dsh-model-accordion 0.1.1 → 0.2.1

package/README.md

@@ -6,9 +6,8 @@ Provider-folded model selector for the DeepSeek Harness Web composer.
 
 - Replaces the single composer model seat with a provider-folded selector.
 - Keeps provider groups collapsed by default and opens models on demand.
-- Reads reasoning-effort choices from each model's resolved catalog metadata instead of maintaining a hard-coded effort list.
-- Sends model and reasoning-effort selections through DSH's shared model-directory service.
-- Wraps vision-router mirror providers in one collapsed `识图镜像（带图会话）` section when those providers are present.
+- Reads reasoning-effort choices from each model's resolved catalog metadata, with a built-in official-effort table as a fallback. A single automatic rule applies: a model with configured/backend-declared efforts uses them, and an unconfigured model falls back to the built-in official table. A user who configured their own levels keeps them; no per-model effort configuration is needed for the official fallback to work.
+- Sends model and reasoning-effort selections through DSH's shared model-directory service; a picked effort the model does not genuinely support is dropped so the request uses the model's default instead of failing with `UNSUPPORTED_REASONING_EFFORT`.
 - Keeps the popup height-bounded with an internal scroll area and wraps effort controls below the model name.
 - Uses DSH semantic theme tokens for colors and supports light/dark theme changes.
 
@@ -34,12 +33,49 @@ dsh plugin --profile web add file:/absolute/path/to/dsh-model-accordion
 
 Refresh the Web UI after installation. The package declares a `dsh.bundle` patch and a Web client entry, so it is installed and loaded as a persistent profile bundle rather than a temporary runtime extension.
 
+## Customizing reasoning effort for a model
+
+DSH persists model/provider configuration in your main `settings.yaml`, and this
+plugin reads the same resolved model metadata the backend produces, so
+**declared/backend effort configuration always wins over the built-in table**.
+The effective priority is:
+
+    model's configured `reasoningEfforts`  >  provider-level default  >  built-in official table
+
+To give a specific model — for one your relay exposes under a name this plugin's
+table does not know — your own effort ladder, declare it on that model in
+`settings.yaml` and the plugin uses exactly those levels:
+
+    providers:
+      my-relay:
+        api: openai-completions
+        baseURL: https://…
+        models:
+          - id: my-unknown-model
+            api: openai-completions
+            contextWindow: 200000
+            maxTokens: 32000
+            input: [text]
+            reasoningEfforts:
+              off: null      # optional: lets the user turn thinking off
+              low: low
+              high: high
+              max: max
+
+The exact fields follow `dsh-llm-pi-ai`'s model schema (see that package's
+docs); `reasoningEfforts` maps each offered level to its wire spelling, and you
+may set a provider-level default with the `reasoning:` key. Because this lives
+in your DSH settings file, an agent — or you — can edit it with the file tools
+and DSH reloads it on save; the browser never sees your API key or relay URL.
+
 ## Behavior and limitations
 
-- The selector submits the exact provider, model, and catalog-declared reasoning effort returned by the DSH model directory.
+The selector submits the exact provider, model, and reasoning effort returned by the DSH model directory (configured efforts first, built-in official table as fallback).
+- A picked reasoning effort that the model does not support is silently dropped so the request falls back to the model's default (no `UNSUPPORTED_REASONING_EFFORT` error). A relay/gateway model configured without an effort level keeps the model but does not force an unsupported level.
 - A selection can still be rejected by the DSH host when the session contains images and the selected model does not declare image input. That is a host/model-capability rule, not a UI override.
-- The optional `识图镜像（带图会话）` section is shown when provider names identify vision-router mirror routes. It is kept collapsed to avoid duplicating every provider in the main list.
-- The plugin currently uses English fallback status text and a Chinese label for the vision-mirror section; full locale integration is planned for a later release.
+- The model search box filters provider lists by name, id, or description while keeping the accordion grouping; a provider group auto-expands while a query is active.
+- The built-in official-effort table keys on the model's name/id and is cross-checked against the installed pi-ai catalog and each vendor's API docs. It covers only models whose vendors expose a real effort ladder: GPT-5.6 (full ladder incl. `max`), gpt-5.2+ (`xhigh`), gpt-5.1, gpt-5/mini/nano, o1/o3/o4, gpt-oss, Grok 4.x, GLM-5.3 (`low/high/max`) and GLM-5.2+ (`xhigh/max`), DeepSeek-V4 (`low/high/max`), DeepSeek-V4.1-Flash (`deepseek-flash` / `deepseek-v4.1-*`, `low/high/max`), Kimi K3, Qwen3.8, Gemini 3+ (`low/high`), Claude 4.6+ adaptive, Gemma 4. Families whose native control is not an effort ladder (Gemini 2.5 / Claude ≤4.5 token budgets, GLM ≤5.1 / Kimi K2 / MiniMax / Mistral thinking switches, Qwen open-source, DeepSeek ≤V3) are deliberately excluded. Unmatched models offer only the efforts their configured catalog declares — see "Customizing reasoning effort for a model" to add your own.
+- The plugin uses English status and search text; full locale integration is planned for a later release.
 
