langgraph-api 0.4.9__tar.gz → 0.4.14__tar.gz

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: langgraph-api
-Version: 0.4.9
+Version: 0.4.14
 Author-email: Nuno Campos <nuno@langchain.dev>, Will Fu-Hinthorn <will@langchain.dev>
 License: Elastic-2.0
 License-File: LICENSE
@@ -11,7 +11,7 @@ Requires-Dist: httpx>=0.25.0
 Requires-Dist: jsonschema-rs<0.30,>=0.20.0
 Requires-Dist: langchain-core>=0.3.64
 Requires-Dist: langgraph-checkpoint>=2.0.23
-Requires-Dist: langgraph-runtime-inmem<0.11.0,>=0.10.0
+Requires-Dist: langgraph-runtime-inmem<0.13.0,>=0.12.0
 Requires-Dist: langgraph-sdk>=0.2.0
 Requires-Dist: langgraph>=0.4.0
 Requires-Dist: langsmith>=0.3.45

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/benchmark/ramp.js

@@ -56,7 +56,7 @@ export let options = {
   },
   thresholds: {
     'run_duration': [`p(95)<${p95_run_duration[MODE]}`],
-    'successful_runs': [`count>${(PLATEAU_DURATION / (p95_run_duration[MODE] / 1000)) * LOAD_SIZE * LEVELS * 2}`],  // Number of expected successful runs per user worst caseduring plateau * max number of users * 2 cause that feels about right
+    'successful_runs': [`count>${(PLATEAU_DURATION / (p95_run_duration[MODE] / 1000)) * LOAD_SIZE * LEVELS * 2}`],  // Number of expected successful runs per user worst case during plateau * max number of users * 2 cause that feels about right
     'http_req_failed': ['rate<0.01'],   // Error rate should be less than 1%
   },
 };
@@ -109,10 +109,16 @@ export default function() {
 
     // Check the response
     const expected_length = MODE === 'single' ? 1 : EXPAND + 1;
-    const success = check(response, {
-      'Run completed successfully': (r) => r.status === 200,
-      'Response contains expected number of messages': (r) => JSON.parse(r.body)?.messages?.length === expected_length,
-    });
+    let success = false;
+    try {
+      success = check(response, {
+        'Run completed successfully': (r) => r.status === 200,
+        'Response contains expected number of messages': (r) => JSON.parse(r.body)?.messages?.length === expected_length,
+      });
+    } catch (error) {
+      console.log(`Error checking response: ${error}`);
+    }
+
 
     if (success) {
       // Record success metrics
@@ -126,34 +132,23 @@ export default function() {
       if (response.status >= 500) {
         serverErrors.add(1);
         console.log(`Server error: ${response.status}`);
-      } else if (response.status === 408 || response.error === 'timeout') {
+      } else if (response.status === 408 || response.error?.includes('timeout')) {
         timeoutErrors.add(1);
         console.log(`Timeout error: ${response.error}`);
-      } else if (response.status === 200 && response?.body?.messages?.length !== expected_length) {
+      } else if (response.status === 200 && response.body?.messages?.length !== expected_length) {
         missingMessageErrors.add(1);
-        console.log(response);
-        console.log(`Missing message error: Status ${response.status}, ${JSON.stringify(response.body)}`);
+        console.log(`Missing message error: Status ${response.status}, ${JSON.stringify(response.body)}, ${response.headers?.['Content-Location']}`);
       } else {
         otherErrors.add(1);
         console.log(`Other error: Status ${response.status}, ${JSON.stringify(response.body)}`);
       }
     }
   } catch (error) {
-    // Handle exceptions (network errors, etc.)
+    // Handle truly unexpected errors
     failedRuns.add(1);
-
-    if (error.message.includes('timeout')) {
-      timeoutErrors.add(1);
-      console.log(`Timeout error: ${error.message}`);
-    } else if (error.message.includes('connection') || error.message.includes('network')) {
-      connectionErrors.add(1);
-      console.log(`Connection error: ${error.message}`);
-    } else {
-      otherErrors.add(1);
-      // Usually we end up with HTML error pages here
-      console.log(response);
-      console.log(`Unexpected error: ${error.message}, Response Body: ${response?.body}`);
-    }
+    otherErrors.add(1);
+    console.log(response);
+    console.log(`Unexpected error: ${error.message}, Response Body: ${response?.body}`);
   }
 
   // Add a small random sleep between iterations to prevent thundering herd

langgraph_api-0.4.14/langgraph_api/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.4.14"

