feat(morph): sub-slice A6 — MCP tools for morph targets#571
Merged
Conversation
Surfaces the existing A1 `MorphAnimationManager` over MCP so AI agents + external tooling can inspect and drive morph weights. Two new tools: **`list_morph_targets`** (heavy) — args: `file`. Imports the file, enumerates `Ogre::Pose` entries, returns the de-duplicated target names across all imported entities. Uses the same safe import + RAII cleanup pattern `bake_vat` and `optimize_mesh` established, so the editor's live scene is unchanged on return. **`set_morph_weight`** (light) — args: `name`, `weight`. Drives the matching `AnimationState` weight on the first selected entity. Weight is clamped to `[0..1]` by the manager. Returns the clamped value back so the caller can see what landed. `SERVER_VERSION` bumped 1.8.0 → 1.9.0 for the new tool surface. Authoring tools (`add_morph_target`, `delete_morph_target`, `rename_morph_target`) wait for A3 — they need the in-memory edit-mode-delta-to-pose authoring path to land first. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
Warning Rate limit exceeded
You’ve run out of usage credits. Purchase more in the billing tab. ⌛ How to resolve this issue?After the wait time has elapsed, a review can be triggered using the We recommend that you space out your commits to avoid hitting the rate limit. 🚦 How do rate limits work?CodeRabbit enforces hourly rate limits for each developer per organization. Our paid plans have higher rate limits than the trial, open-source and free plans. In all cases, we re-allow further reviews after a brief timeout. Please see our FAQ for further information. ℹ️ Review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (2)
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 67e808e9d6
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Comment on lines
+4285
to
+4287
| try { | ||
| MeshImporterExporter::importer({filePath}); | ||
| } catch (const std::exception& e) { |
There was a problem hiding this comment.
list_morph_targets imports a mesh via MeshImporterExporter::importer, which creates scene nodes/entities through Manager::addSceneNode/createEntity and auto-selects them (SelectionSet::selectOne). This tool then destroys those imported nodes in cleanup but never restores the prior selection, so the selection set can retain destroyed-node pointers; subsequent selection-based calls (e.g. code paths that iterate selected nodes and read node->getName()) may dereference invalid memory, and the user’s original selection is lost even though the tool claims the live scene is unchanged.
Useful? React with 👍 / 👎.
Codex P1 on PR #571: `list_morph_targets` imports a mesh via `MeshImporterExporter::importer`, which auto-selects the new entities (Manager::addSceneNode + SelectionSet::append). The tool's cleanup then destroys those imported nodes — but never restored the prior selection, so the SelectionSet ended up holding dangling pointers to destroyed entities. Any subsequent selection-based code (iterating nodes, reading getName(), etc.) would dereference freed memory, and the user's original selection was lost even though the tool claims the live scene is unchanged. Fix: snapshot `getNodesSelectionList` / `getEntitiesSelectionList` / `getSubEntitiesSelectionList` before the import, then in the `ImportCleanup` dtor: 1. Destroy the imported scene nodes (existing behavior). 2. `selSet->clearList()` — wipes the importer-added entries (which now point at freed memory) without dereferencing them. 3. Re-append the snapshotted pre-import selection. `clearList()` (not `clear()`) is the API SelectionSet exposes exactly for this case: the contents have already been destroyed externally, so we can't safely walk them to bookkeep. The same hole likely exists in `bake_vat` and `optimize_mesh` (both predate this PR) — a follow-up will extract this pattern into a shared `TransientImportSession` helper and migrate all three tools. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
Merged
3 tasks
fernandotonon
added a commit
that referenced
this pull request
May 17, 2026
