satif-ai 0.2.8__py3-none-any.whl → 0.2.10__py3-none-any.whl

satif_ai/utils/merge_sdif.py

@@ -0,0 +1,535 @@
+import json
+import logging
+import shutil
+import sqlite3
+from datetime import datetime
+from pathlib import Path
+from typing import Any, Dict, List
+
+from satif_core.types import SDIFPath
+from sdif_db.database import (
+    SDIFDatabase,  # Assuming this is the conventional import path
+)
+
+log = logging.getLogger(__name__)
+
+
+class _SDIFMerger:
+    def __init__(self, target_sdif_path: Path):
+        self.target_db = SDIFDatabase(target_sdif_path, overwrite=True)
+        # Mappings per source_db_idx:
+        self.source_id_map: Dict[int, Dict[int, int]] = {}
+        self.table_name_map: Dict[int, Dict[str, str]] = {}
+        self.object_name_map: Dict[int, Dict[str, str]] = {}
+        self.media_name_map: Dict[int, Dict[str, str]] = {}
+
+    def _get_new_source_id(self, source_db_idx: int, old_source_id: int) -> int:
+        return self.source_id_map[source_db_idx][old_source_id]
+
+    def _get_new_table_name(self, source_db_idx: int, old_table_name: str) -> str:
+        return self.table_name_map[source_db_idx].get(old_table_name, old_table_name)
+
+    def _get_new_object_name(self, source_db_idx: int, old_object_name: str) -> str:
+        return self.object_name_map[source_db_idx].get(old_object_name, old_object_name)
+
+    def _get_new_media_name(self, source_db_idx: int, old_media_name: str) -> str:
+        return self.media_name_map[source_db_idx].get(old_media_name, old_media_name)
+
+    def _generate_unique_name_in_target(self, base_name: str, list_func) -> str:
+        """Generates a unique name for the target DB by appending a suffix if base_name exists."""
+        if base_name not in list_func():
+            return base_name
+        i = 1
+        while True:
+            new_name = f"{base_name}_{i}"
+            if new_name not in list_func():
+                return new_name
+            i += 1
+            if i > 1000:  # Safety break
+                raise RuntimeError(
+                    f"Could not generate a unique name for base '{base_name}' after 1000 attempts."
+                )
+
+    def _merge_properties(self, source_db: SDIFDatabase, source_db_idx: int):
+        source_props = source_db.get_properties()
+        if not source_props:
+            log.warning(
+                f"Source database {source_db.path} has no properties. Skipping properties merge for this source."
+            )
+            return
+
+        if source_props.get("sdif_version") != "1.0":
+            # Or allow a configurable expected version
+            raise ValueError(
+                f"Source database {source_db.path} has unsupported SDIF version: {source_props.get('sdif_version')}. Expected '1.0'."
+            )
+
+        if source_db_idx == 0:  # First database sets the version for the target
+            try:
+                self.target_db.conn.execute(
+                    "UPDATE sdif_properties SET sdif_version = ?",
+                    (
+                        source_props.get("sdif_version", "1.0"),
+                    ),  # Default to 1.0 if somehow missing
+                )
+                self.target_db.conn.commit()
+            except sqlite3.Error as e:
+                log.error(
+                    f"Failed to set sdif_version in target DB from {source_db.path}: {e}"
+                )
+                raise
+        # creation_timestamp will be set at the end of the entire merge process.
+
+    def _merge_sources(self, source_db: SDIFDatabase, source_db_idx: int):
+        self.source_id_map[source_db_idx] = {}
+        source_sources = source_db.list_sources()
+        for old_source_entry in source_sources:
+            old_source_id = old_source_entry["source_id"]
+            new_source_id = self.target_db.add_source(
+                file_name=old_source_entry["original_file_name"],
+                file_type=old_source_entry["original_file_type"],
+                description=old_source_entry.get("source_description"),
+            )
+            # original processing_timestamp is not directly carried over, new one is set by add_source
+            self.source_id_map[source_db_idx][old_source_id] = new_source_id
+
+    def _merge_tables(self, source_db: SDIFDatabase, source_db_idx: int):
+        self.table_name_map[source_db_idx] = {}
+        source_schema = source_db.get_schema()
+        source_tables_schema = source_schema.get("tables", {})
+
+        # Pass 1: Determine new table names for all tables from this source DB
+        # This is to ensure FKs can be remapped correctly to tables from the *same* source db.
