Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

Start the first agent from a chosen project folder

dee9b7bff997904
diff --git a/desktop/lib/screens/swarm_screen.dart b/desktop/lib/screens/swarm_screen.dart
index 029d9a1ad..89c0af307 100644
--- a/desktop/lib/screens/swarm_screen.dart
+++ b/desktop/lib/screens/swarm_screen.dart
@@ -2,6 +2,7 @@ import 'dart:async';
import 'dart:convert';
import 'dart:io' show Platform;
+import 'package:file_selector/file_selector.dart';
import 'package:flutter/material.dart';
import 'package:flutter/services.dart';
@@ -490,6 +491,7 @@ class _SwarmScreenState extends State<SwarmScreen> {
String? folder,
String? swarmId,
PaneSplitRequest? split,
+ bool chooseFolderFirst = false,
}) => _dialog(() async {
final local = app.machineStates.values
.where((m) => m.isLocalMachine)
@@ -502,13 +504,44 @@ class _SwarmScreenState extends State<SwarmScreen> {
await showSwarmLinkDialog(context, app);
return;
}
+ final targetId = swarmId ?? app.activeSwarmId;
+ final machine = app.machineStates[id];
+ var selectedFolder = folder;
+ Future<void>? initialEngineProbe;
+ if (chooseFolderFirst && machine != null && machine.isLocalMachine) {
+ // Read availability while the user chooses a folder, so the form can
+ // prefer an installed agent without adding a second probe or wait.
+ initialEngineProbe = app.probeEngines(id, force: true);
+ try {
+ selectedFolder = await getDirectoryPath(
+ confirmButtonText: 'Use folder',
+ );
+ } catch (error) {
+ if (mounted) {
+ ScaffoldMessenger.of(context).showSnackBar(
+ SnackBar(content: Text('Could not open the folder picker: $error')),
+ );
+ }
+ return;
+ }
+ if (!mounted ||
+ selectedFolder == null ||
+ app.activeSwarmId != targetId ||
+ app.machineStates[id] != machine ||
+ !machine.isLocalMachine ||
+ machine.nodeOnline == false ||
+ machine.needsLink) {
+ return;
+ }
+ }
await showNewAgentDialog(
context,
app,
id,
- source: 'swarm',
- initialFolder: folder,
- swarmId: swarmId,
+ source: chooseFolderFirst ? 'first_folder' : 'swarm',
+ initialFolder: selectedFolder,
+ initialEngineProbe: initialEngineProbe,
+ swarmId: targetId,
split: split,
);
});
@@ -1150,6 +1183,8 @@ class _SwarmScreenState extends State<SwarmScreen> {
notifier: app,
projects: _projects.projects,
onNewAgent: _newAgent,
+ onChooseFirstFolder: () =>
+ _newAgent(chooseFolderFirst: true),
onAgent: (entry) => _activateSearch(
SwarmSearchSelection(
SwarmDestination(
diff --git a/desktop/lib/widgets/new_agent_dialog.dart b/desktop/lib/widgets/new_agent_dialog.dart
index fc386d188..aae765326 100644
--- a/desktop/lib/widgets/new_agent_dialog.dart
+++ b/desktop/lib/widgets/new_agent_dialog.dart
@@ -43,6 +43,7 @@ Future<void> showNewAgentDialog(
String? initialFolder,
String? swarmId,
PaneSplitRequest? split,
+ Future<void>? initialEngineProbe,
}) {
// Reported here rather than at each call site: the doors are four and
// growing, and one that forgets to track is a hole in the funnel that only
@@ -59,6 +60,7 @@ Future<void> showNewAgentDialog(
initialFolder: initialFolder,
swarmId: swarmId ?? notifier.activeSwarmId,
split: split,
+ initialEngineProbe: initialEngineProbe,
),
);
}
@@ -69,6 +71,7 @@ class _NewAgentDialog extends StatefulWidget {
final String? initialFolder;
final String swarmId;
final PaneSplitRequest? split;
+ final Future<void>? initialEngineProbe;
const _NewAgentDialog({
required this.notifier,
@@ -76,6 +79,7 @@ class _NewAgentDialog extends StatefulWidget {
this.initialFolder,
required this.swarmId,
this.split,
+ this.initialEngineProbe,
});
@override
@@ -123,7 +127,7 @@ class _NewAgentDialogState extends State<_NewAgentDialog> {
// answer lands, so nothing blanks.