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/langgraph_api/api/a2a.py

@@ -171,7 +171,7 @@ async def _validate_supports_messages(
     """
     assistant_id = assistant["assistant_id"]
 
-    cached_schemas = _assistant_schemas_cache.get(assistant_id)
+    cached_schemas = await _assistant_schemas_cache.get(assistant_id)
     if cached_schemas is not None:
         schemas = cached_schemas
     else:

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/langgraph_api/api/meta.py

@@ -1,6 +1,7 @@
 from typing import cast
 
 import langgraph.version
+import structlog
 from starlette.responses import JSONResponse, PlainTextResponse
 
 from langgraph_api import __version__, config, metadata
@@ -13,6 +14,8 @@ from langgraph_runtime.ops import Runs
 
 METRICS_FORMATS = {"prometheus", "json"}
 
+logger = structlog.stdlib.get_logger(__name__)
+
 
 async def meta_info(request: ApiRequest):
     plus = plus_features_enabled()
@@ -71,35 +74,44 @@ async def meta_metrics(request: ApiRequest):
                 resp["workers"] = worker_metrics
             return JSONResponse(resp)
     elif metrics_format == "prometheus":
-        async with connect() as conn:
-            queue_stats = await Runs.stats(conn)
-
-            metrics = [
-                "# HELP lg_api_num_pending_runs The number of runs currently pending.",
-                "# TYPE lg_api_num_pending_runs gauge",
-                f'lg_api_num_pending_runs{{project_id="{metadata.PROJECT_ID}", revision_id="{metadata.HOST_REVISION_ID}"}} {queue_stats["n_pending"]}',
-                "# HELP lg_api_num_running_runs The number of runs currently running.",
-                "# TYPE lg_api_num_running_runs gauge",
-                f'lg_api_num_running_runs{{project_id="{metadata.PROJECT_ID}", revision_id="{metadata.HOST_REVISION_ID}"}} {queue_stats["n_running"]}',
-            ]
+        metrics = []
+        try:
+            async with connect() as conn:
+                queue_stats = await Runs.stats(conn)
 
-            if config.N_JOBS_PER_WORKER > 0:
                 metrics.extend(
                     [
-                        "# HELP lg_api_workers_max The maximum number of workers available.",
-                        "# TYPE lg_api_workers_max gauge",
-                        f'lg_api_workers_max{{project_id="{metadata.PROJECT_ID}", revision_id="{metadata.HOST_REVISION_ID}"}} {workers_max}',
-                        "# HELP lg_api_workers_active The number of currently active workers.",
-                        "# TYPE lg_api_workers_active gauge",
-                        f'lg_api_workers_active{{project_id="{metadata.PROJECT_ID}", revision_id="{metadata.HOST_REVISION_ID}"}} {workers_active}',
-                        "# HELP lg_api_workers_available The number of available (idle) workers.",
-                        "# TYPE lg_api_workers_available gauge",
-                        f'lg_api_workers_available{{project_id="{metadata.PROJECT_ID}", revision_id="{metadata.HOST_REVISION_ID}"}} {workers_available}',
+                        "# HELP lg_api_num_pending_runs The number of runs currently pending.",
+                        "# TYPE lg_api_num_pending_runs gauge",
+                        f'lg_api_num_pending_runs{{project_id="{metadata.PROJECT_ID}", revision_id="{metadata.HOST_REVISION_ID}"}} {queue_stats["n_pending"]}',
+                        "# HELP lg_api_num_running_runs The number of runs currently running.",
+                        "# TYPE lg_api_num_running_runs gauge",
+                        f'lg_api_num_running_runs{{project_id="{metadata.PROJECT_ID}", revision_id="{metadata.HOST_REVISION_ID}"}} {queue_stats["n_running"]}',
                     ]
                 )
+        except Exception as e:
+            # if we get a db connection error/timeout, just skip queue stats
+            await logger.awarning(
+                "Ignoring error while getting run stats for /metrics", exc_info=e
+            )
+
+        if config.N_JOBS_PER_WORKER > 0:
+            metrics.extend(
+                [
+                    "# HELP lg_api_workers_max The maximum number of workers available.",
+                    "# TYPE lg_api_workers_max gauge",
+                    f'lg_api_workers_max{{project_id="{metadata.PROJECT_ID}", revision_id="{metadata.HOST_REVISION_ID}"}} {workers_max}',
+                    "# HELP lg_api_workers_active The number of currently active workers.",
+                    "# TYPE lg_api_workers_active gauge",
+                    f'lg_api_workers_active{{project_id="{metadata.PROJECT_ID}", revision_id="{metadata.HOST_REVISION_ID}"}} {workers_active}',
+                    "# HELP lg_api_workers_available The number of available (idle) workers.",
+                    "# TYPE lg_api_workers_available gauge",
+                    f'lg_api_workers_available{{project_id="{metadata.PROJECT_ID}", revision_id="{metadata.HOST_REVISION_ID}"}} {workers_available}',
+                ]
+            )
 
