langgraph-api 0.18.0.dev11__py3-none-any.whl → 0.18.0.dev13__py3-none-any.whl

langgraph_api/__init__.py

@@ -1 +1 @@
-__version__ = "0.18.0.dev11"
+__version__ = "0.18.0.dev13"

langgraph_api/api/__init__.py

@@ -28,10 +28,14 @@ from langgraph_api.api.threads import threads_routes
 from langgraph_api.api.ui import ui_routes
 from langgraph_api.auth.middleware import auth_middleware
 from langgraph_api.config import (
+    A2A_ENABLED,
     FF_PYSPY_PROFILING_ENABLED,
     HTTP_CONFIG,
     LANGGRAPH_ENCRYPTION,
+    MCP_ENABLED,
     MIGRATIONS_PATH,
+    RUNS_ENABLED,
+    THREADS_ENABLED,
 )
 from langgraph_api.feature_flags import (
     FF_V2_EVENT_STREAMING,
@@ -166,9 +170,9 @@ protected_routes: list[BaseRoute] = []
 if HTTP_CONFIG:
     if not HTTP_CONFIG.get("disable_assistants"):
         protected_routes.extend(assistants_routes)
-    if not HTTP_CONFIG.get("disable_runs"):
+    if RUNS_ENABLED:
         protected_routes.extend(runs_routes)
-    if not HTTP_CONFIG.get("disable_threads"):
+    if THREADS_ENABLED:
         protected_routes.extend(threads_routes)
     if not HTTP_CONFIG.get("disable_store"):
         protected_routes.extend(store_routes)
@@ -176,21 +180,16 @@ if HTTP_CONFIG:
         protected_routes.extend(profile_routes)
     if not HTTP_CONFIG.get("disable_ui"):
         protected_routes.extend(ui_routes)
-    if not HTTP_CONFIG.get("disable_mcp"):
+    if MCP_ENABLED:
         protected_routes.extend(mcp_routes)
-    if not HTTP_CONFIG.get("disable_a2a"):
+    if A2A_ENABLED:
         protected_routes.extend(a2a_routes)
-    # ``disable_runs`` also gates event streaming because the v2 transport
-    # exposes run-creating commands (``run.start`` / ``run.cancel`` /
-    # ``run.attach`` over the WS, plus the ``/runs/{run_id}/protocol``
-    # attach route). If an operator hardens a deployment with
-    # ``disable_runs=true``, leaving event streaming open would be a
-    # backdoor that lets clients create runs anyway. ``disable_event_streaming``
-    # is the narrow knob (turn off just v2); ``disable_runs`` is the broad
-    # knob (no run-related functionality at all).
+    # disable_runs also unmounts event streaming: its WebSocket and the
+    # /runs/{run_id}/protocol route can start, cancel and attach runs, so leaving
+    # them mounted would let clients create runs on a deployment that disabled runs.
     if (
         FF_V2_EVENT_STREAMING
-        and not HTTP_CONFIG.get("disable_runs")
+        and RUNS_ENABLED
         and not HTTP_CONFIG.get("disable_event_streaming")
     ):
         protected_routes.extend(event_streaming_routes)

langgraph_api/api/event_streaming.py

@@ -330,7 +330,7 @@ async def _thread_events(request: Request) -> Response:
         finally:
             await manager.close()
 
-    return EventSourceResponse(body_iter())
+    return EventSourceResponse(body_iter(), on_close=manager.close)
 
