Extract actual executed actions from result JSON. Returns keys: - actual_action_count: int | None - actual_action_text: str | None

Extract actual executed actions from result JSON. Returns keys: - actual_action_count: int | None - actual_action_text: str | None

Extract token stats from result JSON metadata token_usage.

Extract token stats from result JSON metadata token_usage.

Extract and sum prompt/completion/total tokens from stdout/stderr text.

Extract and sum prompt/completion/total tokens from stdout/stderr text.

Sum prompt/completion/total tokens printed to stdout/stderr.

JSON

Check if currently recording

A task is complete when it has written an evaluation summary JSON.

Merge token stats, preferring non-empty primary values.

Merge token stats, preferring non-empty primary values.

ID （ ， ） ： - ai2thor_04000 -> ai2thor04000 - ai2thor04000 -> ai2thor04000

action_sequence

Read structured result from result json. Returns keys: - has_result_json: bool - task_result: 'success' | 'failure' | None - fa

Read structured result from result json. Returns keys: - has_result_json: bool - task_result: 'success' | 'failure' | None - fa

Read task IDs from CSV; when only_null, skip rows whose Completed is true/false.

， instruction golden_actions

CSV ：Completed +

CSV Completed Args: csv_path: CSV task_id: ID status: （'true' 'false'） lock:

Write a clear diagnostic file when task folder exists but no result JSON was produced.

Write a clear diagnostic file when task folder exists but no result JSON was produced.

[ ]

Build a predicate for a single condition dict.

； SPATIAL_TASKS_ROOT （ taskabs/）

Remove room/character/structural ids from a visibility result.

bounds: (min_x, max_x, min_y, max_y) （ ）

Generate text state description

Return the currently selected first-person camera id.

Return (position_Vector3r, orientation_Quaternionr) of the drone.

Return (position_Vector3r, orientation_Quaternionr) of the drone.

Return IDs visible in the current first-person camera.

Convert runner state fragments to JSON-safe values.

Check whether a graph node id matches the requested object type exactly.

（ / ） Args: point_x: X point_z: Z polygon: ， {"x": float, "y": int, "z": float}

Convert Unity quaternion [x, y, z, w] to yaw radians.

Rollback a compound WalkForward action to its starting pose.

Sync client yaw cache from live graph, fallback to provided yaw if needed.

prefix: 'benchmark' | 'benchmark_sequential' save_name ：{prefix}_{save_name}_{ }（ ） save_name ：{prefix}_{ }

Returns: Optional[FrameData]: ， None

Classify a failed run into model/external bucket for benchmark stats.

Coerce init rotation payload to a discrete quaternion. Priority: 1) `char_yaw_degrees` when provided 2) `char_rotation` quaternion list [

Per-agent step budget for multi-agent tasks: max_steps = 10 + n.

，

； ，

Create a VLM instance from a config dict.

Read n from ``golden_actions.steps`` for multi-agent tasks.

：JSON.failure_type > fail_reason > >

( 0， ) Args: px, py: polygon: ， [x, y] [x, y, z]

OpenAI ， API Args: response: OpenAI API api_type: API ， None Returns:

Find dual-agent result json in a task output dir.

Find dual-agent result json in a task output dir.