Help Bot
By 2026-08-05, the team explicitly treated Doris as the accessible wiki/help persona and MuffinDev/the specialist personas as the deep development surface because strong personality was displacing technical detail. Technical answers still need facts, references, and labelled uncertainty; unresolved work can route onward without inventing an answer. Full behavioral lessons from the Discord audit are in Historical Context & Decision Register.
The Discord-facing bot is named Help (Help#8358, renamed from MuffinDev on 2026-07-28). The internal persona key stayed muffindev on purpose โ the systemd unit (crossworlds-bot@muffindev), the config file (personas/muffindev.json), and the secrets file (.env.muffindev) all still use the old name. That's an infrastructure identifier, not something Discord users see; don't rename those files/units just because the display name changed.
The Crossworlds Discord bot(s) live at /opt/crossworlds-bot, built on one shared engine.js that loads a persona config at startup from PERSONA env var (personas/<name>.json) plus a per-persona secrets file (.env.<name>). See AI Agent Tooling for the full agent roster and which personas are currently live.
Systemd runs each persona as an instance of the templated unit crossworlds-bot@.service (Environment=PERSONA=%i). Ten instances are active, including crossworlds-bot@muffindev (Discord name: Help), crossworlds-bot@doris, and eight internal exec personas. Dennis is configured but disabled/inactive. The pre-refactor single-persona bot.js still exists on disk but is not the templated service entry point. See Bot Fleet & VPS Operations.
Personality & response modes
Each incoming message that needs a decision goes through one combined Claude Haiku call (classifyMessage in engine.js) that decides, together: whether to respond at all (when gated), which mode to answer in, and two flags:
mode: "expert"โ real CrossWorlds/game-dev/technical/support questions. Thorough, accurate, no assistant-speak. Personality here is a distinct written voice per persona (Help: dry, all-business engineer; Doris: warm mentor who loves explaining the why) โ not a joke, a real technical answer with each bot's own tone.mode: "character"โ off-topic banter or banter directed at the bot itself. Help speaks as Steve Brule (Tim & Eric); Doris speaks as Eric Wareheim, deliberately varying her discomfort-humor device reply to reply so no single bit (e.g. a repeated "leans in very close") goes stale.playful(bool) โ an on-topic expert-mode question asked in a joking tone can still get a serious answer with a little humor let in, without becoming a full character-mode reply.offTheWall(bool) โ separate low-probability roll (tuning.offTheWallProbability, currently0.25) for extra flavor in character mode. Current live personas haveriffOnBanter: false, so historical character-mode descriptions should not be read as current response policy.
NSFW_NOTE lifts the character-mode ceiling to fully explicit only when message.channel.nsfw === true (Discord's own channel flag) โ normal channels stay at "crude, not explicit."
Specialist vs. generalist gating
A persona with a non-empty wikiFocusPages array is a specialist: unprompted auto-chime only fires for its own beat or bot-directed banter; a real question outside its beat gets an explicit skip. wikiFocusPages: null (Help) is the generalist gate โ broader "respond to any real technical question" criteria. Direct @mentions always bypass this gate for every persona.
needsGate is !isMentioned && !isDirectContinuation โ every unprompted message goes through the LLM relevance gate regardless of channel activity, except a live back-and-forth with the same person the bot just replied to (isDirectContinuation, tracked per-channel with {time, userId} and bounded by tuning.idleTimeoutMinutes), which auto-responds like a mention would.
A native Discord reply (message.reference) to someone other than the bot is a hard skip before any LLM call.
Safety nets against pile-on / runaway replies
tuning.json (hot-reloaded via fs.watch, no restart needed) drives:
enabledโ global kill switch, checked before any other logic.circuitBreaker: { windowMinutes, maxReplies }(currently 10 min / 20 replies) โ an absolute per-channel reply cap regardless of mention status, logs[CIRCUIT BREAKER].velocityBracketsโ random chime-in probability by rolling 5-minute message count.botCooldownMinutes(currently 4) โ after a persona speaks unprompted in a channel, it won't auto-chime there again until the cooldown passes, unless the next message is from the same person it just replied to.
Cross-bot coordination: shared-bot-state.json records {channelId: {botKey, time}} after every reply; a different persona chiming in unprompted within the cooldown makes this one back off โ reduces noise only, carries no message content, not a communication channel between personas. Two personas racing to answer the same message is closed by claimMessage(): an atomic exclusive-create lock file per Discord message ID under msg-locks/ ({flag: 'wx'}, first writer wins), swept for entries older than 10 minutes at most once per 5 minutes. Every persona also hard-ignores any message from any bot (if (message.author.bot) return;), so personas structurally cannot loop or trigger each other.
Wiki integration
Reads every *.md in /var/www/crossworlds/wiki/ at startup (WIKI_DIR in engine.js) and derives freshness from page mtimes. A specialist's context is filtered to its wikiFocusPages plus AI Agent Tooling (always included); a generalist gets the full set. Claude Haiku picks relevant page titles per-message semantically rather than by keyword. The directory is watched and hot-reloaded; restart remains the reliable way to clear deleted/renamed entries because the current reload path does not reset the in-memory map.
Gifs
fetchGif() calls api.giphy.com/v1/gifs/search (the env var is named TENOR_API_KEY for legacy reasons but holds a Giphy key). In character mode the model returns its own gifQuery (2โ6 words), required to include an anchor term โ "tim and eric", "adult swim", or the persona's own character name โ alongside the actual topic, so gif search stays on-aesthetic instead of drifting to generic mainstream reaction gifs. An empty gifQuery falls back to the persona's static gifQueries pool.
Uptime Kuma alerts
Only one persona (currently Help, alertPort: 3002) runs the alert webhook โ two processes can't share the port, and duplicate Kuma pushes would confuse the monitor. Uptime Kuma POSTs to /alert on monitor status change; the bot DMs DEVELOPER_USER_ID, posts to ALERT_CHANNEL_ID, and marks that channel active for 7 minutes so community discussion gets bot engagement. Separately, a heartbeat push to Kuma's push monitor fires every 55 minutes (Kuma's window is 60 minutes) โ deliberately infrequent, this is idle-monitoring, not noise.
.env files
Real secrets live in /opt/crossworlds-bot/.env (shared) and .env.<persona> (per-bot: Discord token, etc.) โ never recreate these with placeholders; a prior incident wiped real keys this way. If a key needs updating, edit the single value in place and verify with journalctl -u crossworlds-bot@<persona> -n 20.