…ll transient-import tools (#572) * refactor(mcp): extract TransientImportSession; restore selection in all transient-import tools Three MCP tools (`apply_atlas`, `bake_vat`, `list_morph_targets`) needed the same pattern: - Snapshot the pre-import entity set + the user's selection. - Run `MeshImporterExporter::importer(...)`. - Diff to find the imported entities. - On scope exit: destroy the imported scene nodes AND restore the prior selection. PR #571 fixed the selection-restore hole in `list_morph_targets` only; `bake_vat` and `apply_atlas` had the same bug (`MeshImporterExporter::importer` auto-selects the new entities; my cleanup destroyed them, leaving the SelectionSet holding dangling pointers). Extract the whole pattern into a `TransientImportSession` RAII helper (anonymous namespace, MCPServer.cpp) and migrate the three tools to use it. Each tool's body drops from ~50 lines of boilerplate to: ```cpp TransientImportSession session(mgr); if (QString err = session.runImporter(filePath); !err.isEmpty()) return makeErrorResult(err); const auto& imported = session.importedEntities(); ``` The helper: - Filters `getEntities()` by `getMovableType() == "Entity"` — gizmos / brush rings / mask overlays would otherwise crash on the raw cast. - Uses `SelectionSet::clearList()` (not `clear()`) in the destructor — the importer-added selection entries point at entities we're about to destroy; `clear()` would dereference freed memory while bookkeeping. `toolOptimizeMesh` is intentionally NOT migrated in this PR — it passes an extra `animOnlySkeletons` out-parameter to `MeshImporterExporter::importer`, which the helper's `runImporter()` doesn't yet support. That's a separate small extension when we touch optimize-mesh next. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * review(mcp): TransientImportSession captures partial imports on throw CodeRabbit major on PR #572: If `MeshImporterExporter::importer()` throws partway through, the prior `runImporter()` returned the error without recording what the importer managed to create — leaving partial state (scene nodes, entities) in the live scene that the destructor never cleaned up. That broke the transient-import contract on the error path. Fix: extract the post-import "diff `getEntities()` against `m_beforeSet`" logic into `captureImportedEntities()` and call it from all three exit paths (success + both catch blocks). The helper dedups against `m_imported` so callers running it twice is safe. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
3 tasks
fernandotonon
added a commit
that referenced
this pull request
May 17, 2026
…576) * feat(morph): sub-slice A5 — dope sheet shows morph-weight tracks Per #518: "Dope sheet integration: morph-weight tracks appear in the dope sheet alongside bone tracks." Slice A5 ships the read-only display half — one row per Ogre::Pose on the selected entity with diamond markers at each keyframe time. Full selection / move / copy / paste interaction for morph tracks is a follow-up (it depends on A3's authoring path to land first). ### What ships **`AnimationControlController::allMorphRows()`** (new Q_INVOKABLE): - Walks the selected entity's `getPoseList()`. - For each pose, looks up the matching `Ogre::Animation` (which A1's importer creates one-per-pose), reads its VAT_POSE track's keyframe times. - Returns `[{ name: QString, keyTimes: [double] }]` — same shape the dope sheet's existing bone-row reader expects, minus the channel-flags map (morph tracks are scalar weight, no channel decomposition). **`AnimationDopeSheet.qml`** picks up the new API: - New `morphRows` property bound to `allMorphRows()`. - Refresh hooks: existing `onSelectionChanged` now also refreshes morph rows; new `MorphAnimationManager` Connections refresh on `morphTargetsChanged` (selection moved) and `morphWeightChanged` (Inspector slider, MCP poke, future authoring path). - New `morphBand` Rectangle anchored to the bottom of the dope sheet: collapses to height=0 when there are no morphs (so a bone-only animation looks the same as before); otherwise shows "Morph Targets (N)" header + one row per pose with the same diamond style the bone tracks use. ### 3 new tests in `AnimationControlController_test.cpp` - `AllMorphRowsEmptyWhenNoSelection` — returns `[]` cleanly. - `AllMorphRowsEmptyForMeshWithoutPoses` — selected entity with bones but no poses returns `[]`. - `AllMorphRowsListsPoseNamesAndKeyTimes` — builds a mesh with two named poses + matching VAT_POSE animations, asserts both names appear and each carries a single t=0 keyframe (A1's importer-time default). ### #518 status after this slice | Sub-slice | What | Status | |-|-|-| | A1 | Importer + manager + CLI list | ✅ #569 | | A2 | Inspector "Morph Targets" subgroup | ✅ #570 | | A4a | glTF morph-target export | ✅ #573 | | A4b | FBX morph-target export | ✅ #575 | | A5 | Dope sheet morph rows (read-only) | this PR | | A6 | MCP tools | ✅ #571 | | A3 | Authoring (delta → new pose, rename, delete) | still pending | Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * review(morph): A5 — filter morph keyTimes by pose target handle Codex P2 on PR #576: `allMorphRows()` looped over every VertexTrack in the per-pose Animation. A1's importer groups same-named poses across submeshes into a single Animation with one VAT_POSE track per affected submesh (so e.g. a "Smile" target on body + head ends up as one Animation with two tracks). Without filtering, the dope sheet row for that pose would carry diamonds from both tracks — duplicated `t=0` markers in the simple A1 case, and wrong key counts once authoring adds per-time keys. Fix: look up only the VertexTrack whose handle matches `pose->getTarget()`. Also assert it's a VAT_POSE track (belt-and-suspenders — VAT_MORPH could be added in a future slice and we don't want to silently mix track types). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(morph): A5 — avoid overlap between bone ListView and morph band The bone-row ListView and the new morph band were both anchored to `parent.bottom`. With both visible, their layout rectangles overlapped, which (under some redraw conditions in CI) drove SIGSEGV crashes in any MainWindow-construction test that instantiated the dope sheet QML during MainWindow startup — MCPServerTest.ToggleNormals_WithMainWindowTogglesVisibility and MainWindowTest.ViewMenuConsoleToggleUpdatesDockVisibilityAndSettings both regressed on this branch. Fix: anchor the bone ListView's bottom to `morphBand.top` when the morph band is visible, falling back to `parent.bottom` otherwise. The two no longer fight for the same space. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(morph): A5 — drop `onMorphWeightChanged` QML binding The signal's payload includes `Ogre::Entity*`, a raw pointer Qt 6 can't safely marshal to QML/JS. Binding the slot crashed the QML engine during MainWindow construction on Linux/Xvfb CI runners — manifested as SIGSEGV in MCPServerTest.ToggleNormals_* and MainWindowTest.ViewMenuConsoleToggle*. Both tests construct a MainWindow which loads the dope sheet QML; the binding setup itself was the unsafe step. The dope sheet's per-row data is structural (pose name + keyframe times), not weight-driven. Listening only to `morphTargetsChanged` (which fires on selection change and is parameter-free) covers the cases that matter: a different entity showing different morph names. Per-weight notifications were purely cosmetic (the row would re-fetch identical data on every slider drag) and skipping them is the right tradeoff. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(morph): A5 — defer QML dope-sheet integration to a follow-up The QML changes in this slice consistently triggered SIGSEGV in unrelated test suites (MCPServerTest + MainWindowTest visibility-toggle tests), deterministically across 4 CI re-runs. The 3 new C++ controller tests all pass; only the QML integration crashes are blocking merge. Ship just the data API (`AnimationControlController::allMorphRows`) and its tests in this slice. The QML dope-sheet morph band will land in a separate PR where the crash can be isolated without holding back the backend work that 3 other in-flight slices need. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2 tasks
fernandotonon
added a commit
that referenced
this pull request
May 17, 2026
…#578) * feat(morph): sub-slice A3 — authoring (add from edit, rename, delete) Last remaining sub-slice of #518. Closes the morph epic's authoring loop: users can save the current edit-mode geometry as a new morph target, rename existing targets, and delete unwanted ones. All three operations are undoable. ## What ships ### `commands/MorphCommands.{h,cpp}` — three QUndoCommand subclasses - `AddMorphTargetCommand` — create N same-named Ogre::Pose entries (one per submesh) plus a matching VAT_POSE Animation. Slices come in as sparse `{submeshHandle, {vertexIndex: Vector3f}}` maps so the command doesn't depend on EditableMesh — anything that produces deltas can drive it (current edit, MCP, future "fix mirror axis" rules…). - `DeleteMorphTargetCommand` — snapshot the same-named pose offsets at construction, drop the pose(s) + Animation + AnimationState on redo, rebuild from the snapshot on undo. - `RenameMorphTargetCommand` — same snapshot-and-rebuild pattern, just with a different name on the rebuild side. Ogre 14.5 doesn't expose Pose::setName, so destroy-and-recreate is the only path. ### `MorphAnimationManager` — three