Plugin API 1.3 adds what commit-message generators and before-commit checks need:
- `menu = "commit.message"`: a plugin command, with the plugin's icon, in the toolbar above the commit message of the Commit tool window and the Commit dialog.
- `commit.pending-commit()` (`vcs.read`): what Commit would commit right now, per file its repository, path, status and a `git diff` of only the included files and changed blocks against HEAD (1 MiB in all), plus the message, Amend and a revision. The diff is made by the same selection the commit uses, in a copy of the index, so the index and the working tree stay as they are; unsaved editors of the included files are saved first, as Commit does. Only in trusted projects.
- `commit.set-commit-message(revision, message)` (`vcs.write`): replaces the message only while the draft is still at that revision, so text the user typed meanwhile is never overwritten (`stale`).
- `[[contributes.commitChecks]]` and the optional `commit-checker` export (worlds `plugin-commit-checks` / `commit-checks`): checks listed in Commit Options (on by default, `vcs.commit.disabledPluginChecks` turns them off) and run after the IDE's own; their problems, optionally with a location, stop the commit until Commit Anyway.
The window answers the host through `CommitRequested` / `plugin_answer_commit`, also for remote windows. The Rust SDK gains `host::pending_commit`, `host::set_commit_message`, `openidea_plugin::commit` and their mock; a `commit` test fixture exercises them. All components are rebuilt for the WIT change.
Refs #1650
* fix(lsp): drop language servers whose workspace root was deleted (#1726)
A per-root server (e.g. a vtsls rooted at a nested tsconfig dir) is
spawned with its root as the process working directory. On Windows,
once that directory is deleted every spawn fails with os error 267 and
the crash backoff keeps respawning it, leaving a permanently 'Crashed'
entry in the Language Servers panel and a process tree that can never
work again.
- ensure_started: skip starting a server whose root no longer exists.
- on_instance_closed: after a crash, if the root is gone, drop the
server (LanguageServerRemoved) instead of retrying forever.
- restart: return a clear error when the panel restart is pressed for
a server whose root vanished.
- open_for_file_op: skip sampled candidates already deleted on disk.
Fixes#1726
* fix(lsp): drop dead-root servers claimed by ensure_started
A crash with the root still alive enters the restart backoff; if the root
is deleted during that sleep (or an open-file request arrives later),
ensure_started used to return early and leave a permanently "Crashed"
entry: no instance, no watcher, nothing re-arming it. Claim the Idle
server first (begin_start never claims a live one), then drop it when its
root is gone, mirroring on_instance_closed.
Adds root_deleted_during_crash_backoff_drops_the_server, which fails on
the previous code (server never removed) and passes with this fix.
Plugin API 1.3 gains an optional hover-provider export (worlds plugin-hover
and hover): hover(doc, at) returns Markdown and the range it describes. A
plugin declares the hover kind under [[contributes.providers]] for any
language; the host asks it with the language server, under the edit-path
budget, and shows the server's hover first and each plugin's below a rule.
Declaring the kind needs engines.openidea ^1.3. The Rust SDK gains
openidea_plugin::hover and register_plugin!(T, hover); every shipped
component is rebuilt for the WIT change.
Pest test() and it() calls and describe() blocks get gutter Run and Debug,
nested by describe and filtered by the names Pest prints (`outer` → it does
x, plus a data set suffix). Projects with vendor/bin/pest get a Pest test
configuration of the PHPUnit kind, pest commands are recognised, Pest's
pest_qn:// hints are read and test tree reruns filter Pest tests by name.
Repo::gix() reopened the cached gitoxide handle only when config or packed-refs changed. An external git fetch/pull/repack writes new packs under objects/pack and moves refs, but never touches those files — so the ODB kept serving the snapshot taken at open, and refs whose tip lived only in the new pack failed to peel: branches vanished and status warned "Could not peel reference". Watch the pack directory's mtime too and reopen when it changes, so externally added objects become visible (issue #1919).
Fixes#1919
* fix(vfs): warn when the inotify watch limit leaves folders unwatched
When watch() hits the OS watch limit mid-tree (inotify ENOSPC/EMFILE), the affected directories were silently never watched: the project stayed stale with no hint. Record the failure on DirState and surface it through the new Watcher::unwatched_reason() — the host emits the 'sync may be slow' warning once the project opens and once on the first event batch after a late failure (deferred walk / update_watches).