-            metrics.extend(http_metrics)
-            metrics.extend(pg_redis_stats)
+        metrics.extend(http_metrics)
+        metrics.extend(pg_redis_stats)
 
         metrics_response = "\n".join(metrics)
         return PlainTextResponse(metrics_response)

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/langgraph_api/api/threads.py

@@ -290,7 +290,12 @@ async def patch_thread(
     validate_uuid(thread_id, "Invalid thread ID: must be a UUID")
     payload = await request.json(ThreadPatch)
     async with connect() as conn:
-        thread = await Threads.patch(conn, thread_id, metadata=payload["metadata"])
+        thread = await Threads.patch(
+            conn,
+            thread_id,
+            metadata=payload.get("metadata", {}),
+            ttl=payload.get("ttl"),
+        )
     return ApiResponse(await fetchone(thread))
 
 

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/langgraph_api/asyncio.py

@@ -158,6 +158,7 @@ class SimpleTaskGroup(AbstractAsyncContextManager["SimpleTaskGroup"]):
         self,
         *coros: Coroutine[Any, Any, T],
         cancel: bool = False,
+        cancel_event: asyncio.Event | None = None,
         wait: bool = True,
         taskset: set[asyncio.Task] | None = None,
         taskgroup_name: str | None = None,
@@ -165,6 +166,7 @@ class SimpleTaskGroup(AbstractAsyncContextManager["SimpleTaskGroup"]):
         # Copy the taskset to avoid modifying the original set unintentionally (like in lifespan)
         self.tasks = taskset.copy() if taskset is not None else set()
         self.cancel = cancel
+        self.cancel_event = cancel_event
         self.wait = wait
         if taskset:
             for task in tuple(taskset):
@@ -181,6 +183,8 @@ class SimpleTaskGroup(AbstractAsyncContextManager["SimpleTaskGroup"]):
         try:
             if (exc := task.exception()) and not isinstance(exc, ignore_exceptions):
                 logger.exception("asyncio.task failed in task group", exc_info=exc)
+                if self.cancel_event:
+                    self.cancel_event.set()
         except asyncio.CancelledError:
             pass
 

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/langgraph_api/auth/langsmith/backend.py

@@ -58,7 +58,7 @@ class LangsmithAuthBackend(AuthenticationBackend):
 
         # Check cache first
         cache_key = self._get_cache_key(headers)
-        if cached_entry := self._cache.get(cache_key):
+        if cached_entry := await self._cache.get(cache_key):
             return cached_entry["credentials"], cached_entry["user"]
 
         async with auth_client() as auth:

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/langgraph_api/command.py

@@ -1,3 +1,5 @@
+from typing import cast
+
 from langgraph.types import Command, Send
 
 from langgraph_api.schema import RunCommand
@@ -11,9 +13,9 @@ def map_cmd(cmd: RunCommand) -> Command:
     update = cmd.get("update")
     if isinstance(update, tuple | list) and all(
         isinstance(t, tuple | list) and len(t) == 2 and isinstance(t[0], str)
-        for t in update
+        for t in cast(list, update)
     ):
-        update = [tuple(t) for t in update]
+        update = [tuple(t) for t in cast(list, update)]
 
     return Command(
         update=update,

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/langgraph_api/graph.py

@@ -392,14 +392,14 @@ async def collect_graphs_from_env(register: bool = False) -> None:
 
         if (
             config.HTTP_CONFIG
-            and config.HTTP_CONFIG.get("app")
-            and is_js_path(config.HTTP_CONFIG.get("app").split(":")[0])
+            and (js_app := config.HTTP_CONFIG.get("app"))
+            and is_js_path(js_app.split(":")[0])
         ):
             js_bg_tasks.add(
                 asyncio.create_task(
                     run_js_http_process(
                         paths_str,
-                        config.HTTP_CONFIG.get("app"),
+                        config.HTTP_CONFIG or {},
                         watch="--reload" in sys.argv[1:],
                     ),
                 )

