fold sub-agent prompts via parent_tool_use_id (T-338)

In live stream-json sessions a sub-agent's spawning prompt is tagged
with parent_tool_use_id, not the transcript JSONL's isSidechain +
parentUuid. The parser ignored that field, so the prompt parsed as a
main-thread user turn and rendered as a blue "you" card above the
Activity Agent card instead of folding into it.

Carry parent_tool_use_id onto ConversationItem; its presence now marks
the item as a sidechain message. The sidechain fold resolves ownership
directly by tool-use id (no transcript-only uuid chain to walk), so the
prompt folds into its Agent card and the run nests under it as before.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

.pql/changelog/ticket_history/2026-06.sql

@@ -3186,3 +3186,4 @@ INSERT INTO ticket_history (ticket_record_id, field, old_value, new_value, chang
 INSERT INTO ticket_history (ticket_record_id, field, old_value, new_value, changed_by, changed_at, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06FB3KS499THAD899M0NWN7E3R', 'status', 'backlog', 'ready', NULL, '2026-06-10 13:53:55', '2026-06-10 13:53:55', '2026-06-10 13:53:55', NULL, '72773d8d91af176305bd6c1b7d5d65be', 2) ON CONFLICT(hash) DO NOTHING;
 INSERT INTO ticket_history (ticket_record_id, field, old_value, new_value, changed_by, changed_at, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06FB3JM0AXK1CTWD720RV1AVZ0', 'status', 'backlog', 'ready', NULL, '2026-06-10 13:54:02', '2026-06-10 13:54:02', '2026-06-10 13:54:02', NULL, '5cb488441033ac5a781d4b7d91dbda6f', 2) ON CONFLICT(hash) DO NOTHING;
 INSERT INTO ticket_history (ticket_record_id, field, old_value, new_value, changed_by, changed_at, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06FB0TNQM7F2TBJJV1F2KP7XR8', 'status', 'in_progress', 'done', NULL, '2026-06-10 13:55:30', '2026-06-10 13:55:30', '2026-06-10 13:55:30', NULL, 'f618ee8a3a33dabfa822bd0cebae4e03', 2) ON CONFLICT(hash) DO NOTHING;
+INSERT INTO ticket_history (ticket_record_id, field, old_value, new_value, changed_by, changed_at, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06FB3JAXDKZMS0805MMEYB6820', 'status', 'ready', 'done', NULL, '2026-06-10 14:03:10', '2026-06-10 14:03:10', '2026-06-10 14:03:10', NULL, '96365bc3ef7dbf505448be2dad16e4b6', 2) ON CONFLICT(hash) DO NOTHING;

