i've been vibecoding this too much but it passes all my tests so it's ok

README.md

@@ -81,6 +81,28 @@ $ 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
 
+# 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
+
+you can let the local gemma-4 write the descriptions for you, it needs `pip install pillow` (and ffmpeg for videos), ollama running, and the model has to be able to see images
+```console
+$ python3 describe.py #only the empty ones, --all redoes everything, --model <name> to use another model, --gemini to use the gemini api instead
+```
+
 # 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