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/langgraph_api/js/remote.py

@@ -153,9 +153,9 @@ class RemotePregel(BaseRemotePregel):
 
         async for event in _client_stream("streamEvents", data):
             if event["event"] == "on_custom_event":
-                yield CustomStreamEvent(**event)
+                yield CustomStreamEvent(**event)  # type: ignore[missing-typed-dict-key]
             else:
-                yield StandardStreamEvent(**event)
+                yield StandardStreamEvent(**event)  # type: ignore[missing-typed-dict-key]
 
     async def fetch_state_schema(self):
         return await _client_invoke("getSchema", {"graph_id": self.graph_id})
@@ -187,15 +187,17 @@ class RemotePregel(BaseRemotePregel):
                 )
                 for data in nodes
             },
-            {
-                Edge(
-                    data["source"],
-                    data["target"],
-                    data.get("data"),
-                    data.get("conditional", False),
-                )
-                for data in edges
-            },
+            list(
+                {
+                    Edge(
+                        data["source"],
+                        data["target"],
+                        data.get("data"),
+                        data.get("conditional", False),
+                    )
+                    for data in edges
+                }
+            ),
         )
 
     async def fetch_subgraphs(
@@ -861,6 +863,8 @@ class CustomJsAuthBackend(AuthenticationBackend):
             self.ls_auth = LangsmithAuthBackend()
         self.ttl_cache: LRUCache | None = None
         self.cache_keys: list[str] | None = None
+        if LANGGRAPH_AUTH is None:
+            raise ValueError("LANGGRAPH_AUTH is not set")
         if cache := LANGGRAPH_AUTH.get("cache"):
             keys = cache.get("cache_keys", [])
             if not isinstance(keys, list):
@@ -891,7 +895,7 @@ class CustomJsAuthBackend(AuthenticationBackend):
         if self.cache_keys:
             cache_key = tuple((k, headers[k]) for k in self.cache_keys if k in headers)
             if cache_key and self.ttl_cache is not None:
-                cached = self.ttl_cache.get(cache_key)
+                cached = await self.ttl_cache.get(cache_key)
                 if cached:
                     return cached
 

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/langgraph_api/queue_entrypoint.py

@@ -86,6 +86,7 @@ async def health_and_metrics_server():
         log_level="error",
         access_log=False,
     )
+    # Server will run indefinitely until the process is terminated
     server = uvicorn.Server(config)
 
     logger.info(f"Health and metrics server started at http://0.0.0.0:{port}")
@@ -93,14 +94,15 @@ async def health_and_metrics_server():
 
 
 async def entrypoint(
-    grpc_port: int | None = None, entrypoint_name: str = "python-queue"
+    grpc_port: int | None = None,
+    entrypoint_name: str = "python-queue",
+    cancel_event: asyncio.Event | None = None,
 ):
     from langgraph_api import logging as lg_logging
     from langgraph_api.api import user_router
 
     lg_logging.set_logging_context({"entrypoint": entrypoint_name})
     tasks: set[asyncio.Task] = set()
-    tasks.add(asyncio.create_task(health_and_metrics_server()))
 
     original_lifespan = user_router.router.lifespan_context if user_router else None
 
@@ -113,6 +115,7 @@ async def entrypoint(
             with_cron_scheduler=with_cron_scheduler,
             grpc_port=grpc_port,
             taskset=taskset,
+            cancel_event=cancel_event,
         ):
             if original_lifespan:
                 async with original_lifespan(app):
@@ -123,6 +126,7 @@ async def entrypoint(
     async with combined_lifespan(
         None, with_cron_scheduler=False, grpc_port=grpc_port, taskset=tasks
     ):
+        tasks.add(asyncio.create_task(health_and_metrics_server()))
         await asyncio.gather(*tasks)
 
 
@@ -141,8 +145,14 @@ async def main(grpc_port: int | None = None, entrypoint_name: str = "python-queu
         signal.signal(signal.SIGTERM, lambda *_: _handle_signal())
 
     entry_task = asyncio.create_task(
-        entrypoint(grpc_port=grpc_port, entrypoint_name=entrypoint_name)
+        entrypoint(
+            grpc_port=grpc_port,
+            entrypoint_name=entrypoint_name,
+            cancel_event=stop_event,
+        )
     )
+    # Handle the case where the entrypoint errors out
+    entry_task.add_done_callback(lambda _: stop_event.set())
     await stop_event.wait()
 
     logger.warning("Cancelling queue entrypoint task")

