This PR adds a new app setting to control whether block focus follows
cursor movement:
- `app:focusfollowscursor: "off" | "on" | "term"`
- Default is `"off"` (no behavior change unless enabled)
## What changed
- Added hover-focus behavior on block pointer enter.
- Added guardrails so hover focus does not trigger when:
- touch input is used
- pointer buttons are pressed (drag/select scenarios)
- modal is open
- pointer events are disabled
- block is resizing
- Added config key plumbing across settings types and schema:
- `pkg/wconfig/settingsconfig.go`
- `pkg/wconfig/metaconsts.go`
- `schema/settings.json`
- `pkg/wconfig/defaultconfig/settings.json` ->
`"app:focusfollowscursor": "off"`
- Added docs for the new key and default example:
- `docs/docs/config.mdx`
## Behavior
- `"off"`: do not focus on cursor movement
- `"on"`: focus follows cursor for all block types
- `"term"`: focus follows cursor only for terminal blocks
also updates ROADMAP.md, and fixes a node pruning bug on the FE, and
adds a new diff viewer that we can view the write_text_file and
edit_text_file diffs in. adds a backup file system that can be used to restore AI edited files back to their original states.
Don't allow tabs with active Wave AI sessions to get closed when we
close the last block. Have Cmd-W close Wave AI if it is focused (rather
than a random node). Also fixes some lurking bugs with the pinned tab
functionality (and adds some nice visual feedback when we try to close a
pinned tab).
Massive PR, over 13k LOC updated, 128 commits to implement the first pass at the new Wave AI panel. Two backend adapters (OpenAI and Anthropic), layout changes to support the panel, keyboard shortcuts, and a huge focus/layout change to integrate the panel seamlessly into the UI.
Also fixes some small issues found during the Wave AI journey (zoom fixes, documentation, more scss removal, circular dependency issues, settings, etc)
The current layout system uses a complex bidirectional atom architecture
that forces every layout change to round-trip through the backend
WaveObject, even though **the backend never reads this data** - it only
queues actions via `PendingBackendActions`. By switching to a "write
cache" pattern where local atoms are the source of truth and backend
writes are fire-and-forget, we can eliminate ~70% of the complexity
while maintaining full persistence.
----
Every layout change (split, close, focus, magnify) currently follows
this flow:
```
User action
↓
treeReducer() mutates layoutState
↓
layoutState.generation++ ← Only purpose: trigger the write
↓
Bidirectional atom setter (checks generation)
↓
Write to WaveObject {rootnode, focusednodeid, magnifiednodeid}
↓
WaveObject update notification
↓
Bidirectional atom getter runs
↓
ALL dependent atoms recalculate (every isFocused, etc.)
↓
React re-renders with updated state
```
---
## Proposed "Write Cache" Architecture
### Core Concept
```
User action
↓
Update LOCAL atom (immediate, synchronous)
↓
React re-renders (single tick, all atoms see new state)
↓
[async, fire-and-forget] Persist to WaveObject
```
### Key Principles
1. **Local atoms are source of truth** during runtime
2. **WaveObject is persistence layer** only (read on init, write async)
3. **Backend actions still work** via `PendingBackendActions`
4. **No generation tracking needed** (no need to trigger writes)
Ephemeral blocks can now be added to the LayoutModel for a tab. Only one
ephemeral block can exist at a time. It is placed above all other
blocks, including the magnified blocks.
Updates how magnified and ephemeral blocks overlay the other blocks.
Now, there's a blurred backdrop behind them that will obscure the other
blocks. As a result of this, the overlayed blocks are now translucent.
Rather than try to track the transition state, which was proving
unreliable, I am just directly tracking the node state and determining
whether to debounce the inner rect based on whether the user has the
prefers-reduced-motion setting or query, whether the node is resizing,
and whether it's currently magnified. I'm then using the actual
animation time setting to determine how long to debounce.
Adds a new setting for the gap size between tiles in a layout. Also
updates the resize handle calculations so they are dynamically generated
based on the gap size. Also updates the styling for the resize handles
to be more robust.
This also updates the default gap size to 3px.
This also slims out the Block Frame padding so it is just enough that
the blocks don't overlap when there's no gap.
This handles an edge case where a user deletes all unmagnified nodes,
leaving a final node that is still magnified. Because we ignore
magnify/unmagnify operations when there's only one leaf remaining, this
would result in the last node being stuck magnified until a new node is
added.
Also fixes a bug where the layout would not always update when a new
block was added.
Fixes an infinite loop in the layoutModel atom synchronization that
would cause the atom to update indefinitely when the root node is
deleted.
Also adds a dedicated `disabled` flag for the IconButton decl so we can
disable the onClick handler when the button is disabled.
Also updates the Magnify toggle button to use this new flag, so that
when there's only one leaf in a layout, the magnify button is disabed.
Adds a delay to the resize handle line first showing so it doesn't flicker when the mouse briefly passes over it. Also removes an unnecessary findNode call that was happening on every move. Also adjusts the pointer offset based on the display container bounding rect.
Rather than using a timeout to debounce changes to the
NodeModel.innerRect, I'll use a transition event on the display
container. This way, if the user disables transitions using the reduced
motion setting, changes to the innerRect value will not be debounced.
Adds a flag to the insert layout action to explicitly set the focus of a
newly inserted node. This also adds a flag in the starter layout to
focus on the terminal block.
Make the block content sizing update once when its node moves or becomes
magnified. By manually updating this inner sizing rather than letting
the block flow in the DOM, the animations of the block frames are much
smoother.
This also fixes an issue where two scrollbars were being rendered for
the Directory Preview widget.
This also sets zero padding on nodes when there's only a single node
being rendered.
Fixes the in-memory focused node stack so it can be used to find the currently-focused node. This is necessary for `closeFocusedNode` to work. Also fixes `validateFocusedNode` so it will set the first node to be focused if no focused node is available. Also cleans up leafOrder update pattern.
This adds a new NodeModel, which can be passed from the TileLayout to
contained blocks. It contains all the layout data that the block should
care about, including focus status, whether a drag operation is
underway, whether the node is magnified, etc.
This also adds a focus stack for the layout, which will let the focus
switch to the last-focused node when the currently-focused one is
closed.
This also addresses a regression in the resize handles that caused them
to be offset from the cursor when dragged.
---------
Co-authored-by: sawka <mike.sawka@gmail.com>
This simplifies the rect calculations by making the gap between pixels
into a padding applied directly to the leaf nodes. This means the gaps
won't be present in the overlay layer, so when dragging a node around it
is always over an OverlayNode and the pendingAction won't be unset.
Also simplifies onDrop handling
