From c36477c9cb241b8db39d46041d206a869754198f Mon Sep 17 00:00:00 2001
From: yumoqing <yumoqing@gmail.com>
Date: Thu, 7 May 2026 14:51:44 +0800
Subject: [PATCH] fix: improve LLM call error handling and plan parsing

- Add detailed logging to _llm_call method
- Improve _parse_plan_json to handle more LLM response formats
- Show LLM error messages in reasoning_submit.dspy
- Better error handling and fallback for JSON parsing
---
 harnessed_reasoning/core.py       | 62 ++++++++++++++++++++++++-------
 wwwroot/api/reasoning_submit.dspy | 14 ++++++-
 2 files changed, 60 insertions(+), 16 deletions(-)

diff --git a/harnessed_reasoning/core.py b/harnessed_reasoning/core.py
index 5d80ad5..e518c44 100644
--- a/harnessed_reasoning/core.py
+++ b/harnessed_reasoning/core.py
@@ -258,17 +258,23 @@ class HermesReasoningEngine:
         # Use harnessed_agent's llm_chat if available
         env = ServerEnv()
         if hasattr(env, 'llm_chat'):
-            return await env.llm_chat(
-                messages=messages,
-                model=model,
-                temperature=temperature,
-                max_tokens=max_tokens,
-                **extra,
-            )
-
-        # Fallback: direct config-based call (shouldn't happen in production)
-        error("llm_chat not available on ServerEnv")
-        return {'error': {'message': 'LLM client not available'}}
+            info(f"Calling llm_chat: model={model}, temp={temperature}, max_tokens={max_tokens}")
+            try:
+                result = await env.llm_chat(
+                    messages=messages,
+                    model=model,
+                    temperature=temperature,
+                    max_tokens=max_tokens,
+                    **extra,
+                )
+                info(f"llm_chat returned: success={'error' not in result}")
+                return result
+            except Exception as e:
+                error(f"llm_chat exception: {e}")
+                return {'error': {'message': f'llm_chat exception: {str(e)}', 'type': 'exception'}}
+        else:
+            error("llm_chat not available on ServerEnv")
+            return {'error': {'message': 'LLM client not available (llm_chat not registered)', 'type': 'configuration_error'}}
 
     async def _generate_plan(self, request: str, context: Dict[str, Any],
                              config: Dict[str, Any]) -> Dict[str, Any]:
@@ -321,8 +327,9 @@ class HermesReasoningEngine:
 
     def _parse_plan_json(self, text: str) -> Dict[str, Any]:
         """Extract and parse JSON plan from LLM response."""
-        # Try to find JSON block
         import re
+        
+        # Try to find JSON block in code fence
         json_match = re.search(r'```json\s*\n(.*?)\n```', text, re.DOTALL)
         if json_match:
             text = json_match.group(1)
@@ -340,10 +347,37 @@ class HermesReasoningEngine:
             plan.setdefault('analysis', '')
             plan.setdefault('steps', [])
             plan.setdefault('safety_notes', [])
+            
+            # Validate steps structure
+            valid_steps = []
+            for i, step in enumerate(plan.get('steps', [])):
+                if isinstance(step, dict):
+                    step.setdefault('step_number', i + 1)
+                    step.setdefault('description', '')
+                    step.setdefault('actions', [])
+                    valid_steps.append(step)
+                elif isinstance(step, str):
+                    # Handle case where LLM returns list of strings
+                    valid_steps.append({
+                        'step_number': i + 1,
+                        'description': step,
+                        'actions': []
+                    })
+            plan['steps'] = valid_steps
+            
             return plan
         except json.JSONDecodeError:
-            error(f"Failed to parse LLM plan JSON: {text[:200]}")
-            return {'analysis': '无法解析 LLM 返回的计划', 'steps': [], 'safety_notes': []}
+            error(f"Failed to parse LLM plan JSON: {text[:500]}")
+            # Fallback: create a simple plan from the response text
+            return {
+                'analysis': text[:200],
+                'steps': [{
+                    'step_number': 1,
+                    'description': text[:100],
+                    'actions': []
+                }],
+                'safety_notes': []
+            }
 
     # --------------------------------------------------------
     # Tool execution
diff --git a/wwwroot/api/reasoning_submit.dspy b/wwwroot/api/reasoning_submit.dspy
index 25e0820..952c585 100644
--- a/wwwroot/api/reasoning_submit.dspy
+++ b/wwwroot/api/reasoning_submit.dspy
@@ -21,31 +21,41 @@ try:
         
         if reasoning_result.get('success'):
             # Build result widget showing reasoning output
+            analysis = reasoning_result.get('analysis', '无')
+            llm_error = reasoning_result.get('llm_error', '')
+            
             plan_items = []
             for step in reasoning_result.get('execution_plan', []):
                 plan_items.append(f"步骤{step.get('step_number', '?')}: {step.get('description', '')}")
             
             plan_text = '\n'.join(plan_items) if plan_items else '无执行计划'
             safety_text = '\n'.join(reasoning_result.get('safety_violations', [])) or '无安全风险'
+            status = reasoning_result.get('status', 'unknown')
             
             summary = (
                 f"请求: {request_text}\n\n"
+                f"分析: {analysis}\n\n"
                 f"上下文: {reasoning_result.get('context_summary', '无')}\n\n"
                 f"置信度: {reasoning_result.get('confidence_score', 0):.0%}\n\n"
                 f"执行计划:\n{plan_text}\n\n"
                 f"安全检查:\n{safety_text}\n\n"
-                f"状态: {reasoning_result.get('status', 'unknown')}"
+                f"状态: {status}"
             )
             
+            # Show LLM error if any
+            if llm_error:
+                summary += f"\n\nLLM 错误: {llm_error}"
+            
             if reasoning_result.get('execution_results'):
                 summary += f"\n\n执行结果:\n{json.dumps(reasoning_result['execution_results'], ensure_ascii=False, indent=2)}"
             
+            msg_type = 'success' if not llm_error else 'warning'
             result = {
                 'widgettype': 'Message',
                 'options': {
                     'title': '推理完成',
                     'message': summary,
-                    'type': 'success'
+                    'type': msg_type
                 }
             }
         else: