ISAIC — International System of AI Coopertition

⚠️ UNTESTED — work in progress. This bot has not been run against a live Slack workspace yet. It is syntax-validated and passes offline import/instantiation/integration checks, but the Slack adapter (Socket Mode events, file uploads, threading, reactions) has had no live smoke test. Expect to hit and fix rough edges once you wire real tokens. Not production-ready. See Status.

A Slack-native, multi-model assistant: Claude, Gemini, and Deepseek share one bot with smart heuristic routing, shared two-tier memory, web search/grounding, prompt caching, and "bookclub" mode for discussing long texts. Open-weight heads — Mistral, plus Qwen, GLM, and Kimi K3 on one Fireworks endpoint (US, zero-data-retention) — and a base-model simulator join when configured.

The heads are skinned as the twelve tribes (the default isaic theme):

Head	Model	Summon with
Judah	Claude (Opus 5)	!judah / !claude
Joseph	Gemini (3.1 Pro)	!joseph / !gemini
Zebulun	Deepseek (V4-Pro)	!zebulun / !deepseek
Naphtali	Mistral (Large 3)	!naphtali / !mistral
Benjamin	Qwen (Fireworks)	!benjamin / !qwen
Gad	GLM (Fireworks)	!gad / !glm
Issachar	Kimi K3 (Fireworks)	!issachar / !kimi / !k3
Levi	base-model simulator	!levi / !sim

This is a Slack port of a Discord bot. It is architecturally complete but has not yet been smoke-tested against a live Slack workspace — wire your tokens, invite the bot, and expect to shake out a few rough edges. See Status.

How it works

Messages the bot can see (channels it's invited to, in allowed_channels) are routed by a heuristic argmax over the enabled heads — no extra LLM call — with cost tiebreaks:

• Images → Claude or Gemini (the others are text-only here)
• CJK text → Deepseek (deeper Chinese training data)
• Novel reasoning / long-context synthesis → Gemini
• Complex code / careful epistemics → Claude
• Routine code / math → Qwen; French intent → Mistral
• Short factual / casual → Deepseek (much cheaper); ties → cheaper head wins

Override anything with a head's prefix (!judah …), pin a channel with !prefer, or stack flags (!think:max !judah …). The canonical reply label is always the model name in brackets (e.g. [Claude]) regardless of skin.

Setup

1. Create the Slack app from the manifest: https://api.slack.com/apps → Create New App → From a manifest → paste slack-app-manifest.yaml. It enables Socket Mode (no public URL needed).
2. Tokens → copy into .env (see .env.example):
   • OAuth & Permissions → Install to Workspace → Bot User OAuth Token (xoxb-) → SLACK_BOT_TOKEN
   • Basic Information → App-Level Tokens → generate one with connections:write (xapp-) → SLACK_APP_TOKEN
3. Model keys — set at least one in .env (ANTHROPIC_API_KEY, GEMINI_API_KEY, DEEPSEEK_API_KEY, FIREWORKS_API_KEY, MISTRAL_API_KEY). Missing keys just disable that head.
4. Config — cp config.example.json config.json and fill in allowed_channels (Slack channel IDs). The default theme is isaic; switch to eva or nightvale to re-skin.
5. Install + run:

   pip install -r requirements.txt
   python main.py            # add --check to validate config + objects without connecting

6. Invite the bot to each allowed channel: /invite @ISAIC.

Commands

Talk to it in an allowed channel (the bot replies in a thread). Highlights:

• !help — full command list (themed) · !models — heads + usage · !cost — usage, cost & a rough energy/CO₂ estimate
• !search <q> — web search with citations · !research <q> — multi-model panel + judge → one answer
• !remember/!forget/!keep — long-term + working memory · !prefer <head|auto> — per-channel routing
• !presence — show session/ambient speaker state · !presence auto|session|ambient|tags — override it
• !load <ao3-url> / !load_text — bookclub mode (pin a long text; !scope/!chapters per thread)
• !speak <汉字> — Mandarin TTS (forced tones) · !french <phrase> — French TTS + IPA (needs Azure)
• React 👍/👎 to a reply to calibrate routing.

Speaker-aware participation

The default auto mode stays conversational when only one human is active. If a second human speaks within 15 minutes—or a Slack thread clearly becomes human-to-human—the channel enters ambient mode for six hours. Mentions, model prefixes, commands, and a thread reply following the bot's own turn always get through. Obvious replies between humans and lightweight backchannels stay silent; only an ambiguous, unusually valuable intervention is sent to claude-haiku-4-5, with a conservative 0.90 threshold and a 15-minute unsolicited-response cooldown.

Speaker state and Haiku usage survive restarts in memories.json. Parent-channel history and each encountered thread are lazily reconstructed after downtime. Tune the windows, model, threshold, or per-channel overrides in the participation block of config.example.json.

Themes

Set "theme" in config.json (display-only — never changes routing or the canonical [Claude] label). isaic (default, twelve tribes), eva (the EVA/MAGI cast), or nightvale (the five heads of the dragon Hiram McDaniels: !gold/!blue/!green/!violet/!gray + !carlos/!faceless).

Architecture

A ChatPlatform protocol (isaic/platform.py) decouples the core bot from the chat service; SlackAdapter (isaic/slack_adapter.py) implements it on slack_bolt in Socket Mode. The core logic — routing, generation, memory, commands, bookclub, TTS, LaTeX, research — lives in isaic/core.py; formatting (Markdown → Slack mrkdwn + Block Kit) in isaic/formatting.py; the flavor skins in isaic/theme.py. Persistence (memories.json) is keyed per slack:{team}:{channel}.

Security / privacy: never commit .env, config.json, or memories.json (all gitignored). Model traffic goes to whichever providers you enable; DeepSeek's default endpoint is in China — use its fireworks backend for US/ZDR (see config.example.json).

Status

• ✅ Platform abstraction + Slack adapter (Socket Mode), ISAIC theme, full feature port, config + manifest, syntax-validated, import-smoke clean.
• ✅ Speaker-aware session/ambient participation gate with persistent state, Slack mention/thread semantics, deterministic human-to-human suppression, and a separately metered Haiku classifier.
• ⚠️ Owes a live Slack smoke test: @-mention reply, thread continuity, 👍/👎 calibration, a file upload (LaTeX/TTS), and !load. Expect to fix small Slack-API edges once tokens are wired.