Merge pull request MODSetter#1291 from MODSetter/dev

fix: agent runtime issues

Author: MODSetter

surfsense_backend/app/agents/new_chat/chat_deepagent.py

@@ -24,7 +24,6 @@
 from deepagents.graph import BASE_AGENT_PROMPT
 from deepagents.middleware.patch_tool_calls import PatchToolCallsMiddleware
 from deepagents.middleware.subagents import GENERAL_PURPOSE_SUBAGENT
-from deepagents.middleware.summarization import create_summarization_middleware
 from langchain.agents import create_agent
 from langchain.agents.middleware import TodoListMiddleware
 from langchain_anthropic.middleware import AnthropicPromptCachingMiddleware
@@ -41,6 +40,9 @@
     MemoryInjectionMiddleware,
     SurfSenseFilesystemMiddleware,
 )
+from app.agents.new_chat.middleware.safe_summarization import (
+    create_safe_summarization_middleware,
+)
 from app.agents.new_chat.system_prompt import (
     build_configurable_system_prompt,
     build_surfsense_system_prompt,
@@ -442,7 +444,7 @@ async def create_surfsense_deep_agent(
             created_by_id=user_id,
             thread_id=thread_id,
         ),
-        create_summarization_middleware(llm, StateBackend),
+        create_safe_summarization_middleware(llm, StateBackend),
         PatchToolCallsMiddleware(),
         AnthropicPromptCachingMiddleware(unsupported_model_behavior="ignore"),
     ]
@@ -472,7 +474,7 @@ async def create_surfsense_deep_agent(
             thread_id=thread_id,
         ),
         SubAgentMiddleware(backend=StateBackend, subagents=[general_purpose_spec]),
-        create_summarization_middleware(llm, StateBackend),
+        create_safe_summarization_middleware(llm, StateBackend),
         PatchToolCallsMiddleware(),
         DedupHITLToolCallsMiddleware(agent_tools=tools),
         AnthropicPromptCachingMiddleware(unsupported_model_behavior="ignore"),

surfsense_backend/app/agents/new_chat/middleware/safe_summarization.py

@@ -0,0 +1,123 @@
+"""Safe wrapper around deepagents' SummarizationMiddleware.
+
+Upstream issue
+--------------
+`deepagents.middleware.summarization.SummarizationMiddleware._aoffload_to_backend`
+(and its sync counterpart) call
+``get_buffer_string(filtered_messages)`` before writing the evicted history
+to the backend file. In recent ``langchain-core`` versions, ``get_buffer_string``
+accesses ``m.text`` which iterates ``self.content`` — this raises
+``TypeError: 'NoneType' object is not iterable`` whenever an ``AIMessage``
+has ``content=None`` (common when a model returns *only* tool_calls, seen
+frequently with Azure OpenAI ``gpt-5.x`` responses streamed through
+LiteLLM).
+
+The exception aborts the whole agent turn, so the user just sees "Error during
+chat" with no assistant response.
+
+Fix
+---
+We subclass ``SummarizationMiddleware`` and override
+``_filter_summary_messages`` — the only call site that feeds messages into
+``get_buffer_string`` — to return *copies* of messages whose ``content`` is
+``None`` with ``content=""``. The originals flowing through the rest of the
+agent state are untouched.
+
+We also expose a drop-in ``create_safe_summarization_middleware`` factory
+that mirrors ``deepagents.middleware.summarization.create_summarization_middleware``
+but instantiates our safe subclass.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING
+
+from deepagents.middleware.summarization import (
+    SummarizationMiddleware,
+    compute_summarization_defaults,
+)
+
+if TYPE_CHECKING:
+    from deepagents.backends.protocol import BACKEND_TYPES
+    from langchain_core.language_models import BaseChatModel
+    from langchain_core.messages import AnyMessage
+
+logger = logging.getLogger(__name__)
+
+
+def _sanitize_message_content(msg: AnyMessage) -> AnyMessage:
+    """Return ``msg`` with ``content`` coerced to a non-``None`` value.
+
+    ``get_buffer_string`` reads ``m.text`` which iterates ``self.content``;
+    when a provider streams back an ``AIMessage`` with only tool_calls and
+    no text, ``content`` can be ``None`` and the iteration explodes. We
+    replace ``None`` with an empty string so downstream consumers that only
+    care about text see an empty body.
+
+    The original message is left untouched — we return a copy via
+    pydantic's ``model_copy`` when available, otherwise we fall back to
+    re-setting the attribute on a shallow copy.
+    """
+
+    if getattr(msg, "content", "not-missing") is not None:
+        return msg
+
+    try:
+        return msg.model_copy(update={"content": ""})
+    except AttributeError:
+        import copy
+
+        new_msg = copy.copy(msg)
+        try:
+            new_msg.content = ""
+        except Exception:  # pragma: no cover - defensive
+            logger.debug(
+                "Could not sanitize content=None on message of type %s",
+                type(msg).__name__,
+            )
+            return msg
+        return new_msg
+
+
+class SafeSummarizationMiddleware(SummarizationMiddleware):
+    """`SummarizationMiddleware` that tolerates messages with ``content=None``.
+
+    Only ``_filter_summary_messages`` is overridden — this is the single
+    helper invoked by both the sync and async offload paths immediately
+    before ``get_buffer_string``. Normalising here means we get coverage
+    for both without having to copy the (long, rapidly-changing) offload
+    implementations from upstream.
+    """
+
+    def _filter_summary_messages(self, messages: list[AnyMessage]) -> list[AnyMessage]:
+        filtered = super()._filter_summary_messages(messages)
+        return [_sanitize_message_content(m) for m in filtered]
+
+
+def create_safe_summarization_middleware(
+    model: BaseChatModel,
+    backend: BACKEND_TYPES,
+) -> SafeSummarizationMiddleware:
+    """Drop-in replacement for ``create_summarization_middleware``.
+
+    Mirrors the defaults computed by ``deepagents`` but returns our
+    ``SafeSummarizationMiddleware`` subclass so the
+    ``content=None`` crash in ``get_buffer_string`` is avoided.
+    """
+
+    defaults = compute_summarization_defaults(model)
+    return SafeSummarizationMiddleware(
+        model=model,
+        backend=backend,
+        trigger=defaults["trigger"],
+        keep=defaults["keep"],
+        trim_tokens_to_summarize=None,
+        truncate_args_settings=defaults["truncate_args_settings"],
+    )
+
+
+__all__ = [
+    "SafeSummarizationMiddleware",
+    "create_safe_summarization_middleware",
+]

