OneCAD-Draft/docs/plugin-host-model.md

Scripting host model (API v7)

This note describes how a script controls OCS and what the host guarantees. The per-kind evidence is in the coverage ledger, the increment-by-increment record in the completion plan, and the script-facing reference in plugins/opencad-python/PLUGIN.md. A reviewer new to the change should start with the review guide.

Shape

Python runs in RustPython inside a separate plugin process (plugins/opencad-python), so a script cannot crash the editor. It talks to the host over the versioned IPC in crates/ocs_plugin_api. The host owns the drawing: it validates every request, records the undo step, refreshes scene caches and the shared document view, and marks the document dirty. The Python adapter is thin: it turns dictionaries into typed values and back, and holds no CAD rules of its own. Everything below is additive to API v7; a plugin built without it keeps working, and the default HostApi methods refuse.

What the host offers

Area	HostApi entry	What it does
Read	document(), entity_snapshot	The full CadDocument snapshot; a detached JSON view of any typed entity.
Entities	add_entity, remove_entity, update_entities_transaction, set_reference_only	Create, delete and replace entities. A transaction validates every replacement first and either commits all of them as one undo step or none. set_reference_only designates plugin-owned display geometry (API v9, see below).
Selection	selection, set_selection	Ordered, tab-scoped, exact (missing or duplicate handles reject the request).
Solids	solid_operation	Kernel-backed create, transform, booleans, extrude, region and surface from a profile, picture embedding. Every result is verified to lift back from its ACIS payload without loss; a refusal leaves the operands untouched.
Tables	table_operation	Layers, text and dimension styles, blocks and their contents, linetypes, layouts (see below).
Objects	create_xrecord, read_xrecord, update_xrecord, delete_xrecord, list_xrecords	Name-addressed XRecords under the drawing's named-object dictionary (API v8, see below): a non-graphical record carrier with strict create-vs-update semantics and one undo step per change.
Commands	run_command	Drives the real OCS command one step at a time (see below).
Settings	system_variable, set_system_variable	CLAYER, SNAPANG, and a read-only CTAB (the current layout).

Entity kinds outside a generated schema still read (handle, kind, layer) and take layer changes; nothing is writable that the host has not validated.

Entity validation

The host embeds a coverage catalog generated from its traced EntityType registry and crates/ocs_plugin_api/entity_coverage_policy.json. It classifies all 48 variants: 43 canvas kinds, three internal records (Block, BlockEnd, Seqend) and two opaque fallbacks (Extended, Unknown), and records for each property its source field, type, shape, snapshot readability, Python access (read_write, read_only, unmapped) and validation status.

Creation and edits are validated per kind (finite coordinates, positive radii and scales, unit directions, non-parallel vectors, in-range flags, reference targets that exist and have the right kind, and so on). Creation must be valid; an edit is refused only for problems it introduces, so already-invalid legacy values stay editable. Derived state is computed by the host, never trusted from a script (MLine geometry, Dimension measurement, Helix spline, Table merge dimensions, Spline rational flag, SectionSymbol counts, Viewport id and scale).

Nested records are checked too. A dictionary for a vertex, edge or face may not carry an unknown key, and the fields that define its geometry are required (entity_manifest.json, required_struct_fields); a typo can no longer turn into a zero coordinate. Everything else in a nested record defaults when absent.

Every kind is exercised through the real plugin runner by a per-kind audit that checks create, read, edit, delete, undo, invalid input with an atomic rollback, and a DWG and DXF round trip. A kind is Complete in the ledger only when all of those pass.

Tables

TableOperation is one additive request with typed variants, executed by HostSession::table_operation. Every refusal is decided before the undo step is recorded, so a failure changes nothing and leaves the history alone.

• Layers: create, modify, rename (entities follow), delete (refused while it holds objects unless erasing them is asked for), set current. Layer 0, Defpoints and the current layer are protected.
• Text and dimension styles: create, modify, rename, delete, set current. Dimension style fields travel as a validated JSON object; handles, xref fields and the name are host-managed. Renaming a text style updates the dimension styles that name it. Standard and any style that is current or in use are protected.
• Blocks: create from entities (copied in shifted by the base point, optionally erasing the originals), modify, rename (inserts follow), delete (refused while referenced), and add an entity to a definition through the normal validation path. A block may not nest itself, directly or through other blocks.
• Linetypes: simple dash, gap and dot patterns; rename follows layers and entities; delete is refused while anything uses it.
• Layouts: create with the default page setup and sheet viewport, rename, delete, switch (through OCS's own layout switch), and page setup (paper size, rotation, custom scale). Model is protected.

Style and layout operations act on the active document, because they use the application's shared style and view machinery.

XRecords (API v8)

create_xrecord, read_xrecord, update_xrecord, delete_xrecord and list_xrecords address XRecords by name under a dictionary path walked from the named-object dictionary root (an empty path is the root itself); missing dictionaries are created on create only, so the first call seeds a plugin's namespace. The payload is the codec's typed XRecordEntry pairs — the host puts no opinion on the content. Names and path fragments are symbol-validated, and dictionary addressing follows the codec's case-insensitive lookup. Create never overwrites and update never creates; each refusal is decided before the undo scope is touched, so a failure changes nothing, and each change joins a plugin-opened push_undo scope — the entity-write contract — so an operation composed of several writes (entities and XRecords) is a single undo step; with no scope open a change is one undo step of its own, whose undo restores the objects map byte-exactly, the auto-created dictionaries included. list_xrecords reports only the keys that address XRecords, in the dictionary's own order, so audits need no prior knowledge. Not yet mapped to Python.

Reference-only entities (API v9)

set_reference_only designates plugin-owned display geometry: a reference-only entity renders and snaps normally, but the built-in edit paths treat it like an entity on a locked layer — it is not grip-editable, not click-pickable (a click passes through to the editable entity underneath), and window, box, lasso and fence selections skip it. User-driven writes (grip drags, modify commands, the properties panel, user erase, MCP set_properties) refuse it; the owning plugin keeps full write access: add_entity, update_entity, update_entities_transaction, remove_entity and the XDATA record methods are unaffected, so a plugin can refresh derived display geometry in place. The record-method policy applies first: absent entities and locked layers refuse, and like them every refusal is decided before the undo scope is touched, so a failure changes nothing. A re-designation that would not change the state succeeds without recording an undo step. Undo composition is the entity-write contract: a change inside a plugin-opened push_undo scope joins it, so an operation composed of several writes is a single undo step under the plugin's label; with no scope open a change is one undo step of its own.

Storage is a scene-scoped in-memory set, not a file format. The host writes nothing of its own into the saved DWG/DXF — no APPID registration, no XDATA marker on the entity — so a saved drawing never depends on host state to load correctly. The set lives and dies with the drawing tab: closing a tab forgets its designations, and a freshly opened drawing starts undesignated. Because the designation is per handle, an in-place refresh (clone → edit → replace) keeps it by construction. Undo keeps the contract through an explicit designation channel on each history step (the same pattern the parametric-constraint directory uses): undo/redo restores the designation together with the entity it belongs to — erasing a designated entity inside a scope, then undoing, brings back the entity and its designation. Out of scope today: entities nested inside block instances (the block cache rebuilds instance wires without the flag) and a Python mapping. The MCP entities read exposes the flag as a reference_only boolean on every entity record.

