AbstractRuntime 0.10.2__tar.gz → 0.10.3__tar.gz

{abstractruntime-0.10.2 → abstractruntime-0.10.3}/CHANGELOG.md

@@ -5,6 +5,17 @@ All notable changes to AbstractRuntime will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.10.3] - 2026-10-18
+
+### Fixed
+
+- Image and video specs whose provider is an endpoint profile (`endpoint:<id>`) run in process instead of in
+  the one-shot local media subprocess. The subprocess isolates local pipelines and has no host
+  endpoint-profile resolver, so a gateway profile (for example an OpenAI-compatible relay serving
+  `gpt-image-2` selected on the gateway's image route) could not be resolved there; in process,
+  AbstractVision resolves it through the resolver the host attached to the client. Tests:
+  `tests/test_endpoint_profile_media_in_process.py`.
+
 ## [0.10.2] - 2026-10-16
 
 ### Changed

{abstractruntime-0.10.2 → abstractruntime-0.10.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: AbstractRuntime
-Version: 0.10.2
+Version: 0.10.3
 Summary: AbstractRuntime: a durable graph runner designed to pair with AbstractCore.
 Project-URL: AbstractCore (website), https://www.abstractcore.ai/
 Project-URL: AbstractRuntime (GitHub), https://github.com/lpalbou/abstractruntime

{abstractruntime-0.10.2 → abstractruntime-0.10.3}/docs/api.md

@@ -355,6 +355,7 @@ Multimodal support:
 - with the pooled local client (`MultiLocalAbstractCoreLLMClient`), a request whose outputs are all non-text (image, video, voice, music, transcription) and that carries no tools never builds the text client: it runs on AbstractCore's capability host (`CapabilityHostProvider`, no text model), and so does streaming TTS. A text default that cannot be built (an engine not installed) therefore never blocks speech or image generation
 - media-only normalized results expose `runtime_provider` / `runtime_model` separately from `media_provider` / `media_model`
 - local image/video generation runs in a one-shot subprocess (`execution_mode="local_one_shot_subprocess"`) unless every requested media model is resident (explicitly loaded with `load_model_residency`), in which case it runs in-process on the loaded pipeline (`execution_mode="resident_in_process"`, `resident_load_ids`)
+- a media spec whose `provider` is an endpoint profile (`endpoint:<id>`) never uses the one-shot subprocess: it is an HTTP call to that endpoint, and the host's endpoint-profile resolver (`set_provider_endpoint_profile_resolver`) exists only in this process
 - optional local media residency failures complete with `status_hint="warning"` and `degraded=true`; local media warmup that no capability plugin can serve (`image_generation`, `image_upscale`, `video_generation`, `text_to_video`, `image_to_video`, `tts`, `stt`, `music_generation`) reports `requires_long_lived_server=true`, and image/video tasks also report `execution_mode="local_one_shot_subprocess"`
 - Gateway/hosts remain responsible for explicit Core server URLs, Core server auth headers, provider/model defaults, selected local-inference profiles, and translation of Gateway-owned env/config into explicit Runtime inputs; Runtime persists only JSON-safe routing metadata and artifact refs
 

{abstractruntime-0.10.2 → abstractruntime-0.10.3}/llms-full.txt

@@ -1225,6 +1225,7 @@ Multimodal support:
 - with the pooled local client (`MultiLocalAbstractCoreLLMClient`), a request whose outputs are all non-text (image, video, voice, music, transcription) and that carries no tools never builds the text client: it runs on AbstractCore's capability host (`CapabilityHostProvider`, no text model), and so does streaming TTS. A text default that cannot be built (an engine not installed) therefore never blocks speech or image generation
 - media-only normalized results expose `runtime_provider` / `runtime_model` separately from `media_provider` / `media_model`
 - local image/video generation runs in a one-shot subprocess (`execution_mode="local_one_shot_subprocess"`) unless every requested media model is resident (explicitly loaded with `load_model_residency`), in which case it runs in-process on the loaded pipeline (`execution_mode="resident_in_process"`, `resident_load_ids`)
+- a media spec whose `provider` is an endpoint profile (`endpoint:<id>`) never uses the one-shot subprocess: it is an HTTP call to that endpoint, and the host's endpoint-profile resolver (`set_provider_endpoint_profile_resolver`) exists only in this process
 - optional local media residency failures complete with `status_hint="warning"` and `degraded=true`; local media warmup that no capability plugin can serve (`image_generation`, `image_upscale`, `video_generation`, `text_to_video`, `image_to_video`, `tts`, `stt`, `music_generation`) reports `requires_long_lived_server=true`, and image/video tasks also report `execution_mode="local_one_shot_subprocess"`
 - Gateway/hosts remain responsible for explicit Core server URLs, Core server auth headers, provider/model defaults, selected local-inference profiles, and translation of Gateway-owned env/config into explicit Runtime inputs; Runtime persists only JSON-safe routing metadata and artifact refs
 

{abstractruntime-0.10.2 → abstractruntime-0.10.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "AbstractRuntime"
-version = "0.10.2"
+version = "0.10.3"
 description = "AbstractRuntime: a durable graph runner designed to pair with AbstractCore."
 readme = "README.md"
 requires-python = ">=3.10"

abstractruntime-0.10.3/release-notes.md

@@ -0,0 +1,8 @@
+### Fixed
+
+- Image and video specs whose provider is an endpoint profile (`endpoint:<id>`) run in process instead of in
+  the one-shot local media subprocess. The subprocess isolates local pipelines and has no host
+  endpoint-profile resolver, so a gateway profile (for example an OpenAI-compatible relay serving
+  `gpt-image-2` selected on the gateway's image route) could not be resolved there; in process,
+  AbstractVision resolves it through the resolver the host attached to the client. Tests:
+  `tests/test_endpoint_profile_media_in_process.py`.

{abstractruntime-0.10.2 → abstractruntime-0.10.3}/src/abstractruntime/integrations/abstractcore/llm_client.py

@@ -2141,6 +2141,20 @@ def _decode_subprocess_media_payload(payload: Dict[str, Any]) -> Dict[str, Any]:
     return out
 
 
+def _spec_routes_to_endpoint_profile(spec: Any) -> bool:
+    """True when a media spec names an `endpoint:<id>` provider.
+
+    Such a spec is an HTTP call to a named endpoint connection, not a local
+    pipeline: the one-shot media subprocess has nothing to isolate for it, and
+    it cannot resolve the profile -- the host's endpoint-profile resolver
+    (`set_provider_endpoint_profile_resolver`) lives on THIS process's
+    clients, and a per-user gateway profile has no AbstractCore config row
+    the subprocess could read. So it runs in process, where AbstractVision
+    resolves the profile through the owning provider's resolver.
+    """
+    return isinstance(spec, dict) and str(spec.get("provider") or "").strip().lower().startswith("endpoint:")
+
+
 def _is_subprocess_safe_image_specs(specs: List[Dict[str, Any]], media: Optional[List[Any]]) -> bool:
     if not _env_flag_enabled("ABSTRACTRUNTIME_LOCAL_IMAGE_SUBPROCESS", default=True):
         return False
@@ -2148,7 +2162,7 @@ def _is_subprocess_safe_image_specs(specs: List[Dict[str, Any]], media: Optional
         return False
     normalized_media = list(media or [])
     for spec in specs:
-        if not isinstance(spec, dict):
+        if not isinstance(spec, dict) or _spec_routes_to_endpoint_profile(spec):
             return False
         modality = str(spec.get("modality") or "").strip().lower()
         task = str(spec.get("task") or "").strip().lower()
@@ -2202,6 +2216,8 @@ def _is_subprocess_safe_video_specs(specs: List[Dict[str, Any]], media: Optional
     if len(specs) != 1 or not isinstance(specs[0], dict):
         return False
     spec = specs[0]
+    if _spec_routes_to_endpoint_profile(spec):
+        return False
     modality = str(spec.get("modality") or "").strip().lower()
     task = str(spec.get("task") or "").strip().lower()
     if modality != "video":

abstractruntime-0.10.3/tests/test_endpoint_profile_media_in_process.py

@@ -0,0 +1,45 @@
+"""Media specs routed to an `endpoint:<id>` profile run in process.
+
+The one-shot media subprocess isolates LOCAL pipelines; it has no host
+endpoint-profile resolver, so a gateway profile (e.g. an OpenAI-compatible
+relay serving `gpt-image-2`) could not be resolved there. Regression for the
+operator report of 2026-10-18.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from abstractruntime.integrations.abstractcore.llm_client import (
+    _is_subprocess_safe_image_specs,
+    _is_subprocess_safe_video_specs,
+    _spec_routes_to_endpoint_profile,
+)
+
+
+def test_endpoint_spec_detection() -> None:
+    assert _spec_routes_to_endpoint_profile({"provider": "endpoint:relay"}) is True
+    assert _spec_routes_to_endpoint_profile({"provider": "Endpoint:Relay"}) is True
+    assert _spec_routes_to_endpoint_profile({"provider": "huggingface"}) is False
+    assert _spec_routes_to_endpoint_profile({}) is False
+    assert _spec_routes_to_endpoint_profile(None) is False
+
+
+def test_endpoint_image_specs_skip_the_subprocess(tmp_path: Path) -> None:
+    source = tmp_path / "source.png"
+    source.write_bytes(b"png")
+    media = [{"file_path": str(source), "type": "image", "role": "source"}]
+
+    assert _is_subprocess_safe_image_specs([{"modality": "image", "task": "text_to_image", "provider": "huggingface"}], None) is True
+    assert _is_subprocess_safe_image_specs(
+        [{"modality": "image", "task": "text_to_image", "provider": "endpoint:relay", "model": "gpt-image-2"}], None
+    ) is False
+    assert _is_subprocess_safe_image_specs([{"modality": "image", "task": "image_edit", "provider": "huggingface"}], media) is True
+    assert _is_subprocess_safe_image_specs(
+        [{"modality": "image", "task": "image_edit", "provider": "endpoint:relay", "model": "gpt-image-2"}], media
+    ) is False
+
+
+def test_endpoint_video_spec_skips_the_subprocess() -> None:
+    assert _is_subprocess_safe_video_specs([{"modality": "video", "task": "text_to_video"}], None) is True
+    assert _is_subprocess_safe_video_specs([{"modality": "video", "task": "text_to_video", "provider": "endpoint:relay"}], None) is False

abstractruntime-0.10.2/release-notes.md

@@ -1,19 +0,0 @@
-### Changed
-
-- **Code nodes can import Python's standard library.** In the `sandbox` permission mode, `import json`, `re`,
-  `datetime`, `math`, `collections`, `itertools`, `statistics`, `textwrap`, `hashlib`, `urllib.parse`… now
-  work, at the top of the code and inside functions, in both the RestrictedPython lane and the fallback lane, and
-  inside the function bodies of flow function libraries. The sandbox still refuses third-party packages, private
-  (`_`-prefixed) modules, and the standard-library modules that reach the machine or go around the sandbox's own
-  checks: processes and the operating system (`os`, `subprocess`, `threading`, `signal`, `ctypes`, `platform`…), files (`io`,
-  `pathlib`, `shutil`, `tempfile`, `glob`, `sqlite3`, `logging`, and the parsers that read a file by path: `xml`,
-  `configparser`, `argparse`…), the network (`socket`, `http`, `urllib.request`, `smtplib`…), code loading or
-  interpreter internals (`sys`, `importlib`, `pickle`, `types`, `inspect`…) and attribute access by name,
-  evaluated annotations or frames (`operator`, `typing`, `functools`, `traceback`). A refusal names the module
-  and the reason; `from x import *`, relative imports and the `__import__` call stay refused.
-- **A Code node's return value must be JSON data.** `transform` returning a module, a function, a set, bytes or
-  any other value JSON cannot carry now fails with a sandbox error naming the value and where it sits
-  (`transform returned a value that cannot be sent as JSON: module 'time' at output[2]`), in every permission
-  mode, instead of a serializer crash later in the gateway's simulate reply or the run's persisted state. `full_access` (host policy) imports anything, as before.
-  The rule is applied twice: at compile time by `validate_code`, and at run time by the `__import__` bound into
-  the sandbox builtins. `get_code_execution_policy()` now lists `imports.denied_modules` for thin clients.