OpenJarvis 1.0.4.dev878__tar.gz → 1.0.4.dev883__tar.gz

{openjarvis-1.0.4.dev878 → openjarvis-1.0.4.dev883}/.github/clone-stats/badge.json

@@ -1,7 +1,7 @@
 {
   "schemaVersion": 1,
   "label": "Git Clones",
-  "message": "140,724",
+  "message": "142,750",
   "color": "green",
   "namedLogo": "git"
 }

{openjarvis-1.0.4.dev878 → openjarvis-1.0.4.dev883}/.github/clone-stats/clone-data.json

@@ -1,6 +1,6 @@
 {
-  "total_clones": 140724,
-  "last_updated": "2026-07-02T07:10:21Z",
+  "total_clones": 142750,
+  "last_updated": "2026-07-06T10:14:21Z",
   "daily": {
     "2026-03-27": 2189,
     "2026-03-28": 1874,
@@ -98,6 +98,10 @@
     "2026-06-28": 1028,
     "2026-06-29": 765,
     "2026-06-30": 951,
-    "2026-07-01": 1134
+    "2026-07-01": 1134,
+    "2026-07-02": 593,
+    "2026-07-03": 537,
+    "2026-07-04": 411,
+    "2026-07-05": 485
   }
 }

{openjarvis-1.0.4.dev878 → openjarvis-1.0.4.dev883}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: OpenJarvis
-Version: 1.0.4.dev878
+Version: 1.0.4.dev883
 Summary: OpenJarvis — modular AI assistant backend with composable intelligence primitives
 Project-URL: Homepage, https://github.com/open-jarvis/OpenJarvis
 Project-URL: Documentation, https://open-jarvis.github.io/OpenJarvis/

{openjarvis-1.0.4.dev878 → openjarvis-1.0.4.dev883}/src/openjarvis/agents/errors.py

@@ -2,6 +2,8 @@
 
 from __future__ import annotations
 
+from openjarvis.engine._base import looks_like_context_length_error
+
 
 class AgentTickError(Exception):
     """Base class for agent tick errors."""
@@ -64,6 +66,14 @@ def classify_error(exc: Exception) -> AgentTickError:
 
     msg = str(exc).lower()
 
+    # A context-window overflow is deterministic: retrying the identical
+    # over-length request can never succeed, so fail fast instead of burning
+    # the retry budget on it.
+    if getattr(exc, "is_context_length_error", False) or (
+        looks_like_context_length_error(msg)
+    ):
+        return FatalError(str(exc))
+
     # Check fatal patterns first (more specific)
     if isinstance(exc, PermissionError):
         return FatalError(str(exc))
@@ -90,6 +100,11 @@ def retry_delay(attempt: int) -> int:
 def suggest_action(error: AgentTickError) -> str:
     """Return a human-readable suggested action for the given error."""
     msg = str(error).lower()
+    if looks_like_context_length_error(msg):
+        return (
+            "Conversation too long for the model's context window \u2014 "
+            "start a new chat or shorten the conversation"
+        )
     if any(p in msg for p in ("rate limit", "rate_limit", "429", "too many requests")):
         return "Rate limited \u2014 agent will auto-retry on next tick"
     if any(p in msg for p in ("timeout", "timed out", "connection", "unavailable")):

{openjarvis-1.0.4.dev878 → openjarvis-1.0.4.dev883}/src/openjarvis/agents/hybrid/mini_swe_agent.py

@@ -1403,12 +1403,10 @@ def _loop_local(
             # server walled the call. Compact aggressively (keep_last=1)
             # and retry once. Re-raise on anything else or on a second
             # failure — the runner records the row as errored.
+            from openjarvis.engine._base import looks_like_context_length_error
+
             msg = str(exc)
-            is_ctx = (
-                "maximum context length" in msg
-                or "context length" in msg.lower()
-                and "exceed" in msg.lower()
-            )
+            is_ctx = looks_like_context_length_error(msg)
             if not is_ctx:
                 raise
             _record_event(

