Adventure

Adventure

"Every directory is a room. Every file is a clue. Navigation is investigation."

Turn exploration into a quest — or any simulation into a hybrid LLM/deterministic CLI.

Lineage: Colossal Cave (Crowther & Woods), Scott Adams Adventures, Zork (Infocom), MUD (Bartle), LambdaMOO (Curtis).

Inherits from: simulation/ — all simulation properties plus adventure-specific state.

[!TIP] This is a general pattern. Text adventure is the reference implementation, but the same architecture powers city sims, cloud management tools, board games — anything where deterministic transforms meet creative narration.

[!TIP] Perfect for codebase archaeology. "Find where the auth bug was introduced" — that's a quest!

The Premise

An adventure creates a player with state, places them in a room, and the LLM dungeon masters them around.

# player.yml
name: Alice
location: entrance-hall
inventory:
  refs:                                    # Lightweight pointers (weight: 0)
    - pub/bar/brass-lantern.yml
    - street/acme-catalog.yml#portable-hole
  objects:                                 # Deep copies (has weight)
    - { id: notebook, name: "Notebook", weight: 0.5 }
  fungibles:                               # Stacks
    - { proto: economy/gold.yml, count: 50 }
health: 100
notes: "Looking for the lost artifact"

Inventory Protocol: See skills/inventory/ for full TAKE/DROP/BOX/BEAM operations, pointer syntax, and structural editing.

The core loop:

User: "go north"
  → DM: Updates player.location, describes the new room
User: "look around"
  → DM: Reads room YAML, narrates contents atmospherically  
User: "take the rusty key"
  → DM: Moves key to player.inventory, narrates the action

The mapping:

• Directories = Rooms to enter
• Files = Clues, artifacts, characters
• player.yml = Your state (location, inventory, health)
• Chat = How you control your character
• LLM = Dungeon Master (narrates, adjudicates, surprises)

This is Memory Palace with narrative framing and a player character.

Multi-User, Multi-Agent (Engelbart NLS tradition)

Naturally supports multiple simultaneous participants:

# characters/
├── alice.yml        # Human player 1
├── bob.yml          # Human player 2  
├── merchant.yml     # NPC (DM-controlled)
├── guard-bot.yml    # Autonomous bot (action queue)
└── oracle.yml       # LLM agent with own goals

Character types:

All coexist in the same world:

# library/ROOM.yml
occupants:
  - alice          # Player exploring
  - bob            # Another player
  - librarian      # NPC who answers questions
  - dust-sprite    # Bot that cleans autonomously

Selection: Current Character or Swarm (Sims/Populous tradition)

Like The Sims and Populous, you have a selection — who you're controlling right now:

selection:
  mode: single          # or: group, swarm
  current: alice        # commands go to Alice
  
# Or control multiple at once:
selection:
  mode: group
  current: [alice, bob, charlie]  # "go north" moves all three
  
# Or a whole swarm (Populous/Dungeon Keeper style):
selection:
  mode: swarm
  filter: { type: imp, location: mines }

Selection commands:

Commands apply to selection:

> SELECT alice, bob, charlie
> go north
Alice goes north.
Bob goes north.  
Charlie goes north.

The coherence engine orchestrates all:

• Players get chat turns
• NPCs respond when spoken to
• Bots execute their action queues
• Agents pursue goals in background
• Selection determines who receives your commands

Quest Structure

graph TD
    START[🎯 Quest Objective] --> R1[Enter Room]
    R1 --> LOOK[👀 Look Around]
    LOOK --> EXAMINE[🔍 Examine Objects]
    EXAMINE --> COLLECT[📝 Collect Evidence]
    COLLECT --> DECIDE{What next?}
    DECIDE -->|New room| R1
    DECIDE -->|Solved| END[🏆 Quest Complete]

The Files

quest/
├── ADVENTURE.yml     # Quest state
├── LOG.md            # Narrative journal
├── EVIDENCE/         # Collected clues
└── MAP.yml           # Explored territory

ADVENTURE.yml

adventure:
  quest: "Find the authentication bug"
  status: in_progress
  
  current_room: "src/auth/"
  rooms_explored: 5
  clues_found: 3
  
  hypothesis: "Session cookie not being set"
  confidence: 0.7

LOG.md

# Adventure Log

## Day 1: Entering the Auth Dungeon

I stepped into `src/auth/` — a maze of middleware.

