feat: 支持ms-swift格式数据

AGENTS.md

@@ -161,6 +161,7 @@ input_data = {
 |--------|-------------------|-------------|
 | JSON | `json` | Top-level array, each element is one sample |
 | JSONL | `jsonl` | One JSON object per line |
+| ms-swift | `ms_swift` | Agent trajectory JSONL; preserves raw fields and prepares selected evaluator fields |
 | CSV | `csv` | Comma-separated values with header |
 | Parquet | `parquet` | Apache Parquet columnar format |
 | Plain text | `plaintext` | One line per sample, mapped to `content` |
@@ -170,6 +171,22 @@ input_data = {
 
 MinerU formats automatically flatten blocks into `Data` objects, preserving all original fields (type, bbox, page_idx, etc.) via Pydantic `extra="allow"`. Use `dataset.mineru_config.include_types` to filter by block type.
 
+The `ms_swift` converter lives in `dingo/data/converter/ms_swift.py`. It prepares
+independent `<EvaluatorName>_<field>` fields for all 9 deterministic Agent rules
+and 7 Agent LLM metrics, only when selected in `evaluator[].evals`.
+Each deterministic rule owns its parsing and field assembly in its `prepare_*`
+function; LLM metrics continue sharing `_messages_content()`.
+Map these fields in a separate evaluator block for each metric. Initial LLM
+inputs contain only message `role`/`content`, with empty prompts; Plan Adherence
+also receives an empty context. Tool results are paired by call ID; execution
+status and duration extensions are copied only when explicitly present.
+See `docs/agent_trace/ms_swift.md` and `examples/dataset/ms_swift.json` for mappings
+and limitations.
+Each converted record privately tracks the fields actually generated by its
+selected preparers. Local result output removes these fields from a copy via
+`ResultInfo.without_derived_fields()`, preserving both existing output formats
+and conflict backups. No manually maintained cleanup list is needed.
+
 ### Execution Modes
 
 | Mode | Class | Use Case |

dingo/data/converter/__init__.py

@@ -1,4 +1,5 @@
 import dingo.data.converter.mineru  # noqa: F401  — registers mineru / mineru_v2
+import dingo.data.converter.ms_swift  # noqa: F401 — registers ms_swift
 from dingo.data.converter.base import BaseConverter
 
 converters = BaseConverter.converters

dingo/data/datasource/local.py

@@ -51,7 +51,7 @@ def _find_all_files(self, path: str, file_list: List[str]):
         for _f in os.listdir(path):
             f = os.path.join(path, _f)
             if os.path.isfile(f):
-                if getattr(self.input_args.dataset, "format", None) == "jsonl":
+                if getattr(self.input_args.dataset, "format", None) in ("jsonl", "ms_swift"):
                     if f.lower().endswith(".jsonl"):
                         file_list.append(f)
                 elif getattr(self.input_args.dataset, "format", None) == "json":

dingo/exec/local.py

@@ -90,7 +90,11 @@ def execute(self) -> SummaryModel:
                 futures_results = []
                 for data in batch:
                     dingo_id += 1
-                    r_i = ResultInfo(dingo_id = str(dingo_id), raw_data = data.to_dict())
+                    r_i = ResultInfo(
+                        dingo_id=str(dingo_id),
+                        raw_data=data.to_dict(),
+                        derived_fields=data._derived_fields,
+                    )
                     futures_results.append(r_i)
 
                     for e_p in self.input_args.evaluator:
@@ -419,6 +423,8 @@ def _should_write_to_file(self, input_args: InputArgs, file_path: str) -> bool:
 
     def _build_output_json(self, input_args: InputArgs, result_info: ResultInfo) -> str:
         field_list = self._resolve_field_list(input_args)
+        if input_args.dataset.format == "ms_swift":
+            result_info = result_info.without_derived_fields()
         if input_args.executor.result_save.raw:
             output_data = result_info.to_raw_dict(field_list=field_list)
         else:

dingo/io/input/data.py

@@ -1,6 +1,6 @@
-from typing import Any, Dict
+from typing import Any, Dict, Set
 
-from pydantic import BaseModel
+from pydantic import BaseModel, PrivateAttr
 
 
 class Data(BaseModel):
@@ -9,6 +9,9 @@ class Data(BaseModel):
     Flexible data structure that allows any fields to be configured.
     """
 
+    # Per-record converter metadata; never included in evaluator field mappings.
+    _derived_fields: Set[str] = PrivateAttr(default_factory=set)
+
     class Config:
         extra = "allow"
 

dingo/io/output/result_info.py

@@ -3,7 +3,7 @@
 from collections.abc import Mapping
 from datetime import date, datetime
 from decimal import Decimal
-from typing import Any, Dict, List, Optional
+from typing import Any, Dict, List, Optional, Set
 
 from pydantic import BaseModel, Field
 
@@ -23,6 +23,15 @@ class ResultInfo(BaseModel):
         default_factory=dict,
         exclude=True,
     )
+    derived_fields: Set[str] = Field(default_factory=set, exclude=True)
+
+    def without_derived_fields(self) -> "ResultInfo":
+        """Return an output copy without converter-generated top-level raw fields."""
+        cleaned_raw_data = {
+            key: value for key, value in self.raw_data.items()
+            if key not in self.derived_fields
+        }
+        return self.model_copy(update={"raw_data": cleaned_raw_data})
 
     @staticmethod
     def _eval_detail_to_dict(model_res: EvalDetail) -> Dict[str, Any]: