feat(mcp): assembly mutates in shared cad_submit map by jeffglousher · Pull Request #63 · jackControls/noBS-CAD

jeffglousher · 2026-08-22T06:45:51Z

Summary

Stacks on feat(mcp): UI-owned inbox apply (#11 slice) #60 (feat/mcp-ui-owned-apply / nbcad-mcp-mutate) and brings in the assembly ToolSpecs from feat(mcp): host-neutral assembly component tools #62.
Registers the six assembly modeling mutates in the shared MUTATES map so cad_submit accepts them while attached (UI inbox apply path).
assembly_document / assembly_solution stay read-safe inspect tools (not inbox mutates).
Adds Assembly focus pack + disclosure tags; headless roundtrip + classifier coverage.

Tools added to `crates/mcp-mutate` (`engine_method` matches `host.rs`)

MCP name	engine_method	payload
`assembly_create_component`	`assembly_create_component`	Object
`assembly_update_component`	`assembly_update_component`	Object
`assembly_create_occurrence`	`assembly_create_occurrence`	Object
`assembly_update_occurrence`	`assembly_update_occurrence`	Object
`assembly_set_occurrence_pose`	`assembly_set_occurrence_pose`	Object
`assembly_set_occurrence_grounded`	`assembly_set_occurrence_grounded`	Object

Stacking / merge order

GitHub could not use feat/mcp-ui-owned-apply as base on jackControls/noBS-CAD (branch only exists on the fork / #60 head). Please merge after #60 (this branch is based on that tip). Includes #62 assembly ToolSpecs so ToolSpec ↔ mutate map stay in sync; #62 alone is not required if this lands after #60.

Do not merge before #60.

Test plan

cargo test -p nbcad-mcp-mutate
cargo test in mcp-server (42 passed), including:
- assembly_create_component_is_in_lookup_mutate
- assembly_create_component_accepted_by_cad_submit_classifier
- assembly_component_occurrence_grounded_roundtrip
- tool_spec_mutates_match_shared_inbox_map
- every_shared_mutate_is_accepted_by_cad_submit_classifier
Jack: confirm merge-after-feat(mcp): UI-owned inbox apply (#11 slice) #60

jeffglousher · 2026-08-22T06:45:58Z

Sitrep: stacks on #60 (shared mcp-mutate / cad_submit map) and carries the assembly ToolSpecs from #62. Six assembly mutates are now in MUTATES so attached cad_submit will not reject them. Inspect-only: assembly_document, assembly_solution. Please review; do not merge before #60.

jackControls

Requesting changes on the current stacked head. There are two data-loss blockers and one revision-accounting issue:

The carried assembly update schemas still allow destructive partial replacement: assembly_update_component requires only id/name, and assembly_update_occurrence only id/name/component_id, while serde defaults omitted ownership, transforms, parent, visibility, and grounding before the complete records are replaced. Require full replacement fields or use patch DTOs with preservation tests.
For assembly and other non-solid inbox results, applyInboxNow falls back to loadDocument(). That action unconditionally writes dirty: false, so an MCP-applied live edit can be treated as saved and the user can close without an unsaved-changes warning. Refresh the affected state while preserving/setting dirty: true.
The native inbox apply already increments engine_revision, then the store changes caused by applyInboxNow reach the subscription and call mcp_session_bridge_note_mutation, incrementing it again. Suppress the local mutation note for inbox-originated synchronization (and prefer a targeted refresh over the full document reload).

This also stacks on open #60 and carries #62 functionality; please rebase/deduplicate after the dependencies are corrected. I did not merge.

jeffglousher · 2026-08-22T14:15:42Z

Addressed Jack’s three #63 blockers (local only; not merging):

Assembly update destructive partials — Cherry-picked feat(mcp): host-neutral assembly component tools #62’s b6d6543 (ComponentDefinitionPatchDto / ComponentOccurrencePatchDto). Host merge preserves omitted fields; ToolSpecs require only id. Tests: update_component_and_occurrence_patches_preserve_omitted_fields, assembly_update_component_rename_preserves_bodies_and_lcs, assembly_update_occurrence_rename_preserves_pose_parent_flags.
dirty:false after inbox apply — applyInboxNow no longer calls loadDocument(). Non-solid results use refreshAfterInboxApply (assembly-targeted path) with dirty: true. Solid updates still go through applySolidUpdate (already dirty:true).
Double engine_revision — While inboxApplying, the store subscription skips mcp_session_bridge_note_mutation. Native apply’s bump remains the single advance; inbox-originated UI sync does not note again.

