up 向上导航）。Up 可以一直走到 `/`，因此挂载磁盘上的模型（`/mnt/…`、`/media/…`）无需先登记文件夹即可打开。
 2. 调整 **context**、**KV cache K / V**、**Max MTP head**（固定草稿，0-7；3 是调校好的甜点值 —— 更深的草稿在大 ctx 下会崩溃）、**thinking level** + **preserve thinking** 复选框，以及可选的 **mmproj** 和 **MoE** 设置。
 3. **Load model**，观察状态卡片，通过 **Open terminal** 查看输出。
 4. **Register in dsh** —— 该路由（默认 `local-<alias>`）会出现在 Models 选择器中。

package/README.md

@@ -69,7 +69,7 @@ dsh plugin --profile web remove dsh-local-models
 
 ## Usage
 
-1. **Choose GGUF…** — pick a model file (Home / Models shortcuts, Up navigation).
+1. **Choose GGUF…** — pick a model file (Home / Root buttons, Models shortcuts, Up navigation). Up goes all the way to `/`, so a model on a mounted disk (`/mnt/…`, `/media/…`, a network export) opens without adding its folder to Model folders first.
 2. Tune **context**, **KV cache K / V**, **Max MTP head** (fixed draft, 0-7; 3 is the tuned sweet spot — deeper collapses at large ctx), **thinking level** + **preserve thinking** checkbox, optional **mmproj** and **MoE** settings.
 3. **Load model**, watch the status card, inspect output via **Open terminal**.
 4. **Register in dsh** — the route (default `local-<alias>`) appears in the Models picker.

package/lib/client.js

@@ -326,6 +326,11 @@ window.__ModuleLoader__.load({
 			var err = errPair[0], setErr = errPair[1];
 			var selPair = useStatePair(null);
 			var selected = selPair[0], setSelected = selPair[1];
+			// The server stops a huge folder at 2000 entries. Root browsing
+			// makes /proc and /usr/lib reachable, so say so rather than let the
+			// list look complete when it is not.
+			var truncPair = useStatePair(false);
+			var truncated = truncPair[0], setTrunc = truncPair[1];
 			// Remembered folders (persisted): jump buttons + toggle for the current dir.
 			var LS_FOLDERS = "dsh-local-models:remembered-folders";
 			var readFolders = function () {
@@ -355,6 +360,10 @@ window.__ModuleLoader__.load({
 			// empty-looking folder reads as a broken picker.
 			var dirMode = props.dirMode === true;
 			var allFiles = props.allFiles === true || dirMode;
+			// The listing starts at the filesystem root, so the naive
+			// base + "/" + name would ask for "//etc" and paint a path bar
+			// reading "//etc", with every deeper click stacking a slash.
+			var into = function (base, name) { return (base === "/" ? "" : base) + "/" + name; };
 			var load = useCallback(function (d) {
 				setLoading(true);
 				setErr(null);
@@ -364,6 +373,7 @@ window.__ModuleLoader__.load({
 							setDir(body.path || d);
 							setEntries(body.entries);
 							setParent(body.parent != null ? body.parent : null);
+							setTrunc(Boolean(body.truncated));
 							setSelected(null);
 						} else {
 							setErr((body && body.error) || "could not list directory");
@@ -387,6 +397,7 @@ window.__ModuleLoader__.load({
 						"div",
 						{ style: style.row },
 						h("button", { style: style.secondaryButton, onClick: function () { load(""); } }, "Home"),
+						h("button", { style: style.secondaryButton, title: "the filesystem root — mounted disks hang off it (/mnt, /media, /run/media)", onClick: function () { load("/"); } }, "Root"),
 						(props.shortcuts || []).map(function (sc, i) {
 							return h("button", { key: i, style: style.secondaryButton, onClick: function () { load(sc.path); }, title: sc.path }, sc.label);
 						}),
@@ -403,7 +414,7 @@ window.__ModuleLoader__.load({
 						{ style: style.list },
 						(entries || []).map(function (en, i) {
 							if (en.type === "dir") {
-								return h("li", { key: i, style: Object.assign({}, style.item, style.itemDir), onClick: function () { load(dir + "/" + en.name); } }, en.name + "/");
+								return h("li", { key: i, style: Object.assign({}, style.item, style.itemDir), onClick: function () { load(into(dir, en.name)); } }, en.name + "/");
 							}
 							var sel = selected === en.name;
 							return h("li", {
@@ -414,13 +425,14 @@ window.__ModuleLoader__.load({
 							}, en.name + (en.size != null ? "  (" + fmtGiB(en.size) + ")" : ""));
 						}),
 					),
+					truncated ? h("p", { style: style.warn }, "showing the first 2000 entries — open a subfolder to narrow it down") : null,
 					h(
 						"div",
 						{ style: style.row },
 						h("button", {
 							style: (dirMode ? dir : selected) ? style.button : Object.assign({}, style.button, style.disabled),
 							disabled: dirMode ? !dir : !selected,
-							onClick: function () { if (dirMode) { if (dir) props.onPick(dir); } else if (selected) props.onPick(dir + "/" + selected); },
+							onClick: function () { if (dirMode) { if (dir) props.onPick(dir); } else if (selected) props.onPick(into(dir, selected)); },
 						}, dirMode ? "Use this folder" : "Use this file"),
 						h("button", { style: style.secondaryButton, onClick: props.onClose }, "Cancel"),
 					),