{openjarvis-1.0.4.dev878 → openjarvis-1.0.4.dev883}/src/openjarvis/cli/ask.py

@@ -20,6 +20,7 @@ from openjarvis.core.events import EventBus, EventType
 from openjarvis.core.types import Message, Role
 from openjarvis.engine import (
     EngineConnectionError,
+    EngineContextLengthError,
     discover_engines,
     discover_models,
     get_engine,
@@ -881,6 +882,11 @@ def ask(
                 capability_policy=sec.capability_policy,
                 memory_files_config=effective_mf,
             )
+        except EngineContextLengthError as exc:
+            # Not a reachability problem — pointing the user at server/host
+            # config (hint_no_engine) would be misleading here.
+            console.print(f"[red]{exc}[/red]")
+            sys.exit(1)
         except EngineConnectionError as exc:
             console.print(f"[red]Engine error:[/red] {exc}")
             console.print(hint_no_engine())
@@ -990,6 +996,11 @@ def ask(
                 temperature=temperature,
                 max_tokens=max_tokens,
             )
+    except EngineContextLengthError as exc:
+        # Not a reachability problem — pointing the user at server/host
+        # config (hint_no_engine) would be misleading here.
+        console.print(f"[red]{exc}[/red]")
+        sys.exit(1)
     except EngineConnectionError as exc:
         console.print(f"[red]Engine error:[/red] {exc}")
         console.print(hint_no_engine())

{openjarvis-1.0.4.dev878 → openjarvis-1.0.4.dev883}/src/openjarvis/engine/__init__.py

@@ -9,7 +9,9 @@ import openjarvis.engine.ollama  # noqa: F401
 import openjarvis.engine.openai_compat_engines  # noqa: F401
 from openjarvis.engine._base import (
     EngineConnectionError,
+    EngineContextLengthError,
     InferenceEngine,
+    looks_like_context_length_error,
     messages_to_dicts,
 )
 from openjarvis.engine._discovery import discover_engines, discover_models, get_engine
@@ -23,9 +25,11 @@ for _optional in ("cloud", "litellm", "gemma_cpp"):
 
 __all__ = [
     "EngineConnectionError",
+    "EngineContextLengthError",
     "InferenceEngine",
     "discover_engines",
     "discover_models",
     "get_engine",
+    "looks_like_context_length_error",
     "messages_to_dicts",
 ]

{openjarvis-1.0.4.dev878 → openjarvis-1.0.4.dev883}/src/openjarvis/engine/_base.py

@@ -13,6 +13,46 @@ class EngineConnectionError(Exception):
     """Raised when an engine is unreachable."""
 
 
+class EngineContextLengthError(EngineConnectionError):
+    """The prompt exceeds the served model's maximum context window.
+
+    Subclasses ``EngineConnectionError`` so existing ``except
+    EngineConnectionError`` handlers keep catching it, while callers that want a
+    distinct, user-facing "conversation too long" message can branch on this type
+    (or the ``is_context_length_error`` marker) instead of surfacing a generic
+    engine failure.
+    """
+
+    is_context_length_error: bool = True
+
+
+# Substrings that identify an error body as a context-window overflow (vLLM,
+# SGLang, and OpenAI-compatible servers phrase this a few different ways).
+# Every marker is anchored on "context" on purpose: generic phrases like
+# "please reduce" or "too many tokens" also appear in unrelated 400 bodies
+# (max_tokens validation, rate limiting, oversized images) and would
+# misclassify those as "conversation too long".
+CONTEXT_LENGTH_MARKERS = (
+    "context length",
+    "maximum context",
+    "context window",
+    "maximum_context",
+    "context_length_exceeded",
+)
+
+
+def looks_like_context_length_error(text: str) -> bool:
+    """True when *text* reads like a context-window overflow error.
+
+    The single shared heuristic for recognizing vendor context-overflow
+    phrasings — used by the engine layer (typing upstream 400s), agent error
+    classification, and the server stream bridge, so a new vendor phrasing
+    only ever needs to be added here.
+    """
+    low = (text or "").lower()
+    return any(marker in low for marker in CONTEXT_LENGTH_MARKERS)
+
+
 _REASONING_METADATA_KEYS = ("reasoning_content", "thinking")
 
 
@@ -80,8 +120,11 @@ def estimate_prompt_tokens(messages: Sequence[Message]) -> int:
 