Commits: 6e6cdfa (cherry-pick #62 patch DTOs), 19800b9 (dirty + suppress). Still stacks on #60; please re-review.

Stack STEP import + solid move/copy ToolSpecs (jackControls#61) onto the fixed assembly submit-map tip (jackControls#63 @ 19800b9) so cad_submit accepts solid_import_step, solid_edit_import_step, solid_move_copy, and solid_edit_move_copy while attached. cad_script and cad_compare_solids stay read-safe control helpers (not inbox mutates). Upgrade cad_script to jackControls#61 @ 52266b2 behavior: attach/refresh seeds a cad_load_project_model baseline, and the e2e test actually replays the dumped script on a fresh CadServer (detach after attach so jackControls#60 session_read_only still holds). Bump MODELING_TOOL_COUNT and keep ToolSpec↔map sync.

jackControls

Re-reviewed at 19800b9. The latest PR-local fix looks correct: inbox refresh preserves dirty state and suppresses the duplicate revision note while apply is in progress. I cannot approve the stacked PR yet because it inherits PR60's unresolved atomic revision race and stuck conflicted/malformed head-of-queue behavior. Its MCP check is also failing on the inherited formatting diff. Please resolve/rebase those upstream blockers.

jeffglousher · 2026-08-23T14:23:28Z

Reset onto #60 tip 30785fe (cherry-picked assembly map + patch DTO + dirty/suppress). New tip: 45324f1.

Assembly mutate map entries only (import/move_copy already on feat(mcp): UI-owned inbox apply (#11 slice) #60 — not duplicated).
Kept ComponentDefinitionPatchDto / ComponentOccurrencePatchDto + rename preservation.
Dirty path: refreshAfterInboxApply (never loadDocument); retained feat(mcp): UI-owned inbox apply (#11 slice) #60 native atomic engine_revision / dead-letter (no JS noteEngineRevision).
Tests: nbcad-mcp-mutate 5 passed; mcp-server 47 passed; fmt clean.
History rewritten — please re-review. Still needed on top of feat(mcp): UI-owned inbox apply (#11 slice) #60.

jackControls

Re-reviewed at 45324f1. The PR-local work is sound: assembly mutates use the shared cad_submit map, dirty state remains set after inbox apply, the double revision bump is gone, and all 47 MCP-server tests pass.

This stack still inherits the two P1 coherency races from PR60: a mutation can occur between snapshot export and bridge write so stale geometry is published at the current revision, and switching retained project tabs does not rebind/advance the bridge so an op authored for tab A can dispatch through active_mut() into tab B. Assembly mutations make the second failure especially destructive. Please rebase after PR60 fixes those races and carries deterministic regression tests.

Rebase jackControls#63 onto jackControls#60 tip f241739. Mutate map entries and ToolSpecs already landed with jackControls#60/jackControls#62; keep unique classifier/lookup coverage so cad_submit accepts the six assembly modeling mutates while assembly_document and assembly_solution stay read-safe inspect tools.

jeffglousher · 2026-08-25T14:33:11Z

Rebased onto #60 tip f241739 (P1 publish + project-session races closed). Did not merge. Stayed off Jack #65/#66/#67.

New head: 48222e0

Unique vs #60 (map entries + ToolSpecs already on the #60/#62 stack):

refreshAfterInboxApply with dirty: true — assembly/non-solid inbox apply no longer falls back to loadDocument() (which cleared dirty)
dedicated assembly lookup + cad_submit classifier tests (assembly_document / assembly_solution stay inspect-only)

Tests: cargo test -p nbcad-mcp-mutate 5/5; cargo test in mcp-server 47/47.

Please re-review. Still merge-after-#60.