 
 # ---------------------------------------------------------------------------

langgraph_api/api/mcp/_call_thread.py

@@ -0,0 +1,220 @@
+from __future__ import annotations
+
+from contextlib import asynccontextmanager, contextmanager
+from dataclasses import dataclass
+from http import HTTPStatus
+from typing import TYPE_CHECKING
+
+import httpx
+import jsonschema_rs
+import mcp_types as types
+import structlog
+from mcp.shared.exceptions import MCPError
+from starlette.exceptions import HTTPException
+
+from langgraph_api.api.mcp._constants import (
+    BREAKPOINT_MESSAGE,
+    FINISHED_CALL_TTL_MINUTES,
+    PAUSED_CALL_TTL_MINUTES,
+)
+from langgraph_api.api.mcp._outcome import Failed, Paused, RunOutcome, parse_outcome
+from langgraph_api.api.mcp._request_state import Binding, InvalidRequestState
+from langgraph_api.config import THREAD_TTL
+from langgraph_api.models.run import create_valid_run
+from langgraph_api.validation import RunCreateStateful, sanitize_and_validate
+from langgraph_runtime.database import connect
+from langgraph_runtime.retry import retry_db
+
+if TYPE_CHECKING:
+    from collections.abc import AsyncIterator, Iterator, Mapping
+    from uuid import UUID
+
+    from langgraph_sdk.client import LangGraphClient
+    from langgraph_sdk.schema import Thread
+    from pydantic import JsonValue
+
+    from langgraph_api.api.mcp._calls import ToolCallParams
+    from langgraph_api.api.mcp._request_state import RequestState
+    from langgraph_api.schema import Interrupt
+
+logger = structlog.stdlib.get_logger(__name__)
+
+
+@dataclass(frozen=True, slots=True)
+class CallThread:
+    client: LangGraphClient
+    headers: Mapping[str, str]
+    assistant_id: str
+    thread_id: str
+
+    @classmethod
+    async def create(
+        cls, client: LangGraphClient, headers: Mapping[str, str], assistant_id: str
+    ) -> CallThread:
+        thread = await client.threads.create(
+            ttl=_within_operator_ttl(PAUSED_CALL_TTL_MINUTES), headers=headers
+        )
+        return cls(client, headers, assistant_id, thread["thread_id"])
+
+    async def get(self) -> Thread:
+        with _invalid_state_if_missing():
+            return await self.client.threads.get(self.thread_id, headers=self.headers)
+
+    async def pending(self, thread: Thread) -> list[Interrupt]:
+        if thread["status"] != "interrupted":
+            return []
+        with _invalid_state_if_missing():
+            state = await self.client.threads.get_state(
+                self.thread_id, headers=self.headers
+            )
+        return [
+            interrupt for task in state["tasks"] for interrupt in task["interrupts"]
+        ]
+
+    async def bind(self, binding: Binding, replacing: Binding | None) -> None:
+        if binding == replacing:
+            return
+        if replacing is not None and await self._bound() not in (replacing, binding):
+            return
+        await self.client.threads.update(
+            self.thread_id,
+            metadata=binding.metadata,
+            ttl=_within_operator_ttl(PAUSED_CALL_TTL_MINUTES),
+            headers=self.headers,
+        )
+
+    async def run(self, params: ToolCallParams) -> RunOutcome:
+        value = await self.client.runs.wait(
+            thread_id=self.thread_id,
+            assistant_id=self.assistant_id,
+            input=params.input,
+            context=params.context,
+            durability="exit",
+            headers=self.headers,
+            raise_error=False,
+        )
+        return await self._outcome(value)
+
+    async def resume(
+        self,
+        value: JsonValue,
+        state: RequestState,
+        context: Mapping[str, JsonValue] | None,
+    ) -> RunOutcome:
+        try:
+            await self._create_run(self._resume_payload(value, context), state.run_id)
+        except HTTPException as error:
+            if error.status_code == HTTPStatus.NOT_FOUND:
+                raise InvalidRequestState from None
+            if error.status_code != HTTPStatus.CONFLICT:
+                raise
+            return await self.join(state)
+        return await self._wait_for(state)
+
+    async def join(self, state: RequestState) -> RunOutcome:
+        if not await self.resumed_with(state):
+            raise InvalidRequestState
+        return await self._wait_for(state)
+
+    async def resumed_with(self, state: RequestState) -> bool:
+        try:
+            await self.client.runs.get(
+                self.thread_id, str(state.run_id), headers=self.headers
+            )
+        except httpx.HTTPStatusError as error:
+            if error.response.status_code != HTTPStatus.NOT_FOUND:
+                raise
+            return False
+        return True
+
+    async def delete(self) -> None:
+        async with self._best_effort("Could not delete a finished MCP call's thread"):
+            await self.client.threads.delete(self.thread_id, headers=self.headers)
+
+    async def expire_soon(self) -> None:
+        async with self._best_effort("Could not shorten a finished MCP call's TTL"):
+            await self.client.threads.update(
+                self.thread_id,
+                metadata={},
+                ttl=_within_operator_ttl(FINISHED_CALL_TTL_MINUTES),
+                headers=self.headers,
+            )
+
+    async def _bound(self) -> Binding | None:
+        return Binding.stored((await self.get())["metadata"])
+
+    @asynccontextmanager
+    async def _best_effort(self, failure: str) -> AsyncIterator[None]:
+        try:
+            yield
+        except httpx.HTTPStatusError as error:
+            if error.response.status_code != HTTPStatus.NOT_FOUND:
+                await logger.awarning(
+                    failure,
+                    thread_id=self.thread_id,
+                    status_code=error.response.status_code,
+                )
+
+    @retry_db
+    async def _create_run(self, payload: dict[str, JsonValue], run_id: UUID) -> None:
+        async with connect() as conn:
+            await create_valid_run(
+                conn, self.thread_id, payload, self.headers, run_id=run_id
+            )
+
+    async def _wait_for(self, state: RequestState) -> RunOutcome:
+        with _invalid_state_if_missing():
+            value = await self.client.runs.join(
+                self.thread_id, str(state.run_id), headers=self.headers
+            )
+        return await self._outcome(value)
+
+    async def _outcome(self, value: JsonValue) -> RunOutcome:
+        outcome = parse_outcome(value)
+        if isinstance(outcome, Failed):
+            return outcome
+        thread = await self.get()
+        match thread["status"]:
+            case "busy":
+                raise InvalidRequestState
+            case "interrupted":
+                return await self._paused_or_stopped(thread)
+            case _:
+                return outcome
+
+    async def _paused_or_stopped(self, thread: Thread) -> Paused | Failed:
+        if interrupts := await self.pending(thread):
+            return Paused(tuple(interrupts))
+        return Failed(BREAKPOINT_MESSAGE)
+
+    def _resume_payload(
+        self, value: JsonValue, context: Mapping[str, JsonValue] | None
+    ) -> dict[str, JsonValue]:
+        payload: dict[str, JsonValue] = {
+            "assistant_id": self.assistant_id,
+            "command": {"resume": value},
+            "multitask_strategy": "reject",
+            "durability": "exit",
+        }
+        if context is not None:
+            payload["context"] = dict(context)
+        try:
+            sanitize_and_validate(payload, RunCreateStateful)
+        except jsonschema_rs.ValidationError:
+            raise MCPError(types.INVALID_PARAMS, "Invalid tool call context") from None
+        return payload
+
+
+def _within_operator_ttl(minutes: float) -> float:
+    operator_ttl = THREAD_TTL.get("default_ttl") if THREAD_TTL else None
+    return minutes if operator_ttl is None else min(minutes, operator_ttl)
+
+
+@contextmanager
+def _invalid_state_if_missing() -> Iterator[None]:
+    try:
+        yield
+    except httpx.HTTPStatusError as error:
+        if error.response.status_code != HTTPStatus.NOT_FOUND:
+            raise
+        raise InvalidRequestState from None