 
 __all__ = [
+    "CONTEXT_LENGTH_MARKERS",
     "EngineConnectionError",
+    "EngineContextLengthError",
     "InferenceEngine",
     "estimate_prompt_tokens",
+    "looks_like_context_length_error",
     "messages_to_dicts",
 ]

openjarvis-1.0.4.dev883/src/openjarvis/engine/_http_async.py

@@ -0,0 +1,128 @@
+"""Shared async-HTTP plumbing for engines that stream over httpx.
+
+Home of the pieces the OpenAI-compat and Ollama engines were each hand-rolling:
+the async-client factory (with the configured timeout applied), a cached
+long-lived client so consecutive streams reuse pooled connections instead of
+paying a fresh TCP/TLS handshake per turn, the transport-error set that maps to
+``EngineConnectionError``, and the non-2xx → engine-error translation.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+from typing import NoReturn
+
+import httpx
+
+from openjarvis.engine._base import (
+    EngineConnectionError,
+    EngineContextLengthError,
+    looks_like_context_length_error,
+)
+
+logger = logging.getLogger(__name__)
+
+# Transport failures that map to EngineConnectionError on the streaming paths.
+# ``RemoteProtocolError``/``ReadError`` cover a server dying MID-STREAM (peer
+# closed between tokens); a wedged read trips the configured timeout
+# (TimeoutException). Kept exactly this narrow on purpose:
+# ``asyncio.CancelledError``/``GeneratorExit`` are NOT ``httpx.TransportError``
+# subclasses and must keep propagating for correct cancellation.
+STREAM_TRANSPORT_ERRORS = (
+    httpx.ConnectError,
+    httpx.TimeoutException,
+    httpx.RemoteProtocolError,
+    httpx.ReadError,
+)
+
+_CONTEXT_LENGTH_USER_MESSAGE = (
+    "The conversation is too long for the model's context window. "
+    "Start a new chat or shorten the conversation, then try again."
+)
+
+
+class AsyncHTTPEngineMixin:
+    """Async streaming plumbing shared by httpx-backed engines.
+
+    Expects the engine to provide ``engine_id``, ``_host``, ``_timeout``, an
+    ``_async_transport`` test seam (``httpx.MockTransport`` in tests, ``None``
+    in production), and optionally ``_headers``.
+    """
+
+    engine_id: str
+    _host: str
+    _timeout: float
+    _async_transport: httpx.AsyncBaseTransport | None
+
+    # Set True by engines whose upstream reports context-window overflows in
+    # 400 bodies (OpenAI-compat servers). Ollama has no such signal.
+    _stream_400_signals_context_length: bool = False
+
+    # Lazily-created shared client (and the loop it belongs to). Class-level
+    # ``None`` defaults keep engine ``__init__``s free of mixin bookkeeping.
+    _async_client: httpx.AsyncClient | None = None
+    _async_client_loop: asyncio.AbstractEventLoop | None = None
+
+    def _make_async_client(self) -> httpx.AsyncClient:
+        """Build an async client that honours the configured timeout."""
+        return httpx.AsyncClient(
+            base_url=self._host,
+            timeout=self._timeout,
+            headers=getattr(self, "_headers", None),
+            transport=self._async_transport,
+        )
+
+    def _get_async_client(self) -> httpx.AsyncClient:
+        """Return the shared async client for the running event loop.
+
+        Reusing one client across calls preserves connection pooling — without
+        it every conversation turn pays a fresh TCP (and TLS) handshake. The
+        client is cached per event loop: pooled connections die with their
+        loop, so CLI flows that run ``asyncio.run()`` per turn transparently
+        get a fresh client while a long-lived server loop keeps one pool.
+        """
+        loop = asyncio.get_running_loop()
+        client = self._async_client
+        if client is None or client.is_closed or self._async_client_loop is not loop:
+            # Any previous client belonged to a finished loop; its pooled
+            # connections are already dead, so just drop the reference.
+            client = self._make_async_client()
+            self._async_client = client
+            self._async_client_loop = loop
+        return client
+
+    def _close_async_client(self) -> None:
+        """Best-effort close of the shared async client (for ``close()``)."""
+        client = self._async_client
+        loop = self._async_client_loop
+        self._async_client = None
+        self._async_client_loop = None
+        if client is None or client.is_closed:
+            return
+        try:
+            if loop is not None and not loop.is_closed():
+                if loop.is_running():
+                    loop.create_task(client.aclose())
+                else:
+                    loop.run_until_complete(client.aclose())
+        except Exception:  # noqa: BLE001 — cleanup must never mask the close
+            logger.debug("Async client did not close cleanly", exc_info=True)
+
+    def _raise_stream_http_error(self, status: int, detail: str) -> NoReturn:
+        """Map a non-success streaming HTTP response to a clean engine error."""
+        detail = (detail or "").strip()
+        if (
+            status == 400
+            and self._stream_400_signals_context_length
+            and looks_like_context_length_error(detail)
+        ):
+            raise EngineContextLengthError(_CONTEXT_LENGTH_USER_MESSAGE)
+        detail_suffix = f": {detail}" if detail else ""
+        raise EngineConnectionError(
+            f"{self.engine_id} engine at {self._host} returned HTTP "
+            f"{status}{detail_suffix}"
+        )
+
+
+__all__ = ["AsyncHTTPEngineMixin", "STREAM_TRANSPORT_ERRORS"]

