diff --git a/NucleicRemote/NucleicRemote/Models/HostConnection.swift b/NucleicRemote/NucleicRemote/Models/HostConnection.swift
index 0dba1bd..009af0a 100644
--- a/NucleicRemote/NucleicRemote/Models/HostConnection.swift
+++ b/NucleicRemote/NucleicRemote/Models/HostConnection.swift
@@ -42,6 +42,8 @@ final class HostConnection {
             guard openSessionID != oldValue else { return }
             openMaxSeq = nil
             seenSeq = []
+            transcriptFetchRequestID = nil
+            didStartFullFetch = false
         }
     }
 
@@ -50,6 +52,15 @@ final class HostConnection {
     /// instead of cold-resetting to the 200-event tail (which would truncate a long transcript).
     private var openMaxSeq: UInt64?
 
+    /// The in-flight full-history fetch's request id (mesh full-transcript sync). On open we pull the
+    /// *whole* transcript — the cold `subscribe` only returns a 200-event tail — and merge its chunks
+    /// into the transcript on screen; replies are matched on this id so a late batch from a previous
+    /// open (the session changed underneath us) is dropped. `didStartFullFetch` guards against issuing
+    /// it twice — both `open` and the reconnect nudge call `fetchFullTranscript`, but only the first
+    /// that finds a live, capable connection actually sends.
+    private var transcriptFetchRequestID: String?
+    private var didStartFullFetch = false
+
     // MARK: Callbacks up to RemoteStore (aggregate concerns)
 
     /// How a `HostConnection` talks back to `RemoteStore`. All fire on the main actor.