Usage from a plugin

Build the derived geometry and designate it inside one push_undo scope, so creating a preview plus marking it is a single undo step under the plugin's label:

host.push_undo("CVI: refresh path preview");
let preview = host.add_entity(EntityType::LwPolyline(computed_path));
if let Err(error) = host.set_reference_only(preview, true) {
    host.push_error(&format!("set_reference_only failed: {error}"));
}

Refresh the display data in place through the same API — reads before the scope, mutations inside it (the entity-write contract's ordering rule):

let found = host
    .document()
    .entities()
    .find(|entity| entity.common().handle == preview)
    .cloned();
if let Some(mut entity) = found {
    // edit the clone; the designation travels with the entity
    host.push_undo("CVI: recompute preview");
    host.update_entity(entity);
}

Notes for plugin authors:

• add_entity ignores layer locks; set_reference_only does not — it refuses absent entities and locked layers exactly like the XDATA record methods, every refusal changes nothing, and an idempotent re-mark records no undo step.
• The designation travels with the entity: a clone → edit → replace refresh keeps it, so a plugin marks its display data once and refreshes freely.
• The designation is scene state, not file state: it does not survive DWG/DXF save and reload — a plugin re-marks its display data when the document opens, and idempotent re-marks keep that loop finite.
• Rendering and object snap are unaffected; grips, picking and the built-in write paths are the only observers.
• Plugins that read the document through the IPC client (the out-of-process runner) resolve entities by iterating document().entities(): a received DocumentSnapshot does not serve get_entity(handle) lookups. This is a property of the snapshot path, not of this API — host-side lookups are unaffected.

Commands

run_command takes a CommandRequest (Run, Start, Point, Text, Token, Entity, Selection, Enter, Cancel) and answers with a CommandOutcome: whether the command is done or waiting, its prompt, what input it accepts, its keyword options, how many entities it added, tokens nobody asked for, and any error. It uses the same primitives as the automation channel and finishes each step synchronously on the host thread, so a step never waits for anything that has to arrive from outside; an earlier fire-and-forget replay could hang, this does not.

Guards: refused while another command is active or off the active tab; refused for commands that could end the session or re-enter Python (QUIT, EXIT, CLOSE*, NEW, QNEW, OPEN, SAVE*, RECOVER, SCRIPT, RUNSCRIPT, SCRIPTCOMMANDS, PY_*); an editor or dialog a command opens is reported and closed by Cancel. The user can turn the whole facility off with the saved SCRIPTCOMMANDS preference (on by default; SCRIPTCOMMANDS 0 at the command line): while it is off, every request except Cancel is refused (so a command a script left waiting can still be closed). A script cannot change it, because the runner refuses that command. The runner does not consult the automation on/off switch, which governs the external MCP and serve channels. The Python layer cancels any step sequence that leaves a command waiting.

Python surface

ocs.active_document exposes entities, create_entity, delete_entity, transaction, selection, solids, layers, text_styles, dim_styles, blocks, linetypes, layouts, command, start_command and modify (offset, trim, extend, fillet, chamfer, move, copy, rotate, scale, mirror, erase, rectangular, polar, path and 3-D arrays, explode, join, break, stretch, lengthen and polyline edits). doc.coverage() lists the catalog, and doc.entities[handle].coverage one entry, with the readable and editable keys and the fields the typed snapshot carries but the model does not map.

Interactive picks use a token-based request and poll API because PY_RUN uses a fresh interpreter per invocation and a script cannot suspend while the user clicks. Drawing, selection and command notifications are kept in a bounded 256-event queue per tab between runs and drained with doc.poll_events(); an overflow is reported as {"type": "overflow", "dropped": n}. Tokens are bound to the tab that asked.

Known limits

• Twelve of the 43 canvas kinds are not yet Complete; each has a named blocker in the ledger. Seven are opencadcodec DXF bugs whose fixes are submitted upstream but not yet adopted.
• The DXF codec also drops some table properties (text style generation flags, block descriptions, a dimension style's text-style name) and mis-scales one angle; these are pinned by canaries and listed in cadcodec-reader-gaps.md.
• Not available from a script: complex (text and shape) linetypes, plot devices and named page setups, PEDIT's fit and spline options, and interactive-only commands such as HATCH.
• A command step nests inside the editor's message handling. In a debug build a deep step (PEDIT converting a line to a polyline) needs between 2 and 4 MiB of stack; the native builds link with at least 8 MiB (16 MiB on Windows, .cargo/config.toml), the nesting depth is fixed rather than recursive (PY_* is refused), and a release build uses far less.