agenthub-python 0.4.9__tar.gz → 0.4.11__tar.gz

{agenthub_python-0.4.9 → agenthub_python-0.4.11}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agenthub-python
-Version: 0.4.9
+Version: 0.4.11
 Summary: AgentHub is the LLM API Hub for the Agent era, built for high-precision autonomous agents.
 Keywords: agent,llm,gemini,claude,gpt
 Author: PrismShadow

{agenthub_python-0.4.9 → agenthub_python-0.4.11}/agenthub/ant_messages/client.py

@@ -128,8 +128,15 @@ class AntMessagesClient(LLMClient):
 
         if config.get("thinking_level") is not None:
             ant_config.update(self._convert_thinking_level_to_thinking_config(config["thinking_level"]))
-            if config.get("thinking_summary") and ant_config.get("thinking", {}).get("type") == "adaptive":
-                ant_config["thinking"]["display"] = "summarized"
+
+        if config.get("thinking_summary") is not None:
+            # display lives on the thinking block, so a summary asked for on its own selects
+            # adaptive thinking. A disabled block is the one place it cannot ride along --
+            # "thinking.disabled.display: Extra inputs are not permitted" (400, verified live
+            # 2026-09-03) -- and thinking_level NONE disables thinking, leaving nothing to show.
+            thinking = ant_config.setdefault("thinking", {"type": "adaptive"})
+            if thinking["type"] != "disabled":
+                thinking["display"] = "summarized" if config["thinking_summary"] else "omitted"
 
         # Convert tools to the Messages API tool schema
         if config.get("tools") is not None:

{agenthub_python-0.4.9 → agenthub_python-0.4.11}/agenthub/auto_client.py

@@ -21,7 +21,13 @@ from .types import UniConfig, UniEvent, UniMessage
 
 
 # The generic protocol clients are named explicitly rather than deduced from a model id.
-_PROTOCOL_CLIENT_TYPES = ("openai-chat", "openai-responses", "ant-messages", "openai-embedding")
+_PROTOCOL_CLIENT_TYPES = (
+    "openai-chat",
+    "openai-chat-vllm-adapter",
+    "openai-responses",
+    "ant-messages",
+    "openai-embedding",
+)
 
 
 class AutoLLMClient(LLMClient):
@@ -65,23 +71,25 @@ class AutoLLMClient(LLMClient):
             type[LLMClient] | None: The client class, or None when no client claims the type.
         """
         # every Gemini generation shares the unified client ("gemini-3" also matches the
-        # gemini-3.7/gemini-3.6/gemini-3.5-flash-lite client types)
+        # gemini-3.8/gemini-3.7/gemini-3.6/gemini-3.5-flash-lite client types)
         if any(
             prefix in client_type for prefix in ("gemini-3", "gemini-embedding")
-        ):  # e.g., gemini-3.7-flash, gemini-3-flash-preview, gemini-embedding-2
-            from .gemini3_7 import Gemini3_7Client
+        ):  # e.g., gemini-3.8-flash, gemini-3-flash-preview, gemini-embedding-2
+            from .gemini3_8 import Gemini3_8Client
 
-            return Gemini3_7Client
+            return Gemini3_8Client
         elif "claude" in client_type and (
             "4-6" in client_type or "4-7" in client_type or "4-8" in client_type or "-5" in client_type
         ):  # the whole Claude 4.6+ series shares the unified client, e.g., claude-sonnet-4-6
             from .claude5 import Claude5Client
 
             return Claude5Client
-        elif "gpt-5.4" in client_type or "gpt-5.5" in client_type or "gpt-5.6" in client_type:  # e.g., gpt-5.6
-            from .gpt5_6 import GPT5_6Client
+        elif (
+            "gpt-5.4" in client_type or "gpt-5.5" in client_type or "gpt-5.6" in client_type or "gpt-6" in client_type
+        ):  # e.g., gpt-6-astra
+            from .gpt6 import GPT6Client
 
-            return GPT5_6Client
+            return GPT6Client
         elif "glm-5" in client_type:  # the whole GLM series shares the unified client
             from .glm5_3 import GLM5_3Client
 
@@ -99,6 +107,12 @@ class AutoLLMClient(LLMClient):
             from .deepseek_v4 import DeepSeekV4Client
 
             return DeepSeekV4Client
+        elif client_type == "openai-chat-vllm-adapter":
+            # exact match: "openai-chat-vllm-adapter" contains "openai", so the substring
+            # branches below would otherwise claim it
+            from .openai_chat_vllm_adapter import OpenaiChatVllmAdapterClient
+
+            return OpenaiChatVllmAdapterClient
         elif "ant-messages" in client_type:
             from .ant_messages import AntMessagesClient
 
@@ -131,10 +145,10 @@ class AutoLLMClient(LLMClient):
         if client_class is None:
             raise ValueError(
                 f"{client_type} is not supported. "
-                "Supported client types: minimax-m3, gemini-3.7, gemini-3.6, gemini-3, "
-                "claude-5, claude-4-8, claude-4-7, claude-4-6, gpt-5.6, gpt-5.5, gpt-5.4, "
+                "Supported client types: minimax-m3, gemini-3.8, gemini-3.7, gemini-3.6, gemini-3, "
+                "claude-5, claude-4-8, claude-4-7, claude-4-6, gpt-6, gpt-5.6, gpt-5.5, gpt-5.4, "
                 "glm-5.3, glm-5.2, glm-5.1, kimi-k3, kimi-k2.6, kimi-k2.5, deepseek-v4, "
-                "openai-embedding, ant-messages, openai-responses, openai-chat."
+                "openai-chat-vllm-adapter, openai-embedding, ant-messages, openai-responses, openai-chat."
             )
 
         return client_class(model=model, api_key=api_key, base_url=base_url, default_headers=default_headers)