 ## Development checks
 

package/lib/client.js

@@ -25,7 +25,6 @@ window.__ModuleLoader__.load({
 .dshma-status,.dshma-empty{color:var(--dsw-alias-label-tertiary);padding:10px;font-size:13px;line-height:20px}
 .dshma-error{background:var(--dsw-alias-interactive-bg-hover-danger);color:var(--dsw-alias-state-error-primary);border-radius:8px;margin-bottom:4px;padding:7px 8px;font-size:12px;line-height:18px}
 .dshma-provider{border-radius:10px;overflow:hidden;flex:none}
-.dshma-vision{margin-top:4px;border-top:1px solid var(--dsw-alias-border-l3);padding-top:4px}
 .dshma-provider-head{box-sizing:border-box;width:100%;display:flex;align-items:center;gap:7px;padding:8px;border:none;border-radius:10px;background:transparent;color:var(--dsw-alias-label-secondary);font-size:12px;font-weight:600;line-height:18px;text-align:left;cursor:pointer;font-family:inherit}
 .dshma-provider-head:hover{background:var(--dsw-alias-interactive-bg-hover)}
 .dshma-provider-name{flex:1;min-width:0;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
@@ -46,6 +45,11 @@ window.__ModuleLoader__.load({
 .dshma-effort:hover{color:var(--dsw-alias-label-secondary);background:var(--dsw-alias-interactive-bg-hover)}
 .dshma-effort-selected{color:var(--dsw-alias-label-primary);border-color:var(--dsw-alias-label-secondary)}
 .dshma-failure{color:var(--dsw-alias-label-tertiary);border-radius:8px;padding:7px 8px;font-size:12px;line-height:18px}
+.dshma-search-wrap{flex:none;padding:4px;position:relative}
+.dshma-search{box-sizing:border-box;width:100%;height:28px;padding:0 28px 0 10px;border:1px solid var(--dsw-alias-border-l3);border-radius:8px;background:var(--dsw-alias-bg-module-platform);color:var(--dsw-alias-label-primary);font-size:13px;line-height:20px;outline:none;font-family:inherit}
+.dshma-search:focus-visible{border-color:var(--dsw-alias-border-l3);box-shadow:0 0 0 2px var(--dsw-alias-border-l3)}
+.dshma-search-clear{position:absolute;top:50%;right:7px;transform:translateY(-50%);width:18px;height:18px;border:none;border-radius:50%;background:transparent;color:var(--dsw-alias-label-tertiary);cursor:pointer;display:flex;align-items:center;justify-content:center;padding:0;font-size:13px;line-height:1}
+.dshma-search-clear:hover{color:var(--dsw-alias-label-secondary);background:var(--dsw-alias-interactive-bg-hover)}
 `;
 
     function Chevron(props) {
@@ -62,9 +66,110 @@ window.__ModuleLoader__.load({
       );
     }
 