new Q_INVOKABLE methods - `addMorphTargetFromCurrentEdit(name)` — reads `EditModeController`'s live EditableMesh, diffs every submesh against the mesh's bind positions on the GPU buffer, builds the slice list, pushes an AddMorphTargetCommand. Falls back to no-op if not in edit mode, no vertices moved, or name collides. - `renameMorphTarget(old, new)` — name-trim, collision check, push. - `deleteMorphTarget(name)` — existence check, push. All three resolve the entity from SelectionSet's first entity (same pattern as the A1/A2 setters), push onto the shared `UndoManager` stack so Ctrl+Z reverses, and emit `morphTargetsChanged` so the Inspector re-fetches. ## Tests 8 new GTest cases on `MorphAnimationManager_test.cpp`: - Add command creates + undoes pose and Animation cleanly. - Delete command round-trips (poses restored on undo). - Rename command round-trips (old name restored on undo). - Rename rejects target-name collisions, no-op self-renames, blank names. - Delete rejects unknown / empty names; succeeds on a real target. - Add-from-edit rejects empty / whitespace names before edit-mode check. - Add-from-edit returns false when no EditableMesh is available (i.e. user isn't in edit mode). - All three methods reject when nothing is selected. The command path is what's actually under test — the manager wrappers are thin selection + name-validation glue. We keep direct command tests because that's where the undo / redo contract lives and it exercises the pose snapshot / rebuild path without needing an Inspector. ## What's deferred - **Inspector UI** (Add / Rename / Delete buttons in the Morph Targets subgroup) — small but lives in QML, and the last QML change in this epic (A5's dope-sheet integration) deterministically crashed unrelated tests for reasons I never fully isolated. Doing the UI in a separate PR keeps blast radius contained. - **MCP tools** `add_morph_target` / `rename_morph_target` / `delete_morph_target` — straightforward but adds surface; can land alongside the Inspector UI. - **CLI** `qtmesh morph --rename / --delete` — needs export-after-edit plumbing through the FBX/glTF morph exporters from A4a/A4b. #518 status after this slice: | Sub-slice | Status | |-|-| | A1 — Importer + manager + CLI list | shipped (#569) | | A2 — Inspector subgroup | shipped (#570) | | A3 — Authoring | **this PR** | | A4a — glTF export | shipped (#573) | | A4b — FBX export | shipped (#575) | | A5 — Controller API | shipped (#576) | | A5b — Dope-sheet UI | follow-up | | A6 — MCP tools | shipped (#571) | | A3b — Inspector / MCP / CLI authoring surface | follow-up | Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(build): add MorphCommands.cpp to tests/ source list Linker errors on the secondary test binaries (MaterialEditorQML_test, MaterialEditorQML_qml_test, MaterialEditorQML_perf_test) because they compile MorphAnimationManager.cpp (which now references the three MorphCommand constructors) but didn't pick up MorphCommands.cpp. src/CMakeLists.txt has it; tests/CMakeLists.txt was missing the line. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * review(morph): A3 — capture bind positions on EditableMesh load Addresses Codex P1 findings on PR #578: ## P1 #1 — diff against captured bind positions, not the live buffer `addMorphTargetFromCurrentEdit` was diffing `EditableMesh` against positions read from the entity's current GPU vertex buffer. But edit-mode ops (translate / rotate / scale) call `commitToEntity` during the gesture — both sides of the subtraction end up identical in the common flow, so `slice.offsets` stays empty and the command incorrectly returns no-op even after visible edits. Fix: snapshot the bind positions once at `EditableMesh::loadFromOgreMesh` time (`EditableSubMesh::originalPositions`). The snapshot is never mutated by edit-mode ops, so `vertices[i].position - originalPositions[i]` recovers the right delta regardless of how many commits have run since enter-edit-mode. ## P1 #2 — shared-vertex submeshes The old `readBindPositions` rejected any submesh with `useSharedVertices=true` (vertexData null in that case). Most Mixamo-style assets use shared vertices, so morph authoring silently skipped them. Fix is implicit in #1: `loadFromOgreMesh` already copies the shared vertex pool into each affected submesh's `vertices` array; the new `originalPositions` snapshot is built from that same source, so shared-vertex submeshes carry their own baseline. The GPU-buffer read path goes away