Fixes#1915
* fix(vfs): latch the startup watch-limit warning so batches don't repeat it
gopls builds a module for the host's GOOS/GOARCH only, so usages in files such as foo_windows.go were missing from Find Usages on macOS, and renaming left them behind, breaking the build for that platform. Those files of the searched modules are now resolved the way other modules are: each mention of the name in its own view (gopls builds one for the file's platform), added after gopls' answer and included in Rename.
Incoming Calls now also lists the callers found this way, in other modules and in build-constrained files: the declaration containing each call site.
Finding usages of a common Go name such as New opened every project file that mentions it and asked gopls again, which took up to seconds and found nothing more: gopls answers from the view of the file it is asked about, and opened files only get views of their own. Go searches now skip that widening.
Usages in other modules of the project that use the declaration's module through a local replace (a multi-module repository without go.work) are found the way GoLand finds them: each mention of the name there is resolved in its own module's view, and those resolving to the declaration are added after gopls' own answer. Rename edits them too.
Supersedes #1749.
Co-authored-by: tiube00001 <[email protected]>
Angular component templates in @Component({ template }) were parsed as TypeScript strings, so template highlighting and tag closing never applied. TypeScript files in Angular projects (angular.json or @angular/core) now use an Angular TypeScript mode that mounts the Angular template parser on literal template strings; other strings and non-Angular projects are unchanged. Inline and external templates both close tags on /, with multiple cursors and undo.
Closes#1895
Co-authored-by: jaw52 <[email protected]>
go.mod and go.work opened as plain text: their language definitions had no ui-mode, so no CodeMirror language loaded. Add a GoMod mode (a small stream parser): directives as keywords, module paths as strings, versions as numbers, // comments; shared by go.mod and go.work.
Co-authored-by: elfadl <[email protected]>
Open .ino sketches with C++ highlighting and the C++ file icon. Sketches get their own Arduino language with no language server: clangd can't compile a sketch on its own (implicit Arduino.h, generated prototypes) and reported "Unable to handle compilation" on line 1 of every sketch.
Co-authored-by: mushketer888 <[email protected]>
The New Branch dialog refused any existing local name, so a branch could not be moved to a selected commit. When the name matches a local branch of the target repository, the dialog now offers an unchecked Overwrite existing branch option (git switch -C / git branch --force); the consent clears when the name changes, and Git's checkout checks keep the old tip and local changes if switching fails. New Branch from a Git Log commit also targets that commit's repository in a multi-repository project.
Co-authored-by: shohruh-genius <[email protected]>
Stop flipped the session back to Run as soon as the direct child exited, but a shell, Maven or the debug adapter can exit while the application it started still holds its port, so the next Run failed with "port already in use". The console also went quiet on a debug Stop because the DAP session was disconnected before java-debug drained the program's output, and closing a window killed running programs with SIGKILL at once.
Run and Debug sessions now stay in a stopping state from the Stop press until the whole process group (or the debuggee's pid tree) is gone. As in IntelliJ IDEA, the first Stop only interrupts and waits, and a second Stop kills; the Stop button turns into Kill process meanwhile. A rerun or restart waits for the old process without a time limit, and the termination time limit setting now only bounds a configuration's stop command. Debug launches record the program pid from the adapter's process/processid event, signal its pid tree and keep the adapter attached until the program exits; closing a project also stops debug sessions under its root. Closing a window with running programs shows a Terminating panel with Kill process and closes once they are gone.
Co-authored-by: misswell <[email protected]>
Usage-count lenses and Safe Delete now search only the workspace's sources, and a source type with no implementations there is answered without building the type hierarchy over every dependency jar. These background searches had piled up in jdtls and delayed Show Usages and navigation by seconds in projects with many dependencies.
Inside PHP strings, plugin completions were buried under project file paths and go to definition mixed the plugin's answer with the server's guesses and with IDE helper stubs. Now path items are left out when a plugin answered at that position, and in PHP and Blade paths only show once the string reads like a path. Inside a string literal (checked on the document's syntax tree) plugin definitions replace the server's; elsewhere both are kept, the plugin's first, so e.g. MyBatis mapper navigation still offers jdtls' answer. Generated helpers (_ide_helper*.php, .phpstorm.meta.php, vendor/_laravel_ide*) give way to real declarations, an empty plugin answer is marked incomplete so typing asks again, and a callable accepted before a closing quote no longer gets ().
