@kenkaiiii/ggcoder 5.49.5 → 5.49.7

package/dist/app-sidecar.js

@@ -38,7 +38,7 @@ import { validateKenModelPref, effectiveKenModel } from "./core/ken-model.js";
 import { normalizeAutopilotMarkersForHistory, normalizeAppMarkersForHistory, normalizeKenTurnsForHistory, getHistoryMessageVisibility, replayMessagesInOrder, restoreUserRow, restoreAssistantTexts, resolveRestoredCommand, autopilotMarkerCopySeed, } from "./core/session-history.js";
 import { sessionToMarkdown, defaultExportFilename, } from "./core/session-export.js";
 import { AuthStorage } from "./core/auth-storage.js";
-import { explainPullFailure, isValidHfRepoId, parseOllamaPullLine, pickGgufQuant, toHfSearchRow, } from "./hf-pull.js";
+import { advancesPhase, explainPullFailure, isGgufShard, isValidHfRepoId, parseOllamaPullLine, pickGgufQuant, SHARDED_MESSAGE, toHfSearchRow, } from "./hf-pull.js";
 import { cleanupToolOutputs } from "./tools/overflow.js";
 import { readCappedBody } from "./utils/http-body.js";
 import { fetchSubscriptionUsage, SubscriptionUsageError, XIAOMI_CREDITS_KEY, dualAuthProvider, oauthStorageKey, discoverLocalModels, findProbedModel, formatLocalModelId, parseLocalModelId, probeEndpoint, toModelInfo as localModelInfo, } from "@kenkaiiii/gg-core";
@@ -1452,13 +1452,19 @@ async function createSession(deps, opts) {
     }
     /** Search the Hub for GGUF repos (what Ollama can pull). */
     async function hfSearch(query) {
+        // `filter=gguf` (the tag the Hub applies to repos that actually contain
+        // GGUF files) — `library=gguf` also matches safetensors-only base repos,
+        // which then fail at pull time. `expand[]=gguf` confirms per row.
         const params = new URLSearchParams({
             search: query,
-            library: "gguf",
+            filter: "gguf",
             sort: "downloads",
             direction: "-1",
-            limit: "8",
+            limit: "12",
         });
+        for (const field of ["gguf", "downloads", "likes", "lastModified"]) {
+            params.append("expand[]", field);
+        }
         const data = (await hfHubJson(`/api/models?${params.toString()}`));
         return (Array.isArray(data) ? data : [])
             .map(toHfSearchRow)
@@ -1474,9 +1480,15 @@ async function createSession(deps, opts) {
         if (hfPull?.child) {
             throw Object.assign(new Error("A download is already running."), { status: 409 });
         }
-        const tree = (await hfHubJson(`/api/models/${repo}/tree/main`));
+        // `recursive=true`: many repos (unsloth, mradermacher) keep quants in
+        // per-quant subfolders, and a flat listing reports them as GGUF-less.
+        const tree = (await hfHubJson(`/api/models/${repo}/tree/main?recursive=true`));
         const files = (Array.isArray(tree) ? tree : []).flatMap((entry) => {
             const e = entry;
+            // `type` guards a directory entry (size 0) from beating real files in the
+            // size fallback; the Hub marks folders as "directory".
+            if (e.type !== undefined && e.type !== "file")
+                return [];
             if (typeof e.path !== "string" || !e.path.toLowerCase().endsWith(".gguf"))
                 return [];
             const size = Number(e.lfs?.size ?? e.size ?? 0);
@@ -1484,9 +1496,8 @@ async function createSession(deps, opts) {
         });
         const choice = pickGgufQuant(files);
         if (!choice) {
-            throw Object.assign(new Error("That repo has no GGUF file for Ollama to pull."), {
-                status: 400,
-            });
+            const sharded = files.some((f) => isGgufShard(f.path));
+            throw Object.assign(new Error(sharded ? SHARDED_MESSAGE : "That repo has no GGUF file for Ollama to pull."), { status: 400 });
         }
         const model = `hf.co/${repo}${choice.tag ? `:${choice.tag}` : ""}`;
         const token = await hfToken();
@@ -1504,6 +1515,7 @@ async function createSession(deps, opts) {
         broadcast("hf_pull", hfPullPayload(state));
         // ollama prints progress to stderr (stdout on some builds); parse both.
         let stderrTail = "";
+        let lastBroadcast = "";
         const feed = (chunk) => {
             // Once the pull is terminal (success, failure, or user cancel) stop
             // parsing: a killed ollama dumps a burst of stderr that would otherwise
@@ -1514,12 +1526,23 @@ async function createSession(deps, opts) {
                 const parsed = parseOllamaPullLine(line);
                 if (!parsed)
                     continue;
+                // A redrawn frame repeats `pulling manifest` next to live progress; it
+                // must not drag the modal back to "Contacting Ollama…".
+                if (!advancesPhase(state.phase, parsed.phase))
+                    continue;
                 if (parsed.phase !== "error") {
                     state.phase = parsed.phase;
                     if (parsed.percent !== undefined)
                         state.percent = parsed.percent;
                     state.detail = parsed.detail;
                 }
+                // ollama redraws its TUI frame several times a second, mostly with
+                // identical numbers. Broadcasting each one re-rendered the modal for no
+                // visible change; only a frame that actually reads differently ships.
+                const next = JSON.stringify(hfPullPayload(state));
+                if (next === lastBroadcast)
+                    continue;
+                lastBroadcast = next;
                 broadcast("hf_pull", hfPullPayload(state));
             }
         };
@@ -1554,9 +1577,11 @@ async function createSession(deps, opts) {
                     state.percent = 100;
                     state.detail = undefined;
                     broadcast("hf_pull", hfPullPayload(state));
-                    // The new model only exists to the app once Ollama lists it.
+                    // The new model only exists to the app once Ollama lists it. Ollama's
+                    // library is machine-wide, like the shared auth file, so every window
+                    // gets the refresh — not just the one that ran the download.
                     void scanLocalModels(true)
-                        .then(() => broadcast("models_change", { local: localStatePayload() }))
+                        .then(() => broadcastAll("models_change", { local: localStatePayload() }))
                         .catch(() => undefined);
                 }
                 else {