Reading the world graph

The graph inspector is both a cartographer's table and a continuity debugger. It shows how entities are actually connected, which can differ from what descriptions or memories claim. Open the deployed inspector, authenticate for admin access, and connect to the live server or load a snapshot file.

Read each top-level view

Use the view that matches the design question. A complete map does not prove a valid quest graph, and a clean social graph does not prove characters share a room.

Drill into containment

Enter a room to see occupants, items, containers, and other contents. Enter a character to see inventory and equipment. Open a container to follow nested contents.

Containment review catches common mistakes:

• an item with no physical parent;
• a character in the wrong room;
• one entity appearing under multiple parents;
• inventory accidentally marked as room content;
• a required item hidden inside an inaccessible or locked container;
• a region edge mistaken for physical reachability.

Turn on parent nodes when you need to understand how the selected entity fits into a larger chain.

Inspect components and outgoing edges

Selecting a node opens its ECS details. Compare structured state with the rendered story:

• Does the room description claim daylight while light state says otherwise?
• Is the “portable” wick actually carrying PortableComponent?
• Does a resolved quest still have an active goal on a participant?
• Does an overdue routine point to a place the character cannot know or reach?
• Does a controller badge match the intended human, LLM, scripted, behavioral, or suspended role?

Click relationship targets to follow causality. For a quest, inspect objective and reward entities. For an obligation, follow debtor and creditor. For a character, follow controller, home, social, ownership, and quest relationships.

Use the event feed as a timeline

When connected live, enable events. The feed shows typed domain events independently of the debounced graph refresh. Use it to correlate attempted play with authoritative change:

character action
→ command result event
→ changed components or edges
→ consequence event
→ visible projection update

If dialogue announces a repair but no repair event or state change follows, the repair did not happen. If an event fires repeatedly, inspect the driving component and whether completion or cleanup ever changes it.

The feed is recent operational evidence, not permanent lore. Important outcomes still need persistent state, records, history, or memory.

Audit one expansion seam

Before adding content, select the anchor entity and trace outward:

1. For a new room, inspect the door, containing room, intended direction, and nearby region.
2. For a new character, inspect the room, existing cast, available controller, and needed resources.
3. For a new item, inspect the parent container, current custody, matching actions, and any quest dependency.
4. For a new event, inspect active incidents, open obligations, schedules, and current room population.

This context is exactly what a good LLM patch prompt should summarize.

Look for graph smells

Some entities, such as controllers, history records, obligations, and world services, are intentionally not physically contained. “Uncontained” is only a smell when the entity's kind should have an owner or parent.

Inspector review

• Map, region, social, and quest views agree with the intended world.
• Containment paths explain where every physical entity is.
• Component state agrees with descriptions and claimed outcomes.
• Events correlate with real mutations and stop after completion.
• Controller badges and assignments match the intended cast.
• Expansion anchors have enough nearby context for a bounded addition.
• Unusual graph shapes have an explicit narrative or mechanical reason.

Next, use that context in LLM-assisted world patches.