[eric] agents: an invisible thinking pill no longer breaks the repeat-card dedup

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014wtspwSFzZmjCx9UNPAorQ

backend/apps/agents/manager/run/handle_run_error.py

@@ -34,6 +34,12 @@ logger = logging.getLogger(__name__)
 
 
 
+@typechecked
+def p_renders_as_nothing(msg: Message) -> bool:
+    """An empty thinking pill occupies the tail while showing the user nothing, which silently broke the dedup below (live drill 2026-08-20: three identical cards)."""
+    return msg.role == "thinking" and not str(msg.content or "").strip()
+
+
 @typechecked
 def absorb_repeat_card(session: AgentSession, error_msg: Message) -> None:
     """Append the error card, unless the branch tail is already the IDENTICAL card with nothing
@@ -47,7 +53,8 @@ def absorb_repeat_card(session: AgentSession, error_msg: Message) -> None:
     because each retry's hidden prompt had displaced the previous card from the tail."""
     p_tail = [m for m in session.messages
               if getattr(m, "branch_id", None) in (None, session.active_branch_id)
-              and not getattr(m, "hidden", False)]
+              and not getattr(m, "hidden", False)
+              and not p_renders_as_nothing(m)]
     if p_tail and p_tail[-1].role == "system" and p_tail[-1].content == error_msg.content:
         error_msg.id = p_tail[-1].id
         p_tail[-1].timestamp = error_msg.timestamp

backend/apps/agents/manager/streaming/handle_assistant_message.py

@@ -220,7 +220,10 @@ async def handle_assistant_message(
                 content=p_copy,
                 branch_id=session.active_branch_id,
             )
-            session.messages.append(p_card)
+            # A retry ladder re-failing the same way must bump one card, not stack clones; the live
+            # drill produced three before this line existed.
+            from backend.apps.agents.manager.run.handle_run_error import absorb_repeat_card
+            absorb_repeat_card(session, p_card)
             logger.warning(
                 f"Agent {session_id}: provider returned {p_provider_error.kind} "
                 f"(status={p_provider_error.status}, lane={p_provider_error.lane}) as assistant "

backend/tests/test_error_card_dedup.py

@@ -78,3 +78,34 @@ def test_a_real_user_message_still_earns_a_fresh_card():
 
     shown = [m for m in s.messages if m.role == "system"]
     assert len(shown) == 2, "each real ask gets its own honest answer"
+
+
+def test_an_empty_thinking_pill_does_not_break_the_dedup():
+    """Found by the live provider-error drill 2026-08-20, invisible to every unit test here.
+
+    Each retry leaves a non-hidden `thinking` message with empty content sitting in the tail. It
+    renders as nothing, but it displaced the previous card from the tail scan, so a re-failing
+    ladder stacked three identical cards on screen while this suite stayed green.
+    """
+    s = AgentSession(name="t", model="sonnet")
+    first = p_card("Lost the connection to the model.")
+    absorb_repeat_card(s, first)
+    s.messages.append(Message(role="thinking", content="", branch_id="main"))
+    repeat = p_card("Lost the connection to the model.")
+    absorb_repeat_card(s, repeat)
+    p_cards = [m for m in s.messages if m.role == "system"]
+    assert len(p_cards) == 1, "an invisible pill must not earn the user a duplicate card"
+    assert repeat.id == first.id
+
+
+def test_a_thinking_pill_with_real_content_still_breaks_the_dedup():
+    """NEGATIVE CONTROL. Only the EMPTY pill is invisible; real thinking is content the user saw,
+    so a card after it is genuinely new and must not be absorbed into the older one."""
+    s = AgentSession(name="t", model="sonnet")
+    first = p_card("Lost the connection to the model.")
+    absorb_repeat_card(s, first)
+    s.messages.append(Message(role="thinking", content="Let me try that again.", branch_id="main"))
+    repeat = p_card("Lost the connection to the model.")
+    absorb_repeat_card(s, repeat)
+    p_cards = [m for m in s.messages if m.role == "system"]
+    assert len(p_cards) == 2, "visible thinking separates the two failures"