diff --git a/coworker/engine.py b/coworker/engine.py
index ae34d4a0..7ef75aa9 100644
--- a/coworker/engine.py
+++ b/coworker/engine.py
@@ -1018,8 +1018,10 @@ def _outbound_messages(self) -> list[dict[str, Any]]:
         `ts` — (providers reject unknown keys), unconditionally — whether or not a
         `<system-context>` block is added. When a context
         provider yields a non-empty string, an ephemeral `<system-context>` block is appended to the
-        last user message. Never mutates `self.messages`, so neither the strip nor the block is
-        persisted/replayed.
+        last user message. A partial assistant turn abandoned by a turn that died mid-stream —
+        identified by the error/interrupted notice that follows it — is dropped too: display-only
+        like the notices, and rejected outright by several providers as the final message.
+        Never mutates `self.messages`, so none of this is persisted/replayed.
         """
         # Strip the display-only sidecars — `source` (connector cards), `_display`
         # (e.g. filter-hidden counts), `ts` (append-time timestamps), `reasoning`
@@ -1033,14 +1035,43 @@ def _outbound_messages(self) -> list[dict[str, Any]]:
         source_messages = _compaction.apply_to_outbound(
             self.messages, self.compaction_state
         )
+        # A turn that dies mid-stream (provider error, interrupt) leaves its partial assistant
+        # message in the transcript — so the user can still see what streamed — followed by the
+        # notice saying why it stopped. That partial is display-only in the same sense the
+        # notices are: not a turn the model finished, and several providers reject it as the
+        # FINAL message. Gemini 400s ("Requests ending with a model turn are not supported"),
+        # Claude 4.6+ 400s ("does not support assistant message prefill"), and one ending in
+        # whitespace 400s on older Claude ("final assistant content cannot end with trailing
+        # whitespace"). `retry()` replays the same tail, so the session wedges for good.
+        # The trailing notice is what identifies it: a COMPLETED turn's assistant message looks
+        # identical but has no error/interrupted notice after it, and must be kept (its `extras`
+        # sidecar is replayed back to the owning provider). Trailing `tool_calls` mean a turn
+        # suspended awaiting durable resume (see `resume`), never an abandoned one. Consecutive
+        # failed retries stack one partial each, so walk back through the whole run.
+        abandoned: set[int] = set()
+        failed = False
+        for i in range(len(source_messages) - 1, -1, -1):
+            message = source_messages[i]
+            if message.get("role") == "notice":
+                failed = failed or message.get("kind") in ("error", "interrupted")
+                continue
+            if (
+                failed
+                and message.get("role") == "assistant"
+                and not message.get("tool_calls")
+            ):
+                abandoned.add(i)
+                failed = False  # that notice belongs to this partial
+                continue
+            break
         out = [
             (
                 {k: v for k, v in msg.items() if k not in _SIDECARS}
                 if any(s in msg for s in _SIDECARS)
                 else msg
             )
-            for msg in source_messages
-            if msg.get("role") != "notice"
+            for i, msg in enumerate(source_messages)
+            if msg.get("role") != "notice" and i not in abandoned
         ]
         # PDF attachments (stored as `file` parts) are adapted to the ACTIVE model right
         # here — never in the persisted history — so a mid-session model switch always
diff --git a/tests/test_durable_resume.py b/tests/test_durable_resume.py
index fee2bcad..264f0466 100644
--- a/tests/test_durable_resume.py
+++ b/tests/test_durable_resume.py
@@ -94,6 +94,47 @@ async def scenario():
     assert mgr.inbox.pending(sid) == []  # nothing left pending
 
 
+def test_resume_payload_keeps_the_suspended_tool_call(tmp_path):
+    # `_outbound_messages` drops a TRAILING assistant turn (the partial a mid-stream death
+    # leaves behind). A suspended turn is also stored ending on an assistant turn, so this
+    # pins that resume is unaffected: the pending calls are answered first, making the tail a
+    # `tool` message, and the assistant turn that made the call still reaches the model.
