Artifacts and storage

GTA writes image, plan, and script files into the active workspace. The normal dashboard uses the repository workspace; library working copies have isolated workspace directories. Switching a session does not change the web process’s working directory.

Artifact	Meaning
environment.jpg	Captured/working RGB scene image
resized_img.png	Padded/resized image used for model coordinates
annotated_keypoints.png	Display annotation image
skills_output.json	Parameterized controller recipes
keypoint_locations.json	Named pixel coordinates and descriptions
robot_script.py	Generated Python robot script
generated_plans/	Save Workflow exports
library_checkpoints/<id>/	Versioned saved checkpoint, copied artifacts, agent/workflow state
library_sessions/<id>/	Independent mutable working copy
execution_logs/<run-id>/	Frozen script, structured events, server/script or helper logs

Library execution logs are nested under the session’s workspace. Execution records remain on disk until explicitly removed; there is no automatic retention policy. An application restart returns to checkpoint selection. Browser refresh reconnects to the running application’s current session and status.

Checkpoint versus working copy

A checkpoint preserves the approved task, subtasks, controller recipes, parameters, keypoint definitions and grounding, images, available script, messages/events, and agent context. It also stores each agent’s provider, model, system prompt, and effective generation settings.

Saving requires a valid parameterized plan and no active background handlers. Required state and artifacts are gathered together; the checkpoint becomes visible only after a complete save. A provider-context export failure aborts the save.

Loading copies the snapshot without camera capture, model requests, or pipeline-event replay. OpenAI context is recreated in a new conversation on the next model operation; Gemini chat content is reconstructed independently. Feedback does not append to the preparation conversation. Loading does not make saved coordinates valid for a new physical scene.

Current-machine credentials, camera/calibration, ROS environment, robot address, and device settings apply to live operations. Saved model/prompt settings apply to continuation. Requery and regenerate the script for the current scene and tool configuration.

Workspace revisions and readiness

Image, grounding, and script readiness are tied to workspace revisions. Requery/refinement invalidates affected readiness before work begins. Readiness is restored when the new operation completes successfully; previous artifact files remain on disk.

Reset to Checkpoint makes a new working copy, retaining previous session artifacts and execution records on disk. Deleting a session does not delete its source checkpoint, but it does remove any unsaved feedback and run records in that session. Stop the application/operations before managing live workspaces.

Sensitive data

Checkpoints, exports, and logs can contain camera images, task text, model responses, feedback, conversation content, and local artifact paths. OpenAI snapshots may include available encrypted reasoning content. Configuration serialization excludes API keys and .env. User messages and logs are retained as supplied.

Keep credentials out of feedback, screenshots, and shared reports. Review images, messages, and local paths before sharing checkpoints, exports, or logs. Exclude .env and private runtime artifacts from published source.