{agenthub_python-0.4.9 → agenthub_python-0.4.11}/agenthub/claude5/client.py

@@ -176,8 +176,14 @@ class Claude5Client(LLMClient):
 
         if config.get("thinking_level") is not None:
             claude_config.update(self._convert_thinking_level_to_thinking_config(config["thinking_level"]))
-            if config.get("thinking_summary") and claude_config.get("thinking") is not None:
-                claude_config["thinking"]["display"] = "summarized"
+
+        if config.get("thinking_summary") is not None:
+            # display lives on the thinking block, so a summary asked for on its own selects
+            # adaptive thinking, which is what this family runs by default anyway; a block
+            # carrying display but no output_config is accepted on 4.6 through 5 (verified
+            # live 2026-09-03). NONE omits the block, so the request lands on that default.
+            thinking = claude_config.setdefault("thinking", {"type": "adaptive"})
+            thinking["display"] = "summarized" if config["thinking_summary"] else "omitted"
 
         # Convert tools to Claude's tool schema
         if config.get("tools") is not None:

{agenthub_python-0.4.9 → agenthub_python-0.4.11}/agenthub/deepseek_v4/client.py

@@ -14,6 +14,7 @@
 
 import json
 import os
+import re
 from typing import Any, AsyncIterator
 
 from openai import AsyncOpenAI
@@ -36,6 +37,13 @@ from ..types import (
 from ..utils import is_debug_enabled
 
 
+# The DeepSeek ids that read no image: the current V4 Flash and V4 Pro, bare or with a dated
+# snapshot suffix (deepseek-v4-flash-0731). Every other id forwards its images. Matched against
+# the bare id — the part after the last "/", lowercased — so a gateway prefix (deepseek/,
+# deepseek-ai/) and the spelling a platform uses do not change the verdict.
+_TEXT_ONLY_MODELS = re.compile(r"deepseek-v4-(flash|pro)(-\d{4})?")
+
+
 class DeepSeekV4Client(LLMClient):
     """DeepSeek V4-specific LLM client implementation using the OpenAI-compatible Responses API."""
 
@@ -103,10 +111,17 @@ class DeepSeekV4Client(LLMClient):
                 self.__class__.__name__, "temperature", "DeepSeek V4 does not support setting temperature."
             )
 
-        # a thinking summary is accepted but never generated, so the parameter is left out
         if config.get("thinking_level") is not None:
             deepseek_config["reasoning"] = {"effort": self._convert_thinking_level_to_effort(config["thinking_level"])}
 
+        if config.get("thinking_summary"):
+            # DeepSeek takes reasoning.summary with or without an effort and returns an empty
+            # summary list for now (verified live 2026-09-03 on api.deepseek.com and
+            # OpenRouter), so the request carries the preference instead of dropping it and
+            # picks up summaries as soon as the vendor generates them. False needs no key:
+            # the Responses API returns no summary unless one is asked for.
+            deepseek_config.setdefault("reasoning", {})["summary"] = "concise"
+
         if config.get("tools") is not None:
             deepseek_config["tools"] = [{"type": "function", **tool} for tool in config["tools"]]
 
@@ -135,10 +150,10 @@ class DeepSeekV4Client(LLMClient):
         Returns:
             List of input items for the Responses API
         """
-        # only a vision model reads image parts; every other DeepSeek model answers from a
-        # placeholder instead of failing (llmsdk_docs/deepseek_v4/docs/responses-api.md), so an
-        # image is refused here rather than silently dropped
-        supports_image = "vision" in self._model.lower()
+        # a text-only model answers from a placeholder instead of failing
+        # (llmsdk_docs/deepseek_v4/docs/responses-api.md), so an image is refused here rather
+        # than silently dropped
+        supports_image = not _TEXT_ONLY_MODELS.fullmatch(self._model.lower().rsplit("/", 1)[-1])
         input_list: list[ResponseInputParam] = []
 
         for msg in messages:

{agenthub_python-0.4.9/agenthub/gemini3_7 → agenthub_python-0.4.11/agenthub/gemini3_8}/__init__.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .client import Gemini3_7Client
+from .client import Gemini3_8Client
 
 
-__all__ = ["Gemini3_7Client"]
+__all__ = ["Gemini3_8Client"]

{agenthub_python-0.4.9/agenthub/gemini3_7 → agenthub_python-0.4.11/agenthub/gemini3_8}/client.py

@@ -61,14 +61,14 @@ def _split_function_response_runs(parts: list[types.Part]) -> list[list[types.Pa
     return runs if runs else [parts]
 
 
-class Gemini3_7Client(LLMClient):
-    """Unified client for the Gemini family, named for the newest generation it serves (3.7).
-
-    It serves every generateContent model generation (3.7 back through 3.x text, image, TTS,
-    and embedding models, with the 2.5 series reachable via an explicit client_type). The API
-    deprecated the temperature/top_p/top_k sampling parameters starting with the 3.6
-    generation (silently ignored today, HTTP 400 in future generations), and this client
-    applies that contract to the whole family: temperature is rejected everywhere.
+class Gemini3_8Client(LLMClient):
+    """Unified client for the Gemini family, named for the newest generation it serves (3.8).
+
+    It serves every generateContent model generation (3.8 back through 3.x text, image, TTS,
+    and embedding models). The API deprecated the temperature/top_p/top_k sampling parameters
+    starting with the 3.6 generation (silently ignored today, HTTP 400 in future generations),
+    and this client applies that contract to the whole family: temperature is rejected
+    everywhere.
     """
 
     def __init__(
@@ -78,7 +78,7 @@ class Gemini3_7Client(LLMClient):
         base_url: str | None = None,
         default_headers: dict[str, str] | None = None,
     ):
-        """Initialize Gemini 3.7 client with model and API key."""
+        """Initialize Gemini 3.8 client with model and API key."""
         self._model = model
         api_key = api_key or os.getenv("GEMINI_API_KEY")
         base_url = base_url or os.getenv("GEMINI_BASE_URL")
@@ -139,15 +139,11 @@ class Gemini3_7Client(LLMClient):
     )
 
     def _supported_thinking_levels(self) -> tuple[types.ThinkingLevel, ...]:
-        """Thinking levels the target model accepts (llmsdk_docs/gemini3_7/docs/thinking.md).
+        """Thinking levels the target model accepts (llmsdk_docs/gemini3_8/docs/thinking.md).
 
         An empty tuple means the model rejects the thinking_level parameter
         entirely, so it must be omitted from the request.
         """
-        if "gemini-2.5" in self._model:
-            # The vendor table claims low/medium/high, but the live API rejects
-            # every thinking_level value for the 2.5 series (verified 2026-07-24).
-            return ()
         if "-image" in self._model:
             return (types.ThinkingLevel.MINIMAL, types.ThinkingLevel.HIGH)
         if "gemini-3-pro" in self._model:
@@ -159,8 +155,9 @@ class Gemini3_7Client(LLMClient):
             # would have accepted costs a little accuracy, forwarding an
             # unsupported one is a 400).
             return (types.ThinkingLevel.LOW, types.ThinkingLevel.MEDIUM, types.ThinkingLevel.HIGH)
-        if "gemini-3.7" in self._model:
-            # The 3.7 generation rejects "minimal" with a 400 (verified live 2026-08-13).
+        if "gemini-3.7" in self._model or "gemini-3.8" in self._model:
+            # Both generations reject "minimal" with a 400 (3.7 verified live 2026-08-13;
+            # 3.8 documented at ai.google.dev/gemini-api/docs/latest-model).
             return (types.ThinkingLevel.LOW, types.ThinkingLevel.MEDIUM, types.ThinkingLevel.HIGH)
         return self._GEMINI_LEVEL_ORDER
 
@@ -180,8 +177,9 @@ class Gemini3_7Client(LLMClient):
             return None
         supported = self._supported_thinking_levels()
         if not supported:
-            # The model takes no thinking_level at all; drop the parameter and
-            # let the model use its default instead of forwarding a 400.
+            # A model that takes no thinking_level at all has nothing to clamp onto, so the
+            # parameter is omitted rather than turned into a failed request. thinking_summary
+            # is unaffected -- include_thoughts still rides along.
             return None
         if level in supported:
             return level
@@ -281,7 +279,7 @@ class Gemini3_7Client(LLMClient):
             config_params["system_instruction"] = config["system_prompt"]
 
         # include_thoughts asks for thought summaries, but whether generateContent returns any
-        # is model-dependent (llmsdk_docs/gemini3_7/docs/thinking.md)
+        # is model-dependent (llmsdk_docs/gemini3_8/docs/thinking.md)
         thinking_summary = config.get("thinking_summary")
         thinking_level = config.get("thinking_level")
         if thinking_summary is not None or thinking_level is not None:
@@ -570,7 +568,7 @@ class Gemini3_7Client(LLMClient):
 
             for item in event["content_items"]:
                 if item["type"] == "tool_call":
-                    # gemini 3.7 does not support partial tool call, mock a partial tool call event
+                    # the Gemini API does not stream partial tool calls, mock a partial tool call event
                     yield {
                         "role": "assistant",
                         "event_type": "delta",

{agenthub_python-0.4.9 → agenthub_python-0.4.11}/agenthub/glm5_3/client.py

@@ -75,9 +75,11 @@ class GLM5_3Client(LLMClient):
 
