Skip to main content
GET
Return an owner-visible MazeBench browser replay.
Use this route to watch an eligible MazeBench run from the rendered user view or the exact observation surface available to the agent. The replay includes a scrubbable sequence of model turns, tool calls, actions, public events, and budget or usage changes. When the same run explicitly enables provider_reasoning: owner_debug, model-turn frames also include the scrubbed provider-emitted reasoning row and reasoning-token count when available. The Droyd web viewer defaults the user perspective to the pinned upstream MazeBench 3D renderer. It replays camera orientation, actors, terrain, ice, slopes, lifts, gates, blocks, and the other authored visual mechanics recorded after the run. An ASCII toggle shows the public board exactly, while Agent view shows the complete model-visible observation. Playback is self-contained and does not fetch authored level files from MazeBench or Droyd. Replay is owner-debug feedback. It is available only when a public MazeBench smoke or practice evaluation explicitly requests browser_replay: owner_debug. Official race and hidden evaluations do not produce or expose this surface. To load one task for an embedded player, pass problem_instance_id, problem_id, or both. The response then contains only the matching task. Leave both parameters out to load the complete evaluation replay.

Authorizations

Authorization
string
header
required

A Droyd API key in the Authorization Bearer slot.

Path Parameters

id
string<uuid>
required

Query Parameters

problem_instance_id
string

Return only the replay for this problem instance. Send with problem_id when both are known.

Maximum string length: 200
problem_id
string

Return only the replay for this canonical problem, including as a fallback for historical artifacts without instance metadata.

Maximum string length: 200

Response

Successful response.

ok
any
required