Closes#1336
Co-authored-by: yzPeedro <[email protected]>
A plugin view declared with `location = "services"` (plugin API 1.3, `engines.openidea ^1.3`) shows its tree in the Services tool window, next to Docker and the run configurations. The nodes, the toolbar (acting on one selected node) and their events come from `view-provider`; `viewItem.context` menus give the nodes a right-click menu. The pane beside the tree comes from a new optional export, `service-provider.detail(view, key)` (worlds `plugin-services` / `services`), which returns typed tabs: a message, read-only text or a process log. The IDE switches tabs without calling the plugin.
For a log, the UI names only the node and tab (`plugin_service_log`); the host asks the plugin for the pane again and runs its program with the plugin's `process:` grant in the project root, with the arguments of the toggles that are on, streaming its output into the log pane Docker uses until it exits (the exit code is shown), the pane closes or the Services window is hidden. A cancelled request starts nothing, lines are cut at 16 KiB, and revoking the grant or cancelling kills the program and what it started. The log pane moves to `src/ui/LogPane` and drops output that arrives after its source was stopped. The Rust SDK gains `openidea_plugin::services` (`Services`, `Tab`, `Pane`, `ProcessLog`). All components are rebuilt for the WIT change.
Closes#1736
Co-authored-by: arnaud-ritti <[email protected]>
The Maven path tests map a free drive letter with `subst`, but nextest runs them in separate processes that could pick the same letter at once, so one test's mapping vanished under the other (ERROR_PATH_NOT_FOUND). The helper now holds a file lock in the temp directory from choosing the letter until the drive is unmapped. Both Maven path tests share this helper; no other `subst` callers exist.
The native Claude Code provider turned each AskUserQuestion question into a
permission request with only the option labels, shown in the permission
banner. Descriptions, headers, previews and multi-select were lost, and the
only way to give another answer was Skip, which tells Claude the user didn't
answer.
All questions now go out in one request carrying the full tool input, and
the conversation shows them as a card above the composer: each option with
its description, radio buttons or checkboxes for multi-select, an Other
choice with a text field, the option's preview beside the options on a wide
card (below them on a narrow one, sized to the tallest preview so hovering
never moves the options), number keys to pick, Esc to skip, and one Submit
for every answer.
Ordinary permission requests keep the banner, and Skip answers as before.
The answers go back in Claude Code's own format: question text to answer,
multi-select picks and Other text comma-joined, with an item that contains
the separator or a quote JSON-quoted so it reads back as one item.
Closes#1628
Co-authored-by: NealimeKenna <[email protected]>
* feat(maven): go to declaration for POM properties and parent-managed dependencies
Ctrl/Cmd-click (or Go to Declaration) on a `${property}` reference in a
pom.xml opens the `<properties>` entry that defines it - the POM's own,
else the first one along the parent chain (a child's definition wins; an
active profile's counts as the POM's own) - and a dependency / plugin
whose version a `<dependencyManagement>` / `<pluginManagement>` entry
provides opens that managing declaration. The chain follows
`<relativePath>` in the project and the local repository (Settings >
Maven's, else ~/.m2) for external parents, which open read-only as plain
files; `project.*` built-ins go to their coordinate elements. When
nothing resolves under the caret, nothing navigates: the usual 'Cannot
find declaration to go to' hint shows instead.
The backend answers a caret position (0-based line / UTF-16 column) with
the winner's file, range and kind (maven_goto_declaration), reusing the
build model's POM reader and parent resolver (ide_build::maven) instead
of a new parser. The frontend contribution (build-tools.pomNavigation)
claims only positions that can resolve - inside a `${...}` reference or
a <dependency> / <plugin> element - so the Cmd-hover underline stays out
of the rest of the file.
Refs #1735
* test(maven): normalize expected goto paths like the parent chain does
parent_chain resolves every path it returns through dunce::canonicalize,
but the goto_declaration tests compared those results against raw temp
paths, which differ on the CI runners: on macOS the temp dir sits behind
the /var -> /private/var symlink and on Windows under the RUNNER~1 8.3
short name with backslash separators, so the Rust (macos-latest) and
Rust (windows-latest) jobs failed in the new assertions while the
functionality itself was right.
Feed the expected paths through the same canonicalization call before
comparing, mirroring the production behavior on every platform; on a
plain Linux checkout the canonical form equals the raw path, so the
assertions stay equivalent there.
