Fix/liar177 fix(terminal): fix AI agent input, cursor rendering, and status detection in the terminal - #40
Open
liar177 wants to merge 37 commits into
Open
Fix/liar177 fix(terminal): fix AI agent input, cursor rendering, and status detection in the terminal#40liar177 wants to merge 37 commits into
liar177 wants to merge 37 commits into
Conversation
docs: add Simplified Chinese README translation
…re active On macOS, adds applicationShouldTerminate: to GPUI's delegate via ObjC runtime, returning NSTerminateCancel when sessions exist (triggers the system "failed to quit" dialog). On Windows, subclasses the GPUI window to intercept WM_QUERYENDSESSION with ShutdownBlockReasonCreate. User-initiated quit (Cmd+Q / Ctrl+Q) always proceeds immediately.
Check the WRAPLINE flag on each grid row before inserting a newline during copy. Wrapped rows are continuations of the same logical line and should be joined without a newline, matching standard terminal emulator behavior.
Cast through *const () before isize to satisfy fn-to-numeric-cast and function-casts-as-integer lints.
- Add CI step to stamp workspace version from the release tag before
building, so env!("CARGO_PKG_VERSION") reflects the actual release
- Replace hardcoded "v0.1.0" in settings panel with dynamic version
- Fix Windows rustfmt for shutdown_guard_windows.rs
GPUI's Window has an inherent window_handle() returning AnyWindowHandle that shadows the HasWindowHandle trait method returning Result. Use HasWindowHandle::window_handle(window) to call the trait method.
… resize Empty terminal rows use the default background color and produce no bg_rects_hsla entries, so they showed the panel_bg from the outer div rather than the terminal background color. This was visible as a visual cut-off when dragging a session from a smaller pane into a larger one. Fix by painting the full canvas bounds with terminal_bg at the start of the paint closure, before any cell-specific backgrounds. Non-default background cells still paint their own color on top, as before.
…ling When new output arrives while the viewport is scrolled into scrollback, alacritty's partial damage tracking only reports rows changed in the active area. But new output can push lines into scrollback, shifting all content visible in the scrolled viewport. Using partial damage leaves most rows stale, causing a cycling/repeating display artifact. Skip the partial-damage optimization when display_offset > 0 so all viewport rows are rebuilt from the current grid state.
When multiple signal files target the same session (e.g., parent + subagent), non-deterministic read_dir order could let a stale "idle" from the parent overwrite a newer "working" from the subagent, keeping the status stuck on idle.
On app restart, finalize_restored_session_bootstrap overwrote the workspace session's session_uuid with the old UUID from state.json. But the PTY was already spawned with the new UUID from Session::new() as the CODIRIGENT_SESSION_UUID env var. This mismatch caused hook signal routing to fail silently — signals carried the PTY's UUID but the workspace had the old one — keeping status stuck on Idle.
- Add sign-and-upload.ps1 script for local MSI signing with Certum cert - Add windows-release-signing.md documenting the full signing workflow - Exclude MSI from CI checksums (replaced by signed version post-build) - Reorganize docs into features/, release/, workflows/ subdirectories - Remove obsolete plan files and gitignore specs/
fix: terminal display cutoff and rendering issues
- Release workflow now creates draft releases for manual review - Remove author attribution from README - Update SmartScreen warning to reflect code-signed MSI
On Windows, GPUI 0.2.x's WindowControlArea::Drag causes WM_NCHITTEST to return HTCAPTION, which makes Windows enter a modal drag loop inside DefWindowProc. This re-enters the message pump while GPUI still holds RefCell borrows, causing a panic/freeze (issue oso95#31). Fix: on Windows, do NOT use WindowControlArea::Drag at all. Instead, post WM_NCLBUTTONDOWN(HTCAPTION) asynchronously via PostMessageW from an on_mouse_down handler, so the modal drag loop starts on the next message-pump iteration after GPUI's borrows are released. Also moves logo/title into a dedicated drag_region child div (cleaner separation from caption buttons) and adds double-click-to-maximize support on both platforms. Closes oso95#31
Reuses the existing session dropdown menu (rename, groups, end session) when right-clicking a tab, positioned at the cursor location.
The field was intentionally kept unused after the UUID overwrite fix (8d746d3) but never cleaned up. Removing it along with the stale NOTE comment and all construction sites.
The rename and group assignment modals used append-only text input with no cursor positioning. Add cursor_position tracking to SessionActionModal, reusing the existing cursor helper methods from the task creation modal. Supports: left/right arrows, Home/End, Delete, backspace and insert at cursor position, and Ctrl+V paste.
Covers the full release process: tagging, CI draft creation, local MSI code signing with SimplySign/Certum, checksum updates, and publishing. Includes troubleshooting and recovery steps.