+    //#region Official reasoning-effort divisions
+    // Built-in, provider-independent reasoning-effort divisions keyed on the
+    // model's name/id. This is what lets the selector offer sensible effort
+    // choices even when a model's configured catalog (or a third-party relay /
+    // gateway) declares no `reasoningEfforts` — the common npm-install case.
+    // The divisions below are sourced from each vendor's public API docs, cross-
+    // checked against the installed pi-ai catalog (providers/data/*.json) and the
+    // vendor docs listed in the plugin changelog. Levels are the canonical pi-ai
+    // ids (off/minimal/low/medium/high/xhigh/max — see THINKING_LEVELS in
+    // dsh-llm-pi-ai); `default` is the official default and is used only when the
+    // backend leaves it unset.
+    //
+    // IMPORTANT — only models whose vendors expose a real reasoning-effort ladder
+    // are listed. Families whose native control is NOT an effort ladder (Claude
+    // ≤4.5 & Gemini 2.5 token budgets, GLM ≤5.1 / Kimi K2 / MiniMax / Mistral
+    // two-state thinking, Qwen open-source enable_thinking, DeepSeek ≤V3 / R1) are
+    // deliberately excluded so we never surface a ladder the API would reject.
+    // Official-family models are normally covered by the installed pi-ai catalog
+    // (backend first layer) anyway; this table only backs up manually-declared /
+    // relay models the backend leaves unconfigured.
+    //
+    // Entries are checked top-down; the first match wins.
+    //   gpt-5.6 (Sol/Terra/Luna): full ladder incl. max            (default high ⚠ TBD)
+    //   gpt-5.2+ / 5.3-codex:      low/medium/high/xhigh            (default high ⚠ TBD)
+    //   gpt-5.1:                   minimal/low/medium/high          (default low)
+    //   gpt-5 / mini / nano:       minimal/low/medium/high          (default medium)
+    //   o1/o3/o4 families:         low/medium/high                  (default medium)
+    //   gpt-oss:                   low/medium/high                  (default medium)
+    //   grok 4.x (responses):      low/medium/high                  (default high)
+    //   glm-5.3 / 5.3-FLASH:       low/high/max (thinking un-offable)(default max)
+    //   glm-5.2+:                  low/medium/high/xhigh/max        (default max)
+    //   deepseek-v4:               low/high/max                     (default high)
+    //   deepseek-v4.1-flash / deepseek-flash: low/high/max         (default high)
+    //   kimi-k3:                   low/high/max                     (default max)
+    //   qwen3.8:                   xhigh/medium/low                 (default xhigh)
+    //   gemini-3+ (Pro/Flash):     low/high (always on)             (default high)
+    //   claude 4.6+ adaptive:      low/medium/high                  (default high)
+    //   gemma-4:                   minimal/high                     (default high)
+    const OFFICIAL_EFFORTS = [
+      { test: /gpt-5\.6/i, efforts: ["minimal", "low", "medium", "high", "xhigh", "max"], def: "high" },
+      { test: /gpt-5\.2[^0-9]|gpt-5\.[3-9]|gpt-5\.2\b/i, efforts: ["low", "medium", "high", "xhigh"], def: "high" },
+      { test: /gpt-5\.1/i, efforts: ["minimal", "low", "medium", "high"], def: "low" },
+      { test: /gpt-5(?:-mini|-nano)?\b/i, efforts: ["minimal", "low", "medium", "high"], def: "medium" },
+      { test: /o[134](-mini|-pro)?\b/i, efforts: ["low", "medium", "high"], def: "medium" },
+      { test: /gpt-oss/i, efforts: ["low", "medium", "high"], def: "medium" },
+      { test: /grok/i, efforts: ["low", "medium", "high"], def: "high" },
+      { test: /glm-5\.3/i, efforts: ["low", "high", "max"], def: "max" },
+      { test: /glm-5(?:\.[2-9]|\.\d+)/i, efforts: ["low", "medium", "high", "xhigh", "max"], def: "max" },
+      { test: /deepseek-v4/i, efforts: ["low", "high", "max"], def: "high" },
+      { test: /deepseek-(?:v4\.1|flash)/i, efforts: ["low", "high", "max"], def: "high" },
+      { test: /kimi-k3/i, efforts: ["low", "high", "max"], def: "max" },
+      { test: /qwen3\.8/i, efforts: ["xhigh", "medium", "low"], def: "xhigh" },
+      { test: /gemini-3/i, efforts: ["low", "high"], def: "high" },
+      { test: /claude-(opus|sonnet|haiku|fable)-(4\.[6-9]|5)\b/i, efforts: ["low", "medium", "high"], def: "high" },
+      { test: /gemma-4/i, efforts: ["minimal", "high"], def: "high" }
+    ];
+
+    function officialEffortsFor(model) {
+      if (!model) return void 0;
+      const hay = `${model.name || ""} ${model.id || ""}`;
+      for (const entry of OFFICIAL_EFFORTS) {
+        if (entry.test.test(hay)) return entry;
+      }
+      return void 0;
+    }
+
+    // Reason offered for a model: the backend/configured-catalog efforts when the
+    // model has them, otherwise the built-in official table as a fallback. This
+    // single "auto" rule is all there is — a user who configured their own levels
+    // keeps them, and an unconfigured model still gets sensible official choices.
+    function resolveEfforts(model) {
+      const backend = model && model.reasoning && Array.isArray(model.reasoning.efforts)
+        ? model.reasoning.efforts.filter((e) => e && typeof e.id === "string")
+        : [];
+      if (backend.length > 0) return backend;
+      const official = officialEffortsFor(model);
+      if (!official) return [];
+      return official.efforts.map((id) => ({ id, name: `${id.charAt(0).toUpperCase()}${id.slice(1)}` }));
+    }
+    // The default effort id to surface for a model (trigger caption + default
+    // chip highlight): the backend-declared default wins, and the official table's
+    // default is used only when the backend left it unset.
+    function resolveDefaultEffort(model) {
+      const backendDefault = model && model.reasoning && model.reasoning.defaultEffort != null
+        ? model.reasoning.defaultEffort
+        : void 0;
+      if (backendDefault !== void 0) return backendDefault;
+      const official = officialEffortsFor(model);
+      return official ? official.def : void 0;
+    }
+    // Whether an effort id is genuinely supported by the model — used by the
+    // submit path to decide whether to send reasoningEffort.
+    function effortSupported(model, id) {
+      return resolveEfforts(model).some((e) => e.id === id);
+    }
+
     function effortOptions(model) {
-      const es = model && model.reasoning && Array.isArray(model.reasoning.efforts) ? model.reasoning.efforts : [];
-      return es.filter((e) => e && typeof e.id === "string");
+      return resolveEfforts(model);
+    }
+
+    function effortName(model, id) {
+      const es = resolveEfforts(model);
+      const hit = es.find((e) => e.id === id);
+      return hit && hit.name ? hit.name : id;
     }
 