+        temp_name_map_for_this_source = {}
+        for old_table_name in source_tables_schema.keys():
+            # Use create_table with if_exists='add' to get a unique name, but only for name generation.
+            # This is a bit of a workaround. A dedicated _generate_unique_target_table_name might be cleaner.
+            # The SDIFDatabase.create_table(if_exists='add') will actually create metadata entries.
+            # This might be acceptable if we're careful.
+            # Let's use the simpler approach of generating unique name first.
+            effective_new_name = self._generate_unique_name_in_target(
+                old_table_name, self.target_db.list_tables
+            )
+            temp_name_map_for_this_source[old_table_name] = effective_new_name
+        self.table_name_map[source_db_idx] = temp_name_map_for_this_source
+
+        # Pass 2: Create tables with remapped FKs and copy data
+        for old_table_name, table_detail_from_schema in source_tables_schema.items():
+            new_table_name = self.table_name_map[source_db_idx][old_table_name]
+
+            columns_for_create: Dict[str, Dict[str, Any]] = {}
+            original_columns_detail = table_detail_from_schema.get("columns", [])
+
+            for col_detail in original_columns_detail:
+                col_name = col_detail["name"]
+                col_props = {
+                    "type": col_detail["sqlite_type"],
+                    "not_null": col_detail["not_null"],
+                    "primary_key": col_detail[
+                        "primary_key"
+                    ],  # Assumes single col PK flag
+                    "description": col_detail.get("description"),
+                    "original_column_name": col_detail.get("original_column_name"),
+                    # 'unique' constraint not in get_schema output, assumed not used or handled by primary_key
+                }
+
+                # Remap foreign keys defined for this column
+                table_fks_detail = table_detail_from_schema.get("foreign_keys", [])
+                for fk_info in table_fks_detail:
+                    if fk_info["from_column"] == col_name:
+                        original_fk_target_table = fk_info["target_table"]
+                        # FKs are assumed to target tables within the same source SDIF file.
+                        remapped_fk_target_table = self.table_name_map[
+                            source_db_idx
+                        ].get(original_fk_target_table)
+                        if not remapped_fk_target_table:
+                            log.warning(
+                                f"Could not remap FK target table '{original_fk_target_table}' for column '{col_name}' in table '{old_table_name}'. FK might be dropped or invalid."
+                            )
+                            # Decide: skip FK, or raise error, or create FK pointing to original name (which might conflict or be wrong)
+                            # For now, we'll proceed without this FK if target not found in map (shouldn't happen if all tables from source are processed)
+                            continue
+
+                        col_props["foreign_key"] = {
+                            "table": remapped_fk_target_table,
+                            "column": fk_info["target_column"],
+                            "on_delete": fk_info[
+                                "on_delete"
+                            ].upper(),  # Ensure standard casing
+                            "on_update": fk_info[
+                                "on_update"
+                            ].upper(),  # Ensure standard casing
+                        }
+                        break  # Assuming one FK per 'from_column' for this col_props structure
+                columns_for_create[col_name] = col_props
+
+            source_table_metadata = table_detail_from_schema.get("metadata", {})
+            old_source_id_for_table = source_table_metadata.get("source_id")
+            if old_source_id_for_table is None:
+                raise ValueError(
+                    f"Table '{old_table_name}' from {source_db.path} is missing source_id in its metadata."
+                )
+
+            new_source_id_for_table = self._get_new_source_id(
+                source_db_idx, old_source_id_for_table
+            )
+
+            # Create the table structure in the target database
+            # Using if_exists="fail" because new_table_name should already be unique.
+            # SDIFDatabase.create_table handles complex PKs via table_constraints reconstruction.
+            actual_created_name = self.target_db.create_table(
+                table_name=new_table_name,
+                columns=columns_for_create,
+                source_id=new_source_id_for_table,
+                description=source_table_metadata.get("description"),
+                original_identifier=source_table_metadata.get("original_identifier"),
+                if_exists="fail",
+            )
+            if actual_created_name != new_table_name:
+                # This case should ideally not happen if _generate_unique_target_table_name was correct
+                # and create_table used if_exists='fail'. If create_table internally changes name even with 'fail',
+                # this is an issue. For now, assume 'fail' means it uses the name or errors.
+                log.warning(
+                    f"Table name discrepancy: expected {new_table_name}, created as {actual_created_name}. Using created name."
+                )
+                self.table_name_map[source_db_idx][old_table_name] = (
+                    actual_created_name  # Update map
+                )
+
+            # Copy data
+            try:
+                data_df = source_db.read_table(old_table_name)
+                if not data_df.empty:
+                    # SDIFDatabase.insert_data expects List[Dict].
+                    # SDIFDatabase.write_dataframe is higher level but might re-create table.
+                    # Let's use insert_data.
+
+                    # Handle data type conversions that pandas might do, to align with SQLite expectations
+                    # For example, pandas bools to int 0/1, datetimes to ISO strings.
+                    # The SDIFDatabase.write_dataframe has logic for this.
+                    # We can replicate parts or simplify if read_table and insert_data are robust.
+                    # For now, assume read_table gives compatible data for insert_data
+                    # or insert_data can handle common pandas types.
+                    # A quick check: SDIFDatabase.insert_data does not do type conversion.
+                    # SDIFDatabase.write_dataframe does. So it's safer to go df -> records -> insert
+                    # after manual conversion like in write_dataframe.
+
+                    df_copy = data_df.copy()
+                    for col_name_str in df_copy.columns:
+                        col_name = str(col_name_str)  # Ensure string
+                        if pd.api.types.is_bool_dtype(df_copy[col_name].dtype):
+                            df_copy[col_name] = df_copy[col_name].astype(int)
+                        elif pd.api.types.is_datetime64_any_dtype(
+                            df_copy[col_name].dtype
+                        ):
+                            df_copy[col_name] = df_copy[col_name].apply(
+                                lambda x: x.isoformat() if pd.notnull(x) else None
+                            )
+                        elif pd.api.types.is_timedelta64_dtype(df_copy[col_name].dtype):
+                            df_copy[col_name] = df_copy[col_name].astype(str)
+                        # Handle potential np.nan to None for JSON compatibility if objects were stored as text
+                        if df_copy[col_name].dtype == object:
+                            df_copy[col_name] = df_copy[col_name].replace(
+                                {np.nan: None}
+                            )
+
+                    data_records = df_copy.to_dict("records")
+                    if data_records:  # Ensure there are records to insert
+                        self.target_db.insert_data(actual_created_name, data_records)
+            except Exception as e:
+                log.error(
+                    f"Failed to copy data for table {old_table_name} to {actual_created_name}: {e}"
+                )
+                # Decide: continue with other tables or raise? For robustness, log and continue.
+                # Or add a strict mode flag. For now, log and continue.
+
+    def _merge_objects(self, source_db: SDIFDatabase, source_db_idx: int):
+        self.object_name_map[source_db_idx] = {}
+        for old_object_name in source_db.list_objects():
+            obj_data = source_db.get_object(
+                old_object_name, parse_json=False
+            )  # Get raw JSON strings
+            if not obj_data:
+                log.warning(
+                    f"Could not retrieve object '{old_object_name}' from {source_db.path}. Skipping."
+                )
+                continue
+
+            new_object_name = self._generate_unique_name_in_target(
+                old_object_name, self.target_db.list_objects
+            )
+            self.object_name_map[source_db_idx][old_object_name] = new_object_name
+
+            new_source_id = self._get_new_source_id(
+                source_db_idx, obj_data["source_id"]
+            )
+
+            # Data is already string from parse_json=False. Schema hint also string.
+            # SDIFDatabase.add_object expects data to be Any (serializable) and schema_hint Dict.
+            # So we need to parse them back if they are strings.
+            parsed_json_data = json.loads(obj_data["json_data"])
+            parsed_schema_hint = (
+                json.loads(obj_data["schema_hint"])
+                if obj_data.get("schema_hint")
+                else None
+            )
+
+            self.target_db.add_object(
+                object_name=new_object_name,
+                json_data=parsed_json_data,
+                source_id=new_source_id,
+                description=obj_data.get("description"),
+                schema_hint=parsed_schema_hint,
+            )
+
+    def _merge_media(self, source_db: SDIFDatabase, source_db_idx: int):
+        self.media_name_map[source_db_idx] = {}
+        for old_media_name in source_db.list_media():
+            media_entry = source_db.get_media(
+                old_media_name, parse_json=False
+            )  # Get raw JSON for tech_metadata
+            if not media_entry:
+                log.warning(
+                    f"Could not retrieve media '{old_media_name}' from {source_db.path}. Skipping."
+                )
+                continue
+
+            new_media_name = self._generate_unique_name_in_target(
+                old_media_name, self.target_db.list_media
+            )
+            self.media_name_map[source_db_idx][old_media_name] = new_media_name
+
+            new_source_id = self._get_new_source_id(
+                source_db_idx, media_entry["source_id"]
+            )
+
+            parsed_tech_metadata = (
+                json.loads(media_entry["technical_metadata"])
+                if media_entry.get("technical_metadata")
+                else None
+            )
+
+            self.target_db.add_media(
+                media_name=new_media_name,
+                media_data=media_entry["media_data"],  # Should be bytes
+                media_type=media_entry["media_type"],
+                source_id=new_source_id,
+                description=media_entry.get("description"),
+                original_format=media_entry.get("original_format"),
+                technical_metadata=parsed_tech_metadata,
+            )
+
+    def _remap_element_spec(
+        self, element_type: str, element_spec_json: str, source_db_idx: int
+    ) -> str:
+        if not element_spec_json:
+            return element_spec_json
+
+        try:
+            spec_dict = json.loads(element_spec_json)
+        except json.JSONDecodeError:
+            log.warning(
+                f"Invalid JSON in element_spec: {element_spec_json}. Returning as is."
+            )
+            return element_spec_json
+
+        new_spec_dict = spec_dict.copy()
+
+        # Remap source_id if present (relevant for 'source' element_type in annotations, not directly in semantic_links spec)
+        # Semantic links link to other entities which carry their own source_id.
+        # But if spec itself contains a source_id key (e.g. for target_element_type='source' in annotations)
+        if "source_id" in new_spec_dict and isinstance(new_spec_dict["source_id"], int):
+            new_spec_dict["source_id"] = self._get_new_source_id(
+                source_db_idx, new_spec_dict["source_id"]
+            )
+
+        # Remap names based on element_type
+        if element_type in ["table", "column"]:
+            if "table_name" in new_spec_dict:
+                new_spec_dict["table_name"] = self._get_new_table_name(
+                    source_db_idx, new_spec_dict["table_name"]
+                )
+        elif element_type == "object":  # Direct object reference
+            if "object_name" in new_spec_dict:
+                new_spec_dict["object_name"] = self._get_new_object_name(
+                    source_db_idx, new_spec_dict["object_name"]
+                )
+        elif element_type == "json_path":  # JSONPath typically refers to an object
+            if (
+                "object_name" in new_spec_dict
+            ):  # If the spec identifies the object container
+                new_spec_dict["object_name"] = self._get_new_object_name(
+                    source_db_idx, new_spec_dict["object_name"]
+                )
+        elif element_type == "media":
+            if "media_name" in new_spec_dict:
+                new_spec_dict["media_name"] = self._get_new_media_name(
+                    source_db_idx, new_spec_dict["media_name"]
+                )
+        # 'file' type needs no remapping of spec content.
+        # 'source' type: primary key is 'source_id', remapped above.
+
+        return json.dumps(new_spec_dict)
+
+    def _merge_semantic_links(self, source_db: SDIFDatabase, source_db_idx: int):
+        # SDIFDatabase.list_semantic_links default parses JSON spec. We need this.
+        source_links = source_db.list_semantic_links(parse_json=True)
+
+        for link in source_links:
+            # The specs are already dicts because parse_json=True was used.
+            try:
+                from_spec_dict = link["from_element_spec"]
+                to_spec_dict = link["to_element_spec"]
+
+                # Remap the dicts directly
+                new_from_spec_dict = self._remap_element_spec_dict(
+                    link["from_element_type"], from_spec_dict, source_db_idx
+                )
+                new_to_spec_dict = self._remap_element_spec_dict(
+                    link["to_element_type"], to_spec_dict, source_db_idx
+                )
+
+                self.target_db.add_semantic_link(
+                    link_type=link["link_type"],
+                    from_element_type=link["from_element_type"],
+                    from_element_spec=new_from_spec_dict,  # add_semantic_link takes dict
+                    to_element_type=link["to_element_type"],
+                    to_element_spec=new_to_spec_dict,  # add_semantic_link takes dict
+                    description=link.get("description"),
+                )
+            except Exception as e:
+                link_id = link.get("link_id", "Unknown")
+                log.error(
+                    f"Failed to merge semantic link ID {link_id} from {source_db.path}: {e}. Skipping link."