{openjarvis-1.0.4.dev878 → openjarvis-1.0.4.dev883}/src/openjarvis/engine/_openai_compat.py

@@ -12,18 +12,27 @@ import httpx
 from openjarvis.core.types import Message
 from openjarvis.engine._base import (
     EngineConnectionError,
+    EngineContextLengthError,
     InferenceEngine,
     estimate_prompt_tokens,
     messages_to_dicts,
 )
+from openjarvis.engine._http_async import (
+    STREAM_TRANSPORT_ERRORS,
+    AsyncHTTPEngineMixin,
+)
 from openjarvis.engine._stubs import StreamChunk
 
 logger = logging.getLogger(__name__)
 
 
-class _OpenAICompatibleEngine(InferenceEngine):
+class _OpenAICompatibleEngine(AsyncHTTPEngineMixin, InferenceEngine):
     """Base for engines that serve the OpenAI ``/v1/chat/completions`` API."""
 
+    # vLLM/SGLang report context-window overflows in 400 bodies; the shared
+    # ``_raise_stream_http_error`` types those as ``EngineContextLengthError``.
+    _stream_400_signals_context_length = True
+
     engine_id: str = ""
     _default_host: str = "http://localhost:8000"
     _api_prefix: str = "/v1"
@@ -50,6 +59,16 @@ class _OpenAICompatibleEngine(InferenceEngine):
         headers = (
             {"Authorization": f"Bearer {self._api_key}"} if self._api_key else None
         )
+        # Used by the shared async streaming plumbing (AsyncHTTPEngineMixin) so
+        # the bounded request timeout is applied to streaming reads, not just
+        # the synchronous methods (a wedged token read fails at ``timeout``
+        # rather than hanging the caller for the httpx default).
+        self._timeout = timeout
+        self._headers = headers
+        # Injection seam for tests: an ``httpx.MockTransport`` swapped in here lets
+        # the async stream path be exercised with a mocked transport and no real
+        # server. ``None`` in production so httpx uses its default networking.
+        self._async_transport: httpx.AsyncBaseTransport | None = None
         self._client = httpx.Client(
             base_url=self._host, timeout=timeout, headers=headers
         )
@@ -168,11 +187,26 @@ class _OpenAICompatibleEngine(InferenceEngine):
         # Default to tool_choice=auto when tools are provided
         if "tools" in payload and "tool_choice" not in payload:
             payload["tool_choice"] = "auto"
+        url = f"{self._api_prefix}/chat/completions"
         try:
-            url = f"{self._api_prefix}/chat/completions"
-            with self._client.stream("POST", url, json=payload) as resp:
-                resp.raise_for_status()
-                for line in resp.iter_lines():
+            # ASYNC streaming: ``httpx.AsyncClient`` + ``aiter_lines`` never
+            # blocks the event loop between tokens (the previous SYNC
+            # ``httpx.Client`` + ``iter_lines`` inside this ``async def`` blocked
+            # the single uvicorn worker on every inter-token wait, serializing all
+            # concurrent chats and letting one wedged read freeze the whole API).
+            # The shared client keeps pooled connections across turns.
+            client = self._get_async_client()
+            async with client.stream("POST", url, json=payload) as resp:
+                # ``not is_success`` covers 3xx as well as 4xx/5xx. With
+                # ``follow_redirects`` off (the default) an unexpected redirect
+                # would otherwise fall through to ``aiter_lines`` and surface as
+                # a silent EMPTY stream instead of a clean engine error.
+                if not resp.is_success:
+                    # Load the (short) error body before touching ``.text``:
+                    # a streaming response is otherwise unread.
+                    await resp.aread()
+                    self._raise_stream_http_error(resp.status_code, resp.text)
+                async for line in resp.aiter_lines():
                     if not line.startswith("data:"):
                         continue
                     data_str = line[len("data:") :].strip()
@@ -186,7 +220,11 @@ class _OpenAICompatibleEngine(InferenceEngine):
                     content = delta.get("content")
                     if content:
                         yield content
-        except (httpx.ConnectError, httpx.TimeoutException) as exc:
+        except STREAM_TRANSPORT_ERRORS as exc:
+            # A wedged upstream read trips ``timeout`` (ReadTimeout) and is mapped
+            # here, so the request fails cleanly at the configured bound instead
+            # of hanging indefinitely (see STREAM_TRANSPORT_ERRORS for why the
+            # set is exactly this narrow).
             raise EngineConnectionError(
                 f"{self.engine_id} engine not reachable at {self._host}"
             ) from exc
@@ -212,11 +250,20 @@ class _OpenAICompatibleEngine(InferenceEngine):
         }
         if "tools" in payload and "tool_choice" not in payload:
             payload["tool_choice"] = "auto"
+        url = f"{self._api_prefix}/chat/completions"
         try:
-            url = f"{self._api_prefix}/chat/completions"
-            with self._client.stream("POST", url, json=payload) as resp:
-                resp.raise_for_status()
-                for line in resp.iter_lines():
+            # ASYNC streaming (see ``stream``): non-blocking shared client so
+            # rich streaming never stalls the event loop and honours ``timeout``.
+            client = self._get_async_client()
+            async with client.stream("POST", url, json=payload) as resp:
+                # ``not is_success`` covers 3xx as well as 4xx/5xx. With
+                # ``follow_redirects`` off (the default) an unexpected redirect
+                # would otherwise fall through to ``aiter_lines`` and surface as
+                # a silent EMPTY stream instead of a clean engine error.
+                if not resp.is_success:
+                    await resp.aread()
+                    self._raise_stream_http_error(resp.status_code, resp.text)
+                async for line in resp.aiter_lines():
                     if not line.startswith("data:"):
                         continue
                     data_str = line[len("data:") :].strip()
@@ -240,7 +287,10 @@ class _OpenAICompatibleEngine(InferenceEngine):
                             finish_reason=finish,
                             usage=usage,
                         )
-        except (httpx.ConnectError, httpx.TimeoutException) as exc:
+        except STREAM_TRANSPORT_ERRORS as exc:
+            # See ``stream``: transport failures (incl. a mid-stream server
+            # disconnect) map to a clean error; the set is kept narrow so
+            # cancellation still propagates.
             raise EngineConnectionError(
                 f"{self.engine_id} engine not reachable at {self._host}"
             ) from exc
@@ -279,6 +329,9 @@ class _OpenAICompatibleEngine(InferenceEngine):
 
     def close(self) -> None:
         self._client.close()
+        self._close_async_client()
 
 
-__all__ = ["_OpenAICompatibleEngine"]
+# ``EngineContextLengthError`` moved to ``openjarvis.engine._base``; re-exported
+# here for callers/tests that import it from this module.
+__all__ = ["_OpenAICompatibleEngine", "EngineContextLengthError"]

{openjarvis-1.0.4.dev878 → openjarvis-1.0.4.dev883}/src/openjarvis/engine/litellm.py

@@ -123,8 +123,12 @@ class LiteLLMEngine(InferenceEngine):
             call_kwargs["api_base"] = self._api_base
         call_kwargs.update(kwargs)
 