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/langgraph_api/serde.py

@@ -1,5 +1,4 @@
 import asyncio
-import base64
 import re
 import uuid
 from base64 import b64encode
@@ -178,21 +177,3 @@ class Serializer(JsonPlusSerializer):
 
 mpack_keys = {"method", "value"}
 SERIALIZER = Serializer()
-
-
-# TODO: Make more performant (by removing)
-async def reserialize_message(message: bytes) -> bytes:
-    # Stream messages from golang runtime are a byte dict of StreamChunks.
-    loaded = await ajson_loads(message)
-    converted = {}
-    for k, v in loaded.items():
-        if isinstance(v, dict) and v.keys() == mpack_keys:
-            if v["method"] == "missing":
-                converted[k] = v["value"]  # oops
-            else:
-                converted[k] = SERIALIZER.loads_typed(
-                    (v["method"], base64.b64decode(v["value"]))
-                )
-        else:
-            converted[k] = v
-    return json_dumpb(converted)

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/langgraph_api/state.py

@@ -27,7 +27,7 @@ def runnable_config_to_checkpoint(
         return None
 
     configurable = config["configurable"]
-    checkpoint: Checkpoint = {
+    checkpoint: Checkpoint = {  # type: ignore[typed-dict-item]
         "checkpoint_id": configurable["checkpoint_id"],
         "thread_id": configurable["thread_id"],
     }

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/langgraph_api/thread_ttl.py

@@ -1,6 +1,7 @@
 """Sweeping logic for cleaning up expired threads and checkpoints."""
 
 import asyncio
+from typing import cast
 
 import structlog
 
@@ -23,7 +24,9 @@ async def thread_ttl_sweep_loop():
         raise NotImplementedError(
             f"Unrecognized thread deletion strategy: {strategy}. Expected 'delete'."
         )
-    sweep_interval_minutes = thread_ttl_config.get("sweep_interval_minutes", 5)
+    sweep_interval_minutes = cast(
+        int, thread_ttl_config.get("sweep_interval_minutes", 5)
+    )
     await logger.ainfo(
         f"Starting thread TTL sweeper with interval {sweep_interval_minutes} minutes",
         strategy=strategy,

langgraph_api-0.4.14/langgraph_api/utils/cache.py

@@ -0,0 +1,95 @@
+import asyncio
+import time
+from collections import OrderedDict
+from collections.abc import Awaitable, Callable
+from typing import Generic, TypeVar
+
+T = TypeVar("T")
+
+
+class LRUCache(Generic[T]):
+    """LRU cache with TTL and proactive refresh support."""
+
+    def __init__(
+        self,
+        max_size: int = 1000,
+        ttl: float = 60,
+        refresh_window: float = 30,
+        refresh_callback: Callable[[str], Awaitable[T | None]] | None = None,
+    ):
+        self._cache: OrderedDict[str, tuple[T, float, bool]] = OrderedDict()
+        self._max_size = max_size if max_size > 0 else 1000
+        self._ttl = ttl
+        self._refresh_window = refresh_window if refresh_window > 0 else 30
+        self._refresh_callback = refresh_callback
+
+    def _get_time(self) -> float:
+        """Get current time, using loop.time() if available for better performance."""
+        try:
+            return asyncio.get_event_loop().time()
+        except RuntimeError:
+            return time.monotonic()
+
+    async def get(self, key: str) -> T | None:
+        """Get item from cache, attempting refresh if within refresh window."""
+        if key not in self._cache:
+            return None
+
+        value, timestamp, is_refreshing = self._cache[key]
+        current_time = self._get_time()
+        time_until_expiry = self._ttl - (current_time - timestamp)
+
+        # Check if expired
+        if time_until_expiry <= 0:
+            del self._cache[key]
+            return None
+
+        # Check if we should attempt refresh (within refresh window and not already refreshing)
+        if (
+            time_until_expiry <= self._refresh_window
+            and not is_refreshing
+            and self._refresh_callback
+        ):
+            # Mark as refreshing to prevent multiple simultaneous refresh attempts
+            self._cache[key] = (value, timestamp, True)
+
+            try:
+                # Attempt refresh
+                refreshed_value = await self._refresh_callback(key)
+                if refreshed_value is not None:
+                    # Refresh successful, update cache with new value
+                    self._cache[key] = (refreshed_value, current_time, False)
+                    # Move to end (most recently used)
+                    self._cache.move_to_end(key)
+                    return refreshed_value
+                else:
+                    # Refresh failed, fallback to cached value
+                    self._cache[key] = (value, timestamp, False)
+            except Exception:
+                # Refresh failed with exception, fallback to cached value
+                self._cache[key] = (value, timestamp, False)
+
+        # Move to end (most recently used)
+        self._cache.move_to_end(key)
+        return value
+
+    def set(self, key: str, value: T) -> None:
+        """Set item in cache, evicting old entries if needed."""
+        # Remove if already exists (to update timestamp)
+        if key in self._cache:
+            del self._cache[key]
+
+        # Evict oldest entries if needed
+        while len(self._cache) >= self._max_size:
+            self._cache.popitem(last=False)  # Remove oldest (FIFO)
+
+        # Add new entry (not refreshing initially)
+        self._cache[key] = (value, self._get_time(), False)
+
+    def size(self) -> int:
+        """Return current cache size."""
+        return len(self._cache)
+
+    def clear(self) -> None:
+        """Clear all entries from cache."""
+        self._cache.clear()

