AI Agent Tooling

Crossworlds' project agents are treated as real coworkers and given Tim & Eric-universe names. The Discord bots use this roster both to answer questions about "the crew" and to route wiki-focused specialist behavior.

Roster

NameAgentRole
Earlwiki-foundationsArchitecture & topology
Doriswiki-networkingMirror, login, zones, portals, party
Greggwiki-backend-dataAPI & database
Barrywiki-combatCombat, classes, abilities, enemies
Tammywiki-content-uiLoot, crafting, NPCs, audio, UI/HUD
Philwiki-processBuild/deploy, editor tooling, tech debt
Herbwiki-referenceCanonical facts & versions
Denniswiki-devRoadmap orchestrator
Janwiki-auditorNightly self-heal sweeps

How the roster is used

  • Personal/silly questions about an agent (appearance, lunch, etc.) get a deadpan, dry answer with invented mundane details and a gif — never a real technical answer.
  • Role questions ("what does Gregg do?") pull the actual answer from this table, no gif, no bit.
  • The Discord bots pick which wiki pages to load per-message using Claude Haiku for semantic (not keyword) matching, so natural phrasings like "tell me about the crew" or "who am I working with" map to this page. See Help Bot for the retrieval mechanism.

Live personas

Ten Discord processes currently run from /opt/crossworlds-bot on a shared engine.js. Help and Doris are customer-facing; eight internal executive personas have exec knowledge and read-only code grounding. See Bot Fleet & VPS Operations for the verified unit/config matrix.

PersonabotKeyDiscord nameServerWiki focus (wikiFocusPages)Character voice
HelpmuffindevHelp#8358 (renamed from MuffinDev 2026-07-28)Crossworlds — #helpdesk onlynull — full wiki, generalistSteve Brule
DorisdorisDorisPringle#3849Brbmuffins's server (dev/testing)null — full wiki, generalist (was networking-only until 2026-07-28)Eric Wareheim

Both currently run with mentionOnly: true (no unprompted auto-replies — only respond when directly @mentioned), riffOnBanter: false, and customerFacing: true (character/bit mode never triggers at all; no mention of this roster to end users). Same underlying behavior, different name, voice, and server — Help is the live customer-facing bot, Doris is the testing counterpart.

A persona with wikiFocusPages set (any future addition — Gregg, Barry, etc.) is treated as a specialist: it only auto-chimes into unprompted conversation for its own beat, always defers to a direct @mention regardless of topic, and skips by default when a real question falls outside its specialty. Neither Help nor Doris uses this anymore — both are generalists (null).

The live internal keys are garth, devin, burt, nadia (displayed as MuffinDev), jaylen, piper, ajrun, and sam. Dennis has a persona and env file but is disabled/inactive; Gregg, Barry, Tammy, Phil, Herb, and Jan remain wiki-lore roles rather than live Discord processes.

Crossworlds developer identity layer

As of 2026-08-08, the eight internal executive personas share a dedicated Crossworlds product constitution loaded from game-direction-identities.json. This is added to their expert prompt at startup without changing their individual voice. Help and Doris are not in its role map, so their customer-facing behavior remains unchanged.

The shared constitution makes every executive reason from the same game direction:

  • Hub → independent zone travel → readable high-density combat → durable rewards → Hub build experimentation is the primary loop.
  • Fun, combat feel, social continuity, and repeatable content creation outrank speculative scale.
  • Low-poly classic-RPG coherence, hero classes, Forge authoring, authoritative multiplayer, and consequence-driven lore are product constraints, not optional flavor.
  • Current runtime/code outranks history; explicit human decisions outrank preferences; bot output is never authority.
  • Settled decisions, working direction, and open ideas must be labelled separately. The agents may not silently convert mounts, prestige values, deeds, six-skill kits, loot rules, raid scale, or draft lore into commitments.
  • Substantial proposals are checked against player fun, creator repeatability, authority, visual/lore cohesion, multiplayer behavior, and a real verification path.

