langgraph-executor 0.0.1a4__tar.gz → 0.0.1a6__tar.gz

{langgraph_executor-0.0.1a4 → langgraph_executor-0.0.1a6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: langgraph-executor
-Version: 0.0.1a4
+Version: 0.0.1a6
 Summary: LangGraph python RPC server executable by the langgraph-go orchestrator.
 Requires-Python: >=3.11
 Requires-Dist: grpcio>=1.73.1

langgraph_executor-0.0.1a6/langgraph_executor/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.0.1a6"

langgraph_executor-0.0.1a6/langgraph_executor/client/patch.py

@@ -0,0 +1,305 @@
+import logging
+from typing import Any
+
+import grpc
+from langgraph._internal._config import ensure_config
+from langgraph.errors import GraphInterrupt, GraphRecursionError
+from langgraph.pregel import Pregel
+from langgraph.runtime import get_runtime
+from langgraph.types import Interrupt
+from pydantic import ValidationError
+
+from langgraph_executor.client.utils import (
+    SERDE,
+    config_to_pb,
+    context_to_pb,
+    create_runopts_pb,
+    decode_response,
+    input_to_pb,
+)
+from langgraph_executor.common import var_child_runnable_config
+from langgraph_executor.pb import runtime_pb2
+from langgraph_executor.pb.runtime_pb2 import OutputChunk
+from langgraph_executor.pb.runtime_pb2_grpc import LangGraphRuntimeStub
+
+
+def _patch_pregel(runtime_client: LangGraphRuntimeStub, logger: logging.Logger):
+    async def patched_ainvoke(pregel_self, input, config=None, **kwargs):
+        return await _ainvoke_wrapper(
+            runtime_client, logger, pregel_self, input, config, **kwargs
+        )
+
+    def patched_invoke(pregel_self, input, config=None, **kwargs):
+        return _invoke_wrapper(
+            runtime_client, logger, pregel_self, input, config, **kwargs
+        )
+
+    Pregel.ainvoke = patched_ainvoke  # type: ignore[invalid-assignment]
+    Pregel.invoke = patched_invoke  # type: ignore[invalid-assignment]
+
+
+async def _ainvoke_wrapper(
+    runtime_client: LangGraphRuntimeStub,
+    logger: logging.Logger,
+    pregel_self: Pregel,  # This is the actual Pregel instance
+    input,
+    config=None,
+    context=None,
+    stream_mode=["values"],
+    output_keys=None,
+    interrupt_before=None,
+    interrupt_after=None,
+    durability=None,
+    debug=None,
+    subgraphs=False,
+) -> dict[str, Any] | Any:
+    """Wrapper that handles the actual invoke logic."""
+
+    # subgraph names coerced when initializing executor
+    graph_name = pregel_self.name
+
+    logger.info(f"SUBGRAPH INVOKE ENCOUNTERED: {graph_name}")
+
+    # TODO: Hacky way of retrieving runtime from runnable context
+    if not context:
+        try:
+            runtime = get_runtime()
+            if runtime.context:
+                context = runtime.context
+        except Exception as e:
+            logger.error(f"failed to retrive parent runtime for subgraph: {e}")
+
+    if parent_config := var_child_runnable_config.get({}):
+        config = ensure_config(config, parent_config)
+
+    try:
+        # create request
+        invoke_request = runtime_pb2.InvokeRequest(
+            graph_name=graph_name,
+            input=input_to_pb(input),
+            config=config_to_pb(config),
+            context=context_to_pb(context),
+            run_opts=create_runopts_pb(
+                stream_mode,
+                output_keys,
+                interrupt_before,
+                interrupt_after,
+                durability,
+                debug,
+                subgraphs,
+            ),
+        )
+
+        # get response - if this blocks, you might need to make it async
+        try:
+            # Option 1: If runtime_client.Invoke is synchronous and might block:
+            import asyncio
+
+            loop = asyncio.get_event_loop()
+            response = await loop.run_in_executor(
+                None, runtime_client.Invoke, invoke_request
+            )
+
+            if response.WhichOneof("message") == "error":
+                error = response.error.error
+
+                if error.WhichOneof("error_type") == "graph_interrupt":
+                    graph_interrupt = error.graph_interrupt
+
+                    interrupts = []
+
+                    for interrupt in graph_interrupt.interrupts:
+                        interrupts.append(
+                            Interrupt(
+                                value=SERDE.loads_typed(
+                                    (
+                                        interrupt.value.base_value.method,
+                                        interrupt.value.base_value.value,
+                                    )
+                                ),
+                                id=interrupt.id,
+                            )
+                        )
+
+                    raise GraphInterrupt(interrupts)
+
+                else:
+                    raise ValueError(
+                        f"Unknown subgraph error from orchestrator: {error!s}"
+                    )
+
+        except grpc.RpcError as e:
+            # grpc_message is inside str(e)
+            details = str(e)
+            if details and "recursion limit exceeded" in details.lower():
+                raise GraphRecursionError
+            if details and "invalid context format" in details.lower():
+                raise TypeError
+            if details and "invalid pydantic context format" in details.lower():
+                import json
+
+                # Extract the JSON error data from the error message
+                error_msg = str(e)
+                if ": {" in error_msg:
+                    json_part = "{" + error_msg.split(": {")[1]
+                    try:
+                        error_data = json.loads(json_part)
+                        raise ValidationError.from_exception_data(
+                            error_data["title"], error_data["errors"]
+                        )
+                    except (json.JSONDecodeError, KeyError) as e:
+                        logger.error(f"JSONDecodeError: {e}")
+                # Fallback if parsing fails
+                raise ValidationError.from_exception_data(
+                    "ValidationError",
+                    [
+                        {
+                            "type": "value_error",
+                            "loc": ("context",),
+                            "msg": "invalid pydantic context format",
+                            "input": None,
+                        }
+                    ],
+                )
+            raise
+
+        # decode response
+        return decode_response(response, stream_mode)
+
+    except Exception as e:
+        if isinstance(e, grpc.RpcError):
+            logger.error(f"gRPC client/runtime error: {e!s}")
+        raise e
+
+
+def _invoke_wrapper(
+    runtime_client: LangGraphRuntimeStub,
+    logger: logging.Logger,
+    pregel_self: Pregel,  # This is the actual Pregel instance
+    input,
+    config=None,
+    context=None,
+    stream_mode=["values"],
+    output_keys=None,
+    interrupt_before=None,
+    interrupt_after=None,
+    durability=None,
+    debug=None,
+    subgraphs=False,
+) -> dict[str, Any] | Any:
+    """Wrapper that handles the actual invoke logic."""
+
+    # subgraph names coerced when initializing executor
+    graph_name = pregel_self.name
+
+    logger.info(f"SUBGRAPH INVOKE ENCOUNTERED: {graph_name}")
+
+    # TODO: Hacky way of retrieving runtime from runnable context
+    if not context:
+        try:
+            runtime = get_runtime()
+            if runtime.context:
+                context = runtime.context
+        except Exception as e:
+            logger.error(f"failed to retrive parent runtime for subgraph: {e}")
+
+    # need to get config of parent because wont be available in orchestrator
+    if parent_config := var_child_runnable_config.get({}):
+        config = ensure_config(config, parent_config)
+
+    try:
+        # create request
+        invoke_request = runtime_pb2.InvokeRequest(
+            graph_name=graph_name,
+            input=input_to_pb(input),
+            config=config_to_pb(config),
+            context=context_to_pb(context),
+            run_opts=create_runopts_pb(
+                stream_mode,
+                output_keys,
+                interrupt_before,
+                interrupt_after,
+                durability,
+                debug,
+                subgraphs,
+            ),
+        )
+
+        try:
+            response: OutputChunk = runtime_client.Invoke(invoke_request)
+
+            if response.WhichOneof("message") == "error":
+                error = response.error.error
+
+                if error.WhichOneof("error_type") == "graph_interrupt":
+                    graph_interrupt = error.graph_interrupt
+
+                    interrupts = []
+
+                    for interrupt in graph_interrupt.interrupts:
+                        interrupts.append(
+                            Interrupt(
+                                value=SERDE.loads_typed(
+                                    (
+                                        interrupt.value.base_value.method,
+                                        interrupt.value.base_value.value,