surfsense_backend/app/services/llm_router_service.py

@@ -133,6 +133,44 @@ def _sanitize_content(content: Any) -> Any:
 }
 
 
+# Default ``api_base`` per LiteLLM provider prefix.  Used as a safety net when
+# a global LLM config does *not* specify ``api_base``: without this, LiteLLM
+# happily picks up provider-agnostic env vars (e.g. ``AZURE_API_BASE``,
+# ``OPENAI_API_BASE``) and routes, say, an ``openrouter/anthropic/claude-3-haiku``
+# request to an Azure endpoint, which then 404s with ``Resource not found``.
+# Only providers with a well-known, stable public base URL are listed here —
+# self-hosted / BYO-endpoint providers (ollama, custom, bedrock, vertex_ai,
+# huggingface, databricks, cloudflare, replicate) are intentionally omitted
+# so their existing config-driven behaviour is preserved.
+PROVIDER_DEFAULT_API_BASE = {
+    "openrouter": "https://openrouter.ai/api/v1",
+    "groq": "https://api.groq.com/openai/v1",
+    "mistral": "https://api.mistral.ai/v1",
+    "perplexity": "https://api.perplexity.ai",
+    "xai": "https://api.x.ai/v1",
+    "cerebras": "https://api.cerebras.ai/v1",
+    "deepinfra": "https://api.deepinfra.com/v1/openai",
+    "fireworks_ai": "https://api.fireworks.ai/inference/v1",
+    "together_ai": "https://api.together.xyz/v1",
+    "anyscale": "https://api.endpoints.anyscale.com/v1",
+    "cometapi": "https://api.cometapi.com/v1",
+    "sambanova": "https://api.sambanova.ai/v1",
+}
+
+
+# Canonical provider → base URL when a config uses a generic ``openai``-style
+# prefix but the ``provider`` field tells us which API it really is
+# (e.g. DeepSeek/Alibaba/Moonshot/Zhipu/MiniMax all use ``openai`` compat but
+# each has its own base URL).
+PROVIDER_KEY_DEFAULT_API_BASE = {
+    "DEEPSEEK": "https://api.deepseek.com/v1",
+    "ALIBABA_QWEN": "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
+    "MOONSHOT": "https://api.moonshot.ai/v1",
+    "ZHIPU": "https://open.bigmodel.cn/api/paas/v4",
+    "MINIMAX": "https://api.minimax.io/v1",
+}
+
+
 class LLMRouterService:
     """
     Singleton service for managing LiteLLM Router.
@@ -224,6 +262,16 @@ def initialize(
         # hits ContextWindowExceededError.
         full_model_list, ctx_fallbacks = cls._build_context_fallback_groups(model_list)
 