WidgetsBinding.instance.addPostFrameCallback((_) {
if (mounted) {
- unawaited(_probeEngines());
+ unawaited(_probeEngines(initialProbe: widget.initialEngineProbe));
}
});
}
@@ -140,10 +144,11 @@ class _NewAgentDialogState extends State<_NewAgentDialog> {
_engine;
}
- Future<void> _probeEngines() async {
+ Future<void> _probeEngines({Future<void>? initialProbe}) async {
final machineId = _machineId;
final revision = _machineRevision;
- await widget.notifier.probeEngines(machineId, force: true);
+ await (initialProbe ??
+ widget.notifier.probeEngines(machineId, force: true));
if (!mounted ||
revision != _machineRevision ||
_submitting ||
diff --git a/desktop/lib/widgets/swarm_welcome.dart b/desktop/lib/widgets/swarm_welcome.dart
index b05e18114..59b053872 100644
--- a/desktop/lib/widgets/swarm_welcome.dart
+++ b/desktop/lib/widgets/swarm_welcome.dart
@@ -20,12 +20,14 @@ class SwarmWelcome extends StatelessWidget {
required this.onLinkMachine,
required this.onMachine,
required this.onProject,
+ this.onChooseFirstFolder,
this.onAgent,
this.onProjectAgents,
});
final AppNotifier notifier;
final List<SavedSwarmProject> projects;
final VoidCallback onNewAgent;
+ final VoidCallback? onChooseFirstFolder;
final Widget searchField;
final VoidCallback onAddProject;
final VoidCallback onLinkMachine;
@@ -53,6 +55,7 @@ class SwarmWelcome extends StatelessWidget {
AgentLoadStatus.loaded &&
app.machineStates.values.first.nodeOnline != false &&
!app.machineStates.values.first.needsLink;
+ final chooseFirstFolder = noAgents && onChooseFirstFolder != null;
final readyAgents =
agents
.where(
@@ -113,9 +116,20 @@ class SwarmWelcome extends StatelessWidget {
Expanded(child: searchField),
const SizedBox(width: 12),
FilledButton.icon(
- onPressed: onNewAgent,
- icon: const Icon(Icons.add, size: 17),
- label: const Text('New agent'),
+ onPressed: chooseFirstFolder
+ ? onChooseFirstFolder
+ : onNewAgent,
+ icon: Icon(
+ chooseFirstFolder
+ ? Icons.folder_open_outlined
+ : Icons.add,
+ size: 17,
+ ),
+ label: Text(
+ chooseFirstFolder
+ ? 'Choose folder…'
+ : 'New agent',
+ ),
style: FilledButton.styleFrom(
minimumSize: const Size(126, 44),
backgroundColor: grid.AppPalette.swarmAccent,
diff --git a/desktop/test/first_workspace_test.dart b/desktop/test/first_workspace_test.dart
index c2d936f7f..fd7752a07 100644
--- a/desktop/test/first_workspace_test.dart
+++ b/desktop/test/first_workspace_test.dart
@@ -15,6 +15,7 @@ import 'package:harness/terminal/terminal_binary.dart';
import 'package:xterm/xterm.dart';
import 'swarm_screen_test.dart' show mount, terminal;
+import 'swarm_interactions_test.dart' show chord;
class _FirstUseApp extends AppNotifier {
_FirstUseApp()
@@ -37,6 +38,7 @@ class _FirstUseApp extends AppNotifier {
}
Completer<void>? probe;
+ var probes = 0;
Completer<String?>? creation;
final launches =
<({String machine, String engine, String folder, bool bypass})>[];
@@ -44,6 +46,7 @@ class _FirstUseApp extends AppNotifier {
@override
Future<void> probeEngines(String machineId, {bool force = false}) async {
+ probes++;
await probe?.future;
}
@@ -99,7 +102,7 @@ void main() {
FileSelectorPlatform.instance = _FolderPicker();
addTearDown(() => FileSelectorPlatform.instance = oldPicker);
await mount(tester, app);
- await tester.tap(find.text('New agent'));
+ await chord(tester, LogicalKeyboardKey.keyN);
await tester.pump();
await tester.tap(find.text('Browse…'));
await tester.pump();
@@ -164,7 +167,7 @@ void main() {
FileSelectorPlatform.instance = picker;
addTearDown(() => FileSelectorPlatform.instance = oldPicker);
await mount(tester, app);
- await tester.tap(find.text('New agent'));
+ await chord(tester, LogicalKeyboardKey.keyN);
await tester.pump();
bool fieldFocused(String key) => tester
@@ -251,7 +254,7 @@ void main() {
expect(find.text('Machines'), findsNothing);
expect(find.text('Projects'), findsNothing);
expect(app.launches, isEmpty);
- await tester.tap(find.text('New agent'));
+ await tester.tap(find.text('Choose folder…'));
await tester.pump();
final engine = tester.widget<AppSelectField<String>>(
find.byKey(const Key('new-agent-engine-field')),
@@ -265,9 +268,9 @@ void main() {
tester.getTopLeft(find.text('Project folder')).dy,
lessThan(tester.getTopLeft(find.text('Coding agent')).dy),
);
- expect(picker.opened, 0);
- await tester.tap(find.text('Browse…'));
- await tester.pump();
+ expect(picker.opened, 1);
+ expect(app.probes, 1);
+ expect(find.text('/work/my-project'), findsOneWidget);
expect(app.launches, isEmpty);
await tester.tap(find.text('Create agent'));
await tester.pump();
@@ -312,13 +315,109 @@ void main() {
app.dispose();
});
+ testWidgets(
+ 'first-folder selection reuses discovery and ignores repeated activation',
+ (tester) async {
+ final app = _FirstUseApp()..probe = Completer<void>();
+ final oldPicker = FileSelectorPlatform.instance;
+ final picker = _FolderPicker()..pending = Completer<String?>();
+ FileSelectorPlatform.instance = picker;
+ addTearDown(() => FileSelectorPlatform.instance = oldPicker);
+ await mount(tester, app);
+
+ await tester.tap(find.text('Choose folder…'));
+ await tester.pump();
+ await tester.tap(find.text('Choose folder…'));
+ await tester.pump();
+ expect(picker.opened, 1);
+ expect(app.probes, 1);
+ expect(find.byType(AlertDialog), findsNothing);
+ expect(app.launches, isEmpty);
+
+ app.machineStates['m']!.engines.replace(const [
+ EngineAvailability(engine: 'claude', installed: false),
+ EngineAvailability(engine: 'codex', installed: true),
+ ]);
+ app.probe!.complete();
+ await tester.pump();
+ picker.pending!.complete('/work/chosen');
+ await tester.pump();
+ await tester.pump();
+ expect(find.byType(AlertDialog), findsOneWidget);
+ expect(find.text('/work/chosen'), findsOneWidget);
+ expect(
+ tester
+ .widget<AppSelectField<String>>(
+ find.byKey(const Key('new-agent-engine-field')),
+ )
+ .value,
+ 'codex',
+ );
+ expect(app.probes, 1);
+ expect(app.launches, isEmpty);
+ expect(app.input, isEmpty);
+ await tester.pumpWidget(const SizedBox());
+ app.dispose();
+ },
+ );
+
+ for (final change in [
+ 'cancel',
+ 'swarm',
+ 'machine',
+ 'remote',
+ 'offline',
+ 'link',
+ ]) {
+ testWidgets('first-folder result is discarded after $change', (
+ tester,
+ ) async {
+ final app = _FirstUseApp();
+ final oldPicker = FileSelectorPlatform.instance;
+ final picker = _FolderPicker()..pending = Completer<String?>();
+ FileSelectorPlatform.instance = picker;
+ addTearDown(() => FileSelectorPlatform.instance = oldPicker);
+ await mount(tester, app);
+ await tester.tap(find.text('Choose folder…'));
+ await tester.pump();
+ switch (change) {
+ case 'swarm':
+ app.newSwarm();
+ case 'machine':
+ app.machineStates['m'] = MachineState(app.machines.single)
+ ..localOnly = true;
+ case 'remote':
+ app.machineStates['m']!.localOnly = false;
+ case 'offline':
+ app.machineStates['m']!.nodeOnline = false;
+ case 'link':
+ app.machineStates['m']!.needsLink = true;
+ }
+ picker.pending!.complete(change == 'cancel' ? null : '/work/stale');
+ await tester.pump();
+ await tester.pump();
+ expect(find.byType(AlertDialog), findsNothing);
+ expect(app.allPanes, isEmpty);
+ expect(app.launches, isEmpty);
+ expect(app.input, isEmpty);
+ // Dismissing the native chooser also releases the ordinary New agent
+ // command; no abandoned pending flag should trap the workspace.
+ await chord(tester, LogicalKeyboardKey.keyN);
+ expect(find.byType(AlertDialog), findsOneWidget);
+ expect(find.text('/work/stale'), findsNothing);
+ expect(tester.takeException(), isNull);
+ await tester.pumpWidget(const SizedBox());
+ app.dispose();
+ });
+ }
+
testWidgets(
'a late probe uses an installed agent without overriding a choice',
(tester) async {
for (final chooseExplicitly in [false, true]) {
final app = _FirstUseApp()..probe = Completer<void>();
await mount(tester, app);
- await tester.tap(find.text('New agent'));
+ await chord(tester, LogicalKeyboardKey.keyN);
await tester.pump();
if (chooseExplicitly) {
await tester.tap(find.byKey(const Key('new-agent-engine-field')));
@@ -332,6 +431,8 @@ void main() {
]);
app.probe!.complete();
await tester.pump();
+ await tester.pump();
+ expect(app.probes, 1);
final engine = tester.widget<AppSelectField<String>>(
find.byKey(const Key('new-agent-engine-field')),
);
diff --git a/docs/harness-v2-appearance-onboarding.md b/docs/harness-v2-appearance-onboarding.md
index 08b73088a..5a4c8b32a 100644
--- a/docs/harness-v2-appearance-onboarding.md
+++ b/docs/harness-v2-appearance-onboarding.md
@@ -52,3 +52,7 @@ The audit reproduced a pending-launch bug: Escape dismissed New agent despite it
Known missing/invalid-folder failures now name the target machine and ask the user to choose another folder. Missing tmux gets its required recovery instead of a wire code. Other server details remain visible. A request timeout is described as an unconfirmed outcome and directs the user to Search before creating another agent; it is never automatically retried. Protocol-level creation idempotency and late-reply reconciliation remain separate future reliability work.
Validation: the new pending-launch regression failed before the fix and passed afterward. Forty-six focused onboarding/profile/split/async checks passed, plus isolated 880×560 renders of pending, invalid-folder and timeout states. These checks use synthetic state and requests; they do not install software, launch a provider, or send a task to a real agent.
+
+The next pass removes one intermediate click for a known empty local workspace. **Choose folder…** opens the native directory chooser directly. The ordinary form then shows that folder and prefers an installed coding agent. Availability probing runs while the user chooses, and the form reuses the same probe. Cmd-N and the returning-user New agent action keep their existing behavior. This is our adaptation of the folder-first flow, not a measured claim about user activation.
+
+The chooser is opened only by deliberate activation. Duplicate activation, cancellation, changed target identity and offline/unlinked state are checked before showing its result. Sixty-five focused onboarding/profile/engine/split/search checks pass, including completion with no automatic agent input, keyboard-only ordinary creation, late availability, cancellation and stale-target cases. Isolated 880×560 welcome/form renders were reviewed. The native OS chooser itself was stubbed; live first-install and provider sign-in remain separate validation.
diff --git a/docs/harness-v2-developer-tools-research.md b/docs/harness-v2-developer-tools-research.md
index 49f90892a..8838d4b38 100644
--- a/docs/harness-v2-developer-tools-research.md
+++ b/docs/harness-v2-developer-tools-research.md
@@ -21,7 +21,8 @@ Updated 2026-09-13 after user review. Keep the existing simple Command shortcuts
| Rebuilt; live search/project review passed | One top-right Cmd-P search for agents, Swarms, machines and projects, with Open and explicit Add here actions. Follow-up in validation: New swarm searches in place, and machine/project results open their swarm directly. Separate Add picker, floating pane + and titlebar bell removed. New swarm uses only the selected mountain/lake wallpaper; rotation and other assets removed. Native typing, Cmd-P select-all and Escape were verified from an existing empty New swarm after the focus correction. |
| Implemented; measured in isolation | Unified search construction and ranking avoid repeated scans/metadata work. Median catalog construction fell 50% in the 2,059-destination debug fixture; native input-to-display measurement remains separate. |
| Implemented and regression checked | [A coherent agent-workspace key system with file-first remapping](harness-v2-keyboard-system.md). The optional config loads at startup and reloads on save. Workspace and picker dispatch, native menus, search actions and shortcut help share effective bindings; invalid edits keep the last working map. Existing simple defaults remain unchanged. See the progress log for build status and validation. |
-| Calibration in progress; foreground runs paused | Native Release input/navigation fixture with synthetic terminals. No validated latency result yet; resume after the keyboard work stabilizes. Live keyboard/tab/drag and real remote interruption checks remain. |