package/lib/index.js

@@ -100,11 +100,14 @@ const MMPROJ_CPU = process.env.LOCAL_MODELS_MMPROJ_CPU !== "0";
  * above the raw threshold) — and with xhigh + preserveThinking every turn
  * adds 10-20K thinking tokens, so compaction fires again on the next step:
  * the "compaction loop". The honest fix for mid-size windows is a per-model
- * headroom override in the dsh profile (headroom ~6-16K moves the 131K
- * threshold to ~82-92K), not a smaller O — shrinking O truncates the xhigh
- * thinking blocks this cap exists to protect. Heavy xhigh thinking that
- * needs more must raise per-request maxTokens explicitly (which honestly
- * moves the compaction threshold earlier instead of silently disabling it).
+ * headroom override on the agent-preset row that mounts the running engine
+ * (`preset-standard` → `plugins` → `compaction` group → `compaction-basic`;
+ * headroom ~6-16K moves the 131K threshold to ~82-92K) — a patch on the
+ * top-level `compaction-basic` row is dead config (that row is disabled),
+ * and a smaller O would truncate the xhigh thinking blocks this cap
+ * exists to protect. Heavy xhigh thinking that needs more must raise
+ * per-request maxTokens explicitly (which honestly moves the compaction
+ * threshold earlier instead of silently disabling it).
  */
 const MAX_OUTPUT_TOKENS = 32768;
 /** Default output share of the window: never offer more than half as max
@@ -1563,19 +1566,36 @@ function sendJson(res, code, body) {
 	res.end(payload);
 }
 
+/** The filesystem root, in the form resolve() produces (posix "/"). Every
+ * absolute path is under it, so isPathAllowed() treats it as containing the
+ * whole tree — the one root that makes "Up" from any folder land somewhere
+ * real. */
+const FILESYSTEM_ROOT = resolve(sep);
+
+/** Single wording for a path the file endpoints refuse. With the root in
+ * allowedRoots() this is unreachable for any absolute path; it stays as the
+ * one place the wording lives, so narrowing the roots again has something to
+ * narrow. */
+const NOT_READABLE = "this tab may not read that path";
+
 /** Roots the file endpoints (browse/gguf-meta/run, and profile saves) may
- * touch: the user's home, the browser shortcut dirs (Runtime card, else
- * LOCAL_MODELS_SHORTCUTS), the directories of the models already declared in
- * saved profiles — a saved profile must always stay loadable, whatever the
- * shortcut list says — and the llama.cpp dir when it really holds a
- * llama-server. Resolves lexically and requires containment so inputs can't
- * wander the filesystem.
+ * touch. The list LEADS WITH THE FILESYSTEM ROOT: the picker walks up to "/"
+ * on purpose, so a model can live on a mounted disk the user never declared
+ * here (/mnt/data, /media/<user>/<label>, a network export) and reaching it
+ * must not depend on having typed its path into Model folders first.
+ *
+ * Everything below the root is therefore readable, which makes the remaining
+ * entries — home, the shortcut dirs (Runtime card, else
+ * LOCAL_MODELS_SHORTCUTS), the model dirs of saved profiles, the llama.cpp
+ * dir — jump buttons and back-compat rather than a confinement. They stay
+ * because the picker offers them, and because a saved profile must stay
+ * loadable whatever else happens.
  *
- * Nothing here is taken on faith: a path only widens the roots once it is a
- * directory the user configured, a profile that already exists, or a binary
- * that passes isLlamaServerPath(). */
+ * The guard that is still load-bearing is isLlamaServerPath(): it, not this
+ * list, decides what may be spawned (POST /settings, /runtime/check and
+ * requireBin()), so nothing a browse opens can turn into the executable. */
 function allowedRoots() {
-	const roots = [homedir(), ...shortcutDirs().map((s) => s.path)];
+	const roots = [FILESYSTEM_ROOT, homedir(), ...shortcutDirs().map((s) => s.path)];
 	try {
 		for (const profile of cachedProfiles()) {
 			for (const declared of [profile?.modelPath, profile?.mmprojPath]) {
@@ -1613,9 +1633,10 @@ function allowedRoots() {
 	return roots;
 }
 
-/** A path we are willing to register as a browsable root: absolute, no ".."
- * segment, and not a filesystem root (which would list everything). Exported
- * for tests. */
+/** A path we are willing to show as a picker button: absolute, no ".."
+ * segment, and naming one folder rather than the whole tree. The root is
+ * browsable on its own (there is a Root button and an Up that reaches it), so
+ * this is button hygiene, not confinement. Exported for tests. */
 export function isSafeRootPath(p) {
 	if (typeof p !== "string" || !p.trim()) return false;
 	const raw = p.trim();
@@ -1641,7 +1662,9 @@ function isPathAllowed(p) {
 		} catch {
 			return false;
 		}
-		return abs === r || abs.startsWith(r + sep);
+		// The root needs its own clause: the prefix test below would only ever
+		// match "/" itself (every other path starts "/…", not "//").
+		return abs === r || r === FILESYSTEM_ROOT || abs.startsWith(r + sep);
 	});
 }
 
@@ -1865,12 +1888,12 @@ function cachedProfiles() {
 export function normalizeProfileConfig(raw) {
 	const modelPath = typeof raw?.modelPath === "string" && raw.modelPath ? raw.modelPath : "";
 	if (!modelPath || !existsSync(modelPath)) throw new Error("profile needs a modelPath that exists on disk");
-	if (!isPathAllowed(modelPath)) throw new Error("modelPath is outside the folders this tab may read: " + modelPath + " — add its folder under Model folders in the Runtime card first");
+	if (!isPathAllowed(modelPath)) throw new Error("modelPath " + NOT_READABLE + ": " + modelPath);
 	const ctx = Number.isInteger(raw?.ctx) && raw.ctx > 0 ? raw.ctx : 8192;
 	const mtpHeads = Number.isInteger(raw?.mtpHeads) && raw.mtpHeads >= 0 ? Math.min(raw.mtpHeads, MTP_MAX) : 0;
 	const mmprojPath = typeof raw?.mmprojPath === "string" && raw.mmprojPath ? raw.mmprojPath : null;
 	if (mmprojPath && !existsSync(mmprojPath)) throw new Error("mmprojPath does not exist: " + mmprojPath);
-	if (mmprojPath && !isPathAllowed(mmprojPath)) throw new Error("mmprojPath is outside the folders this tab may read: " + mmprojPath + " — add its folder under Model folders in the Runtime card first");
+	if (mmprojPath && !isPathAllowed(mmprojPath)) throw new Error("mmprojPath " + NOT_READABLE + ": " + mmprojPath);
 	const effort = normalizeEffort(raw?.effort);
 	const preserveThinking = normalizePreserveThinking(raw);
 	// Multi-GPU placement. Legacy profiles have neither field and keep
@@ -2085,11 +2108,16 @@ export function apply(ctx) {
 				const url = new URL(req.url ?? "/", "http://x");
 				const dir = url.searchParams.get("dir");
 				const allFiles = url.searchParams.get("all") === "1";
-				const target = dir && dir.indexOf("/") === 0
+				const raw = dir && dir.indexOf("/") === 0
 					? dir
 					: dir ? join(homedir(), dir) : homedir();
+				// Canonical before anything else. The picker joins with "/" and
+				// starts at the root, so a click on /etc arrives as "//etc" — and
+				// the path bar would then read "//etc", with every deeper click
+				// stacking another slash. resolve() also collapses "..".
+				const target = resolve(raw);
 				if (!isPathAllowed(target)) {
-					sendJson(res, 400, { error: "path is outside the folders this tab may read (home + Model folders + the llama.cpp dir): " + target });
+					sendJson(res, 400, { error: NOT_READABLE + ": " + target });
 					return;
 				}
 				const listing = listDir(target, allFiles);
@@ -2117,7 +2145,7 @@ export function apply(ctx) {
 					return;
 				}
 				if (!isPathAllowed(path)) {
-					sendJson(res, 400, { error: "path is outside the folders this tab may read (home + Model folders + the llama.cpp dir)" });
+					sendJson(res, 400, { error: NOT_READABLE + ": " + path });
 					return;
 				}
 				try {
@@ -2210,11 +2238,11 @@ export function apply(ctx) {
 					return;
 				}
 				if (!isPathAllowed(path)) {
-					sendJson(res, 400, { error: "model path is outside the folders this tab may read: " + path + " — add its folder under Model folders in the Runtime card" });
+					sendJson(res, 400, { error: "model " + NOT_READABLE + ": " + path });
 					return;
 				}
 				if (mmproj && !isPathAllowed(mmproj)) {
-					sendJson(res, 400, { error: "mmproj path is outside the folders this tab may read: " + mmproj + " — add its folder under Model folders in the Runtime card" });
+					sendJson(res, 400, { error: "mmproj " + NOT_READABLE + ": " + mmproj });
 					return;
 				}
 				const result = await manager.run(path, Number.isFinite(ctx) ? ctx : 8192, Number.isFinite(mtp) ? mtp : 0, mmproj, effort, typeof body.mmprojCpu === "boolean" ? body.mmprojCpu : MMPROJ_CPU, moe, normalizePreserveThinking(body), kvTypeK, kvTypeV, normalizeSplitMode(body.splitMode), normalizeTensorSplit(body.tensorSplit));
@@ -2359,8 +2387,8 @@ export function apply(ctx) {
 						}
 						patch.binPath = wanted;
 					}
-					// File-browser shortcuts are browsable roots, so they are the one
-					// place where a bad value widens the file endpoints: absolute,
+					// File-browser shortcuts become one button each in the picker,
+					// so a value here has to name a folder a person would: absolute,
 					// no "..", never the filesystem root, at most 12 entries.
 					// Missing dirs are still kept (an unmounted disk must not
 					// silently lose its shortcut); the tab marks them as missing.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-local-models",
-  "version": "0.3.1",
+  "version": "0.4.0",
   "description": "dsh addon: a Local Models settings tab that starts and stops a llama-server child process living with the dsh host process.",
   "type": "module",
   "main": "lib/index.js",

