How it works

One clear path, from event to reply.

From a Discord event to the router, then the action gate or the model chain, then the output filter — on a single SQLite archive.

  1. The event arrives

    Discord events enter the bot, and an intent stage routes each one to the archive or to the model as the request requires.

  2. Context is read

    The router reads the persona and the memory the member pinned, then decides: a direct reply, the model chain when generated text is needed, or an action that needs the gate.

  3. The action gate

    The gate checks the actor’s immutable Discord id, the required permission, and a word that signals intent, before anything reaches the Discord API.

  4. The output filter

    It strips @everyone, token-like strings and leftover tags, and corrects a reply that claims an action still sitting in the queue was executed.

  5. The archive

    One SQLite file: server_messages for raw messages and an FTS5 index for search. No external queue, cache or worker.

Angel's message pipeline Five stages right to left: a Discord event, the language router, memory and archive, a proposed decision, then the confirmation gate before execution. An unconfirmed decision executes nothing. بوابة التأكيد بلا تأكيد: لا تنفيذ 1 حدث من Discord on_message · interaction 2 الموجّه اللغوي intent router 3 الذاكرة والأرشيف archive + index 4 قرار مقترح model stage 5 تنفيذ بعد التأكيد commands النموذج يقترح، والبوابة تُقرّر — ولا إجراء إشرافي يُنفَّذ بلا تأكيد.
مسار الرسالة: من الحدث إلى التنفيذ، وبوابة تأكيد في المنتصف.

Operations