# -*- coding: utf-8 -*-
"""AgentScope bindings for tool specifications."""
from __future__ import annotations
import json
import os
from typing import Any, Callable, Dict
from pydantic import ValidationError
from data_juicer_agents.core.tool import ToolContext, ToolResult, ToolSpec
from data_juicer_agents.utils.runtime_helpers import (
to_bool,
to_text_response,
truncate_text,
)
from .schema_utils import normalize_tool_schema
_MODEL_TEXT_LIMIT = 2400
_MODEL_LIST_LIMIT = 12
_MODEL_OPERATOR_LIMIT = 10
_MODEL_PARAM_LIMIT = 20
[文档]
def build_agentscope_json_schema(spec: ToolSpec) -> Dict[str, Any]:
parameters = normalize_tool_schema(spec.input_model.model_json_schema())
return {
"type": "function",
"function": {
"name": spec.name,
"description": spec.description,
"parameters": parameters,
},
}
def _preview_value(value: Any, *, limit: int = 800) -> Any:
if isinstance(value, str):
return truncate_text(value, limit=limit)
if isinstance(value, (dict, list)):
try:
return truncate_text(json.dumps(value, ensure_ascii=False), limit=limit)
except Exception:
return truncate_text(str(value), limit=limit)
return value
[文档]
def default_arg_preview(_spec: ToolSpec, raw_kwargs: Dict[str, Any]) -> Dict[str, Any]:
return {key: _preview_value(value) for key, value in raw_kwargs.items()}
def _short_text(value: Any, *, limit: int = _MODEL_TEXT_LIMIT) -> str:
return truncate_text(str(value or ""), limit=limit).strip()
def _compact_list(items: Any, *, limit: int = _MODEL_LIST_LIMIT) -> list[Any]:
return list(items[:limit]) if isinstance(items, list) else []
def _compact_mapping(value: Any, *, text_limit: int = 400) -> Any:
if isinstance(value, str):
return _short_text(value, limit=text_limit)
if isinstance(value, list):
return [
_compact_mapping(item, text_limit=text_limit)
for item in value[:_MODEL_LIST_LIMIT]
]
if isinstance(value, dict):
return {
str(key): _compact_mapping(item, text_limit=text_limit)
for key, item in value.items()
}
return value
def _compact_operator(item: Any, *, include_parameters: bool = False) -> Dict[str, Any]:
if not isinstance(item, dict):
return {"value": _short_text(item, limit=300)}
compact = {
key: item.get(key)
for key in (
"operator_name",
"resolved_name",
"operator_type",
"score",
"source",
"source_path",
"test_path",
)
if item.get(key) not in (None, "")
}
if item.get("tags"):
compact["tags"] = _compact_list(item.get("tags"), limit=8)
if item.get("description"):
compact["description"] = _short_text(item.get("description"), limit=360)
if include_parameters and isinstance(item.get("parameters"), list):
compact["parameters"] = [
_compact_mapping(param, text_limit=260)
for param in item["parameters"][:_MODEL_PARAM_LIMIT]
if isinstance(param, dict)
]
return compact
def _base_payload(payload: Dict[str, Any]) -> Dict[str, Any]:
base: Dict[str, Any] = {}
for key in ("ok", "action", "message", "error_type"):
if key in payload:
base[key] = payload[key]
for key in ("warnings", "validation_errors", "requires"):
if isinstance(payload.get(key), list):
base[key] = _compact_list(payload.get(key), limit=8)
return base
def _recipe_summary(recipe: Any) -> Dict[str, Any]:
if not isinstance(recipe, dict):
return {}
process = recipe.get("process", [])
operators = []
if isinstance(process, list):
for step in process[:_MODEL_OPERATOR_LIMIT]:
if isinstance(step, dict) and step:
operators.extend(str(name) for name in step.keys())
return {
"dataset_path": recipe.get("dataset_path"),
"export_path": recipe.get("export_path"),
"text_keys": recipe.get("text_keys"),
"image_key": recipe.get("image_key"),
"audio_key": recipe.get("audio_key"),
"video_key": recipe.get("video_key"),
"operator_names": operators,
"np": recipe.get("np"),
"executor_type": recipe.get("executor_type"),
}
[文档]
def compact_payload_for_model(
tool_name: str, payload: Dict[str, Any]
) -> Dict[str, Any]:
"""Return a semantically faithful, smaller payload for ReAct memory.