package/skills/dsh-local-models-ops/SKILL.md

@@ -37,9 +37,10 @@ cpuMoe, nCpuMoe, expertUsed, arch, updatedAt }`
 
 - **Paths must exist and live on the fast mount**. Referencing the failing/
   legacy `/mnt/disco1` is a red flag - targets are under `/mnt/raid0/GGUF/`.
-  A path outside the tab's roots (home + the Runtime card's Model folders + the
-  folders saved profiles already declare) is refused on save and on launch, so
-  a profile pointing somewhere else needs its folder added in the tab first.
+  The tab can reach any absolute path (the picker's Up goes all the way to `/`,
+  so mounted disks are browsable without declaring them), so a path pointing
+  somewhere else is an operational mistake the guard will not catch - check it
+  by eye.
 - **MTP depth must be <= 7** (the tab's ceiling; upstream clamps the
   effective depth to the model's nextn depth) and **3 is the tuned value** —
   deeper collapses at large ctx. Old profiles may still carry 4-6 from the
@@ -111,21 +112,21 @@ so a 131072 window with O = 32768 thresholds at **32768**, firing at
 ~32-38K pressure tokens and looping with xhigh + preserveThinking (every
 turn re-adds 10-20K thinking tokens). A 98304 window is worse: pressure
 exactly 0, proactive compaction disabled entirely. The fix is a per-model
-headroom override in the dsh profile (needs a `dsh web` restart — loader
-normalization happens at boot), NOT a smaller O (that truncates xhigh
-thinking). Compute it with `compactionBudgetFor(ctx).recommendedHeadroom`
-(~6.5K for 131K → ~92K threshold; ~4K floor for 96K → ~61K). Example for
-the Swift 131K route (also add siblings as needed):
-
-```yaml
-- id: compaction-basic
-  name: '@deepseek-ai/dsh-compaction-basic'
-  config:
-    modelPolicies:
-      - provider: local-router
-        model: swift-1-5-qwen3-8-27b
-        headroomTokens: 6554
-```
+headroom override — but placement matters: the engines that run live
+inside the agent-preset scopes (`preset-standard` → `plugins` →
+`compaction` group → `compaction-basic`), NOT the top-level
+`compaction-basic` row (disabled there; a patch on it is dead config that
+dump-config will happily show). So the override must ride a wholesale
+`config` on the preset row (row config is replaced whole, not merged):
+copy the installed preset file
+(`dsh-web-app/presets/standard.patch.yml` → `insert[0].config`), add
+`modelPolicies` to the nested engine, paste as `- id: preset-standard`.
+Compute values with `compactionBudgetFor(ctx).recommendedHeadroom`
+(~6.5K for 131K → ~92K threshold; ~4K floor for 96K → ~61K). Because it
+is wholesale, the row freezes: re-generate after any dsh upgrade that
+touches the preset file, or upstream changes to that preset are silently
+dropped. Only patch presets actually used (`standard` here); `ptc` /
+`cordis` / `minimal` need the same treatment if adopted.
 
 Shortcut with zero config edits: run the bigger window instead — the
 200K/250K profiles threshold at ~106K/158K under defaults.