+    class Recording(ScriptedProvider):
+        def __init__(self, turns):
+            super().__init__(turns)
+            self.payloads = []
+
+        def complete(self, *, model, messages, tools=None, **settings):
+            self.payloads.append(messages)
+            return super().complete(
+                model=model, messages=messages, tools=tools, **settings
+            )
+
+    provider = Recording(
+        [
+            _tool("ask_user", {"question": "Which region?", "options": ["a", "b"]}, "call_q"),
+            _text("You chose b."),
+        ]
+    )
+    mgr = SessionManager(workspace=tmp_path, provider=provider)
+    sid = "dur-payload"
+
+    async def scenario():
+        engine = mgr.get_engine(sid, agent="cowork", workspace=str(tmp_path))
+        item = await _run_until_pending(mgr, sid, engine)
+        await mgr.resolve_inbox(item.id, "b")
+
+    asyncio.run(scenario())
+
+    resumed = next(
+        p for p in provider.payloads if any(m.get("role") == "tool" for m in p)
+    )
+    assert [m.get("role") for m in resumed][-2:] == ["assistant", "tool"]
+    assert any(
+        m.get("role") == "assistant" and m.get("tool_calls") for m in resumed
+    ), "the assistant turn that made the call must still reach the model"
+
+
 def test_durable_resume_approval_executes_tool(tmp_path):
     # The model wants a write (needs approval); on durable resume "allow" must RE-EXECUTE the tool.
     target = tmp_path / "scratch_marker.txt"
diff --git a/tests/test_engine.py b/tests/test_engine.py
index 90e38fdb..7e6cb25c 100644
--- a/tests/test_engine.py
+++ b/tests/test_engine.py
@@ -319,6 +319,88 @@ def test_streaming_emits_deltas(tmp_path):
     assert events[-1].type == EventType.TURN_END
 
 
+class DyingStreamProvider(ProviderClient):
+    """Streams a delta, then dies — the shape of a provider 503 landing mid-turn."""
+
+    def __init__(self):
+        self.payloads = []
+
+    def complete(self, **kwargs):  # pragma: no cover - streamed instead
+        raise NotImplementedError
+
+    def capabilities(self, model):
+        return ModelCapabilities()
+
+    def stream(self, *, model, messages, tools=None, **settings):
+        self.payloads.append(messages)
+        yield StreamChunk(text_delta="I will request permission to ")
+        raise RuntimeError("Error code: 503 - {'error': {'type': 'overloaded_error'}}")
+
+
+def test_retry_after_mid_stream_death_does_not_replay_the_partial(tmp_path):
+    # The partial stays in the transcript so the user can see what streamed, but several
+    # providers reject a payload whose FINAL message is an assistant turn — Gemini ("Requests
+    # ending with a model turn are not supported"), Claude 4.6+ ("does not support assistant
+    # message prefill"), older Claude when it ends in whitespace. Replaying it wedges retry.
+    provider = DyingStreamProvider()
+    engine = TurnEngine(
+        provider=provider,
+        registry=ToolRegistry(),
+        permissions=PermissionEngine(workspace_root=tmp_path),
+        model="gpt-5.5",
+    )
+    _collect(engine, "arrange my visa documents")
+
+    # kept in history (display), followed by the error notice
+    assert engine.messages[-2]["role"] == "assistant"
+    assert engine.messages[-2]["content"] == "I will request permission to "
+    assert engine.messages[-1]["role"] == "notice"
+
+    async def _retry():
+        return [ev async for ev in engine.retry()]
+
+    asyncio.run(_retry())
+
+    sent = provider.payloads[-1]
+    assert sent, "retry never reached the provider"
+    assert sent[-1]["role"] != "assistant"
+    assert all(m.get("content") != "I will request permission to " for m in sent)
+
+
+def test_completed_assistant_turn_is_never_dropped(tmp_path):
+    # The counterpart to the test above: a finished turn's assistant message is structurally
+    # identical to an abandoned partial (trailing, no tool_calls). Only the error notice tells
+    # them apart, so a completed turn with no notice after it must survive intact.
+    engine, _ = _engine(tmp_path, [_text_turn("all done")])
+    _collect(engine, "hi")
+
+    assert engine.messages[-1]["role"] == "assistant"
+    outbound = engine._outbound_messages()
+    assert outbound[-1]["role"] == "assistant"
+    assert outbound[-1]["content"] == "all done"
+
+
+def test_consecutive_deaths_do_not_stack_partials_in_the_payload(tmp_path):
+    # Each failed retry appends its own partial, so one drop is not enough.
+    provider = DyingStreamProvider()
+    engine = TurnEngine(
+        provider=provider,
+        registry=ToolRegistry(),
+        permissions=PermissionEngine(workspace_root=tmp_path),
+        model="gpt-5.5",
+    )
+    _collect(engine, "go")
+
+    async def _retry():
+        return [ev async for ev in engine.retry()]
+
+    asyncio.run(_retry())
+    asyncio.run(_retry())
+
+    assert sum(1 for m in engine.messages if m.get("role") == "assistant") == 3
+    assert provider.payloads[-1][-1]["role"] != "assistant"
+
+
 def _pdf_file_part():
     import base64
     import io