-        resp = litellm.completion(**call_kwargs)
-        for chunk in resp:
+        # ``acompletion`` + ``async for``: the sync ``litellm.completion`` used
+        # before made a blocking network call (and blocking per-chunk reads)
+        # inside this ``async def``, stalling the whole event loop between
+        # tokens — the same bug the httpx engines' streaming paths had.
+        resp = await litellm.acompletion(**call_kwargs)
+        async for chunk in resp:
             delta = chunk.choices[0].delta if chunk.choices else None
             if delta and delta.content:
                 yield delta.content

{openjarvis-1.0.4.dev878 → openjarvis-1.0.4.dev883}/src/openjarvis/engine/ollama.py

@@ -18,6 +18,10 @@ from openjarvis.engine._base import (
     estimate_prompt_tokens,
     messages_to_dicts,
 )
+from openjarvis.engine._http_async import (
+    STREAM_TRANSPORT_ERRORS,
+    AsyncHTTPEngineMixin,
+)
 from openjarvis.engine._stubs import StreamChunk
 
 logger = logging.getLogger(__name__)
@@ -80,11 +84,15 @@ def _default_num_ctx() -> int:
 
 
 @EngineRegistry.register("ollama")
-class OllamaEngine(InferenceEngine):
+class OllamaEngine(AsyncHTTPEngineMixin, InferenceEngine):
     """Ollama backend via its native HTTP API."""
 
     engine_id = "ollama"
 
+    # Ollama has no context-length overflow signal in its 400 bodies, so the
+    # shared ``_raise_stream_http_error`` keeps its default (no
+    # ``EngineContextLengthError`` branch, unlike the OpenAI-compat engines).
+
     _DEFAULT_HOST = "http://localhost:11434"
 
     def __init__(
@@ -98,6 +106,14 @@ class OllamaEngine(InferenceEngine):
             env_host = os.environ.get("OLLAMA_HOST")
             host = env_host or self._DEFAULT_HOST
         self._host = host.rstrip("/")
+        # Used by the shared async streaming plumbing (AsyncHTTPEngineMixin) so a
+        # wedged token read is bounded by ``timeout`` instead of hanging the
+        # single event loop for the httpx default.
+        self._timeout = timeout
+        # Injection seam for tests: an ``httpx.MockTransport`` swapped in here drives
+        # the async stream path with no real Ollama server. ``None`` in production so
+        # httpx uses its default networking.
+        self._async_transport: httpx.AsyncBaseTransport | None = None
         self._client = httpx.Client(base_url=self._host, timeout=timeout)
         # Last stream usage — captured from Ollama's final chunk
         self._last_stream_usage: Dict[str, int] = {}
@@ -263,9 +279,26 @@ class OllamaEngine(InferenceEngine):
         elif kwargs["think"] is not None:
             payload["think"] = kwargs["think"]
         try:
-            with self._client.stream("POST", "/api/chat", json=payload) as resp:
-                resp.raise_for_status()
-                for line in resp.iter_lines():
+            # ASYNC streaming: ``httpx.AsyncClient`` + ``aiter_lines`` never
+            # blocks the event loop between tokens (the previous SYNC
+            # ``self._client`` + ``iter_lines`` inside this ``async def`` blocked
+            # the single uvicorn worker on every inter-token wait, serializing all
+            # concurrent chats and letting one wedged read freeze the whole API).
+            # The shared client keeps pooled connections across turns.
+            client = self._get_async_client()
+            async with client.stream("POST", "/api/chat", json=payload) as resp:
+                # ``not is_success`` covers 3xx as well as 4xx/5xx and maps
+                # to ``EngineConnectionError`` (matching the OpenAI-compat
+                # path) instead of leaking a raw ``httpx.HTTPStatusError``.
+                # With redirects off (the default) an unexpected 3xx would
+                # otherwise fall through to ``aiter_lines`` and surface as a
+                # silent EMPTY stream rather than a clean engine error.
+                if not resp.is_success:
+                    # Read the (short) error body before touching ``.text``:
+                    # a streaming response is otherwise unread.
+                    await resp.aread()
+                    self._raise_stream_http_error(resp.status_code, resp.text)
+                async for line in resp.aiter_lines():
                     if not line.strip():
                         continue
                     try:
@@ -290,7 +323,10 @@ class OllamaEngine(InferenceEngine):
                             "total_tokens": full_prompt + comp,
                         }
                         break
-        except (httpx.ConnectError, httpx.TimeoutException) as exc:
+        except STREAM_TRANSPORT_ERRORS as exc:
+            # Transport failures (incl. a mid-stream server disconnect) map to a
+            # clean error; the set is kept narrow (see STREAM_TRANSPORT_ERRORS)
+            # so cancellation still propagates.
             raise EngineConnectionError(
                 f"Ollama not reachable at {self._host}"
             ) from exc
@@ -356,19 +392,34 @@ class OllamaEngine(InferenceEngine):
     ) -> AsyncIterator[StreamChunk]:
         """Execute the streaming request and yield parsed StreamChunks."""
         try:
-            with self._client.stream("POST", "/api/chat", json=payload) as resp:
+            # ASYNC streaming (see ``stream``): shared ``AsyncClient`` +
+            # ``aiter_lines`` so rich streaming never stalls the event loop and
+            # honours ``timeout``.
+            client = self._get_async_client()
+            async with client.stream("POST", "/api/chat", json=payload) as resp:
                 if resp.status_code == 400 and retry_without_tools:
                     # Model doesn't support tools — retry without them.
+                    # PRESERVED: this specific 400 path must still trigger the
+                    # tools-less retry; only OTHER non-2xx responses map to
+                    # EngineConnectionError below.
                     payload.pop("tools", None)
                     async for c in self._run_stream(
                         payload, messages, retry_without_tools=False
                     ):
                         yield c
                     return
-                resp.raise_for_status()
+                # ``not is_success`` covers 3xx as well as 4xx/5xx and maps
+                # to ``EngineConnectionError`` (matching the OpenAI-compat
+                # path) instead of leaking a raw ``httpx.HTTPStatusError``.
+                # With redirects off (the default) an unexpected 3xx would
+                # otherwise fall through to ``aiter_lines`` and surface as a
+                # silent EMPTY stream rather than a clean engine error.
+                if not resp.is_success:
+                    await resp.aread()
+                    self._raise_stream_http_error(resp.status_code, resp.text)
 