+                                    )
+                                ),
+                                id=interrupt.id,
+                            )
+                        )
+
+                    raise GraphInterrupt(interrupts)
+
+                else:
+                    raise ValueError(
+                        f"Unknown subgraph error from orchestrator: {error!s}"
+                    )
+
+        except grpc.RpcError as e:
+            # grpc_message is inside str(e)
+            details = str(e)
+            if details and "recursion limit exceeded" in details.lower():
+                raise GraphRecursionError
+            if details and "invalid context format" in details.lower():
+                raise TypeError
+            if details and "invalid pydantic context format" in details.lower():
+                import json
+
+                # Extract the JSON error data from the error message
+                error_msg = str(e)
+                if ": {" in error_msg:
+                    json_part = "{" + error_msg.split(": {")[1]
+                    try:
+                        error_data = json.loads(json_part)
+                        raise ValidationError.from_exception_data(
+                            error_data["title"], error_data["errors"]
+                        )
+                    except (json.JSONDecodeError, KeyError) as e:
+                        logger.error(f"JSONDecodeError: {e}")
+                # Fallback if parsing fails
+                raise ValidationError.from_exception_data(
+                    "ValidationError",
+                    [
+                        {
+                            "type": "value_error",
+                            "loc": ("context",),
+                            "msg": "invalid pydantic context format",
+                            "input": None,
+                        }
+                    ],
+                )
+            raise
+
+        # decode response
+        return decode_response(response, stream_mode)
+
+    except Exception as e:
+        if isinstance(e, grpc.RpcError):
+            logger.error(f"gRPC client/runtime error: {e!s}")
+        raise e
+
+
+__all__ = [
+    "_patch_pregel",
+]

langgraph_executor-0.0.1a6/langgraph_executor/client/utils.py

@@ -0,0 +1,340 @@
+import base64
+import copy
+import re
+from collections.abc import Sequence
+from typing import Any, cast
+
+from google.protobuf.json_format import MessageToDict
+from langchain_core.messages import AIMessageChunk, BaseMessage
+from langchain_core.messages.utils import convert_to_messages
+from langchain_core.runnables import RunnableConfig
+from langgraph._internal._config import _is_not_empty
+from langgraph._internal._constants import (
+    CONFIG_KEY_CHECKPOINT_ID,
+    CONFIG_KEY_CHECKPOINT_MAP,
+    CONFIG_KEY_CHECKPOINT_NS,
+    CONFIG_KEY_DURABILITY,
+    CONFIG_KEY_RESUMING,
+    CONFIG_KEY_TASK_ID,
+    CONFIG_KEY_THREAD_ID,
+)
+from langgraph.checkpoint.serde.jsonplus import JsonPlusSerializer
+from langgraph.pregel.debug import CheckpointMetadata
+from langgraph.types import StateSnapshot
+
+from langgraph_executor.common import reconstruct_config, val_to_pb
+from langgraph_executor.pb import runtime_pb2, types_pb2
+
+SERDE = JsonPlusSerializer()
+
+
+def input_to_pb(input):
+    return val_to_pb(None, input)
+
+
+def _is_present_and_not_empty(config: RunnableConfig, key: Any) -> bool:
+    return key in config and _is_not_empty(config[key])
+
+
+def maybe_update_reserved_configurable(
+    key: str, value: Any, reserved_configurable: types_pb2.ReservedConfigurable
+) -> bool:
+    if key == CONFIG_KEY_RESUMING:
+        reserved_configurable.resuming = bool(value)
+    elif key == CONFIG_KEY_TASK_ID:
+        reserved_configurable.task_id = str(value)
+    elif key == CONFIG_KEY_THREAD_ID:
+        reserved_configurable.thread_id = str(value)
+    elif key == CONFIG_KEY_CHECKPOINT_MAP:
+        reserved_configurable.checkpoint_map.update(cast(dict[str, str], value))
+    elif key == CONFIG_KEY_CHECKPOINT_ID:
+        reserved_configurable.checkpoint_id = str(value)
+    elif key == CONFIG_KEY_CHECKPOINT_NS:
+        reserved_configurable.checkpoint_ns = str(value)
+    # elif key == CONFIG_KEY_PREVIOUS:
+    #     serde = JsonPlusSerializer()
+    #     meth, ser = serde.dumps_typed(value)
+    #     reserved_configurable.previous = types_pb2.SerializedValue(method=meth, value=bytes(ser))
+    elif key == CONFIG_KEY_DURABILITY:
+        reserved_configurable.durability = str(value)
+    else:
+        return False
+
+    return True
+
+
+def merge_configurables(