Refs #1735
Extends the VenvTool set with a "Conda Environment" option: the dialog
runs 'conda create -p <location> -y', pinning 'python=<major.minor>' from
the chosen base interpreter when it reports a version. An existing
conda-meta directory is refused before conda runs, and a missing conda
binary reports an install hint instead of a spawn failure.
Refs l0ng-ai/openidea#1724
* feat(sdk): create Python environments with uv and Poetry in Add Interpreter
Add a VenvTool parameter to sdk_python_create_venv so the Add Python
Interpreter dialog can build a new environment with virtualenv, uv or
Poetry, matching what the environment discovery already finds. Poetry
keeps envs in its own cache, so the dialog disables the location field
for it and resolves the interpreter via 'poetry env info --path'.
Fixesl0ng-ai/openidea#1776
* fix(i18n): update text catalogs for the new env-tool strings
The Add Interpreter dialog gained uv/Poetry labels and per-tool notes,
and its spinner text changed from 'Creating virtual environment' to
'Creating environment'. Add the new zh entries and rename the stale
key in every catalog so the text-catalog test passes.
* fix(run): use system python per OS and pytest templates, detect pixi envs
Gutter/context pytest configs on Windows used `python3`, which is not a
command there; fall back to `python` on Windows. Gutter-created pytest
configs now also inherit "Edit configuration templates... > pytest"
values via a template cache populated with the run configurations.
Pixi environments (`.pixi/envs/<env>`) are detected as the project's
Python: as run-configuration env (`PyEnv::Pixi`, tools via
`pixi run` or the env's own executables), as the project interpreter
for scripts/tests and the debugger, and as selectable SDKs, so the
interpreter picked in the IDE feeds completion.
Fixesl0ng-ai/openidea#1885
* fix(run): drop the pytest template's own target in gutter configs
testConfigFor always appends the gutter file, so a target stored on the
template would run twice (template target plus the file). The gutter file is
the target. Also corrects pixi_env's doc comment to the implemented name
order.
* fix(agent): don't release an idle native agent while background tasks are running
A turn's background tasks (Claude Code's run_in_background commands) outlive
the turn: release_when_idle only checked for a running turn and open
permission requests, so five minutes after the turn ended it stopped the
agent process and orphaned whatever still ran, its result never reaching
the conversation.
Claude Code reports its live background tasks in stream-json mode: a
system/background_tasks_changed event carries the full list on every
change. The driver now translates it into an openidea_background_tasks
session update (like openidea_retry), the task keeps the list, and the
idle release waits while it isn't empty: it checks again after the usual
delay, and when the last task is done its report voids the pending
schedule and starts the delay over from then. A (re)started process
starts with no tasks, so a stale list can never hold the agent up.
Signed-off-by: Inf1mum <[email protected]>
* feat(agents): show the agent's background tasks above the composer
A grey bar above the composer lists what the agent runs in the
background (shells, monitors), with how long each has been running and
a Stop button per task. The task now carries the background tasks the
agent reports, with when each was first seen, and clears them when the
process ends. Stop asks Claude Code to end that one task through a new
_openidea/stop_task request; the composer's Stop keeps ending only the
turn (perTaskStopAffordance).
Closes#1788
---------
Signed-off-by: Inf1mum <[email protected]>
Co-authored-by: Inf1mum <[email protected]>
Co-authored-by: NealimeKenna <[email protected]>
Plugin API 1.3 adds an optional semantic-tokens-provider export (world plugin-semantic-tokens). A plugin declares the semantic-tokens kind under [[contributes.providers]] for any language, including ones it does not contribute, and returns single-line tokens with LSP token kinds and modifiers. The host asks it with the other editor providers (50 ms / 500 ms) when the editor requests semantic tokens and joins the result with the language server's: plugin tokens replace the server tokens they overlap, so the colour scheme styles them like any semantic token. Declaring the kind needs engines.openidea ^1.3.
API 1.2 shipped with 0.1.38, so wit/ and the SDK's copy, ENGINE_VERSION and both API_VERSION constants move to 1.3 and every shipped component is rebuilt. The SDK gains openidea_plugin::semantic_tokens, and register_plugin! accepts any combination of optional exports (providers, views, indexing, syntax, semantic_tokens) instead of a fixed list of pairs.