.pql/changelog/tickets/2026-06.sql

@@ -2582,3 +2582,12 @@ INSERT INTO tickets (record_id, type, parent_record_id, title, description, stat
 - Markdown rendering is clide-owned custom CustomPaint/widgets (not a package), so detection + hit-testing lands in the conversation/markdown render path.
 - Honors User/Claude parity (D-6): the click maps to the existing `clide editor open` verb.
 - Open questions to settle during design: how aggressive path detection should be (existence check vs. heuristic), handling of non-existent / external paths, and visual affordance (underline/hover) for a linkified ref.', 'done', 'medium', NULL, NULL, NULL, '2026-06-09 20:02:41', '2026-06-10 13:55:30', NULL, 'd4022c088022495e2a0bda625a45c636', 2) ON CONFLICT(record_id) DO UPDATE SET type=excluded.type, parent_record_id=excluded.parent_record_id, title=excluded.title, description=excluded.description, status=excluded.status, priority=excluded.priority, assigned_to=excluded.assigned_to, team=excluded.team, decision_ref=excluded.decision_ref, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at > tickets.updated_at OR (excluded.updated_at = tickets.updated_at AND excluded.hash > tickets.hash);
+INSERT INTO tickets (record_id, type, parent_record_id, title, description, status, priority, assigned_to, team, decision_ref, created_at, updated_at, deleted_at, hash, canonical_version) VALUES ('06FB3JAXDKZMS0805MMEYB6820', 'bug', '06FB0TNQM5TWC00GW0P3X02HZW', 'Agent prompt renders as a ''you'' card instead of folding into the Activity Agent card', 'When Claude spawns a sub-agent (Agent/Task tool, rendered as an ''Activity Agent'' card), the prompt handed to the sub-agent renders as a separate top-level UserMessage card labelled ''you'' (blue, globalFocus accent) above the card, instead of folding into the Activity Agent card as a muted ''agent prompt'' (the T-263 behaviour).
+
+Repro: launch any sub-agent from the Claude pane; the task prompt appears as a ''you'' turn.
+
+Root cause: the sidechain fold keys entirely off the transcript envelope. In transcript_reader.dart the prompt''s UserMessage carries isSidechain/parentUuid straight from envelope[''isSidechain''] and envelope[''parentUuid''] (transcript_reader.dart:558-560, parsed in _parseUserInto at :593-643). conversation_view.dart _sidechainFold() (:179-233) only claims an item when it.isSidechain is true AND resolveOwner() walks its parentUuid chain up to the Agent tool_use message (or the nearest preceding Agent as fallback). If the spawning prompt''s envelope arrives WITHOUT isSidechain==true (or without a parentUuid that roots at the Agent), the item is never added to ownedSidechainUuids, _visibleItems() (:116-160) never drops it, and _ConversationTurn.build() falls through past the ''when i.injected || i.isSidechain'' guard into the default UserMessage() case => label: ''you'' (the blue user card).
+
+Need to confirm whether the prompt envelope is missing isSidechain (likely — the spawning prompt is the FIRST user turn of the sidechain and may be emitted by the harness without the flag, or with a parentUuid that points at the spawning turn rather than the Agent tool_use). Check stream_json_session.dart emission too.
+
+Fix direction (TBD after confirming the envelope shape): either (a) ensure the prompt envelope is correctly flagged/parented, or (b) in _sidechainFold associate the first sidechain-adjacent user prompt with its Agent tool_use even when the flag/chain is incomplete. Related: T-263 (prompt folds into the call), T-264 (run nests under the card).', 'done', 'medium', NULL, NULL, NULL, '2026-06-10 13:47:04', '2026-06-10 14:03:10', NULL, 'ee474dfe790eb6e72bd6a7381cd900e2', 2) ON CONFLICT(record_id) DO UPDATE SET type=excluded.type, parent_record_id=excluded.parent_record_id, title=excluded.title, description=excluded.description, status=excluded.status, priority=excluded.priority, assigned_to=excluded.assigned_to, team=excluded.team, decision_ref=excluded.decision_ref, updated_at=excluded.updated_at, deleted_at=excluded.deleted_at, hash=excluded.hash, canonical_version=excluded.canonical_version WHERE excluded.updated_at > tickets.updated_at OR (excluded.updated_at = tickets.updated_at AND excluded.hash > tickets.hash);

CHANGELOG.md

@@ -16,6 +16,15 @@ heading, and (b) bumping `pubspec.yaml` `version:` in the same commit.
 
 ## [Unreleased]
 
+### Fixed
+
+- **Sub-agent prompts no longer render as a blue "you" card.** In live
+  (stream-json) sessions a sub-agent's spawning prompt is tagged with
+  `parent_tool_use_id` rather than the transcript's `isSidechain`/`parentUuid`,
+  so it slipped past the sidechain fold and showed as a top-level user turn. The
+  parser now treats `parent_tool_use_id` as a sidechain marker and routes the
+  item straight to its Agent card by tool-use id. (T-338)
+
 ### Added
 