Rebase jackControls#63 onto jackControls#60 tip f241739. Mutate map entries and ToolSpecs already landed with jackControls#60/jackControls#62; keep unique classifier/lookup coverage so cad_submit accepts the six assembly modeling mutates while assembly_document and assembly_solution stay read-safe inspect tools.

jeffglousher · 2026-08-25T14:38:20Z

Rebased onto new #60 tip 5526bf0 (itself on #65/#66). Clean rebase — leftover is still just assembly cad_submit map + dirty-refresh (refreshAfterInboxApply, no loadDocument). Jack's project-tab / i18n shell untouched.

New tip: 9cb5a63. mcp-mutate 5, mcp-server 47 green. Not merging.

cad_submit writes inbox/<seq>.json; the desktop applies via host::handle and publishes. MCP does not write model.json (no last-writer-wins).

Address Jack's three jackControls#60 blockers: 1. call_tool enforces session_read_only for direct mutates while attached (cad_submit / inspect / export stay allowed); writeback:true attach is rejected. Regression covers attach → direct mutate rejected → cad_submit accepted → detach restores headless mutation. 2. WindowPublisher.engine_revision is the authoritative OCC gate, advanced atomically with live engine apply and on immediate UI mutation notes (not heartbeat-debounce alone). Deterministic tests cover stale-base after UI note and two same-base ops. 3. Shared nbcad-mcp-mutate name→engine-method+payload map drives both cad_submit accept-list and Tauri inbox dispatch; unsupported names are rejected before enqueue; failed applies dead-letter to inbox/failed/ so the queue cannot wedge. Table-driven coverage of accepted mutates.

Advance session engine_revision in run_ui_mutation under the publisher lock across the live engine call so inbox OCC cannot race a later JS note. Dead-letter conflicting and malformed inbox heads to inbox/failed/ so the queue cannot wedge; next sequence can apply. Sync jackControls#61 mutates (move_copy/import_step) into the shared map after rebase onto main.

Reserve captures engine_revision and write rejects a snapshot if it moved during export. Bind the bridge to the active native project-session identity so an A inbox op cannot apply to B.

Rebase jackControls#63 onto jackControls#60 tip f241739. Mutate map entries and ToolSpecs already landed with jackControls#60/jackControls#62; keep unique classifier/lookup coverage so cad_submit accepts the six assembly modeling mutates while assembly_document and assembly_solution stay read-safe inspect tools.

jeffglousher · 2026-08-25T14:59:45Z

Rebased onto new #60 tip 25bdf45 (itself on a0015f0 / #67). Clean rebase — leftover is still dirty-refresh (refreshAfterInboxApply, no loadDocument) + assembly cad_submit classifier tests. Jack's constraint-glyph / viewport shell untouched. New tip: 5c2ebc7. mcp-mutate 5, mcp-server 47 green. Not merging.

jackControls

Thanks for the updates. I rechecked the current head (5c2ebc7) and the dirty-state correction looks good. One refresh-state issue remains, and this branch also inherits the unresolved bridge blockers from PR60:

A targeted assembly refresh leaves an existing jointMotionPreview active. refreshSessionFromMcp() advances jointMotionPreviewGeneration and clears jointPreviewSolution/mechanismPreview, but does not clear the already stored jointMotionPreview. Both viewport paths prefer jointMotionPreview.solution over the newly refreshed assemblySolution, so an external assembly mutation can be applied successfully while the old motion pose continues to render and its controls remain stale. Please set jointMotionPreview: null in the assembly-refresh path and add a state-level regression test that begins with an active motion preview.
This PR is stacked on the current PR60 head and therefore includes PR60's two outstanding P1 concurrency issues: the cross-tab publication reservation/write identity race and the non-atomic MCP inbox sequence allocation. Please rebase once those are corrected.

The focused build/tests pass, but they do not cover the stored-preview refresh case.

assembly_create_joint/update return a joint DTO, so applyInboxNow no longer falls through to loadDocument() (dirty:false). Port jackControls#63 refreshAfterInboxApply with dirty:true and clear stale jointMotionPreview.

Drop the extra helper; applyInboxNow uses the same scene+document guard as jackControls#63 so joint DTOs take refreshAfterInboxApply (dirty:true).

Carry reserved session_id + project_session_id on the write path so a delayed A export cannot consume B's reservation. cad_submit allocates inbox sequences with create_new so concurrent submits cannot share a seq.

