Replay · WitnessBench

0 / 0 speed

Levels: filter action history:

Game CRT

↑ ← ↓ →

⟲ ✓

faded path = the path just submitted (display overlay; the game resets the path to the start on a failed submit)

red rings (if any) = cells the game flagged as violations

Now

Action history ◆ LLM = LLM call ; dim = plan exec

Plan

Memory

Active rules

running state at this step

● verified ● proposed ● transferred

Compacted (last LLM-fed snapshot)

Last change

Memory growth

Reasoning trace

What the LLM saw (semantic ASCII board)

Rules fed to the reflection (## Current Knowledge)

Observations (recent action effects)

LLM response (meta = thinking, add/del = rule churn, plan = next actions)

What is the agent doing right now?

The agent alternates between LLM reflection calls and plan execution. The current stage shows in the Now panel and is color-banded in the timeline above:

reconnaissance Short probing of the available actions before planning (no LLM calls).
planning The bulk of the run. At each reflection the model reads the current board, the recent action effects and its rule set, then updates its rules and proposes a short action plan (which may include 0 = RESET). The ◆ LLM rows in the action history are the reflection calls; the plain rows between them are that plan being carried out.
reset RESET is an action the agent may choose inside a plan (⟲ in the plan pane / D-pad): the board returns to the level start; the level, the timeline segment and the action count all continue.