     function Accordion(props) {
@@ -75,14 +180,26 @@ window.__ModuleLoader__.load({
       const [failure, setFailure] = React.useState(null);
       const groups = Array.isArray(state.groups) ? state.groups : [];
       const rootRef = React.useRef(null);
-      const [visionOpen, setVisionOpen] = React.useState(false);
-      // Vision-router twin detection: only treat a group as the auto-vision
-      // wrapper when its display name carries the 自动识图 token or its provider
-      // id ends in "-vision" (official vision-router uses route id
-      // "deepseek-vision" + display "DeepSeek + 自动识图").
-      const isTwin = (g) => /自动识图/.test(g.name || "") || /(?:-|\b)vision$/i.test((g.id || "").trim());
-      const mainGroups = groups.filter((g) => !isTwin(g));
-      const visionGroups = groups.filter(isTwin);
+      const [query, setQuery] = React.useState("");
+      // Roving focus registry: every visible model-row button registers itself here
+      // (keyed by group+model), so arrow/Home/End navigate in accordion order.
+      const itemKeys = React.useRef({});
+      const registerItem = (key, el) => { if (el) itemKeys.current[key] = el; else delete itemKeys.current[key]; };
+      const searchRef = React.useRef(null);
+
+      // When the menu opens with NO provider expanded yet, surface the provider
+      // holding the current model so a fresh open is not a blank, fully-collapsed
+      // list. Crucially this only fires while `expanded` is empty: re-running it on
+      // every open (selecting a model closes the menu, so compare-testing reopens
+      // repeatedly) would otherwise accumulate every tested provider into
+      // `expanded`, leaving them all open once the search filter is cleared. When
+      // the user has already opened providers (browse state) we leave them alone.
+      React.useEffect(() => {
+        if (!open) return;
+        if (Object.values(expanded).some(Boolean)) return;
+        const g = groups.find((gg) => gg.id === state.current?.provider);
+        if (g) setExpanded((old) => ({ ...old, [g.id]: true }));
+      }, [open]);
 
