diff --git a/strix/interface/main.py b/strix/interface/main.py
index bc88da673..37a6de5b4 100644
--- a/strix/interface/main.py
+++ b/strix/interface/main.py
@@ -629,6 +629,7 @@ def main() -> None:  # noqa: PLR0912, PLR0915
     finally:
         tracer = get_global_tracer()
         if tracer:
+            tracer.cleanup()
             posthog.end(tracer, exit_reason=exit_reason)
 
     results_path = Path("strix_runs") / args.run_name
diff --git a/strix/llm/llm.py b/strix/llm/llm.py
index 5e6a01f73..b0b6b22dd 100644
--- a/strix/llm/llm.py
+++ b/strix/llm/llm.py
@@ -239,7 +239,9 @@ def _prepare_messages(self, conversation_history: list[dict[str, Any]]) -> list[
         conversation_history.extend(compressed)
         messages.extend(compressed)
 
-        if messages[-1].get("role") == "assistant" and not self.config.interactive:
+        if messages[-1].get("role") == "assistant" and (
+            not self.config.interactive or self._is_anthropic()
+        ):
             messages.append({"role": "user", "content": "<meta>Continue the task.</meta>"})
 
         if self._is_anthropic() and self.config.enable_prompt_caching:
diff --git a/strix/tools/finish/finish_actions.py b/strix/tools/finish/finish_actions.py
index 79f48e79b..6a3037257 100644
--- a/strix/tools/finish/finish_actions.py
+++ b/strix/tools/finish/finish_actions.py
@@ -99,19 +99,30 @@ def finish_scan(
     if active_agents_error:
         return active_agents_error
 
-    validation_errors = []
-
-    if not executive_summary or not executive_summary.strip():
-        validation_errors.append("Executive summary cannot be empty")
-    if not methodology or not methodology.strip():
-        validation_errors.append("Methodology cannot be empty")
-    if not technical_analysis or not technical_analysis.strip():
-        validation_errors.append("Technical analysis cannot be empty")
-    if not recommendations or not recommendations.strip():
-        validation_errors.append("Recommendations cannot be empty")
-
-    if validation_errors:
-        return {"success": False, "message": "Validation failed", "errors": validation_errors}
+    _NOT_PROVIDED = "[Not provided by model]"
+    placeholder_fields = []
+    if not (executive_summary or "").strip():
+        placeholder_fields.append("executive_summary")
+    if not (methodology or "").strip():
+        placeholder_fields.append("methodology")
+    if not (technical_analysis or "").strip():
+        placeholder_fields.append("technical_analysis")
+    if not (recommendations or "").strip():
+        placeholder_fields.append("recommendations")
+
+    executive_summary = (executive_summary or "").strip() or _NOT_PROVIDED
+    methodology = (methodology or "").strip() or _NOT_PROVIDED
+    technical_analysis = (technical_analysis or "").strip() or _NOT_PROVIDED
+    recommendations = (recommendations or "").strip() or _NOT_PROVIDED
+
+    if placeholder_fields:
+        import logging
+
+        logging.warning(
+            "finish_scan: model omitted required field(s) %s; "
+            "saving partial report with placeholder text",
+            placeholder_fields,
+        )
 
     try:
         from strix.telemetry.tracer import get_global_tracer
diff --git a/tests/llm/test_prepare_messages.py b/tests/llm/test_prepare_messages.py
new file mode 100644
index 000000000..2566f8ee9
--- /dev/null
+++ b/tests/llm/test_prepare_messages.py
@@ -0,0 +1,50 @@
+"""Tests for LLM._prepare_messages trailing-assistant-message handling."""
+from strix.llm.config import LLMConfig
+from strix.llm.llm import LLM
+
+
+def _make_llm(monkeypatch, model_name: str, interactive: bool) -> LLM:
+    monkeypatch.setenv("STRIX_LLM", model_name)
+    config = LLMConfig(model_name=model_name, interactive=interactive, enable_prompt_caching=False)
+    return LLM(config, agent_name=None)
+
+
+def _history_ending_with_assistant() -> list[dict]:
+    return [
+        {"role": "user", "content": "Scan this target."},
+        {"role": "assistant", "content": "I found a vulnerability."},
+    ]
+
+
+def test_non_interactive_anthropic_adds_user_message(monkeypatch) -> None:
+    """Non-interactive mode always appends a user message when history ends with assistant."""
+    llm = _make_llm(monkeypatch, "claude-sonnet-4-6", interactive=False)
+    history = _history_ending_with_assistant()
+    messages = llm._prepare_messages(history)
+    assert messages[-1]["role"] == "user"
+    assert messages[-1]["content"] == "<meta>Continue the task.</meta>"
+
+
+def test_interactive_anthropic_adds_user_message(monkeypatch) -> None:
+    """Interactive mode with Anthropic model must also append a user message.
+
+    Anthropic API rejects messages where the last entry has role 'assistant'
+    (no assistant prefill support). This should hold regardless of interactive mode.
+    """
+    llm = _make_llm(monkeypatch, "claude-sonnet-4-6", interactive=True)
+    history = _history_ending_with_assistant()
+    messages = llm._prepare_messages(history)
+    assert messages[-1]["role"] == "user"
+    assert messages[-1]["content"] == "<meta>Continue the task.</meta>"
+
+
+def test_interactive_non_anthropic_does_not_add_user_message(monkeypatch) -> None:
+    """Interactive mode with a non-Anthropic model keeps the trailing assistant message.
+
+    Non-Anthropic models may support assistant prefill; in interactive mode the
+    caller (TUI) is responsible for appending the next user message.
+    """
+    llm = _make_llm(monkeypatch, "openai/gpt-5.4", interactive=True)
+    history = _history_ending_with_assistant()
+    messages = llm._prepare_messages(history)
+    assert messages[-1]["role"] == "assistant"