langgraph_api/api/mcp/_calls.py

@@ -1,6 +1,20 @@
 import mcp_types as types
 from pydantic import JsonValue
 
+from langgraph_api.api.mcp._digest import digest
+
 
 class ToolCallParams(types.CallToolRequestParams):
+    arguments: dict[str, JsonValue] | None = None
     context: dict[str, JsonValue] | None = None
+
+    @property
+    def input(self) -> dict[str, JsonValue]:
+        return self.arguments or {}
+
+    @property
+    def answers(self) -> types.InputResponses:
+        return self.input_responses or {}
+
+    def digest(self, assistant_id: str) -> str:
+        return digest(assistant_id, self.name, self.input)

langgraph_api/api/mcp/_constants.py

@@ -9,3 +9,17 @@ PAUSED_TOOL_MESSAGE = (
 RUN_FAILED_MESSAGE = "The tool's run failed."
 INVALID_CURSOR_MESSAGE = "Invalid cursor"
 INTERNAL_ERROR_MESSAGE = "Internal server error"
+UNRESUMABLE_PAUSE_MESSAGE = (
+    "This tool paused to ask for human input, and this call can't be resumed."
+)
+ELICITATION_REQUIRED_MESSAGE = (
+    "This tool asks for human input and needs {modes} elicitation."
+)
+INVALID_REQUEST_STATE = "Invalid or expired requestState"
+UNASKABLE_INPUT_MESSAGE = "This tool needs input that can't be asked over MCP"
+BREAKPOINT_MESSAGE = (
+    "This tool stopped at a breakpoint (interrupt_before or interrupt_after), "
+    "which can't be resumed over MCP."
+)
+PAUSED_CALL_TTL_MINUTES = 24 * 60
+FINISHED_CALL_TTL_MINUTES = 10

langgraph_api/api/mcp/_digest.py

@@ -0,0 +1,11 @@
+import hashlib
+
+import orjson
+from pydantic import JsonValue
+
+KEY_DIGEST_LENGTH = 16
+
+
+def digest(*parts: JsonValue) -> str:
+    encoded = orjson.dumps(parts, option=orjson.OPT_SORT_KEYS)
+    return hashlib.sha256(encoded).hexdigest()[:KEY_DIGEST_LENGTH]