fix: use json.loads(strict=False) for all LLM output parsing paths

CHANGELOG.md

@@ -11,6 +11,7 @@ Only write entries that are worth mentioning to users.
 
 ## Unreleased
 
+- Core: Fix JSON parsing error when LLM tool call arguments contain unescaped control characters — use `json.loads(strict=False)` across all LLM output parsing paths to prevent tool execution failure and session corruption
 - Shell: Auto-trigger agent when background tasks complete while idle — the shell now detects when a background bash command or agent task finishes and automatically starts a new agent turn to process the results, instead of waiting for the user to type something
 - Core: Fix `QuestionRequest` hanging in print mode — `AskUserQuestion`, `EnterPlanMode`, and `ExitPlanMode` now auto-resolve when running in non-interactive (yolo) mode, preventing indefinite tool call hangs in `--print` sessions
 - Core: Fix background agent task output not visible until completion — `/task` browser and `TaskOutput` tool now show real-time output while background agent tasks are running, by tee-writing to the task log during execution instead of copying on completion

docs/en/release-notes/changelog.md

@@ -4,6 +4,7 @@ This page documents the changes in each Kimi Code CLI release.
 
 ## Unreleased
 
+- Core: Fix JSON parsing error when LLM tool call arguments contain unescaped control characters — use `json.loads(strict=False)` across all LLM output parsing paths to prevent tool execution failure and session corruption
 - Shell: Auto-trigger agent when background tasks complete while idle — the shell now detects when a background bash command or agent task finishes and automatically starts a new agent turn to process the results, instead of waiting for the user to type something
 - Core: Fix `QuestionRequest` hanging in print mode — `AskUserQuestion`, `EnterPlanMode`, and `ExitPlanMode` now auto-resolve when running in non-interactive (yolo) mode, preventing indefinite tool call hangs in `--print` sessions
 - Core: Fix background agent task output not visible until completion — `/task` browser and `TaskOutput` tool now show real-time output while background agent tasks are running, by tee-writing to the task log during execution instead of copying on completion

docs/zh/release-notes/changelog.md

@@ -4,6 +4,7 @@
 
 ## 未发布
 
+- Core：修复 LLM 工具调用参数包含未转义控制字符时 JSON 解析失败的问题——在所有 LLM 输出解析路径使用 `json.loads(strict=False)`，防止工具执行失败和会话永久损坏
 - Shell：空闲时自动响应后台任务完成——Shell 现在会检测后台 Bash 命令或 Agent 任务的完成，并自动发起新的 Agent 轮次处理结果，无需等待用户输入
 - Core：修复 Print 模式下 `QuestionRequest` 导致挂起的问题——`AskUserQuestion`、`EnterPlanMode` 和 `ExitPlanMode` 在非交互（yolo）模式下自动处理，避免 `--print` 会话中工具调用无限挂起
 - Core：修复后台 Agent 任务运行期间无法查看输出的问题——`/task` 浏览器和 `TaskOutput` 工具现在可实时显示后台 Agent 任务的输出，通过在执行期间同步写入任务日志替代完成后拷贝的方式实现

packages/kosong/CHANGELOG.md

@@ -2,6 +2,7 @@
 
 ## Unreleased
 
+- Core: Use `json.loads(strict=False)` when parsing tool call arguments to tolerate unescaped control characters from LLM output
 - Core: Treat `httpx.ProtocolError` as `APIConnectionError` in shared `convert_httpx_error()` mapping so streaming protocol disconnects now participate in existing retry logic
 - Anthropic: Fix `httpx.ReadTimeout` leaking through `_convert_stream_response` during streaming — the exception is now caught and converted to `APITimeoutError`, enabling retry logic that was previously bypassed
 - Anthropic: Fix `_convert_error` ordering — `AnthropicAPITimeoutError` is now checked before `AnthropicAPIConnectionError` to avoid misclassification due to inheritance

packages/kosong/src/kosong/contrib/chat_provider/anthropic.py

@@ -340,7 +340,7 @@ def _convert_message(self, message: Message) -> MessageParam:
         for tool_call in message.tool_calls or []:
             if tool_call.function.arguments:
                 try:
-                    parsed_arguments = json.loads(tool_call.function.arguments)
+                    parsed_arguments = json.loads(tool_call.function.arguments, strict=False)
                 except json.JSONDecodeError as exc:  # pragma: no cover - defensive guard
                     raise ChatProviderError("Tool call arguments must be valid JSON.") from exc
                 if not isinstance(parsed_arguments, dict):

packages/kosong/src/kosong/contrib/chat_provider/google_genai.py

@@ -651,7 +651,7 @@ def message_to_google_genai(message: Message) -> Content:
     for tool_call in message.tool_calls or []:
         if tool_call.function.arguments:
             try:
-                parsed_arguments = json.loads(tool_call.function.arguments)
+                parsed_arguments = json.loads(tool_call.function.arguments, strict=False)
             except json.JSONDecodeError as exc:  # pragma: no cover - defensive guard
                 raise ChatProviderError("Tool call arguments must be valid JSON.") from exc
             if not isinstance(parsed_arguments, dict):

