Sonar scene authoring from natural language

A full simulation takes minutes of GPU time, and a heightmap preview takes tens of seconds. Studio renders the terrain first, marks the band that the sonar will image, and reports when a requested feature falls outside that band. The scene is corrected in conversation before the simulation runs.

The agent reads the same scene definition as the simulator and runs the same validator that blocks an invalid run in the desktop app. A validation failure is reported as a failure; the agent does not substitute a different scene.

Once the scene is accepted, the job runs on the GPU and returns the full product set: the dynamic-range-compressed image for viewing, the complex SLC for processing, and the raw HDF5. The pipeline, formats, and imagery are those of a local run of the same scene file.

An agent that can run a simulator runs code on a workstation, so Studio restricts what the agent can do. It has no shell and no ability to read, write, or edit files. It talks to exactly one in-process tool server, and it loads none of the operator's own settings, hooks, or extensions.

The documentation tool shows how the restriction works. Studio answers questions about sonar theory and the scene schema by reading the project's reference documentation and nothing else. The lookup resolves the path first and then checks that the result is still inside the public reference directory, so no phrasing of a question reaches anything else in the repository.

These restrictions are properties of the tool surface given to the model; none of them depends on instructions in the prompt.