       React.useEffect(() => {
         if (!open) return;
@@ -93,6 +210,11 @@ window.__ModuleLoader__.load({
         return () => { document.removeEventListener("mousedown", onDown); document.removeEventListener("keydown", onKey); };
       }, [open]);
 
+      // Populate store.current as soon as the model slot becomes available
+      // (mount/startup), so the trigger shows the last-used model instead of
+      // falling back to "Select model" until the user opens the menu.
+      React.useEffect(() => { if (props.available) props.load(); }, [props.available, props.load]);
+      // Refresh the store whenever the menu is opened (mirrors official reload).
       React.useEffect(() => { if (open && props.available) props.load(); }, [open, props.available]);
 
       const select = (group, model, effort) => {
@@ -100,7 +222,14 @@ window.__ModuleLoader__.load({
         setBusy(true);
         setFailure(null);
         const sel = { provider: group.id, model: model.id };
-        if (effort !== void 0) sel.reasoningEffort = effort;
+        // Submit-degradation: only send reasoningEffort when the picked level is
+        // one the model genuinely supports. When the backend catalog declares
+        // efforts, an "official" level it does not list is dropped so the request
+        // falls back to the model's default instead of failing with
+        // UNSUPPORTED_REASONING_EFFORT. NOTE: when the backend declares no efforts
+        // at all the official table is the sole authority, so any official level is
+        // sent as-is — an under-configured relay may still reject it.
+        if (effort !== void 0 && effortSupported(model, effort)) sel.reasoningEffort = effort;
         props.select(sel).then((ok) => {
           if (!ok) throw new Error("model selection failed");
           setOpen(false);
@@ -113,12 +242,41 @@ window.__ModuleLoader__.load({
       const curGroup = groups.find((g) => g.id === current?.provider);
       const curModel = curGroup ? curGroup.models.find((m) => m.id === current?.model) : void 0;
       const label = curModel ? curModel.name : (current ? `${current.provider}/${current.model}` : "Select model");
-      const curEffort = curModel && current && current.provider === curGroup.id && current.model === curModel.id
-        ? (current.reasoningEffort || (curModel.reasoning && curModel.reasoning.defaultEffort))
+      // Show a `· effort` caption ONLY when a reasoning effort is actually pinned
+      // on the current selection (persisted in `current.reasoningEffort`). The
+      // official default is deliberately NOT surfaced here: showing it before the
+      // user pins anything reads as "already set", which is misleading.
+      const curEffort = curModel && current && current.provider === curGroup.id && current.model === curModel.id && current.reasoningEffort
+        ? effortName(curModel, current.reasoningEffort)
         : void 0;
 
