Merge branch 'fixbug-incorrectly-escaped-commands' into 'mgx_ops'

Fixbug: 指令解析时遇到转义字符解析错误

See merge request pub/MetaGPT!298

metagpt/roles/di/role_zero.py

@@ -35,7 +35,11 @@ from metagpt.tools.libs.editor import Editor
 from metagpt.tools.tool_recommend import BM25ToolRecommender, ToolRecommender
 from metagpt.tools.tool_registry import register_tool
 from metagpt.utils.common import CodeParser, any_to_str
-from metagpt.utils.repair_llm_raw_output import RepairType, repair_llm_raw_output
+from metagpt.utils.repair_llm_raw_output import (
+    RepairType,
+    repair_escape_error,
+    repair_llm_raw_output,
+)
 from metagpt.utils.report import ThoughtReporter
 
 
@@ -315,10 +319,20 @@ class RoleZero(Role):
             if commands.endswith("]") and not commands.startswith("["):
                 commands = "[" + commands
             commands = json.loads(repair_llm_raw_output(output=commands, req_keys=[None], repair_type=RepairType.JSON))
-        except json.JSONDecodeError:
+        except json.JSONDecodeError as e:
             logger.warning(f"Failed to parse JSON for: {self.command_rsp}. Trying to repair...")
-            commands = await self.llm.aask(msg=JSON_REPAIR_PROMPT.format(json_data=self.command_rsp))
-            commands = json.loads(CodeParser.parse_code(block=None, lang="json", text=commands))
+            commands = await self.llm.aask(
+                msg=JSON_REPAIR_PROMPT.format(json_data=self.command_rsp, json_decode_error=str(e))
+            )
+            try:
+                commands = json.loads(CodeParser.parse_code(block=None, lang="json", text=commands))
+            except json.JSONDecodeError:
+                # repair escape error of code and math
+                commands = CodeParser.parse_code(block=None, lang="json", text=self.command_rsp)
+                new_command = repair_escape_error(commands)
+                commands = json.loads(
+                    repair_llm_raw_output(output=new_command, req_keys=[None], repair_type=RepairType.JSON)
+                )
         except Exception as e:
             tb = traceback.format_exc()
             print(tb)

metagpt/utils/repair_llm_raw_output.py

@@ -347,3 +347,44 @@ def extract_state_value_from_output(content: str) -> str:
     matches = list(set(matches))
     state = matches[0] if len(matches) > 0 else "-1"
     return state
+
+
+def repair_escape_error(commands):
+    """
+    Repaires escape errors in command responses.
+    When RoleZero parses a command, the command may contain unknown escape characters.
+
+    This function has two steps:
+    1. Transform unescaped substrings like "\d" and "\(" to "\\\\d" and "\\\\(".
+    2. Transform escaped characters like '\f' to substrings like "\\\\f".
+
+    Example:
+        When the original JSON string is " {"content":"\\\\( \\\\frac{1}{2} \\\\)"} ",
+        The "content" will be parsed correctly to "\( \frac{1}{2} \)".
+
+        However, if the original JSON string is " {"content":"\( \frac{1}{2} \)"}" directly.
+        It will cause a parsing error.
+
+        To repair the wrong JSON string, the following transformations will be used:
+        "\("   --->  "\\\\("
+        '\f'   --->  "\\\\f"
+        "\)"   --->  "\\\\)"
+
+    """
+    escape_repair_map = {
+        "\a": "\\\\a",
+        "\b": "\\\\b",
+        "\f": "\\\\f",
+        "\r": "\\\\r",
+        "\t": "\\\\t",
+        "\v": "\\\\v",
+    }
+    new_command = ""
+    for index, ch in enumerate(commands):
+        if ch == "\\" and index + 1 < len(commands):
+            if commands[index + 1] not in ["n", '"', " "]:
+                new_command += "\\"
+        elif ch in escape_repair_map:
+            ch = escape_repair_map[ch]
+        new_command += ch
+    return new_command