flesh out the Claude prompt UX: options, stepper, collapse

Builds on the in-composer prompt surface (D-78):

- Permission prompts (T-175): Allow / Allow-and-don't-ask-again / Deny.
  "Don't ask again" appears only when the request carries a
  permission_suggestion and echoes it back as updatedPermissions. An
  optional note rides Deny as the message, or Allow as a follow-up user
  message (the protocol has no allow-with-message).
- AskUserQuestion picker (T-176): a single question renders bare; 2-4
  questions step one at a time (nav shows "N · Header", ✓ when answered)
  then a review/confirm screen. Each question offers an "Other" free-text
  choice and a per-choice note; multi-select joins labels. A "chat
  instead" escape denies the prompt so the user can type freely. On
  submit the answer is echoed into the log, since the card is ephemeral.
- Collapsed tool cards (T-177): multi-line tool_use / tool_result start
  collapsed behind a one-line summary; one-line output renders inline.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-05-25 09:30:50 +02:00
co-authored by Claude Opus 4.7
parent fbf5a720a8
commit de264e7141
8 changed files with 324 additions and 16 deletions
+10 -5
View File
@@ -18,14 +18,19 @@ heading, and (b) bumping `pubspec.yaml` `version:` in the same commit.
### Added
- Native permission & AskUserQuestion prompts (T-166, D-78) — when Claude
needs tool approval or asks a question, an inline card appears in the
conversation with Allow/Deny or selectable options; the decision is
returned over the stream-json control channel. Closes the prompt gap
the tmux model couldn't surface.
- Native permission & AskUserQuestion prompts (T-166, T-175, T-176, D-78) —
when Claude needs tool approval or asks a question, the composer is
replaced by a prompt: Allow / Allow-and-don't-ask-again / Deny for
permissions (with an optional note), and a single-question or
stepped-with-review option picker for AskUserQuestion (with "Other"
free-text and per-choice notes). Closes the prompt gap the tmux model
couldn't surface.
- Conversation message cards (T-173) — every turn in the Claude pane now
renders through one card template with a copy button on hover and a
collapse/expand caret for tool calls, results, and thinking.
- Collapsed-by-default tool cards (T-177) — multi-line tool calls and
results start collapsed behind a one-line summary; one-line output stays
inline so a caret never hides a single line.
- Claude meta sidebar (T-141, T-157) — an always-pickable left-panel tab
showing Claude activity (the latest day's messages/sessions/tool-calls
plus lifetime totals, from `stats-cache.json`) and, when a tmux team is
+16 -1
View File
@@ -37,6 +37,7 @@ class ConversationCard extends StatefulWidget {
this.actions = const [],
this.collapsible = false,
this.collapsedByDefault = false,
this.collapsedSummary,
this.borderColor,
});
@@ -54,6 +55,11 @@ class ConversationCard extends StatefulWidget {
final bool collapsible;
final bool collapsedByDefault;
/// One-line gist shown next to the label while collapsed (e.g. the tool's
/// key arg, or a result's first line), so a collapsed card still says what
/// it holds. Null → just the label.
final String? collapsedSummary;
/// Border colour for the bordered variant (e.g. error red); defaults to the
/// panel border.
final Color? borderColor;
@@ -127,11 +133,20 @@ class _ConversationCardState extends State<ConversationCard> {
}
Widget _header(SurfaceTokens tokens) {
final summary = widget.collapsedSummary;
return Row(
children: [
if (widget.collapsible) _caret(tokens),
ClideText(widget.label, fontSize: clideFontSmall, color: widget.accent, fontFamily: clideMonoFamily),
const Spacer(),
// While collapsed, show a one-line gist next to the label so the card
// still says what it holds.
if (_collapsed && summary != null) ...[
const SizedBox(width: 10),
Expanded(
child: ClideText(summary, fontSize: clideFontMeta, color: tokens.globalTextMuted, fontFamily: clideMonoFamily, maxLines: 1),
),
] else
const Spacer(),
// Hover-revealed actions. (Always-reachable keyboard a11y for these is
// a follow-up detail; the collapse caret above is always visible.)
if (_hover) ..._actions(tokens),
+25 -3
View File
@@ -149,26 +149,34 @@ class _ConversationTurn extends StatelessWidget {
Widget _toolUse(AssistantToolUse t) {
final pretty = const JsonEncoder.withIndent(' ').convert(t.input);
// Collapse only the bulky multi-line form; a trivial one-liner just shows.
final multiline = pretty.contains('\n');
return ConversationCard(
variant: ConversationCardVariant.bordered,
accent: tokens.globalFocus,
label: t.name,
copyText: pretty,
collapsible: true,
collapsible: multiline,
collapsedByDefault: multiline,
collapsedSummary: multiline ? _toolUseSummary(t) : null,
body: ClideCodeBlock(source: pretty, language: 'json'),
);
}
Widget _toolResult(ToolResultMessage t) {
final accent = t.isError ? tokens.statusError : tokens.globalTextMuted;
// A one-line result is all chrome to collapse — show it inline. Only fold
// away multi-line output, behind a summary of its first line.
final multiline = t.content.contains('\n');
return ConversationCard(
variant: ConversationCardVariant.bordered,
accent: accent,
borderColor: t.isError ? tokens.statusError : tokens.panelBorder,
label: t.isError ? 'error' : 'result',
copyText: t.content,
collapsible: true,
collapsedByDefault: true,
collapsible: multiline,
collapsedByDefault: multiline,
collapsedSummary: multiline ? _firstLine(t.content) : null,
body: ClideText(
t.content,
fontSize: clideFontMeta,
@@ -177,4 +185,18 @@ class _ConversationTurn extends StatelessWidget {
),
);
}
/// A compact one-liner for a collapsed tool-use card: the most telling arg.
String _toolUseSummary(AssistantToolUse t) {
final input = t.input;
final key =
input['file_path'] ?? input['command'] ?? input['path'] ?? input['pattern'] ?? input['url'] ?? (input.values.isNotEmpty ? input.values.first : null);
final s = key?.toString().replaceAll('\n', ' ') ?? '';
return s.length > 80 ? '${s.substring(0, 80)}' : s;
}
String _firstLine(String content) {
final line = content.split('\n').first.trim();
return line.length > 80 ? '${line.substring(0, 80)}' : line;
}
}
Binary file not shown.
@@ -91,6 +91,7 @@ class ToolPrompt {
required this.input,
this.description,
this.toolUseId = '',
this.permissionSuggestions = const [],
});
/// The control_request `request_id` — the key passed to [StreamJsonSession.resolvePrompt].
@@ -111,6 +112,11 @@ class ToolPrompt {
/// The tool's proposed input — echoed back (possibly modified) on allow.
final Map<String, dynamic> input;
/// Permission-rule suggestions from the request (e.g. a `setMode` /
/// `localSettings` entry). Non-empty → an "allow & don't ask again" path is
/// available; echo a chosen entry back as `updatedPermissions` (D-78).
final List<dynamic> permissionSuggestions;
/// AskUserQuestion is answered through the same channel (D-78).
bool get isQuestion => toolName == 'AskUserQuestion';
}
@@ -124,11 +130,22 @@ sealed class ToolDecision {
/// Allow the tool. [updatedInput] is REQUIRED by the protocol — pass the
/// request's input unchanged to allow as-is, or modified to alter the call.
/// For AskUserQuestion, include the `answers` map (question text → label).
///
/// [updatedPermissions] echoes a permission suggestion back to skip future
/// prompts ("don't ask again"). [followUpNote] is NOT part of the protocol —
/// the protocol has no allow-with-message — so the session sends it as a
/// separate user message right after allowing (D-78).
final class AllowTool extends ToolDecision {
const AllowTool(this.updatedInput);
const AllowTool(this.updatedInput, {this.updatedPermissions, this.followUpNote});
final Map<String, dynamic> updatedInput;
final List<dynamic>? updatedPermissions;
final String? followUpNote;
@override
Map<String, dynamic> toJson() => {'behavior': 'allow', 'updatedInput': updatedInput};
Map<String, dynamic> toJson() => {
'behavior': 'allow',
'updatedInput': updatedInput,
if (updatedPermissions != null && updatedPermissions!.isNotEmpty) 'updatedPermissions': updatedPermissions,
};
}
/// Deny the tool with a user-facing [message] (required by the protocol).
@@ -218,6 +235,7 @@ class StreamJsonSession {
description: request['description'] as String?,
toolUseId: request['tool_use_id'] as String? ?? '',
input: input,
permissionSuggestions: (request['permission_suggestions'] as List?) ?? const [],
));
_pendingCtl.add(pendingPrompt);
return; // awaits resolvePrompt
@@ -232,13 +250,27 @@ class StreamJsonSession {
/// [ToolPrompt.promptId]. No-op if unknown or already resolved. Advances the
/// queue so the next pending prompt (if any) surfaces.
void resolvePrompt(String promptId, ToolDecision decision) {
final before = _queue.length;
_queue.removeWhere((p) => p.promptId == promptId);
if (_queue.length == before) return; // unknown / already resolved
final idx = _queue.indexWhere((p) => p.promptId == promptId);
if (idx < 0) return; // unknown / already resolved
final prompt = _queue.removeAt(idx);
_proc.writeLine(jsonEncode({
'type': 'control_response',
'response': {'subtype': 'success', 'request_id': promptId, 'response': decision.toJson()},
}));
if (decision is AllowTool) {
// The prompt card is ephemeral (it vanishes once resolved), so leave a
// compact record of an answered question in the conversation log (D-78).
if (prompt.isQuestion) {
final answers = decision.updatedInput['answers'];
if (answers is Map && answers.isNotEmpty) {
final summary = answers.entries.map((e) => '${e.key}${e.value}').join('; ');
_items.add(UserMessage(uuid: 'local-${_localSeq++}', timestamp: DateTime.now(), isSidechain: false, text: '✓ answered: $summary'));
}
}
// The protocol has no allow-with-message, so an allow note rides as a
// follow-up user message right after the approval (D-78).
if (decision.followUpNote?.trim().isNotEmpty ?? false) send(decision.followUpNote!.trim());
}
_pendingCtl.add(pendingPrompt);
}
@@ -170,6 +170,24 @@ void main() {
expect(find.text('error'), findsOneWidget);
});
testWidgets('a one-line tool result renders inline (no collapse caret)', (tester) async {
await pumpWith(tester, [_result('hello-from-spike')]);
expect(find.text('hello-from-spike'), findsOneWidget);
expect(find.byType(ClideIcon), findsNothing); // not collapsible → no caret
});
testWidgets('a multi-line tool result starts collapsed with a first-line summary', (tester) async {
await pumpWith(tester, [_result('first line\nsecond line\nthird line')]);
// Collapsed: caret present, summary (first line) shown, full body hidden.
expect(find.byType(ClideIcon), findsOneWidget);
expect(find.text('first line'), findsOneWidget);
expect(find.text('first line\nsecond line\nthird line'), findsNothing);
await tester.tap(find.byType(ClideIcon));
await tester.pump();
expect(find.text('first line\nsecond line\nthird line'), findsOneWidget);
});
testWidgets('select-all + copy spans multiple cards', (tester) async {
final clipboard = _MockClipboard();
TestDefaultBinaryMessengerBinding.instance.defaultBinaryMessenger.setMockMethodCallHandler(SystemChannels.platform, clipboard.handleMethodCall);
+151 -2
View File
@@ -1,16 +1,18 @@
import 'package:clide/builtin/claude/src/prompt_card.dart';
import 'package:clide/builtin/claude/src/stream_json_session.dart';
import 'package:flutter/widgets.dart';
import 'package:flutter_test/flutter_test.dart';
import '../../helpers/kernel_fixture.dart';
import '../../helpers/widget_harness.dart';
ToolPrompt permissionPrompt() => const ToolPrompt(
ToolPrompt permissionPrompt({List<dynamic> suggestions = const []}) => ToolPrompt(
promptId: 'req-1',
toolName: 'Write',
displayName: 'Write',
description: 'banana.txt',
input: {'file_path': '/tmp/banana.txt', 'content': 'banana'},
input: const {'file_path': '/tmp/banana.txt', 'content': 'banana'},
permissionSuggestions: suggestions,
);
ToolPrompt questionPrompt({bool multi = false}) => ToolPrompt(
@@ -32,6 +34,34 @@ ToolPrompt questionPrompt({bool multi = false}) => ToolPrompt(
},
);
ToolPrompt twoQuestionPrompt() => const ToolPrompt(
promptId: 'req-2q',
toolName: 'AskUserQuestion',
displayName: 'AskUserQuestion',
input: {
'questions': [
{
'question': 'Which pet?',
'header': 'Pet',
'multiSelect': false,
'options': [
{'label': 'Cats', 'description': ''},
{'label': 'Dogs', 'description': ''},
],
},
{
'question': 'How eaten?',