* fix(clipboard): write file names as the copied files' text flavor (#1878)
Ctrl+C on project-view files now pastes the file names into editors and
chat boxes, not the absolute paths. The file list itself (CF_HDROP /
file URLs) still carries the full paths, so pasting into Explorer or a
terminal that reads files keeps working as before. Paths without a file
name (a drive root) keep their whole path.
* test(clipboard): exercise the no-file-name fallback portably
C:/ is a root only on Windows; the Unix and macOS runners read its
file_name as "C:". Use the real filesystem root, which has no file
name on any platform.
In a Vue CLI project without jsconfig.json, tsserver only knew the files that
were opened: usage counts said "no usages" for code used from unopened .vue
files, and `@/` imports resolved to nothing, so a mixin imported that way made
the component's whole `this` `any` and Go to Declaration failed in the same file.
The tsserver project registered for config-less trees now also serves Vue
(vtsls of vue.toml), includes the files the Vue tsserver plugin handles, and maps
the webpack aliases that @vue/cli-service resolves (else Vue CLI's `@` -> `src`).
Until vtsls has diagnostics for an open file it answers from its syntax server,
so the server reports loading meanwhile and empty usage counts are asked again.
A PHP without the Xdebug extension ignores the -dxdebug.* options php-debug
starts it with, so a PHP debug session ran past every breakpoint and nothing
said why. Starting one now checks the PHP that will run (with XDEBUG_MODE=off,
in the session's folder, where a version manager's php shim picks the PHP) and
warns when Xdebug is not loaded, with Open PHP Settings; the session still
starts. Settings > PHP > CLI Interpreter shows the Xdebug version of the
selected PHP, or Not installed.
Co-authored-by: arookieofc <[email protected]>
Project scope visits hidden files and folders (the search engine's
walk_files documents exactly that), but find-files left every dot-folder
out on both paths: the index filter dropped them and the disk walk ran
with hidden(true). Hidden folders are now visited like anywhere else;
only .git internals stay out.
Fixes#1850
File history passed one ^<sha>^@ exclusion per created file, and path-filtered
log passed every selected path, as git arguments. A large module blew past
Windows' ~32K command line (os error 206). Pass them through git log --stdin.
Following renames only makes sense for a file: on a directory each created file
cut the history short, leaving a single commit. Directories now use plain path
filtering.
Plugin API 1.2 adds `ui.pick`, `ui.pick-many`, `ui.input`, `ui.confirm` and `views.reveal`. The dialogs show in the window whose command or event asked and block that call until the user answers; the wait is added to the call's deadline, so it doesn't use up the time budget. They need no capability and return `invalid` without a window (events without a client, rendering, editor providers). Prompts are withdrawn when the call is cancelled or the plugin stops, and reach remote windows through the dispatcher (`plugin_answer_prompt`). `views.reveal` brings one of the plugin's declared views to the front. The Rust SDK and its mock gain `pick`, `pick_many`, `input`, `confirm` and `reveal_view`. All components are rebuilt for the WIT change.
Closes#1768
Co-authored-by: arnaud-ritti <[email protected]>
Plugin API 1.2 adds `ui.set-context(key, value)` and `ui.remove-context(key)`: a plugin sets `plugin.<id>.<key>` for `when` clauses, typed as `context-value` (flag, number or text), in the window the call came from or every window outside a command or event; a window's value wins over the global one. Keys are 1-64 of `A-Za-z0-9_-`, at most 100 per plugin, kept across window reloads and idle reclaim, dropped when the plugin stops or the window closes; remote windows get them through the dispatcher. The Rust SDK takes `impl Into<ContextValue>`. All components are rebuilt for the WIT change.
Closes#1622
Co-authored-by: arnaud-ritti <[email protected]>
Plugin API 1.2 adds `invocation.selected-paths`: the files and folders a Project view or editor tab menu ran the command on, as absolute paths (the command's argument stays the first of them, so existing plugins see no change). The frontend passes the menu's selection to `plugin_run_command` on the desktop and through the remote dispatcher; the Rust SDK and its mock gain `Invocation::selected_paths`. All components are rebuilt for the WIT change.
Closes#1621
Co-authored-by: arnaud-ritti <[email protected]>
A plugin can declare `[[contributes.menus]]` with `menu = "viewItem.context"` and `view = "<its view>"`. Right-clicking an item of that tree or list view lists the matching commands, each run with the item's key as argument; `when` sees `viewItem` (the key) and `viewItemKind` (the key up to its first `:`). Validation refuses the menu without `view`, with an undeclared or web view, `view` on another menu, and (new in API 1.2) a manifest whose engines.openidea still admits older IDEs.
Closes#1618
Co-authored-by: arnaud-ritti <[email protected]>
A plugin can set `icon` (and `icon-dark`) in plugin.toml. The SVGs go through the existing sanitiser (16 KiB each), ship with the package, and show in Settings › Plugins (list rows and detail header) and the marketplace, whose index entries carry them; the IDE sanitises index icons again and drops unsafe ones.
IDEs before plugin API 1.2 refuse a manifest with `icon`, so Manifest::needs_api refuses one whose engines.openidea still admits them (check and install), keeping the index from offering such a package to older IDEs.
Closes#1732
Co-authored-by: arnaud-ritti <[email protected]>
* feat(trust): let users add trusted folders from settings
The trusted-locations check already matches projects by path prefix, but
the only way to populate it was the "Trust projects in…" checkbox on the
first-open dialog — a folder could never be trusted ahead of time.
Add a project_trust_add_location command that canonicalizes the picked
directory, rejects non-directories, and deduplicates, plus an "Add
folder…" button on the Project Trust settings page backed by the native
directory picker.
Fixes#1743
* fix(trust): translate the trusted-location additions into Chinese
The zh source-text catalog must cover every literal UI text; the three
new strings in the trusted-locations settings page were missing.
m2e resolves and downloads into maven.repo.local, which neither the
userSettings init option nor MAVEN_OPTS covers. Without it the embedded
Maven fell back to ~/.m2 even when every mvn run the IDE launches used
the configured repository, splitting the cache across two locations.
A -Dmaven.repo.local in the user's own jdtls VM options wins.
Fixes#1739
On Linux a watcher creation can fail with a bare OS error when the
kernel's inotify limits are exhausted: one instance per watcher
(EMFILE, os error 24) or one watch per directory (ENOSPC, mapped by
notify to MaxFilesWatch). Name the sysctl knob to raise in the error
message instead of leaving users to search engines (issue #788).
Delve compiles the debuggee to __debug_bin in the session's cwd (on
Windows __debug_bin.exe plus __debug_bin.exe<digits> in DAP mode) and
removes it on a graceful exit. When shutdown doesn't reach that cleanup
the binaries pile up in the project directory.
After the session is down — the adapter disconnected and its process
reaped — remove the leftover files matching Delve's own names from the
session's cwd. Only a locally started adapter (no debugServer) is swept:
a remote Delve's cwd may be on another machine. Files that merely share
the prefix (__debug_bin.go sources, in-progress names with other
suffixes, directories) are left alone.
Fixes#1790
* fix(vfs): tolerate fsync ENOTTY/EINVAL on filesystems without fsync
virtio-fs (macOS VMs) answers fsync/F_FULLFSYNC with "Inappropriate
ioctl for device" (os error 25), failing the whole atomic write. The
write itself already went through, so degrade like SQLite: log and
continue to the rename instead of propagating.
* fix(vcs): tolerate filesystems without fsync when updating ignore files
* fix(http): join multipart request bodies with CRLF
Multipart parts are delimited by CRLF per RFC 2046; prepare() joined
body lines with LF, so strict servers (Tomcat) never found the boundary
and rejected the request as 413. Join with CRLF when the request's
Content-Type is multipart/*; other bodies keep LF.
Fixes#1700
* ci: retrigger after flaky mybatis test on windows
* ci: retrigger after flaky java debug smoke
Opening a workspace with nested repositories, or a large repository on Linux, could stall open_project for minutes. GitignoreStack cached nested repositories per parent directory and cleared that cache on every created or removed path, so the startup walk rediscovered and recompiled the same nested repository's ignore rules for each directory; the per-directory watch walk also ran to completion on the open_project thread.
Nested stacks are now cached by repository root and invalidated only under the changed path (a `.git` appearing or disappearing included). The startup walk registers up to 256 directories synchronously and hands the rest to the watcher thread; children of a verified directory are matched by their own name only.
Tests cover the deferred startup watches and a repository created inside the workspace becoming an ignore boundary.
Co-authored-by: dengliu <[email protected]>
l0ng-ai