hello

README.md

@@ -64,6 +64,23 @@ $ python3 bot.py
 ```
 she only reads `general` and `announcements`, and every dm gets its own memory in `./files/log/dm/` so it never leaks into the server
 
+# fallback
+if you're using gemini you can have a local ollama `main.py` take over when gemini says you're out of quota. the wizard asks about it (it's the third line of `./files/config/connection.txt`), it needs to run on another port so run it as
+```console
+$ PORT=8766 python3 main.py
+```
+the bot goes back to trying gemini after an hour. if there's no fallback (or it's down) she just says "usage limit reached, see you in a month"
+
+# admin
+`admin.py` is the admin panel, it only edits files on the box so nobody on discord can touch it. the bot re-reads everything on every message, no restart needed
+```console
+$ python3 admin.py list
+$ python3 admin.py edit <blocked|prompt|settings|reactions|params> #opens $EDITOR, vim if you don't have one
+$ python3 admin.py block <user id>
+$ python3 admin.py unblock <user id>
+```
+`reactions.json` is where the stickers / emojis / media she can send live, the bot adds new ones with an empty description and you fill them in. media with no description never gets offered
+
 # colab
 if your pc can't run the model you can host `main.py` on colab, run `setup_wizard.py` on both sides and say yes to the colab question
 ```console

admin.py

@@ -0,0 +1,107 @@
+#!/usr/bin/env python3
+"""
+admin panel for faggot-chan. runs on the box that has the files, so there's nothing on discord to reach
+(not even with admin perms). the bot re-reads all of these on every message, no restart needed.
+
+    admin.py list
+    admin.py edit <blocked|prompt|settings|reactions|params>   opens it in $EDITOR (vim if unset)
+    admin.py block <user id> [note]
+    admin.py unblock <user id>
+"""
+import json
+import os
+import shlex
+import subprocess
+import sys
+from pathlib import Path
+
+SCRIPT_DIR = Path(__file__).resolve().parent
+BOT_CONFIG = SCRIPT_DIR / "files" / "config"
+
+TARGETS = {
+    "blocked":   BOT_CONFIG / "blocked.txt",
+    "prompt":    BOT_CONFIG / "system_prompt.txt",
+    "settings":  BOT_CONFIG / "settings.json",
+    "reactions": BOT_CONFIG / "reactions.json",
+    "params":    SCRIPT_DIR / "config" / "params.json",  # main.py side, ollama options
+}
+
+
+def edit(name: str):
+    path = TARGETS[name]
+    path.parent.mkdir(parents=True, exist_ok=True)
+    original = path.read_bytes() if path.exists() else None  # to put back if the json ends up broken and they give up
+    if original is None:
+        path.write_text("{}\n" if path.suffix == ".json" else "", encoding="utf-8")
+    editor = shlex.split(os.environ.get("EDITOR", "vim"))
+    while True:
+        subprocess.call([*editor, str(path)])
+        if path.suffix != ".json":
+            return
+        try:
+            json.loads(path.read_text(encoding="utf-8"))
+            return
+        except json.JSONDecodeError as e:
+            print(f"that's not valid json: {e}")
+            if input("open it again? (y/n): ").strip().lower() in ("y", "yes"):
+                continue
+            if original is None:
+                path.unlink()
+            else:
+                path.write_bytes(original)
+            print("reverted")
+            return
+
+
+def blocked_ids(lines: list[str]) -> set[str]:
+    return {l.split("#")[0].strip() for l in lines if l.split("#")[0].strip()}
+
+
+def block(user_id: str, note: str = ""):
+    path = TARGETS["blocked"]
+    lines = path.read_text(encoding="utf-8").splitlines() if path.exists() else []
+    if user_id in blocked_ids(lines):
+        print(f"{user_id} is already blocked")
+        return
+    lines.append(f"{user_id}  # {note}" if note else user_id)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text("\n".join(lines) + "\n", encoding="utf-8")
+    print(f"blocked {user_id}")
+
+
+def unblock(user_id: str):
+    path = TARGETS["blocked"]
+    lines = path.read_text(encoding="utf-8").splitlines() if path.exists() else []
+    kept = [l for l in lines if l.split("#")[0].strip() != user_id]
+    if len(kept) == len(lines):
+        print(f"{user_id} wasn't blocked")
+        return
+    path.write_text("\n".join(kept) + "\n", encoding="utf-8")
+    print(f"unblocked {user_id}")
+
+
+def usage():
+    print(__doc__.strip())
+    sys.exit(1)
+
+
+def main():
+    args = sys.argv[1:]
+    if not args:
+        usage()
+    cmd = args[0]
+    if cmd == "list":
+        for name, path in TARGETS.items():
+            print(f"{name:10} {path}{'' if path.exists() else '  (not made yet)'}")
+    elif cmd == "edit" and len(args) == 2 and args[1] in TARGETS:
+        edit(args[1])
+    elif cmd == "block" and len(args) >= 2 and args[1].isdigit():
+        block(args[1], " ".join(args[2:]))
+    elif cmd == "unblock" and len(args) == 2 and args[1].isdigit():
+        unblock(args[1])
+    else:
+        usage()
+
+
+if __name__ == "__main__":
+    main()

bot.py

@@ -37,6 +37,11 @@ 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
+SETTINGS_FILE  = CONFIG_DIR / "settings.json"
+REACTIONS_FILE = CONFIG_DIR / "reactions.json"  # sticker/emoji/media descriptions she picks from
+
+DEFAULT_SETTINGS = {"admin_id": "1000000000000001", "moralfag_id": "1000000000000002", "faglimit": 2}  # faglimit: ignore every x reply
 
