molstar-scene

name	molstar-scene
description	Use when operating an active Mol* scene through allowlisted Burrete actions and typed observe/action results.

Molstar Scene

Use this workflow for scene operations:

• focus ligand;
• label the active selection or a precise component;
• show ligands;
• select residues;
• focus selection;
• compute lightweight contacts;
• hide or show waters;
• show molecular surface;
• color by chain;
• reset camera;
• apply a compact declarative scene schema for target/component requests such as highlighting protein, selecting active-site loops, and focusing a named selection;
• load a complete MolViewSpec scene payload when the agent already has one;
• translate user requests into MolViewSpec-informed selectors, components, representations, color, opacity, labels, camera, focus, and canvas concepts;
• export image or screenshot when supported.

Contract

Use one typed action surface:

bun scripts/burrete-agent.mjs act --session-dir <dir> '{"type":"hide_waters"}' --wait-ms 12000

For browser-agent-shell, use the sessionDir returned by open --mode browser-agent-shell; the browser shell polls that session over its local dev-server endpoint and relays actions to the active Mol* iframe. Do not use Browser DOM clicks for scene edits when this session contract is available. If the user is already looking at a browser-agent-shell URL and the session directory is not in context, pass that URL directly with --url; the CLI will resolve the shell session through /__burette/agent-session/session.json when the shell is alive.

Do not add one endpoint per Mol* command. The action body must be allowlisted and serializable.

MolViewSpec-informed scene language

Use MolViewSpec as the mental model for agent scene requests, but keep two execution paths separate:

• apply_scene edits the currently active Burrete/Mol* viewer by relaying allowlisted commands. Use it for highlight, select, focus, contact neighborhood, water/surface/theme toggles, screenshot, and reset camera.
• load_mvs loads a complete MolViewSpec state (mvsj JSON or mvsx archive) through Mol* loadMvsData. Use it when the request really needs a full MVS tree: download/parse/structure/transform/component/representation/color/ opacity/label/tooltip/focus/camera/canvas/primitives/volume/animation.

Map natural language to these MVS concepts:

• component: what atoms/residues/chains are affected.
• representation: how that component is drawn, e.g. cartoon, ball_and_stick, line, spacefill, surface.
• color/opacity: visual emphasis for the component or a sub-selection.
• label/tooltip: textual annotation for a component.
• focus: camera fit to a component.
• camera: explicit root-level camera orientation (target, position, up).
• canvas: global view settings such as background color.
• transform/instance: coordinate movement or duplicated transformed instances; only promise this through a full MVS scene, not through quick apply_scene.

Selectors follow MolViewSpec vocabulary. Prefer these static selectors when they match the request:

"all" | "polymer" | "protein" | "nucleic" | "branched" | "ligand" | "ion" | "water"

For precise requests, use an object selector. Prefer label_* identifiers when the file exposes them, but accept auth_* when the user names author/PDB numbering:

{"label_asym_id":"A"}
{"label_asym_id":"A","beg_label_seq_id":45,"end_label_seq_id":58}
{"auth_asym_id":"A","beg_auth_seq_id":100,"end_auth_seq_id":112}
{"label_comp_id":"PYZ"}
{"comp_id":"PYZ"}
{"label_asym_id":"A","label_seq_id":377,"label_atom_id":"CA"}

Use an array of selector objects as a union, e.g. multiple chains or separated loop ranges.

apply_scene accepts MVS-like components or operations with selector/target, optional label, optional visual hints, and operation flags:

bun scripts/burrete-agent.mjs act --session-dir <dir> '{
  "type": "apply_scene",
  "components": [
    {"selector": "protein", "label": "Protein", "highlight": true, "color": "#4f8cff"},
    {"selector": {"chain": "A", "range": [45, 58]}, "label": "Active loop", "select": true, "focus": true}
  ]
}' --wait-ms 12000

Use these request patterns:

# Highlight the whole protein.
bun scripts/burrete-agent.mjs act --session-dir <dir> '{
  "type": "apply_scene",
  "components": [
    {"selector": "protein", "label": "Protein", "highlight": true, "color": "#4f8cff"}
  ]
}' --wait-ms 12000

# Select and focus an active loop by residue range.
bun scripts/burrete-agent.mjs act --session-dir <dir> '{
  "type": "apply_scene",
  "components": [
    {"selector": {"label_asym_id": "A", "beg_label_seq_id": 45, "end_label_seq_id": 58}, "label": "Active loop", "select": true, "focus": true}
  ]
}' --wait-ms 12000

# Focus a ligand and ask for a local protein neighborhood/contacts.
bun scripts/burrete-agent.mjs act --session-dir <dir> '{
  "type": "focus_ligand",
  "selector": {"comp_id": "PYZ"},
  "allowAmbiguous": true,
  "showNeighborhood": true,
  "radiusA": 5
}' --wait-ms 12000

# Add a visible 3D label to the last focused/selected ligand.
bun scripts/burrete-agent.mjs act --session-dir <dir> '{
  "type": "label_selection",
  "selection": "last",
  "text": "PYZ A:378"
}' --wait-ms 12000

# Reset camera after manual movement.
bun scripts/burrete-agent.mjs act --session-dir <dir> '{"type":"reset_camera","args":{"durationMs":250}}' --wait-ms 12000

When a user says "move the scene", decide which meaning applies:

• camera movement/orientation: use focus_selection, focus_ligand, reset_camera, or a future explicit camera action. Do not simulate this with Browser drag events.
• structure movement/rotation/instances: build and load a complete MVS scene with transform or instance nodes via load_mvs.

Use load_mvs only for complete MolViewSpec (mvsj or mvsx) payloads that should be handed to Mol* loadMvsData as scene state:

bun scripts/burrete-agent.mjs act --session-dir <dir> '{
  "type": "load_mvs",
  "json": {
    "metadata": {"title": "Protein and ligand view", "version": "1"},
    "root": {"kind": "root", "children": []}
  },
  "options": {"replaceExisting": true}
}' --wait-ms 12000

Upstream References

Ground new scene language in:

• MolViewSpec docs: https://molstar.org/mol-view-spec-docs/
• Tree schema: https://molstar.org/mol-view-spec-docs/tree-schema/
• Selectors: https://molstar.org/mol-view-spec-docs/selectors/
• Camera settings: https://molstar.org/mol-view-spec-docs/camera-settings/
• MolViewSpec repository: https://github.com/molstar/mol-view-spec
• Mol* examples: https://github.com/molstar/molstar/tree/master/src/examples

Relevant local Mol* examples when node_modules is installed:

• node_modules/molstar/lib/examples/basic-wrapper/index.js
• node_modules/molstar/lib/examples/proteopedia-wrapper/index.js
• node_modules/molstar/lib/examples/interactions/index.js
• node_modules/molstar/lib/examples/mvs-stories/stories/
• node_modules/molstar/lib/extensions/mvs/camera.js

Verification

After scene actions, run observe and inspect the action result. Use Browser or Computer only for visual confirmation. If the active fixture has no waters, hide_waters should return a successful typed no-op such as componentCount: 0 instead of a false failure.

For browser-agent-shell and desktop sessions, check observe.scene.selection first. It reports the last known selection/focused ligand with selectionId, ligand identity, and counts when the active viewer returned them.

Ligand focus can be a partial success. If showNeighborhood cannot resolve a protein/contact target, the action should still succeed when the ligand was selected/focused and return neighborhood.ok: false with a typed error. Do not retry the whole focus just because contacts failed.

Unsupported Cases

If a selector cannot resolve, return or report typed failures such as SELECTION_EMPTY or NO_STRUCTURE. Do not guess from screenshots.

Persistent overpaint and arbitrary representation graph rewriting are not yet a general apply_scene guarantee. If durable representation, opacity, labels, tooltips, primitives, volumes, animations, or coordinate transforms are required, prefer a full load_mvs payload or implement a new allowlisted BurreteSceneActions handler first.