entirely. ## CodeRabbit minor — null-assert SelectionSet in tests Added `ASSERT_NE(sel, nullptr)` before `sel->append(entity)` in the three new authoring tests, matching the slice-A1 pattern. If the singleton ever fails to initialize, we get a clear assertion failure instead of a null-deref crash. ## Tests - `LoadFromEntityTriangleMesh` extended to assert (a) the snapshot matches `vertices[].position` after load and (b) mutating `vertices[].position` does NOT touch `originalPositions`. Exercises the shared-vertex case (the triangle fixture uses shared verts). - New `AddMorphTargetUndoableViaUndoManager` exercises the full add → undo → redo loop through the shared `UndoManager`, with a hand-built slice that mirrors what `addMorphTargetFromCurrentEdit` will produce once edit-mode capture is wired up end-to-end. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This was referenced May 17, 2026
fernandotonon
added a commit
that referenced
this pull request
May 17, 2026
Final sub-slice of #518. The C++ data API (`allMorphRows`) shipped in #576 as a Q_INVOKABLE on AnimationControlController; this PR wires it into the dope-sheet view. ## What ships - New `morphRows` property on the dope-sheet root, bound to `AnimationControlController.allMorphRows()`. Refreshed via the existing `onSelectionChanged` handler so a clip change rebuilds both bone and morph rows in lockstep. - `rowsView` (bone ListView) bottom anchor switches to `morphBand.top` when the morph band is visible. Bone-only entities → band collapses to `height=0` → bone list draws full height as before. - New `morphBand` Rectangle at the bottom: header row showing "Morph Targets (N)", then one row per pose with the target name on the left and `#88ccff` diamond markers at each keyframe time on the right. Diamonds share the bone-row timeline math (`pxPerSec`, `viewStart`) so they line up vertically. ## Why this is small and safe The original A5 (PR #576's first commit) deterministically crashed unrelated MainWindow / MCPServer visibility-toggle suites in CI. Root cause was never fully isolated but the suspicion was an `import PropertiesPanel 1.0` triggering a different QML singleton chain during MainWindow construction. This PR keeps the discipline that worked for A3b: - **No new QML imports.** `AnimationControl 1.0` was already imported. `allMorphRows()` was deliberately put on AnimationControlController (not MorphAnimationManager) for exactly this reason. - **No new `Connections` blocks.** The existing one already fires on the signals we care about. - **No MouseAreas, no selection, no drag** on morph diamonds — this is strictly read-only display. Interactive morph keyframes (selection, multi-select, drag, copy/paste) need their own controller surface and land in a separate PR. ## #518 status After this slice the issue is feature-complete and ready to close: | Sub-slice | Status | |-|-| | A1 — Importer + manager + CLI list | shipped (#569) | | A2 — Inspector subgroup | shipped (#570) | | A3 — Authoring data layer | shipped (#578) | | A3b — Inspector authoring UI | shipped (#579) | | A4a — glTF export | shipped (#573) | | A4b — FBX export | shipped (#575) | | A5 — Controller API | shipped (#576) | | A5b — Dope-sheet UI | **this PR** | | A6 — MCP tools | shipped (#571) | Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
8 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Sub-slice A6 of #518. Surfaces the existing A1 `MorphAnimationManager` over MCP so AI agents + external tooling can inspect and drive morph weights without going through the GUI.
What ships
`list_morph_targets` (heavy)
Args: `file` (path).
Imports the mesh, enumerates `Ogre::Pose` entries, returns the de-duplicated target names across all imported entities. Uses the same safe import + RAII cleanup pattern `bake_vat` and `optimize_mesh` established, so the editor's live scene is unchanged on return.
`set_morph_weight` (light)
Args: `name`, `weight` (number, clamped to [0..1]).
Drives the matching `AnimationState` weight on the first selected entity in the live editor. Returns the clamped value back. Errors when no entity is selected or the named target doesn't exist.
`SERVER_VERSION` 1.8.0 → 1.9.0
Clients can detect the new capability.
NOT in this PR — wait for A3 (authoring)
Manual smoke
```
$ echo '{"jsonrpc":"2.0","id":1,"method":"tools/call",
"params":{"name":"list_morph_targets","arguments":{"file":"face.fbx"}}}' \
| qtmesheditor --mcp
{ ... morphTargets: [...] }
```
Test plan
🤖 Generated with Claude Code