R3.2.8.6-R1 Reasoning-Only Exhaustion Recovery

core/pipeline_agent.py

@@ -44,6 +44,10 @@ class DeepSeekRequestError(DeepSeekAPIError):
     pass
 
 
+class DeepSeekReasoningOnlyExhaustionError(DeepSeekAPIError):
+    """Reasoning tokens were emitted but no final answer content arrived."""
+
+
 class PipelineCancelledError(RuntimeError):
     """Raised when the Worker or local task state requests cancellation."""
 
@@ -869,6 +873,7 @@ class SSciPipelineAgent:
         reasoning_effort=None,
         temp=0.3,
         allow_fallback=None,
+        reasoning_only_recovery=False,
     ):
         (
             requested_model,
@@ -1032,6 +1037,11 @@ class SSciPipelineAgent:
                         yield "\n</think>\n"
 
                 if not emitted_content:
+                    if reasoning_only_recovery and reasoning_chunks > 0:
+                        raise DeepSeekReasoningOnlyExhaustionError(
+                            "DeepSeek reasoning-only exhaustion: "
+                            f"reasoning_chunks={reasoning_chunks}; final content absent."
+                        )
                     raise DeepSeekAPIError(
                         "DeepSeek stream ended without final content."
                     )
@@ -1048,6 +1058,9 @@ class SSciPipelineAgent:
 
             except GeneratorExit:
                 raise
+            except DeepSeekReasoningOnlyExhaustionError:
+                self._touch_execution_progress("deepseek_reasoning_only_exhaustion")
+                raise
             except (
                 DeepSeekAuthenticationError,
                 DeepSeekBalanceError,
@@ -1192,13 +1205,40 @@ class SSciPipelineAgent:
         user_prompt = APP_CONFIG["prompt_templates"]["synthesis_user"].replace(
             "{topic}", topic).replace("{user_injection}", user_injection).replace(
             "{refs_text}", refs_text).replace("{directive}", APP_CONFIG['deep_synthesis_directive'])
-        return self.stream_deepseek(
-            sys_prompt,
-            user_prompt,
-            model_role="reasoning",
-            thinking=True,
-            allow_fallback=False,
-        )
+
+        def _stream_with_reasoning_exhaustion_recovery():
+            try:
+                yield from self.stream_deepseek(
+                    sys_prompt,
+                    user_prompt,
+                    model_role="reasoning",
+                    thinking=True,
+                    allow_fallback=False,
+                    reasoning_only_recovery=True,
+                )
+                return
+            except DeepSeekReasoningOnlyExhaustionError as exc:
+                logging.warning(
+                    "Literature synthesis reasoning-only exhaustion; retrying once with thinking disabled: %s",
+                    exc,
+                )
+                self._touch_execution_progress("literature_synthesis_direct_recovery")
+
+            recovery_system = (
+                sys_prompt
+                + "\n\nRECOVERY MODE: Produce the final literature synthesis directly. "
+                  "Do not perform extended hidden reasoning. Preserve the original evidence, "
+                  "citation boundaries, uncertainty, and scope; do not invent facts."
+            )
+            yield from self.stream_deepseek(
+                recovery_system,
+                user_prompt,
+                model_role="reasoning",
+                thinking=False,
+                allow_fallback=False,
+            )
+
+        return _stream_with_reasoning_exhaustion_recovery()
 
     # Generate abstract and conclusion anchors from literature plus evidence.
     def stream_generate_research_logic(