         GLM-5.3 accepts only low/high/max and errors on anything else, so the client
         clamps to the closest value; NONE rides on low because 5.3 cannot disable
-        thinking. GLM-5.2 accepts the full vocabulary and maps it server-side
-        (low/medium to high, xhigh to max); NONE disables thinking there instead.
-        Models before 5.2 take no reasoning_effort parameter at all.
+        thinking. Every earlier generation takes the vocabulary unchanged: 5.2 maps it
+        server-side (low/medium to high, xhigh to max), and 5.1 and below accept the
+        parameter and ignore it (verified live 2026-09-03 on Z.AI, OpenRouter and
+        SiliconFlow), so the level is forwarded there rather than dropped. Outside 5.3
+        NONE disables thinking outright, which leaves no effort to send.
         """
         model = self._model.lower()  # provider-hosted ids keep their own casing
         if "glm-5.3" in model:
@@ -90,17 +92,15 @@ class GLM5_3Client(LLMClient):
                 ThinkingLevel.MAX: "max",
             }
             return mapping.get(thinking_level)
-        if "glm-5.2" in model:
-            mapping = {
-                ThinkingLevel.NONE: None,
-                ThinkingLevel.LOW: "low",
-                ThinkingLevel.MEDIUM: "medium",
-                ThinkingLevel.HIGH: "high",
-                ThinkingLevel.XHIGH: "xhigh",
-                ThinkingLevel.MAX: "max",
-            }
-            return mapping.get(thinking_level)
-        return None
+        mapping = {
+            ThinkingLevel.NONE: None,
+            ThinkingLevel.LOW: "low",
+            ThinkingLevel.MEDIUM: "medium",
+            ThinkingLevel.HIGH: "high",
+            ThinkingLevel.XHIGH: "xhigh",
+            ThinkingLevel.MAX: "max",
+        }
+        return mapping.get(thinking_level)
 
     def _convert_tool_choice(self, tool_choice: ToolChoice) -> str:
         """Convert ToolChoice to OpenAI's tool_choice format."""

{agenthub_python-0.4.9/agenthub/gpt5_6 → agenthub_python-0.4.11/agenthub/gpt6}/__init__.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .client import GPT5_6Client
+from .client import GPT6Client
 
 
-__all__ = ["GPT5_6Client"]
+__all__ = ["GPT6Client"]

{agenthub_python-0.4.9/agenthub/gpt5_6 → agenthub_python-0.4.11/agenthub/gpt6}/client.py

@@ -36,8 +36,8 @@ from ..types import (
 from ..utils import is_debug_enabled, openai_image_detail
 
 
-class GPT5_6Client(LLMClient):
-    """GPT-5.6-specific LLM client implementation (also serves GPT-5.4 and GPT-5.5)."""
+class GPT6Client(LLMClient):
+    """GPT-6-specific LLM client implementation (also serves GPT-5.6, GPT-5.5 and GPT-5.4)."""
 
     def __init__(
         self,
@@ -46,7 +46,7 @@ class GPT5_6Client(LLMClient):
         base_url: str | None = None,
         default_headers: dict[str, str] | None = None,
     ):
-        """Initialize GPT-5.6 client with model and API key."""
+        """Initialize GPT-6 client with model and API key."""
         self._model = model
         api_key = api_key or os.getenv("OPENAI_API_KEY")
         base_url = base_url or os.getenv("OPENAI_BASE_URL")
@@ -55,6 +55,13 @@ class GPT5_6Client(LLMClient):
 
     def _convert_thinking_level_to_effort(self, thinking_level: ThinkingLevel) -> str:
         """Convert ThinkingLevel enum to OpenAI's reasoning effort."""
+        if thinking_level == ThinkingLevel.NONE and "gpt-6" in self._model:
+            # GPT-6 rejects both "none" and "minimal" with a 400 (verified live 2026-09-09:
+            # "Unsupported value: 'none' is not supported with the 'gpt-6-astra' model.
+            # Supported values are: 'low', 'medium', 'high', 'xhigh', and 'max'."), so NONE
+            # degrades to the lowest effort the generation accepts.
+            return "low"
+
         mapping = {
             ThinkingLevel.NONE: "none",
             ThinkingLevel.LOW: "low",
@@ -101,13 +108,17 @@ class GPT5_6Client(LLMClient):
 
         if config.get("temperature") is not None and config["temperature"] != 1.0:
             raise UnsupportedParameterError(
-                self.__class__.__name__, "temperature", "GPT-5.6 does not support setting temperature."
+                self.__class__.__name__, "temperature", "GPT-6 does not support setting temperature."
             )
 
         if config.get("thinking_level") is not None:
             openai_config["reasoning"] = {"effort": self._convert_thinking_level_to_effort(config["thinking_level"])}
-            if config.get("thinking_summary"):
-                openai_config["reasoning"]["summary"] = "concise"
+
+        if config.get("thinking_summary"):
+            # reasoning.summary stands on its own, with or without an effort (verified live
+            # 2026-09-03 on the OpenAI and OpenRouter endpoints). False needs no key: the
+            # Responses API returns no summary unless one is asked for.
+            openai_config.setdefault("reasoning", {})["summary"] = "concise"
 
         if config.get("tools") is not None:
             openai_config["tools"] = [{"type": "function", **tool} for tool in config["tools"]]
@@ -120,7 +131,7 @@ class GPT5_6Client(LLMClient):
 
         if config.get("prompt_caching") is not None and config["prompt_caching"] != PromptCaching.ENABLE:
             raise UnsupportedParameterError(
-                self.__class__.__name__, "prompt_caching", "prompt_caching must be ENABLE for GPT-5.6."
+                self.__class__.__name__, "prompt_caching", "prompt_caching must be ENABLE for GPT-6."
             )
 
         return openai_config
@@ -302,7 +313,12 @@ class GPT5_6Client(LLMClient):
                 # the completed item carries the canonical wire fields to send back on the
                 # next turn (identical to the response.completed copy, but adjacent to the
                 # thinking deltas so the fidelity lands on the item that carried the text);
-                # record the channel plus the fields the server demands back
+                # record the channel plus the fields the server demands back. This event is the
+                # only source of encrypted_content, because the streaming-events reference says
+                # of response.output_item.added: "For reasoning items, encrypted_content may be
+                # incomplete while the item is in progress. Use the reasoning item from the
+                # corresponding response.output_item.done event when passing it as input to a
+                # subsequent request."
                 event_type = "delta"
                 fidelity = {}
                 if getattr(model_output.item, "summary", None):

{agenthub_python-0.4.9 → agenthub_python-0.4.11}/agenthub/openai_chat/client.py

@@ -173,7 +173,7 @@ class OpenaiChatClient(LLMClient):
                     if "tool_call_id" not in item:
                         raise ValueError("tool_call_id is required for tool result.")
 
-                    content = [{"type": "text", "text": item["text"]}]
+                    image_parts = []
 
                     if "images" in item and item["images"]:
                         for image_url in item["images"]:
@@ -182,7 +182,12 @@ class OpenaiChatClient(LLMClient):
                                 # siliconflow does not support image_url in tool result
                                 content_parts.append(part)
                             else:
-                                content.append(part)
+                                image_parts.append(part)
+
+                    # a plain string is the form every OpenAI-compatible server accepts for a text
+                    # result; the content-part list is reserved for results carrying images, which
+                    # only servers with multimodal tool messages take
+                    content = [{"type": "text", "text": item["text"]}, *image_parts] if image_parts else item["text"]
 
                     # Tool results are sent as separate messages
                     openai_messages.append(

agenthub_python-0.4.11/agenthub/openai_chat_vllm_adapter/__init__.py

@@ -0,0 +1,4 @@
+from .client import OpenaiChatVllmAdapterClient
+
+
+__all__ = ["OpenaiChatVllmAdapterClient"]

agenthub_python-0.4.11/agenthub/openai_chat_vllm_adapter/client.py

@@ -0,0 +1,122 @@
+# Copyright 2025 Prism Shadow. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any
+
+from ..openai_chat import OpenaiChatClient
+from ..types import ThinkingLevel, UniConfig
+
+
+# vLLM passes chat_template_kwargs straight to the served model's chat template, so the
+# switch that turns thinking on is whatever that template happens to read. Each profile
+# below maps an AgentHub level onto one family's kwargs; an empty mapping means the
+# request carries no chat_template_kwargs at all.
+#
+# The upstream artifacts these profiles are read off, and the clamping those artifacts
+# force, are snapshotted in llmsdk_docs/openai_chat_vllm_adapter/. Update that snapshot
+# whenever a model is added here.
+
+# Qwen3 templates read a single enable_thinking boolean and no effort key at all.
+_QWEN3_THINKING: dict[ThinkingLevel, dict[str, Any]] = {
+    ThinkingLevel.NONE: {"enable_thinking": False},
+    ThinkingLevel.LOW: {"enable_thinking": True},
+    ThinkingLevel.MEDIUM: {"enable_thinking": True},
+    ThinkingLevel.HIGH: {"enable_thinking": True},
+    ThinkingLevel.XHIGH: {"enable_thinking": True},
+    ThinkingLevel.MAX: {"enable_thinking": True},
+}
+
+# Qwen3.8-27B and Qwen3.8-Flash-Next ship the same chat template, byte for byte, so they
+# share a profile. It keeps enable_thinking as the off switch and takes its adaptive modes
+# as reasoning_effort, validated against low/medium/xhigh, so high and max clamp to xhigh.
+# The template defaults the key to xhigh, so a model on this template that is not sent the
+# key runs every level at full effort.
+_QWEN3_8_THINKING: dict[ThinkingLevel, dict[str, Any]] = {
+    ThinkingLevel.NONE: {"enable_thinking": False},
+    ThinkingLevel.LOW: {"reasoning_effort": "low"},
+    ThinkingLevel.MEDIUM: {"reasoning_effort": "medium"},
+    ThinkingLevel.HIGH: {"reasoning_effort": "xhigh"},
+    ThinkingLevel.XHIGH: {"reasoning_effort": "xhigh"},
+    ThinkingLevel.MAX: {"reasoning_effort": "xhigh"},
+}
+
+# DeepSeek V4 publishes no chat template; vLLM reads a thinking flag paired with
+# reasoning_effort, and thinking is off whenever the flag is absent, which is what NONE
+# sends. DeepSeek-V4-Pro and DeepSeek-V4-Flash share an encoding module that asserts
+# reasoning_effort in ['max', None, 'high'], so low is a failed request rather than a
+# weaker answer and high is the lowest value they take. That module then branches on 'max'
+# alone, which means LOW through XHIGH all render the same prompt on these two models.
+_DEEPSEEK_V4_PRO_FLASH_THINKING: dict[ThinkingLevel, dict[str, Any]] = {
+    ThinkingLevel.NONE: {},
+    ThinkingLevel.LOW: {"thinking": True, "reasoning_effort": "high"},
+    ThinkingLevel.MEDIUM: {"thinking": True, "reasoning_effort": "high"},
+    ThinkingLevel.HIGH: {"thinking": True, "reasoning_effort": "high"},
+    ThinkingLevel.XHIGH: {"thinking": True, "reasoning_effort": "high"},
+    ThinkingLevel.MAX: {"thinking": True, "reasoning_effort": "max"},
+}
+
+# DeepSeek-V4-Flash-Vision-Exp ships a different copy of that encoding module, one that
+# validates reasoning_effort against a low/high/max table, so it keeps the finer scale;
+# medium and xhigh clamp to high.
+_DEEPSEEK_V4_VISION_EXP_THINKING: dict[ThinkingLevel, dict[str, Any]] = {
+    ThinkingLevel.NONE: {},
+    ThinkingLevel.LOW: {"thinking": True, "reasoning_effort": "low"},
+    ThinkingLevel.MEDIUM: {"thinking": True, "reasoning_effort": "high"},
+    ThinkingLevel.HIGH: {"thinking": True, "reasoning_effort": "high"},
+    ThinkingLevel.XHIGH: {"thinking": True, "reasoning_effort": "high"},
+    ThinkingLevel.MAX: {"thinking": True, "reasoning_effort": "max"},
+}
+
+# Keys are matched as substrings of the lowercased model id, so a served id keeps whatever
+# prefix the deployment gave it (Qwen/Qwen3.6-35B-A3B, deepseek-ai/DeepSeek-V4-Pro). The
+# first match wins, so a key that contains another must come first: deepseek-v4-flash is a
+# prefix of deepseek-v4-flash-vision-exp.
+_MODEL_THINKING_PROFILES: tuple[tuple[str, dict[ThinkingLevel, dict[str, Any]]], ...] = (
+    ("qwen3.8-flash-next", _QWEN3_8_THINKING),
+    ("qwen3.8-27b", _QWEN3_8_THINKING),
+    ("qwen3.6-35b-a3b", _QWEN3_THINKING),
+    ("qwen3.5-0.8b", _QWEN3_THINKING),
+    ("qwen3.5-9b", _QWEN3_THINKING),
+    ("deepseek-v4-flash-vision-exp", _DEEPSEEK_V4_VISION_EXP_THINKING),
+    ("deepseek-v4-pro", _DEEPSEEK_V4_PRO_FLASH_THINKING),
+    ("deepseek-v4-flash", _DEEPSEEK_V4_PRO_FLASH_THINKING),
+)
+
+
+class OpenaiChatVllmAdapterClient(OpenaiChatClient):
+    """Models served through vLLM's OpenAI-compatible Chat Completions API."""
+
+    def _thinking_chat_template_kwargs(self, thinking_level: ThinkingLevel) -> dict[str, Any]:
+        """Return the chat_template_kwargs this model's template reads for the level.
+
+        A model outside the table falls back to Qwen3's enable_thinking, the most
+        widespread of the conventions and inert on a template that ignores the key.
+        """
+        model = self._model.lower()
+        for name, profile in _MODEL_THINKING_PROFILES:
+            if name in model:
+                return dict(profile[thinking_level])
+
+        return dict(_QWEN3_THINKING[thinking_level])
+
+    def transform_uni_config_to_model_config(self, config: UniConfig) -> dict[str, Any]:
+        """Map AgentHub's level onto the thinking switch this model's chat template reads."""
+        vllm_config = super().transform_uni_config_to_model_config(config)
+
+        if config.get("thinking_level") is not None:
+            chat_template_kwargs = self._thinking_chat_template_kwargs(config["thinking_level"])
+            if chat_template_kwargs:
+                vllm_config["chat_template_kwargs"] = chat_template_kwargs
+
+        return vllm_config

{agenthub_python-0.4.9 → agenthub_python-0.4.11}/agenthub/openai_responses/client.py

@@ -55,6 +55,12 @@ class OpenaiResponsesClient(LLMClient):
 
     def _convert_thinking_level_to_effort(self, thinking_level: ThinkingLevel) -> str:
         """Convert ThinkingLevel enum to the Responses API reasoning effort."""
+        if thinking_level == ThinkingLevel.NONE and "gpt-6" in self._model:
+            # a gateway serving GPT-6 forwards the effort to OpenAI, which rejects "none" and
+            # "minimal" with a 400 (verified live 2026-09-09 against api.openai.com), so NONE
+            # degrades to the lowest effort the generation accepts.
+            return "low"
+
         mapping = {
             ThinkingLevel.NONE: "none",
             ThinkingLevel.LOW: "low",
@@ -100,6 +106,11 @@ class OpenaiResponsesClient(LLMClient):
         if config.get("temperature") is not None:
             openai_config["temperature"] = config["temperature"]
 
+        # Unlike the model-specific Responses clients, the summary stays inside this branch:
+        # OpenRouter reads a reasoning object carrying no effort as "reasoning disabled" and
+        # refuses it on a forced-thinking model -- "Reasoning is mandatory for this endpoint
+        # and cannot be disabled" (400, verified live 2026-09-03 with z-ai/glm-5.3) -- so a
+        # summary sent on its own would turn a dropped value into a failed request.
         if config.get("thinking_level") is not None:
             openai_config["reasoning"] = {"effort": self._convert_thinking_level_to_effort(config["thinking_level"])}
             if config.get("thinking_summary"):
@@ -197,13 +208,20 @@ class OpenaiResponsesClient(LLMClient):
                         raise ValueError("tool_call_id is required for tool result.")
 
