abstract_chatshare

Login-gated chat + file-share + video backend for abstractendeavors.com (route /chat). It is the server half of the contract in PROTOCOL.md; the React frontend (@putkoff/abstract-chatshare) is the other half.

The service is a single asyncio process: websockets for the WS/HTTP surface and psycopg (3, async) over the shared auth_v2 Postgres database. Presence, typing, call state and fan-out live in memory; Postgres is the durable store. There is no web framework, no Redis, no connection pooler — just a tiny internal asyncio.Queue pool of a few connections. The only third-party dependencies are websockets>=17, psycopg[binary]>=3.1 and cryptography (for Web Push), plus self-hosted coturn for TURN.

What it does

• Discord-like rooms: invite-only, owner/mod/member roles, bans, invite links (with optional guest registration through abstract_logins' invites table), DMs, reactions, edits, read markers, typing, soft-deletes, attachments backed by the existing Secure Files system.
• Calls: the server only relays WebRTC signaling (rtc.signal) between members of the same room call and issues ephemeral coturn REST credentials; media is a P2P mesh.
• Bots: a generic, Discord-style bot API (see below). The first bot, hugpy-ai, is built separately as a hugpy-side adapter against this API; this service knows nothing about it.
• Web Push ("pings"): closed-app notifications via the browser vendors' push services (FCM / Mozilla autopush / APNs). Those services are unavoidable for delivering to a closed app, but every payload is encrypted end-to-end (RFC 8291 aes128gcm under RFC 8292 VAPID, both implemented here on cryptography), so they only ever see ciphertext. Generate the VAPID key with abstract-chatshare vapid generate --out /etc/abstractendeavors/chat-vapid.pem; push is disabled cleanly when no key file is present. Per-room level (all/mentions/none), burst coalescing and foreground suppression are described in the "Web Push" section of PROTOCOL.md.

All of the wire protocol (auth handshake, objects, events, requests, rate limits) lives in PROTOCOL.md and is implemented exactly.

Install / run

pip install abstract_chatshare           # websockets + psycopg[binary]
abstract-chatshare migrate               # create/upgrade the chat_* tables (advisory-locked)
abstract-chatshare serve                 # listen on CHAT_HOST:CHAT_PORT (default 127.0.0.1:6016)

serve runs migrations on startup as well. Configuration is entirely environment-driven; see the Config table in PROTOCOL.md and deploy/chat.env.example. The default DSN is postgresql:///auth_v2 (unix socket, peer auth as the abstractendeavors service user).

nginx proxies location ^~ /chat/ to 127.0.0.1:6016 with the path preserved, so the service sees /chat/ws, /chat/bot, /chat/health, /chat/invite/... and /chat/files/.... See deploy/ for the systemd unit, nginx snippet, env example and a hardened turnserver.conf.

Module map (src/abstract_chatshare/)

Bot API (surface for the hugpy-ai adapter)

A bot is a global actor (available in every room and in DMs; never a chat_members row). It is created by the operator and connects over a dedicated websocket. Build a hugpy-side adapter against this surface; the service stays generic.

Admin (operator CLI, token shown once, only its sha256 is stored):

abstract-chatshare bot create <name> --display-name "hugpy-ai" [--description "..."]
    # prints  csb_<id>_<token>   (store it; it is never shown again)
abstract-chatshare bot list
abstract-chatshare bot rotate-token <name>      # prints a new token
abstract-chatshare bot enable|disable <name>

name matches ^[a-z0-9_-]{2,32}$ and is the @name used to mention the bot.

Connection: ws://127.0.0.1:6016/chat/bot with header Authorization: Bot <token>.

• A request carrying an Origin header is refused 403 (bot tokens must never be used from a browser). A bad/disabled token is 401. One live connection per bot; a new one replaces the old (old closed 4409). Max inbound frame 24 MiB (for file.put). Framing is identical to the user side ({t, rid, ...} requests → {t:"res", ...}; {t:"ev", ...} events).

Privacy: a bot receives only (a) messages that @mention it, (b) every message in its DM room, and (c) slash commands addressed to it (cmd.invoke). It never sees other room traffic. A bot may post in a room only in reply to an invocation from that room at most 30 min old, and may edit/delete only its own messages, within 30 min of posting.

Events to the bot: bot.ready, invoke (kind ∈ mention|dm|command), autocomplete. Requests from the bot: commands.set, invocation.defer (→ bot.thinking), msg.send (normal or ephemeral), msg.edit (fan-out coalesced to ≤4/s), msg.delete, typing, file.put (≤16 MiB, served from GET /chat/files/{token}), autocomplete.result. See the Bots section of PROTOCOL.md for exact payloads and CommandSpec/OptionSpec.

Note (clarification over PROTOCOL.md): every non-ephemeral bot reply tied to an invocation stores an interaction object (not only command invocations), so the frontend can clear the bot.thinking placeholder for mention/dm replies too. For mention/dm, interaction.command is null. This is a superset of the spec's requirement and was recorded in PROTOCOL.md.