+        # Build a general-purpose fallback list so NotFound/timeout/rate-limit
+        # style failures on one deployment don't bubble up as hard errors —
+        # the router retries with a sibling deployment in ``auto-large``.
+        # ``auto-large`` is the large-context subset of ``auto``; if it is
+        # empty we fall back to ``auto`` itself so the router at least picks a
+        # different deployment in the same group.
+        fallbacks: list[dict[str, list[str]]] | None = None
+        if ctx_fallbacks:
+            fallbacks = [{"auto": ["auto-large"]}]
+
         try:
             router_kwargs: dict[str, Any] = {
                 "model_list": full_model_list,
@@ -237,15 +285,18 @@ def initialize(
             }
             if ctx_fallbacks:
                 router_kwargs["context_window_fallbacks"] = ctx_fallbacks
+            if fallbacks:
+                router_kwargs["fallbacks"] = fallbacks
 
             instance._router = Router(**router_kwargs)
             instance._initialized = True
             logger.info(
                 "LLM Router initialized with %d deployments, "
-                "strategy: %s, context_window_fallbacks: %s",
+                "strategy: %s, context_window_fallbacks: %s, fallbacks: %s",
                 len(model_list),
                 final_settings.get("routing_strategy"),
                 ctx_fallbacks or "none",
+                fallbacks or "none",
             )
         except Exception as e:
             logger.error(f"Failed to initialize LLM Router: {e}")
@@ -348,10 +399,11 @@ def _config_to_deployment(cls, config: dict) -> dict | None:
                 return None
 
             # Build model string
+            provider = config.get("provider", "").upper()
             if config.get("custom_provider"):
-                model_string = f"{config['custom_provider']}/{config['model_name']}"
+                provider_prefix = config["custom_provider"]
+                model_string = f"{provider_prefix}/{config['model_name']}"
             else:
-                provider = config.get("provider", "").upper()
                 provider_prefix = PROVIDER_MAP.get(provider, provider.lower())
                 model_string = f"{provider_prefix}/{config['model_name']}"
 
@@ -361,9 +413,19 @@ def _config_to_deployment(cls, config: dict) -> dict | None:
                 "api_key": config.get("api_key"),
             }
 
-            # Add optional api_base
-            if config.get("api_base"):
-                litellm_params["api_base"] = config["api_base"]
+            # Resolve ``api_base``. Config value wins; otherwise apply a
+            # provider-aware default so the deployment does not silently
+            # inherit unrelated env vars (e.g. ``AZURE_API_BASE``) and route
+            # requests to the wrong endpoint.  See ``PROVIDER_DEFAULT_API_BASE``
+            # docstring for the motivating bug (OpenRouter models 404-ing
+            # against an Azure endpoint).
+            api_base = config.get("api_base")
+            if not api_base:
+                api_base = PROVIDER_KEY_DEFAULT_API_BASE.get(provider)
+            if not api_base:
+                api_base = PROVIDER_DEFAULT_API_BASE.get(provider_prefix)
+            if api_base:
+                litellm_params["api_base"] = api_base
 
             # Add any additional litellm parameters
             if config.get("litellm_params"):

surfsense_backend/pyproject.toml

@@ -74,7 +74,7 @@ dependencies = [
     "deepagents>=0.4.12",
     "stripe>=15.0.0",
     "azure-ai-documentintelligence>=1.0.2",
-    "litellm>=1.83.0",
+    "litellm>=1.83.4",
     "langchain-litellm>=0.6.4",
 ]
 

surfsense_backend/tests/unit/test_error_contract.py

@@ -202,9 +202,7 @@ def test_503_preserves_detail(self, client):
         # Intentional 503s (e.g. feature flag off) must surface the developer
         # message so the frontend can render actionable copy.
         body = _assert_envelope(client.get("/http-503"), 503)
-        assert (
-            body["error"]["message"] == "Page purchases are temporarily unavailable."
-        )
+        assert body["error"]["message"] == "Page purchases are temporarily unavailable."
         assert body["error"]["message"] != GENERIC_5XX_MESSAGE
 
     def test_502_preserves_detail(self, client):

surfsense_backend/uv.lock

@@ -200,8 +200,8 @@ export function DesktopContent() {
 						Launch on Startup
 					</CardTitle>
 					<CardDescription className="text-xs md:text-sm">
-						Automatically start SurfSense when you sign in to your computer so global
-						shortcuts and folder sync are always available.
+						Automatically start SurfSense when you sign in to your computer so global shortcuts and
+						folder sync are always available.
 					</CardDescription>
 				</CardHeader>
 				<CardContent className="px-3 md:px-6 pb-3 md:pb-6 space-y-3">
@@ -232,8 +232,7 @@ export function DesktopContent() {
 								Start minimized to tray
 							</Label>
 							<p className="text-xs text-muted-foreground">
-								Skip the main window on boot — SurfSense lives in the system tray until you need
-								it.
+								Skip the main window on boot — SurfSense lives in the system tray until you need it.
 							</p>
 						</div>
 						<Switch

surfsense_web/app/dashboard/[search_space_id]/user-settings/components/DesktopContent.tsx

@@ -126,9 +126,7 @@ export function PurchaseHistoryContent() {
 		return [
 			...pagePurchases.map(normalizePagePurchase),
 			...tokenPurchases.map(normalizeTokenPurchase),
-		].sort(
-			(a, b) => new Date(b.created_at).getTime() - new Date(a.created_at).getTime()
-		);
+		].sort((a, b) => new Date(b.created_at).getTime() - new Date(a.created_at).getTime());
 	}, [pagesQuery.data, tokensQuery.data]);
 
 	if (isLoading) {