 def init_files():
     for d in (LOG_DIR, MEDIA_DIR, CONFIG_DIR, DM_DIR):
@@ -47,6 +52,10 @@ def init_files():
         THREADS_FILE.write_text("6\n")
     if not PROMPT_FILE.exists():  # otherwise every single message dies on a missing file
         PROMPT_FILE.write_text("you are faggot-chan, a chaotic discord bot. edit files/config/system_prompt.txt\n", encoding="utf-8")
+    if not BLOCKED_FILE.exists():
+        BLOCKED_FILE.write_text("# one discord user id per line, edit with admin.py edit blocked\n1000000000000003  # fuck these two in particular\n1000000000000004\n", encoding="utf-8")
+    if not SETTINGS_FILE.exists():
+        SETTINGS_FILE.write_text(json.dumps(DEFAULT_SETTINGS, indent=2) + "\n", encoding="utf-8")
 
 init_files()
 
@@ -57,23 +66,27 @@ server = int(_config_lines[1])
 _conn_lines = [l.strip() for l in (CONFIG_DIR / "connection.txt").read_text(encoding="utf-8").splitlines() if l.strip() and not l.strip().startswith("#")]
 MAIN_URI   = _conn_lines[0]  # ws:// or wss:// address of main.py
 MAIN_TOKEN = _conn_lines[1]  # shared secret main.py checks on every payload
+FALLBACK_URI = _conn_lines[2] if len(_conn_lines) > 2 else None  # optional, local main.py (ollama) that takes over when gemini runs out of quota, same token
 
 intents = discord.Intents.default()
 intents.message_content = True
 intents.members = True
 client = discord.Client(intents=intents)
 
-block_id = ["1000000000000003", "1000000000000004"] #fuck these two in particular 
-
-moralfag_id = "1000000000000002"
 moralfag_counter = 0
 moralfag_skip = False
-faglimit = 2  #ignore every x reply
 
 last_server_message = 0.0  # unix timestamp of last message anyone sent in the server
 dm_sent = False
 interaction_count = 0
 
+QUOTA_MSG = "usage limit reached, see you in a month"
+QUOTA_COOLDOWN = 3600  # after gemini says it's out of quota go straight to the fallback for this long (seconds), then try it again
+quota_until = 0.0  # unix timestamp
+
+MAX_REACTION_OPTIONS = 25  # per kind (stickers/emojis/media), random sample so the prompt stays small
+REACTION_TAG = re.compile(r'\[(sticker|emoji|media):\s*([^\]]+?)\s*\]', re.I)
+
 QUIET_COOLDOWN = 600  # seconds of silence before she's allowed to send a random message (10 mins)
 
 MEDIA_MIMES = ("image/png", "image/jpeg", "image/webp", "image/heic", "image/heif",
@@ -85,10 +98,27 @@ def log_error(text: str):
     with ERROR_FILE.open("a", encoding="utf-8") as f:
         f.write(text + "\n")
 
+def load_blocked() -> set[str]:
+    # re-read on every message so admin.py edits apply without restarting
+    try:
+        lines = [l.split("#")[0].strip() for l in BLOCKED_FILE.read_text(encoding="utf-8").splitlines()]
+        return {l for l in lines if l}
+    except Exception as e:
+        log_error(f"[{now_str()}] error reading blocked list: {e}")
+        return set()
+
+def load_settings() -> dict:
+    try:
+        return {**DEFAULT_SETTINGS, **json.loads(SETTINGS_FILE.read_text(encoding="utf-8"))}
+    except Exception as e:
+        log_error(f"[{now_str()}] error reading settings: {e}")
+        return dict(DEFAULT_SETTINGS)
+
 def log_message(text: str, file: Path = MEMORY_FILE):
     text = " ".join(text.splitlines())  # one message = one line, otherwise multiline messages break the history parsing
     with file.open("a", encoding="utf-8") as f:
         f.write(text + "\n")
+    return text.strip()  # what shortterm will read back for this line, build_payload uses it to find the message she's answering
 
 def load_shortterm(file: Path = MEMORY_FILE) -> list[str]:
     try:
@@ -105,21 +135,23 @@ def load_longterm() -> list[str]:
         log_error(f"[{now_str()}] error reading longterm: {e}")
         return []
 
-def build_payload(user_input: str, rebuild: bool = False, media: list = None, logged: bool = False, dm_file: Path = None) -> str:
+def build_payload(user_input: str, rebuild: bool = False, media: list = None, logged: str = None, dm_file: Path = None, tools: str = "") -> str:
     global vector_db_ready
     shortterm = load_shortterm(dm_file or MEMORY_FILE)
-    if logged and shortterm:
-        shortterm = shortterm[:-1]  # the message she's answering is already the last line, it goes in the final turn instead
+    if logged and logged in shortterm:
+        # cut at the message she's answering (it goes in the final turn instead), found by content and not by position:
+        # people keep talking while she's busy so it's not always the last line, and whatever got logged after it isn't part of this reply
+        shortterm = shortterm[:len(shortterm) - 1 - shortterm[::-1].index(logged)]
     if dm_file:
         longterm = []  # dms stay out of the shared rag so private stuff can't come up in a public channel
         should_rebuild = False
     else:
-        longterm = load_longterm()
         should_rebuild = rebuild or not vector_db_ready  # always rebuild on first ever message
+        longterm = load_longterm() if should_rebuild else []  # only sent when it's needed, the backend asks for it (need_longterm) if it lost it
         if should_rebuild:
             vector_db_ready = True
     return json.dumps({
-        "system_prompt":   load_system_prompt(),
+        "system_prompt":   load_system_prompt() + tools,
         "user_input":      user_input,
         "shortterm":       shortterm,
         "longterm":        longterm,
@@ -128,34 +160,56 @@ def build_payload(user_input: str, rebuild: bool = False, media: list = None, lo
         "auth_token":      MAIN_TOKEN,
     })
 
