Output Files

Every episode writes one run directory (--out).

File	What it holds
result.json	The verdict (environment_success), the arm, how the agent finished, token usage and LLM calls, and the settings of the run.
episode.mp4	The episode video (RPent robots; on RoboDojo the run directory keeps each step's images under steps/, and RoboDojo records the videos).
workspace/	Every cell or program the agent ran, and the files it wrote.
tool_turns.json	Each tool call: which tool, how long it took, whether it moved the robot, the simulator steps, and anything the sandbox flagged.
calls.jsonl	Every primitive call: arguments, outcome, wall time and simulator steps.
prompt.txt, card.json, card/*.png	The prompt and the task card the model received.
codex_events.jsonl / claude_events.jsonl / openai_events.jsonl	The runtime's own record of the conversation (with --runtime openai, also messages.json).
states.json, per-step images	The robot stack's own records (RPent robots).
policy.py, policy_output.txt	The program and what it printed (single shot).
host.log	The log of the robot host that ran the episode: where to look when play reports that it failed.

On RoboDojo, RoboDojo's evaluation client keeps its own records as well (RoboDojo).

Tables

pyrualean summarize runs/demo runs/other           # one row per episode
pyrualean summarize --json runs/demo runs/other    # the same rows, then the totals as JSON: success rate, statuses, mean calls and tokens

Give summarize run directories, not a folder that also holds other things.

A crash, a timeout or an episode that did not boot counts as a failed episode.