release

antgrid.yaml

@@ -17,3 +17,8 @@ worktree:
         workingDir: packages/antgrid_relay_client
     timeoutMs: 900000
     onFailure: warn
+    # The agent is useful before the installs finish — reading, planning and
+    # searching all work in a tree that is only checked out — and the five
+    # steps here are minutes on a cold worktree. It does race step 1: the
+    # service `.env` files can land after the agent is already up.
+    startAgent: immediate

app/CLAUDE.md

@@ -40,7 +40,7 @@ app's relay layer lives outside this tree: see
 
 **Never start async work from a `void` callback without `detached` (`util/detached.dart`).** A tap handler, a post-frame callback and a `ref.listen` all DISCARD the future they start, so a rejection reaches `PlatformDispatcher.onError` as a FATAL with no in-app frames to point at the site — which is how a `session:*` reply dropped while the transport re-establishes (`PendingReply`'s routine 15s `TimeoutException`, already reconciled by the next `session:updated` push) shipped as a crash in 1.20668.151. `await`ing INSIDE such a callback is not a fix — the callback is the boundary, and an `async` closure passed where a `VoidCallback` is expected is the same bug wearing an await. Where the user pressed something and is owed an answer, also catch the `TimeoutException` at the call and say so in the UI; a log line alone makes a confirmed action indistinguishable from a dropped tap.
 
-**A queued start is not a stopped session.** While `worktree.setup` runs the bridge HOLDS the session's `session:start`, so the entry reports `running: false` for the whole run with `setup.pendingStart` set. Every auto-start path must gate on `sessionStartQueued` (`providers/session_setup.dart`) or it sends a SECOND start carrying no `initialPrompt` — a start the user did not ask for, into a workspace that is not provisioned yet. `WorkspaceShellState._bootstrapSessions` and the drawer row's own tap (`session_row.dart`) are both such paths. The bridge keeps an already-queued prompt rather than letting a promptless start replace it, so the user's typing is no longer at stake, but that is a backstop and not the contract. For the same reason the create flow navigates on the SEND, not on the start reply.
+**A queued start is not a stopped session.** While `worktree.setup` runs the bridge HOLDS the session's `session:start`, so the entry reports `running: false` for the whole run with `setup.pendingStart` set. Every auto-start path must gate on `sessionStartQueued` (`providers/session_setup.dart`) or it sends a SECOND start carrying no `initialPrompt` — a start the user did not ask for, into a workspace that is not provisioned yet. `WorkspaceShellState._bootstrapSessions` and the drawer row's own tap (`session_row.dart`) are both such paths. The bridge keeps an already-queued prompt rather than letting a promptless start replace it, so the user's typing is no longer at stake, but that is a backstop and not the contract. For the same reason the create flow navigates on the SEND, not on the start reply. **Rendering paths owe the same gate**, not only auto-start ones: `TerminalScreen` branches to a provisioning pane whose body is the setup transcript itself, because calling a queued session stopped — over a Start button whose press only re-enters the gate — is the one account of itself the workspace must never give. That pane and `SessionSetupBanner` are two views of one run, so the banner drops its own tail line and log disclosure while a start is queued (both derived from `sessionStartQueued`, never from the two widgets knowing about each other). `setup.stepNames` is the ledger's ONLY source and is optional on the wire — a state recovered from disk knows how many steps ran but not what they were called, and `SetupStepLedger` renders nothing rather than a column of blanks. It is also the one `SessionSetup` field whose equality is a pairwise walk (`_sameNames`): two structurally equal lists are different objects, so `==` on the reference would make every re-parse of an unchanged entry look like a change.
 
 **`AbConfig` must round-trip every top-level `antgrid.yaml` key it does not model.** `ProjectSettingsScreen`'s Save re-serializes the whole config through `models/ab_config.dart`, so a key the model drops is DELETED from the user's file — `worktree` is carried verbatim as a raw map for exactly that reason, and a new block in `bridge/src/config.ts` with no home here is a data-loss bug nothing type-checks. A write also re-emits the YAML through the bridge's serializer, losing the user's comments and formatting: weigh that before adding another surface that edits the config on the user's behalf.
 

app/lib/design/ab_tokens.dart

@@ -239,6 +239,16 @@ abstract final class AbTokens {
   /// out by scrolling internally rather than widening this measure.
   static const transcriptMaxWidth = 680.0;
 
+  /// Max measure for a rendered markdown document in the file viewer. Wider
+  /// than [transcriptMaxWidth] because document body is [fontBody] where the
+  /// transcript's is [fontMd] — the same ~70ch, one step up the scale. Code
+  /// fences and tables opt out by scrolling internally.
+  static const documentMaxWidth = 720.0;
+
+  /// Width of the heading-outline rail beside a markdown document. Sized to
+  /// hold a nested `h3` label at [fontXs] without ellipsizing every entry.
+  static const documentOutlineWidth = 208.0;
+
   static const sidebarWidth = 48.0;
   static const commandTrayHeight = 44.0;
   static const bottomNavHeight = 56.0;

app/lib/models/ab_message.dart

@@ -259,23 +259,28 @@ class HandlerStatusMessage {
   /// What an absent per-session judge tool resolves to for PTY slots (the
   /// project's agent tool); chat slots resolve from their own session entry.
   final String? defaultTool;
-  final bool defaultNotifyOnly;
   final List<Map<String, dynamic>> sessions;
 
-  /// Every §5.2 snapshot the project still knows about, replayed like the
+  /// Every snapshot the project still knows about, replayed like the
   /// escalations so an app that restarted between the advert and the tap can
   /// still reach the undo. Project-level, not per session: the offer matters
   /// most once the session that took it has wrapped up.
   final List<Map<String, dynamic>> snapshots;
 
+  /// The morning-after wrap-up reports, replayed on the same terms as the
+  /// snapshots — the session they describe is disarmed by the time anyone reads
+  /// one, so the replay is the only path that survives an app restart between
+  /// the wrap-up and the read.
+  final List<Map<String, dynamic>> wrapUps;
+
   const HandlerStatusMessage({
     required this.id,
     required this.timestamp,
     required this.projectId,
     this.defaultTool,
-    required this.defaultNotifyOnly,
     required this.sessions,
     this.snapshots = const [],
+    this.wrapUps = const [],
   });
 }
 
@@ -1425,16 +1430,25 @@ Object? parseAbMessage(Map<String, dynamic> json) {
             if (s is Map<String, dynamic>) snapshots.add(s);
           }
         }
+        // Same guard, same reason — and here absent and empty genuinely mean
+        // the same thing, so presence is never read as a capability signal.
+        final wrapUpsJson = json['wrapUps'];
+        final wrapUps = <Map<String, dynamic>>[];
+        if (wrapUpsJson is List) {
+          for (final w in wrapUpsJson) {
+            if (w is Map<String, dynamic>) wrapUps.add(w);
+          }
+        }
         return HandlerStatusMessage(
           id: id,
           timestamp: timestamp,
           projectId: projectId,
           defaultTool: json['defaultTool'] is String
               ? json['defaultTool'] as String
               : null,
-          defaultNotifyOnly: json['defaultNotifyOnly'] == true,
           sessions: sessions,
           snapshots: snapshots,
+          wrapUps: wrapUps,
         );
       }