                 finish_reason: str | None = None
-                for line in resp.iter_lines():
+                async for line in resp.aiter_lines():
                     if not line.strip():
                         continue
                     try:
@@ -441,7 +492,10 @@ class OllamaEngine(InferenceEngine):
                             usage=dict(self._last_stream_usage),
                         )
                         break
-        except (httpx.ConnectError, httpx.TimeoutException) as exc:
+        except STREAM_TRANSPORT_ERRORS as exc:
+            # See ``stream``: transport failures (incl. a mid-stream server
+            # disconnect) map to a clean error; the set is kept narrow so
+            # cancellation still propagates.
             raise EngineConnectionError(
                 f"Ollama not reachable at {self._host}"
             ) from exc
@@ -474,6 +528,7 @@ class OllamaEngine(InferenceEngine):
 
     def close(self) -> None:
         self._client.close()
+        self._close_async_client()
 
 
 __all__ = ["OllamaEngine"]

{openjarvis-1.0.4.dev878 → openjarvis-1.0.4.dev883}/src/openjarvis/server/api_routes.py

@@ -741,8 +741,11 @@ async def websocket_chat_stream(websocket: WebSocket):
                                 )
                     except TypeError:
                         # stream() didn't return an iterable; fall back to
-                        # generate()
-                        result = engine.generate(messages, model=model)
+                        # generate(). It makes a blocking upstream call, so run
+                        # it in a worker thread to keep the event loop free.
+                        result = await asyncio.to_thread(
+                            engine.generate, messages, model=model
+                        )
                         content = (
                             result.get("content", "")
                             if isinstance(
@@ -767,8 +770,11 @@ async def websocket_chat_stream(websocket: WebSocket):
                         ended_at=_time.time(),
                     )
                 else:
-                    # No stream method — single-shot generate
-                    result = engine.generate(messages, model=model)
+                    # No stream method — single-shot generate. Blocking upstream
+                    # call, so run in a worker thread to keep the event loop free.
+                    result = await asyncio.to_thread(
+                        engine.generate, messages, model=model
+                    )
                     content = (
                         result.get("content", "")
                         if isinstance(