PostMessageW(WM_NCLBUTTONDOWN, HTCAPTION, 0) passed lparam=0 meaning cursor position (0,0). Some Windows builds silently ignore the message when the position falls outside the window bounds. Now calls GetCursorPos and packs screen coordinates into lparam.
canvas_metrics Rc<Cell<>> is only populated during prepaint, so content_height is always 0 when render_terminal_scrollbar runs. Use rows * cell_height as the render-time estimate. Mouse handlers still prefer the real content_height from prepaint when available.
High-resolution touchpad momentum scrolling produced delta_y values of 500+ pixels, translating to 30+ lines per event — causing page-sized jumps. Cap to half the viewport height. Also reduce snap-to-bottom threshold from rows+lines (~30 rows) to 3+lines so it only triggers when genuinely near the bottom, fixing the infinite scroll feel.
… Windows GPUI 0.2.x re-dispatches WM_NCLBUTTONDOWN through its element tree before DefWindowProc can process it. When the title bar's on_mouse_down handler catches the event, DefWindowProc never runs — breaking both window drag (PostMessageW loop) and top-edge resize (drag region overlaps resize zone). Replace the direct PostMessageW(WM_NCLBUTTONDOWN) approach with a Win32 window subclass (SetWindowSubclass) that intercepts NC messages before GPUI's WndProc: - Drag: on_mouse_down posts a custom WM_APP message; the subclass catches it and calls DefWindowProc(WM_NCLBUTTONDOWN, HTCAPTION) directly, starting the OS drag loop when no GPUI borrows are held. - Resize: WM_NCLBUTTONDOWN for resize edges (HTTOP, HTLEFT, …) routes straight to DefWindowProc, bypassing GPUI's element dispatch. - Buttons: WM_NCLBUTTONDOWN for HTMINBUTTON/HTMAXBUTTON/HTCLOSE passes through to GPUI for normal handling. Fixes window drag not working on some Windows machines and top-edge resize being eaten by the title bar's mouse handler.
After a double-click, the WM_APP_DRAG_WINDOW from the first click may still be in the message queue when titlebar_double_click() has already maximized/restored the window. Check GetAsyncKeyState(VK_LBUTTON) before starting the drag — if the button is released, skip the stale message to avoid a spurious drag on the now-toggled window.
…n restart The resume command (claude --resume) was sent to the PTY immediately after bootstrapping, before the shell had completed its DSR handshake. The buffered command bytes corrupted PowerShell's cursor position query, leaving terminals blank. Now resume commands are enqueued and dispatched only after the shell produces its first output, with a 3-second fallback timeout.
fix: include Nerd Font family fonts in Windows terminal font detection
1. Normalize and validate Windows hook executable paths before registration. 2. Ignore PowerShell shell prompts in generic attention detection. 3. Rebuild terminal row caches when switching alternate screens. 4. Accumulate pending dirty rows across consecutive snapshots.
1. Classify visible terminal semantics for approval and response-ready states 2. Preserve legacy prompt patterns and support extensible status rules 3. Map Claude idle notifications to response ready 4. Add detector, UI, hook regressions and documentation
1. Register input handlers for all task creation fields. 2. Route IME composition and committed Unicode text correctly. 3. Preserve direct English, digit, and space input with regression tests.
1. Exclude INVERSE-flagged blank cells from the default-background skip path so TUIs that draw their caret as a reverse-video blank (e.g. claude's Ink `\e[7m \e[27m`) render it as a background rect instead of being discarded. 2. Add a regression test asserting a reverse-video blank produces a caret background rect.
1. Occlude the update toast from terminal mouse input.\n2. Add GPUI hit-testing regression coverage.
1. Anchor IME preedit to hidden-cursor software carets.\n2. Synchronize ConPTY size before restored agents start.\n3. Forward terminal editing control keys while preserving app shortcuts.\n4. Add cursor, resize, resume, focus, and key regression coverage.
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
4 participants
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.
Overview
This PR fixes a set of issues exposed when running AI agents such as Claude and Kimi inside the codirigent terminal,
centered on: cursor/input rendering errors, broken IME text input, inaccurate agent state detection, and a few minor
UI defects. Every change is covered by regression tests.
Changes
default-background blank cell, causing the caret to vanish and the input position to jump (e.g. to the far
right). We now keep INVERSE blanks and render them as caret blocks.
rows across consecutive snapshots for correct dynamic refresh.
terminal semantics, and map Claude's idle notification to "response ready".
before registration.
for every input field, correctly route IME composition and committed Unicode text, while preserving direct
English/digit/space input.
agents start; forward terminal editing control keys (arrows, backspace, etc.) while preserving app shortcuts.
add hit-testing regression coverage.
Testing
Each commit adds corresponding regression tests covering cursor rendering, size sync, session resume, focus, key
handling, state detection, IME input, and hit-testing.