+                )
+
+    def _remap_element_spec_dict(
+        self, element_type: str, spec_dict: Dict, source_db_idx: int
+    ) -> Dict:
+        # Helper for _merge_semantic_links that works with dicts directly
+        new_spec_dict = spec_dict.copy()
+
+        if "source_id" in new_spec_dict and isinstance(new_spec_dict["source_id"], int):
+            new_spec_dict["source_id"] = self._get_new_source_id(
+                source_db_idx, new_spec_dict["source_id"]
+            )
+
+        if element_type in ["table", "column"]:
+            if "table_name" in new_spec_dict:
+                new_spec_dict["table_name"] = self._get_new_table_name(
+                    source_db_idx, new_spec_dict["table_name"]
+                )
+        elif element_type == "object" or (
+            element_type == "json_path" and "object_name" in new_spec_dict
+        ):
+            if "object_name" in new_spec_dict:
+                new_spec_dict["object_name"] = self._get_new_object_name(
+                    source_db_idx, new_spec_dict["object_name"]
+                )
+        elif element_type == "media":
+            if "media_name" in new_spec_dict:
+                new_spec_dict["media_name"] = self._get_new_media_name(
+                    source_db_idx, new_spec_dict["media_name"]
+                )
+        return new_spec_dict
+
+    def merge_all(self, source_sdif_paths: List[SDIFPath]):
+        # Import pandas and numpy here to avoid making them a hard dependency of the module if not used.
+        # However, SDIFDatabase itself uses them. So they are effectively dependencies.
+        global pd, np
+        import numpy as np
+        import pandas as pd
+
+        for idx, source_path_item in enumerate(source_sdif_paths):
+            source_path = Path(source_path_item)  # Ensure Path object
+            log.info(
+                f"Processing source SDIF ({idx + 1}/{len(source_sdif_paths)}): {source_path}"
+            )
+            source_db = SDIFDatabase(source_path, read_only=True)
+            try:  # Ensure source_db is closed
+                self._merge_properties(source_db, idx)
+                self._merge_sources(source_db, idx)
+                self._merge_tables(source_db, idx)  # This needs pandas for data reading
+                self._merge_objects(source_db, idx)
+                self._merge_media(source_db, idx)
+                self._merge_semantic_links(source_db, idx)
+                # Not merging sdif_annotations in this version.
+            finally:
+                source_db.close()
+
+        # Finalize target DB properties
+        try:
+            current_timestamp_utc_z = datetime.utcnow().strftime("%Y-%m-%dT%H:%M:%SZ")
+            self.target_db.conn.execute(
+                "UPDATE sdif_properties SET creation_timestamp = ?",
+                (current_timestamp_utc_z,),
+            )
+            self.target_db.conn.commit()
+        except sqlite3.Error as e:
+            log.error(f"Failed to update final creation_timestamp in target DB: {e}")
+            # Non-fatal, proceed.
+
+        self.target_db.close()
+        log.info(
+            f"Successfully merged {len(source_sdif_paths)} SDIF files into {self.target_db.path}"
+        )
+        return self.target_db.path
+
+
+def merge_sdif_files(sdif_paths: List[SDIFPath], output_path: Path) -> Path:
+    """
+    Merges multiple SDIF files into a single new SDIF file.
+
+    Args:
+        sdif_paths: A list of paths to the SDIF files to merge.
+        output_path: The full path where the merged SDIF file should be saved.
+                     Its parent directory will be created if it doesn't exist.
+                     If output_path is an existing file, it will be overwritten.
+                     If output_path is an existing directory, a ValueError is raised.
+
+    Returns:
+        Path to the newly created merged SDIF file (same as output_path).
+
+    Raises:
+        ValueError: If no SDIF files are provided, or output_path is invalid (e.g., an existing directory).
+        FileNotFoundError: If a source SDIF file does not exist.
+        sqlite3.Error: For database-related errors during merging.
+        RuntimeError: For critical errors like inability to generate unique names.
+    """
+    if not sdif_paths:
+        raise ValueError("No SDIF files provided for merging.")
+
+    output_path = Path(output_path).resolve()
+
+    if output_path.is_dir():
+        raise ValueError(
+            f"Output path '{output_path}' is an existing directory. Please provide a full file path."
+        )
+
+    # Ensure parent directory of output_path exists
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+
+    # Ensure all source paths are Path objects and exist
+    processed_sdif_paths: List[Path] = []
+    for p in sdif_paths:
+        path_obj = Path(p).resolve()
+        if not path_obj.exists():
+            raise FileNotFoundError(f"Source SDIF file not found: {path_obj}")
+        if not path_obj.is_file():
+            raise ValueError(f"Source SDIF path is not a file: {path_obj}")
+        processed_sdif_paths.append(path_obj)
+
+    if len(processed_sdif_paths) == 1:
+        source_file = processed_sdif_paths[0]
+
+        # If the source and target are the same file, no copy is needed.