 - **Clickable file references in the Claude conversation.** Workspace file paths

lib/builtin/claude/src/conversation_view.dart

@@ -182,6 +182,13 @@ class _ConversationViewState extends State<ConversationView> {
       for (final it in items)
         if (it is AssistantToolUse && _isAgentTool(it.name)) it.uuid: it,
     };
+    // Stream-json tags sidechain items with the spawning Agent's tool-use id
+    // directly (T-338), so map Agent cards by tool-use id for a direct lookup
+    // that doesn't depend on the (transcript-only) parentUuid chain.
+    final agentByToolUseId = <String, AssistantToolUse>{
+      for (final it in items)
+        if (it is AssistantToolUse && _isAgentTool(it.name)) it.toolUseId: it,
+    };
     // Envelope-level chain info (consistent across items sharing a uuid).
     final parentByUuid = <String, String?>{};
     final sidechainByUuid = <String, bool>{};
@@ -193,6 +200,13 @@ class _ConversationViewState extends State<ConversationView> {
     }
 
     AssistantToolUse? resolveOwner(ConversationItem item, AssistantToolUse? nearest) {
+      // Direct route: stream-json hands us the spawning Agent's tool-use id on
+      // the item itself (T-338) — no chain to walk.
+      final byTool = item.parentToolUseId;
+      if (byTool != null) {
+        final agent = agentByToolUseId[byTool];
+        if (agent != null) return agent;
+      }
       var cur = item.uuid;
       final seen = <String>{};
       while (seen.add(cur)) {

lib/builtin/claude/src/transcript_reader.dart

@@ -39,7 +39,7 @@ import 'dart:isolate';
 
 /// Discriminated union of conversation items the reader can emit.
 sealed class ConversationItem {
-  const ConversationItem({required this.uuid, required this.timestamp, required this.isSidechain, this.parentUuid});
+  const ConversationItem({required this.uuid, required this.timestamp, required this.isSidechain, this.parentUuid, this.parentToolUseId});
 
   final String uuid;
   final DateTime timestamp;
@@ -50,6 +50,14 @@ sealed class ConversationItem {
   /// branches off the assistant message that issued its spawning Agent/Task
   /// tool-use, so this links the prompt to the right Agent card (T-263).
   final String? parentUuid;
+
+  /// The `parent_tool_use_id` from the stream-json wire (T-338): the tool-use
+  /// id of the Agent/Task call that spawned this sub-agent message. Stream-json
+  /// tags every sidechain item with it — the transcript JSONL instead uses
+  /// [isSidechain] + [parentUuid]. When present it routes the item straight to
+  /// its Agent card by tool-use id, no uuid-chain walk needed, and on its own
+  /// marks the item as a sidechain message.
+  final String? parentToolUseId;
 }
 
 /// A user-typed message (plain text, possibly multi-part).
@@ -59,6 +67,7 @@ final class UserMessage extends ConversationItem {
     required super.timestamp,
     required super.isSidechain,
     super.parentUuid,
+    super.parentToolUseId,
     required this.text,
     this.injected = false,
   });
@@ -82,6 +91,7 @@ final class ToolResultMessage extends ConversationItem {
     required super.timestamp,
     required super.isSidechain,
     super.parentUuid,
+    super.parentToolUseId,
     required this.toolUseId,
     required this.content,
     required this.isError,
@@ -102,6 +112,7 @@ final class AssistantTextMessage extends ConversationItem {
     required super.timestamp,
     required super.isSidechain,
     super.parentUuid,
+    super.parentToolUseId,
     required this.text,
   });
 
@@ -118,6 +129,7 @@ final class AssistantThinkingMessage extends ConversationItem {
     required super.timestamp,
     required super.isSidechain,
     super.parentUuid,
+    super.parentToolUseId,
     required this.thinking,
   });
 
