fix: improve JSON parsing error handling with bad_json mechanism

.vs/GenericAgent/v17/DocumentLayout.json

@@ -0,0 +1,12 @@
+{
+  "Version": 1,
+  "WorkspaceRootPath": "C:\\Users\\Ljq\\Documents\\mywork\\MyTools\\AutoOperation\\webagent\\GenericAgent\\",
+  "Documents": [],
+  "DocumentGroupContainers": [
+    {
+      "Orientation": 0,
+      "VerticalTabListWidth": 256,
+      "DocumentGroups": []
+    }
+  ]
+}

.vs/ProjectSettings.json

@@ -0,0 +1,3 @@
+{
+  "CurrentProjectSetting": null
+}

.vs/VSWorkspaceState.json

@@ -0,0 +1,6 @@
+{
+  "ExpandedNodes": [
+    ""
+  ],
+  "PreviewInSolutionExplorer": false
+}

agent_loop.py

@@ -24,6 +24,8 @@ class BaseHandler:
             ret = yield from try_call_generator(getattr(self, method_name), args, response)
             _ = yield from try_call_generator(self.tool_after_callback, tool_name, args, response, ret)
             return ret
+        elif tool_name == 'bad_json':
+            return StepOutcome(None, next_prompt=args.get('msg', 'bad_json'), should_exit=False)
         else:
             yield f"❌ 未知工具: {tool_name}\n"
             return StepOutcome(None, next_prompt=f"未知工具 {tool_name}", should_exit=False)

agentmain.py

@@ -30,15 +30,14 @@ class GeneraticAgent:
         if not os.path.exists('temp'): os.makedirs('temp')
         from sidercall import sider_cookie, oai_configs, claude_configs
         llm_sessions = []
+        for cfg in claude_configs.values():
+            llm_sessions += [ClaudeSession(api_key=cfg['apikey'], api_base=cfg['apibase'], model=cfg['model'])]
         if sider_cookie: llm_sessions += [SiderLLMSession(default_model=x) for x in \
                                     ["gemini-3.0-flash", "claude-haiku-4.5", "kimi-k2"]]
         for cfg in oai_configs.values():
             llm_sessions += [LLMSession(api_key=cfg['apikey'], api_base=cfg['apibase'], model=cfg['model'])]
-        for cfg in claude_configs.values():
-            llm_sessions += [ClaudeSession(api_key=cfg['apikey'], api_base=cfg['apibase'], model=cfg['model'])]
         if len(llm_sessions) > 0: 
-            llmclient = ToolClient(llm_sessions, auto_save_tokens=True)
+            self.llmclient = ToolClient(llm_sessions, auto_save_tokens=True)
-            self.llmclient = llmclient
         else:
             self.llmclient = None
         self.lock = threading.Lock()

ga.py

@@ -84,20 +84,10 @@ def code_run(code, code_type="python", timeout=60, cwd=None, code_cwd=None, stop
 
 
 def ask_user(question: str, candidates: list = None):
+    """question: 向用户提出的问题。candidates: 可选的候选项列表。需要保证should_exit为True
     """
-    构造一个中断请求。
+    return {"status": "INTERRUPT", "intent": "HUMAN_INTERVENTION",
-    question: 向用户提出的问题。
+        "data": {"question": question, "candidates": candidates or []}}
-    candidates: 可选的候选项列表。
-    需要保证should_exit为True
-    """
-    return {
-        "status": "INTERRUPT",
-        "intent": "HUMAN_INTERVENTION",
-        "data": {
-            "question": question,
-            "candidates": candidates or []
-        }
-    }
 
 from simphtml import execute_js_rich, get_html
 

sidercall.py

@@ -56,18 +56,18 @@ class GeminiSession:
         return iter([full_text]) if stream else full_text
   
 class ClaudeSession:
-    def __init__(self, api_key, api_base, model="claude-opus", context_win=32000):
+    def __init__(self, api_key, api_base, model="claude-opus", context_win=24000):
         self.api_key, self.api_base, self.default_model, self.context_win = api_key, api_base.rstrip('/'), model, context_win
         self.raw_msgs, self.lock = [], threading.Lock()
     def _trim_messages(self, messages):
         total = sum(len(m['prompt'])//4 for m in messages)
         if total <= self.context_win: return messages
-        trimmed = []
+        target, current, result = self.context_win * 0.9, 0, []
         for msg in reversed(messages):
-            if sum(len(m['prompt'])//4 for m in trimmed) + len(msg['prompt'])//4 <= self.context_win * 0.9:
+            if (msg_len := len(msg['prompt'])//4) + current <= target:
-                trimmed.insert(0, msg)
+                result.append(msg); current += msg_len
             else: break
-        return trimmed if trimmed else messages[-2:]
+        return result[::-1] or messages[-2:]
     def raw_ask(self, messages, model=None, temperature=0.5, max_tokens=4096):
         model = model or self.default_model
         headers = {"x-api-key": self.api_key, "Content-Type": "application/json"}
@@ -315,9 +315,9 @@ class ToolClient:
                 args = data.get('arguments') or data.get('args') or data.get('params') or data.get('parameters')
                 if args is None: args = data
                 if func_name: tool_calls = [MockToolCall(func_name, args)]
-            except json.JSONDecodeError:
+            except json.JSONDecodeError as e:
                 print("[Warn] Failed to parse tool_use JSON:", json_str)
-                remaining_text += f"[Warning] JSON 解析失败，模型输出了无效的 JSON."
+                tool_calls = [MockToolCall('bad_json', {'msg': f'Failed to parse tool_use JSON: {str(e)}'})]
             except Exception as e:
                 print("[Error] Exception during tool_use parsing:", str(e), data)