The full payload is still passed through runtime events; this function only
controls what is written into AgentScope tool_result memory.
"""
if not isinstance(payload, dict):
return {"ok": True, "result": _compact_mapping(payload)}
if payload.get("ok") is False:
compact = _base_payload(payload)
for key in ("error", "error_message", "stderr", "stdout"):
if payload.get(key):
compact[key] = _short_text(payload.get(key), limit=1200)
return compact
compact = _base_payload(payload)
if tool_name == "inspect_dataset":
for key in (
"dataset",
"inspected_path",
"sampled_records",
"scanned_lines",
"modality",
"keys",
"candidate_text_keys",
"candidate_image_keys",
):
if key in payload:
compact[key] = _compact_mapping(payload[key], text_limit=300)
if isinstance(payload.get("key_stats"), dict):
compact["key_stats"] = {
str(key): _compact_mapping(value, text_limit=180)
for key, value in payload["key_stats"].items()
}
if isinstance(payload.get("sample_preview"), list):
compact["sample_preview"] = [
_compact_mapping(item, text_limit=180)
for item in payload["sample_preview"][:2]
]
return compact
if tool_name in {"retrieve_operators", "retrieve_operators_api"}:
for key in ("intent", "mode", "source", "candidate_names", "requested_tags"):
if key in payload:
compact[key] = _compact_mapping(payload[key], text_limit=240)
if isinstance(payload.get("candidates"), list):
compact["candidates"] = [
_compact_operator(item)
for item in payload["candidates"][:_MODEL_OPERATOR_LIMIT]
]
compact["candidate_count"] = len(payload["candidates"])
return compact
if tool_name == "list_operator_catalog":
for key in (
"total_count",
"returned_count",
"op_type_filter",
"requested_tags",
"include_parameters",
"limit",
):
if key in payload:
compact[key] = payload[key]
if isinstance(payload.get("operators"), list):
include_parameters = bool(payload.get("include_parameters"))
compact["operators"] = [
_compact_operator(item, include_parameters=include_parameters)
for item in payload["operators"][:_MODEL_OPERATOR_LIMIT]
]
compact["compacted_count"] = len(compact["operators"])
return compact
if tool_name == "get_operator_info":
for key in (
"requested_name",
"resolved_name",
"resolved",
"exact_match",
"operator_type",
"tags",
"source_path",
"test_path",
):
if key in payload:
compact[key] = _compact_mapping(payload[key], text_limit=260)
if payload.get("description"):
compact["description"] = _short_text(payload.get("description"), limit=500)
if isinstance(payload.get("parameters"), list):
compact["parameters"] = [
_compact_mapping(param, text_limit=300)
for param in payload["parameters"][:_MODEL_PARAM_LIMIT]
]
return compact
if tool_name == "assemble_plan" and isinstance(payload.get("plan"), dict):
# Keep the exact plan for follow-up plan_validate / plan_save calls.
compact.update(
{
"plan": payload["plan"],
"plan_id": payload.get("plan_id"),
"operator_names": payload.get("operator_names"),
"modality": payload.get("modality"),
"plan_summary": _recipe_summary(payload["plan"].get("recipe")),
}
)
return compact
if tool_name in {"execute_bash", "execute_python_code"}:
for key in ("returncode", "timeout_sec", "command", "message"):
if key in payload:
compact[key] = payload[key]
for key in ("stdout", "stderr"):
if payload.get(key):
compact[key] = _short_text(payload.get(key), limit=1800)
return compact
if tool_name == "view_text_file":
for key, value in payload.items():
if key in {"content", "text"}:
compact[key] = _short_text(value, limit=1800)
else:
compact[key] = value
return compact
return _compact_mapping(payload, text_limit=800)
__all__ = [
"build_agentscope_json_schema",
"build_agentscope_tool_function",
"compact_payload_for_model",
"default_arg_preview",
"invoke_tool_spec",
"tool_result_compaction_enabled",
]