@@ -134,6 +146,7 @@ final class AssistantToolUse extends ConversationItem {
     required super.timestamp,
     required super.isSidechain,
     super.parentUuid,
+    super.parentToolUseId,
     required this.toolUseId,
     required this.name,
     required this.input,
@@ -555,9 +568,14 @@ void _parseLineInto(String line, List<ConversationItem> out, List<String> warnin
   if (_skipTypes.contains(type)) return;
 
   final uuid = envelope['uuid'] as String? ?? '';
-  final isSidechain = envelope['isSidechain'] as bool? ?? false;
   final rawParent = envelope['parentUuid'] as String?;
   final parentUuid = (rawParent != null && rawParent.isNotEmpty) ? rawParent : null;
+  // Stream-json tags sub-agent messages with `parent_tool_use_id` (the spawning
+  // Agent/Task tool-use), not the transcript's isSidechain/parentUuid (T-338).
+  // Treat its presence as a sidechain marker so the fold + de-emphasis kick in.
+  final rawParentTool = envelope['parent_tool_use_id'] as String?;
+  final parentToolUseId = (rawParentTool != null && rawParentTool.isNotEmpty) ? rawParentTool : null;
+  final isSidechain = (envelope['isSidechain'] as bool? ?? false) || parentToolUseId != null;
 
   DateTime timestamp;
   try {
@@ -568,9 +586,9 @@ void _parseLineInto(String line, List<ConversationItem> out, List<String> warnin
 
   switch (type) {
     case 'user':
-      _parseUserInto(envelope, uuid, timestamp, isSidechain, parentUuid, out);
+      _parseUserInto(envelope, uuid, timestamp, isSidechain, parentUuid, parentToolUseId, out);
     case 'assistant':
-      _parseAssistantInto(envelope, uuid, timestamp, isSidechain, parentUuid, out);
+      _parseAssistantInto(envelope, uuid, timestamp, isSidechain, parentUuid, parentToolUseId, out);
       _extractAssistantStatus(envelope, status);
     default:
       break; // unknown type — degrade gracefully
@@ -596,6 +614,7 @@ void _parseUserInto(
   DateTime timestamp,
   bool isSidechain,
   String? parentUuid,
+  String? parentToolUseId,
   List<ConversationItem> out,
 ) {
   final message = envelope['message'] as Map?;
@@ -609,7 +628,14 @@ void _parseUserInto(
 
   if (content is String) {
     if (content.isNotEmpty) {
-      out.add(UserMessage(uuid: uuid, timestamp: timestamp, isSidechain: isSidechain, parentUuid: parentUuid, text: content, injected: injected));
+      out.add(UserMessage(
+          uuid: uuid,
+          timestamp: timestamp,
+          isSidechain: isSidechain,
+          parentUuid: parentUuid,
+          parentToolUseId: parentToolUseId,
+          text: content,
+          injected: injected));
     }
     return;
   }
@@ -629,6 +655,7 @@ void _parseUserInto(
           timestamp: timestamp,
           isSidechain: isSidechain,
           parentUuid: parentUuid,
+          parentToolUseId: parentToolUseId,
           toolUseId: item['tool_use_id'] as String? ?? '',
           content: rawContent is String ? rawContent : jsonEncode(rawContent),
           isError: item['is_error'] as bool? ?? false,
@@ -648,6 +675,7 @@ void _parseAssistantInto(
   DateTime timestamp,
   bool isSidechain,
   String? parentUuid,
+  String? parentToolUseId,
   List<ConversationItem> out,
 ) {
   final message = envelope['message'] as Map?;
@@ -661,12 +689,14 @@ void _parseAssistantInto(
       case 'text':
         final text = item['text'] as String? ?? '';
         if (text.isNotEmpty) {
-          out.add(AssistantTextMessage(uuid: uuid, timestamp: timestamp, isSidechain: isSidechain, parentUuid: parentUuid, text: text));
+          out.add(AssistantTextMessage(
+              uuid: uuid, timestamp: timestamp, isSidechain: isSidechain, parentUuid: parentUuid, parentToolUseId: parentToolUseId, text: text));
         }
       case 'thinking':
         final thinking = item['thinking'] as String? ?? '';
         if (thinking.isNotEmpty) {
-          out.add(AssistantThinkingMessage(uuid: uuid, timestamp: timestamp, isSidechain: isSidechain, parentUuid: parentUuid, thinking: thinking));
+          out.add(AssistantThinkingMessage(
+              uuid: uuid, timestamp: timestamp, isSidechain: isSidechain, parentUuid: parentUuid, parentToolUseId: parentToolUseId, thinking: thinking));
         }
       case 'tool_use':
         final rawInput = item['input'];
@@ -675,6 +705,7 @@ void _parseAssistantInto(
           timestamp: timestamp,
           isSidechain: isSidechain,
           parentUuid: parentUuid,
+          parentToolUseId: parentToolUseId,
           toolUseId: item['id'] as String? ?? '',
           name: item['name'] as String? ?? '',
           input: rawInput is Map ? rawInput.cast<String, dynamic>() : <String, dynamic>{},