packages/kosong/src/kosong/contrib/context/linear.py

@@ -96,7 +96,7 @@ def _restore():
                 for line in f:
                     if not line.strip():
                         continue
-                    line_json = json.loads(line)
+                    line_json = json.loads(line, strict=False)
                     if "token_count" in line_json:
                         self._token_count = line_json["token_count"]
                         continue

packages/kosong/src/kosong/tooling/simple.py

@@ -119,7 +119,7 @@ def handle(self, tool_call: ToolCall) -> HandleResult:
         tool = self._tool_dict[tool_call.function.name]
 
         try:
-            arguments: JsonType = json.loads(tool_call.function.arguments or "{}")
+            arguments: JsonType = json.loads(tool_call.function.arguments or "{}", strict=False)
         except json.JSONDecodeError as e:
             return ToolResult(tool_call_id=tool_call.id, return_value=ToolParseError(str(e)))
 

src/kimi_cli/session.py

@@ -69,7 +69,7 @@ def is_empty(self) -> bool:
                     line = line.strip()
                     if not line:
                         continue
-                    role = json.loads(line).get("role")
+                    role = json.loads(line, strict=False).get("role")
                     if isinstance(role, str) and not role.startswith("_"):
                         return False
         except FileNotFoundError:

src/kimi_cli/soul/context.py

@@ -39,7 +39,7 @@ async def restore(self) -> bool:
             async for line in f:
                 if not line.strip():
                     continue
-                line_json = json.loads(line)
+                line_json = json.loads(line, strict=False)
                 if line_json["role"] == "_system_prompt":
                     self._system_prompt = line_json["content"]
                     continue
@@ -160,7 +160,7 @@ async def revert_to(self, checkpoint_id: int):
                 if not line.strip():
                     continue
 
-                line_json = json.loads(line)
+                line_json = json.loads(line, strict=False)
                 if line_json["role"] == "_checkpoint" and line_json["id"] == checkpoint_id:
                     break
 

src/kimi_cli/soul/toolset.py

@@ -122,7 +122,7 @@ def handle(self, tool_call: ToolCall) -> HandleResult:
             tool = self._tool_dict[tool_call.function.name]
 
             try:
-                arguments: JsonType = json.loads(tool_call.function.arguments or "{}")
+                arguments: JsonType = json.loads(tool_call.function.arguments or "{}", strict=False)
             except json.JSONDecodeError as e:
                 return ToolResult(tool_call_id=tool_call.id, return_value=ToolParseError(str(e)))
 

src/kimi_cli/tools/__init__.py

@@ -20,7 +20,7 @@ def extract_key_argument(json_content: str | streamingjson.Lexer, tool_name: str
     else:
         json_str = json_content
     try:
-        curr_args: JsonType = json.loads(json_str)
+        curr_args: JsonType = json.loads(json_str, strict=False)
     except json.JSONDecodeError:
         return None
     if not curr_args:

src/kimi_cli/ui/shell/debug.py

@@ -70,7 +70,7 @@ def _format_tool_call(tool_call: ToolCall) -> Panel:
     """Format a tool call."""
     args = tool_call.function.arguments or "{}"
     try:
-        args_formatted = json.dumps(json.loads(args), indent=2)
+        args_formatted = json.dumps(json.loads(args, strict=False), indent=2)
         args_syntax = Syntax(args_formatted, "json", theme="monokai", padding=(0, 1))
     except json.JSONDecodeError:
         args_syntax = Text(args, style="red")

src/kimi_cli/ui/shell/visualize.py

@@ -332,7 +332,7 @@ def _extract_full_url(arguments: str | None, tool_name: str) -> str | None:
         if tool_name != "FetchURL" or not arguments:
             return None
         try:
-            args = json.loads(arguments)
+            args = json.loads(arguments, strict=False)
         except (json.JSONDecodeError, TypeError):
             return None
         if isinstance(args, dict):

src/kimi_cli/utils/export.py

@@ -55,7 +55,7 @@ def _extract_tool_call_hint(args_json: str) -> str:
     short string value.  Returns ``""`` when nothing useful is found.
     """
     try:
-        parsed: object = json.loads(args_json)
+        parsed: object = json.loads(args_json, strict=False)
     except (json.JSONDecodeError, TypeError):
         return ""
     if not isinstance(parsed, dict):
@@ -104,7 +104,8 @@ def _format_tool_call_md(tool_call: ToolCall) -> str:
         title += f" (`{hint}`)"
 
     try:
-        args_formatted = json.dumps(json.loads(args_raw), indent=2, ensure_ascii=False)
+        parsed = json.loads(args_raw, strict=False)
+        args_formatted = json.dumps(parsed, indent=2, ensure_ascii=False)
     except json.JSONDecodeError:
         args_formatted = args_raw
 
@@ -404,7 +405,7 @@ def _stringify_tool_calls(tool_calls: Sequence[ToolCall]) -> str:
     for tc in tool_calls:
         args_raw = tc.function.arguments or "{}"
         try:
-            args = json.loads(args_raw)
+            args = json.loads(args_raw, strict=False)
             args_str = json.dumps(args, ensure_ascii=False)
         except (json.JSONDecodeError, TypeError):
             args_str = args_raw