+| Calibration paused on native focus | A fresh isolated Release fixture built, but macOS reported its visible window inactive and not key. No latency samples were accepted. Resume only after new foreground-access evidence; retain the guard and keep real agents out of timing tests. Live keyboard/tab/drag and real remote interruption checks remain. |
+| First-folder path implemented and regression checked | The empty local workspace directly opens the folder chooser, then a prefilled New agent form. Installed-agent discovery overlaps the chooser. Cancellation/stale targets are guarded, and creation remains explicit. Existing agents still open directly; returning users and Cmd-N keep the normal form. [Onboarding notes](harness-v2-appearance-onboarding.md) record recovery, source research and validation limits. |
| Resizing, splits, commands and preview implemented | `>` commands work in both search fields. Controlled resizing uses the existing gaps, saves proportions, protects usable sizes and retains terminal elements. Explicit right/down splits use the normal New agent form and preserve the surrounding rectangles. Cmd-S returns to a preset. Cmd-I or the eye control enables a compact snapshot of already-retained output for the selected agent; preview starts hidden and performs no network or terminal attachment. Both search fields now emphasize matching names and visible context, using the same field preference as ranking and preserving Unicode graphemes; no permanent control is added. The [progress log](harness-v2-progress.md) records validation and the close-layout fallback. [The six-prototype review](harness-v2-prototype-review.md) records what to borrow and what stays out. |
| Keep deferred | Workflow canvases, autonomous manager layers, a full IDE/Git client and a marketplace. |
diff --git a/docs/harness-v2-performance.md b/docs/harness-v2-performance.md
index 64809eff9..5fc6f5d11 100644
--- a/docs/harness-v2-performance.md
+++ b/docs/harness-v2-performance.md
@@ -10,6 +10,14 @@ flutter test --no-pub --reporter expanded test/benchmarks/swarm_benchmark.dart
It prints `SWARM_BENCH` JSON records. Its filename deliberately does not end in `_test.dart`, so ordinary correctness runs do not include timing measurements.
+## Native calibration remains unmeasured (2026-09-13)
+
+A fresh disposable Release copy of the existing native benchmark built at `/private/tmp/harness-native-benchmark-4t74nfhn`. Its separate bundle identity, temporary store, blocked networking and synthetic retained terminals isolate it from real agents. AppKit events are posted only to that fixture's own queue. The planned metric joins each observed input to Flutter's matching frame number and raster-finish timestamp; it excludes network and physical display latency. [Flutter's FrameTiming reference](https://api.flutter.dev/flutter/dart-ui/FrameTiming-class.html) describes those fields and recommends profile/release performance collection. The distinction between input latency under load and throughput follows the measurement discussion in [Dan Luu's terminal study](https://danluu.com/term-latency/); its old terminal rankings are not treated as current comparisons.
+
+The first calibration stopped before samples because content focus could not be established. Inspection found that its driver requested focus on the wrapper view; the production titlebar correctly uses the Flutter view controller. Only the disposable copy was corrected to match that responder and to report active/key state separately. The rebuilt calibration then stopped with **key=false, active=false, visible=true**. Both runs exited without accepted timings. The guards were retained, the original untracked benchmark source was left unchanged, and the real preview reopened after the fixture exited.
+
+Artifacts: `/private/tmp/harness-v2-native-prepare-current.log`, `/private/tmp/harness-native-benchmark-4t74nfhn/{build,rebuild-controller}.log`, `/private/tmp/harness-v2-native-calibration-20260913{.json,.log,-run.log}` and `/private/tmp/harness-v2-native-calibration-controller{.json,.log,-run.log}`. Do not retry the calibration until new evidence establishes that the disposable native window can become active/key. These failures provide no p50/p95/p99 or input-to-display result.
+
## Search match emphasis (2026-09-13)
Highlighting runs only for built result rows, sharing ranking's field scores and rune-based subsequence matching. The visible labels preserve grapheme boundaries and do not change ranking. A label longer than 1,024 UTF-16 units stays plain; up to twelve distinct terms from the first twelve query terms are considered, each at most 128 units, against fields at most 4,096 units. Empty queries stay on the plain-text path.
diff --git a/docs/harness-v2-progress.md b/docs/harness-v2-progress.md
index 1d215f8c6..18edcf35a 100644
--- a/docs/harness-v2-progress.md
+++ b/docs/harness-v2-progress.md
@@ -2,6 +2,15 @@
Updated 2026-09-13 with coordinated palettes, Models, combined closed-work History and faster unified search. This is a working preview, not a release.
+## First-folder checkpoint
+
+- A genuinely empty, discovered local workspace now offers **Choose folder…**. That action opens the native folder chooser directly, then the ordinary New agent form with the selected folder. Installed-agent discovery overlaps folder selection, and the form reuses that request. The returning-user button and Cmd-N still open New agent normally.
+- Repeated activation cannot stack native choosers. Cancellation, a changed swarm/machine, loss of local identity, disconnection or a newly required link discards the late result. No agent starts until Create agent is selected; no task is sent automatically.
+- Validation: **65 focused onboarding, engine/profile, split and inline-search checks passed**, followed by an isolated visual render of the welcome and prefilled form at 880×560. Analyzer has zero errors/warnings and the existing 12 vendored infos. Artifacts: `/private/tmp/harness-v2-first-folder-{tests,render,analyze}.log` and `/private/tmp/harness-v2-first-folder-{welcome,form}.png`.
+- The preceding search-highlighting and first-agent-recovery changes were rebuilt successfully (`/private/tmp/harness-v2-search-onboarding-build.log`). The first-folder change is awaiting its own rebuild. Live capture is still unavailable; visual evidence here is from isolated renders.
+- Native latency calibration is paused: the disposable Release benchmark window was visible, but macOS reported it inactive and not key. No timings were accepted. See [performance notes](harness-v2-performance.md) for the failed calibration artifacts and restart condition. The real preview was reopened after the fixture exited.
+- Continue local commits. Public-origin approval remains pending; do not retry pushing or include the unrelated naming edits, collaboration draft or native benchmark directory in this checkpoint.
+
## First-agent recovery checkpoint
- A reproduced Escape/backdrop dismissal during agent creation now keeps the form and its result visible. Pending creation has a readable progress label; controls return after a response. Failures preserve the user's choices, and a new folder/engine choice clears the obsolete error.
@@ -410,12 +419,12 @@ Earlier logs contain superseded failures. Temporary logs and toolchains are loca
## Remaining work
-1. Measure native terminal input and tab-switch responsiveness under output load, including p50/p95/p99 and cold/warm paths. Live Cmd+P destination focus/quick return, branch headers, project associations, New swarm cleanup and native History are verified. Keep typing/transport measurements in disposable fixtures; the user's real agent terminals are not test inputs.
+1. Measure native terminal input and tab-switch responsiveness under output load, including p50/p95/p99 and cold/warm paths. Calibration is paused until there is new evidence that the isolated window can become active/key; do not bypass its focus guard. Earlier live Cmd+P destination focus/quick return, branch headers, project associations, New swarm cleanup and native History were verified. Keep typing/transport measurements in disposable fixtures; the user's real agent terminals are not test inputs.
2. Audit native drag/reorder, overflow, close-last-tab, renaming and accessibility without changing the user's saved agent memberships. Keep any integration runner separate and out of the foreground review app.
3. Verify real remote reconnect, agent input, paste, selection and IME against disposable local/remote processes. The AppNotifier/keyframe and stale completion regressions now pass; they do not prove every network or native editing condition. Preserve immediate first-input flush and shared retained renderers.
4. Visually confirm app-menu modal behavior and overflow accessibility in AppKit; route/shortcut behavior is covered by passing Flutter tests.
5. Build/review Linux and Windows when their toolchains are available. Remote full project/branch metadata requires daemons running the new wire format; do not upgrade them automatically.
-6. Follow the bounded prototype recommendation: optional retained-context preview/highlights in Cmd+P, then deliberate right/down splits with compatible layout persistence. Integrate remappable keys and existing-command search; ordinary-shell creation is outside the current agent-workspace scope. Keep the rejected shelf/hierarchy/dashboard/manager surfaces absent.
+6. Optional retained-context preview/highlights, deliberate right/down splits, controlled resizing, file-based keyboard customization and existing-command search are implemented and regression checked. Continue targeted real-workflow verification; ordinary-shell creation is outside the current agent-workspace scope. Keep the rejected shelf/hierarchy/dashboard/manager surfaces absent.
7. Audit local Git watch failure/removal and relocated-worktree invalidation before broadening that compatibility path. It deliberately reads a small subset of Git metadata; prefer richer daemon metadata as available. Continue the active goal using the current research status list, distinguishing shipped reductions from proposed follow-ups.
## Toolchain and workflow