test/builtin/claude/conversation_view_test.dart

@@ -430,6 +430,24 @@ void main() {
       expect(find.text('find all the widgets'), findsOneWidget);
     });
 
+    testWidgets('a stream-json prompt folds via parentToolUseId, not "you" (T-338)', (tester) async {
+      // The live stream-json wire flags sub-agent prompts with parent_tool_use_id
+      // (the Task tool-use id) and NO isSidechain/parentUuid — the prompt must
+      // still fold into the Agent card rather than render as a blue "you" turn.
+      await pumpWith(tester, [
+        AssistantToolUse(uuid: 'agt-msg', timestamp: _t, isSidechain: false, toolUseId: 'task1', name: 'Task', input: const {'description': 'explore'}),
+        UserMessage(uuid: 'sp', timestamp: _t, isSidechain: true, parentToolUseId: 'task1', text: 'find all the widgets'),
+      ]);
+      expect(find.text('you'), findsNothing);
+      expect(find.text('agent prompt'), findsNothing);
+      expect(find.text('Task'), findsOneWidget);
+      expect(find.text('find all the widgets'), findsNothing); // folded, collapsed
+
+      await tester.tap(find.bySemanticsLabel('Task, 1 step, collapsed'));
+      await tester.pumpAndSettle();
+      expect(find.text('find all the widgets'), findsOneWidget);
+    });
+
     testWidgets('an orphan sidechain prompt renders as muted "agent prompt", never "you" (T-263)', (tester) async {
       // No Agent tool-use to attach to → stays standalone, but relabelled.
       await pumpWith(tester, [

test/builtin/claude/transcript_reader_test.dart

@@ -378,6 +378,30 @@ void main() {
       expect(items[0].parentUuid, 'msg-A');
       expect(items[1].parentUuid, isNull); // '' → null
     });
+
+    test('stream-json parent_tool_use_id marks a sidechain message (T-338)', () {
+      // The stream-json wire tags sub-agent messages with parent_tool_use_id and
+      // NO isSidechain flag — we must treat it as a sidechain item anyway.
+      final raw = envelope(
+        type: 'user',
+        uuid: 'sp',
+        message: {'role': 'user', 'content': 'go explore the codebase'},
+      )..['parent_tool_use_id'] = 'toolu_task1';
+      final items = parseAll([raw]);
+      expect(items.first.isSidechain, isTrue);
+      expect(items.first.parentToolUseId, 'toolu_task1');
+    });
+
+    test('empty parent_tool_use_id normalises to null and stays main-thread (T-338)', () {
+      final raw = envelope(
+        type: 'user',
+        uuid: 'mt',
+        message: {'role': 'user', 'content': 'a normal turn'},
+      )..['parent_tool_use_id'] = '';
+      final items = parseAll([raw]);
+      expect(items.first.isSidechain, isFalse);
+      expect(items.first.parentToolUseId, isNull);
+    });
   });
 
   // -------------------------------------------------------------------------