changed how the fallback works

bot.py

@@ -34,6 +34,7 @@ DM_DIR       = LOG_DIR    / "dm"  # one memory file per person so dms never end
 
 ERROR_FILE   = LOG_DIR    / "errorlog.txt"
 MEMORY_FILE  = LOG_DIR    / "memory.txt"
+MEDIA_FILE   = MEDIA_DIR  / "media.txt"
 THREADS_FILE = CONFIG_DIR / "threads.txt"
 PROMPT_FILE  = CONFIG_DIR / "system_prompt.txt"
 BLOCKED_FILE   = CONFIG_DIR / "blocked.txt"  # edited with admin.py, re-read on every message
@@ -45,7 +46,7 @@ DEFAULT_SETTINGS = {"admin_id": "1000000000000001", "moralfag_id": "100000000000
 def init_files():
     for d in (LOG_DIR, MEDIA_DIR, CONFIG_DIR, DM_DIR):
         d.mkdir(parents=True, exist_ok=True)
-    for f in (ERROR_FILE, MEMORY_FILE):
+    for f in (ERROR_FILE, MEMORY_FILE, MEDIA_FILE):
         f.touch(exist_ok=True)
     if not THREADS_FILE.exists():
         THREADS_FILE.write_text("6\n")
@@ -91,7 +92,7 @@ QUIET_COOLDOWN = 600  # seconds of silence before she's allowed to send a random
 MEDIA_MIMES = ("image/png", "image/jpeg", "image/webp", "image/heic", "image/heif",
                "audio/", "video/", "application/pdf", "text/plain")  # what gemini can actually read
 MEDIA_MAX_BYTES = 10 * 1024 * 1024  # total per message, gemini caps inline request size (~20mb and base64 adds a third)
-MEDIA_SAVE_MAX_BYTES = 19 * 1024 * 1024  # per file we keep and offer, has to fit the server's upload limit or she can't send it back
+MEDIA_SAVE_MAX_BYTES = 8 * 1024 * 1024  # per file we keep, stays under discord's 10mb upload limit so she can send it back
 
 def log_error(text: str):
     with ERROR_FILE.open("a", encoding="utf-8") as f:
@@ -196,12 +197,17 @@ async def call_main(payload: str) -> tuple[bool, str]:
     async with _main_lock:
         if not (FALLBACK_URI and time.time() < quota_until):  # skip gemini while it's known to be out of quota
             ok, result, quota = await call_backend(MAIN_URI, payload)
-            if not quota:
+            if ok:
                 return ok, result
+            if quota:
                 quota_until = time.time() + QUOTA_COOLDOWN
                 if not FALLBACK_URI:
                     return False, QUOTA_MSG
                 print(f"[{now_str()}] gemini is out of quota, using the fallback for the next {QUOTA_COOLDOWN // 60} min")
+            elif not FALLBACK_URI:
+                return False, result
+            else:
+                print(f"[{now_str()}] main backend failed ({result}), trying the fallback")  # dead server / 502 / any other error
         ok, result, _ = await call_backend(FALLBACK_URI, payload)
         return ok, result
 
@@ -251,13 +257,14 @@ async def send_reply(message, text: str):
 
 vector_db_ready = False  # gets set to True after first rebuild so we don't keep rebuilding
 
-MEDIA_EXTS = (".gif", ".png", ".jpg", ".jpeg", ".webp", ".mp4", ".webm", ".mov", ".mkv")
+saved_media = set()  # file names inside MEDIA_DIR
 
-def scan_media() -> set:
-    # whatever is in the media folder is what she can send, skips media.txt and anything too big to upload
-    return {p.name for p in MEDIA_DIR.iterdir() if p.is_file() and p.suffix.lower() in MEDIA_EXTS and p.stat().st_size <= MEDIA_SAVE_MAX_BYTES}
-
-saved_media = scan_media()  # file names inside MEDIA_DIR
+with MEDIA_FILE.open("r", encoding="utf-8") as f:
+    for line in f:
+        if "saved media:" in line:
+            name = line.split("saved media:", 1)[1].strip()
+            if (MEDIA_DIR / name).is_file():  # old entries were urls (those expire), they don't match a file so they get dropped
+                saved_media.add(name)
 
 async def steal_media(url: str, name: str):
     # discord cdn links are signed and die after a day or so, so we keep our own copy
@@ -281,6 +288,8 @@ async def steal_media(url: str, name: str):
         saved_media.discard(name)
         log_error(f"[{now_str()}] failed to save media {url}: {e}")
         return
+    with MEDIA_FILE.open("a", encoding="utf-8") as f:
+        f.write(f"[{now_str()}] saved media: {name}\n")
 
 async def steal_attachments(message):
     #steal images and gifs and stuff (always do this regardless of what else fires)
@@ -429,8 +438,6 @@ def sync_reactions(guild):
 def reaction_tool_prompt(guild) -> str:
     # the "tool": instead of (or along with) words she can write a [sticker: name] / [emoji: name] / [media: name] tag and
     # split_reactions picks it back out of the reply. same call as the reply, so no extra latency or gemini quota.
-    saved_media.clear()  # ls the folder again so files added / renamed (describe.py) while she's running show up
-    saved_media.update(scan_media())
     sync_reactions(guild)
     data = load_reactions()
     have = {