Traces

Each episode directory holds manifest.json, events.jsonl, summary.json, kernel.log and, when the worker records video, videos/. A suite directory adds suite.json (experiment, agent, identity, agent version, deploy, tasks, seeds, layouts), suite.log, workers/, and, per task, unstable.json / displaced.json when evaluation layouts were replaced.

manifest.json

Key	What it holds
schema_version, kind	1, program
config	environment, agent_version, main (the main model role), runtime, verifier, system_prompt, task, seeds, max_model_calls, max_wall_seconds, task_description, max_cost_usd, agent_id (the identity), agent_lock, placement
config_sha256	the configuration’s hash
seed
embodiment_profile, embodiment_sha256	the worker’s profile after reset (without task-specific keys)
provenance	Python, git commit and dirty state, the package’s files and hash

events.jsonl

One JSON object per line with seq (from 0), time (epoch seconds), event and its fields.

Event	Fields
reset	observation: episode_id, step, instruction, state, images (data URLs)
library_loaded	version, status, error
model_request	call (from 0), reason, total_messages, sent_messages, new_messages, context (one row per message: role, text start, characters, images, omitted)
context_shortened	history_turns, sent_messages (after a model ran out of GPU memory)
model_reply	call (from 1), text, usage (input_tokens, output_tokens, cost, finish_reason), response_id, seconds, reasoning when the provider returns it
protocol_error	message, truncated
malformed_dropped	messages
verify	claim, answer
cell	code; action (direct mode); result: status, error, escalation, episode_ended, finish, stdout, seconds, events (one per execution: skill, event, ticks, handled), state, usage, report (the skills’ turn report: text, the cameras it outlined), strategy; attachments (count)
policy_to_end	status, error: the handed-over task run on to the end once the calls after the policy’s run were used up
finish	text; source: "program" when a program called finish(), "driver" when the driver ended the episode after policy_to_end
wrap_up	status, error, events
kernel_died	message
outcome	outcome: success, reward, process_score, terminated, truncated, steps
episode_video	files (camera → path)
interrupted, error, cleanup_error	message

summary.json

Key
run, task, seed
status	finished, episode_ended, budget_exhausted, wall_time_exhausted, cost_exhausted, interrupted, error
success, reward, process_score, steps	from the outcome
model_calls, verifier_calls, cells
input_tokens, output_tokens, cost_usd	model and verifier, plus the vision calls programs made
agent_seconds, kernel_seconds, wall_seconds
decisions	model calls by reason
events_handled, events_escalated
vision_calls, vision_cost_usd
strategy	the choice of start(strategy, kind): strategy (A, B, C), kind, policy, step; policy_ran once the handed-over task ran; after when B handed over later
error	when the status is error

Execution events

cell.result.events lists one entry per bounded execution:

{"skill": "move_eef", "event": {"kind": "COMPLETED", "source": "move_eef", "detail": {}, "tick": 22}, "ticks": 22}

Event kinds: COMPLETED, PREDICATE_SATISFIED (nominal); TIMEOUT, BLOCKED, GUARD_TRIGGERED (or a guard’s own event name), REGION_EXIT, SHIELD_REJECTED, REACTIVE_FAULT, SKILL_ERROR, EPISODE_ENDED. An escalation carries the whole trace: skill, parameters, start_tick, end_tick, event, samples (joints every 5 ticks).