fixes

annexual_promot

@@ -0,0 +1,6 @@
+--- HIGH PRIORITY
+Being abble to :
+Disable features / Enable it:
+ie : thinking, deep search, etc...
+
+Downloading models throught a catalog : ollama only

another_prompt

@@ -1,3 +1,4 @@
+--- VERY HIGH PRIORITY
 fait en sorte de pouvoir upload un avatar (Servir un cdn en local via fichier upload sur le serveur dans un dossier à côté de la codebase)
 
 cdn comme sa
@@ -21,3 +22,13 @@ Rate-limit per change / per account de:
 file length maxi: 5MB
 
 SYSTEM DE CHANGEMENT DE MDP:
+Donc avoir une upload box, propre, adapative etc
+
+---- MEDIUM PRIORITY
+ADAPTATION MOBILE:
+GÉRER LES MODALS, LES BOXES, LES BOARDING, LES CHAT BOXES AUSSI POUR PETIT ÉCRAN, ADAPTER LEUR WEBVIEW PORTAGE, DIALOG BOX, CONFIRMATON, TRAY CHANNELS SIDE LIST... 
+DRAG AND SNAP LIKE PHONES, SCROLLING FOR OPENGING/EXITING ETC..
+
+--- LOW PRIORITY
+Système de service mails aussi pour les comptes (optionel) pour recovery passwords, dans le page de login:
+  - Recover password with emails. smtp server,

docs/ARCHITECTURE.md

@@ -22,7 +22,20 @@ browser ──► Bun.serve (src/index.ts)
    are auto-titled from it.
 4. Non-streaming → `driver.chat()` → JSON. Streaming → `driver.chatStream()`
    → SSE events `token` / `done` / `error`. The full reply is persisted after a
-   successful stream (partial text + `[interrupted]` marker on failure).
+   successful stream; on error the partial text is kept as-is (no internal
+   markers). Server logs each stream open/done/abort per driver when
+   `DRIVER_DEBUG=true`.
+
+Streaming guarantees (ChatGPT-style, no buffering):
+
+- Tokens are forwarded via `controller.enqueue` the moment they arrive from the
+  LLM; disconnects are detected (`req.signal`, guarded `enqueue`,
+  stream `cancel()`) and propagated upstream via an `AbortController` passed as
+  `opts.signal`, so the model stops generating when the tab closes.
+- `Bun.serve` runs with `idleTimeout: 0` — the default 10 s would kill slow
+  streams (cold model load, thinking pauses) mid-generation.
+- Responses carry `X-Accel-Buffering: no` and the NixOS nginx vhost sets
+  `proxy_buffering off`, otherwise the proxy would buffer the whole reply.
 
 ## Database schema
 

src/index.ts

@@ -114,6 +114,10 @@ async function readJson(req: Request): Promise<{ ok: boolean; body: Record<strin
 const server = Bun.serve({
   port: config.port,
   hostname: config.host,
+  // SSE chat streams can legitimately go silent for a while (cold model load,
+  // slow/thinking models). Bun's default 10s idle timeout would kill them
+  // mid-generation — disable it (0). Nginx in front has its own timeouts.
+  idleTimeout: 0,
   async fetch(req, server) {
     const url = new URL(req.url);
     const path = url.pathname;