                     # NOTE: tool results are input items
-                    tool_result = [{"type": "input_text", "text": item["text"]}]
+                    image_parts = []
                     if "images" in item:
                         for image_url in item["images"]:
-                            tool_result.append(self._convert_image_url(image_url))
+                            image_parts.append(self._convert_image_url(image_url))
+
+                    # a plain string is the form every OpenAI-compatible server accepts for a text
+                    # result; the content-part list is reserved for results carrying images, which
+                    # only servers with multimodal tool messages take
+                    output = (
+                        [{"type": "input_text", "text": item["text"]}, *image_parts] if image_parts else item["text"]
+                    )
 
                     input_list.append(
-                        {"type": "function_call_output", "call_id": item["tool_call_id"], "output": tool_result}
+                        {"type": "function_call_output", "call_id": item["tool_call_id"], "output": output}
                     )
                 else:
                     raise ValueError(f"Unknown item: {item}")

{agenthub_python-0.4.9 → agenthub_python-0.4.11}/agenthub/registry.py

@@ -42,6 +42,9 @@ class SupportedModel(TypedDict):
     ``AutoLLMClient(model=entry["model"], base_url=entry["base_url"], client_type=entry["client"])``.
     Modalities describe what is usable through that client; ``context_window`` and
     ``pricing`` are omitted where the platform publishes no authoritative value.
+
+    ``pricing`` is always the LIST price. A running promotion is deliberately not recorded:
+    the registry's job is the catalog price, and applying a promotion is the consumer's.
     """
 