Rebase jackControls#63 onto jackControls#60 tip f241739. Mutate map entries and ToolSpecs already landed with jackControls#60/jackControls#62; keep unique classifier/lookup coverage so cad_submit accepts the six assembly modeling mutates while assembly_document and assembly_solution stay read-safe inspect tools.

applyInboxNow no longer falls back to loadDocument() (which cleared dirty). Assembly ops get a targeted refreshAfterInboxApply with dirty:true; other non-solid results refresh live state the same way. inboxApplying still guards re-entry; native apply already owns the engine_revision bump.

jeffglousher · 2026-08-25T19:24:45Z

Rebased onto #60 5447175 (identity-bound publish write + exclusive inbox seq). Leftover still dirty-refresh + assembly classifier tests.

assembly_create_joint/update return a joint DTO, so applyInboxNow no longer falls through to loadDocument() (dirty:false). Port jackControls#63 refreshAfterInboxApply with dirty:true and clear stale jointMotionPreview.

Drop the extra helper; applyInboxNow uses the same scene+document guard as jackControls#63 so joint DTOs take refreshAfterInboxApply (dirty:true).

jackControls

Thanks for the update—the inherited publish/inbox races and dirty-state path are now addressed. One stale-preview blocker remains: refreshAfterInboxApply invalidates jointMotionPreviewGeneration but does not clear jointMotionPreview. The viewport resolves jointMotionPreview?.solution before assemblySolution, so after an attached assembly mutation it can continue rendering the old motion pose despite the refreshed solution. Please add jointMotionPreview: null beside jointPreviewSolution: null (as the combined PR already does) and ideally cover the targeted refresh path with a regression test.

refreshAfterInboxApply invalidated jointMotionPreviewGeneration but left the stored preview in place. Viewport prefers jointMotionPreview.solution over assemblySolution, so an attached assembly mutate could keep rendering the old motion pose. Clear it beside jointPreviewSolution.

jeffglousher · 2026-08-25T23:55:36Z

Addressed the remaining stale-preview blocker.

refreshAfterInboxApply now sets jointMotionPreview: null beside jointPreviewSolution: null on the targeted assembly path (viewport prefers jointMotionPreview?.solution over the refreshed assemblySolution). Generation is still invalidated.

Cheap regression: assembly_inbox_refresh_clears_joint_motion_preview locks the assembly inbox-refresh block in appStore.ts.

New tip: e94e27e. mcp-mutate 5/5; new source-lock + classifier tests green. Not merging until Approved.

assembly_create_joint/update return a joint DTO, so applyInboxNow no longer falls through to loadDocument() (dirty:false). Port jackControls#63 refreshAfterInboxApply with dirty:true and clear stale jointMotionPreview.

Drop the extra helper; applyInboxNow uses the same scene+document guard as jackControls#63 so joint DTOs take refreshAfterInboxApply (dirty:true).

Sixth-pass: no applyInboxAll loadDocument path; leftover refreshAfterInboxApply stays the single jackControls#63 dirty:true contract. Lock cad_refresh vs cad_load_project_model joint parity and a part-document joint inbox typed reject.

The merge-base changed after approval.

jeffglousher requested a review from jackControls as a code owner August 22, 2026 06:45

jeffglousher mentioned this pull request Aug 22, 2026

feat(mcp): import/move_copy mutates in shared cad_submit map #64

Closed

2 tasks

jackControls requested changes Aug 22, 2026

View reviewed changes

jeffglousher requested a review from jackControls August 22, 2026 14:15

jackControls requested changes Aug 22, 2026

View reviewed changes

jeffglousher force-pushed the feat/mcp-assembly-submit-map branch from 19800b9 to 45324f1 Compare August 23, 2026 14:23

jeffglousher requested a review from jackControls August 23, 2026 14:23

jackControls requested changes Aug 23, 2026

View reviewed changes

jeffglousher force-pushed the feat/mcp-assembly-submit-map branch from 45324f1 to 48222e0 Compare August 25, 2026 14:33

jeffglousher requested a review from jackControls August 25, 2026 14:33

jeffglousher force-pushed the feat/mcp-assembly-submit-map branch from 48222e0 to 9cb5a63 Compare August 25, 2026 14:38

jeffglousher mentioned this pull request Aug 25, 2026

feat(mcp): host-neutral assembly joint create/update #68

Closed

2 tasks

jeffglousher added 4 commits August 25, 2026 14:55

feat(mcp): UI-owned inbox apply for attached sessions

97a1862

cad_submit writes inbox/<seq>.json; the desktop applies via host::handle and publishes. MCP does not write model.json (no last-writer-wins).