+      // Model search: filter each provider's models by the query while keeping
+      // the accordion grouping structure. Matches on name, id, and description.
+      const q = query.trim().toLowerCase();
+      // When the filter is cleared (✕ / Escape / manual delete), collapse the
+      // accordion to the provider holding the CURRENT model. Selecting a model
+      // never writes `expanded`, so without this the menu would fall back to
+      // whatever was expanded before the search — not the model you just picked.
+      // Because selection does not accumulate into `expanded`, this stays focused
+      // on one group and never re-introduces the "everything open" problem.
+      const prevQuery = React.useRef("");
+      React.useEffect(() => {
+        const prev = prevQuery.current;
+        prevQuery.current = q;
+        if (q === "" && prev !== "") {
+          const cur = state.current;
+          setExpanded(cur && groups.some((g) => g.id === cur.provider) ? { [cur.provider]: true } : {});
+        }
+      }, [q]);
+      const matches = (m) => q === "" || [m.name, m.id, m.description].some((s) => typeof s === "string" && s.toLowerCase().includes(q));
+      const visibleGroups = q === ""
+        ? groups
+        : groups.map((g) => ({ ...g, models: g.models.filter(matches) })).filter((g) => g.models.length > 0);
+      const hasAnyModel = visibleGroups.some((g) => g.models.length > 0);
+
       const renderGroup = (group) => {
-        const isOpen = expanded[group.id] === true;
+        const isOpen = expanded[group.id] === true || q !== "";
+        if (!group.models.length) return null;
         return React.createElement("div", { className: "dshma-provider", key: group.id },
           React.createElement("button", { type: "button", className: "dshma-provider-head", onClick: () => setExpanded((old) => ({ ...old, [group.id]: !isOpen })), "aria-expanded": isOpen },
             React.createElement(Chevron, { className: `dshma-provider-chevron${isOpen ? " dshma-provider-chevron-open" : ""}` }),
@@ -128,17 +286,16 @@ window.__ModuleLoader__.load({
           isOpen ? React.createElement("div", { className: "dshma-models" }, group.models.map((model) => {
             const selected = current && current.provider === group.id && current.model === model.id;
             const efforts = effortOptions(model);
-            const defEffort = model.reasoning && model.reasoning.defaultEffort;
-            // A "provider default" chip clears a pinned effort so the provider's
-            // built-in default applies again (mirrors the official provider-default
-            // entry in dsh-client-ui-model-selection @315-324).
+            const defEffort = resolveDefaultEffort(model);
+            // A "default" chip clears a pinned effort so the model's effective
+            // default applies again (mirrors the official provider-default entry
+            // in dsh-client-ui-model-selection @315-324). It appears only when a
+            // default effort is known (backend-declared or official-table).
             const defaultChip = defEffort !== void 0 ? [{ id: "__provider_default__", name: "默认" }] : [];
             const chips = defaultChip.concat(efforts);
-            // reason about the currently active effort: a pinned explicit effort
-            // wins over the provider's default; otherwise provider default applies.
             const pinnedEffort = selected ? current.reasoningEffort : void 0;
             return React.createElement("div", { className: `dshma-option${selected ? " dshma-option-selected" : ""}`, key: model.id },
-              React.createElement("button", { type: "button", className: "dshma-option-row", onClick: () => {
+              React.createElement("button", { type: "button", className: "dshma-option-row", ref: (el) => registerItem(`${group.id}\u0000${model.id}`, el), onClick: () => {
                 // Re-clicking the already-selected model must NOT re-submit and
                 // thereby wipe a custom (pinned) effort — just close the menu.
                 // Mirrors dsh-client-ui-model-selection @397-400 / @718.
@@ -157,28 +314,48 @@ window.__ModuleLoader__.load({
         );
       };
 
+      // Roving arrow-key navigation across the rendered model rows (mirrors the
+      // official ModelSelect moveFocus). Enter/Space activate the focused row
+      // natively; Arrow keys / Home / End move between rows. ArrowDown from the
+      // search field lands on the first row.
+      const onMenuKey = (e) => {
+        // Escape: first press with text clears the search (keeping focus), second
+        // press closes. stopPropagation prevents the document-level handler from
+        // closing the menu while we are only clearing.
+        if (e.key === "Escape" && query !== "") {
+          e.preventDefault(); e.stopPropagation(); setQuery(""); searchRef.current && searchRef.current.focus(); return;
+        }
+        const els = Object.values(itemKeys.current)
+          .filter((el) => el && el.offsetParent !== null);
+        if (!els.length) return;
+        const idx = els.indexOf(document.activeElement);
+        let next = null;
+        if (e.key === "ArrowDown") { e.preventDefault(); next = idx === -1 ? els[0] : els[(idx + 1) % els.length]; }
+        else if (e.key === "ArrowUp") { e.preventDefault(); next = idx === -1 ? els[els.length - 1] : els[(idx - 1 + els.length) % els.length]; }
+        else if (e.key === "Home") { e.preventDefault(); next = els[0]; }
+        else if (e.key === "End") { e.preventDefault(); next = els[els.length - 1]; }
+        if (next) next.focus();
+      };
+
       return React.createElement("div", { className: "dshma-root", ref: rootRef },
         React.createElement("button", { type: "button", className: "dshma-trigger", disabled: props.locked || !props.available || busy, onClick: () => setOpen((v) => !v), "aria-expanded": open, "aria-haspopup": "menu" },
           React.createElement("span", { className: "dshma-trigger-label" }, label),
           curEffort ? React.createElement("span", { className: "dshma-trigger-effort" }, `· ${curEffort}`) : null,
           React.createElement(Chevron, { className: `dshma-chevron${open ? " dshma-chevron-open" : ""}` })
         ),
-        open ? React.createElement("div", { className: "dshma-menu", role: "menu" },
+        open ? React.createElement("div", { className: "dshma-menu", role: "menu", "aria-label": "Model selector", onKeyDown: onMenuKey },
+          React.createElement("div", { className: "dshma-search-wrap" },
+            React.createElement("input", { type: "text", className: "dshma-search", placeholder: "Search models…", value: query, onChange: (e) => setQuery(e.target.value), "aria-label": "Search models", autoFocus: true, ref: searchRef }),
+            query !== "" ? React.createElement("button", { type: "button", className: "dshma-search-clear", onClick: () => { setQuery(""); searchRef.current && searchRef.current.focus(); }, "aria-label": "Clear search", title: "Clear search" }, "✕") : null
+          ),
           React.createElement("div", { className: "dshma-menu-scroll" },
             failure ? React.createElement("div", { className: "dshma-error" }, failure) : null,
             state.error ? React.createElement("div", { className: "dshma-error" }, state.error) : null,
             state.status === "loading" ? React.createElement("div", { className: "dshma-status" }, "Loading models…") : null,
             (state.status === "selecting" || busy) ? React.createElement("div", { className: "dshma-status" }, "Selecting…") : null,
-            state.status !== "loading" && mainGroups.length === 0 && visionGroups.length === 0 ? React.createElement("div", { className: "dshma-empty" }, "No models available") : null,
-            mainGroups.map(renderGroup),
-            visionGroups.length ? React.createElement("div", { className: "dshma-vision", key: "__vision__" },
-              React.createElement("button", { type: "button", className: "dshma-provider-head", onClick: () => setVisionOpen((v) => !v), "aria-expanded": visionOpen },
-                React.createElement(Chevron, { className: `dshma-provider-chevron${visionOpen ? " dshma-provider-chevron-open" : ""}` }),
-                React.createElement("span", { className: "dshma-provider-name" }, "识图镜像（带图会话）"),
-                React.createElement("span", { className: "dshma-count" }, `${visionGroups.length}`)
-              ),
-              visionOpen ? React.createElement("div", { className: "dshma-models" }, visionGroups.map(renderGroup)) : null
-            ) : null
+            state.status !== "loading" && !hasAnyModel && q === "" ? React.createElement("div", { className: "dshma-empty" }, "No models available") : null,
+            q !== "" && !hasAnyModel ? React.createElement("div", { className: "dshma-empty" }, "No matches") : null,
+            visibleGroups.map(renderGroup)
           )
         ) : null
       );

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-model-accordion",
-  "version": "0.1.1",
+  "version": "0.2.1",
   "description": "Provider-folded model selector for the DSH Web composer with catalog-driven reasoning effort choices.",
   "keywords": [
     "deepseek-harness",
@@ -46,7 +46,7 @@
   },
   "repository": {
     "type": "git",
-    "url": "https://github.com/lakabuji-arch/dsh-model-accordion.git"
+    "url": "git+https://github.com/lakabuji-arch/dsh-model-accordion.git"
   },
   "homepage": "https://github.com/lakabuji-arch/dsh-model-accordion#readme",
   "bugs": {