'header': 'Eaten',
'multiSelect': false,
'options': [
{'label': 'Fresh', 'description': ''},
{'label': 'Smoothie', 'description': ''},
],
},
],
},
);
void main() {
late KernelFixture f;
setUp(() async => f = await KernelFixture.create());
@@ -79,6 +109,51 @@ void main() {
expect((decision as DenyTool).message, isNotEmpty);
});
testWidgets('permission: no "don\'t ask again" button without a suggestion', (tester) async {
await tester.pumpWidget(harness(f, ToolPromptCard(prompt: permissionPrompt(), onResolve: (_, __) {})));
await tester.pump();
expect(find.text("Allow & don't ask again"), findsNothing);
});
testWidgets('permission: "don\'t ask again" shows with a suggestion and returns updatedPermissions', (tester) async {
ToolDecision? decision;
const sugg = [
{'type': 'setMode', 'mode': 'acceptEdits', 'destination': 'session'}
];
await tester.pumpWidget(harness(
f,
ToolPromptCard(prompt: permissionPrompt(suggestions: sugg), onResolve: (_, d) => decision = d),
));
await tester.pump();
expect(find.text("Allow & don't ask again"), findsOneWidget);
await tester.tap(find.text("Allow & don't ask again"));
await tester.pump();
expect((decision as AllowTool).updatedPermissions, hasLength(1));
});
testWidgets('permission: a typed note rides Deny as the message', (tester) async {
ToolDecision? decision;
await tester.pumpWidget(harness(f, ToolPromptCard(prompt: permissionPrompt(), onResolve: (_, d) => decision = d)));
await tester.pump();
await tester.enterText(find.byType(EditableText), 'write it under docs/ instead');
await tester.pump();
await tester.tap(find.text('Deny'));
await tester.pump();
expect((decision as DenyTool).message, 'write it under docs/ instead');
});
testWidgets('permission: a typed note rides Allow as a follow-up note', (tester) async {
ToolDecision? decision;
await tester.pumpWidget(harness(f, ToolPromptCard(prompt: permissionPrompt(), onResolve: (_, d) => decision = d)));
await tester.pump();
await tester.enterText(find.byType(EditableText), 'fyi: sandbox only');
await tester.pump();
await tester.tap(find.text('Allow'));
await tester.pump();
expect((decision as AllowTool).followUpNote, 'fyi: sandbox only');
});
testWidgets('question card: Submit is gated until an option is picked, then returns answers', (tester) async {
ToolDecision? decision;
await tester.pumpWidget(harness(
@@ -122,4 +197,78 @@ void main() {
final answers = (decision as AllowTool).updatedInput['answers'] as Map;
expect(answers['Do you prefer cats or dogs?'], 'Cats, Dogs');
});
testWidgets('question card: "Other…" free-text becomes the answer value', (tester) async {
ToolDecision? decision;
await tester.pumpWidget(harness(f, ToolPromptCard(prompt: questionPrompt(), onResolve: (_, d) => decision = d)));
await tester.pump();
await tester.tap(find.text('○ Other…'));
await tester.pump();
// Two fields now: [0] = the Other free-text, [1] = the per-choice note.
await tester.enterText(find.byType(EditableText).first, 'Kiwi');
await tester.pump();
await tester.tap(find.text('Submit'));
await tester.pump();
final answers = (decision as AllowTool).updatedInput['answers'] as Map;
expect(answers['Do you prefer cats or dogs?'], 'Kiwi'); // not the word "Other"
});
testWidgets('question card: a per-choice note is appended to the label', (tester) async {
ToolDecision? decision;
await tester.pumpWidget(harness(f, ToolPromptCard(prompt: questionPrompt(), onResolve: (_, d) => decision = d)));
await tester.pump();
await tester.tap(find.textContaining('Dogs'));
await tester.pump();
await tester.enterText(find.byType(EditableText), 'only big ones'); // the note field
await tester.pump();
await tester.tap(find.text('Submit'));
await tester.pump();
final answers = (decision as AllowTool).updatedInput['answers'] as Map;
expect(answers['Do you prefer cats or dogs?'], 'Dogs — only big ones');
});
testWidgets('multi-question: steps through to review, then submits both answers', (tester) async {
ToolDecision? decision;
await tester.pumpWidget(harness(f, ToolPromptCard(prompt: twoQuestionPrompt(), onResolve: (_, d) => decision = d)));
await tester.pump();
// Stepper nav shows numbered headers; only question 1 is visible.
expect(find.textContaining('1 · Pet'), findsOneWidget);
expect(find.text('Which pet?'), findsOneWidget);
expect(find.text('How eaten?'), findsNothing);
await tester.tap(find.textContaining('Dogs'));
await tester.pump();
await tester.tap(find.text('Next '));
await tester.pump();
expect(find.text('How eaten?'), findsOneWidget);
await tester.tap(find.textContaining('Fresh'));
await tester.pump();
await tester.tap(find.text('Review '));
await tester.pump();
// Review screen lists both answers; submit delivers them.
expect(find.text('Review your answers'), findsOneWidget);
await tester.tap(find.text('Submit answers'));
await tester.pump();
final answers = (decision as AllowTool).updatedInput['answers'] as Map;
expect(answers['Which pet?'], 'Dogs');
expect(answers['How eaten?'], 'Fresh');
});
testWidgets('question card: "chat instead" denies the prompt', (tester) async {
ToolDecision? decision;
await tester.pumpWidget(harness(f, ToolPromptCard(prompt: questionPrompt(), onResolve: (_, d) => decision = d)));
await tester.pump();
await tester.tap(find.text('chat instead'));
await tester.pump();
expect(decision, isA<DenyTool>());
});
}
@@ -179,6 +179,50 @@ void main() {
expect((decision['updatedInput'] as Map)['content'], 'banana');
});
test('a permission request carries its permission_suggestions', () async {
proc.emit(jsonEncode({
'type': 'control_request',
'request_id': 'rs',
'request': {
'subtype': 'can_use_tool',
'tool_name': 'Write',
'input': {'file_path': '/tmp/x'},
'permission_suggestions': [
{'type': 'setMode', 'mode': 'acceptEdits', 'destination': 'session'}
],
},
}));
await Future<void>.delayed(Duration.zero);
expect(session.pendingPrompt!.permissionSuggestions, hasLength(1));
});
test('resolvePrompt(allow with updatedPermissions) echoes them in the response', () async {
proc.emit(canUseTool('rp'));
await Future<void>.delayed(Duration.zero);
session.resolvePrompt(
'rp',
AllowTool(const {
'x': 1
}, updatedPermissions: const [
{'type': 'setMode'}
]));
final decision = ((jsonDecode(proc.writes.single) as Map)['response'] as Map)['response'] as Map;
expect(decision['behavior'], 'allow');
expect(decision['updatedPermissions'], hasLength(1));
});
test('resolvePrompt(allow with a follow-up note) sends the note as a user message', () async {
proc.emit(canUseTool('rn'));
await Future<void>.delayed(Duration.zero);
session.resolvePrompt('rn', AllowTool(const {'x': 1}, followUpNote: 'use docs/ instead'));
// first write = control_response (allow), second = the follow-up message
expect(proc.writes, hasLength(2));
final follow = jsonDecode(proc.writes[1]) as Map;
expect(follow['type'], 'user');
expect((follow['message'] as Map)['content'], 'use docs/ instead');
});
test('resolvePrompt(deny) writes a deny decision with a message', () async {
proc.emit(canUseTool('req-3'));
await Future<void>.delayed(Duration.zero);
@@ -189,6 +233,29 @@ void main() {
expect(decision['message'], 'nope');
});
test('resolving an AskUserQuestion leaves an answered echo in the log', () async {
proc.emit(jsonEncode({
'type': 'control_request',
'request_id': 'aq',
'request': {
'subtype': 'can_use_tool',
'tool_name': 'AskUserQuestion',
'input': {'questions': <dynamic>[]},
},
}));
await Future<void>.delayed(Duration.zero);
session.resolvePrompt(
'aq',
AllowTool(const {
'answers': {'Pet': 'Dogs'}
}));
await Future<void>.delayed(Duration.zero);
final echo = items.whereType<UserMessage>().toList();
expect(echo, hasLength(1));
expect(echo.single.text, contains('Pet → Dogs'));
});
test('prompts queue: resolving the head surfaces the next', () async {
proc.emit(canUseTool('q1'));
proc.emit(canUseTool('q2'));