+        if source_file == output_path:
+            return source_file
+
+        shutil.copy(source_file, output_path)
+        log.info(f"Copied single SDIF file to '{output_path}' as no merge was needed.")
+        return output_path
+
+    # For multiple files, merge them into the output_path
+    merger = _SDIFMerger(output_path)
+    return merger.merge_all(processed_sdif_paths)

satif_ai/utils/openai_mcp.py

@@ -0,0 +1,97 @@
+import logging
+from typing import Any
+
+from agents.mcp.server import CallToolResult, MCPServer, MCPTool
+from fastmcp import FastMCP
+
+logger = logging.getLogger(__name__)
+
+
+class OpenAICompatibleMCP(MCPServer):
+    def __init__(self, mcp: FastMCP):
+        self.mcp = mcp
+        self._is_connected = False  # Track connection state
+
+    async def connect(self):
+        """Connect to the server.
+        For FastMCP, connection is managed externally when the server is run.
+        This method marks the wrapper as connected.
+        """
+        # Assuming FastMCP instance is already running and configured.
+        # No specific connect action required for the FastMCP instance itself here,
+        # as its lifecycle (run, stop) is managed outside this wrapper.
+        logger.info(
+            f"OpenAICompatibleMCP: Simulating connection to FastMCP server '{self.mcp.name}'."
+        )
+        self._is_connected = True
+
+    @property
+    def name(self) -> str:
+        """A readable name for the server."""
+        return self.mcp.name
+
+    async def cleanup(self):
+        """Cleanup the server.
+        For FastMCP, cleanup is managed externally. This method marks the wrapper as disconnected.
+        """
+        # Similar to connect, actual server cleanup is external.
+        logger.info(
+            f"OpenAICompatibleMCP: Simulating cleanup for FastMCP server '{self.mcp.name}'."
+        )
+        self._is_connected = False
+
+    async def list_tools(self) -> list[MCPTool]:
+        """List the tools available on the server."""
+        if not self._is_connected:
+            # Or raise an error, depending on desired behavior for disconnected state
+            raise RuntimeError(
+                "OpenAICompatibleMCP.list_tools called while not connected."
+            )
+
+        # FastMCP's get_tools() returns a dict[str, fastmcp.tools.tool.Tool]
+        # Each fastmcp.tools.tool.Tool has a to_mcp_tool(name=key) method
+        # MCPTool is an alias for mcp.types.Tool
+        try:
+            fastmcp_tools = await self.mcp.get_tools()
+            mcp_tools_list = [
+                tool.to_mcp_tool(name=key) for key, tool in fastmcp_tools.items()
+            ]
+            return mcp_tools_list
+        except Exception as e:
+            logger.error(
+                f"Error listing tools from FastMCP server '{self.mcp.name}': {e}",
+                exc_info=True,
+            )
+            raise e
+
+    async def call_tool(
+        self, tool_name: str, arguments: dict[str, Any] | None
+    ) -> CallToolResult:
+        """Invoke a tool on the server."""
+        if not self._is_connected:
+            logger.warning(
+                f"OpenAICompatibleMCP.call_tool '{tool_name}' called while not connected."
+            )
+            # Return an error CallToolResult
+            return CallToolResult(
+                content=[{"type": "text", "text": "Server not connected"}], isError=True
+            )
+
+        try:
+            # FastMCP's _mcp_call_tool is a protected member, but seems to be what we need.
+            # It returns: list[TextContent | ImageContent | EmbeddedResource]
+            # This matches the 'content' part of CallToolResult.
+            # We need to handle potential errors and wrap the result.
+            content = await self.mcp._mcp_call_tool(tool_name, arguments or {})
+            return CallToolResult(content=content, isError=False)
+        except Exception as e:
+            logger.error(
+                f"Error calling tool '{tool_name}' on FastMCP server '{self.mcp.name}': {e}",
+                exc_info=True,
+            )
+            error_message = f"Error calling tool '{tool_name}': {type(e).__name__}: {e}"
+            # Ensure content is a list of valid MCP content items, even for errors.
+            # A TextContent is a safe choice.
+            return CallToolResult(
+                content=[{"type": "text", "text": error_message}], isError=True
+            )