-_main_ws = None  # kept open across messages instead of reconnecting every time
-_main_lock = None  # one send/recv at a time on the shared socket, made on first use so it lives in discord's loop
+_main_ws = {}  # one connection per backend uri (gemini / fallback), kept open across messages instead of reconnecting every time
+_main_lock = None  # one send/recv at a time on the shared sockets, made on first use so it lives in discord's loop
+
+async def call_backend(uri: str, payload: str) -> tuple[bool, str, bool]:
+    # last bool is True when the backend said it's out of quota
+    resent = False
+    for attempt in range(2):  # retry once with a fresh connection if the cached one turned out dead
+        try:
+            ws = _main_ws.get(uri)
+            if ws is None:
+                ws = _main_ws[uri] = await websockets.connect(uri, open_timeout=None, ping_timeout=None)
+            await ws.send(payload)
+            data = json.loads(await ws.recv())
+            if data.get("need_longterm") and not resent:  # backend has no vector db (restarted / first time as fallback), send it the longterm once
+                resent = True
+                p = json.loads(payload)
+                p["longterm"] = load_longterm()
+                p["rebuild_vectors"] = True
+                payload = json.dumps(p)
+                await ws.send(payload)
+                data = json.loads(await ws.recv())
+            if "error" in data:
+                return False, data["error"], bool(data.get("quota"))
+            return True, data["ok"], False
+        except Exception as e:
+            _main_ws.pop(uri, None)
+            if attempt == 0:
+                continue  # one retry with a fresh connection before giving up
+            return False, f"[{now_str()}] websocket error: {e}", False
 
 async def call_main(payload: str) -> tuple[bool, str]:
-    global _main_ws, _main_lock
+    global _main_lock, quota_until
     if _main_lock is None:
         _main_lock = asyncio.Lock()
     async with _main_lock:
-        for attempt in range(2):  # retry once with a fresh connection if the cached one turned out dead
-            try:
-                if _main_ws is None:
-                    _main_ws = await websockets.connect(MAIN_URI, open_timeout=None, ping_timeout=None)
-                await _main_ws.send(payload)
-                raw = await _main_ws.recv()
-                data = json.loads(raw)
-                if "error" in data:
-                    return False, data["error"]
-                return True, data["ok"]
-            except Exception as e:
-                _main_ws = None
-                if attempt == 0:
-                    continue  # one retry with a fresh connection before giving up
-                return False, f"[{now_str()}] websocket error: {e}"
+        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:
+                return ok, result
+            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")
+        ok, result, _ = await call_backend(FALLBACK_URI, payload)
+        return ok, result
 
 async def handle_error(error_msg: str, notify_channel=None, notify_user=None):
     print(error_msg)
     log_error(error_msg)
-    notice = "someone tell <@1000000000000001> there is a problem with my ai"
+    notice = QUOTA_MSG if error_msg == QUOTA_MSG else f"someone tell <@{load_settings()['admin_id']}> there is a problem with my ai"
     try:
         if notify_channel:
             await notify_channel.send(notice)
@@ -347,71 +401,83 @@ async def random_message():
 @client.event
 async def on_ready():
     asyncio.get_running_loop().set_exception_handler(ignore_connection_closed_ok)
+    for guild in client.guilds:
+        sync_reactions(guild)
     client.loop.create_task(random_message())
     client.loop.create_task(random_dm())
     print(f"faggot-chan is active as {client.user}")
 
 
-async def on_random(message) -> bool:
-    # single weighted pick between sticker/emoji/media, used as a substitute for an actual
-    # llm reply — only called once on_message already decided this message gets a response.
-    # returns True if it fired (caller should skip the llm call), False if it should proceed normally.
-    stickers = list(message.guild.stickers)
-    emojis = [str(e) for e in message.guild.emojis]
-    media_files = [n for n in saved_media if (MEDIA_DIR / n).is_file()]  # in case something got deleted by hand
+def load_reactions() -> dict:
+    # re-read every time so admin.py edit reactions applies without a restart
+    try:
+        data = json.loads(REACTIONS_FILE.read_text(encoding="utf-8"))
+    except (FileNotFoundError, json.JSONDecodeError):
+        data = {}
+    for k in ("stickers", "emojis", "media"):
+        data.setdefault(k, {})  # name -> short description
+    return data
 
-    options = []
-    if stickers:
-        options.append(("sticker", 0.10))
-    if emojis:
-        options.append(("emoji", 0.13))
-    if media_files:
-        options.append(("media", 0.07))  # ~30% combined chance of firing when all three are available
+def sync_reactions(guild):
+    # anything new gets added with an empty description, fill them in with admin.py edit reactions
+    data = load_reactions()
+    changed = False
+    for key, names in (("stickers", [s.name for s in guild.stickers]), ("emojis", [e.name for e in guild.emojis]), ("media", sorted(saved_media))):
+        for n in names:
+            if n not in data[key]:
+                data[key][n] = ""
+                changed = True
+    if changed:
+        REACTIONS_FILE.write_text(json.dumps(data, indent=2, ensure_ascii=False) + "\n", encoding="utf-8")
 
-    if not options:
-        return False
-
-    roll = random.random()
-    cumulative = 0.0
-    for kind, weight in options:
-        cumulative += weight
-        if roll >= cumulative:
+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.
+    sync_reactions(guild)
+    data = load_reactions()
+    have = {
+        "stickers": {s.name for s in guild.stickers},
+        "emojis":   {e.name for e in guild.emojis},
+        "media":    {n for n in saved_media if (MEDIA_DIR / n).is_file()},
+    }
+    sections = []
+    for key in ("stickers", "emojis", "media"):
+        options = [(n, d) for n, d in data[key].items() if n in have[key] and (d or key != "media")]  # media without a description tells her nothing
+        if not options:
             continue
+        options = random.sample(options, min(len(options), MAX_REACTION_OPTIONS))
+        sections.append(f"{key}:\n" + "\n".join(f"- {n}: {d}" if d else f"- {n}" for n, d in options))
+    if not sections:
+        return ""
+    return ("\n\nREACTIONS: you can send a sticker, an emoji reaction or a media file instead of (or along with) words, but only when it really fits the moment, most of the time just reply normally. "
+            "to do it put ONE tag at the end of your message: [sticker: name], [emoji: name] or [media: name], using exactly one of the names below. "
+            "the tag alone is fine if you have nothing to say. never invent names.\n" + "\n".join(sections))
 
+def split_reactions(text: str) -> tuple[str, list]:
+    return REACTION_TAG.sub("", text).strip(), [(k.lower(), n) for k, n in REACTION_TAG.findall(text)]
+
+async def send_reaction(message, kind: str, name: str):
+    # looks the name up for real, a made up name just does nothing
+    name = name.strip().strip(":").lower()
+    try:
         if kind == "sticker":
-            chosen = random.choice(stickers)
-            try:
+            chosen = next((s for s in message.guild.stickers if s.name.lower() == name), None)
+            if chosen:
                 await message.channel.send(stickers=[chosen])
                 log_message(f"[{now_str()}] Faggot-chan sent sticker: {chosen.name}")
                 print(f"[{now_str()}] Faggot-chan sent sticker: {chosen.name}")
-            except Exception as e:
-                log_error(f"[{now_str()}] error: {e}")
-                await message.reply("someone tell <@1000000000000001> there is a problem with my ai")
-                print(f"Sticker send failed: {e}")
-            return True
-
-        if kind == "emoji":
-            try:
-                await message.add_reaction(random.choice(emojis))
-            except Exception as e:
-                log_error(f"[{now_str()}] error: {e}")
-                await message.reply("someone tell <@1000000000000001> there is a problem with my ai")
-                print(f"Emoji reaction failed: {e}")
-            return True
-
-        if kind == "media":
-            chosen = random.choice(media_files)
-            try:
+        elif kind == "emoji":
+            chosen = next((e for e in message.guild.emojis if e.name.lower() == name), None)
+            if chosen:
+                await message.add_reaction(chosen)
+        elif kind == "media":
+            chosen = next((n for n in saved_media if n.lower() == name and (MEDIA_DIR / n).is_file()), None)  # only names we saved ourselves
+            if chosen:
                 await message.channel.send(file=discord.File(MEDIA_DIR / chosen))
                 log_message(f"[{now_str()}] Faggot-chan sent media: {chosen}")
                 print(f"[{now_str()}] Faggot-chan sent media: {chosen}")
-            except Exception as e:
-                log_error(f"[{now_str()}] error: {e}")
-                await message.reply("someone tell <@1000000000000001> there is a problem with my ai")
-                print(f"Media send failed: {e}")
-            return True
-
-    return False
+    except Exception as e:
+        await handle_error(f"[{now_str()}] failed to send {kind} {name}: {e}", notify_channel=message.channel)
 
 
 @client.event
@@ -426,7 +492,7 @@ async def on_message(message):
     # update last activity so random_message knows the server isn't quiet
     last_server_message = time.time()
 
-    if str(message.author.id) in block_id:
+    if str(message.author.id) in load_blocked():
          return
 
     # drop guild messages outside allowed channels before doing anything else, thank you claude
@@ -452,15 +518,15 @@ async def on_message(message):
                 if replied_message.author == client.user:
                     replied_to_bot = True
                 replied_content = resolve_mentions(replied_message.content, message.guild)
-                log_message(f'[{now_str()}] {message.author} replied to {replied_message.author} saying: "{replied_content}" with reply: "{content}"', mem_file)
+                logged_line = log_message(f'[{now_str()}] {message.author} replied to {replied_message.author} saying: "{replied_content}" with reply: "{content}"', mem_file)
             else:
-                log_message(f'[{now_str()}] {message.author} said: {content}', mem_file)
+                logged_line = log_message(f'[{now_str()}] {message.author} said: {content}', mem_file)
         except Exception as e:
             replied_message = None  # fetch failed (deleted), treat it like a normal message
             log_error(f"[{now_str()}] failed to resolve reply: {e}")
-            log_message(f'[{now_str()}] {message.author} said: {content}', mem_file)
+            logged_line = log_message(f'[{now_str()}] {message.author} said: {content}', mem_file)
     else:
-        log_message(f'[{now_str()}] {message.author} said: {content}', mem_file)
+        logged_line = log_message(f'[{now_str()}] {message.author} said: {content}', mem_file)
 
     interaction_count += 1
     rebuild = (interaction_count % 150 == 0)
@@ -471,7 +537,7 @@ async def on_message(message):
             user_input = f'{message.author} replied to {replied_message.author} saying: "{replied_content}" with reply: "{content}"'
         else:
             user_input = f"{message.author} said in DM: {content}"
-        payload = build_payload(user_input, rebuild, await grab_media(message), True, mem_file)  # a dm is always aimed at her
+        payload = build_payload(user_input, rebuild, await grab_media(message), logged_line, mem_file)  # a dm is always aimed at her
         stop_typing = asyncio.Event()
         asyncio.create_task(keep_typing(message.channel, stop_typing))
         ok, result = await call_main(payload)
@@ -498,22 +564,20 @@ async def on_message(message):
     if not mentioned and random.random() > 0.03: #3% chance of replying if not mentiond
         return
 
-    if str(message.author.id) == moralfag_id:  # only checked once she was actually going to reply, otherwise the skip gets eaten by chatter
+    settings = load_settings()
+    if str(message.author.id) == settings["moralfag_id"]:  # only checked once she was actually going to reply, otherwise the skip gets eaten by chatter
         global moralfag_skip, moralfag_counter
         if moralfag_skip:
             moralfag_skip = False
             return
 
-    if await on_random(message):  # ~30% chance she just sends a sticker/emoji/media instead
-        return
-
     if replied_message:
         user_input = f'{message.author} replied to {replied_message.author} saying: "{replied_content}" with reply: "{content}"'
     else:
         user_input = f"{message.author} said: {content}"
 
     media = await grab_media(message) if client.user in message.mentions else []  # only direct @mentions, not replies or random 3% ones
-    payload = build_payload(user_input, rebuild, media, True)
+    payload = build_payload(user_input, rebuild, media, logged_line, tools=reaction_tool_prompt(message.guild))
     stop_typing = asyncio.Event()
     asyncio.create_task(keep_typing(message.channel, stop_typing))
     ok, result = await call_main(payload)
@@ -521,18 +585,22 @@ async def on_message(message):
     if not ok:
         await handle_error(result, notify_channel=message.channel)
         return
-    if not result.strip():
+    result, reactions = split_reactions(result)  # she can answer with a [sticker: x] style tag instead of words
+    if not result.strip() and not reactions:
         return
 
     try:
-        await send_reply(message, result)
-        print(f"[{now_str()}] {message.author} said: {content}\n[{now_str()}] Faggot-chan said: {result}")
-        log_message(f"[{now_str()}] Faggot-chan said: {result}")
-        if str(message.author.id) == moralfag_id:
-            moralfag_counter += 1
-            if moralfag_counter >= faglimit:
-                moralfag_skip = True
-                moralfag_counter = 0
+        if result.strip():
+            await send_reply(message, result)
+            print(f"[{now_str()}] {message.author} said: {content}\n[{now_str()}] Faggot-chan said: {result}")
+            log_message(f"[{now_str()}] Faggot-chan said: {result}")
+            if str(message.author.id) == settings["moralfag_id"]:
+                moralfag_counter += 1
+                if moralfag_counter >= settings["faglimit"]:
+                    moralfag_skip = True
+                    moralfag_counter = 0
+        for kind, name in reactions[:1]:  # one is enough
+            await send_reaction(message, kind, name)
     except Exception as e:
         await handle_error(f"[{now_str()}] failed to send reply: {e}", notify_channel=message.channel)
 

gemini.py

@@ -9,12 +9,16 @@ import websockets
 from pathlib import Path
 from websockets.server import serve
 from google import genai
-from google.genai import types
-from utils import nots, now_str
+from google.genai import types, errors
+from utils import nots, now_str, make_chunks, pick_chunks
 
 chatmodel = "gemini-3.5-flash-lite"
 embedmodel = "gemini-embedding-001"
 embeddim = 768  # default is 3072, the cosine loop below is pure python so smaller = way faster
+RAG_THRESHOLD = 0.6  # chunks scoring under this aren't injected at all, watch the "best rag score" print and tweak it
+PORT = int(os.environ.get("PORT", 8765))
+RETRIES = 5  # how many times we retry a 503 (5s apart) before giving up
+QUOTA_MSG = "usage limit reached, see you in a month"
 
 # resolve config relative to this script, not whatever dir it happens to get launched from
 SCRIPT_DIR = Path(__file__).resolve().parent
@@ -51,6 +55,8 @@ SAFETY_SETTINGS = [
 
 # Each element in the VECTOR_DB will be a tuple (chunk, embedding)
 VECTOR_DB = []
+RAG_CHUNKS = []  # chunks from the last rebuild, so a partial (quota limited) rebuild can be finished without the bot resending everything
+rag_loaded = False  # False until the bot sends longterm at least once, after a restart we ask for it
 
 def embed(texts, task_type):
     # batches of 100 since that's the api limit per call
@@ -77,7 +83,9 @@ def retrieve(query, top_n=5):
         similarity = cosine_similarity(query_embedding, embedding)
         similarities.append((chunk, similarity))
     similarities.sort(key=lambda x: x[1], reverse=True)
-    return similarities[:top_n]
+    if similarities:
+        print(f"[{now_str()}] best rag score: {similarities[0][1]:.2f}")
+    return pick_chunks(similarities, top_n, RAG_THRESHOLD)
 
 def rebuild_vector_db(longterm_lines: list[str]):
     global VECTOR_DB
@@ -97,6 +105,7 @@ def rebuild_vector_db(longterm_lines: list[str]):
 
 
 async def geminichat(websocket):
+    global RAG_CHUNKS, rag_loaded
     # loop over every message on this connection instead of handling one and returning —
     # returning from this function is what makes the websockets lib close the connection
     async for raw in websocket:
@@ -110,14 +119,21 @@ async def geminichat(websocket):
             system_prompt = payload["system_prompt"]
             user_input    = payload["user_input"]
             shortterm     = payload.get("shortterm", [])   # list of strings
-            longterm      = payload.get("longterm", [])    # list of strings
+            longterm      = payload.get("longterm", [])    # list of strings, only there on a rebuild
             media         = payload.get("media", [])       # list of {"mime_type", "data" (base64)}
 
-            # rebuild vector db if bot signals it (every 150 interactions bot sends full longterm),
-            # or if we come up empty but longterm isn't — e.g. we just restarted and lost VECTOR_DB
+            # bot only sends longterm on a rebuild, if we never got it (just restarted) ask for it and the bot resends
+            if payload.get("rebuild_vectors"):
+                RAG_CHUNKS = make_chunks(longterm)
+                rag_loaded = True
+            elif not rag_loaded:
+                await websocket.send(json.dumps({"need_longterm": True}))
+                continue
+
+            # rebuild vector db if bot signals it (every 150 interactions bot sends full longterm)
             # also keeps going while the db is still catching up after a partial (quota limited) rebuild
-            if payload.get("rebuild_vectors") or len(VECTOR_DB) < len(longterm):
-                await asyncio.to_thread(rebuild_vector_db, longterm)
+            if payload.get("rebuild_vectors") or len(VECTOR_DB) < len(RAG_CHUNKS):
+                await asyncio.to_thread(rebuild_vector_db, RAG_CHUNKS)
                 print(f"[{now_str()}] Vector DB reconciled, {len(VECTOR_DB)} chunks")
 
             user_query = nots(user_input)
@@ -131,7 +147,7 @@ async def geminichat(websocket):
             # rag chunks go in system so they don't pollute turn history
             rag_block = ""
             if retrieved_knowledge:
-                rag_block = "\n\nOlder relevant messages for context:\n" + "\n".join(f"- {chunk}" for chunk, _ in retrieved_knowledge)
+                rag_block = "\n\nOlder relevant messages for context:\n" + "\n".join(f"- {chunk.replace(chr(10), ' | ')}" for chunk, _ in retrieved_knowledge)
 
             # lines to skip — not real dialogue
             skip_fragments = ("sent sticker:", "sent media:", "sent a random dm to ", "saved media:")
@@ -160,7 +176,8 @@ async def geminichat(websocket):
                     text = rest.rsplit(' with reply: "', 1)[-1]
                     if text.endswith('"'):
                         text = text[:-1]
-                    return speaker.strip(), text.strip()
+                    target = rest[len("replied to "):].split(" saying: ", 1)[0]  # keep who they were talking to, otherwise replies lose it
+                    return speaker.strip(), f"(replying to {target}) {text.strip()}"
                 if rest.startswith("said: "):
                     return speaker.strip(), rest[len("said: "):].strip()
                 return None, None
@@ -196,7 +213,7 @@ async def geminichat(websocket):
                 messages.append({"role": "user", "content": "\n".join(current_user_lines)})
 
             # final isolated turn — the specific message she needs to reply to
-            messages.append({"role": "user", "content": f"[responding to]: {nots(user_input)}"})
+            messages.append({"role": "user", "content": f"[responding to]: {nots(user_input)}\n(that is the newest message and the only one you answer, the lines above are just context. only the name that starts it is speaking to you, everyone else is someone else)"})
 
             # make sure we start with a user turn (gemini rejects empty parts so no empty string here)
             if messages and messages[0]["role"] != "user":
@@ -220,13 +237,31 @@ async def geminichat(websocket):
                 safety_settings=SAFETY_SETTINGS,
             )
 
+            quota_hit = False
             async with chat_semaphore:
-                response = await asyncio.to_thread(
-                    client.models.generate_content,
-                    model=chatmodel,
-                    contents=contents,
-                    config=config,
-                )
+                for attempt in range(RETRIES + 1):
+                    try:
+                        response = await asyncio.to_thread(
+                            client.models.generate_content,
+                            model=chatmodel,
+                            contents=contents,
+                            config=config,
+                        )
+                        break
+                    except errors.APIError as e:
+                        if e.code == 429 and "PerMinute" not in str(e):  # out of quota for real, not just a rate limit blip
+                            quota_hit = True
+                            break
+                        if (e.code == 503 or e.code == 429) and attempt < RETRIES:  # overloaded / rate limited, usually clears up in a few secs
+                            print(f"[{now_str()}] gemini {e.code}, retrying the same request in 5s ({attempt + 1}/{RETRIES})")
+                            await asyncio.sleep(5)  # sleeping with the semaphore held on purpose, the retry goes first
+                            continue
+                        raise
+
+            if quota_hit:
+                print(f"[{now_str()}] gemini is out of quota")
+                await websocket.send(json.dumps({"error": QUOTA_MSG, "quota": True}))  # bot.py switches to its fallback on this
+                continue
 
             # response.text is None when gemini blocks the reply anyway
             if not response.text:
@@ -253,8 +288,8 @@ async def geminichat(websocket):
 
 
 async def main():
-    print(f"[{now_str()}] hosting on ws://0.0.0.0:8765")
-    async with serve(geminichat, "0.0.0.0", 8765, ping_interval=None, ping_timeout=None, max_size=32 * 1024 * 1024):  # default 1mb would kill any media payload
+    print(f"[{now_str()}] hosting on ws://0.0.0.0:{PORT}")
+    async with serve(geminichat, "0.0.0.0", PORT, ping_interval=None, ping_timeout=None, max_size=32 * 1024 * 1024):  # default 1mb would kill any media payload
         await asyncio.Future()
 
 

main.py

@@ -3,15 +3,18 @@ import asyncio
 import base64
 import hmac
 import json
+import os
 import re  # thank you claude
 import ollama
 import websockets
 from pathlib import Path
 from websockets.server import serve
-from utils import nots, now_str
+from utils import nots, now_str, make_chunks, pick_chunks
 
 chatmodel = "hf.co/mradermacher/Huihui-gemma-4-E4B-it-abliterated-GGUF:IQ4_XS"
 embedmodel = 'hf.co/CompendiumLabs/bge-base-en-v1.5-gguf'
+RAG_THRESHOLD = 0.65  # chunks scoring under this aren't injected at all, watch the "best rag score" print and tweak it
+PORT = int(os.environ.get("PORT", 8765))  # PORT=8766 python3 main.py when it runs as the fallback next to gemini.py
 
 # resolve config relative to this script, not whatever dir it happens to get launched from
 SCRIPT_DIR = Path(__file__).resolve().parent
@@ -31,6 +34,7 @@ chat_semaphore = asyncio.Semaphore(1)
 # Each element in the VECTOR_DB will be a tuple (chunk, embedding)
 # The embedding is a list of floats, for example: [0.1, 0.04, -0.34, 0.21, ...]
 VECTOR_DB = []
+rag_loaded = False  # False until the bot sends longterm at least once, after a restart we ask for it
 
 def add_chunk_to_database(chunk):
     embedding = ollama.embed(model=embedmodel, input=chunk)['embeddings'][0]
@@ -49,7 +53,9 @@ def retrieve(query, top_n=5):
         similarity = cosine_similarity(query_embedding, embedding)
         similarities.append((chunk, similarity))
     similarities.sort(key=lambda x: x[1], reverse=True)
-    return similarities[:top_n]
+    if similarities:
+        print(f"[{now_str()}] best rag score: {similarities[0][1]:.2f}")
+    return pick_chunks(similarities, top_n, RAG_THRESHOLD)
 
 def rebuild_vector_db(longterm_lines: list[str]):
     global VECTOR_DB
@@ -67,6 +73,7 @@ def rebuild_vector_db(longterm_lines: list[str]):
 
 
 async def ollamachat(websocket):
+    global rag_loaded
     # loop over every message on this connection instead of handling one and returning —
     # returning from this function is what makes the websockets lib close the connection,
     # which was breaking bot.py's persistent connection after every single reply
@@ -81,13 +88,17 @@ async def ollamachat(websocket):
             system_prompt = payload["system_prompt"]
             user_input    = payload["user_input"]
             shortterm     = payload.get("shortterm", [])   # list of strings
-            longterm      = payload.get("longterm", [])    # list of strings
+            longterm      = payload.get("longterm", [])    # list of strings, only there on a rebuild
             media         = payload.get("media", [])       # list of {"mime_type", "data" (base64)}
 
             # rebuild vector db if bot signals it (every 150 interactions bot sends full longterm),
-            # or if we come up empty but longterm isn't — e.g. we just restarted and lost VECTOR_DB
-            if payload.get("rebuild_vectors") or (not VECTOR_DB and longterm):
-                await asyncio.to_thread(rebuild_vector_db, longterm)
+            # if we never got longterm (just restarted, or first time as a fallback) ask for it and the bot resends
+            if payload.get("rebuild_vectors"):
+                await asyncio.to_thread(rebuild_vector_db, make_chunks(longterm))
+                rag_loaded = True
+            elif not rag_loaded:
+                await websocket.send(json.dumps({"need_longterm": True}))
+                continue
                 print(f"[{now_str()}] Vector DB reconciled, {len(VECTOR_DB)} chunks")
 
             user_query = nots(user_input)
@@ -96,7 +107,7 @@ async def ollamachat(websocket):
             # rag chunks go in system so they don't pollute turn history
             rag_block = ""
             if retrieved_knowledge:
-                rag_block = "\n\nOlder relevant messages for context:\n" + "\n".join(f"- {chunk}" for chunk, _ in retrieved_knowledge)
+                rag_block = "\n\nOlder relevant messages for context:\n" + "\n".join(f"- {chunk.replace(chr(10), ' | ')}" for chunk, _ in retrieved_knowledge)
 
             # lines to skip — not real dialogue
             skip_fragments = ("sent sticker:", "sent media:", "sent a random dm to ", "saved media:")
@@ -125,7 +136,8 @@ async def ollamachat(websocket):
                     text = rest.rsplit(' with reply: "', 1)[-1]
                     if text.endswith('"'):
                         text = text[:-1]
-                    return speaker.strip(), text.strip()
+                    target = rest[len("replied to "):].split(" saying: ", 1)[0]  # keep who they were talking to, otherwise replies lose it
+                    return speaker.strip(), f"(replying to {target}) {text.strip()}"
                 if rest.startswith("said: "):
                     return speaker.strip(), rest[len("said: "):].strip()
                 return None, None
@@ -161,7 +173,7 @@ async def ollamachat(websocket):
                 messages.append({"role": "user", "content": "\n".join(current_user_lines)})
 
             # final isolated turn — the specific message she needs to reply to
-            messages.append({"role": "user", "content": f"[responding to]: {nots(user_input)}"})
+            messages.append({"role": "user", "content": f"[responding to]: {nots(user_input)}\n(that is the newest message and the only one you answer, the lines above are just context. only the name that starts it is speaking to you, everyone else is someone else)"})
 
             # ollama only takes images (no audio/video/pdf), and only on the final turn so history stays text
             images = [base64.b64decode(m["data"]) for m in media if m["mime_type"].startswith("image/")]
@@ -205,8 +217,8 @@ async def ollamachat(websocket):
 
 
 async def main():
-    print(f"[{now_str()}] hosting on ws://0.0.0.0:8765")
-    async with serve(ollamachat, "0.0.0.0", 8765, ping_interval=None, ping_timeout=None, max_size=32 * 1024 * 1024):  # default 1mb would kill any media payload
+    print(f"[{now_str()}] hosting on ws://0.0.0.0:{PORT}")
+    async with serve(ollamachat, "0.0.0.0", PORT, ping_interval=None, ping_timeout=None, max_size=32 * 1024 * 1024):  # default 1mb would kill any media payload
         await asyncio.Future()
 
 

setup_wizard.py

@@ -63,11 +63,12 @@ def write_tokens_file(config_dir: Path):
     print(f"wrote {config_dir / 'tokens.txt'}")
 
 
-def write_connection_file(config_dir: Path, default_uri: str = "ws://127.0.0.1:8765"):
+def write_connection_file(config_dir: Path, default_uri: str = "ws://127.0.0.1:8765", ask_fallback: bool = False):
     config_dir.mkdir(parents=True, exist_ok=True)
     old = read_lines(config_dir / "connection.txt")
     old_uri = old[0] if len(old) > 0 else default_uri
     old_token = old[1] if len(old) > 1 else None
+    fallback = old[2] if len(old) > 2 else None
     print("\n-- connection to main.py --")
     uri = ask_keep("main.py address (ws:// or wss://)", old_uri)
     if old_token:
@@ -78,7 +79,14 @@ def write_connection_file(config_dir: Path, default_uri: str = "ws://127.0.0.1:8
         if not token:
             token = secrets.token_urlsafe(32)
             print(f"generated token: {token}")
-    (config_dir / "connection.txt").write_text(f"{uri}\n{token}\n", encoding="utf-8")
+    if ask_fallback:
+        print("\n-- fallback (optional) --")
+        if ask_yn("use a local ollama main.py as the fallback when gemini runs out of quota"):
+            fallback = ask_keep("fallback main.py address", fallback or "ws://127.0.0.1:8766")
+            print("start it with: PORT=8766 python3 main.py (it uses the same auth token, from config/auth_token.txt)")
+        else:
+            fallback = None
+    (config_dir / "connection.txt").write_text(f"{uri}\n{token}\n" + (f"{fallback}\n" if fallback else ""), encoding="utf-8")
     print(f"wrote {config_dir / 'connection.txt'}")
     return token
 
@@ -124,7 +132,7 @@ def main():
         main_config = SCRIPT_DIR / "config"
         backend = ask_backend()
         write_tokens_file(bot_config)
-        token = write_connection_file(bot_config, default_uri="ws://127.0.0.1:8765")
+        token = write_connection_file(bot_config, default_uri="ws://127.0.0.1:8765", ask_fallback=(backend == "gemini"))
         write_auth_token_file(main_config, token=token)
         if backend == "gemini":
             write_gemini_key_file(main_config)

utils.py

@@ -36,3 +36,44 @@ def tail_lines(path: Path, n: int) -> list[str]:
                 break
             block_size *= 2
     return [l.decode("utf-8", errors="ignore").strip() for l in lines[-n:] if l.strip()]
+
+
+CHUNK_LINES = 4  # consecutive messages per rag chunk
+MIN_CHUNK_CHARS = 40  # windows shorter than this are basically "lol" "yeah", not worth embedding
+SKIP_FRAGMENTS = ("sent sticker:", "sent media:", "sent a random dm to ", "saved media:")
+
+
+def make_chunks(lines: list[str]) -> list[str]:
+    # rag chunks are windows of consecutive messages instead of single lines, one-liners embed badly.
+    # the step is 1 so when the tail of memory.txt slides, the old windows keep the exact same text and don't get re-embedded
+    kept = []
+    for l in lines:
+        c = l.strip()
+        if any(f in c for f in SKIP_FRAGMENTS):
+            continue
+        if "said:" in c and c.split("said:", 1)[1].strip() == "":
+            continue
+        kept.append(c)
+    if len(kept) <= CHUNK_LINES:
+        windows = ["\n".join(kept)] if kept else []
+    else:
+        windows = ["\n".join(kept[i:i + CHUNK_LINES]) for i in range(len(kept) - CHUNK_LINES + 1)]
+    return [w for w in windows if len(w) >= MIN_CHUNK_CHARS]
+
+
+def pick_chunks(scored: list, top_n: int, threshold: float) -> list:
+    # scored is (chunk, similarity) sorted best first. drops anything under the threshold and skips windows
+    # that share a line with one already picked (they overlap, otherwise the top 5 would be the same messages)
+    picked = []
+    used = set()
+    for chunk, sim in scored:
+        if sim < threshold:
+            break
+        lines = set(chunk.split("\n"))
+        if lines & used:
+            continue
+        used |= lines
+        picked.append((chunk, sim))
+        if len(picked) >= top_n:
+            break
+    return picked