+    config: RunnableConfig, pb_config: types_pb2.RunnableConfig
+) -> None:
+    if not _is_present_and_not_empty(config, "configurable"):
+        return
+
+    configurable = pb_config.configurable
+    reserved_configurable = pb_config.reserved_configurable
+
+    for k, v in config["configurable"].items():
+        if not maybe_update_reserved_configurable(k, v, reserved_configurable):
+            try:
+                configurable.update({k: v})
+            except ValueError:  # TODO handle this
+                print(f"could not pass config field {k}:{v} to proto")
+
+
+def config_to_pb(config: RunnableConfig) -> types_pb2.RunnableConfig:
+    if not config:
+        return types_pb2.RunnableConfig()
+
+    # Prepare kwargs for construction
+    kwargs = {}
+
+    if _is_present_and_not_empty(config, "run_name"):
+        kwargs["run_name"] = config["run_name"]
+
+    if _is_present_and_not_empty(config, "run_id"):
+        kwargs["run_id"] = str(config["run_id"]) if config["run_id"] else ""
+
+    if _is_present_and_not_empty(config, "max_concurrency"):
+        kwargs["max_concurrency"] = int(config["max_concurrency"])
+
+    if _is_present_and_not_empty(config, "recursion_limit"):
+        kwargs["recursion_limit"] = config["recursion_limit"]
+
+    # Create the config with initial values
+    pb_config = types_pb2.RunnableConfig(**kwargs)
+
+    # Handle collections after construction
+    if _is_present_and_not_empty(config, "tags"):
+        if isinstance(config["tags"], list):
+            pb_config.tags.extend(config["tags"])
+        elif isinstance(config["tags"], str):
+            pb_config.tags.append(config["tags"])
+
+    if _is_present_and_not_empty(config, "metadata"):
+        pb_config.metadata.update(config["metadata"])
+
+    merge_configurables(config, pb_config)
+
+    return pb_config
+
+
+def context_to_pb(context: dict[str, Any] | Any) -> types_pb2.Context | None:
+    if context is None:
+        return None
+
+    # Convert dataclass or other objects to dict if needed
+    if hasattr(context, "__dict__") and not hasattr(context, "items"):
+        # Convert dataclass to dict
+        context_dict = context.__dict__
+    elif hasattr(context, "items"):
+        # Already a dict-like object
+        context_dict = context
+    else:
+        # Try to convert to dict using vars()
+        context_dict = vars(context) if hasattr(context, "__dict__") else {}
+
+    return types_pb2.Context(context=context_dict)
+
+
+# TODO
+def create_runopts_pb(
+    stream_mode,
+    output_keys,
+    interrupt_before,
+    interrupt_after,
+    durability,
+    debug,
+    subgraphs,
+) -> runtime_pb2.RunOpts:
+    # Prepare kwargs for construction
+    kwargs = {}
+
+    if durability is not None:
+        kwargs["durability"] = durability
+
+    if debug is not None:
+        kwargs["debug"] = debug
+
+    if subgraphs is not None:
+        kwargs["subgraphs"] = subgraphs
+
+    if output_keys is not None:
+        string_or_slice_pb = None
+        if isinstance(output_keys, str):
+            string_or_slice_pb = types_pb2.StringOrSlice(
+                is_string=True, values=[output_keys]
+            )
+        elif isinstance(output_keys, list[str]):
+            string_or_slice_pb = types_pb2.StringOrSlice(
+                is_string=False, values=output_keys
+            )
+
+        if string_or_slice_pb is not None:
+            kwargs["output_keys"] = string_or_slice_pb
+
+    # Create the RunOpts with initial values
+    run_opts = runtime_pb2.RunOpts(**kwargs)
+
+    # Handle repeated fields after construction
+    if stream_mode is not None:
+        if isinstance(stream_mode, str):
+            run_opts.stream_mode.append(stream_mode)
+        elif isinstance(stream_mode, list):
+            run_opts.stream_mode.extend(stream_mode)
+
+    if interrupt_before is not None:
+        run_opts.interrupt_before.extend(interrupt_before)
+
+    if interrupt_after is not None:
+        run_opts.interrupt_after.extend(interrupt_after)
+
+    # Note: checkpoint_during field doesn't exist in RunOpts proto
+    # Ignoring it as it's not in the proto definition
+
+    return run_opts
+
+
+def decode_response(response, stream_mode):
+    which = response.WhichOneof("message")
+    if which == "error":
+        raise ValueError(response.error)
+    if which == "chunk":
+        return decode_chunk(response.chunk, stream_mode)
+    if which == "chunk_list":
+        return [
+            decode_chunk(chunk.chunk, stream_mode)
+            for chunk in response.chunk_list.chunks
+        ]
+
