fixes

src/db.ts

@@ -387,7 +387,7 @@ export function deleteShare(convId: string): void {
   getDb().query("DELETE FROM shares WHERE conv_id = ?").run(convId);
 }
 
-export function getShareByPublic(publicId: string): { conv: ConversationRow; messages: MessageRow[]; authorName: string; sharedAt: number } | null {
+export function getShareByPublic(publicId: string): { conv: ConversationRow; messages: MessageRow[]; authorName: string; authorAvatarUrl: string; sharedAt: number } | null {
   const d = getDb();
   const s = d.query("SELECT * FROM shares WHERE public_id = ?").get(publicId) as { conv_id: string; created_at: number } | null;
   if (!s) return null;
@@ -398,6 +398,7 @@ export function getShareByPublic(publicId: string): { conv: ConversationRow; mes
     conv,
     messages: getMessages(conv.id).filter((m) => m.role !== "system"),
     authorName: author?.display_name || "Someone",
+    authorAvatarUrl: author?.avatar_url || "",
     sharedAt: s.created_at,
   };
 }

src/drivers/registry.ts

@@ -10,8 +10,16 @@ import type { DriverModel, LLMDriver } from "./types.ts";
  */
 const PRIORITY = ["ollama", "mistral", "glm", "puppeteer-openai", "deepseek", "anthropic", "openai"];
 
+/** How long the user-facing model list stays in memory before re-hitting upstream. */
+export const MODEL_CACHE_TTL_MS = 60_000;
+
 export class DriverRegistry {
   private drivers: Map<string, LLMDriver>;
+  // In-memory model list cache: user requests are served from here so we
+  // don't hit Ollama (/api/tags) on every login. Refreshed at most once
+  // per MODEL_CACHE_TTL_MS; admins can force a refresh.
+  private modelsCache: { models: DriverModel[]; at: number } | null = null;
+  private modelsInflight: Promise<DriverModel[]> | null = null;
 
   constructor() {
     const all: LLMDriver[] = [
@@ -49,6 +57,36 @@ export class DriverRegistry {
     return out;
   }
 
+  /**
+   * Cached variant of listAllModels for user-facing requests.
+   * - Serves the in-memory list when fresher than MODEL_CACHE_TTL_MS.
+   * - Otherwise refetches upstream once (concurrent callers share the
+   *   same in-flight promise instead of hammering Ollama).
+   * - `force=true` bypasses the cache (admin "Ollama models" view).
+   */
+  async listAllModelsCached(force = false): Promise<DriverModel[]> {
+    const now = Date.now();
+    if (!force && this.modelsCache && now - this.modelsCache.at < MODEL_CACHE_TTL_MS) {
+      return this.modelsCache.models;
+    }
+    if (!force && this.modelsInflight) return this.modelsInflight;
+    const p = this.listAllModels().then((models) => {
+      this.modelsCache = { models, at: Date.now() };
+      return models;
+    }).finally(() => {
+      if (this.modelsInflight === p) this.modelsInflight = null;
+    });
+    // Only share the promise for non-forced loads; a forced admin refresh
+    // always goes upstream even if a user load is in flight.
+    if (!force) this.modelsInflight = p;
+    return p;
+  }
+
+  /** Drop the cached model list (e.g. after an admin pull/delete). */
+  invalidateModelsCache(): void {
+    this.modelsCache = null;
+  }
+
   /**
    * Split a fully-qualified "driver:model" id.
    * "ollama:llama3.1:8b" -> driver "ollama", model "llama3.1:8b".

src/index.ts

@@ -198,6 +198,7 @@ const server = Bun.serve({
         topic: s.conv.topic,
         model: s.conv.model,
         authorName: s.authorName,
+        authorAvatarUrl: s.authorAvatarUrl,
         sharedAt: s.sharedAt,
         messages: s.messages.map((m) => ({ role: m.role, content: m.content })),
       });
@@ -768,16 +769,21 @@ const server = Bun.serve({
               // Passthrough of the NDJSON progress lines (status/digest/completed/total).
               const dec = new TextDecoder();
               let buf = "";
+              let pulledOk = true;
               for await (const chunk of res.body) {
-                if (req.signal.aborted) break;
+                if (req.signal.aborted) { pulledOk = false; break; }
                 buf += dec.decode(chunk, { stream: true });
                 let nl: number;
                 while ((nl = buf.indexOf("\n")) >= 0) {
                   const line = buf.slice(0, nl).trim();
                   buf = buf.slice(nl + 1);
-                  if (line && !send(JSON.parse(line))) break;
+                  if (line && !send(JSON.parse(line))) { pulledOk = false; break; }
                 }
+                if (!pulledOk) break;
               }
+              // A new model landed on the Ollama host: drop the cached list
+              // so the next user request picks it up (subject to TTL).
+              if (pulledOk && !req.signal.aborted) registry.invalidateModelsCache();
               send({ status: "done" });
             } catch (e) {
               if (!req.signal.aborted && !clientGone) send({ error: (e as Error).message });
@@ -812,6 +818,7 @@ const server = Bun.serve({
             signal: AbortSignal.timeout(15_000),
           });
           if (!res.ok) return json({ error: `ollama /api/delete failed: ${res.status}` }, 502);
+          registry.invalidateModelsCache();
           return json({ ok: true });
         } catch (e) {
           return json({ error: (e as Error).message }, 502);
@@ -967,7 +974,13 @@ const server = Bun.serve({
       }
 
       if (path === "/api/models" && req.method === "GET") {
-        const models = await registry.listAllModels();
+        // User-facing list is served from an in-memory cache (TTL 60s) so
+        // every login doesn't hit Ollama. Admins pass ?refresh=1 to force
+        // a fresh upstream fetch (Ollama models view).
+        const wantRefresh = /^(1|true|yes)$/i.test(
+          url.searchParams.get("refresh") ?? url.searchParams.get("force") ?? "",
+        );
+        const models = await registry.listAllModelsCached(wantRefresh && admin);
         return json({ models });
       }