fix(agents): clear the first-run flag when the catch-up completes (#122)

src/pepper_cli/claude_channel.py

@@ -499,7 +499,6 @@ def _flush_backlog(self) -> None:
         self._push(content, meta)
         self.skipped.clear()
         self.skipped_times.clear()
-        self.first_run = False
 
     def _live_edge(self) -> None:
         """Tell Claude, once, that the channel started at the live edge: there is no backlog to
@@ -549,6 +548,9 @@ def _topic_words(self, topic_id: str) -> str:
     def _caught_up(self) -> None:
         self._flush_backlog()
         self._flush_resync()
+        # The catch-up is over whether or not anything was skipped. Left set when nothing was,
+        # ``first_run`` would go on suppressing resyncs for the rest of the process.
+        self.first_run = False
         self._persist()
 
     def _resync(self, cause: str) -> None:
@@ -829,7 +831,10 @@ def _handle_frame(self, frame: api_client.FeedFrame) -> str | None:
         if frame.kind == "error":
             return self._frame_problem(frame)
         if frame.kind == "comment":
-            if frame.payload.get("comment") != "connected":
+            # ``: ping`` is the contract's replay boundary: the server sends one only while it
+            # has nothing past this position. ``: connected`` opens a connection, and any other
+            # comment (a proxy keepalive, a newer server's frame) says neither has happened.
+            if frame.payload.get("comment") == "ping":
                 self._caught_up()
             return None
         scoped = self.config.room is not None

tests/test_claude_channel.py

@@ -888,6 +888,43 @@ def test_a_resync_names_its_cause_gives_room_ids_and_is_safe_on_a_first_run(monk
     assert f"{_TOPIC} (room_id {_ROOM}; Kitchen › Dinner)" in note["content"]
 
 
+def test_an_old_event_replayed_after_a_first_run_still_owes_a_resync(monkeypatch):
+    """A first run whose window holds no messages pushes no backlog, so nothing but the completed
+    catch-up can clear ``first_run``. The stream then drops, and the reconnect replays an older
+    unrecognised event, which was missed and owes a resync. Left set, the flag takes it for
+    first-run history and drops it."""
+    old = _now() - timedelta(days=3)
+    fake = _Fake(
+        _CONNECTED + _scope() + _PING,
+        503,
+        _CONNECTED + _scope() + _event(7, kind="group.renamed", created=old) + _PING,
+    )
+
+    result = run_channel(monkeypatch, fake, ["--room", _ROOM, "--start", "beginning"])
+
+    assert result.kinds() == ["resync"]
+
+
+def test_a_non_ping_comment_does_not_end_the_first_run_catch_up(monkeypatch):
+    """Only ``: ping`` marks replay completion; ``: connected`` opens the connection, and any
+    other comment (a proxy keepalive, a newer server's frame) proves neither. Left unfiltered,
+    such a comment ends the catch-up early, clears ``first_run``, flushes the backlog it owed,
+    and makes an old event replayed after it owe a spurious resync."""
+    old = _now() - timedelta(days=3)
+    fake = _Fake(
+        _CONNECTED + _scope() + _event(3, created=old) + ": keepalive\n\n",
+        503,
+        _CONNECTED + _scope() + _event(7, kind="group.renamed", created=old) + _PING,
+    )
+
+    result = run_channel(monkeypatch, fake, ["--room", _ROOM, "--start", "beginning"])
+
+    assert result.kinds() == ["backlog"]
+    (note,) = result.notifications
+    assert note["meta"]["first_run"] == "true"
+    assert "live_edge" not in note["meta"]
+
+
 def test_a_first_run_replay_of_old_room_events_owes_no_resync(monkeypatch):
     """The retained window on a first run holds old room events of every kind (a rename, an
     app install); none was missed, and the backlog already says to catch up."""