+    raise ValueError("No stream response")
+
+
+VAL_KEYS = {"method", "value"}
+
+
+def deser_vals(chunk: dict[str, Any]):
+    return _deser_vals(copy.deepcopy(chunk))
+
+
+def _deser_vals(current_chunk):
+    if isinstance(current_chunk, list):
+        return [_deser_vals(v) for v in current_chunk]
+    if not isinstance(current_chunk, dict):
+        return current_chunk
+    if set(current_chunk.keys()) == VAL_KEYS:
+        return SERDE.loads_typed(
+            (current_chunk["method"], base64.b64decode(current_chunk["value"]))
+        )
+    for k, v in current_chunk.items():
+        if isinstance(v, dict | Sequence):
+            current_chunk[k] = _deser_vals(v)
+    return current_chunk
+
+
+def decode_state_history_response(response):
+    if not response:
+        return
+
+    return [reconstruct_state_snapshot(state_pb) for state_pb in response.history]
+
+
+def decode_state_response(response):
+    if not response:
+        return
+
+    return reconstruct_state_snapshot(response.state)
+
+
+# TODO finish reconstructing these
+def reconstruct_state_snapshot(state_pb: types_pb2.StateSnapshot) -> StateSnapshot:
+    return StateSnapshot(
+        values=deser_vals(MessageToDict(state_pb.values)),
+        next=tuple(state_pb.next),
+        config=reconstruct_config(state_pb.config),
+        metadata=CheckpointMetadata(**MessageToDict(state_pb.metadata)),
+        created_at=state_pb.created_at,
+        parent_config=reconstruct_config(state_pb.parent_config),
+        tasks=tuple(),
+        interrupts=tuple(),
+    )
+
+
+def decode_chunk(chunk, stream_mode):
+    d = cast(dict[str, Any], deser_vals(MessageToDict(chunk)))
+    stream_mode = stream_mode or ()
+    mode = d.get("mode")
+    ns = d.get("ns")
+    # Handle messages mode specifically - we don't always send the stream mode in the chunk
+    # Because if user only has 1 mode, we exclude it since it is implied
+    if mode == "messages" or (mode is None and "messages" in stream_mode):
+        return (ns, extract_message_chunk(d["payload"]))
+
+    # Handle custom mode primitive extraction
+    payload = d.get("payload")
+
+    # For custom mode, unwrap primitives from "data" wrapper
+    if mode == "custom" or (mode is None and "custom" in stream_mode):
+        if isinstance(payload, dict) and len(payload) == 1 and "data" in payload:
+            payload = payload["data"]
+
+    # Regular logic for all modes
+    if ns:
+        if mode:
+            return (ns, mode, payload)
+        return (ns, payload)
+    if mode:
+        return (mode, payload)
+
+    return payload
+
+
+class AnyStr(str):
+    def __init__(self, prefix: str | re.Pattern = "") -> None:
+        super().__init__()
+        self.prefix = prefix
+
+    def __eq__(self, other: object) -> bool:
+        return isinstance(other, str) and (
+            other.startswith(self.prefix)
+            if isinstance(self.prefix, str)
+            else self.prefix.match(other)
+        )
+
+    def __hash__(self) -> int:
+        return hash((str(self), self.prefix))
+
+
+def extract_message_chunk(
+    payload: dict[str, Any],
+) -> tuple[BaseMessage, dict[str, Any]]:
+    """Extract (BaseMessage, metadata) tuple from messages mode payload"""
+
+    # Extract writes from payload and deserialize the message data
+    message_data = payload.get("message", {}).get("message", {})
+    metadata = payload.get("metadata", {})
+    message_type = message_data.get("type", "ai")
+    if message_type.endswith("Chunk"):
+        message_id = message_data.get("id")
+        content = message_data.get("content", "")
+        additional_kwargs = message_data.get("additional_kwargs", {})
+        usage_metadata = message_data.get("usage_metadata", None)
+        tool_calls = message_data.get("tool_calls", [])
+        name = message_data.get("name")
+        tool_call_chunks = message_data.get("tool_call_chunks", [])
+        response_metadata = message_data.get("response_metadata", {})
+        if message_type == "AIMessageChunk":
+            message = AIMessageChunk(
+                content=content,
+                id=message_id,
+                additional_kwargs=additional_kwargs,
+                tool_calls=tool_calls,
+                name=name,
+                usage_metadata=usage_metadata,
+                tool_call_chunks=tool_call_chunks,
+                response_metadata=response_metadata,
+            )
+            return (message, metadata)
+        else:
+            raise ValueError(f"Unknown message type: {message_type}")
+
+    else:
+        return convert_to_messages([message_data])[0], metadata