A record of
the run.

A game leaves a trail.

A card is played. A room changes. A deck gains or loses a card. The logger catches these events and records the surrounding state, including player resources, cards, and enemy intentions.

Selected event fields from the repository's example Defect run. This is a recorded excerpt, not a live game feed.

[
  {
    "event_index": 1,
    "type": "run_started",
    "summary": "run started"
  },
  {
    "event_index": 2,
    "type": "room_entered",
    "summary": "room entered"
  },
  {
    "event_index": 3,
    "type": "deck_card_removed",
    "summary": "deck card removed: CARD.STRIKE_DEFECT"
  },
  {
    "event_index": 4,
    "type": "combat_setup",
    "summary": "combat setup: ENCOUNTER.CORPSE_SLUGS_WEAK"
  },
  {
    "event_index": 5,
    "type": "room_entered",
    "summary": "room entered"
  },
  {
    "event_index": 6,
    "type": "turn_started",
    "summary": "turn started: round 1 Player"
  }
]

From action to record

One event at a time.

1. 01

   Listen

   Patches and runtime hooks capture actions and changes in the game.

2. 02

   Snapshot

   Record the relevant player, card, room, and combat state.

3. 03

   Write

   Append structured JSONL events with identifiers and an event index.

Each line is a standalone JSON record. Session and run identifiers keep the records tied to their context, while the event index makes their order explicit.

A resumed run is not a new run.

Continuing a save introduces a less obvious problem: the game can revisit state the logger has already seen. EventLogger includes handling for continuing existing run logs and detecting state rewinds. It also tracks deck mutations to avoid recording the same change more than once.

These checks matter before any analysis or training. A repeated event or an unexplained jump backward can change what a sequence of actions appears to mean.

Check the record before using it.

A validation script checks the log structure, event ordering, timestamp progression, and the pairing of card-play start and end events. The example log gives those checks something concrete to run against.

• Event indices make gaps or ordering problems visible.
• Paired action events help identify incomplete sequences.
• State snapshots preserve the context needed to interpret an action.

The experiment so far.

The project collects structured game data for later analysis and possible AI training. The work shown here is the logging mod and its validation tools. It does not include a trained game-playing agent.