fix(mcp): close publish and project-session coherency races

25bdf45

Reserve captures engine_revision and write rejects a snapshot if it moved during export. Bind the bridge to the active native project-session identity so an A inbox op cannot apply to B.

jeffglousher force-pushed the feat/mcp-assembly-submit-map branch from 9cb5a63 to 5c2ebc7 Compare August 25, 2026 14:59

jackControls requested changes Aug 25, 2026

View reviewed changes

jeffglousher mentioned this pull request Aug 25, 2026

feat(mcp): assembly joint mutates in shared cad_submit map #69

Merged

2 tasks

jeffglousher added 3 commits August 25, 2026 19:21

fix(mcp): bind publish write to reserved identity; exclusive inbox seq

5447175

Carry reserved session_id + project_session_id on the write path so a delayed A export cannot consume B's reservation. cad_submit allocates inbox sequences with create_new so concurrent submits cannot share a seq.

jeffglousher force-pushed the feat/mcp-assembly-submit-map branch from 5c2ebc7 to c7aba14 Compare August 25, 2026 19:24

jeffglousher requested a review from jackControls August 25, 2026 19:24

jackControls requested changes Aug 25, 2026

View reviewed changes

jeffglousher requested a review from jackControls August 25, 2026 23:55

jackControls previously approved these changes Aug 26, 2026

View reviewed changes

jeffglousher closed this pull request by merging all changes into jackControls:main in df875a0 Aug 26, 2026

Navigation Menu

feat(mcp): assembly mutates in shared cad_submit map - #63

feat(mcp): assembly mutates in shared cad_submit map#63
8 commits merged into
jackControls:mainfrom
jeffglousher:feat/mcp-assembly-submit-map

jeffglousher commented Aug 22, 2026 •

edited by jackControls

Loading

Uh oh!

jeffglousher commented Aug 22, 2026

Uh oh!

jackControls left a comment

Uh oh!

jeffglousher commented Aug 22, 2026

Uh oh!

jackControls left a comment

Uh oh!

jeffglousher commented Aug 23, 2026

Uh oh!

jackControls left a comment •

edited

Loading

Uh oh!

jeffglousher commented Aug 25, 2026

Uh oh!

jeffglousher commented Aug 25, 2026

Uh oh!

jeffglousher commented Aug 25, 2026

Uh oh!

jackControls left a comment

Uh oh!

jeffglousher commented Aug 25, 2026

Uh oh!

jackControls left a comment •

edited

Loading

Uh oh!

jeffglousher commented Aug 25, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

jeffglousher commented Aug 22, 2026 • edited by jackControls Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Tools added to crates/mcp-mutate (engine_method matches host.rs)

Stacking / merge order

Test plan

Uh oh!

jeffglousher commented Aug 22, 2026

Uh oh!

jackControls left a comment

Choose a reason for hiding this comment

Uh oh!

jeffglousher commented Aug 22, 2026

Uh oh!

jackControls left a comment

Choose a reason for hiding this comment

Uh oh!

jeffglousher commented Aug 23, 2026

Uh oh!

jackControls left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jeffglousher commented Aug 25, 2026

Uh oh!

jeffglousher commented Aug 25, 2026

Uh oh!

jeffglousher commented Aug 25, 2026

Uh oh!

jackControls left a comment

Choose a reason for hiding this comment

Uh oh!

jeffglousher commented Aug 25, 2026

Uh oh!

jackControls left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jeffglousher commented Aug 25, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

jeffglousher commented Aug 22, 2026 •

edited by jackControls

Loading

Tools added to `crates/mcp-mutate` (`engine_method` matches `host.rs`)

jackControls left a comment •

edited

Loading

jackControls left a comment •

edited

Loading