     model: str
@@ -95,20 +98,31 @@ def _cny(prompt: float, output: float, cached: float | None = None) -> ModelPric
 # /models APIs on 2026-07-22, SiliconFlow CNY prices from the vendors' official price lists.
 _SUPPORTED_MODELS: list[SupportedModel] = [
     # official vendor endpoints
+    {
+        "model": "gemini-3.8-flash",
+        "base_url": _GOOGLE,
+        "client": "gemini-3.8",
+        "input_modalities": ["Text", "Image", "Video", "Audio"],
+        "output_modalities": ["Text"],
+        "context_window": 1048576,
+        # Google runs a launch discount through 2026-12-31 on this row and on the two flash
+        # rows below; the list price is stored regardless, because applying a running
+        # promotion belongs to the consumer, not to the registry.
+        "pricing": _usd(1.5, 7.5, cached=0.15),
+    },
     {
         "model": "gemini-3.7-flash",
         "base_url": _GOOGLE,
-        "client": "gemini-3.7",
+        "client": "gemini-3.8",
         "input_modalities": ["Text", "Image", "Video", "Audio"],
         "output_modalities": ["Text"],
         "context_window": 1048576,
-        # official list price; a launch discount halves all three rates through 2026-12-31
         "pricing": _usd(1.5, 7.5, cached=0.15),
     },
     {
         "model": "gemini-3.6-flash",
         "base_url": _GOOGLE,
-        "client": "gemini-3.7",
+        "client": "gemini-3.8",
         "input_modalities": ["Text", "Image", "Video", "Audio"],
         "output_modalities": ["Text"],
         "context_window": 1048576,
@@ -117,7 +131,7 @@ _SUPPORTED_MODELS: list[SupportedModel] = [
     {
         "model": "gemini-3.5-flash-lite",
         "base_url": _GOOGLE,
-        "client": "gemini-3.7",
+        "client": "gemini-3.8",
         "input_modalities": ["Text", "Image", "Video", "Audio"],
         "output_modalities": ["Text"],
         "context_window": 1048576,
@@ -126,7 +140,7 @@ _SUPPORTED_MODELS: list[SupportedModel] = [
     {
         "model": "gemini-3.5-flash",
         "base_url": _GOOGLE,
-        "client": "gemini-3.7",
+        "client": "gemini-3.8",
         "input_modalities": ["Text", "Image", "Video", "Audio"],
         "output_modalities": ["Text"],
         "context_window": 1048576,
@@ -135,21 +149,21 @@ _SUPPORTED_MODELS: list[SupportedModel] = [
     {
         "model": "gemini-3.1-flash-image",
         "base_url": _GOOGLE,
-        "client": "gemini-3.7",
+        "client": "gemini-3.8",
         "input_modalities": ["Text", "Image"],
         "output_modalities": ["Image"],
     },
     {
         "model": "gemini-3.1-flash-tts-preview",
         "base_url": _GOOGLE,
-        "client": "gemini-3.7",
+        "client": "gemini-3.8",
         "input_modalities": ["Text"],
         "output_modalities": ["Audio"],
     },
     {
         "model": "gemini-embedding-2",
         "base_url": _GOOGLE,
-        "client": "gemini-3.7",
+        "client": "gemini-3.8",
         "input_modalities": ["Text"],
         "output_modalities": ["Embed"],
     },
@@ -198,11 +212,24 @@ _SUPPORTED_MODELS: list[SupportedModel] = [
         "context_window": 1000000,
         "pricing": _usd(3.0, 15.0, cached=0.3),
     },
+    {
+        # official list price per 1M tokens: $1 cached input, $10 uncached input, $12.5 cache
+        # writes, $50 output, all doubled above 272K input tokens except output, which is 1.5x.
+        # The prompt bucket carries the cache-write rate: the usage buckets do not separate
+        # cache-written input from plain input.
+        "model": "gpt-6-astra",
+        "base_url": _OPENAI,
+        "client": "gpt-6",
+        "input_modalities": ["Text", "Image"],
+        "output_modalities": ["Text"],
+        "context_window": 1050000,
+        "pricing": _usd(12.5, 50.0, cached=1.0),
+    },
     {
         # official standard-tier list price; the bare gpt-5.6 alias also routes here
         "model": "gpt-5.6-sol",
         "base_url": _OPENAI,
-        "client": "gpt-5.6",
+        "client": "gpt-6",
         "input_modalities": ["Text", "Image"],
         "output_modalities": ["Text"],
         "context_window": 1050000,
@@ -211,7 +238,7 @@ _SUPPORTED_MODELS: list[SupportedModel] = [
     {
         "model": "gpt-5.6-terra",
         "base_url": _OPENAI,
-        "client": "gpt-5.6",
+        "client": "gpt-6",
         "input_modalities": ["Text", "Image"],
         "output_modalities": ["Text"],
         "context_window": 1050000,
@@ -220,7 +247,7 @@ _SUPPORTED_MODELS: list[SupportedModel] = [
     {
         "model": "gpt-5.6-luna",
         "base_url": _OPENAI,
-        "client": "gpt-5.6",
+        "client": "gpt-6",
         "input_modalities": ["Text", "Image"],
         "output_modalities": ["Text"],
         "context_window": 1050000,
@@ -309,6 +336,20 @@ _SUPPORTED_MODELS: list[SupportedModel] = [
         "context_window": 262144,
         "pricing": _cny(6.5, 27.0, cached=1.1),
     },
+    {
+        "model": "deepseek-v4.1-flash",
+        "base_url": _DEEPSEEK,
+        "client": "deepseek-v4",
+        # announced by DeepSeek for release after 2026-09-10 and not yet served on
+        # 2026-09-09; multimodal per the announcement
+        "input_modalities": ["Text", "Image"],
+        "output_modalities": ["Text"],
+        # assumed equal to deepseek-v4-flash until the official model page lists it
+        "context_window": 1000000,
+        # priced as the V4 Flash series: official off-peak list price effective 2026-09-10,
+        # and peak-hour rates (Beijing 9:00-12:00, 14:00-18:00) are double
+        "pricing": _cny(1.0, 4.0, cached=0.02),
+    },
     {
         "model": "deepseek-v4-flash",
         "base_url": _DEEPSEEK,
@@ -316,9 +357,10 @@ _SUPPORTED_MODELS: list[SupportedModel] = [
         "input_modalities": ["Text"],
         "output_modalities": ["Text"],
         "context_window": 1000000,
-        # official off-peak list price (verified 2026-08-18); peak-hour rates
-        # (Beijing 9:00-12:00, 14:00-18:00) are double
-        "pricing": _cny(1.5, 4.5, cached=0.05),
+        # official off-peak list price effective 2026-09-10 (verified 2026-09-08 against
+        # the official adjustment); peak-hour rates (Beijing 9:00-12:00, 14:00-18:00) are
+        # double
+        "pricing": _cny(1.0, 4.0, cached=0.02),
     },
     {
         "model": "deepseek-v4-flash-vision-exp",
@@ -327,9 +369,10 @@ _SUPPORTED_MODELS: list[SupportedModel] = [
         "input_modalities": ["Text", "Image"],
         "output_modalities": ["Text"],
         "context_window": 1000000,
-        # priced as deepseek-v4-flash; official off-peak list price (verified 2026-08-21), and
-        # peak-hour rates (Beijing 9:00-12:00, 14:00-18:00) are double
-        "pricing": _cny(1.5, 4.5, cached=0.05),
+        # priced as deepseek-v4-flash; official off-peak list price effective 2026-09-10
+        # (verified 2026-09-08 against the official adjustment), and peak-hour rates
+        # (Beijing 9:00-12:00, 14:00-18:00) are double
+        "pricing": _cny(1.0, 4.0, cached=0.02),
     },
     {
         "model": "deepseek-v4-pro",
@@ -451,6 +494,17 @@ _SUPPORTED_MODELS: list[SupportedModel] = [
         "context_window": 1000000,
         "pricing": _usd(0.0, 0.0),
     },
+    {
+        # models API 2026-09-09, default OpenAI endpoint, no discount: $10 input, $50 output,
+        # $1 cache read, $12.5 cache write per 1M tokens
+        "model": "openai/gpt-6-astra",
+        "base_url": _OPENROUTER,
+        "client": "openai-responses",
+        "input_modalities": ["Text", "Image"],
+        "output_modalities": ["Text"],
+        "context_window": 1050000,
+        "pricing": _usd(12.5, 50.0, cached=1.0),
+    },
     {
         "model": "openai/gpt-5.6-sol",
         "base_url": _OPENROUTER,

{agenthub_python-0.4.9 → agenthub_python-0.4.11}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "agenthub-python"
-version = "0.4.9"
+version = "0.4.11"
 description = "AgentHub is the LLM API Hub for the Agent era, built for high-precision autonomous agents."
 keywords = [
     "agent",

{agenthub_python-0.4.9 → agenthub_python-0.4.11}/pyproject.toml.orig

@@ -1,6 +1,6 @@
 [project]
 name = "agenthub-python"
-version = "0.4.9"
+version = "0.4.11"
 description = "AgentHub is the LLM API Hub for the Agent era, built for high-precision autonomous agents."
 keywords = ["agent", "llm", "gemini", "claude", "gpt"]
 readme = "README.md"