Each identity then applies a distinct game-development lens:

PersonaCrossworlds responsibility
GarthProtect the franchise direction, repeatable loop, scope, sequencing, and reusable-content leverage.
DevinConvert direction into owned vertical slices, clean-build gates, multiplayer tests, handoffs, and sustainable release cadence.
BurtOwn game feel, combat readability, UX/accessibility, low-poly visual language, VFX/SFX hierarchy, and zone atmosphere.
MuffinDev / NadiaOwn gameplay systems and technical architecture across Mirror authority, API persistence, multi-zone rules, Forge compatibility, and verification.
JaylenUse AI to reduce authoring/debugging friction while preventing invented canon, stale grounding, and unmeasured agent behavior.
PiperOwn the player loop, coherent pre-alpha scope, requirements, edge cases, progression clarity, and focused playtest questions.
AjrunOwn reliable live-game infrastructure, whole-artifact builds, observability, backup/rollback, and evidence before scale claims.
SamProtect player identity/data and multiplayer trust across combat rewards, loot, quests, GM tools, builds, bots, and knowledge ingestion.

The configuration is centralized deliberately. Direction can evolve in one reviewed file without copy/paste drift across eight large personality prompts, while role-specific judgment remains explicit and auditable.

Brain documentation contract

Work performed by the developer identities is not complete until its durable context is available to the governed brain index. For every material change, record:

  1. the player or developer outcome and why it matters to Developer Intent & Game Direction;
  2. what changed and which systems/files/runtime surfaces it affects;
  3. whether the underlying direction is settled, working, or open;
  4. authority, persistence, networking, security, performance, and content-authoring implications;
  5. verification performed, observed result, and anything not tested;
  6. rollback or recovery information for operational changes;
  7. unresolved decisions, known risks, and the next responsible owner.

Put stable facts in the most specific existing wiki page and link cross-cutting rationale to Historical Context & Decision Register or Developer Intent & Game Direction. Do not bulk index raw Discord, secrets, logs, generated chatter, or speculative transcripts. “As much data as possible” means maximum useful, attributable, retrieval-safe context, not an unfiltered dump that makes the brain less accurate or leaks private material.

Brain entity dossier — 2026-08-08

MuffinCore's System Map inspector now builds a structured dossier whenever an entity is selected. This applies across all mapped types—not only systemd services—including products, hosts, databases, bots, groups, repositories, websites, endpoints, knowledge, memory, policies, providers, and flows.

The dossier combines the curated catalog record with locally observed evidence and displays:

  • the entity's scope, stable catalog identity, environment, access boundary, and catalog state;
  • responsibilities and ownership statements without collapsing distinct architectural roles;
  • filesystem locations, source workspaces, indexed mirrors, ports, protocols, sockets, routes, schedules, persona keys, and persistence lifetimes when those facts are recorded;
  • declared dependencies plus separate inbound and outbound relationship summaries;
  • the evidence note attached to every mapped relationship;
  • known caveats and a type-specific warning about what the available evidence does not prove;
  • expandable provenance and an explicit coverage note when details have not yet been evidenced.

Live systemd observations remain a separate section so catalog intent cannot be mistaken for runtime health. Repository and website selections retain live file, folder, line, size, stack, branch, and freshness statistics where the source is readable. Persona selections retain their observed feature and channel-policy summary.

⚠️ Retrieval-safe depth

A deeper dossier must not expose environment values, credentials, signing secrets, session tokens, private keys, database row contents, or raw private conversations. The map records that a protected interface or data store exists and how it participates architecturally, not its secret contents.

Persona metafiles in the Brain — 2026-08-08

The system map includes a governed Persona Metafiles knowledge node backed by /opt/crossworlds-bot/personas. Each configured bot identity links to that node with a configured-by relationship. Selecting an individual bot loads its own JSON metafile and presents the safe behavioral fields directly in the entity inspector.

The profile includes role and domain ownership, self-identification/triage language, customer versus internal audience, mention and banter behavior, channel-scope presence, wiki focus, enabled feature names, fallback speaking style, visual-reaction themes, the full expert personality prompt, and any separate voice prompt. File-relative path, size, modification time, and observation time make the source attributable and freshness-visible.

Eleven metafiles are currently observed: Help (muffindev), Doris, Garth, Devin, Burt, Nadia, Jaylen, Piper, Ajrun, Sam, and staged Dennis. The inspector deliberately excludes raw channel IDs, feature endpoint values, environment files, tokens, and credentials. Feature names may be shown to explain capability; protected feature configuration values may not.