**Clues found:**
- `session.ts` — handles cookie creation
- `middleware.ts` — checks auth state

**Suspicion:** The cookie is created but never sent...

Commands

The rung selector

Rule: every describable thing answers at any rung, and the reader picks the rung — not the author.

Adventure games shipped this in 1977. VERBOSE, BRIEF, and SUPERBRIEF are viewspecs: per-reader, persistent across the session, set by the reader and applying to everything thereafter. That is the same control Engelbart had in NLS a decade earlier, and the same thing OPML later stored as expansionState. Three independent traditions concluded that a document needs a view knob, and only the web shipped without one.

MOOLLM generalizes it past rooms. Any object — room, character, item, idea, design doc, commit, scene — must be renderable at every rung, because the rung is a property of the reading, not of the thing:

GLYPH and INFODUMP are the new ends. The glyph rung is the one that can be apprehended in parallel — a map or contact sheet of glyphs is scanned pre-attentively, where a page of descriptions is read serially. INFODUMP is what a curious reader or an agent wants: not a better summary, all of it.

Two details from 1977 worth honoring

BRIEF is adaptive, and that is the sophisticated part. It is not a fixed detail level — it is full once, then name only, which makes detail a function of what this reader has already seen. A rung with memory. Nothing in the modern summary-ladder discussion has this, and it is strictly better than a static setting: it spends words exactly where they are new. Implement it by reading visit history, so BRIEF on a revisit is not a shorter description but a different one.

LOOK is the one-shot override. In the classic games LOOK redisplays the room in full regardless of the current mode, without changing the setting. That is a peek: a temporary excursion to a deeper rung that does not disturb your global preference — the same contract as a link popup, and the same as reselection, where you browse a consequence before committing to it. Keep LOOK doing exactly this, and let EXAMINE be the object-scoped version.

Consequences for authoring

• A room, character, or object whose YAML supplies only a full description is incomplete — it cannot answer at SUPERBRIEF or GLYPH. GLANCE.yml is the glance rung; ROOM.yml carries the name and the sign; README.md is the body. The three-file contract is the rung contract.
• Missing rungs may be generated rather than blocking, which is what makes the requirement cheap — see designs/TAGSONOMY-COMPILER.md for the build-time version, and ADVENTURE-COMPILER.md for compiling a world.
• The reader's rung choice is state worth saving, which makes it part of a view record rather than a hidden preference: VIEW-STATE-ANCESTORS.md.

Debug Mode

Toggle technical output with DEBUG-ON and DEBUG-OFF.

When debug is ON, logs include collapsible sections showing:

• File operations (creates, edits, deletes, moves)
• State changes with before/after values
• YAML data islands with abbreviated data
• Markdown links to all referenced files
• Technical narrative explaining HOW and WHY

Example debug output:

<details open>
<summary>📂 <strong>Editing CHARACTER.yml to update player location from start/ to coatroom/</strong></summary>

```yaml
# State change (CHARACTER.yml is canonical)
player:
  location: start/  →  coatroom/  # Character owns their location

The character file owns location state. ADVENTURE.yml mirrors it for convenience.

Files affected:

• CHARACTER.yml — canonical location updated
• ADVENTURE.yml — mirror updated

When debug is OFF, output is clean narrative without technical sections.

Customize with natural language:

> DEBUG-FORMAT Show only file operations, skip YAML, use 🔧 emoji

The format field in ADVENTURE.yml accepts natural language instructions for how to format debug output.

Integration with Cards

Trading cards can be your adventure companions:

cards_in_play:
  - card: "Index Owl 🦉"
    goal: "Search for cookie-related code"
  - card: "Git Goblin 🧌"
    goal: "Find when session handling changed"

Sister Script Integration

Vision: Python CLI handles deterministic operations; LLM focuses on narrative. See README.md for full CLI vision and development plan.

Evidence Types

Room Protocol

When entering any directory:

1. DESCRIBE — List contents, note what's here
2. EXAMINE — Read interesting files
3. COLLECT — Note evidence in adventure log
4. EXITS — Note paths to other rooms
5. DECIDE — Choose next direction

Codebase Archaeology

Adventures work for code exploration:

Live Examples

Best example: examples/adventure-4/ — The gold standard.

The Pub (Crown Jewel)

examples/adventure-4/pub/ — A complete social space:

pub/
├── ROOM.yml              # Themeable tavern (6 themes!)
├── bartender.yml         # NPC with 6 identity variants
├── pie-table.yml         # Octagonal debate table
├── gong.yml              # Gong of Gezelligheid
├── bar/
│   ├── bartender.yml     # The omniscient bartender
│   ├── budtender-marieke.yml
│   └── cat-cave/         # TARDIS-like cat sanctuary
│       ├── ROOM.yml
│       └── 10 cats (Terpie, Stroopwafel, kittens...)
├── arcade/               # Pacman, Pong, Pinball, Fruit Machine
├── games/                # Chess, Darts, Cards
├── stage/
│   └── palm-nook/        # Multi-room character space
│       ├── study/        # Infinite typewriters, infinity desk
│       ├── gym/          # Infinite climb
│       ├── play/
│       └── rest/         # Hammock, silence cushion
└── menus/                # Drinks, snacks, buds, games

Key Patterns from adventure-4

Themeable NPCs (bartender.yml):

identity:
  classic_adventure:
    name: Grim
    appearance: "Weathered human, salt-and-pepper beard..."
  space_cantina:
    name: Z-4RT
    appearance: "Multi-armed service droid..."
  cyberpunk_bar:
    name: Nyx
    appearance: "Chrome-implanted bartender..."

Themeable Rooms (pub/ROOM.yml):

theme:
  current: classic_adventure
  themes:
    classic_adventure:
      name: "The Gezelligheid Grotto"
      bartender: "Grim, a weathered human"
      menu: ["Ale (1 gold)", "Mystery meat pie (3 gold)"]
    space_cantina:
      name: "The Rusty Hyperdrive"
      bartender: "Z-4RT, a droid with too many arms"
      menu: ["Blue milk (1 credit)", "Bantha burger"]

Rich Activities:

activities:
  PERFORM: { venue: stage, effects: [tips, drinks_thrown] }
  DEBATE: { venue: pie_table, rules: roberts_rules }
  RING-GONG: { protocols: [once: attention, twice: emergency, thrice: mercy] }
  CELEBRATE: { effects: [free_round, +morale, everyone_toasts] }

Framing Protocol (for tribute performances):

framing:
  mode: [performance, celebration, tribute]
  tribute_protocol:
    invocation: "Before they arrive, acknowledge we're summoning them"
    performance: "Depicting them as we imagine their best selves"
    acknowledgment: "After they depart, note this was a tribute"

Other Examples

• examples/adventure-4/ — Current walkable instance
• examples/old/adventure-3/ — Earlier version
• examples/old/adventure-2/ — Extended exploration
• examples/old/adventure-1/ — Minimal starting point

The Intertwingularity

graph LR
    AP[⚔️ adventure] -->|IS-A| R[🚪 room]
    AP -->|companions| TC[🎴 card]
    AP -->|logs to| SL[📜 session-log]
    AP -->|similar to| DB[🔧 debugging]
    MP[🏛️ memory-palace] -->|sibling of| AP

Temporal abstraction (Sutton one-step trap)

Adventure commands are options in Sutton's sense — semi-MDP macros, not iterated micro-transitions.

• GO north jumps to an abstract room; the DM does not simulate every footstep.
• LOOK / EXAMINE use pickDescription LOD — detail on demand.
• exit edges connect graph nodes; guards are option preconditions.
• time TICK fires on meaningful moves, not every chat line.

Anti-pattern: rolling out N implicit one-step predictions to answer "what happens if I keep going?"

Field case: Urban eBike Safari — voice options ("next Invader") then SET DESTINATION; Bosch turn-by-turn is micro layer only.

Full treatment: simulation/ONE-STEP-TRAP.md · examples/one-step-trap.yml

Field ride game: continuous AI guess/suggest; pie-menu network by biking; VoyStick + steering law — urban-safari-ride-game.yml · WWSFF room

Future Vision

CLI Uplift Plan, Browser Compilation, Scott Adams History, Owl Simulation See README.md for complete development roadmap and inspiration.

Dovetails With

Sister Skills

• simulation/ — Base class (adventure inherits this)
• simulation/ONE-STEP-TRAP.md — Options vs one-step rollout
• exit — Graph edges between rooms
• time — Turns vs LLM iterations
• memory-palace/ — Method of loci (iLoci / spatial safari)
• speech/ — Voice command parser; field STT
• simulator-effect/ — Implication beats simulation

Kernel

• kernel/context-assembly-protocol.md — Working set loading