langgraph_api-0.4.14/langgraph_api/utils/retriable_client.py

@@ -0,0 +1,74 @@
+import asyncio
+
+import httpx
+import structlog
+
+logger = structlog.stdlib.get_logger(__name__)
+
+
+async def _make_http_request_with_retries(
+    url: str,
+    headers: dict,
+    method: str = "GET",
+    json_data: dict | None = None,
+    max_retries: int = 3,
+    base_delay: float = 1.0,
+) -> httpx.Response | None:
+    """
+    Make an HTTP request with exponential backoff retries.
+
+    Args:
+        url: The URL to request
+        headers: Headers to include in the request
+        method: HTTP method ("GET" or "POST")
+        json_data: JSON data for POST requests
+        max_retries: Maximum number of retry attempts
+        base_delay: Base delay in seconds for exponential backoff
+
+    Returns:
+        httpx.Response: The successful response
+
+    Raises:
+        httpx.HTTPStatusError: If the request fails after all retries
+        httpx.RequestError: If the request fails after all retries
+    """
+    for attempt in range(max_retries + 1):
+        try:
+            async with httpx.AsyncClient(timeout=10.0) as client:
+                response = await client.request(
+                    method, url, headers=headers, json=json_data
+                )
+                response.raise_for_status()
+                return response
+
+        except (
+            httpx.TimeoutException,
+            httpx.NetworkError,
+            httpx.RequestError,
+            httpx.HTTPStatusError,
+        ) as e:
+            if isinstance(e, httpx.HTTPStatusError) and e.response.status_code < 500:
+                # Don't retry on 4xx errors, but do on 5xxs
+                raise e
+
+            # Back off and retry if we haven't reached the max retries
+            if attempt < max_retries:
+                delay = base_delay * (2**attempt)  # Exponential backoff
+                logger.warning(
+                    "HTTP %s request attempt %d to %s failed: %s. Retrying in %.1f seconds...",
+                    method,
+                    attempt + 1,
+                    url,
+                    e,
+                    delay,
+                )
+                await asyncio.sleep(delay)
+            else:
+                logger.exception(
+                    "HTTP %s request to %s failed after %d attempts. Last error: %s",
+                    method,
+                    url,
+                    max_retries + 1,
+                    e,
+                )
+                raise e

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/langgraph_api/worker.py

@@ -153,7 +153,7 @@ async def worker(
                 raise UserTimeout(e) from e
             raise
 
-    async with Runs.enter(run_id, run["thread_id"], main_loop) as done:
+    async with Runs.enter(run_id, run["thread_id"], main_loop, resumable) as done:
         # attempt the run
         try:
             if attempt > BG_JOB_MAX_RETRIES:

{langgraph_api-0.4.9 → langgraph_api-0.4.14}/openapi.json

@@ -5261,11 +5261,30 @@
             "type": "object",
             "title": "Metadata",
             "description": "Metadata to merge with existing thread metadata."
+          },
+          "ttl": {
+            "type": "object",
+            "title": "TTL",
+            "description": "The time-to-live for the thread.",
+            "properties": {
+              "strategy": {
+                "type": "string",
+                "enum": [
+                  "delete"
+                ],
+                "description": "The TTL strategy. 'delete' removes the entire thread.",
+                "default": "delete"
+              },
+              "ttl": {
+                "type": "number",
+                "description": "The time-to-live in minutes from now until thread should be swept."
+              }
+            }
           }
         },
         "type": "object",
         "title": "ThreadPatch",
-        "description": "Payload for creating a thread."
+        "description": "Payload for updating a thread."
       },
       "ThreadStateCheckpointRequest": {
         "properties": {